text
stringlengths
2.85k
2.55M
label
class label
11 classes
Acceleration of Mean Square Distance Calculations with Floating Close Structure in Metadynamics Simulations arXiv:1801.02362v1 [cs.CE] 8 Jan 2018 Jana Pazúrikováa,∗, Jaroslav Oľhaa , Aleš Křeneka , Vojtěch Spiwokb a Institute of Computer Science, Masaryk University, Botanická 554/68a, 602 00 Brno, Czech Republic b Department of Biochemistry and Microbiology, University of Chemistry and Technology, Technická 5, 166 28 Prague 6, Czech Republic Abstract Molecular dynamics simulates the movements of atoms. Due to its high cost, many methods have been developed to “push the simulation forward”. One of them, metadynamics, can hasten the molecular dynamics with the help of variables describing the simulated process. However, the evaluation of these variables can include numerous mean square distance calculations that introduce substantial computational demands, thus jeopardize the benefit of the approach. Recently, we proposed an approximative method that significantly reduces the number of these distance calculations. Here we evaluate the performance and the scalability on two molecular systems. We assess the maximal theoretical speed-up based on the reduction of distance computations and Ahmdal’s law and compare it to the practical speed-up achieved with our implementation. Keywords: approximation, acceleration, molecular dynamics simulation, metadynamics 1. Introduction Molecular dynamics [1] is a well-known computational method which simulates the behavior of a molecule at the atomic scale by integrating Newton’s ∗ Corresponding author Email address: [email protected] (Jana Pazúriková) 1 equations of motion for all atoms over time. Due to a high frequency of movements at this scale, the integration step must not exceed femtoseconds, which yields realistic computational speed to be at most nanoseconds of simulated time per day of computation time. On the other hand, biologically relevant phenomena take milliseconds and more at the macromolecular level, hence their complete simulations are still not feasible nowadays. The necessity of long simulation timescales can be traced down to high energy barriers between the distinct states of the macromolecule, which require a considerable momentum to be gained at the nanoscale before they are overcome. However, various classes of problems do not require a realistic trace of the simulation, it is sufficient to explore the state space in which the simulation would travel eventually [2]. Therefore, the simulation run can be “pushed forward” by additional steering, e.g. in the method called metadynamics by adding bias potential in already visited areas, or by “filling energy minima with computational sand”, as illustrated in Figure 1. Figure 1: Filling potential energy minima with a bias potential. The simulation starts with x = 0, and it accumulates the bias potential in this area first. After 20 steps the energy barrier at x = −1.75 is overcome and the bias potential starts filling the minimum at x = −4. After approx. 300 steps the state space is explored. Taken from [3], image copyright (2002) National Academy of Sciences. 2 Many methods have been developed to accelerate molecular dynamics simulations by artificial intervensions [4]. Unfortunately, the most promising ones often introduce a non-negligible computational overhead to calculate the bias potential, hence jeopardize the benefit of the whole approach. Recently, we proposed a method [5] which considerably reduces the number of required calls to the most expensive routine of such calculation—mean square distance (MSD) between two molecular shapes, while not affecting the overall accuracy. Here we introduce the ready-to-use implementation of the method (replacing a rather inefficient prototype from [5]) and evaluate the practical benefits—speed-up of the real computation, including the use of multiple CPU cores. 2. Background Simulations of molecular dynamics (MD) calculate the movements of atoms caused by their interactions [1]. First, the potential energy of the system has to be computed. In this work, we modeled the interatomic potential using molecular mechanics model, considering atoms as mass charged points and approximating both bonded and nonbonded interactions with a series of empirical functions and parameters. Second, the forces (determined by the potential) are integrated with Newton’s law of motion to obtain the positions of atoms in the next step. Computational demands of molecular dynamics grow with the number of atoms and simulation steps. The calculation of the interatomic potential scales as O(Natoms log(Natoms )) with common methods. The parallelization through the spatial domain makes it possible to simulate even millions of atoms. On the other hand, the timescales of simulations are more difficult to handle. A tiny timestep of the integration scheme (in the order of 10−15 s) makes it hardly feasible to simulate more than a few microseconds [6], a millisecond at most [7], although many biologically or chemically relevant processes take longer than that. We need longer simulation times because we need to explore the whole energy landscape, ideally all possible states of the molecular system. However, molecules tend to stay in the local energy minima and they cross the high-energy barrier between them only once in a while. Many methods have been developed to accelerate the rare events of those crossings, for an overview see [8, 4, 9]. They enhance the sampling of the energy landscape and explore it much faster using various artificial interven3 tions to the simulation. One of the most common, metadynamics [3] (MTD), fills the minima in the energy landscape with an artifical bias potential. It “spreads the sand into the valleys”, and thus makes it easier to cross the peaks. To properly function, metadynamics needs a few collective variables (CV)— preselected degrees of freedom that differentiate between the states of the simulated process and encompass all the relevant motions. Moreover, they have to be limited in number, otherwise, both the performance and the accuracy of metadynamics will decline [10, 11]. Collective variables are usually computed from the coordinates of the molecular system at the given time, further referred to as the current structure or x. Metadynamics accumulates a bias potential in the form of Gaussian hills centered in the values of collective variables, see Equation (1). In Figure 1, the sum of Gaussian hills forms curvy lines depicting the borders of the summed bias potential in various time points. Well-chosen CVs direct to the right spot for “the sand to be spread on”. The bias potential expresses as the bias force, so collective variables need to be differentiable with respect to coordinates. Vbias (S(x), t) = X Y t0 =τG ,2τG ,... t<t0 i   0 (Si (x(t)) − sti )2 w exp − 2σ 2 (1) Fbias (t) = ∂Vbias (S(x), t) ∂x where x are coordinates of the current structure, S(x) is the vector of functions computing collective variables; τG is the frequency of Gaussian hills 0 addition; w and σ is the height and the width of the Gaussian hill; sti is the value of i-th CV at time t0 , i.e. Si (x(t0 )). The choice or design of CVs requires physico-chemical expertise and naturally gets more complicated with complex processes. A distance between specific atoms, angles or their combinations work well for simple processes. More complex ones, like protein folding, require a more sophisticated approach such as path CVs [12] or their generalization, property map CVs [13]. They are based on comparison to reference landmark structures, a series of snapshots capturing various states of the process. Reference structures can be generated with any non-continuous simulations, non-physical actions, high temperatures, annealing techniques or even manual construction in a visualization software. 4 The value of the property map collective variable corresponds to a weighted average of properties of the closest reference structures, where the weight is determined by the distance, see Equation (2). For example, if we take an index of the reference structure as a property, the value of CV would be close to the index of the closest reference structure. N P S(x) = qi exp (−λD(x, ai )) i=1 N P (2) exp (−λD(x, ai )) i=1 where N is the number of reference structures, ai are coordinates of i-th reference structure, qi is an arbitrary property of the given reference structure, D() is the distance function, λ is a tuning parameter. Clearly, the evaluation of property map CVs requires as many distance computations in every step of the simulation as there are reference structures. For smooth processes, we can reduce the number of distance computations by employing a neighbourlist (NL). Since the contribution to a CV’s value decreases exponentially with the distance, without sudden changes in system’s structure it suffices to compute distance only to the few closest reference structures in the NL and update the list regularly. However, in processes with abrupt changes, we need to compute the distances of the current structure to all the reference structures in each step (referred to as no neighbourlist). In both cases, numerous distance computations introduce high computational cost. 3. Floating Close Structure The standard distance measure between molecular systems in computational sciences is the (root) mean square distance between superimposed, fitted structures [14, ch.16]. The fitting between structures minimizes the distance and consists of two tasks: translation and rotation. Shifting both structures to the same location in the coordinate system means subtracting coordinates of the center of mass from coordinates of all atoms. As this is both simple and fast, we omit it from further equations for simplicity and assume that it was done in advance for all the structures. On the contrary, obtaining the rotation matrix presents the most expensive task in the distance computation. It involves demanding matrix operations, such as the 5 matrix diagonalization from quaternions as in [15, 16] or the singular value decomposition as in [17]. To get a closer look on the bottlenecks, we analyzed the performance of simulations with a profiling software. We ran simulations using Gromacs [18] as the molecular dynamics software and Plumed [19] as the metadynamics tool, details in section 5. Even with the neighbourlist employed, distance computation takes up to 43% of miniprotein Trp-cage simulation walltime. For small 24-atom cyclooctane simulation without the neighbourlist, it takes 93% of walltime, see Table 1. The majority of that time is spent in BLAS function DSYEVR, diagonalizing matrix in Kearsley’s method [15] implemented in Plumed. As BLAS is already a finely-tuned routine, we assume no further code optimizations would bring a substantial improvement. Offloading to GPU brings its own issues, as we outlined in [20]. Therefore, we aim to modify the entire method in a way that reduces the number of demanding distance computations and thus decreases the computational cost. 3.1. Exact MSD Calculation The original method for exact MSD computation calculates accurately but expensively the rotation matrix for all the reference structures in the neighbourlist in every step, see Algorithm 1 and Figure 2. Then it uses rotation matrices to compute the distance, see Equation (3), and derivatives with respect to coordinates (to evaluate the bias force), see Equation (4). x ai Figure 2: Original method expensively computes distances from the current structure (x) to reference structures in neighbourlist (ai ) in every step. Arrows denote passing simulation time. 6 Algorithm 1 Computation of the collective variables in the original method 1: function compute_cvs(x, ref_structures) 2: loop through ai in neighlist 3: Rxai ← compute_R(x, ai ) 4: distance ← Equation (3) 5: derivatives ← Equation (4) D(x, ai ) = NX atoms wj kdj k2 (3) j where dj = xj − Rai x aij ∂D(x, ai ) = 2wk dk + wk0 ∂xk ! X −2wj dj ! X −2wj dj × aij j + j ⊗ ∂Rai x ∂xk (4) where x are coordinates of the current structure, ai are coordinates of i-th reference structure, Rai x is the rotation matrix fitting the reference structure ai onto the current structure x, w and w0 are weights, ⊗ is element-wise multiplication. Two sets of weights influence the computation: displacement w that describes the contribution of an atom’s displacement to the total distance, and alignment w0 that describes the contribution to the center of mass which influences the translation. During a usual step, we need to compute the expensive rotation matrix M times, where M is the neighbourlist size. During an update of NL, every L steps, we calculate the distance (thus also the rotation matrix) to all the reference structures. More details about the computational cost in section 6.2. 3.2. Approximative MSD Calculation Our method is based on the assumption that the molecular system moves reasonably smoothly in dynamics simulations due to a tiny timestep. As its 7 coordinates alter only slowly, the values in the rotation matrices should also change only slightly every step. Therefore, we reuse them in a few consecutive steps, until the current structure moves too far. At the beginning of the simulation, we start as usual, see Figure 3 and Algorithm 2. We assign current coordinates to a close structure y and save computed rotation matrices Rai y . In most steps, we reuse these saved matrices to approximate Rai x ≈ Rxy Rai y while the current structure x and the close structure y stay close to each other. If the distance D(x, y) exceeds a given threshold ε, we reassign the close structure, y ← x and recompute the matrices Rai y . Algorithm 2 Approximative computation of the collective variables with the close structure 1: function compute_cvs(x, ref_structures) 2: Rxy ← compute_R(x,y) 3: Dxy ← Equation (3) 4: ∂Rxy /∂xk ← Equation(4) 5: if Dxy > ε then 6: loop through all ai 7: Rai x ← compute_R(x, ai ) 8: save Rai x 9: distance ← Equation (3) 10: derivatives ← Equation (4) 11: else 12: loop through ai in neighlist 13: Rai x ← Rxy Rai y 14: distance ← Equation (5) 15: derivatives ← Equation (6) We use exact MSD calculation, Equations (3) and (4), in the reassignment steps and for the computation of D(x, y). In the majority of steps, we use Equation (5) for distance and Equation (6) for derivatives. Rxy Rai y approximates Rai x , omitting the demanding accurate computation of the rotation matrix. 8 D(x,y) > y⬅x y⬅x x ai Figure 3: Close structure method expensively computes distances only between the current (x) and the close structure (y) and then during reassignments when the distances to all reference structures (ai ) are accurately calculated. Arrows denote passing simulation time. D̃(x, ai ) = NX atoms wj kd̃j k2 (5) j d̃j = xj − Rxy Rai y aij ∂ D̃(x, ai ) = 2wk d˜k + wk0 ∂xk ! X −2wj d̃j j ! X −2wj d̃j × Rai y aij j + (6) ∂Rxy ⊗ ∂xk where y is the close structure, Rxy is the rotation matrix that fits the close structure to the current structure, Rai y is the saved rotation matrix that fits the close structure to the reference structure ai . The neighbourlist technique, not detailed in Algorithm 2 for simplicity, can be applied as usual in conjunction with the close structure method. However, we do not recompute the distance to all reference structures in the NL update step, as it is in the original method. Instead, we recompute during the reassignment of the close structure. Therefore, in the close structure method, we select the structures during the NL update according to the approximated distance. 9 Close structure approximation reduces the number of distance computations in a usual step to one—evaluating D(x, y). When reassigning the close structure, we expensively calculate the distance to all N reference structures. More details about the computational cost in section 6.2. 4. Implementation We implemented the close structure method into Plumed [19], a standard tool for metadynamics. Plumed works as a plugin into many molecular dynamics softwares, we used it with Gromacs [18]. We changed two classes, one that represents property map and path CVs, and another that encompasses RMSD computation. Since the prototype implementation of our method, we have dealt with issues regarding code optimization, data structures, memory access patterns, communication through MPI and scalability issues. The code ready for production use is available in the version v2.4 of Plumed, see https://github.com/plumed/plumed2. For performance evaluation, we combined Gromacs v5.1.4 with Plumed v2.3. We applied the same level of minor code optimizations, enforcing the vectorization of loops and inlining of template functions, both on our modified and the original code in parts regarding the evaluation of property map collective variables. We did not include any aggressive optimizations that would transform data structures since such heavy changes to software design would make it difficult to incorporate the code into the official software [21]. We compiled the code with Intel compiler version 17 and SIMD instruction set AVX_256. 5. Computational Details In this paper, we focus on the evaluation of the acceleration and the scalability. We reuse two molecular systems and their simulation setup from [5] where we have thoroughly evaluated the accuracy. Thus, we describe datasets only shortly, for further details see [5]. The first molecular system, a non-symmetric trans,trans-1,2,4-triflourocyclooctane (referred to as cyclooctane throughout the paper) can form different conformations with rapid transitions. As it contains only 24 atoms and is simulated in the vacuum (ergo no further molecules of the solvent), the usually high computational cost of molecular dynamics stays low for this molecule. However, abrupt and rare changes in structure require the computation of 10 distances to all 521 reference structures in every step, because the neighbourlist would not be able to keep up. As distance computations take a vast majority of the computation time, it serves as an example of a molecule that should greatly benefit from the acceleration of CV computation. Second, 304-atom miniprotein Trp-cage, PDB ID 1L2Y, simulated in an implicit solvent smoothly diffuses between several conformations in its slow folding. The molecular dynamics, especially long-range electrostatic interactions, takes the majority of the computation time. Even with the neighbourlist, the closest 50 structures chosen from all 2120 reference structures every 50 steps, the contribution of metadynamics is not negligible, it takes almost half of the computation time. The trp-cage simulation shows a more common example where the acceleration of metadynamics speeds up the whole simulation only modestly. However, reducing computational overhead brought by metadynamics enables us to simulate at the speed of classical molecular dynamics with the advantage of rare events’ acceleration with metadynamics. We have set up the simulation as described in [5]. All simulations ran for 100,000 steps. Metadynamics simulation has been done without a neighbourlist for cyclooctane, i.e.distances to all 521 reference structures calculated every step, with 50-item NL updated every 50 steps for Trp-cage. In all simulations, the bias potential has been computed with the grid [22], diminishing the cost of a Gaussian hill summation growing with longer simulations. In close structure simulations, we reassign the close structure if D(x, y) > ε, where ε = 0.01 nm2 . Experiments for this paper differ from those in [5] in versions of software and our method’s implementation. Accuracy evaluation there has been done on Gromacs 4.5.7 and Plumed 2.1 with the prototype implementation of our method. Performance evaluation here has been done on Gromacs 5.1.4 and Plumed 2.3 with ready-to-use, optimized implementation of our method. All performance evaluation experiments were executed on one machine with 2x 8-core Intel Xeon E5-2650 v2 2.6 GHz, even multiple MPI processes were assigned to the same machine. 6. Results and Discussion The main aim of the close structure method is to accelerate metadynamics simulations by reducing the number of distance computations. Therefore, in its evaluation, we focused both on the theoretical and the practical speed-up. 11 We have evaluated the accuracy of the close structure method meticulously in [5], here we only shortly outline the main findings. 6.1. Accuracy Because of the chaotic character of molecular dynamics, even the smallest changes caused by the close structure approximation cause subsequent significant changes in trajectory. Since the step-by-step comparison of trajectories would be meaningless, we inspected the similarity of explored energy landscapes. First, we evaluated both Equation (3) and (5) on a long series of x and corresponding Rxy , finding almost perfect correlation. Second, we visualized property map collective variables computed during the simulation with the original method and with the close structure method. Both showed a great resemblance. Finally, we analyzed essential coordinates [23] of both trajectories. We found a match not only between the essential movements of the system but also in minor harmonic oscillations. We concluded that the trajectory of the close structure simulation explored the same energy landscape as the simulation with the original method. 6.2. Theoretical Speedup For theoretical evaluation of speed-up, we inspected the number of expensive distance computations and conditions for speed-up.In the next section, we calculate the theoretical speed-up according to Ahmdal’s law for both evaluated datasets. By an expensive distance computation we mean finding the rotation matrix with Kearsley’s method [15] and calculating the distance and derivatives with Equations (3) and (4). In the original method, all distances are calculated expensively. In the close structure method, the majority of distances are just cheaply approximated with saved rotation matrices using Equations (5) and (6). In the original method, in a usual step, we need M (size of the neighbourlist, or M = 0 in case of no neighborlist) distance computations. In a neighbourlist update step, every L steps (L = 1 in case of no neighborlist), we compute N (number of reference structures) distances. On average, we expensive distance computations per step. need M + N −M L In the close structure method, in a usual step, we need only one MSD computation, the rest is approximated. Furthermore, in a close structure reassignment step, we expensively recompute N rotation matrices, so that we keep the accuracy of distance computations for neighborlist updates. On 12 N expensive distance computations per step, where average, we need 1 + K the close structure is reassigned (on average) every K steps. The close structure method will accelerate metadynamics if M + N −M > L N 1 + K . From our experiments, if ε is not too small, then K  L, as the close structure was reassigned on average every few thousand steps for ε = 0.01 nm2 (10000 for cyclooctane, 3000 for Trp-cage) in the 100,000step simulations we did. The average number of reassignments depends on how often and how fast the structure of molecule changes, so it varies with the threshold ε, but also with the length of the simulation, the starting structure, temperature and other simulation parameters. For the simulations without the neighbourlist, the comparison is reduced N which clearly applies as K  1. Therefore, the number of to N > 1 + K distance computations is reduced approximately by an order of N . For the simulations with the neighbourlist, the specific values of M , L, N and K determine the acceleration. It is safe to assume M  N and L  N . If K > L, which is quite easily achievable for reasonable values of ε, then N −M N > K . Thus, the number of distance computations in simulations with L the NL can be reduced up to an order of M . Naturally, the speed-up of distance computation accelerates the whole simulation only to some extent. This can be assessed using Ahmdal’s law, see Equation (7). The portion p represents the percentage of time spent in functions we have accelerated, i.e. the distance computations, in the original implementation. The speed-up s denotes the acceleration of this portion in our modified implementation. The whole computation can achieve only the speed-up S. For example, even an infinite speed-up of distance computation would accelerate the whole simulation only ten times, if the distance computations originally took up 90% of computation time. S= 1 (1 − p) + p s (7) where p is the original portion of time spent in the given function, s is the speed-up of the accelerated function. More general performance behavior can be deduced from the analysis above. First, the acceleration brought by the close structure method decreases with the increased cost of molecular dynamics due to Ahmdal’s law. This applies especially in the case of larger molecules. The cost of computing MSD with the original method also grows with the number of atoms, but to a much lesser extent than the demands of MD. Therefore, a higher 13 acceleration is expected with smaller molecules. With larger molecules, the best we can achieve is to get closer to the performance of molecular dynamics without metadynamics. Second, the acceleration brought by the close structure method increases with the decreasing stride of neighborlist’s update (L), but does not depend much on its size. Therefore, a higher acceleration is expected in the case of no neighborlist. And finally, the acceleration brought by the close structure method decreases with the increased number of reference structures, especially with the neighborlist updated often. 6.3. Theoretical Speedup for Evaluated Datasets We evaluated the theoretical speed-up expected in simulations with our datasets. Moreover, we assessed the speed-up of the whole simulation using Ahmdal’s law and information from profiling. For cyclooctane, we need 512 MSD computations each step with the original method and an average of 1 + 512/10000 = 1.05 with the close structure method. The theoretical speed-up for MSD computation reaches 512/1.05 = 488. For Trp-cage, we need on average 50 + (2120 − 50)/50 = 91.4 MSD computations with the original method and 1 + 2120/3000 = 1.7 with the close structure method. The theoretical speed-up for MSD computation reaches 91.4/1.7 = 54. The theoretical speed-up for the whole simulation can be assessed using Ahmdal’s law and information from profiling. We combined the information from log files (time spent with whole metadynamics and evaluation of collective variables) and from Intel VTune Profiler (time spent computing the distance) to assess the percentage of computation time spent in various functions in cyclooctane and Trp-cage simulations with one MPI process and one OpenMP thread, see Table 1. Metadynamics (MTD) includes the evaluation of collective variables (CV), initialization, logging and addition of Gaussian hills. CV evaluation requires distance computations (MSD) before the calculation of its own value. MSD computation consists of matrix diagonalization (DSYEVR), the calculation of the distance and the derivatives. The speed-up SM SD represents the maximal theoretical speed-up of the whole simulation according to Ahmdal’s law (see Equation (7)) taking into account the speed-up of MSD computation as assessed above. As expected, cyclooctane simulation would greatly benefit from the acceleration due to a smaller size and no neighborlist. Trp-cage simulation can 14 Table 1: Time portions spent in various metadynamics functions and the maximal speedup of the whole simulation MTD CV MSD DSYEVR SM SD cyclooctane 99.7% 98.5% 93% 70% 14 Trp-cage 53% 50% 43% 23% 1.7 achieve only modest improvement as the most of the computational cost lies in MD code. 6.4. Practical Speedup and Scalability For the practical evaluation of speed-up and scalability, we ran simulations without metadynamics (Gromacs), with metadynamics and the original method for MSD computation (Gromacs+Plumed) and with metadynamics modified with the close structure method (Gromacs+modified Plumed). Furthermore, we executed them for various numbers of MPI processes and OpenMP threads. As a speed measure, we took ns/day stated by Gromacs in its logs. As a reference point for the speed-up, we considered Gromacs with original Plumed ran with one MPI process and one OpenMP thread. Out of four runs for each variant and MPI/OpenMP configuration, we considered the minimal running time, i.e. the highest speed, to eliminate the random interference of the operating system’s background activity. The upper parts of Figures 4 and 5 show the speed-ups for the basic case of one MPI process and one OpenMP thread. For both molecular systems, the close structure method almost reaches the maximal theoretical speed-up as calculated in the previous subsection. Cyclooctane simulations with the close structure method achieved a speedup of an order of magnitude, notice the logarithmic scale on the x-axis of Figure 4. The close structure method almost reached to the maximal theoretical performance, albeit it still lags behind Gromacs without metadynamics. That is caused by Plumed overhead like initialization, logging, Gaussian hills addition and CV evaluation. Trp-cage simulations with the close structure method achieved a more modest speed-up due to the major cost of molecular dynamics. Furthermore, the practical and theoretical speed-up differ more than in cyclooctane simulation. We suggest it is caused by a smaller portion of time spent in matrix diagonalization function, as shown in Table 1. The distance computation takes 43% of computation time. Out of that, only half is spent in diagonal15 Figure 4: Speedup and strong scalability of cyclooctane simulations. Notice the logarithmic scale of x axis, the reference point is the blue bar with the original method. The grey bar with molecular dynamics without metadynamics puts into perspective how metadynamics increases the cost. ization, as opposed to three-quarters in the case of cyclooctane. The rest is spent in the computation of distance and especially derivatives, which scales with the size of the molecule. We reduced the number of calls to DSYEVR 16 Figure 5: Speedup and strong scalability of the Trp-cage simulations. The reference point is the blue bar with the original method. The grey bar with molecular dynamics without metadynamics puts into perspective how metadynamics increases the cost. with the close structure method, but the other half of computation remained unchanged: both Gromacs+Plumed and Gromacs+modified Plumed contain the same code optimizations. Nevertheless, we got much closer to the performance of Gromacs alone, thus gaining the advantage of enhanced sampling 17 for only a small cost. The lower parts of Figures 4 and 5 show the speed-ups of strong scaling for various combinations of MPI processes and OpenMP threads. First, compare the performances with the same total number of threads, i.e. 1/4 with 2/2, 1/8 with 2/4 and 1/16 with 2/8. In all cases for the original method, two MPI processes perform better than one with twice as many threads. The reason for that stems from the implementation of CV calculation in Plumed. The loop over reference structures in the neighbourlist (line 2 in Algorithm 1) is parallelized only with MPI. As the close structure method accelerates the computation within that loop, the second MPI process does not bring any advantage. In the case of Trp-cage, it even slows down the computation due to the communication and synchronization cost. Decreasing speed with more resources (8 total threads for cyclooctane, 16 for Trp-cage) suggests that the overhead caused by parallelization exceeds its benefits. Second, compare performances for one MPI process, i.e. 1/4, 1/8 and 1/16. Any speed-up here can come only from faster molecular dynamics. For cyclooctane simulations, a small amount of computation saturates even a single core, thus performance stays about the same for all three variants. For Trp-cage simulation, some improvement appears as molecular dynamics requires more computation due to the protein’s size. The simulations with the close structure method also exhibit this trend. Overall, the close structure method reduces the gain of another MPI process but does not interfere with the general scalability behavior. 7. Conclusion In this work, we address high computational demands of distance calculations in path/property map collective variables in metadynamics. The original method expensively evaluates the rotation matrices needed to superimpose the current structure to the reference structures in each step for many reference structures. By introducing a close structure, we reuse those matrices in next steps and cheaply approximate the vast majority of distance computations. We thoroughly evaluated accuracy elsewhere [5] and concluded that simulations with the close structure method explore the same energy landscape as simulations with the original method, despite different trajectories. 18 Here, we presented the tuned implementation of our method ready for production use and evaluated its performance. We calculated the number of distance computations for both the original and close structure method and assessed the theoretical speed-up of the whole simulation. The number of expensive MSD computations decreased approximately by the order of the neighbourlist size or the number of all the reference structures in case of no neighborlist. Clearly, the accelerated distance computation is expressed in the speed-up of the whole simulation only to the extent proportional to the time spent on distance computation in the original method. That depends on the neighborlist size and update stride, but mostly on the cost of the molecular dynamics part, determined by the molecule’s size. The experiments showed that the actual speed-up closely approaches the theoretical one for both evaluated molecular systems. Apart from the close structure approximation, minor code optimizations in distance computation contributed to this, including enforcing of loop vectorization and inlining of frequently called functions. These optimizations did not help the original method, as it spent the majority of time in the finely-tuned BLAS routine. However, unvectorized loops became prominent after the employment of the close structure, so even minor code optimizations pushed the practical speed-up further. The close structure method affects the scalability of Plumed code, diminishing the gain of an additional MPI process. This was expected as we reduced the amount of computation so that even one core suffices. Overall, the positive effects of the close structure approximation are prominent especially in smaller molecules with frequently updated neighborlist or no neighborlist at all. With larger molecules, the method diminishes the overhead of metadynamics and approaches the performance of sole molecular dynamics simulation with the advantage of enhanced sampling. The error brought by the approximation of the close structure method does not negatively influence the simulation, the explored landscapes remain the same. Here we presented a typical application of the close structure method, molecular dynamics with metadynamics using property map collective variables to calculate the bias potential. The close structure method would benefit any CV-based method of enhanced sampling, provided it uses mean square distance as a measure. Our method significantly reduces the overhead of metadynamics in simulations, getting closer to the ideal situation: molecular dynamics with the advantage of enhanced sampling—but without 19 significant additional cost. Acknowledgments This work was supported by Czech Science Foundation (15-17269S) and LM2015047 Czech National Infrastructure for Biological Data. Computational resources were provided by the CESNET LM2015042 and the CERIT Scientific Cloud LM2015085, provided under the programme “Projects of Large Research, Development, and Innovations Infrastructures”. References [1] F. Jensen, Introduction to computational chemistry, 2nd Edition, John Wiley & Sons Ltd, Great Britain, 2007. [2] R. C. Bernardi, M. C. Melo, K. Schulten, Enhanced sampling techniques in molecular dynamics simulations of biological systems, Biochimica et Biophysica Acta (BBA) - General Subjects 1850 (5) (2015) 872–877. doi:10.1016/j.bbagen.2014.10.019. [3] A. Laio, M. Parrinello, Escaping free-energy minima., Proceedings of the National Academy of Sciences of the United States of America 99 (20) (2002) 12562–12566. arXiv:0208352, doi:10.1073/pnas.202427399. URL http://www.pnas.org/content/99/20/12562 [4] P. Tiwary, A. van de Walle, A review of enhanced sampling approaches for accelerated molecular dynamics, in: Springer Series in Materials Science, Vol. 245, Springer International Publishing, 2016, pp. 195–221. doi:10.1007/978-3-319-33480-6_6. [5] J. Pazúriková, A. Křenek, V. Spiwok, M. Šimková, Reducing the number of mean-square deviation calculations with floating close structure in metadynamics, Journal of Chemical Physics 146 (11) (2017) 115101. doi:10.1063/1.4978296. URL http://aip.scitation.org/doi/10.1063/1.4978296 [6] J. D. Durrant, R. M. Bush, R. E. Amaro, Microsecond Molecular Dynamics Simulations of Influenza Neuraminidase Suggest a Mechanism for the Increased Virulence of Stalk-Deletion Mutants, Journal of Physical Chemistry B 120 (33) (2016) 8590–8599. doi:10.1021/acs.jpcb. 6b02655. 20 [7] K. Lindorff-Larsen, P. Maragakis, S. Piana, D. E. Shaw, Picosecond to Millisecond Structural Dynamics in Human Ubiquitin, Journal of Physical Chemistry B 120 (33) (2016) 8313–8320. doi:10.1021/acs. jpcb.6b02024. [8] V. Spiwok, Z. Sucur, P. Hosek, Enhanced sampling techniques in biomolecular simulations (nov 2015). doi:10.1016/j.biotechadv. 2014.11.011. [9] A. Laio, F. Gervasio, Metadynamics: a method to simulate rare events and reconstruct the free energy in biophysics, chemistry and material science, Reports on Progress in Physics 71 (12) (2008) 126601:1–126601:22. doi:10.1088/0034-4885/71/12/126601. [10] A. Laio, A. Rodriguez-Fortea, F. L. Gervasio, M. Ceccarelli, M. Parrinello, Assessing the accuracy of metadynamics, Journal of Physical Chemistry B 109 (14) (2005) 6714–6721. doi:10.1021/jp045424k. [11] A. Barducci, M. Bonomi, M. Parrinello, Metadynamics, Wiley Interdisciplinary Reviews: Computational Molecular Science 1 (5) (2011) 826–843. doi:10.1002/wcms.31. [12] D. Branduardi, F. L. Gervasio, M. Parrinello, From A to B in free energy space, Journal of Chemical Physics 126 (5) (2007) 054103. doi:10. 1063/1.2432340. [13] V. Spiwok, B. Králová, Metadynamics in the conformational space nonlinearly dimensionally reduced by Isomap, Journal of Chemical Physics 135 (22) (2011) 1–7. doi:10.1063/1.3660208. [14] J. Gu, P. E. Bourne, Structural bioinformatics, Wiley-Blackwell, 2009. [15] S. Kearsley, On the orthogonal transformation used for structural comparisons, Acta Crystallographica Section A Foundations of Crystallography 45 (2) (1989) 208–210. doi:10.1107/S0108767388010128. [16] E. A. Coutsias, C. Seok, K. A. Dill, Using quaternions to calculate RMSD, Journal of Computational Chemistry 25 (15) (2004) 1849–1857. arXiv:arXiv:1011.1669v3, doi:10.1002/jcc.20110. 21 [17] W. Kabsch, A solution for the best rotation to relate two sets of vectors, Acta Crystallographica Section A 32 (5) (1976) 922–923. arXiv: 05677394, doi:10.1107/S0567739476001873. [18] S. Páll, M. J. Abraham, C. Kutzner, B. Hess, E. Lindahl, Tackling exascale software challenges in molecular dynamics simulations with GROMACS, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 8759 (2015) 3–27. arXiv:1506.00716, doi:10.1007/978-3-319-15976-8_ 1. [19] G. Tribello, M. Bonomi, D. Branduardi, C. Camilloni, G. Bussi, PLUMED 2: New feathers for an old bird, Computer Physics Communications 185 (2) (2014) 604–613. arXiv:1310.0980, doi:10.1016/ j.cpc.2013.09.018. [20] J. Filipovič, J. Pazúriková, A. Křenek, V. Spiwok, Accelerated RMSD Calculation for Molecular Metadynamics, in: J. Évora-Gómez, J. J. Hernandéz-Cabrera (Eds.), 30th European Simulation and Modelling Conference, Reproduct NV, Ghent, Belgium, 2016, pp. 278–280. URL https://www.eurosis.org/cms/index.php?q=taxonomy/term/ 58 [21] Plumed2. URL https://github.com/plumed/plumed2 [22] User documention on feature METAD. URL https://plumed.github.io/doc-v2.2/user-doc/html/_m_e_ t_a_d.html [23] A. Amadei, A. B. Linssen, H. J. Berendsen, Essential dynamics of proteins., Proteins 17 (4) (1993) 412–25. doi:10.1002/prot.340170408. 22
5cs.CE
Quantum query complexity of entropy estimation Tongyang Li∗ Xiaodi Wu† arXiv:1710.06025v1 [quant-ph] 16 Oct 2017 Department of Computer Science, Institute for Advanced Computer Studies, and Joint Center for Quantum Information and Computer Science, University of Maryland, USA Abstract Estimation of Shannon and Rényi entropies of unknown discrete distributions is a fundamental problem in statistical property testing and an active research topic in both theoretical computer science and information theory. Tight bounds on the number of samples to estimate these entropies have been established in the classical setting, while little is known about their quantum counterparts. In this paper, we give the first quantum algorithms for estimating αRényi entropies (Shannon entropy being 1-Renyi entropy). In particular, we demonstrate a quadratic quantum speedup for Shannon entropy estimation and a generic quantum speedup for α-Rényi entropy estimation for all α ≥ 0, including a tight bound for the collision-entropy (2-Rényi entropy). We also provide quantum upper bounds for extreme cases such as the Hartley entropy (i.e., the logarithm of the support size of a distribution, corresponding to α = 0) and the min-entropy case (i.e., α = +∞), as well as the Kullback-Leibler divergence between two distributions. Moreover, we complement our results with quantum lower bounds on α-Rényi entropy estimation for all α ≥ 0. Our approach is inspired by the pioneering work of Bravyi, Harrow, and Hassidim (BHH) [13] on quantum algorithms for distributional property testing, however, with many new technical ingredients. For Shannon entropy and 0-Rényi entropy estimation, we improve the performance of the BHH framework, especially its error dependence, using Montanaro’s approach to estimating the expected output value of a quantum subroutine with bounded variance [41] and giving a fine-tuned error analysis. For general α-Rényi entropy estimation, we further develop a procedure that recursively approximates α-Rényi entropy for a sequence of αs, which is in spirit similar to a cooling schedule in simulated annealing. For special cases such as integer α ≥ 2 and α = +∞ (i.e., the min-entropy), we reduce the entropy estimation problem to the α-distinctness and the dlog ne-distinctness problems, respectively. We exploit various techniques to obtain our lower bounds for different ranges of α, including reductions to (variants of) existing lower bounds in quantum query complexity as well as the polynomial method inspired by the celebrated quantum lower bound for the collision problem. ∗ † [email protected] [email protected] 1 Introduction Motivations. Property testing is a rapidly developing field in theoretical computer science (e.g. see the survey [55]). It aims to determine properties of an object with the least number of independent samples of the object. Property testing is a theoretically appealing topic with intimate connections to statistics, learning theory, and algorithm design. One important topic in property testing is to estimate statistical properties of unknown distributions (e.g., [61]), which are fundamental questions in statistics and information theory, given that much of science relies on samples furnished by nature. The Shannon [56] and Rényi [54] entropies are central measures of randomness compressibility. In this paper, we focus on estimating these entropies for an unknown distribution. Specifically, given a distribution p over a set X of size n (w.l.o.g. let X = [n]) where px denotes the probability of x ∈ X, the Shannon entropy H(p) of this distribution p is defined by 1 X H(p) := px log . (1.1) px x∈X: px >0 A natural question is to determine the sample complexity (i.e., the necessary number of independent samples from p) to estimate H(p), with error  and high probability. This problem has been intensively studied in the classical literature. For multiplicative error , Batu et al. [7, Theorem 2 2] provided the upper bound of O(n(1+o(1))/(1+) log n), while an almost matching lower bound of 2 Ω(n(1−o(1))/(1+) ) was shown by Valiant [61, Theorem 1.3]. For additive errors, Paninski gave a nonconstructive proof of the existence of sublinear estimators in [49, 50], while an explicit construction using Θ(n/ log n) samples was shown by Valiant and Valiant in [60] when  > n−0.03 ; for the case  ≤ n−0.03 , Wu and Yang [64] and Jiao et al. [34] gave the optimal estimator with (log n)2 n Θ(  log ) samples. A sequence of works in information theory [33, 34, 64] studied the n + 2 minimax mean-squared error, which becomes O(1) also using Θ(n/ log n) samples. One important generalization of Shannon entropy is the Rényi entropy of order α > 0, denoted Hα (p), which is defined by ( P 1 log x∈X pαx , when α 6= 1. 1−α (1.2) Hα (p) := limα→1 Hα (p), when α = 1. The Rényi entropy of order 1 is simply the Shannon entropy, i.e., H1 (p) = H(p). General Rényi entropy can be used as a bound on Shannon entropy, making it useful in many applications (e.g., [6, 17]). Rényi entropy is also of interest in its own right. One prominent example is the Rényi entropy of order 2, H2 (p) (also known as the collision entropy), which measures the quality of random number generators (e.g., [62]) and key derivation in cryptographic applications (e.g., [11, 32]). Motivated by these and other applications, the estimation of Rényi entropy has also been actively studied [4, 33, 34]. In particular, Acharya et al. [4] have shown almost tight bounds on the classical query complexity of computing Rényi entropy. Specifically, for any non-integer α > 1, the classical query complexity of α-Rényi entropy is Ω(n1−o(1) ) and O(n). Surprisingly, for any integer α > 1, the classical query complexity is Θ(n1−1/α ), i.e., sublinear in n. When 0 ≤ α < 1, the classical query complexity is Ω(n1/α−o(1) ) and O(n1/α ), which is always superlinear. The extreme case (α → ∞) is known as the min-entropy, denoted H∞ (p), which is defined by H∞ (p) := lim Hα (p) = − log max pi . α→∞ (1.3) i∈[n] Min-entropy plays an important role in the randomness extraction (e.g., [59]) and characterizes the maximum number of uniform bits that can be extracted from a given distribution. Classically, the query complexity of min-entropy estimation is Θ(n/ log n), which follows directly from [60]. 1 Another extreme case (α = 0), also known as the Hartley entropy [29], is the logarithm of the support size of distributions, where the support of any distribution p is defined by Supp(p) := |{x : x ∈ X, px > 0}|. (1.4) It is a natural and fundamental quantity of distributions with various applications (e.g., [20, 22, 26, 31, 36, 51, 58]). However, estimating the support size is impossible in general because elements with negligible but nonzero probability, which are very unlikely to be sampled, could still contribute to Supp(p). Two related quantities (support coverage and support size) have hence been considered as alternatives of 0-Rényi entropy with roughly Θ(n/ log(n)) complexity. (See details in Section 8.) Besides the entropic measures of a discrete distribution, we also briefly discuss an entropic measure between two distributions, namely the Kullback-Leibler (KL) divergence. Given two discrete distributions p and q with cardinality n, the KL divergence is defined as X pi pi log . DKL (pkq) = (1.5) qi i∈[n] KL divergence is a key measure with many applications in information theory [18, 37], data compression [15], and learning theory [35]. Classically, under the assumption that pqii ≤ f (n) ∀ i ∈ [n] for some f (n), DKL (pkq) can be approximated within constant additive error with high success (n) probability if Θ( logn n ) samples are taken from p and Θ( nf log n ) samples are taken from q. Main question. In this paper, we study the impact of quantum computation on estimation of general Rényi entropies. Specifically, we aim to characterize quantum speed-ups for estimating Shannon and Rényi entropies. Our question aligns with the emerging topic called “quantum property testing” (see the survey [43]) and focuses on investigating the quantum advantage in testing classical statistical properties. To the best of our knowledge, the first research paper on distributional quantum property testing is by Bravyi, Harrow, and Hassidim (BHH) [13], where they discovered quantum speedups for testing uniformity, orthogonality, and statistical difference on unknown distributions. Some of these results were subsequently improved by Chakraborty et al. [16]. Reference [13] also claimed that Shannon √ entropy could be estimated with query complexity O( n), however, without details and explicit error dependence. Indeed, our framework is inspired by [13], but with significantly new ingredients to achieve our results. There is also a related line of research on spectrum testing or tomography of quantum states [25, 45–47]. However, these works aim to test properties of general quantum states, while we focus on using quantum algorithms to test properties of classical distributions (i.e., diagonal quantum states)1 . Distributions as oracles. The sampling model in the classical literature assumes that a tester is presented with independent samples from an unknown distribution. One of the contributions of BHH is an alternative model that allows coherent quantum access to unknown distributions. Specifically, BHH models a discrete distribution p = (pi )ni=1 on [n] by an oracle Op : [S] → [n] for some S ∈ N. The probability pi (i ∈ [n]) is proportional to the size of pre-image of i under Op . Namely, an oracle Op : [S] → [n] generates p if and only if for all i ∈ [n], pi = |{s ∈ [S] : Op (s) = i}|/S. (1.6) (note that we assume pi s to be rational numbers). If one samples s uniformly from [S], then the output Op (s) is from distribution p. Instead of considering sample complexity—that is, the number 1 Note that one can also leverage the results of [25, 45–47] to test properties of classical distributions. However, they are less efficient because they deal with a much harder problem involving general quantum states. 2 of used samples—we consider the query complexity in the oracle model that counts the number of oracle uses. Note that a tester interacting with an oracle can potentially be more powerful due to the possibility of learning the internal structure of the oracle as opposed to the sampling model. However, it is shown in [13] that the query complexity of the oracle model and the sample complexity of the sampling model are in fact the same classically. A significant advantage of the oracle model is that it naturally allows coherent access when extended to the quantum case, where we transform Op into a unitary operator Ôp acting on CS ⊗ Cn+1 such that Ôp |si|0i = |si|Op (s)i ∀ s ∈ [S]. (1.7) Moreover, this oracle model can also be readily obtained in some algorithmic settings, e.g., when distributions are generated by some classical or quantum sampling procedure. Thus, statistical property testing results in this oracle model can be potentially leveraged in algorithm design. Our Results. Our main contribution is a systematic study of both upper and lower bounds for the quantum query complexity of estimation of Rényi entropies (including Shannon entropy as a special case). Specifically, we obtain the following quantum speedups for different ranges of α. Theorem 1.1. There are quantum algorithms that approximate Hα (p) of distribution p on [n] within an additive error 0 <  ≤ O(1) with success probability at least 2/3 using2 √  • Õ 1.5n quantum queries when α = 0, i.e., Hartley entropy. See Theorem 8.2.3 1/α−1/2  • Õ n 2 quantum queries4 when 0 < α < 1. See Theorem 5.2. √  • Õ 2n quantum queries when α = 1, i.e., Shannon entropy. See Theorem 3.1. ν(1−1/α)  quantum queries when α > 1, α ∈ N for some ν < 43 . See Theorem 6.1. • Õ n 2 1−1/2α  quantum queries when α > 1, α ∈ / N. See Theorem 5.1. • Õ n 2    n n • Õ Q( 16 log -distinctness) quantum queries when α = ∞, where Q( 16 log -distinctness) is 2 2  16 log n  the quantum query complexity of the -distinctness problem. See Theorem 7.1. 2 Our quantum testers demonstrate advantages over classical ones for all 0 < α < ∞; in particular, our quantum tester has a quadratic speedup in the case of Shannon entropy. When α = ∞, our quantum upper bound depends on the quantum query complexity of the dlog ne-distinctness problem, which is open to the best of our knowledge5 and might demonstrate a quantum advantage. As a corollary, we also obtain quadratic quantum speedup for estimating KL divergence: Corollary 1.1 (see Theorem 4.1). Assuming p and q satisfies pqii ≤ f (n) ∀ i ∈ [n] for some function f : N → R+ , DKL (pkq), there is a quantum algorithm that √ approximates DKL (pkq) within  2 an additive error  > 0 with success probability at least 3 using Õ 2n quantum queries to p and √ (n)  quantum queries to q. Õ nf 2 2 It should be understood that the success probability 2/3 can be boosted to close to 1 without much overhead, e.g., see Lemma 5.5 in Section 5.1.5. 3 0-Rényi entropy estimation is intractable without any assumption, both classically and quantumly. Here, the results are based on the assumption that nonzero probabilities are at least 1/n. See Section 8 for more information. 4 Õ hides factors that are polynomial in log n and log 1/. 5 Existing quantum algorithms for the k-distinctness problem (e.g., [5] has query complexity O(k2 nk/k+1 ) and [9] 2 has query complexity O(2k nν ) for some ν < 3/4) do not behave well for super-constant ks. 3 α classical bounds α=0 Θ( logn n ) [48, 63] 1 1 −o(1) nα α O( log ) [4] n ), Ω(n n Θ( log n ) [34, 60, 64] O( logn n ), Ω(n1−o(1) ) [4] 0<α<1 α=1 α > 1, α ∈ /N α=2 α > 2, α ∈ N α=∞ √ Θ( n) [4] Θ(n1−1/α ) [4] Θ( logn n ) [60] quantum bounds (this paper) √ 1 Õ( n), Ω(n 3 ) 1 1 1 1 Õ(n α − 2 ), Ω(max{n 7α −o(1) , n 3 }) √ 1 Õ( n), Ω(n 3 ) 1  1 1 1 Õ(n1− 2α , Ω(max{n 3 , n 2 − 2α }) 1 Θ̃(n 3 ) 1 1 ν(1−1/α) Õ(n ), Ω(n 2 − 2α ), ν < 3/4 √ Õ(Q(dlog ne-distinctness)), Ω( n) Table 1: Summary of classical and quantum query complexity of Hα (p), assuming  = Θ(1). We also obtain corresponding quantum lower bounds on entropy estimation as follows. We summarize both bounds in Table 1 and visualize them in Figure 1. Theorem 1.2 (See Theorem 9.1). Any quantum algorithm that approximates Hα (p) of distribution p on [n] within additive error  with success probability at least 2/3 must use 1 1 • Ω(n 3 / 6 ) quantum queries when α = 0, assuming 1/n ≤  ≤ 1. 1 2 • Ω̃(n 7α −o(1) / 7 ) quantum queries when 0 < α < 37 . 1 1 1 1 • Ω(n 3 / 6 ) quantum queries when 3 7 ≤ α ≤ 3, assuming 1/n ≤  ≤ 1. • Ω(n 2 − 2α /) quantum queries when 3 ≤ α < ∞. √ • Ω( n/) quantum queries when α = ∞. exponent of n Classical tight bounds 2.0 Quantum upper bounds Quantum lower bounds 1.5 1.0 0.5 0 1 2 3 4 5 α Figure 1: Visualization of classical and quantum query complexity of Hα (p). The x-axis represents α and the y-axis represents the exponent of n. Red curves and points represent quantum upper bounds. Green curves and points represent classical tight bounds. The Blue curve represents quantum lower bounds. Techniques. At a high level, our upper bound is inspired by BHH [13], where we formulate a framework (in Section P 2) that generalizes the technique in BHH and makes it applicable in our case. Let F (p) = x px f (px ) for some function f (·) and distribution p. Similar to BHH, we design a master algorithm that samples x from p and then use the quantum counting primitive [12] to obtain an estimate p̃x of px and outputs f (p̃x ). It is easy to see that the expectation of the 4 output of the master algorithm is roughly6 F (p). By choosing appropriate f (·)s, one can recover H(p) or Hα (p) as well as the ones used in BHH. It suffices then to obtain a good estimate of the output expectation of the master algorithm, which was achieved by multiple independent runs of the master algorithm in BHH. The performance of the above framework (and its analysis) critically depends on how close the expectation of the algorithm is to F (p) and how concentrated the output distribution is around its expectation, which in turn heavily depends on the specific f (·) in use. Our first contribution is a finetuned error analysis for specific f (·)s, such as in the case of Shannon entropy (i.e., f (px ) = − log(px )) whose values could be significant for boundary cases of px . Instead of only considering the case when p̃x is a good estimate of px as in BHH, we need to analyze the entire distribution of p̃x using quantum counting. We also leverage a generic quantum speedup for estimating the expectation of the output of any quantum procedure with additive errors [41], which significantly improves our error dependence as compared to BHH. These improvements already give a quadratic quantum speedup for Shannon (Section 3) and 0-Rényi (Section 8) entropy estimation. As an application, it also gives a quadratic speedup for estimating the KL-divergence between two distributions (see Section 4). For general α-Rényi entropy Hα (p), we choose f (px ) = pα−1 and let Pα (p) = F (p) so that x Hα (p) ∝ log Pα (p). Instead of estimating F (p) with additive errors in the case of Shannon entropy, we switch to working with multiplicative errors which is harder since the aforementioned quantum algorithm [41] is much weaker in this setting. Indeed, by following the same technique, we can only obtain quantum speedups for α-Rényi entropy when 1/2 < α < 2. For general α > 0, our first observation is that if one knew the output expectation E[X] is within [a, b] such that b/a = Θ(1), then one can slightly modify the technique in [41] (as shown in Theorem 2.2) and obtain a quadratic quantum speedup similar to the additive error setting. This approach, however, seems circular since it is unclear how to obtain such a, b in advance. Our second observation is that for any close enough α1 , α2 , Pα1 (p) can be used to bound Pα2 (p). Precisely, when α1 /α2 = 1 ± 1/ log(n), we have Pα1 (p) = Θ(Pα2 (p)α1 /α2 ) (see Lemma 5.3). As a result, when estimating Pα (p), we can first estimate Pα0 to provide a bound on Pα , where α0 , α differ by a 1 ± 1/ log(n) factor and α0 moves toward 1. We apply this strategy recursively on estimating Pα0 until α0 is very close to 1 from above when initial α > 1 or from below when initial α < 1, where a quantum speedup is already known. At a high level, we recursively estimate a sequence (of size O(log n)) of such αs that eventually converges to 1, where in each iteration we establish some quantum speedup which leads to an overall quantum speedup. We remark that our approach is in spirit similar to the cooling schedules in simulated annealing (e.g. [57]). (See Section 5.) For integer α ≥ 2, we observe a connection between Pα (p) and the α-distinctness problem which leads to a more significant quantum speedup. Precisely, let Op : [S] → [n] be the oracle in (1.7), we observe that Pα (p) is proportional to the α-frequency moment of Op (1), . . . , Op (S) which can be solved quantumly [42] based on any quantum algorithm for the α-distinctness problem (e.g., [9]). However, there is a catch that a direct application of [42] will lead to a dependence on S rather than n. We remedy this situation by tweaking the algorithm and its analysis in [42] to remove the dependence on S for our specific setting. (See Section 6.) The integer α algorithm fails to extend to the min-entropy case (i.e., α = +∞) because the hidden constant in O(·) has a poor dependence on α (see Remark 6.1). Instead, we develop another reduction to the dlog ne-distinctness problem by exploiting the so-called “Poissonized sampling” technique [34, 39, 60]. At a high level, we construct Poisson distributions that are parameterized by pi s and leverage the “threshold” behavior of Poisson distributions (see Lemma 7.1). Roughly, if 6 The accurate expectation is P x px E[f (p̃x )]. Intuitively, we expect p̃x to be a good estimate of px . 5 maxi pi passes some threshold, with high probability, these parameterized Poisson distributions will lead to a collision of size dlog ne that will be caught by the dlog ne-distinctness algorithm. Otherwise, we run again with a lower threshold until the threshold becomes trivial. (See Section 7.) Some of our lower bounds come from reductions to existing ones in quantum query complexity, such as the quantum-classical separation of symmetric boolean functions [1], the collision problem [2, 38], and the Hamming weight problem [44], for different ranges of α. We also obtain lower bounds with a better error dependence by the polynomial method, which is inspired by the celebrated quantum lower bound for the collision problem [2, 38]. (See Section 9.) Open questions. Our paper raises a few open questions. A natural question is to close the gaps between our quantum upper and lower bounds. Our quantum techniques on both ends are actually quite different from the state-of-the-art classical ones (e.g., [60]). It is interesting to see whether one can incorporate classical ideas to improve our quantum results. It is also possible to achieve better lower bounds by improving our application of the polynomial method or exploiting the quantum adversary method (e.g., [10, 30]). Finally, our result motivates the study of the quantum algorithm for the k-distinctness problem with super-constant k, which might also be interesting by itself. Notations. Throughout the paper, we consider a discrete distribution {pi }ni=1 on [n], and Pα (p) := Pn α i=1 pi represents the α-power sum of p. In the analyses of our algorithms, ‘log’ is natural logarithm; ‘≈’ omits lower order terms. 2 Master algorithm Let p = (pi )ni=1 be a discrete distribution on [n] encoded by the quantum oracle Ôp defined in (1.7). Inspired by BHH, P we develop the following master algorithm to estimate a property F with the form F (p) := i∈[n] pi f (pi ) for a function f : (0, 1] → R. Algorithm 1: Estimate F (p) = P i pi f (pi ) of a discrete distribution p = (pi )ni=1 on [n]. 5 Set l, M ∈ N; Regard the following subroutine as A: Draw a sample i ∈ [n] according to p; Use EstAmp or EstAmp0 with M queries to obtain an estimation p̃i of pi ; Output X = f (p̃i ); 6 Use A for l executions in Theorem 2.1 or Theorem 2.2 and output F̃ (p) to estimate F (p); 1 2 3 4 Comparing to BHH, we introduce a few new technical ingredients in the design of Algorithm 1 and its analysis, which significantly improve the performance of Algorithm 1 especially for specific f (·)s in our case, e.g., f (px ) = − log(px ) (Shannon entropy) and f (px ) = pα−1 (Rényi entropy). x The first one is a generic quantum speedup of Monte Carlo methods [41], in particular, a quantum algorithm that approximates the output expectation of a subroutine with additive errors that has a quadratic better sample complexity than the one implied by Chebyshev’s inequality. Theorem 2.1 (Additive error; Theorem 5 of [41]). Let A be a quantum algorithm with output X such that Var[X] ≤ σ 2 . Then for  where 0 <  < 4σ, by using O((σ/) log3/2 (σ/) log log(σ/)) executions of A and A−1 , Algorithm 3 in [41] outputs an estimate Ẽ[X] of E[X] such that   Pr Ẽ[X] − E[X] ≥  ≤ 1/5. (2.1) 6 It is worthwhile mentioning that classically one needs to use Ω(σ 2 /2 ) executions of A [19] to estimate E[X]. Theorem 2.1 demonstrates a quadratic improvement on the error dependence. In the case of approximating Hα (p), we need to work with multiplicative errors while existing results (e.g. [41]) have a worse error dependence which is insufficient for our purposes. Instead, inspired by [41], we prove the following theorem (our second ingredient) that takes auxiliary information about the range of E[X] into consideration, which might be of independent interest. Theorem 2.2 (Multiplicative error; Appendix A). Let A be a quantum algorithm with output X such that Var[X] ≤ σ 2 E[X]2 for a known σ. Assume that E[X] ∈ [a, b]. Then for  where 0 <  < 24σ, by using A and A−1 for O((σb/a) log3/2 (σb/a) log log(σb/a)) executions, Algorithm 10 (given in Appendix A) outputs an estimate Ẽ[X] of E[X] such that   Pr Ẽ[X] − E[X] ≥ E[X] ≤ 1/10. (2.2) The third ingredient is a fine-tuned error analysis due to the specific f (·)s. Similar to BHH, we rely on quantum counting (named EstAmp) [12] to estimate the pre-image size of a Boolean function, which provides another source of quantum speedup. In particular, we approximate any probability px in the query model ((1.7)) by p̃x by estimating the size of the pre-image of a Boolean function χ : [S] → {0, 1} with χ(s) = 1 if O(s) = i and χ(s) = 0 otherwise. However, for cases in BHH, it suffices to only consider the probability when px and p̃x are close, while in our case, we need to analyze the whole output distribution of quantum counting. Specifically, letting t = χ−1 (1) and a = t/S = sin2 (ωπ) for some ω, we have Theorem 2.3 ([12]). For any k, M ∈ N, there is a quantum algorithm (named EstAmp) with M lπ quantum queries to χ that outputs ã = sin2 M for some l ∈ {0, . . . , M − 1} such that  lπ i h 1 sin2 (M ∆π) ≤ Pr ã = sin2 = 2 2 , (2.3) M (2M ∆)2 M sin (∆π) √ a(1−a) π2 8 where ∆ = |ω − Ml |. This promises |ã − a| ≤ 2πk M + k2 M 2 with probability at least π 2 for 1 k = 1 and with probability greater than 1 − 2(k−1) for k ≥ 2. If a = 0 then ã = 0 with certainty. Moreover, we also need to slightly modify EstAmp to avoid outputting p̃x = 0 in estimating Shannon entropy. This is because f (p̃x ) = log(p̃x ) is not well-defined at p̃x = 0. Let EstAmp0 be π the modified algorithm. It is required that EstAmp0 outputs sin2 ( 2M ) when EstAmp outputs 0 and outputs EstAmp’s output otherwise. By leveraging Theorem 2.1, Theorem 2.2, Theorem 2.3, and carefully setting parameters in Algorithm 1, we have the following corollaries that describe the complexity of estimating any F (p).    Corollary 2.1 (additive error). Given  > 0. If l = Θ σ log3/2 σ log log σ where Var[X] ≤ σ 2 and M is large enough such that E[X] − F (p) ≤ , then Algorithm 1 approximates F (p) with an additive error  and success probability 2/3 using O M · l) quantum queries to p. Corollary 2.2 (multiplicative error). Assume a procedure using Ca,b quantum queries that returns an estimated range [a,  b], and that E[X] ∈2 [a, b] 2 with probability at least 0.9. Let l = 3/2 σb σb Θ ( σb ) log ( ) log log( ) where Var[X]/(E[X]) ≤ σ and  > 0. For large enough M such a a a that E[X] − F (p) ≤ , Algorithm 1 estimates F (p) with a multiplicative error  and success probability 2/3 with O M · l + Ca,b ) queries. 7 Algorithm 2: Estimate the Shannon entropy of p = (pi )ni=1 on [n]. 2)  2 )  log(n/2 ) 1 Set l = Θ log3/2 log(n/ log log log(n/ ;    2 Regard the following subroutine as A: 3 Draw a sample i ∈ [n] according to p; √ 4 Use EstAmp0 with M = 2dlog2 ( n/)e queries to obtain an estimation p̃i of pi ; 5 Output x̃i = log(1/p̃i ); 6 3 Use A for l executions in Theorem 2.1 and output an estimation H̃(p) of H(p); Shannon entropy estimation We develop Algorithm 2 for Shannon entropy estimation with EstAmp0 in Line 4, which provides quadratic quantum speedup in n. Theorem 3.1. Algorithm 2 approximates H(p) within an additive error 0 <  ≤ O(1) with success √  n 2 probability at least 3 using Õ 2 quantum queries to p. Proof. We prove this theorem in two steps. The first step is to show that the P expectation of the P subroutine A’s output (denoted Ẽ := i∈[n] pi · log(1/p̃i )) is close to E := i∈[n] pi · log(1/pi ) = H(p). To that end, we divide [n] into partitions based on the corresponding probabilities. Let √ m = dlog2 ( n/)e and S0 = {i : pi ≤ sin2 (π/2m+1 )}, S1 = {i : sin2 (π/2m+1 ) < pi ≤ sin2 (π/2m )}, S2 = {i : sin2 (π/2m ) < pi ≤ sin2 (π/2m−1 )}, . . . , Sm = {i : sin2 (π/4) < pi ≤ sin2 (π/2)}. For convenience, denote s0 = |S0 |, s1 = |S1 |, . . . , sm = |Sm |. Then m X m X 22j sj = Θ(1). 22m sj = n, j=0 (3.1) j=0 Our main technical contribution is the following upper bound on the expected difference between log p̃i and log pi in terms of the partition Si , i = 1, · · · , n:  2j s  X   j pi E log p̃i − log pi = O 2m ∀ j ∈ {1, . . . , m}. (3.2) 2 i∈Sj By linearity of expectation, we have |Ẽ − E| ≤ X i∈[n] m X m  2j s   X   X  j O 2m . pi E log p̃i − log pi = pi E log p̃i − log pi = 2 j=0 j=0 i∈Sj As a result, by applying (3.1) and Cauchy-Schwartz inequality to (3.3), we have v ! u X m m  2j s   X  2j X u m 1 2 j |Ẽ − E| = O 2m ≤ O t sj sj = O(). 2 22m 22m j=0 j=0 (3.3) (3.4) j=0 Because a constant overhead does not influence the query complexity, we may rescale Algorithm 2 by a large enough constant so that |Ẽ − E| ≤ /2. The second step is to bound the variance of the random variable, which is X 2 X X pi (log p̃i )2 − pi log p̃i ≤ pi (log p̃i )2 . (3.5) i∈[n] i∈[n] i∈[n] 8 P π 2 ) ≥ M12 ≥ 4n , we have i∈[n] pi (log p̃i )2 ≤ Since for any i, EstAmp0 outputs p̃i such that p̃i ≥ sin2 ( 2M  2 P 4n 2 = log 4n . As a result, by Corollary 2.1 we can approximate Ẽ up to additive 2 i∈[n] pi · log 2  error /2 with failure probability at most 1/3 using  log(n/2 )   log(n/2 )   √n  2  √ 3/2 log(n/ ) dlog2 ( n/)e O log log log ·2 = Õ 2 (3.6)     quantum queries. Together with |Ẽ −E| ≤ /2, Algorithm 2 approximates E = H(p) up to additive error  with failure probability at most 1/3. It remains to prove (3.2). We prove: X i∈S0  s    0 pi E log p̃i − log pi = O 2m . 2 (3.7) For j ∈ {1, 2, . . . , m} in (3.2), the proof is similar because the dominating term has the angles of p̃i and pi fall into the same interval of length 21m , and as a result | log p̃i − log pi = O( 21j ). Proof of (3.7). For convenience, denote h(x) := x(log t−log x) ≤ t/e where 0 < t ≤ 1 and x ∈ (0, t]. Because h0 (x) = log t − log x − 1, when x ∈ (0, t/e), h0 (x) > 0 hence h(x) is an increasing function; when x ∈ (t/e, t), h0 (x) < 0 hence h(x) is a decreasing function; when x = t/e, h0 (x) = 0 and h reaches its maximum t/e. Since i ∈ S0 , we can write pi = sin2 (θi π) where 0 < θi ≤ 1/2m+1 . By Theorem 2.3, for any l ∈ {1, . . . , 2m−1 }, the output of EstAmp0 when taking 2m queries satisfies  π i h Pr p̃i = sin2 m+1 = 2  lπ i h Pr p̃i = sin2 m = 2 sin2 (2m θi π) ≤ 1; 22m sin2 (θi π) l sin2 (2m ( 2m − θi )π) l 22m sin2 (( 2m − θi )π) (3.8) ≤ 1 l (2m+1 ( 2m − θi ))2 . (3.9) Combining (3.8), (3.9), and the property of function h discussed above, for any i ∈ S0 we have   pi E log p̃i − log pi  m−1   π   2X pi log sin2 ( 2lπm ) − log pi 2 ≤ 1 · pi log sin − log pi + (3.10) l 2m+1 (2m+1 ( 2m − θi ))2 l=1 ≤ π sin2 ( 2m+1 ) + e m−1 2X l=1  π     π  1 2 2 lπ 2 · sin log sin − log sin (2l − 1)2 2m+1 2m 2m+1 2m−1  sin( lπ ) 2 1 π2 1 1 π2 X 2m  + · log ≤ π 2m 2m 2 4e 2 2 4 (2l − 1) sin 2m+1 l=1 (3.11) (3.12) m−1 2 π2 1 1 π 2 X log 2l ≤ + · 4e 22m 22m 2 (2l − 1)2 l=1  1  = O 2m , 2 (3.13) (3.14) where (3.10) comes from (3.8) and (3.9), (3.11) comes from the property of h, (3.12) holds because  π π2 π sin2 ( 2m+1 ) ≤ 22m+2 , (3.13) holds because sin2 ( 2lπm ) ≤ 4l2 sin2 2m+1 , and (3.14) holds because 9 P∞ log l l=1 (2l−1)2 = O(1). Consequently, X i∈S0 4  1   s    0 pi E log p̃i − log pi = O 2m · s0 = O 2m . 2 2 (3.15) Application: KL divergence estimation Classically, there does not exist any consistent estimator that guarantees asymptotically small error over the set of all pairs of distributions [14, 27]. These two papers then consider pairs of distributions with bounded probability ratios specified by a function f : N → R+ , namely all pairs of distributions in the set as follows:  pi Un,f (n) := (p, q) : |p| = |q| = n, ≤ f (n) ∀ i ∈ [n] . (4.1) qi Denote the number of samples from p and q to be Mp and Mq , respectively. References [14, 27] shows that classically, DKL (pkq) can be approximated within constant additive error with high (n) success probability if and only if Mp = Ω( logn n ) and Mq = Ω( nf log n ). Quantumly, we are given unitary oracles Ôp and Ôq defined by (1.7). Algorithm 3 below estimates the KL-divergence between p and q, which is similar to Algorithm 2 that uses EstAmp0 , while adapts f to be mutually defined by p and q. Algorithm 3: Estimate the KL divergence of p = (pi )ni=1 and q = (qi )ni=1 on [n]. 2 2 2  2  log2 (nf (n)/2 ) log3/2 log (nf(n)/ ) log log log (nf(n)/ ) ; 1 Set l = Θ  2 Regard the following subroutine as A: 3 Draw a sample i ∈ [n] according to p; √ 4 Use the modified amplitude estimation procedure EstAmp0 with 2dlog2 ( n/)e and √ 2dlog2 ( nf (n)/)e quantum queries to p and q to obtain estimates p̃i and q̃i , respectively; 5 Output x̃i = log p̃i − log q̃i ; 6 Use A for l times in Theorem 2.1 and outputs an estimation D̃KL (pkq) of DKL (pkq); Theorem 4.1. For (p, q) ∈ Un,f (n) , Algorithm 3 approximates DKL (pkq) within an additive error √ √  (n)   > 0 with success probability at least 23 using Õ 2n quantum queries to p and Õ nf quantum 2  queries to q, where Õ hides polynomials terms of log n, log 1/, and log f (n). Proof. If the estimates p̃i and q̃i were precisely accurate, the expectation of the subroutine’s output P would be E := i∈[n] pi ·(log pi −log qi ) = DKL (pkq). On the one hand, we bound how far the actual expectation of the subroutine’s output Ẽ is from its exact value E. By linearity of expectation, X   |Ẽ − E| ≤ pi E (log p̃i − log pi ) + (log q̃i − log qi ) (4.2) i∈[n] ≤ X   X   pi E log p̃i − log pi + pi E log q̃i − log qi i∈[n] ≤ X (4.3) i∈[n] X     pi E log p̃i − log pi + f (n) qi E log q̃i − log qi , i∈[n] i∈[n] 10 (4.4) where (4.4) comes from the definition of Un,f (n) in (4.1). By the proof of Theorem 3.1, in particular √ √ equation (3.4), 2dlog2 ( n/)e and 2dlog2 ( nf (n)/)e quantum queries to p and q give    X X     qi E log q̃i − log qi = O pi E log p̃i − log pi = O() and , (4.5) f (n) i∈[n] i∈[n] respectively. Plugging them into (4.4) and rescaling Algorithm 3 by a large enough constant, we get |Ẽ − E| ≤ 2 . On the other hand, the variance of the random variable is at most X X X pi (log p̃i − log q̃i )2 = pi (log q̃i − log p̃i )2 + pi (log p̃i − log q̃i )2 . (4.6) i: p̃i ≥q̃i i: p̃i <q̃i i∈[n] For the first term in (4.6), because EstAmp0 outputs p̃i such that p̃i ≥ sin2 ( dlog2 (√πn/)e+1 ) ≥ 2 for any i, we have X pi (log q̃i − log p̃i )2 ≤ X i: p̃i <q̃i i: p̃i <q̃i  2 2  4n 2 pi log 1 − log ≤ log 2 . 4n  For the second term in (4.6), we have X X pi (log p̃i − log q̃i )2 ≤ pi (log f (n))2 ≤ (log f (n))2 . i: p̃i ≥q̃i 2 4n (4.7) (4.8) i: p̃i ≥q̃i Plugging (4.7) and (4.8) into (4.6), the variance of the random variable is at most  log  4n 2 nf (n) 2  2 + (log f (n)) = O log . 2 2 (4.9) As a result, by Corollary 2.1 we can approximate Ẽ up to additive error /2 with success probability √  √ √ n 1 n/)e dlog ( 2 = Õ 2 quantum queries to p and Õ( 1 ) · 2dlog2 ( nf (n)/)e = at least 2/3 using Õ(  ) · 2 √ (n)  Õ nf quantum queries to q, respectively. Together with |Ẽ − E| ≤ /2, Algorithm 3 approxi2 mates E = DKL (pkq) up to additive error  with success probability at least 2/3. 5 Non-integer Rényi entropy estimation Recall the classical query complexity of non-integer and integer Rényi entropy estimations are different [4]. Quantumly, we also consider them separately; in this section, we consider α-Rényi entropy estimation Pfor general non-integer α1 > 0. Let Pα (p) := ni=1 pαi . Since Hα (p) = 1−α log Pα (p), to approximate Hα (p) within an additive error  > 0 it suffices to approximate Pα (p) within a multiplicative error e(α−1) − 1 = Θ(). 5.1 Case 1: α > 1, α ∈ /N We develop Algorithm 4 to approximate Pα (p) with a multiplicative error . Theorem 5.1. The output of Algorithm 4 approximates Pα (p) within a multiplicative error 0 < 1− 1   ≤ 1/4 with success probability at least 1 − δ for some δ > 0 using Õ n 22α quantum queries to p, where Õ hides polynomials terms of log n, log 1/, and log 1/δ. 11 Algorithm 4: Estimate the α-power sum Pα (p) of p = (pi )ni=1 on [n], α > 1, α ∈ / N. Regard the following subroutine as A: Draw a sample i ∈ [n] according to p; √ Use the amplitude estimation procedure EstAmp with M = 2dlog2 ( to obtain an estimate p̃i of pi ; Output x̃i = p̃α−1 ; i 1 2 3 4 5 6 7 8 n  √ log( n ))e+1  queries Input parameters (α, , δ), where  is the multiplicative error and δ is the failure probability; if α < 1 + log1 n then Take a = 1e and b = 1 as lower and upper bounds on Pα (p), respectively; else Recursively call Algorithm 4 with α0 = α(1 + log1 n )−1 ,  = 1/4, and δ = 12 log 1n log α therein to give an estimate P̃α0 (p) of Pα0 (p). For simplicity, denote P := P̃α0 (p). Take 1+ 1 1+ 1 log n log n a = (3P/4)e and b = 5P as lower and upper bounds on Pα (p), respectively; 4 1− 1 1− 1 1− 1  Set l = Θ n 2  2α log3/2 ( n 2  2α ) log log( n 2  2α ) ; Use A for l executions in Theorem 2.2 using a and b as auxiliary information and output an estimation of Pα (p); Run Line 1 to Line 7 for d48 log 1δ e executions and take the median of all outputs in Line 7, denoted as P̃α (p). Output P̃α (p); Proof of Theorem 5.1. First, we design a subroutine A in Algorithm 4 to approximate Pα (p) following the same principle asPin Algorithm 2. If the estimate p̃i in A were precisely accurate, its expectation would be E := i∈[n] pi · pα−1 = Pα (p). To be precise, we bound how far the actual i expectation of the subroutine’s output Ẽ is from the exact value Pα (p). In Lemma 5.1, we show √ √ n n that when taking M = 2dlog2 (  log(  ))e+1 queries in EstAmp, we have |Ẽ − E| = O(E). As a result, to approximate Pα (p) within multiplicative error Θ(), it is equivalent to approximate Ẽ within multiplicative error Θ(). Recall Theorem 2.2 showed that if the variance of the random variable output by A is at most σ 2 Ẽ 2 for a known σ, and if we can obtain two values a, b such that Ẽ ∈ [a, b], then Õ(σb/a) executions of A suffice to approximate Ẽ within multiplicative error  with success probability at least 9/10. In the main body of the algorithm (Line 1 to Line 8), we use Theorem 2.2 to approximate Ẽ. On the one hand, in Lemma 5.2, we show that for α > 1 and √ large enough n, the variance is at most 5n1−1/α Ẽ 2 with probability at least π82 . This gives σ = 5n1−1/α = O(n1/2−1/2α ). On the other hand, we need to compute the lower bound a and upper bound b. A key observation (Lemma 5.3) is that for any 0 < α1 < α2 , we have X pαi 2  α1 α2 i∈[n] Because n1/ log n = e, if α2 α1 X ≤ pαi 1 α ≤n 1− α1 i∈[n] 2 X pαi 2  α1 α2 . (5.1) i∈[n] = 1 + O( log1 n ), then X i∈[n] pαi 1  X  α1  α 2 α2 =Θ pi . i∈[n] 12 (5.2) As a result, we compute a and b by recursively calling Algorithm 4 to estimate Pα0 (p) for α0 = α/(1 + 1/ log n), which is used to compute the lower bound a and upper bound b in Line 5; the recursive call keeps until α < 1 + log1 n , when a = 1e and b = 1 (as in Line 3) are simply lower and upper bounds on Pα (p) by (5.1). To be precise, in Lemma 5.4, we prove that b/a < 4e = O(1), and with probability at least 1/e1/12 > 0.92, a and b are indeed lower and upper bounds on Pα (p), respectively; furthermore, in Line 5, Algorithm 4 is recursively called by at most log n log α times, and each recursive call 1 takes at most Õ(n1− 2α ) queries. This promises that when we apply Corollary 2.2, the cost Ca,b is dominated by the query cost from Algorithm 10. Combining all points above, Corollary 2.2 approximates Ẽ up to multiplicative error Θ() with success probability at least π82 · 0.92 · 9/10 > 2/3 using log n log α · Õ  4e · √5n1−1/α   √ dlog2 ( ·2 n  log( √ n ))e+1  = Õ  n1−1/2α  2 (5.3) quantum queries. Together with |Ẽ − E| = O(E) and rescale l, M by a large enough constant, Line 1 to Line 7 in Algorithm 4 approximates E = Pα (p) up to multiplicative error  with success probability at least 2/3. Finally, in Lemma 5.5, we show that after repeating the procedure for d48 log 1δ e executions and taking the median P̃α (p) (as in Line 8), the success probability that P̃α (p) approximates Pα (p) within multiplicative error  is boosted to 1 − δ. It remains to prove the lemmas mentioned above. 5.1.1 Expectation of A is -close to Pα (p) Lemma 5.1. |Ẽ − E| = O(E). √ √ Proof of Lemma 5.1. For convenience, denote m = dlog2 ( n/ log( n/))e + 1, and S0 , S1 , . . . , Sm the same as in Section 3. We still have (3.1). By linearity of expectation, |Ẽ − E| ≤ X m X    X  α−1 pi E p̃α−1 − p pi E p̃α−1 − pα−1 = . i i i i (5.4) j=0 i∈Sj i∈[n] Therefore, to prove |Ẽ − E| = O(E) it suffices to show m X X  X    α−1 pi E p̃α−1 − p = O  pαi . i i j=0 i∈Sj (5.5) i∈[n] For each i ∈ [n] we write pi = sin2 (θi π). Assume k ∈ Z such that k ≤ 2m θi < k + 1. By Theorem 2.3, for any l ∈ {1, 2, . . . , max{k − 1, 2m − k − 1}} the output of EstAmp taking 2m queries satisfies h h i 1 (k + l + 1)π i 2 (k − l − 1)π , Pr p̃ = sin ≤ 2. Pr p̃i = sin2 i m m 2 2 4l (5.6) Furthermore, because sin θi = θi −O(θi3 ), cos θi = 1−O(θi2 ), and (1+θi )2α−1 = 1+(2α−1)θi +o(θi ), sin((θi +  l  2(α−1) 2(α−1) l 2α−3 − sin(θ π) = O . )π) (θ π) i i 2m 2m 13 (5.7) Combining (5.6), (5.7), and the fact that m X X j=0 i∈Sj P2m 1 l=1 l = Θ(m), we have m m 2 X   2j 2 X 1 l 2j 2α−3  α−1 pi E p̃α−1 − p = O s · π · 2 π j i i 2m 4l2 2m 2m j=0 (5.8) l=1 m  π 2α−1 m X  (2α−1)j =O · s 2 . j 22αm (5.9) j=0 On the other side,  X i∈[n] m m  X   · π 2α X  2j 2α  pαi = Θ  = Θ 2αm sj · m π sj 22αj . 2 2 j=0 (5.10) j=0 Therefore, to prove equation (5.5), by (5.9) and (5.10) it suffices to prove m X sj 2(2α−1)j = O j=0 n  log( n  ))e + 1, we have m X (5.11) j=0 √ √ Since m = dlog2 ( m  X  sj 22αj . m sj 2 2m m (2α−1)j √ n  , ≥ =O thus m  √n X 2m j=0 √  m ≥ n 2m . Therefore, it suffices to show  sj 22αj . (5.12) j=0 If α ≥ 3/2, by Hölder’s inequality we have m X sj  1 2α m X j=0 sj 22αj  2α−1 2α ≥ j=0 m X sj 2(2α−1)j . (5.13) j=0 By equation (3.1), this gives n 1 2α−1 m X sj 2 2αj  ≥ m X j=0 sj 2 (2α−1)j m  X j=0 sj 2 (2α−1)j  1 2α−1 . (5.14) j=0 By Hölder’s inequality and also equation (3.1), we have m X sj m  2α−3  X 2α−1 j=0 sj 2(2α−1)j  2 2α−1 ≥ m X sj 22j = Θ(22m ). (5.15) j=0 j=0 This is equivalent to 2α−3 n 2(2α−1) m X sj 2(2α−1)j  1 2α−1 ≥ Θ(2m ). (5.16) j=0 Combining (5.14) and (5.16), we get exactly (5.12). If 1 < α < 3/2, by Hölder’s inequality we have m X sj 22αj j=0 m X j=0 sj 2 2j m 1X α j=0 m 2α−1  X 2 sj sj  α−1 α  3−2α j=0 2 ≥ ≥ m X j=0 m X j=0 14 sj 22j ; (5.17) sj 2(2α−1)j . (5.18) By equation (3.1), the two inequalities above give m X sj 22αj ≥ n1−α 22αm m X and j=0 sj 2(2α−1)j ≤ n1.5−α 2(2α−1)m , (5.19) j=0 which give (5.12). 5.1.2 Bound the variance of A by the square of its expectation Lemma 5.2. With probability at least most 5n1−1/α Ẽ 2 . 8 , π2 the variance of the random variable output by A is at Pn α−1 Proof of Lemma 5.2. The expectation and variance of the output by A are Ẽ = i=1 pi · p̃i  P P α−1 2 and i∈[n] pi · (p̃iα−1 )2 − , respectively. Therefore, it suffices to show that with i∈[n] pi · p̃i 8 probability at least π2 , X pi · (p̃α−1 )2 ≤ 5n1−1/α i n X pi · p̃α−1 i 2 . (5.20) i=1 i∈[n] 8 , π2 By Theorem 2.3, with probability at least we have √ √ 2π pi π pi |p̃i − pi | ≤ ≤ √ 2m n i ∈ [n]. (5.21) For convenience, denote p := pi∗ to be the maximal one among p1 , . . . , pn , i.e., p = maxi∈{1,...,n} pi . We also denote p̃ := p̃i∗ . Then we have Pn i=1 pi P i∈[n] pi · p̃iα−1 2 · (p̃α−1 )2 i Pn ≥ i=1 pi p̃α−1 · P · p̃α−1 i i∈[n] pi 2 · p̃α−1 i Pn = α−1 i=1 pi · p̃i . p̃α−1 (5.22) Furthermore, because xα is a convex function in [0, 1], by (5.21) and Jensen’s inequality we have P Pn α−1 p · p̃α−1 + i6=i∗ pi · p̃α−1 i i=1 pi · p̃i = (5.23) p̃α−1 p̃α−1 √    √  π (1−p)/(n−1) α−1 π p α−1 1−p 1−p √ p p + √n + (n − 1) · n−1 − n−1 n ≥ (5.24)  √ α−1 π p p + √n  1 − p − π √1 − p α−1 ≈ p + (1 − p) . (5.25) √ np + π np Therefore, it suffices to show that for large enough n,  1 − p − π √1 − p α−1 p + (1 − p) ≥ 0.2n−(1−1/α) . √ np + π np 15 (5.26) If p ≥ 0.2n−(1−1/α) , equation (5.26) directly follows. If p < 0.2n−(1−1/α) ,  1 − p − π √1 − p α−1 1−1/α lim n · (1 − p) √ n→∞ np + π np  1 − 0.2n−(1−1/α) − π α−1 √ ≥ lim n1−1/α (1 − 0.2n−(1−1/α) ) n→∞ 0.2n1/α + 0.2πn1/2α  n1/α (1 − 0.2n−(1−1/α) − π) α−1 √ = lim (1 − 0.2n−(1−1/α) ) n→∞ 0.2n1/α + 0.2πn1/2α  1 − π α−1 =1· > 1 > 0.2, 0.2 (5.27) (5.28) (5.29) 1−3.2/4 where (5.29) is true because 1−π = 1. Because (5.25) only omits lower order terms and 0.2 > 0.2 the limit in (5.29) is a constant larger than 0.2, Lemma 5.2 follows. 5.1.3 Give tight bounds on Pα (p) by Pα0 (p) Lemma 5.3. For any distribution (pi )ni=1 and 0 < α1 < α2 , we have X pαi 2  α1 α2 ≤ i∈[n] X pαi 1 ≤ n α 1− α1 2 X i∈[n]  α1 α2 pαi 2 . (5.30) i∈[n] Proof of Lemma 5.3. On the one hand, by the generalized mean inequality, we have α2 i∈[n] pi P ! 1 α2 P ≥ n α1 i∈[n] pi ! 1 α1 n , (5.31) which gives the second inequality in (5.30). On the other hand, since αα21 ≤ 1 and 0≤ P pαi 2 α2 j∈[n] pj ≤1 ∀ i ∈ [n], (5.32) we have α1 X i∈[n] pi P  α1 = α 2 α2 i∈[n] j∈[n] pj P pαi 2 α2 j∈[n] pj P ! α1 α2 ≥ pαi 2 X α2 j∈[n] pj P i∈[n] = 1, (5.33) which is equivalent to the first inequality in (5.30). 5.1.4 Analyze the recursive calls Lemma 5.4. With probability at least 0.92, the a and b in Line 3 or Line 5 of Algorithm 4 are indeed lower and upper bounds on Pα (p), respectively, and b/a = O(1); furthermore, in Line 5, Algorithm 4 is recursively called for at most log n log α executions, and each recursive call takes at 1 most Õ(n1− 2α ) queries. Proof of Lemma 5.4. We decompose the proof into two parts: 16 • In Line 5, Algorithm 4 is recursively called for at most log n log α executions, and 1 each recursive call takes at most Õ(n1− 2α ) queries: 1 −1 and the log n ) log α most log(1+ ≤ 1 ) log n Because each recursive call of Algorithm 4 reduces α by multiplying (1 + recursion ends when α < 1 + log1 n , the total number of recursive calls is at log n log α. When α < 1 + log1 n , a and b are set in Line 3 and no extra queries are needed; when Line 5 calls α(1 + log1 n )−k -power sum estimation for some k ∈ N, by induction on k, we see that (1+1/ log n)k  1 2α ≤ Õ(n1− 2α ) queries. As a result, when we apply this call takes at most Õ n1− Corollary 2.2, the cost Ca,b is dominated by the query cost from Algorithm 10. • With probability at least 0.92, a and b are lower and upper bounds on Pα (p) respectively, and b/a = O(1): P P When 1 < α < 1 + log1 n , on the one hand we have ni=1 pαi ≤ ni=1 pi = 1; on the other hand, 1 because n log n = e, by Lemma 5.3 we have n X pαi Pn i=1 pi nα−1 ≥ i=1 α 1 ≥ . e (5.34) Therefore, a = 1/e and b = 1 in Line 3 are lower and upper bounds on Pα (p) respectively, and b/a = e = O(1). When α > 1 + log1 n , for convenience denote α0 = α(1 + log1 n )−1 . As justified above, the total number of recursive calls in Line 5 is at most log n log α. Because we take δ = 12 log 1n log α in Line 5, with probability at least  log n log α 1 1 1− (5.35) ≥ 1/12 > 0.92, 12 log n log α e the output of every recursive call is within 1/4-multiplicative error. As a result, the P in Line 1 P 0 5 satisfies 3P/4 ≤ ni=1 pαi ≤ 5P/4. Combining this with Lemma 5.3 and using n log n = e, we have 1+ log1 n (3P/4) e ≤ n X i=1 pαi ≤  5P 1+ 1 log n 4 . (5.36) In other words, a and b are indeed lower and upper bounds on Pα (p), respectively. Furthermore, b/a = O(1) because  5 1+ 1 b log n < 4e = O(1). =e· a 3 5.1.5 (5.37) Boost the success probability Lemma 5.5. By repeating Line 1 to Line 7 in Algorithm 4 for d48 log 1δ e executions and taking the median P̃α (p), the success probability is boosted to 1 − δ. 17 Proof of Lemma 5.5. Denote the outputs after running Line 1 to Line 7 for d48 log 1δ e executions 1 as P̃α (p)(1) , . . . , P̃α (p)(d48 log δ e) , respectively. Based on the correctness of Lemma 5.1, Lemma 5.2, and Lemma 5.4, for each i ∈ {1, . . . , d48 log 1δ e}, with probability at least 2/3 we have |P̃α (p)(i) − Pα (p)| ≤ Pα (p). (5.38) For each i ∈ {1, . . . , d48 log 1δ e}, denote Xi to be a Boolean random variable such that Xi = 1 if (5.38) holds, and Xi = 0 otherwise. Then Pr[Xi = 1] ≥ 2/3. Because in Line 8 the output P̃α (p) Pd48 log 1 e 1 is the median of all P̃α (p)(1) , . . . , P̃α (p)(d48 log δ e) , |P̃α (p) − Pα (p)| > Pα (p) leads to i=1 δ Xi < d48 log 1δ e/2. On the other hand, by Chernoff bound we have 1 Pr log δ e h d48X i=1 Xi <  2/3d48 log 1 e · (1/4)2  d48 log 1δ e i δ ≤ exp − ≤ δ. 2 2 (5.39) Therefore, with probability at least 1 − δ, we have |P̃α (p) − Pα (p)| ≤ Pα (p). 5.2 Case 2: 0 < α < 1 When 0 < α < 1, our quantum algorithm follows the same structure as Algorithm 4: Algorithm 5: Estimate the α-power sum Pα (p) of p = (pi )ni=1 on [n], 0 < α < 1. Regard the following subroutine as A: Draw a sample i ∈ [n] according to p; Use the amplitude estimation procedure EstAmp with M = 2dlog2 ( queries to obtain an estimate p̃i of pi ; Output x̃i = p̃α−1 ; i 1 2 3 4 5 6 7 8 n1/2α  log( n 1/2α ))e+1  Input parameters (α, , δ), where  is the multiplicative error and δ is the failure probability; if α > 1 − log1 n then Take a = 1 and b = e as lower and upper bounds on Pα (p), respectively; else Recursively call Algorithm 5 with α0 = α(1 − log1 n )−1 ,  = 1/2, and δ = 12 log n1log 1/α therein to give an estimate P̃α0 (p) of Pα0 (p). For simplicity, denote P := P̃α0 (p). Take 1− 1 1− 1 a = (P/2) log n and b = e(2P ) log n as lower and upper bounds on Pα (p), respectively; 1 −1 1 −1  1 −1 Set l = Θ n 2α 2 log3/2 ( n 2α 2 ) log log( n 2α 2 ) ; Use A for l executions in Theorem 2.2 using a and b as auxiliary information and output an estimation of Pα (p); Run Line 1 to Line 7 for d48 log 1δ e executions and take the median of all outputs in Line 7, denoted as P̃α (p). Output P̃α (p); The main difference is that, in the case α > 1, Algorithm 4 makes α0 smaller and smaller by multiplying (1 + log1 n )−1 each time, whereas in the case 0 < α < 1, Algorithm 5 makes α0 larger and larger by multiplying (1 − log1 n )−1 each time; nevertheless, both recursions end when α0 is close enough to 1. On the more technical level, they have different M in A, different upper bounds on the variance of A, and different expressions for a and b in Line 3 and Line 5. 18 Theorem 5.2. The output of Algorithm 5 approximates Pα (p) within a multiplicative error 0 < 1 −1   ≤ O(1) with success probability at least 1 − δ for some δ > 0 using Õ n α2 2 quantum queries to p, where Õ hides polynomials terms of log n, log 1/, and log 1/δ. Before we give the formal proof of Theorem 5.2, we compare the similarities and differences between Algorithm 4 and Algorithm 5, listed below: • In both algorithms, the subroutine A has the same structure, and is designed to estimate Pα (p). However, to make the expectation of A -close to Pα (p), the EstAmp in Algorithm 4 √ √ suffices to take M = 2dlog2 ( n  n ))e+1  log( dlog2 ( n Algorithm 5 needs to take M = 2 queries (see Lemma 5.1), whereas the EstAmp in 1/2α  log( n 1/2α ))e+1  queries (see Lemma 5.6); • In both algorithms, we use Theorem 2.2 to approximate the expectation of A (denoted Ẽ), hence they both need to upper-bound the variance of A by a multiple of Ẽ 2 . However, technically the proofs are different, and we obtain different upper bounds in Lemma 5.2 and Lemma 5.7, respectively; • Since both algorithms use Theorem 2.2, they both need to compute a lower bound a and upper bound b on Pα (p). Both algorithms achieve this by observing Lemma 5.3, and they both compute a and b by recursively call the estimation of Pα0 (p) for some α0 closer to 1. However, in the case α > 1, Algorithm 4 makes α0 smaller and smaller by multiplying (1 + log1 n )−1 each time, and ends the recursion when α0 < 1 + log1 n ; in the case 0 < α < 1, Algorithm 5 makes α0 larger and larger by multiplying (1 − log1 n )−1 each time, and ends the recursion when α0 > 1− log1 n . This leads to different expressions for a and b in Line 3 and Line 5 of both algorithms, and technically the proofs for Lemma 5.4 and Lemma 5.8 is different; • Both algorithms boost the success probability to 1−δ by repeating the algorithm for d48 log 1δ e executions and taking the median, and their correctness is both promised by Lemma 5.5. Proof of Theorem 5.2. First, if the estimate p̃i in the subroutine AP of Algorithm 5 were precisely α−1 := = Pα (p). To be accurate, the expectation of the subroutine’s output would be E i∈[n] pi · pi precise, we bound how far the actual expectation of the subroutine’s output Ẽ is from the exact n1/2α n1/2α value Pα (p). In Lemma 5.6, we show that when taking M = 2dlog2 (  log(  ))e+1 queries in EstAmp, we have |Ẽ − E| = O(E). As a result, to approximate Pα (p) within multiplicative error Θ(), it is equivalent to approximate Ẽ within multiplicative error Θ(). Recall Theorem 2.2 showed that if the variance of the random variable output by A is at most σ 2 Ẽ 2 for a known σ, and if we can obtain two values a, b such that Ẽ ∈ [a, b], then Õ(σb/a) executions of A suffice to approximate Ẽ within multiplicative error  with success probability at least 9/10. In the main body of the algorithm (Line 1 to Line 8), we use Theorem 2.2 to approximate Ẽ. On the one hand, in Lemma 5.7, we show that for√any 0 < α < 1, the variance is at most 1/α−1 2n Ẽ 2 with probability at least π82 . This gives σ = 2n1/α−1 = O(n1/2α−1/2 ). On the other hand, we need to compute the lower bound a and upper bound b. As stated in the proof of Theorem 5.1, for any 0 < α1 < α2 with αα21 = 1 + O( log1 n ), X i∈[n] pαi 1  X  α1  α 2 α2 =Θ pi . i∈[n] 19 (5.40) As a result, we compute a and b by recursively calling Algorithm 5 to estimate Pα0 (p) for α0 = α/(1 − 1/ log n), which is used to compute the lower bound a and upper bound b in Line 5; the recursive call keeps until α > 1 − log1 n , when a = 1 and b = e (as in Line 3) are simply lower and upper bounds on Pα (p). To be precise, in Lemma 5.8, we prove that b/a ≤ 4e = O(1), and with probability at least 1/e1/12 > 0.92, a and b are indeed lower and upper bounds on Pα (p), respectively; furthermore, in Line 5, Algorithm 5 is recursively called by at most log n log α1 times, and each recursive call 1 1 takes at most Õ(n α − 2 ) queries. This promises that when we apply Corollary 2.2, the cost Ca,b is dominated by the query cost from Algorithm 10. Combining all points above, Corollary 2.2 approximates Ẽ up to multiplicative error Θ() with success probability at least π82 · 0.92 · 9/10 > 2/3 using  4e · 1 log n log · Õ α √ 1 1 nα−2  2n1/α−1  dlog2 ( n1/2α log( n1/2α ))e+1   = Õ ·2  2 (5.41) quantum queries. Together with |Ẽ − E| = O(E) and rescale l, M by a large enough constant, Line 1 to Line 7 in Algorithm 5 approximates E = Pα (p) up to multiplicative error  with success probability at least 2/3. Finally, following from Lemma 5.5, after repeating the procedure for d48 log 1δ e executions and taking the median P̃α (p) (as in Line 8), the success probability that P̃α (p) approximates Pα (p) within multiplicative error  is boosted to 1 − δ. It remains to prove the lemmas mentioned above. 5.2.1 Expectation of A is -close to Pα (p) Lemma 5.6. |Ẽ − E| = O(E). 1/2α 1/2α Proof of Lemma 5.6. For convenience, denote m = dlog2 ( n  log( n  ))e + 1, and S0 , S1 , . . . , Sm the same as previous definitions. We still have (3.1). By linearity of expectation, X |Ẽ − E| ≤ h pi E i∈[n] 1 p̃1−α i − 1 i p1−α i = m X X   pi E p̃α−1 − pα−1 . i i (5.42) j=0 i∈Sj Therefore, to prove |Ẽ − E| = O(E) it suffices to show m X X  X    α−1 pi E p̃α−1 − p = O  pαi . i i j=0 i∈Sj (5.43) i∈[n] Similar to the proof of Lemma 5.1, we have m X X m  π 2α−1 m X    (2α−1)j pi E p̃iα−1 − pα−1 = O · s 2 . j i 22αm (5.44) m m   · π 2α X   X 2j 2α  pαi = Θ  sj · m π = Θ 2αm sj 22αj . 2 2 (5.45) j=0 i∈Sj j=0 On the other side,  X i∈[n] j=0 j=0 20 Therefore, to prove Equation (5.43), by (5.44) and (5.45) it suffices to prove m X sj 2(2α−1)j = O 1/2α  1/2α log( n  2m m ))e + 1, we have m X sj 2 (2α−1)j =O 1/α Pm 2j j=0 sj 2 Pm j=0 sj 2  2αj 1/α n1/2α  , 2m thus  m ≥ n1/2α 2m .  sj 22αj . Therefore, it suffices (5.47) j=0 ; as a result, Pm ≤ ≥ m  n1/2α X j=0 Since sj ∈ N, sj ≤ sj (5.46) j=0 j=0 Since m = dlog2 ( n to prove m   X sj 22αj . m 2αj )1/α j=0 (sj 2  Pm 2αj 1/α j=0 sj 2 = m  X m 1/α X sj 22αj s 22αj Pm j P = 1. (5.48) ≤ m 2αk 2αk k=0 sk 2 k=0 sk 2 j=0 j=0 Plugging (3.1) into the inequality above, we have m X sj 22αj  1 2α = Ω(2m ). (5.49) j=0 On the other side, by Hölder’s inequality we have m X sj  j=0 1 2α m X sj 2 2αj  2α−1 2α ≥ j=0 m X sj 2(2α−1)j . (5.50) j=0 Combining (3.1), (5.49), and (5.50), we get exactly (5.47). 5.2.2 Bound the variance of A by the square of its expectation Lemma 5.7. With probability at least π82 , the variance of the random variable output by A is at most 2n1/α−1 Ẽ 2 . Pn P P )2 − Proof of Lemma 5.7. Because Ẽ = ni=1 pi · p̃α−1 and the variance is ni=1 pi · (p̃α−1 i=1 pi · i i  Pn α−1 2 α−1 2 p̃i ≤ i=1 pi · (p̃i ) , it suffices to show that n X pi · (p̃α−1 )2 i ≤ 2n 1/α−1 i=1 By Theorem 2.3, with probability at least n X pi · p̃α−1 i 2 . (5.51) i=1 8 , π2 we have √ √ 2π pi π pi |p̃i − pi | ≤ ≤ 1/2α 2m n 21 i ∈ [n]. (5.52) As a result, n X √ n  X π pi −2(1−α) pi pi − 1/2α n i=1 n −2(1−α)  X π = p2α−1 1 − √ i n1/2α pi i=1 n   X π ≈ p2α−1 1 + 2(1 − α) √ i n1/2α pi i=1 pi (p̃iα−1 )2 ≤ i=1 = n X p2α−1 + i i=1 (5.53) (5.54) (5.55) n 2(1 − α)π X 2α−0.5 pi . n1/2α i=1 (5.56) Furthermore, n n n n n  X X √ X √  X 2α−1  √ 2α−1 X 2α−0.5 2α−1 ≥ n pi pj pj pi = ≥ pi pi . i=1 j=1 i=1 i=j=1 (5.57) i=1 Plugging this into (5.56), we have n X pi (p̃iα−1 )2 i=1 n 2(1 − α)π  X 2α−1 pi . ≤ 1 + 1/2α−1/2 n i=1  (5.58) Using similar techniques, we can show n X pi · p̃α−1 i 2 i=1 n 2(1 − α)π  X α 2 pi . ≥ 1− n1/α−1 i=1  (5.59) Since 0 < α < 1, lim 1 + n→∞ 2(1 − α)π = 1, n1/2α−1/2 lim 1 − n→∞ 2(1 − α)π = 1. n1/α−1 (5.60) Because (5.55) only omits lower order terms and the limits in (5.60) are both 1, to prove (5.51) it suffices to prove that for large enough n, n X p2α−1 i ≤n 1/α−1 i=1 n X pαi 2 . (5.61) i=1 By generalized mean inequality, we have n 1 X n p2α−1 i  1 2α−1 i=1 ≤ n 1 X n pαi 1 α . (5.62) i=1 Therefore, n X p2α−1 ≤ n1− i 2α−1 α i=1 n X pαi  2α−1 α = n1/α−1 n X i=1 i=1 Hence the result follows. 22 pαi 2−1/α ≤ n1/α−1 n X i=1 pαi 2 . (5.63) 5.2.3 Analyze the recursive calls Lemma 5.8. With probability at least 0.92, the a and b in Line 3 or Line 5 of Algorithm 5 are indeed lower and upper bounds on Pα (p), respectively, and b/a = O(1); furthermore, in Line 5, Algorithm 5 is recursively called for at most log n log α1 executions, and each recursive call takes at 1 1 most Õ(n α − 2 ) queries. Proof of Lemma 5.8. Similar to Lemma 5.4, we decompose the proof into two parts: • In Line 5, Algorithm 5 is recursively called for at most log n log α1 executions, and 1 1 each recursive call takes at most Õ(n α − 2 ) queries: Because each recursive call of Algorithm 5 increases α by multiplying (1 − log1 n )−1 and the recursion ends when α > 1 − log1 n , the total number of recursive calls is at most log α log(1− log1 n ) ≤ log n log α1 . When α > 1 − log1 n , a and b are set in Line 3 and no extra queries are needed; when Line 5 calls α(1 − log1 n )−k -power sum estimation for some k ∈ N, by induction on k, we see that  (1−1/ log n)k 1 1 − 12 α this call takes at most Õ n ≤ Õ(n α − 2 ) queries. As a result, when we apply Corollary 2.2, the cost Ca,b is dominated by the query cost from Algorithm 10. • With probability at least 0.92, a and b are lower and upper bounds on Pα (p) respectively, and b/a = O(1): P P When 1 − log1 n < α < 1, on the one hand we have ni=1 pαi ≥ ni=1 pi = 1; on the other hand, 1 because n log n = e, by Lemma 5.3 we have n X pαi ≤ n1−α n X α pi ≤ e. i=1 (5.64) i=1 Therefore, a = 1 and b = e in Line 3 are lower and upper bounds on Pα (p) respectively, and b/a = e = O(1). When α < 1 − log1 n , for convenience denote α0 = α(1 − log1 n )−1 . As justified above, the total number of recursive calls in Line 5 is at most log n log α1 . Because we take δ = 12 log n1log 1/α in Line 5, with probability at least  log n log 1/α 1 1 1− ≥ 1/12 > 0.92, (5.65) 12 log n log 1/α e the output of every recursive call is within 1/2-multiplicative error. As a result, the P in Line 1 P 0 5 satisfies P/2 ≤ ni=1 pαi ≤ 2P . Combining this with Lemma 5.3 and using n log n = e, we have n X 1− log1 n 1− 1 (P/2) ≤ pαi ≤ e(2P ) log n . (5.66) i=1 In other words, a and b are indeed lower and upper bounds on Pα (p), respectively. Furthermore, b/a = O(1) because b 1− 1 = e · 4 log n ≤ 4e = O(1). a 23 (5.67) 6 Integer Rényi entropy estimation Recall the classical query complexity of α-Rényi entropy estimation for α ∈ N, α ≥ 2 is Θ(n1−1/α ) [4], which is smaller than non-integer cases. Quantumly, we also provide a more significant speedup. Given the oracle Op : [S] → [n] in (1.7), we denote the occurrences of 1, 2, . . . , n among Op (1), . . ., Op (S) as m1 , . . . , mn , respectively. A key observation is that by (1.6), we have n n X X α −α Pα (p) = (mi /S) = S mαi . i=1 (6.1) i=1 P α Therefore, it suffices to approximate i∈[n] mi , which is known as the α-frequency moment of Op (1), . . . , Op (S). Based on the quantum algorithm for α-distinctness [9], Montanaro [42] proved: Fact 6.1 ([42], Step 3b-step 3e in Algorithm 2; Lemma 4). Fix l where l ∈ {1, . . . , n}. Let s1 , . . . , sl ∈ [S] be picked uniformly at random, and denote the number of α-wise collisions in {Op (s1 ), . . . , Op (sl )} as C(s1 , . . . , sl ). Then: • C(s1 , . . . , sl ) can be computed using O(lν log(l/2 )) queries to Ôp with failure probability at most O(2 /l), where ν := 1 − 2α−2 /(2α − 1) < 34 ;  • E[C(s1 , . . . , sl )] = αl Pα (p) and Var[C(s1 , . . . , sl )] = O(1). However, a direct application of [42] will lead to a complexity depending on S (in particular, l in Fact 6.1 can be as large as S) rather than n. Our solution is Algorithm 6 that is almost the same as Algorithm 2 in [42] except Line 1 and Line 2, where we set 2dlog2 αne as an upper bound on l. We claim that such choice of l is valid because by the pigeonhole principle, αn elements Op (s1 ), . . . , Op (sαn ) in [n] must have an α-collision, so the first for-loop must terminate at some i ≤ dlog2 αne. With this modification, we have Theorem 6.1 for integer Rényi entropy estimation. Algorithm 6: Estimate the α-power sum Pα (p) of p = (pi )ni=1 on [n], α > 1, α ∈ N. 1 2 3 4 5 6 7 8 9 10 Set l = 2dlog2 αne ; for i = 0, . . . , dlog2 αne do Pick s1 , . . . , s2i ∈ [S] uniformly at random and let S be the sequence Op (s1 ), . . . , Op (s2i ); Apply the α-distinctness algorithm in [9] to S with failure probability 10dlog1 αne ; If it returns a set of α equal elements, set l = 2i and terminate the loop; 2 Set M = dK/2 e for some K = Θ(1) ; for r = 1, . . . , M do Pick s1 , . . . , sl ∈ [S] uniformly at random; Apply the first bullet in Fact 6.1 to give an estimate C (r) of the number of α-wise collisions in {Op (s1 ), . . . , Op (sl )}; PM (r) Output P̃α (p) = 1 l r=1 C ; M ( α) Theorem 6.1. Assume α > 1, α ∈ N. Algorithm 6 approximates Pα (p) within a multiplicative 3 (1−1/α)  ν(1−1/α)  error 0 <  ≤ O(1) with success probability at least 23 using Õ n 2 = o n 4 2 quantum 3 α−2 α := queries to p, where ν 1−2 /(2 − 1) < 4 . 24 Our proof of Theorem 6.1 is inspired by the proof of Theorem 5 in [42]. Proof. Because Op takes values in [n], by pigeonhole principle, for any s1 , . . . , sαn ∈ [S] there exists a α-wise collision among Op (s1 ), . . . , Op (sαn ). Therefore, Line 5 terminates the first loop with some l ≤ 2dlog2 αne with probability at least (1 − 1/10dlog2 αne)dlog2 αne ≥ e−1/10 > 0.9. Moreover, tighter bounds on l are established next. On the one hand, by Chebyshev’s inequality B and Fact 6.1, the probability that the first for-loop fails to terminate when l ≤ P (p) 1/α for some α constant B > 0 is at most    1    Var[C(s1 , . . . , sl )] 1 = O Pr C(s1 , . . . , sl ) = 0 ≤ = O . (6.2) E[C(s1 , . . . , sl )]2 l2α Pα (p)2 B 2α  1 Therefore, taking a large enough B ensures that l = O P (p) with failure probability at most 1/α α 1/20. On the other hand, by Markov’s inequality and Fact 6.1, we have   Pr C(s1 , . . . , sl ) ≥ 1 ≤ E[C(s1 , . . . , sl )] = O(lα Pα (p)). (6.3) As a result, the probability that the first for-loop terminates when l ≤ A > 0 is at most   log2 O(Pα (p)) · A Pα (p)1/α for some constant A Pα (p)1/α X 2iα = O(Aα ). (6.4) i=0  1 Therefore, taking a small enough A > 0 ensures that l = Ω P (p) with failure probability at 1/α α  1 most 1/20. In all, we have l = Θ P (p)1/α with probability at least 0.9. α By Fact 6.1, the output E[P̃α (p)] in Line 10 of Algorithm 6 satisfies E[P̃α (p)] = M X 1 M  l α E[C (r) ] = Pα (p), Var[P̃α (p)] = r=1 Therefore, by Chebyshev’s inequality and recall l = Θ M X 1 (M  l α 1 Pα (p)1/α )2  Var[C (r) ] = O  r=1 1  . (6.5) M l2α , we have  1 1 = O . (6.6) M l2α 2 Pα (p)2 K   Taking a large enough constant K in Line 6 of Algorithm 6, we have Pr |P̃α (p)−Pα (p)| ≤ Pα (p) ≥ 0.9. In all, with probability at least 0.9 × 0.9 × 0.9 > 2/3, P̃α (p) approximates Pα (p) within multiplicative error . For the rest of the proof, it suffices to compute the quantum query complexity of Algorithm 6. Because the α-distinctness algorithm on m elements in [9] takes O(mν log(1/δ)) quantum queries Plog l when the success probability is 1−δ, the first for-loop in Algorithm 6 takes i=02 O(2νi logdlog2 αne) = Õ(lν ) = Õ(nν(1−1/α) ) quantum queries because   1 = O(n1−1/α ), (6.7) l=Θ Pα (p)1/α    Pr |P̃α (p) − Pα (p)| ≥ Pα (p) ≤ O ν(1−1/α) 2 nν(1−1/α) 2 following from Pα (p) ≥ n1−α . The second for-loop takes dK/2 e · O(lν log(l/2 )) = Õ( n quantum queries by Fact 6.1 and (6.7). In total, the number of quantum queries is Õ( 25  .  ν(1−1/α)  Remark 6.1. In Theorem 6.1, we regard α as a constant, i.e., the query complexity Õ( n 2 hides the multiple in α. In fact, by analyzing the dependence on α carefully in the above proof, the query complexity of Algorithm 6 is actually  nν(1−1/α)  2 . (6.8) Õ α8α · 2 The dependence on α is super-exponential; therefore, Algorithm 6 is not good enough to approximate min-entropy (i.e., α = ∞). As a result, we give the quantum algorithm for estimating min-entropy separately (see Section 7). 7 Min-entropy estimation Since the min-entropy of p is H∞ (p) = − log maxi∈[n] pi by (1.3), it is equivalent to approximate maxi∈[n] pi within multiplicative error . We propose Algorithm 7 below to achieve this task. Algorithm 7: Estimate maxi∈[n] pi of a discrete distribution p = (pi )ni=1 on [n]. 1 2 3 4 Set λ = 1; while λ ≤ n do n Take M ∼ Poi( 16λlog ). Pick s1 , . . . , sM ∈ [S] uniformly at random and let S be the 2 sequence Op (s1 ), . . . , Op (sM ); n Apply a d 16 log e-distinctness quantum algorithm to S with failure probability at most 2  ; 2 log n n If Line 4 outputs a d 16 log e-collision of elements i∗ ∈ [n], apply Theorem 2.3 to 2 approximate pi∗ with multiplicative error  and output its result; if not, set √ λ ← λ · 1 +  and jump to the start of the loop; 5 6 If λ > n and no output has been given, output 1/n; A key property of the Poisson distribution is that if we take M ∼ Poi(ν) samples from p (as in Line 3), then for each j ∈ [n], the number of occurrences of j in Op (s1 ), . . . , Op (sM ) follows the Poisson distribution Mj ∼ Poi(νpj ), and Mj , Mj 0 are independent for all j 6= j 0 . Furthermore: Lemma 7.1. Let X ∼ Poi(µ). Then, if µ < If µ ≥ 16 log n , 2 √1 1+ · 16 log n , 2 we have h 16 log n i 1 Pr X ≥ ≤ 2; 2  n (7.1) h 16 log n i Pr X ≥ > 0.15. 2 (7.2) we have n n Based on Lemma 7.1, our strategy is to set 16 log as a threshold, take ν = 16λlog as in Line 2 2 3, and gradually increase the parameter λ. For convenience, denote pi∗ = maxi pi . As long as n n ν · pi∗ < 16 log , with high probability there is no d 16 log e-collision in S, the distinctness quantum 2 2 √ algorithm in Line 4 rejects, and λ increases by multiplying 1 +  in Line 5; right after the first n n time when ν · pi∗ ≥ 16 log , with probability at least 0.15, i∗ has a d 16 log e-collision in S, while all 2 2 2 other entries in [n] do not (with failure probability at most 1/n ). In this case, with probability at least Ω(1), the distinctness quantum algorithm in Line 4 captures i∗ , and the quantum counting (Theorem 2.3) in Line 5 computes pi∗ within multiplicative error . 26 Theorem 7.1. Algorithm 7 approximates maxi∈[n] pi within a multiplicative error 0 <  ≤ 1 with  n success probability at least Ω(1) using Õ Q(d 16 log e-distinctness) quantum queries to p, where 2 n n Q(d 16 log e-distinctness) is the quantum query complexity of the d 16 log e-distinctness problem. 2 2 We first prove Lemma 7.1. output Proof of Lemma 7.1. First, we prove (7.1)7 . In [23], it is shown that if λ > 0 and X ∼ Poi(λ), then for any ν > 1 we have Pr[X ≥ νλ] ≤ Taking λ = √1 1+ · 16 log n 2 and ν = √ e−λ λνλ . (νλ)!(1 − 1/ν) (7.3) 1 + , by Sterling’s formula we have r h 16 log n i 2 e−λ λνλ 2 1 e−λ λνλ √ Pr X ≥ ≈ √ ≈ ≤ 2 (νλ)!(1 − 1/ν)  2πνλ(νλ/e)νλ π λ e/2 (1 + 2 )1+/2 !λ . (7.4) Because e/2 (1 + 2 )1+/2 lim →0 !8/2  i 4  ln 1 + →0  2  2 h4  8 i 4   2 = lim exp − 2+ − + O(3 ) →0    2 8 = lim exp[−1 + O()] = e−1 , = lim exp h4 − 8 + →0 (7.5) (7.6) (7.7) we have e/2 (1 + 2 )1+/2 !λ 2 ≈ e− 8 λ ≈ 1 . n2 (7.8) Plugging (7.8) into (7.4), we have r h 16 log n i 2 1 1 1 √ Pr X ≥ . ≤ 2. 2 2  π 16 log n n n (7.9) Now we prove (7.2). A theorem of Ramanujan [28, Question 294] states that for any positive integer M , M −1 X 1 M Mm MM e = + θ(M ) · , 2 m! M! (7.10) m=0 where 1 3 ≤ θ(M ) ≤ 1 2 ∀ M ∈ N. Because ∞ X m=M +1 7 Mm m=0 m! P∞ = eM , by (7.10) we have Mm 2 MM 1 + · ≥ eM . m! 3 M! 2 The tail bound of Poisson distributions is also studied elsewhere, for example, in [40, Exercise 4.7]. 27 (7.11) By Stirling’s formula, M ! ≥ √ 2πM  M M . e As a result, ∞ X  1 1 2 MM 1 Mm √ ≥ eM − · √ − eM . =  M M m! 2 3 2 4.5πM 2πM m=M +1 e (7.12) n We take M = b 16 log c. By (7.12), we have 2 h 16 log n 16 log n i Pr X ≥ = e− 2 2  ∞ X m=M +1 n m ( 16 log ) 2 ≥ e−M −1 m! ∞ X 1 1 Mm . (7.13) ≥ − √ m! 2e e 4.5πM m=M +1 Because 0 <  ≤ 1, we have M ≥ b16 log 2c = 11. Therefore, h 16 log n i 1 1 Pr X ≥ ≥ − √ > 0.15. 2  2e e 4.5π · 11 (7.14) Proof of Theorem 7.1. Denote σ to be the permutation on [n] such that pσ(1) ≥ pσ(2) ≥ · · · ≥ pσ(1) pσ(n) . Without loss of generality, we assume that pσ(2) ≤ 1+ ; otherwise, pσ(2) is close enough to pσ(1) in the sense that applying quantum counting to pσ(2) within multiplicative error  gives an approximation to pσ(1) within multiplicative error 2. We may assume that every call of the n n d 16 log e-distinctness quantum algorithm in Line 4 of Algorithm 7 succeeds if and only if a d 16 log e2 2  log√1+ n  −1 collision exists, because this happens with probability at least 1 − 2 log n ≥ e = Ω(1); for convenience, this is always assumed in the result of the proof. On the one hand, when pσ(1) · 16λ log n 1 16 log n <√ , (7.15) · 2  2 1+  n by Lemma 7.1 we have Pr Mi ≥ 16 log ≤ n12 ∀ i ∈ [n], where Mi is the occurences of i. Therefore, 2 n -collision in S. by the union bound, with probability at least 1 − n · n12 = 1 − n1 , there is no 16 log 2 log n 1 log n/ √ Since the while loop only has at most log 1+ n = O(  ) rounds and (1 − n ) = 1 − o(1), we may assume that as long as (7.15) holds, Line 4 of Algorithm 7 always has a negative output and √ Line 5 enforces λ ← λ · 1 +  and jumps to the start of the while loop. The while loop keeps iterating until (7.15) is violated. In the second iteration after (7.15) is violated, we have pσ(1) · 16λ log n √ 16 log n 16 log n ≤ < 1+· ; 2 2   2 since pσ(2) ≤ pσ(1) 1+ , (7.16) we have pσ(2) · 16λ log n 1 16 log n <√ . 2  1 +  2 (7.17) As a result, by Lemma 7.1 we have h 16 log n i > 0.15; Pr Mσ(1) ≥ 2 h 16 log n i 1 Pr Mi ≥ ≤ 2 2 n 28 ∀ i ∈ [n]/{σ(1)}. (7.18)   Therefore, Pr Line 4 outputs σ(1) in the second iteration after (7.15) ish violated ≥i 0.15 · 1 −  n n−1 n−1 . In the first iteration after (7.15) is violated, we still have Pr Mi ≥ 16 log ≤ n12 ∀ i ∈ 2 n 2 [n]/{σ(1)}. Therefore,   Pr Line 4 outputs σ(1) in the first or second iteration after (7.15) is violated  n − 1 n−1 0.15 n − 1 n−1  · 1 − ≥ 2 = Ω(1). (7.19) ≥ 0.15 · 1 − n2 n2 e In all, with probability Ω(1), Line 4 of Algorithm 7 outputs σ(1) correctly in the first or second iteration after (7.15) is violated; after that, the quantum counting in Line 5 approximates pσ(1) = maxi∈[n] pi within multiplicative error . This establishes the correctness of Algorithm 7.  n It remains to show that the quantum query complexity of Algorithm 7 is Õ Q(d 16 log e-distinctness) . 2  n Because there are at most log√1+ n = O( log n ) iterations in the while loop, the d 16 log e-distinctness 2 n algorithm in Line 4 is called for at most O( log n ) times; if it gives a d 16 log e-collision, because 2 √ maxi∈[n] pi ≥ 1/n, the quantum query complexity caused by Line 5 is O( n ) by Theorem 2.3, which is smaller than the Ω(n2/3 ) quantum lower bound on the distinctness problems [2]. As a result, the query complexity of Algorithm 7 in total is at most  log n  l 16 log n m   √n   l 16 log n m  O ·Q -distinctness + O = Õ Q -distinctness . (7.20)  2  2 Remark 7.1. In some special cases, Algorithm 7 already demonstrates provable quantum speedup. Recall the state-of-the-art quantum algorithm for k-distinctness is [9] by Belovs, which has query 2 k−2 k complexity O(2k n1−2 /(2 −1) ); however, this is superlinear √ when k = Θ(log n). Nevertheless, if we are promised that H∞ (p) ≤ f (n) for some f (n) = o( log n), then we can replace the n in n Line 2 of Algorithm 7 by ef (n) and replace every d 16 log e by d 16f2(n) e, and it can be shown that the 2  3 quantum query complexity of min-entropy estimation is Õ e( 4 +o(1))·f (n) , whereas the best classical algorithm takes Θ̃(ef (n) ) queries. In this case, we obtain a ( 43 + o(1))-quantum speedup, but the √ √ classical query complexity is already small (e log n = n1/ log n = o(nc ) for any c > 0). 8 0-Rényi entropy estimation Motivations. Estimating the support size of distributions (i.e., the 0-Rényi entropy) is also important in various fields, ranging from vocabulary size estimation [20, 58], database attribute variation [26], password and security [22], diversity study in microbiology [31, 36, 51], etc. The study of support estimation was initiated by naturalist Corbet in 1940s, who spent two years at Malaya for trapping butterflies and recorded how many times he had trapped various butterfly species. He then asked the leading statistician at that time, Fisher, to predict how many new species he would observe if he returned to Malaya for another two years of butterfly trapping. Fisher answered by alternatively putting plus or minus sign for the number of species that showed up one, two, three times, and so on, which was proven to be an unbiased estimator [21]. Formally, assuming n independent samples are drawn from an unknown distribution, the goal of [21] is to estimate the number of hitherto unseen symbols that would be observed if t · n (t being a pre-determined parameter) additional independent samples were collected from the same distribution. Reference [21] solved the case t = 1, which was later improved to t ≤ 1 [24] and 29 Problem Support coverage Support size Θ Θ n log n m log m classical bounds  · log 1 [3, 48] [ = Ω(n−0.2 )]  · log2 1 [48, 63] [ = Ω(1/m)] quantum bounds (this paper)  1/3  √  Õ 1.5n , Ω n1/6  1/3  √  m Õ 1.5 ,Ω m 1/6 Table 2: Summary of the classical and quantum query complexity of support coverage and size estimation. t = O(log n) [48]; the last work also showed that t = Θ(log n) is the largest possible range to give an estimator with provable guarantee. However, such estimation always assumes n samples; a more natural question is, can we estimate the support of a distribution per se? Specifically, given a discrete distribution p over a finite set X where px denotes the probability of x ∈ X, can we estimate its support, defined by Supp(p) := |{x : x ∈ X, px > 0}|, (8.1) with high precision and success probability? Unfortunately, this is impossible in general because elements with negligible but nonzero probability will be very unlikely to appear in the samples, while still contribute to Supp(p). As an evidence, Supp(p) is the exponent of the 0-Rényi entropy of p, but the sample complexity of αRényi entropy goes to infinity when α → 0+ by Theorem 9.1, both classically and quantumly. To circumvent this difficulty, two related properties have been considered as an alternative to estimate 0-Rényi entropy:  P • Support coverage: Sn (p) := x∈X 1 − (1 − px )n , the expected number of elements observed when taking n samples. To estimate Sn (p) within ±n, [24] showed that n/2 samples from p suffices for any constant ; recently, [65] improved the sample complexity to O logn n , and  [3, 48] also considered the dependence in  by showing that Θ logn n · log 1 is a tight bound, as long as  = Ω(n−0.2 ). • Support size: Supp(p), under the assumption that for any x ∈ X, px = 0 or px ≥ 1/m for some given m ∈ N. Reference [53] proposed the problem and gave a lower bound Ω(m1−o(1) ), and  [60] gave an upper bound O logmm · 12 . Recently, [63] and [48] both proved that Θ logmm ·log2 1 is the tight bound for the problem (both optimal in m and ). Quantumly, we give upper and lower bounds on both support coverage and support size estimation, summarized in Table 2. Support coverage estimation. We give the following upper bound on support coverage estimation; its lower bound is given in Proposition 9.2. Algorithm 8: Estimate the support coverage Sn (p). 1 2 3 Regard the following subroutine as A: Draw a sample i ∈ X according to p; √ Use EstAmp with M = 2dlog2 ( n/)e queries to obtain an estimation p̃i of pi ; 1−(1−p̃i )n if p̃i 6= 0; otherwise, output n; p̃i   3/2 1 1 1 executions in Theorem  log  log log  4 Output x̃i = 5 Use A for Θ S̃n (p) of Sn (p); 30 2.1 and output an estimation Theorem 8.1. Algorithm 8 approximates Sn (p) n := P  ≤ O(1) with success probability at least 2/3 using Õ n x∈X (1−(1−px ) ) √ n n 1.5 within an additive error 0 < quantum queries to p. Proof. We prove this theorem in two steps. step is to show that the expectation of P The first 1−(1−p̃i )n the subroutine A’s output (denoted Ẽ := i∈X pi · ) satisfies |Ẽ − E| = O(n), where p̃i P 1−(1−pi )n E := i∈X pi · = Sn (p). pi To achieve this, it suffices to prove that for each i ∈ X, h 1 − (1 − p̃ )n 1 − (1 − p )n i i i − E = O(n). p̃i pi (8.2) We write pi = sin2 (θi π). Assume k ∈ Z such that k ≤ M θi < k + 1. By Theorem 2.3, for any l ∈ {1, 2, . . . , max{k − 1, M − k − 1}}, the output of EstAmp taking M queries satisfies h h (k + l + 1)π i (k − l − 1)π i 1 Pr p̃i = sin2 , Pr p̃i = sin2 ≤ 2. (8.3) M M 4l  for some l ∈ N. For convenience, We first consider the case when p̃i > pi , and p̃i = sin2 (k+l+1)π M 1−(1−x)n where x ∈ (0, 1]. Because denote f (x) = x f 0 (x) = nx(1 − x)n−1 + (1 − x)n − 1 nx + (1 − nx) − 1 ≤ = 0, x2 x2 (8.4) f is a decreasing function on (0, 1]. Therefore,  2 (k+l+1)π n    1 − (1 − p̃i )n 1 − (1 − pi )n 2 kπ 1 − 1 − sin 2 kπ n M ≤ sin · (8.5) − − 1 − 1 − sin p̃i pi M M sin2 (k+l+1)π M (k+l+1)π   kπ 1 − cos2n 2n kπ M = sin2 · − 1 − cos (8.6) M M sin2 (k+l+1)π M   2 (k+l+1)π 2 kπ 2n (k+l+1)π sin2 kπ − sin + sin2 (k+l+1)π cos2n kπ M M M M − sin M cos M . (8.7) = sin2 (k+l+1)π M By Taylor expansion, we have sin2  k6  kπ k2 π2 = + O , M M2 M6 sin2  (k + l)6  (k + l + 1)π (k + l + 1)2 π 2 = + O , M M2 M6 (8.8) and cos2n  kπ  k2 π2 k 4 2n  k2 π2  k 2 2 2n = 1− + O = 1 − + O = 1 − k 2 π 2  + O 2 k 2 ; (8.9) 2 4 2 M 2M M 2n n similarly cos2n  (k + l + 1)π = 1 − (k + l + 1)2 π 2  + O 2 (k + l)2 . M 31 (8.10) Plugging (8.8), (8.9), and (8.10) into (8.7) and noticing that the tail in (8.8) has 1/M 6 , much smaller than that of (8.9) and (8.10), we have 1 − (1 − p̃i )n 1 − (1 − pi )n − p̃i pi 2 π2 (k+l+1)2 π 2  k2 π 2 − + (k+l+1) (1 − k 2 π 2  + O(2 k 2 )) − M2 M2 M2 ≤ (k+l+1)2 π 2 k2 π 2 (1 M2  − (k + l + 1)2 π 2  + O(2 (k + l)2 )) M2 (8.11) k 2 − (k + l + 1)  2 + (k + l + 1)2 (1 − k 2 π 2 ) − k 2 (1 − (k + l + 1)2 π 2 ) (k + l + 1)2 k2 + O(2 k 2 ) − O(2 (k + l)2 ) (k + l + 1)2  = 0 + O 2 (k + l)2 = (8.12) (8.13) ≤ O(n), (8.14) p where (8.14) holds because k + l ≤ M and M = Θ( n/). Similarly, for the case p̃i < pi , we have 1 − (1 − p̃i )n 1 − (1 − pi )n − ≤ O(n). p̃i pi (8.15) In all, summing all l ∈ {1, 2, . . . , max{k − 1, M − k − 1}} in cases p̃i > pi and p̃i < pi and by (8.3), the expectation of the deviation in (8.2) is at most M h 1 − (1 − p̃ )n 1 − (1 − p )n i X 1 π2 i i − ≤2· · O(n) ≤ · O(n) = O(n). E 2 p̃i pi 4l 3 (8.16) l=1 Therefore, (8.2) follows and |Ẽ − E| = O(n). By rescaling M by a constant, without loss of generality we have |Ẽ − E| ≤ n/2. The second step is to bound the variance of the random variable, which is  1 − (1 − p̃ )n 2  X X 1 − (1 − p̃i )n 2 X i pi · − pi · ≤ pi · n2 = n2 , (8.17) p̃i p̃i i∈X i∈X i∈X because 1 − (1 − p̃i )n ≤ np̃i by 0 ≤ p̃i ≤ 1. As a result of Theorem 2.1, we can approximate Ẽ up to additive error n/2 with failure probability at most 1/3 using n n  n   √n  √ 3/2 dlog2 ( n/)e O log log log ·2 = Õ 1.5 (8.18) n n n  quantum queries. Together with |Ẽ − E| ≤ n/2, Algorithm 8 approximates E = Sn (p) up to additive error n with failure probability at most 1/3; in other words, Algorithm 8 approximates E = Sn (p)/n up to  with success probability at least 2/3. Support size estimation. We give the following upper bound on support size estimation; its lower bound is given in Proposition 9.3. Theorem 8.2. Under the promise that for any x ∈ X, px = 0 or px ≥ 1/m, Algorithm 9 approximates Supp(p)/m within an additive error 0 <  ≤ O(1) with success probability at least 2/3 using √  m Õ 1.5 quantum queries to p. 32 Algorithm 9: Estimate Supp(p), under the promise that px = 0 or px ≥ 1/m for any x ∈ X. 1 2  Call Algorithm 8 with n = dm log(2/)e and error 2 log(2/) , and denote the output as S̃n (p); ] ] := dS̃n (p)e. Output Supp(p) Denote Supp(p) as an estimation of Supp(p); Proof. For convenience, denote X1/m := {x ∈ X : px ≥ 1/m}. Then Supp(p) = |X1/m | by the promise, and 1−  ≤ 1 − (1 − px )n ≤ 1 2 1 − (1 − px )n = 0 ∀ x ∈ X1/m ; (8.19) ∀x ∈ / X1/m ; (8.20) As a result, Sn (p) = X 1 − (1 − px )n ∈ h 1− x∈X i  Supp(p), Supp(p) . 2 (8.21) Furthermore, by the correctness of Algorithm 8, with probability at least 2/3 we have |S̃n (p) − Sn (p)| ≤  m ·n= . 2 log(2/) 2 Together with (8.21), h m m i h m i  Supp(p) − , Supp(p) + ⊆ Supp(p) − m, Supp(p) + . S̃n (p) ∈ 1 − 2 2 2 2 dS̃n (p)e m Therefore, with probability at least 2/3, at least 2/3. 9 approximates Supp(p) m (8.22) (8.23) up to  with success probability Quantum lower bounds In this section, we prove Theorem 1.2, which is rewritten below: Theorem 9.1. Any quantum algorithm that approximates Hα (p) of distribution p on [n] within additive error  with success probability at least 2/3 must use 1 1 • Ω(n 3 / 6 ) quantum queries when α = 0, assuming 1/n ≤  ≤ 1. 1 2 • Ω̃(n 7α −o(1) / 7 ) quantum queries when 0 < α < 37 . 1 1 1 1 • Ω(n 3 / 6 ) quantum queries when 3 7 ≤ α ≤ 3, assuming 1/n ≤  ≤ 1. • Ω(n 2 − 2α /) quantum queries when 3 ≤ α < ∞. √ • Ω( n/) quantum queries when α = ∞. Because we use different techniques for different ranges of α, we divide the proofs into three categories. 33 9.1 Reduction from classical lower bounds (0 < α < 37 ) We prove that the quantum lower bound when 0 < α < Theorem 9.1. 1 2 is indeed Ω(n 7α −o(1) / 7 ), as claimed in 3 7 1 Proof. First, by [4], we know that Ω(n α −o(1) /2 ) is a lower bound on the classical query complexity of α-Rényi entropy estimation. On the other hand, reference [1] shows that for any problem that is invariant under permuting inputs and outputs and that has sufficiently many outputs, the quantum query complexity is at least the seventh root of the classical randomized query complexity (up to poly-logarithmic factors). Our query oracle Op : [S] → [n] has n outputs with tend to infinity when n is large; the distribution p is invariant under permutations on [S] since pi = |{s ∈ [S] : Op (s) = i}|/S is invariant for all i; Rényi entropy is invariant under permutations on [n] since it does not depend 1 2 on the order of pi . Therefore, our problem satisfies the requirements from [1], and Ω(n 7α −o(1) / 7 ) is a lower bound on the quantum query complexity of α-Rényi entropy estimation. 9.2 Exploitation of the collision lower bound (α = 0 and 3 7 ≤ α ≤ 3) We prove lower bounds on entropy estimation by further exploiting the famous collision lower bound [2, 38]. First, we define the following problem: Definition 9.1 (l-pairs distinctness). Given positive integers n and l such that 1 ≤ l ≤ n/2, and a function f : [n] → [n]. Under the promise that either f is 1-to-1 or their exists l pairwise different pairs (xi1 , yi1 ), . . . , (xil , yil ) ∈ [n] × [n] such that xij 6= yij but f (xij ) = f (yij ) for all j ∈ [l], the l-pairs distinctness problem is to determine which is the case, with success probability at least 2/3. Note that when l = 1, l-pairs distinctness reduces to the element distinctness problem, whose quantum query complexity is Θ(n2/3 ) [2, 5]; when l = n/2, l-pairs distinctness reduces to the collision problem, whose quantum query complexity is Θ(n1/3 ) [2, 38]. Inspired by the reduction from the collision lower bound to the element distinctness lower bound in [2], we prove a more general quantum lower bound for l-pairs distinctness: Proposition 9.1. The quantum query complexity of l-pairs distinctness is at least Ω(nα ), where 2 lα = n 3 −α . Proof. Assume the contrary that the quantum query complexity of l-pairs distinctness is o(nα ). Consider a function f : [n] → [n] that is promised to be either 1-to-1 or 2-to-1. By [38], it takes Ω(n1/3 ) quantum queries to decide whether f is 1-to-1 or 2-to-1. √ Denote S to be a subset of [n], where |S| = d2 nle and the elements in S are chosen uniformly at random. If f is 1-to-1, then f restricted on S, denoted f |S , is still 1-to-1 on S. If f is 2-to-1, denote the set of its images as {a1 , . . . , an/2 }. For any j ∈ [n/2], denote Xj to be a binary random variable that equals to 1 when the collision pair of aj appears in S, and equals to 0 otherwise. Then |S| 2 n 2  Pr[Xj = 1] = Denote X = Pn/2 j=1 Xj , |S| 4 n 4  , Pr[Xj Xk = 1] = ∀ j, k ∈ [n], j 6= k. (9.1) which is the number of collision pairs in S. By linearity of expectation, E[X] = n |S|(|S| − 1) · & 2l. 2 n(n − 1) 34 (9.2) On the other hand, Var[X] = E[X 2 ] − E[X]2 = n/2 X E[Xj ] + j=1 n ≤ · 2 X (9.3) E[Xj Xk ] − E[X]2 (9.4) j6=k |S| 2 n 2   nn + −1 · 2 2 |S| 4 n 4  n2 − · 4  |S| 2 2  n 2 2 . 2l. (9.5) (9.6) Therefore, by Chebyshev’s inequality, √   Pr[X < l] ≤ Pr X ≤ E[X] − 2 2l ≤ 1/4. (9.7) In other words, with probability at least 3/4, f |S on S has at least l collision pairs. By our assumption, it takes o(|S|α ) = o(nα/2 · n1/3−α/2 ) = o(n1/3 ) quantum queries to decide whether f |S is 1-to-1 or has l collision pairs, which suffices to decide whether f is 1-to-1 or 2-to-1. However, this contradicts with the Ω(n1/3 ) quantum lower bound for the collision problem [38]. 9.2.1 α=0 For 0-Rényi entropy estimation, we use Proposition 9.1 to give quantum lower bounds for both support coverage estimation and support size estimation (both defined in Section 8). 1/3  Proposition 9.2. The quantum query complexity of support coverage estimation is Ω n1/6 , for 1 all n1 ≤  ≤ 12 . 1 Proof. Because n1 ≤  ≤ 12 , we may denote  = nr where r ∈ [−1, 0]. Consider two distributions p1 and p2 encoded by Op1 , Op2 : [n] → X (S = n in (1.7)), where the nonzero probabilities in p1 are 1/n for n times, and the nonzero probabilities in p2 are 2/n for l = d6ne times and 1/n for n − 2l times. In other words, Op1 is injective, and Op2 has l collision pairs but otherwise injective. On the one hand, by Proposition 9.1, it takes Ω(nα ) quantum queries to distinguish between Op1 and Op2 , where  1  1 r 2 2 +O ≥ − ∀ r ∈ [−1, 0]. (9.8) lα = n 3 −α ⇒ α = 3(2 + r) log n 3 6 As a result, nα ≥ n1/3−r/6 = On the other hand, n1/3 . 1/6  n · 1 − (1 − 1/n)n Sn (p1 ) 1 = ≈1− ; n n e   l · 1 − (1 − 2/n)n + (n − 2l) · 1 − (1 − 1/n)n Sn (p2 ) 1 l(1 − 1/e)2 = ≈1− − . n n e n As a result, (9.9) (9.10) Sn (p1 ) Sn (p2 ) l(1 − 1/e)2 − ≈ > 2. (9.11) n n n Therefore, if a quantum algorithm can estimate support coverage with error , it can distinguish between p1 and p2 with success probability at least 2/3. In conclusion, the quantum query complexity 1/3  of support coverage estimation is Ω n1/6 . 35 Similar to the proof of Proposition 9.2, we can prove (with details omitted): Proposition 9.3. The quantum query complexity of support size estimation is Ω 1 1 m ≤  ≤ 4. 9.2.2 3 7 m1/3 1/6  , for all ≤α≤3 Using Proposition 9.1, we show that the quantum query complexity of entropy estimation when 1 1 3 3 6 7 ≤ α ≤ 3 is also Ω(n / ). Proof. We consider the case α = 1, i.e., Shannon entropy estimation; the proof for other α ∈ [ 37 , 3] is basically identical. Consider two distributions p1 and p2 encoded by Op1 , Op2 : [n] → X (S = n in (1.7)), where the nonzero probabilities in p1 are 1/n for n times, and the nonzero probabilities in p2 are 2/n for l = dn/ log 2e times and 1/n for n − 2l times. In other words, Op1 is injective, and Op2 has l collision pairs but otherwise injective. On the one hand, similar to the proof of Proposition 9.2, it takes it takes Ω(n1/3−/6 ) quantum queries to distinguish between Op1 and Op2 . On the other hand, 1 log n = log n; n 2 n 1 2l H(p2 ) = l · log + (n − 2l) · log n = log n − log 2. n 2 n n H(p1 ) = n · (9.12) (9.13) As a result, |H(p1 ) − H(p2 )| = 2l log 2 ≥ 2. n (9.14) Therefore, if a quantum algorithm can estimate support coverage with error , it can distinguish between p1 and p2 with success probability at least 2/3. In conclusion, the quantum query complexity 1/3  of support coverage estimation is Ω n1/6 . 9.3 Polynomial method (3 ≤ α ≤ ∞) We use the polynomial method [8] to show quantum lower bounds for entropy estimation when 3 ≤ α ≤ ∞. Inspired by the symmetrization technique in [38], we obtain a bivariate polynomial whose degree is at most two times the corresponding quantum query complexity. Next, similar to [44], we apply Paturi’s lemma [52] to give a lower bound on the degree of the polynomial. To be more specific, we prove: √ Proposition 9.4. The quantum query complexity of estimating min-entropy with error  is Ω( n  ). Proposition 9.5. When the constant α satisfies 1 < α < ∞, the quantum query complexity of 1− 1 2α estimating α-Rényi entropy with error  is Ω( αn 2 ). Without loss of generality, we assume that the oracle Op in (1.7) satisfies n|S, otherwise consider the oracle Op0 : [Sn] → [n] such that Op0 (s + Sl) = Op (s) for all s ∈ [S] and l ∈ [n]; this gives an oracle for the same distribution. We consider the special case where the probabilities {pi }ni=1 takes at most two different values; to integrate the probabilities, we assume the existence of two integers c, d where c ∈ {1, . . . , n − 1}, such that pi = n1 − Sd for n − c different i’s in {1, . . . , n}, and pi = n1 + (n−c)d for the other c i’s in cS {1, . . . , n}. 36 Proof of Proposition 9.4. Following the symmetrization technique in [38], we obtain a bivariate polynomial Q(c, d) where such that the degree of Q is at most two times the query complexity of min-entropy estimation, and:  Sc  • c ∈ {1, . . . , n − 1} and d ∈ {− n(n−c) , . . . , Sn }. This is because pi ≥ 0 for all i ∈ [n]. • 0 ≤ Q(c, d) ≤ 1 if c|nd. Only if c|nd, S · is valid under our model in (1.7). 1 n  is an integer and the distribution {pi }ni=1 + (n−c)d cS Furthermore, we consider the property testing problem of determining whether maxi pi = n1 or maxi pi ≥ 1+ n , where the accept probability should be at most 1/3 for the former case and at least 2/3 for the latter case. As a result, • 0 ≤ Q(c, 0) ≤ 1/3: In this case, pi = • 2/3 ≤ Q(c, d) ≤ 1 if c|nd, (n−c)d Sc 1 n for all i ∈ [n]. ≥ n : In this case, ∃ i such that pi = 1 n + (n−c)d cS 1 n − d S ≥ • 2/3 ≤ Q(c, d) ≤ 1 if c|nd, d ≤ − S n : In this case, ∃ i such that pi = ≥ 1+ n . 1+ n . Therefore, we have  S  • 0 ≤ Q(1, d) ≤ 1 for d ∈ {− n(n−1) , . . . , Sn }; • 0 ≤ Q(1, 0) ≤ 1/3;  S   S    S • 2/3 ≤ Q(1, d) ≤ 1 for d ∈ {− n(n−1) , . . . , − S n } ∪ { n(n−1) , . . . , n }. Using Paturi’s lower bound [52], we have q S n(n−1)  · S n  degd Q(1, d) ≥ Ω  S/n(n − 1) !  √n  =Ω .  (9.15) √ Therefore, deg Q(c, d) ≥ degd Q(1, d) = Ω( n/). Proof of Proposition 9.5. The proof is similar to that of Proposition 9.4. Following the symmetrization technique, we still obtain a bivariate polynomial Q(c, d) where such that the degree of Q is at most two the query complexity of min-entropy estimation, and c ∈ {1, . . . , n − 1}, d ∈  times Sc {− n(n−c) , . . . , Sn }, 0 ≤ Q(c, d) ≤ 1 if c|nd. Furthermore, we consider the property testing probP P 2 lem of determining whether i∈[n] pαi ≤ nα−1 or i∈[n] pαi ≥ n2+2 α−1 , where the accept probability should be at most 1/3 for the former case and at least 2/3 for the latter case. We also assume  1/α  1 c = 1. On the one hand, when 0 ≤ d ≤ n n−1−1 · Sn , we have n1 + (n−1)d ≤ n1−1/α , and S X i∈[n] pαi ≤  1 n1−1/α α  + (n − 1) 1  2 1 α 1 − 1−1/α ≤ α−1 . n−1 n n 37 (9.16) On the other hand, because (1 + m)α ≈ 1 + mα when m = o(1), we have  1   1 3 α 3 α 2 + 2 1 + + (n − 1) − − α−1 1 − n−1 n n1−1/α αn1−1/α n1−1/α αn1−1/α  3 α 2 + 2 1 3 α 1 1  − α−1 + 1 − 1−1/α − = α−1 1 + n α (n − 1)α−1 n n αn1−1/α  1  α 3α ≈ α−1 1 + 3 + 1 − 1−1/α − − (2 + 2) n n αn1−1/α 1  α + 3  = α−1  − 1−1/α n n ≥0 for large enough n. As a result, when d ≥ Therefore, we have  (1+3/α)n1/α −1 n−1 · S n  , we have · S n P α i∈[n] pi ≥ (9.17) (9.18) (9.19) (9.20) 2+2 . nα−1 • 0 ≤ Q(1, d) ≤ 1 for d ∈ {0, . . . , Sn };   1/α • 0 ≤ Q(1, d) ≤ 1/3 for d ∈ {0, . . . , n n−1−1 · Sn };   1/α −1 • 2/3 ≤ Q(1, d) ≤ 1 for d ∈ { (1+3/α)n · Sn , . . . , Sn }. n−1 Using Paturi’s lower bound [52], we have q degd Q(1, d) ≥ Ω n1/α −1 n−1 · S n  S n −  (3/α)n1/α n−1 1  n1/α −1 n−1 · S n   ! 1   αn 12 − 2α . =Ω  (9.21) 1 Therefore, deg Q(c, d) ≥ degd Q(1, d) = Ω(αn 2 − 2α /). Technically, our proofs only focus on the degree in d for c = 1, but in general it is possible to prove a better lower bound when analyzing the degree of the polynomial in c and d together. We leave this as an open problem. Acknowledgements We thank Andrew M. Childs for discussions that inspired the proof of Theorem 6.1, and general suggestions on our manuscript; we thank Yanjun Han for introducing us classical references related to Shannon and Rényi entropy estimation, in particular his papers [27, 33, 34]. We also thank anonymous reviewers for helpful comments on an earlier version of this paper. TL acknowledges support from NSF CCF-1526380. References [1] Scott Aaronson and Andris Ambainis, The need for structure in quantum speedups, Theory of Computing 10 (2014), no. 6, 133–166, arXiv:0911.0996. [2] Scott Aaronson and Yaoyun Shi, Quantum lower bounds for the collision and the element distinctness problems, Journal of the ACM (JACM) 51 (2004), no. 4, 595–605. [3] Jayadev Acharya, Hirakendu Das, Alon Orlitsky, and Ananda Theertha Suresh, A unified maximum likelihood approach for optimal distribution property estimation, (2017). 38 [4] Jayadev Acharya, Alon Orlitsky, Ananda Theertha Suresh, and Himanshu Tyagi, Estimating Renyi entropy of discrete distributions, IEEE Transactions on Information Theory 63 (2017), no. 1, 38–56, arXiv:1408.1000. [5] Andris Ambainis, Quantum walk algorithm for element distinctness, SIAM Journal on Computing 37 (2007), no. 1, 210–239, arXiv:quant-ph/0311001. [6] Erdal Arikan, An inequality on guessing and its application to sequential decoding, IEEE Transactions on Information Theory 42 (1996), no. 1, 99–105. [7] Tugkan Batu, Sanjoy Dasgupta, Ravi Kumar, and Ronitt Rubinfeld, The complexity of approximating the entropy, SIAM Journal on Computing 35 (2005), no. 1, 132–150. [8] Robert Beals, Harry Buhrman, Richard Cleve, Michele Mosca, and Ronald de Wolf, Quantum lower bounds by polynomials, Journal of the ACM (JACM) 48 (2001), no. 4, 778–797, arXiv:quant-ph/9802049. [9] Aleksandrs Belovs, Learning-graph-based quantum algorithm for k-distinctness, 53rd Annual Symposium on Foundations of Computer Science, pp. 207–216, IEEE, 2012, arXiv:1205.1534. [10] Aleksandrs Belovs and Ansis Rosmanis, Adversary lower bounds for the collision and the set equality problems, arXiv:1310.5185 (2013). [11] Charles H. Bennett, Gilles Brassard, Claude Crépeau, and Ueli M. Maurer, Generalized privacy amplification, IEEE Transactions on Information Theory 41 (1995), no. 6, 1915–1923. [12] Gilles Brassard, Peter Høyer, Michele Mosca, and Alain Tapp, Quantum amplitude amplification and estimation, Contemporary Mathematics 305 (2002), 53–74, arXiv:quant-ph/0005055. [13] Sergey Bravyi, Aram W. Harrow, and Avinatan Hassidim, Quantum algorithms for testing properties of distributions, IEEE Transactions on Information Theory 57 (2011), no. 6, 3971– 3981, arXiv:0907.3920. [14] Yuheng Bu, Shaofeng Zou, Yingbin Liang, and Venugopal V. Veeravalli, Estimation of KL divergence: Optimal Minimax Rate, arXiv:1607.02653 (2016). [15] Olivier Catoni, Statistical learning theory and stochastic optimization: Ecole d’eté de probabilités de saint-flour xxxi-2001, Springer, 2004. [16] Sourav Chakraborty, Eldar Fischer, Arie Matsliah, and Ronald de Wolf, New results on quantum property testing, Thirtieth International Conference on Foundations of Software Technology and Theoretical Computer Science, p. 145, 2010, arXiv:1005.0523. [17] Imre Csiszár, Generalized cutoff rates and Rényi’s information measures, IEEE Transactions on Information Theory 41 (1995), no. 1, 26–34. [18] Imre Csiszar and János Körner, Information theory: coding theorems for discrete memoryless systems, Cambridge University Press, 2011. [19] Paul Dagum, Richard Karp, Michael Luby, and Sheldon Ross, An optimal algorithm for Monte Carlo estimation, SIAM Journal on Computing 29 (2000), no. 5, 1484–1496. [20] Bradley Efron and Ronald Thisted, Estimating the number of unseen species: How many words did Shakespeare know?, Biometrika 63 (1976), no. 3, 435–447. 39 [21] Ronald A. Fisher, Alexander S. Corbet, and Carrington B. Williams, The relation between the number of species and the number of individuals in a random sample of an animal population, The Journal of Animal Ecology (1943), 42–58. [22] Dinei Florencio and Cormac Herley, A large-scale study of web password habits, Proceedings of the 16th International Conference on World Wide Web, pp. 657–666, ACM, 2007. [23] Peter W. Glynn, Upper bounds on Poisson tail probabilities, Operations Research Letters 6 (1987), no. 1, 9–14. [24] I. J. Good and G. H. Toulmin, The number of new species, and the increase in population coverage, when a sample is increased, Biometrika 43 (1956), no. 1-2, 45–63. [25] Jeongwan Haah, Aram W. Harrow, Zhengfeng Ji, Xiaodi Wu, and Nengkun Yu, Sampleoptimal tomography of quantum states, Proceedings of the 48th Annual ACM Symposium on Theory of Computing, pp. 913–925, ACM, 2016, arXiv:1508.01797. [26] Peter J. Haas, Jeffrey F. Naughton, S. Seshadri, and Lynne Stokes, Sampling-based estimation of the number of distinct values of an attribute, Proceedings of 21th International Conference on Very Large Data Bases, vol. 95, pp. 311–322, 1995. [27] Yanjun Han, Jiantao Jiao, and Tsachy Weissman, Minimax rate-optimal estimation of divergences between discrete distributions, arXiv:1605.09124 (2016). [28] Godfrey H. Hardy, P. V. Seshu Aiyar, and Bertram M. Wilson, Collected papers of Srinivasa Ramanujan, AMS, 1927. [29] Ralph V. L. Hartley, Transmission of information, Bell Labs Technical Journal 7 (1928), no. 3, 535–563. [30] Peter Høyer, Troy Lee, and Robert Spalek, Negative weights make adversaries stronger, Proceedings of the 39th Annual ACM Symposium on Theory of Computing, pp. 526–535, ACM, 2007, arXiv:quant-ph/0611054. [31] Jennifer B. Hughes, Jessica J. Hellmann, Taylor H. Ricketts, and Brendan J. M. Bohannan, Counting the uncountable: statistical approaches to estimating microbial diversity, Applied and Environmental Microbiology 67 (2001), no. 10, 4399–4406. [32] Russell Impagliazzo and David Zuckerman, How to recycle random bits, 30th Annual Symposium on Foundations of Computer Science, pp. 248–253, IEEE, 1989. [33] Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman, Maximum likelihood estimation of functionals of discrete distributions, arXiv:1406.6959 (2014). [34] , Minimax estimation of functionals of discrete distributions, IEEE Transactions on Information Theory 61 (2015), no. 5, 2835–2885, arXiv:1406.6956. [35] Diederik P. Kingma and Max Welling, Auto-encoding variational bayes, arXiv:1312.6114 (2013). [36] Ian Kroes, Paul W. Lepp, and David A. Relman, Bacterial diversity within the human subgingival crevice, Proceedings of the National Academy of Sciences 96 (1999), no. 25, 14547–14552. [37] Solomon Kullback, Information theory and statistics, Courier Corporation, 1997. 40 [38] Samuel Kutin, Quantum lower bound for the collision problem with small range, Theory of Computing 1 (2005), no. 1, 29–36. [39] Lucien Le Cam, Asymptotic methods in statistical decision theory, Springer Science & Business Media, 2012. [40] Michael Mitzenmacher and Eli Upfal, Probability and computing: Randomization and probabilistic techniques in algorithms and data analysis, Cambridge University Press, 2005. [41] Ashley Montanaro, Quantum speedup of Monte Carlo methods, Proc. R. Soc. A, vol. 471, p. 20150301, The Royal Society, 2015, arXiv:1504.06987. [42] , The quantum complexity of approximating the frequency moments, Quantum Information & Computation 16 (2016), no. 13&14, 1169–1190, arXiv:1505.00113. [43] Ashley Montanaro and Ronald de Wolf, A survey of quantum property testing, arXiv:1310.2035 (2013). [44] Ashwin Nayak and Felix Wu, The quantum query complexity of approximating the median and related statistics, Proceedings of the 31st Annual ACM Symposium on Theory of Computing, pp. 384–393, ACM, 1999, arXiv:quant-ph/9804066. [45] Ryan O’Donnell and John Wright, Quantum spectrum testing, Proceedings of the 47th Annual ACM on Symposium on Theory of Computing, pp. 529–538, ACM, 2015, arXiv:1501.05028. [46] , Efficient quantum tomography, Proceedings of the 48th Annual ACM Symposium on Theory of Computing, pp. 899–912, ACM, 2016, arXiv:1508.01907. [47] , Efficient quantum tomography II, Proceedings of the Forty-ninth Annual ACM SIGACT Symposium on Theory of Computing, pp. 962–974, ACM, 2017, arXiv:1612.00034. [48] Alon Orlitsky, Ananda Theertha Suresh, and Yihong Wu, Optimal prediction of the number of unseen species, Proceedings of the National Academy of Sciences 113 (2016), no. 47, 13283– 13288. [49] Liam Paninski, Estimation of entropy and mutual information, Neural Computation 15 (2003), no. 6, 1191–1253. [50] , Estimating entropy on m bins given fewer than m samples, IEEE Transactions on Information Theory 50 (2004), no. 9, 2200–2203. [51] Bruce J. Paster, Susan K. Boches, Jamie L. Galvin, Rebecca E. Ericson, Carol N. Lau, Valerie A. Levanos, Ashish Sahasrabudhe, and Floyd E. Dewhirst, Bacterial diversity in human subgingival plaque, Journal of Bacteriology 183 (2001), no. 12, 3770–3783. [52] Ramamohan Paturi, On the degree of polynomials that approximate symmetric Boolean functions (preliminary version), Proceedings of the 24th Annual ACM Symposium on Theory of Computing, pp. 468–474, ACM, 1992. [53] Sofya Raskhodnikova, Dana Ron, Amir Shpilka, and Adam Smith, Strong lower bounds for approximating distribution support size and the distinct elements problem, SIAM Journal on Computing 39 (2009), no. 3, 813–842. 41 [54] Alfréd Rényi, On measures of entropy and information, Proceedings of the 4th Berkeley Symposium on Mathematical Statistics and Probability, vol. 1, pp. 547–561, 1961. [55] Dana Ron, Algorithmic and analysis techniques in property testing, Foundations and Trends in Theoretical Computer Science 5 (2010), no. 2, 73–205. [56] Claude E. Shannon, A mathematical theory of communication, Bell System Technical Journal 27 (1948), no. 3, 379–423. [57] Daniel Štefankovič, Santosh Vempala, and Eric Vigoda, Adaptive simulated annealing: A nearoptimal connection between sampling and counting, Journal of the ACM (JACM) 56 (2009), no. 3, 18, arXiv:cs/0612058. [58] Ronald Thisted and Bradley Efron, Did Shakespeare write a newly-discovered poem?, Biometrika 74 (1987), no. 3, 445–455. [59] Salil P. Vadhan, Pseudorandomness, Foundations and Trends R in Theoretical Computer Science 7 (2012), no. 1–3, 1–336. [60] Gregory Valiant and Paul Valiant, Estimating the unseen: an n/log(n)-sample estimator for entropy and support size, shown optimal via new CLTs, Proceedings of the 43rd Annual ACM Symposium on Theory of Computing, pp. 685–694, ACM, 2011. [61] Paul Valiant, Testing symmetric properties of distributions, SIAM Journal on Computing 40 (2011), no. 6, 1927–1968. [62] Paul C. van Oorschot and Michael J. Wiener, Parallel collision search with cryptanalytic applications, Journal of Cryptology 12 (1999), no. 1, 1–28. [63] Yihong Wu and Pengkun Yang, Chebyshev polynomials, moment matching, and optimal estimation of the unseen, arXiv:1504.01227 (2015). [64] , Minimax rates of entropy estimation on large alphabets via best polynomial approximation, IEEE Transactions on Information Theory 62 (2016), no. 6, 3702–3720, arXiv:1407.0381. [65] James Zou, Gregory Valiant, Paul Valiant, Konrad Karczewski, Siu-On Chan, Kaitlin Samocha, Monkol Lek, Shamil Sunyaev, Mark Daly, and Daniel G. MacArthur, Quantifying unobserved protein-coding variants in human populations provides a roadmap for large-scale sequencing projects, Nature communications 7 (2016). A Theorem 2.2: Multiplicative quantum Chebyshev inequality The main technique that we use is Lemma 4 in [41], which approximates a random variable with an additive error as long as its second-moment is bounded: Lemma A.1 (Lemma 4 in [41]). Assume A is a quantum algorithm that outputs a random variable X. Then for  where 0 <  < 1/2 (multiplicative error), by using O((1/) log3/2 (1/) log log(1/)) executions of A and A−1 , Algorithm 2 in [41] outputs an estimate Ẽ[X] of E[X] such that8 p   Pr Ẽ[X] − E[X] ≥ ( E[X 2 ] + 1)2 ≤ 1/50. (A.1) 8 The original error probability in (A.1) is 1/5, but it can be improved to 1/50 by rescaling the parameters in Lemma 4 in [41] up to a constant. 42 Based on Lemma A.1 and inspired by Algorithm 3 and Theorem 5 in [41], we propose Algorithm 10. Algorithm 10: Estimate E[X] within multiplicative error . 1 2 3 4 5 6 Run the algorithm that gives a, b such that E[X] ∈ [a, b]; Set A0 = A/σb; Run A0 once and denote m e to be the output. Set B = A0 − m; e Let B− be the algorithm that calls B once; if B outputs x ≥ 0 then B− outputs 0, and if B outputs x < 0 then B− outputs x. Similarly, let B+ be the algorithm such that if B outputs x < 0 then B+ outputs 0, and if B outputs x ≥ 0 then B+ outputs x; a Apply Lemma A.1 to −B− /6 and B+ /6 with error 48σb and failure probability 1/50, and obtain estimates µ e− and µ e+ , respectively; Output Ẽ[X] = σb(m e − 6e µ− + 6e µ+ ); Proof of Theorem 2.2. Because Var[X] ≤ σ 2 E[X]2 ≤ σ 2 b2 , by Chebyshev’s inequality we have   Pr m e − E[X/σb] ≥ 4 ≤ 1/16. (A.2) X − m, e Therefore, with probability at least 15/16 we have |m e − E[X/σb]| ≤ 4. Denote XB = σb which is the random variable output by B; XB,+ := max{XB , 0} is then the output of B+ and XB,− := min{XB , 0} is the output of B− . Assuming |m e − E[X/σb]| ≤ 4, we have h X  X  X   2 i 2 ]=E E[XB −E + E −m e (A.3) σb σb σb h  X  h X 2 i  X  2 i + 2E E ≤ 2E −E −m e (A.4) σb σb σb ≤ 2(12 + 42 ) = 34. (A.5)       Therefore, E (XB /6)2 ≤ 34/36 < 1, hence E (XB,+ /6)2 < 1 and E (−XB,− /6)2 < 1. By Lemma A.1, we have µ e− − E[−XB,− /6] ≤ a 12σb and µ e+ − E[XB,+ /6] ≤ a 12σb both with failure probability at most 1/50. Because   E[X] = σb m e + E[XB ] = σb m e + E[XB,+ ] − E[−XB,− ] , (A.6) (A.7) with probability at least 15/16 · (1 − 1/50)2 > 9/10, we have Ẽ[X] − E[X] ≤ σb · 6 µ e− − E[−XB,− /6] + 6 µ e+ − E[XB,+ /6] a ≤ σb · 2 · 6 · = a ≤ E(X). 12σb 43  (A.8) (A.9)
7cs.IT
arXiv:1605.09332v4 [cs.LG] 10 Jan 2018 Parametric Exponential Linear Unit for Deep Convolutional Neural Networks Ludovic Trottier Philippe Giguère Brahim Chaib-draa Department of Computer Science and Software Engineering Laval University Québec, Canada Email: [email protected] Department of Computer Science and Software Engineering Laval University Québec, Canada Email: [email protected] Department of Computer Science and Software Engineering Laval University Québec, Canada Email: [email protected] Abstract—Object recognition is an important task for improving the ability of visual systems to perform complex scene understanding. Recently, the Exponential Linear Unit (ELU) has been proposed as a key component for managing bias shift in Convolutional Neural Networks (CNNs), but defines a parameter that must be set by hand. In this paper, we propose learning a parameterization of ELU in order to learn the proper activation shape at each layer in the CNNs. Our results on the MNIST, CIFAR-10/100 and ImageNet datasets using the NiN, Overfeat, All-CNN and ResNet networks indicate that our proposed Parametric ELU (PELU) has better performances than the non-parametric ELU. We have observed as much as a 7.28% relative error improvement on ImageNet with the NiN network, with only 0.0003% parameter increase. Our visual examination of the non-linear behaviors adopted by Vgg using PELU shows that the network took advantage of the added flexibility by learning different activations at different layers. I. I NTRODUCTION Recognizing objects using light from the visible spectrum is a essential ability for performing complex scene understanding with a visual system. Vision-based applications, such as face verification, robotic grasping or autonomous driving, require the fundamental skill of object recognition for carrying out their tasks. They must first identify the different elements in their surrounding environment in order to create a highlevel representation of the scene. Since scene understanding is performed by analyzing the spatial relations and the taxonomy of the representation, the overall performance of the visual system depends on the capability of recognizing objects. Integrating novel object recognition advances for building fully-automated vision systems is one of the first steps towards general visual perception. Over the past few years, Convolutional Neural Networks (CNNs) have become the leading approach in computer vision [1], [2], [3], [4], [5], [6]. Through a series of non-linear transformations, CNNs can process high-dimensional input observations into simple low-dimensional concepts. The key principle in CNNs is that features at each layer are composed of features from the layer below, which creates a hierarchical organization of increasingly abstract concepts. Since levels of organization are often seen in complex biological structures, CNNs are particularly well-adapted for capturing high-level abstractions in real-world observations. The activation function plays a crucial role for learning representative features. The recently proposed Exponential Linear Unit (ELU) has the interesting property of reducing bias shift [7]. Defined as the change of a neuron’s mean value due to weight update, bias shift can lead to oscillations and impede learning when not taken into account [7]. Clevert et al. [7] have shown that either centering the neuron values with a Batch Normalization layer [8] or using activation functions with negative values helps to manage this problem. Defined as identity for positive arguments and a(exp(h) − 1) for negative ones (where a = 1 in [7]), ELU’s negative values for negative inputs make the activation function a well-suited candidate for reducing bias shift. Choosing a proper ELU parameterization can however be relatively cumbersome considering that certain parameterizations are more suitable in some networks than others. The objective of this paper is to alleviate this limitation by learning a parameterization of the activation function, which we refer to as the Parametric ELU (PELU). We contribute in the following ways: 1) We define parameters controlling different aspects of the function and show how to learn them during backpropagation. Our parameterization preserves differentiability by acting on both the positive and negative parts of the function. It has the same computational complexity as ELU and adds only 2L additional parameters, where L is the number of layers. 2) We perform an experimental evaluation on the MNIST, CIFAR-10/100 and ImageNet tasks using the ResNet [9], Network in Network [10], All-CNN [11], Vgg [12] and Overfeat [13] networks. Our results indicates that PELU has better performances than ELU. 3) We evaluate the effect of using Batch Normalization (BN) before our PELU activation, and show that BN increases the error rate of ResNet. 4) We experiment with different PELU parameterizations, and show that the proposed one obtains the best performance among the possible parameterizations. 5) We finally show different PELU non-linear behaviors adopted during training by the VGG network. These results highlight the effects of our parameterization in order to better understand the advantage of the activation. The rest of the paper is organized as follows. We present related works in Section II and described our proposed approach in Section III. We detail our experimentations in Section IV and discuss the results in Section V. We conclude the paper in Section VI. II. R ELATED W ORK Our proposed PELU activation function is related to other parametric approaches in the literature. Parametric ReLU (PReLU) [14] learns a parameterization of the Leaky ReLU (LReLU) [15] activation, defined as max{h, 0} + a min{h, 0} where a > 0. PReLU learns a leak parameter a in order to find a proper positive slope for negative inputs. This prevents negative neurons from dying, i.e. neurons that are always equal to zero, which is caused by a null derivative that blocks the back-propagated error signal. Based on the empirical evidence that learning the leak parameter a rather than setting it to a predefined value (as done in LReLU) improves performance [14], our goal is further improving the performance of ELU by learning a proper parameterization of the function. The Adaptive Piecewise Linear (APL) unit aims learning a weighted sum of S parametrized Hinge functions [16]. One drawback of APL is that the number of points at which the function is non-differentiable increase linearly with S. Differentiable activation functions usually give better parameter updates during back-propagation than activation functions with non-differentiable points [2]. Moreover, although APL has the flexibility to be either a convex or non-convex function, the rightmost linear function is forced to have unit slope and zero bias. This may be an inappropriate constraint which could affect the CNN ability to learn representative features. Another activation function is Maxout, which outputs the maximum over K affine functions for each input neuron [17]. The main drawback of Maxout is that it multiplies by K the amount of weights to be learned in each layer. In the context of CNNs where the max operator is applied over the feature maps of each K convolutional layers, the increased computational burden can be too demanding for deep network. Unlike Maxout, our PELU adds only 2L parameters, where L is the number of layers, which makes our activation as computationally demanding as the original ELU function. The S-Shaped ReLU (SReLU) imitates the Webner-Fechner law and the Stevens law by learning a combination of three linear functions [18]. Although this parametric function can be either convex or non-convex, SReLU has two points at which it is non-differentiable. Unlike SReLU, our PELU is fully differentiable, since our parameterization acts on both the positive and negative sides of the function. This in turn improves the back-propagation weight and bias updates. III. PARAMETRIC E XPONENTIAL L INEAR U NIT The standard Exponential Linear Unit (ELU) is defined as identity for positive arguments and a(exp(h) − 1) for negative a=1 b=1 c=1 a = 0.1 b=1 c=1 a=2 b=1 c=1 a=1 b = 0.1 c=1 a=1 b=2 c=1 a=1 b=1 c = 0.1 a=1 b=1 c=2 Fig. 1. Effects of parameters a, b and c on the Exponential Linear Unit (ELU) activation function. The original ELU is shown at the top, where a = b = c = 1. We show the effect of a on the second row, the effect of b on the third row, and the effect of c on the fourth row. The saturation point decreases when a increases, the function saturates faster when b decreases, and the slope of the linear part increases when c increases. arguments (h < 0) [7]. Although the parameter a can be any positive value, Clevert et al. [7] proposed using a = 1 to have a fully differentiable function. For other values a 6= 1, the function is non-differentiable at h = 0. Directly learning parameter a would break differentiability at h = 0, which could impede back-propagation [2]. For this reason, we first start by adding two additional parameters to ELU: ( ch if h ≥ 0 f (h) = , a, b, c > 0 , (1) h a(exp( b ) − 1) if h < 0 We have ch for positive arguments (h ≥ 0) and a(exp( hb )− 1) for negative arguments (h < 0). The original ELU can be recovered when a = b = c = 1. As shown in Figure 1, each parameter controls different aspects of the activation. Parameter c changes the slope of the linear function in the 2.5 MNIST 1e−4 BN-ReLU ELU PELU 2.0 MSE positive quadrant (the larger c, the steeper the slope), parameter b affects the scale of the exponential decay (the larger b, the smaller the decay), while a acts on the saturation point in the negative quadrant (the larger a, the lower the saturation point). Constraining the parameters in the positive quadrant forces the activation to be a monotonic function, such that reducing the weight magnitude during training always lowers the neuron contribution. Using this parameterization, the network can control its nonlinear behavior throughout the course of the training phase. It may increase the slope with c, the decay with b or lower the saturation point with a. However, a standard gradient update on parameters a, b, c would make the function nondifferentiable at h = 0 and impair back-propagation. Instead of relying on a projection operator to restore differentiability after each update, we constrain our parameterization to always have differentiability at h = 0. By equaling the derivatives on both sides of zero, solving for c gives c = ab as solution. The proposed Parametric ELU (PELU) is then as follows: ( a h if h ≥ 0 f (h) = b , a, b > 0 (2) a(exp( hb ) − 1) if h < 0 1.5 1.0 0 50 100 150 Epoch 200 250 300 Fig. 2. Auto-encoder results on the MNIST task. We compare PELU to ELU, and include BN-ReLU as additional reference. Compared to ELU, PELU obtained a lower test mean squared error. With this parameterization, in addition to changing the saturation point and exponential decay respectively, both a and b adjust the slope of the linear function in the positive part to ensure differentiability at h = 0. PELU is trained simultaneously with all the network parameters during back-propagation. Using the chain rule of derivation, the gradients of f with respect to a, b is given by: ( h if h ≥ 0 ∂f (h) = b (3) ∂a exp(h/b) − 1 if h < 0 ( if h ≥ 0 − ah ∂f (h) b2 = . (4) a ∂b − b2 exp(h/b) if h < 0 For preserving parameter positivity after the updates, we constrain them to always be greater than 0.1. IV. E XPERIMENTATIONS In this section, we present our experiments in supervised learning on the CIFAR-10/100 and ImageNet tasks. A. MNIST Auto-Encoder As first experiment, we performed unsupervised learning, which is the task of learning feature representations from unlabeled observations. Unsupervised learning can be useful in cases like deep learning data fusion [19]. For evaluating our proposed PELU activation, we trained a deep auto-encoder on unlabeled MNIST images [20]. We refer to this network as DAA-net. The encoder has four fully connected layers of sizes 1000, 500, 250, 30, and the decoder is symmetrical to the encoder (the weights are not tied). We used Dropout with probability 0.2 after each activation [21]. For ReLU, we put a Batch Normalization (BN) layer before the activation. We trained DAA-Net with RMSProp [22] at a learning rate of 0.001, smoothing constant of 0.9 and a batch size of 128. Fig. 3. Residual network building block structure. On the left, the main basic block structure, and on the right, the transition block structure for reducing the input spatial dimensions and increasing the number of filters. For our CIFAR experiments, we opted for sub-sampling followed by zero concatenation as transition block, while for our ImageNet experiments, we opted for strided convolution followed by batch normalization as transition block. Figure 2 presents the progression of test mean squared error averaged over five tries of DAA-Net on MNIST dataset. These results show that PELU outperformed ELU and ReLU for both convergence speed and reconstruction error. PELU converged approximatively at epoch 75 with a MSE of 1.04e−4, while ELU converged at epoch 100 with a MSE of 1.12e−4 and ReLU at epoch 100 with a MSE of 1.49e−4. B. CIFAR-10/100 Object Recognition We performed object classification on the CIFAR-10 and CIFAR-100 datasets (60,000 32x32 colored images, 10 and 100 classes respectively) [1]. We trained a 110-layer residual network (ResNet) following Facebook’s Torch implementation CIFAR-10 20 BN-ReLU ELU PELU BN-PReLU 18 Error (%) 16 14 12 10 8 6 0 50 100 150 Epoch CIFAR-100 45 BN-ReLU ELU PELU BN-PReLU 40 Error (%) 200 35 30 25 0 50 100 150 Epoch 200 Fig. 4. ResNet 110 layers test error (in %) medians over five tries on both CIFAR-10 and CIFAR-100 datasets. We compare PELU to ELU, and also include BN-ReLU and BN-PReLU as additional references. PELU has a better convergence and lower recognition errors than ELU. TABLE I R ES N ET 110 LAYERS TEST ERROR ( IN %) ON BOTH CIFAR-10 AND CIFAR-100 DATASETS . W E REPORT THE MEAN ERROR OVER THE LAST FIVE EPOCHS AND THE MINIMUM ERROR OVER ALL EPOCHS ( INSIDE PARENTHESIS ) OF THE MEDIAN ERROR OVER FIVE TRIES . O UR ELU PARAMETERIZATION IMPROVES PERFORMANCE , BUT USING BN BEFORE THE ACTIVATION WORSEN THE PERFORMANCE . ACT BN-ReLU ELU PELU BN-PReLU BN-PELU BN-ELU CIFAR-10 5.67 6.55 5.51 5.61 (5.41) (5.99) (5.36) (5.36) 6.24 (5.85) 11.20 (10.39) CIFAR-100 25.92 26.59 25.02 25.83 (24.99) (25.08) (24.55) (25.50) 26.04 (25.38) 35.51 (34.75) fb.resnet.torch1. In order not to favor PELU to the detriment of ELU and BN+ReLU, we performed minimal changes by only replacing the activation function. The building block structure for the network is shown in Figure 3. We show the basic block structure on the left of Figure 3 and the transition block structure on the right of Figure 3. The ACT module can be PELU, ELU, ReLU or PReLU, with or without BN. The network contains mainly 1 https://github.com/facebook/fb.resnet.torch basic blocks, and a few transition blocks for reducing the spatial dimensions of the input image and increasing the number of filters. The ResNet for our CIFAR experiments has a transition block structure with spatial sub-sampling and zero concatenation, while the ResNet for our ImageNet experiments (see Section IV-D) has a transition block structure with a strided convolution followed by Batch Normalization. To train the network, we used stochastic gradient descent with a weight decay of 1e−3, momentum of 0.9 and mini batch-size of 256. The learning rate starts at 0.1 and is divided by 10 after epoch 81, and by 10 again after epoch 122. We performed standard center crop + horizontal flip for data augmentation: four pixels were added on each side of the image, and a random 32 x 32 crop was extracted, which was randomly flipped horizontally. Only color-normalized 32 x 32 images were used during the test phase. Figure 4 presents ResNet test error (in %) medians over five tries on both CIFAR datasets. ResNet obtained a minimum median error rate on CIFAR-10 of 5.41% with BN+ReLU, 5.99% with ELU, 5.36% with PELU and 5.26% with BNPReLU, while ResNet obtained a minimum median error rate on CIFAR-100 of 24.99% with BN+ReLU, 25.08% with ELU, 24.55% with PELU and 25.50% with BN+PReLU. In comparison to ELU, PELU obtained a relative improvement of 10.52% and 2.11% on CIFAR-10 and CIFAR-100 respectively. It is interesting to note that PELU only adds 112 additional parameters, a negligible increase of 0.006% over the total number of parameters. We observed that PELU has a better convergence behavior than ELU. As shown in Figure 4, ELU has a large test error rate increase at the end of the second stage of the training phase on both CIFAR-10 and CIFAR-100 datasets. Although PELU has also a test error rate increase at the end of the second stage, it does not increase as high as ELU. We further observe a small test error rate increase at the end of the training phase for ELU, while PELU converges in a steady way without a test error rate increase. These results show that training a ResNet with our parameterization can improve the performance and the convergence behavior over a ResNet with ELU activation. Compared to ReLU, PReLU obtained a smaller minimum median error rate on CIFAR-10 and a smaller average median error rate on CIFAR-100. As shown in Table I, PReLU obtained a minimum median error rate of 5.36 compared to 5.41 on CIFAR-10, and an average median error rate of 25.83 compared to 25.92 on CIFAR-100. Although PReLU obtained the same minimum median error rate than PELU on CIFAR10, it is significantly higher on CIFAR-100. Note that our main contribution is showing performance improvement over ELU, and that we only add PReLU as an additional reference. Nonetheless, we observe that our PELU parameterization of ELU obtains higher relative improvements than the PReLU parameterization of ReLU. 20 18 16 14 12 10 8 6 CIFAR-10 0 50 100 Epoch 150 CIFAR-10 0 50 100 Epoch 150 200 ELU BN-ELU 40 35 30 25 200 PELU BN-PELU CIFAR-100 45 Error (%) ELU BN-ELU 0 50 100 150 Epoch 200 CIFAR-100 45 Error (%) Error (%) Error (%) 20 18 16 14 12 10 8 6 PELU BN-PELU 40 35 30 25 0 50 100 150 Epoch 200 Fig. 5. Effect of using BN before ELU (first row) and PELU (second row) activations in a ResNet with 110 layers on both CIFAR-10 and CIFAR-100 datasets. We show the convergence behavior of the median test error over five tries. In both cases, BN worsen performance of ELU and PELU. Note that we still use BN after the second conv layer, as seen in Figure 3. C. Understanding the effect of Batch Normalization In this section, we show that using BN before our PELU activation has a detrimental effect on its performance. Figure 5 presents the influence of using BN before ELU and PELU in a ResNet with 110 layers on both CIFAR-10 and CIFAR-100 datasets. We trained the networks using the same framework as in Section IV-B, but added BN before each activate. Note that in all cases, we use BN after the second convolutional layer in the basic block (see Figure 3). The results show a large error rate increase on both CIFAR10 and CIFAR-100 dataset for each ELU and PELU activation. The minimum median test error for ELU increases from 5.99% and 25.08% to 10.39% and 34.75% on CIFAR-10 and CIFAR100 respectively, while for PELU it increases from 5.36% and 24.55% to 5.85% and 25.38%. We also observe that the relative error rate increase for our PELU is smaller than for ELU. Indeed, ELU has a relative minimum test error rate increase of 73% and 39% on CIFAR-10 and CIFAR100 respectively, while PELU has 9% and 3%. Although this shows that our PELU parameterization reduces the detrimental effect of using BN before the activation, PELU should not be preceded by BN. D. ImageNet Object Recognition We tested the proposed PELU on the ImageNet 2012 task (ILSVRC2012) using four different network architectures: ResNet18 [9], Network in Network (NiN) [10], All-CNN [11] and Overfeat [13]. The ResNet18 building block structure is shown in Figure 3. In order not to favor PELU to the detriment of ELU and BN+ReLU, we performed minimal changes by only replacing the activation function. We used either PELU, ELU or BN+ReLU for the activation module. Each network was trained following Chintala’s Torch implementation imagenet-multiGPU.torch 2 with the training regimes 2 https://github.com/soumith/imagenet-multiGPU.torch TABLE II I MAGE N ET TRAINING REGIMES FOR MODIFYING THE LEARNING RATE AND THE WEIGHT DECAY. W E TRAINED N I N AND R ES N ET 18 USING THE FIRST ONE , AND TRAINED A LL -CNN AND OVERFEAT USING THE SECOND ONE . T HE FIRST REGIME STARTS AT A HIGHER LEARNING RATE , BUT HAS A LARGER DECAY THAN THE SECOND REGIME . Regime #1 (ResNet18, NiN) Epoch Learning Rate Weight Decay 1 1e-1 5e-4 10 1e-2 5e-4 20 1e-3 0 25 1e-4 0 Regime #2 (Overfeat, AllCNN) Epoch Learning Rate Weight Decay 1 1e-2 5e-4 19 5e-3 5e-4 30 1e-3 0 44 5e-4 0 53 1e-4 0 shown in Table II. Regime #1 starts at a higher learning rate than regime #2, and has a larger learning rate decay. Figure 6 presents the TOP-1 error rate (in %) of all four networks on ImageNet 2012 validation dataset. In all cases, the networks using PELU outperformed the networks using ELU. NiN obtained the best result of 36.06% with PELU, which corresponds to a relative improvement of 7.29% compared to ELU (40.40%). Since only 24 additional parameters were added to the network, this performance improvement indicates that PELU’s parameterization acts in a different way than the weights and biases. Adding 24 additional weights throughout the network would not have been sufficient to increase the representative ability enough to get the observed performance improvement. Since such a low number of weights cannot significantly increase the expressive power of the network, these results indicate that the networks benefit from PELU. As shown in Figure 6, the training regimes have an interesting effect on the convergence of the networks. The performance of PELU is closer to the performance of ELU for regime #2, but is significantly better than ELU for regime #1. ImageNet - NiN 70 BN-ReLU ELU PELU 65 50 45 55 50 45 40 40 35 35 0 5 10 15 Epoch 20 25 30 30 35 ImageNet - All-CNN 80 70 Error (%) 55 45 Epoch 40 50 20 25 30 60 35 ImageNet - Overfeat BN-ReLU ELU PELU 55 45 30 Epoch 60 50 20 15 65 50 10 10 70 60 0 5 75 65 40 0 80 ReLU ELU PELU 75 Error (%) 60 55 30 BN-ReLU ELU PELU 65 Error (%) Error (%) 60 ImageNet - ResNet18 70 40 0 10 20 30 Epoch 40 50 60 Fig. 6. TOP-1 error rate progression (in %) of ResNet18, NiN, Overfeat and All-CNN on ImageNet 2012 validation set. NiN and ResNet18 (top row) used training regime #1, while All-CNN and Overfeat (bottom row) used training regime #2 (see Table II). PELU has the lowest error rates for all networks. Regime #1 shows a greater performance gap between ELU and PELU than regime #2. TABLE III R ES N ET 110 LAYERS TEST ERROR ( IN %) ON BOTH CIFAR-10 AND CIFAR-100 DATASETS . W E REPORT THE MEAN ERROR OVER THE LAST FIVE EPOCHS AND THE MINIMUM ERROR OVER ALL EPOCHS ( INSIDE PARENTHESIS ) OF THE MEDIAN ERROR OVER FIVE TRIES . W E COMPARE EACH FOUR PELU CONFIGURATIONS . O UR PROPOSED PELU CONFIGURATION (a, 1/b) OBTAINED THE BEST PERFORMANCE . Configuration (a, 1/b) (1/a, 1/b) (1/a, b) (a, b) CIFAR-10 5.51 5.73 6.51 6.74 (5.36) (5.60) (6.00) (6.12) CIFAR-100 25.02 25.68 26.33 26.20 (24.55) (25.17) (25.48) (25.24) We also observe that the error rates of All-CNN and Overfeat with PELU increase by a small amount starting at epoch 44, but stay steady for ELU and ReLU. These results suggest that training regimes with larger learning rates and decays help PELU to obtain a better performance improvement. E. Experimenting with Parameter Configuration The proposed PELU activation function (2) has two parameters a and b, where a is used with a multiplication and b with a division. A priori, any of the four configurations (a, b), (a, 1/b), (1/a, b) or (1/a, 1/b) could be used as param- eterization. Note that these configurations are not reciprocal due to weight decay, which favors low weight magnitude. For instance, favoring low magnitude for parameter b with the (a, b) configuration favors a high PELU slope a/b. On the contrary, favoring low magnitude for parameter b with the (a, 1/b) favors a low PELU slope ab. In order to better understand the difference between each configuration, we performed an experimental evaluation on the CIFAR-10 and CIFAR-100 datasets using the 110-layers ResNet as defined in Section IV-B. As shown in Table III, our proposed parameterization (a, 1/b) obtained the best accuracy. Parameterization (a, 1/b) obtained minimum test error medians of 5.36% and 24.55% on CIFAR-10 and CIFAR-100 respectively, while (1/a, 1/b) obtained 5.60% and 25.17%, (1/a, b) obtained 6.00% and 25.48%, and (a, b) obtained 6.12% and 25.24%. These results also show that the two parameterizations with 1/b obtained a significantly lower error rate than the two parameterizations with b. From the convergence behavior in Figure 7, we see that the parameterizations with b have a larger error increase during the second stage of the training phase than parameterizations with 1/b, and converge to lower error rates. These results concur with our observations in Section III of the effect of the parameters. Since weight decay pushes the weight magnitude CIFAR-10 5 4 3 2 1 14 0 12 0 1 2 3 4 6 7 8 1e4 Layer 3 6 10 5 Update Steps Value Value 3 2 6 100 Epoch 150 200 0 CIFAR-100 1 2 3 4 a, 1/b 1/a, 1/b 1/a, b a, b 40 6 7 a/b (slope) a (negative saturation) 6 4 2 0 0 1 2 3 4 6 7 8 1e4 Layer 7 2.5 Value 2.0 1.5 Fig. 7. Experimenting with the PELU parameter configuration in a ResNet with 110 layers on both CIFAR-10 and CIFAR-100 datasets. We show the convergence behavior of the median test error over five tries. Our proposed parameterization (a, 1/b) obtained the best performance. towards zero, the function saturates slower as b decreases with parameterizations using b. This encourages the function to have an almost linear shape on all its input values, which removes the non-linear characteristic of the activation. On the contrary, the function saturates faster a b decreases with parameterizations using 1/b. This helps the activation to keep its non-linear nature, which explains the observed performance gaps between b and 1/b. 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 0 1 2 3 5 Update Steps 6 7 6 7 8 1e4 2 0 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 1 2 3 4 5 Update Steps 6 7 8 1e4 Layer 6 a/b (slope) a (negative saturation) 0 1 2 3 4 5 Update Steps 6 7 8 1e4 Layer 8 a/b (slope) a (negative saturation) 6 4 0 1 2 3 4 5 Update Steps 6 7 8 1e4 Layer 10 a/b (slope) a (negative saturation) 3.0 2.5 2.0 1.5 1.0 0.5 0 1 2 3 4 5 Update Steps 6 7 0.0 8 1e4 Layer 11 3 2 0 1 2 3 4 6 7 8 1e4 Layer 12 a/b (slope) a (negative saturation) 2.0 1 5 Update Steps 2.5 a/b (slope) a (negative saturation) 4 1.5 1.0 0.5 0 1 2 3 4 5 Update Steps 6 7 0.0 8 1e4 Layer 13 3 4 5 Update Steps 6 7 8 1e4 Layer 14 a/b (slope) a (negative saturation) 10 8 6 4 0.5 0.0 2 12 Value 1.0 1 14 2.0 1.5 0 16 a/b (slope) a (negative saturation) 2.5 Value 5 a/b (slope) a (negative saturation) 3.5 a/b (slope) a (negative saturation) 5 0 4 Layer 4 0 8 1e4 Layer 9 3.0 F. Parameter Progression We perform a visual evaluation of the non-linear behaviors adopted by a Vgg network during training on the CIFAR10 dataset [12]. Figure 8 shows the progression of the slope ( ab ) and the negative of the saturation point (parameter a) for PELU at each layer of Vgg. We can see different behaviors. At layers 2, 4, 7 and 10, the slope quickly increased to a large value, then decreased and converged at a value near 1. As for parameter a, it quickly converged to a value near 0. A slope near 1 and a negative saturation near 0 indicates that the network learned activations having the same shape as ReLU. This is an interesting result because ReLU has the important effect of promoting activation sparsity [23], [15]. Although we do not have a clear understanding to why the network increases the slope then decreases 4 Value 200 Value Epoch 150 0.0 Value 100 3 Update Steps 2 0.5 50 2 8 1.0 0 1 10 a/b (slope) a (negative saturation) 3.0 Value 5 Update Steps 3.5 25 0 3 0 8 1e4 Layer 5 8 35 30 5 Update Steps 10 Value 45 0 Value 50 a/b (slope) a (negative saturation) 1 1 0 Layer 2 4 4 8 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 5 a/b (slope) a (negative saturation) 5 Value Error (%) 16 a/b (slope) a (negative saturation) 6 Value a, 1/b 1/a, 1/b 1/a, b a, b 18 Error (%) Layer 1 7 Value 20 2 0 1 2 3 4 5 Update Steps 6 7 8 1e4 0 0 1 2 3 4 5 Update Steps 6 7 8 1e4 Fig. 8. PELU parameter progression at each layer of Vgg trained on CIFAR10. We present the variation of the slope ( ab ) and the negative saturation (parameter a). The network adopted different non-linear behaviors throughout the training phase. it before converging to ReLU, we believe that increasing the slope helps early during training to disentangle redundant neurons. Since peak activations scatter more the inputs than flat ones, spreading values may allow the network to declutter neurons activating similarly to the same input patterns. Another interesting observation is that, apart from the ReLU layers (layers 2, 4, 7 and 10), the negative saturations of all layers converged at values other than 0. For instance, parameter a converges to a value near 0.5 at layer 1, while it converges to a value near 2 at layer 14. A negative saturation other than zero indicates that the learned PELU activations outputs negative values for negative arguments. The Vgg network had the possibility to learn all activation functions with a zero negative saturation (i.e. shaped like ReLU), but opted for a majority of activations with a non-zero negative saturation. Having activation functions with negative values has been previously analyzed in the context of the standard ELU activation, and it has been proposed that it helps to manage bias shift [7]. These results constitute an additional experimental evidence that this characteristic is important for the network. V. D ISCUSSION During all our experiments with ELU and PELU, we did not use Batch Normalization (BN) before the activations. This is due to the detrimental effect of preceding PELU and ELU with BN, as we have observed in Section IV-C with our ResNet experiments on CIFAR-10 and CIFAR-100. Although this detrimental effect has also been previously observed with ELU by Clevert and his coworkers [7], it is unclear why BN before ELU and PELU increases error rate, but reduces error rate before ReLU. One important difference is that ReLU is positively scale invariant and ELU is not. Indeed, for ReLU we have max{0, kx} = k max{0, x}, where k ≥ 0, while for ELU, which can be expressed as max{0, x} + min{0, exp{x} − 1}, we have min{0, exp{kx} − 1} 6= k min{0, exp{x} − 1}. The fact that ReLU is positively scale invariant and ELU is not may be part of the reason why BN before ReLU helps but harms before ELU. Given that BN performs mean and standard deviation scaling, followed by an affine transformation (scaled by γ and shifted by β), using a positively scale invariant activation function may be essential for BN to properly reduce internal covariate shift [8] or manage bias shift [7]. We could validate this hypothesis by experimenting with a new positively scale invariant activation function and observing whether BN helps or not. We leave this idea as future work. VI. C ONCLUSION Object recognition is an essential ability for improving visual perception in automated vision systems performing complex scene understanding. In a recent work, the Exponential Linear Unit (ELU) has been proposed as a key element in Convolutional Neural Networks (CNNs) for reducing bias shift, but has the inconvenience of defining a parameter that must be set by hand. In this paper, we proposed the Parametric ELU (PELU) that alleviates this limitation by learning a parameterization of the ELU activation function. Our results on the CIFAR-10/100 and ImageNet datasets using the ResNet, NiN, All-CNN and Overfeat networks show that CNNs with PELU have better performance than CNNs with ELU. Our experiments with Vgg have shown that the network uses the added flexibility provided by PELU by learning different activation shapes at different locations in the network. Parameterizing other activation functions, such as Softplus, Sigmoid or Tanh, could be worth investigating. ACKNOWLEDGEMENTS We gratefully acknowledge the support of NVIDIA Corporation for providing the Tesla K20, K80 and Titan X for our experiments. R EFERENCES [1] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classification with deep convolutional neural networks,” in NIPS, 2012, pp. 1097–1105. [2] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015. [3] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A neural image caption generator,” in CVPR, 2015, pp. 3156–3164. [4] M. Jaderberg, K. Simonyan, A. Zisserman et al., “Spatial transformer networks,” in NIPS, 2015, pp. 2017–2025. [5] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards realtime object detection with region proposal networks,” in NIPS, 2015, pp. 91–99. [6] J. Hosang, R. Benenson, P. Dollár, and B. Schiele, “What makes for effective detection proposals?” PAMI, vol. 38, no. 4, pp. 814–830, 2016. [7] D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (elus),” arXiv preprint arXiv:1511.07289, 2015. [8] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in ICML, 2015, pp. 448–456. [9] A. Shah, E. Kadam, H. Shah, and S. Shinde, “Deep residual networks with exponential linear unit,” arXiv preprint arXiv:1604.04112, 2016. [10] M. Lin, Q. Chen, and S. Yan, “Network in network,” arXiv preprint arXiv:1312.4400, 2013. [11] J. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller, “Striving for simplicity: The all convolutional net,” in ICLR (workshop track), 2015. [12] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. [13] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, “Overfeat: Integrated recognition, localization and detection using convolutional networks,” arXiv preprint arXiv:1312.6229, 2013. [14] K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,” in ICCV, 2015, pp. 1026–1034. [15] A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlinearities improve neural network acoustic models,” in ICML Workshop on Deep Learning for Audio, Speech and Language Processing. Citeseer, 2013. [16] F. Agostinelli, M. Hoffman, P. Sadowski, and P. Baldi, “Learning activation functions to improve deep neural networks,” arXiv preprint arXiv:1412.6830, 2014. [17] I. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Bengio, “Maxout networks,” in ICML, 2013, pp. 1319–1327. [18] X. Jin, C. Xu, J. Feng, Y. Wei, J. Xiong, and S. Yan, “Deep learning with s-shaped rectified linear activation units,” arXiv preprint arXiv:1512.07030, 2015. [19] N. Srivastava and R. R. Salakhutdinov, “Multimodal learning with deep boltzmann machines,” in NIPS, 2012, pp. 2222–2230. [20] Y. Lecun and C. Cortes, “The MNIST database of handwritten digits.” [Online]. Available: http://yann.lecun.com/exdb/mnist/ [21] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from overfitting,” JMLR, vol. 15, no. 1, pp. 1929–1958, 2014. [22] G. H. Tijmen Tieleman, “Lecture 6.5 - rmsprop: Divide the gradient by a running average of its recent magnitude.” COURSERA: Neural Networks for Machine Learning, vol. 4, 2012. [23] V. Nair and G. Hinton, “Rectified linear units improve restricted boltzmann machines,” in ICML, 2010, pp. 807–814.
9cs.NE
Technical report Magnetometer calibration using inertial sensors arXiv:1601.05257v3 [cs.SY] 14 Jul 2016 Manon Kok and Thomas B. Schön • Please cite this version: Manon Kok and Thomas B. Schön. Magnetometer calibration using inertial sensors. IEEE Sensors Journal, Volume 16, Issue 14, Pages 5679–5689, 2016. Abstract In this work we present a practical algorithm for calibrating a magnetometer for the presence of magnetic disturbances and for magnetometer sensor errors. To allow for combining the magnetometer measurements with inertial measurements for orientation estimation, the algorithm also corrects for misalignment between the magnetometer and the inertial sensor axes. The calibration algorithm is formulated as the solution to a maximum likelihood problem and the computations are performed offline. The algorithm is shown to give good results using data from two different commercially available sensor units. Using the calibrated magnetometer measurements in combination with the inertial sensors to determine the sensor’s orientation is shown to lead to significantly improved heading estimates. Keywords: Magnetometers, calibration, inertial sensors, maximum likelihood, grey-box system identification, sensor fusion. Magnetometer calibration using inertial sensors Manon Kok? and Thomas B. Schön† ? Department of Electrical Engineering, Linköping University, SE-581 83 Linköping, Sweden E-mail: [email protected] † Department of Information Technology, Uppsala University, Sweden E-mail: [email protected] July 15, 2016 Abstract In this work we present a practical algorithm for calibrating a magnetometer for the presence of magnetic disturbances and for magnetometer sensor errors. To allow for combining the magnetometer measurements with inertial measurements for orientation estimation, the algorithm also corrects for misalignment between the magnetometer and the inertial sensor axes. The calibration algorithm is formulated as the solution to a maximum likelihood problem and the computations are performed offline. The algorithm is shown to give good results using data from two different commercially available sensor units. Using the calibrated magnetometer measurements in combination with the inertial sensors to determine the sensor’s orientation is shown to lead to significantly improved heading estimates. Keywords: Magnetometers, calibration, inertial sensors, maximum likelihood, grey-box system identification, sensor fusion. 1 Introduction Nowadays, magnetometers and inertial sensors (gyroscopes and accelerometers) are widely available, for instance in dedicated sensor units and in smartphones. Magnetometers measure the local magnetic field. When no magnetic disturbances are present, the magnetometer measures a constant local magnetic field vector. This vector points to the local magnetic north and can hence be used for heading estimation. Gyroscopes measure the angular velocity of the sensor. Integration of the gyroscope measurements gives information about the change in orientation. However, it does not provide absolute orientation estimates. Furthermore, the orientation estimates suffer from integration drift. Accelerometers measure the sensor’s acceleration in combination with the earth’s gravity. In the case of small or zero acceleration, the measurements are dominated by the gravity component. Hence, they can be used to estimate the inclination of the sensor. Inertial sensors and magnetometers have successfully been used to obtain accurate 3D orientation estimates for a wide range of applications. For this, however, it is imperative that the sensors are properly calibrated and that the sensor axes are aligned. Calibration is specifically of concern for the magnetometer, which needs recalibration whenever it is placed in a (magnetically) different environment. When the magnetic disturbance is a result of the mounting of the magnetometer onto a magnetic object, the magnetometer can be calibrated to compensate for the presence of this disturbance. This is the focus of this work. Our main contribution is a practical magnetometer calibration algorithm that is designed to improve orientation estimates when combining calibrated magnetometer data with inertial data. The word practical refers to the fact that the calibration does not require specialized additional equipment and can therefore be performed by any user. More specifically, this means that the orientation of the sensor is not assumed to be known. Instead, the calibration problem is formulated as an orientation estimation problem in the presence of unknown parameters and is posed as a maximum likelihood (ML) problem. The algorithm calibrates the magnetometer for the presence of magnetic disturbances, for magnetometer sensor errors and for misalignment between the magnetometer and the inertial sensor axes. Using the calibrated magnetometer measurements to estimate the sensor’s orientation is experimentally shown to lead to significantly improved heading estimates. We aggregate and extend the work from [1] and [2] 1 1 0 −1 −2 −1 0 1 2 Figure 1: Example calibration results with an ellipsoid of magnetometer data before calibration (red) and a unit sphere of data after calibration (blue). with improvements on the implementation of the algorithm. Furthermore, we include a more complete description and analysis, more experimental results and a simulation study illustrating the heading accuracy that can be obtained with a properly calibrated sensor. To perform the calibration, the sensor needs to be rotated in all possible orientations. A perfectly calibrated magnetometer would in that case measure rotated versions of the local magnetic field vector. Hence, the magnetometer data would lie on a sphere. In practice, however, the magnetometer will often measure an ellipsoid of data instead. The calibration maps the ellipsoid of data to a sphere as illustrated in Figure 1. The alignment of the inertial and magnetometer sensor axes determines the orientation of the sphere. Since we are interested in improving the heading estimates, the actual magnitude of the local magnetic field is of no concern. Hence, we assume without loss of generality that the norm is equal to 1, i.e. the sphere in Figure 1 is a unit sphere. 2 Related work Traditional magnetometer calibration approaches assume that a reference sensor is available which is able to provide accurate heading information. A well-known example of this is compass swinging [3]. To allow for any user to perform the calibration, however, a large number of approaches have been developed that remove the need for a source of orientation information. One class of these magnetometer calibration algorithms focuses on minimizing the difference between the magnitude of the measured magnetic field and that of the local magnetic field, see e.g. [4]. This approach is also referred to as scalar checking [5]. Another class formulates the calibration problem as an ellipsoid fitting problem, i.e. as the problem of mapping an ellipsoid of data to a sphere, see e.g. [6, 7, 8]. The benefit of using this formulation, is that there is a vast literature on solving ellipsoid fitting problems, see e.g. [9, 10]. Outside of these two classes, a large number of other calibration approaches is also available, for instance [11], where different formulations of the calibration problem in terms of an ML problem are considered. The benefit of the approaches discussed above is that they can be used with data from a magnetometer only. Our interest, however, lies in calibrating a magnetometer for improved heading estimation in combination with inertial sensors. Alignment of the sensor axes of the inertial sensors and the magnetometer is in this case crucial. This alignment can be seen as determining the orientation of the blue sphere of calibrated magnetometer data in Figure 1. Algorithms that only use magnetometer data can map the red ellipsoid of data to a sphere, but without additional information, the rotation of this sphere remains unknown. A number of recent approaches include a second step in the calibration algorithm to determine the misalignment [6, 12, 13, 14] between different sensor axes. A common choice to align the magnetometer and inertial sensor axes, is to use accelerometer measurements from periods of fairly small accelerations [12, 13]. The downside of this approach is that a threshold for using accelerometer measurements needs to be determined. Furthermore, data from the gyroscope is hereby omitted. In [15] on the other hand, the problem is reformulated in terms of the change in orientation, allowing for direct use of the 2 gyroscope data. In our algorithm we instead formulate the magnetometer calibration problem as a problem of estimating the sensor’s orientation in the presence of unknown (calibration) parameters. This formulation naturally follows from the fact that the problem of orientation estimation and that of magnetometer calibration are inherently connected: If the magnetometer is properly calibrated, good orientation estimates can be obtained. Reversely, if the orientation of the sensor is known accurately, the rotation of the sphere in Figure 1 can accurately be determined, resulting in a good magnetometer calibration. In this formulation, data from the accelerometer and the gyroscope is used to aid the magnetometer calibration. Our formulation of the calibration problem requires solving a non-convex optimization problem to obtain ML estimates of the calibration parameters. To obtain good initial values of the parameters, an ellipsoid fitting problem and a misalignment estimation problem are solved. Solving the calibration problem as a two-step procedure is similar to the approaches in [12, 13]. We analyze the quality of the initial estimates and of the ML estimates in terms of their heading accuracy, both for experimental and simulated data. Based on this analysis, we show that significant heading accuracy improvements can be obtained by using the ML estimates of the parameters. 3 Problem formulation Our magnetometer calibration algorithm is formulated as a problem of determining the sensor’s orientation in the presence of unknown model parameters θ. It can hence be considered to be a grey-box system identification problem. A nonlinear state space model on the following form is used xt+1 = ft (xt , ωt , eω,t , θ),     ya,t ha,t (xt ) yt = = + et (θ), ym,t hm,t (xt , θ) (1a) (1b) where the state xt represents the sensor’s orientation at time t. We use the change in orientation, i.e. the angular velocity ωt , as an input to the dynamic model ft (·). The angular velocity is measured by the gyroscope. However, the measurements yω,t are corrupted by a constant bias δω and Gaussian i.i.d. measurement noise with zero mean and covariance Σω , i.e. eω,t ∼ N (03×1 , Σω ). The measurement models ha,t (·) and hm,t (·) in (1b) describe the accelerometer measurements ya,t and the magnetometer measurements ym,t , respectively. The accelerometer measurement model assumes that the acceleration of the sensor is small compared to the earth gravity. Since the magnetometer is not assumed to be properly calibrated, the magnetometer measurement model hm,t (·) depends on the parameter vector θ. The exact details of the magnetometer measurement model will be introduced in Section 4. The accelerometer and magnetometer measurements are corrupted by Gaussian i.i.d. measurement noise      ea,t Σa 03×3 et = ∼ N 06×1 , . (2) em,t 03×3 Σm The calibration problem is formulated as an ML problem. Hence, the parameters θ in (1) are found by maximizing the likelihood function pθ (y1:N ), θbML = arg max pθ (y1:N ), (3) θ∈Θ where y1:N = {y1 , . . . , yN } and Θ ⊆ Rnθ . Using conditional probabilities and the fact that the logarithm is a monotonic function we have the following equivalent formulation of (3), θbML = arg min − θ∈Θ N X t=1 log pθ (yt | y1:t−1 ), (4) where we use the convention that y1:0 , ∅. The ML estimator (4) enjoys well-understood theoretical properties including strong consistency, asymptotic normality, and asymptotic efficiency [16]. The state space model (1) is nonlinear, implying that there is no closed form solution available for the one step ahead predictor pθ (yt | y1:t−1 ) in (4). This can systematically be handled using sequential Monte Carlo methods (e.g. particle filters and particle smoothers), see e.g. [17, 18]. However, for the 3 magnetometer calibration problem it is sufficient to make use of a more pragmatic approach; we simply approximate the one step ahead predictor using an extended Kalman filter (EKF). The result is  pθ (yt | y1:t−1 ) ≈ N yt ; ybt|t−1 (θ), St (θ) , (5) where the mean value ybt|t−1 (θ) and the covariance St (θ) are obtained from the EKF [19]. Inserting (5) into (4) and neglecting all constants not depending on θ results in the following optimization problem, N 1X kyt − ybt|t−1 (θ)k2S −1 (θ) + log det St (θ), min t θ∈Θ 2 t=1 (6) which we can solve for the unknown parameters θ. The problem (6) is non-convex, implying that a good initial value for θ is required. 4 Magnetometer measurement model In the case of perfect calibration, a magnetometer measures the local magnetic field and its measurements will therefore lie on a sphere with a radius equal to the local magnetic field. Since we are interested in using the magnetometer measurements to improve the orientation estimates from the state space model (1), the actual magnitude of the local magnetic field is of no concern. Hence, we assume without loss of generality that its norm is equal to one. We denote the normalized local magnetic field by mn . Ideally, the magnetometer measurements then lie on a sphere with radius equal to one as hm,t = mbt = Rtbn mn , (7) where hm,t is defined in (1b). The explicit dependence on xt and θ has been omitted for notational simplicity. The matrix Rtbn is the rotation matrix representation of the orientation at time t. The superscript bn denotes that the rotation is from the navigation frame n to the body frame b. The body frame b is aligned with the sensor axes. The navigation frame n is aligned with the earth’s gravity and the local magnetic field. In case the coordinate frame in which a vector is defined can be ambiguous, we explicitly indicate in which coordinate frame the vector is expressed by adding a superscript b or n. Hence, mn denotes the normalized local magnetic field in the navigation frame n while mbt denotes the normalized local magnetic field in the body frame b. The latter is time-dependent and therefore also has a subscript t. Note that the rotation from navigation frame to body frame is denoted Rtnb and Rtbn = (Rtnb )T . In outdoor environments, the local magnetic field is equal to the local earth magnetic field. Its horizontal component points towards the earth’s magnetic north pole. The ratio between the horizontal and vertical component depends on the location on the earth and can be expressed in terms of the dip angle δ. In indoor environments, the magnetic field can locally be assumed to be constant and points towards a local magnetic north. This is not necessarily the earth’s magnetic north pole. Choosing the navigation frame n such that the x-axis is pointing towards the local magnetic north, mn can be parametrized in terms of its vertical component mnz mn = q 2 1 − (mnz ) 0 mnz T , (8a) or in terms of the dip angle δ mn = cos δ − sin δ 0 T . (8b) Note that the two parametrizations do not encode exactly the same knowledge about the magnetic field; the first component of mn in (8a) is positive by construction while this is not true for (8b). However, both parametrizations will be used in the remainder. It will be argued that no information is lost by using (8b) if the parameter estimates are properly initialized. The main need for magnetometer calibration arises from the fact that a magnetometer needs recalibration each time it is placed in a magnetically different environment. Specifically, a magnetometer measures a superposition of the local magnetic field and of the magnetic field due to the presence of magnetic material in the vicinity of the sensor. In case this magnetic material is rigidly attached to the magnetometer, it is possible to calibrate the magnetometer measurements for this. The magnetic 4 material can give rise to both hard and soft iron contributions to the magnetic field. Hard iron effects are due to permanent magnetization of the magnetic material and lead to a constant 3 × 1 offset vector ohi . Soft iron effects are due to magnetization of the material as a result of an external magnetic field and therefore depend on the orientation of the material with respect to the local magnetic field. We model this in terms of a 3 × 3 matrix Csi . Hence, the magnetometer measurements do not lie on a sphere as in (7), but instead, they lie on a translated ellipsoid as hm,t = Csi Rtbn mn + ohi . (9) As discussed in Section 2, when calibrating the magnetometer to obtain better orientation estimates, it is important that the magnetometer and the inertial sensor axes are aligned. Let us now be more specific about the definition of the body frame b and define it to be located in the center of the accelerometer triad and aligned with the accelerometer sensor axes. Furthermore, let us assume that the accelerometer and gyroscope axes are aligned. Defining the rotation between the body frame b and the magnetometer sensor frame bm as Rbm b , the model (9) can be extended to hm,t = Csi Rbm b Rtbn mn + ohi . (10) Finally, the magnetometer calibration can also correct for the presence of sensor errors in the magnetometer. These errors are sensor-specific and can differ for each individual magnetometer. They can be subdivided into three components, see e.g. [8, 7, 6]: 1. Non-orthogonality of the magnetometer axes, represented by a matrix Cno . 2. Presence of a zero bias or null shift, implying that the magnetometer will measure a non-zero magnetic field even if the magnetic field is zero, defined by ozb . 3. Difference in sensitivity of the three magnetometer axes, represented by a diagonal matrix Csc . We can therefore extend the model (10) to also include the magnetometer sensor errors as  hm,t = Csc Cno Csi Rbm b Rtbn mn + ohi + ozb . (11) To obtain a correct calibration, it is fortunately not necessary to identify all individual contributions of the different components in (11). Instead, they can be combined into a 3 × 3 distortion matrix D and a 3 × 1 offset vector o where D = Csc Cno Csi Rbm b , (12a) o = Csc Cno ohi + ozb . (12b) The resulting magnetometer measurement model in (1b) can be written as ym,t = DRtbn mn + o + em,t . (13) In deriving the model we have made two important assumptions: Assumption 1. The calibration matrix D and offset vector o in (12) are assumed to be time-independent. This implies that we assume that the magnetic distortions are constant and rigidly attached to the sensor. Also, the inertial and the magnetometer sensor axes are assumed to be rigidly attached to each other, i.e. their misalignment is represented by a constant rotation matrix. Additionally, in our algorithm we will assume that their misalignment can be described by a rotation matrix, i.e. that their axes are not mirrored with respect to each other. Assumption 2. The local magnetic field mn is assumed to be constant. In outdoor environments, this is typically a physically reasonable assumption. In indoor environments, however, the local magnetic field can differ in different locations in the building and care should be taken to fulfill the assumption. 5 Calibration algorithm In our magnetometer calibration algorithm we solve the optimization problem (6) to estimate the parameter vector θ. In this section we introduce the resulting calibration algorithm which is summarized in Algorithm 1. In Section 5.1, we first discuss our optimization strategy. A crucial part of this optimization strategy is the evaluation of the cost function. Some details related to this are discussed in Section 5.2. Finally, in Section 5.3 we introduce the parameter vector θ in more detail. 5 Algorithm 1 Magnetometer and inertial calibration b 0 , ob0 , m b ω,0 , Σ b a,0 , Σ b m,0 using three steps 1. Determine an initial parameter estimate D b n0 , δbω,0 , Σ b ω,0 , Σ b a,0 , Σ b m,0 . (a) Initialize δbω,0 , Σ e 0 and ob0 based on ellipsoid fitting (see Section 6.1). (b) Obtain an initial D b 0 , ob0 and m (c) Obtain initial D b n0 by initial determination of the sensor axis misalignment (see Section 6.2). 2. Set i = 0 and repeat, b i , obi , m b ω,i , Σ b a,i , Σ b m,i to obtain (a) Run the EKF using the current estimates D b ni , δbω,i , Σ {b yt|t−1 (θbi ), St (θbi )}N t=1 and evaluate the cost function in (6). (b) Determine θbi+1 using the numerical gradient of the cost function in (6), its approximate Hessian and a backtracking line search algorithm. b ω,i+1 , Σ b a,i+1 , Σ b m,i+1 from θbi+1 . b i+1 , obi+1 , m (c) Obtain D b n , δbω,i+1 , Σ i+1 (d) Set i := i + 1 and repeat from Step 2a until convergence. 5.1 Optimization algorithm The optimization problem (6) is solved in Step 2 of Algorithm 1. Standard unconstrained minimization techniques are used, which iteratively update the parameter estimates as −1 θi+1 = θi − αi [H(θi )] G(θi ), (14) −1 where the direction of the parameter update at iteration i is determined by [H(θi )] G(θi ). The step length of the update at iteration i is denoted by αi . Typical choices for the search direction include choosing G(θi ) to be the gradient of the cost function in (6) and H(θi ) to be its Hessian. This leads to a Newton optimization algorithm. However, computing the gradient and Hessian of (6) is not straightforward. Possible approaches are discussed in [20, 21] for the case of linear models. In the case of nonlinear models, however, they only lead to approximate gradients, see e.g. [22, 23]. For this reason we make use of a numerical approximation of G(θi ) instead and use a Broyden-Fletcher-Goldfarb-Shanno (BFGS) method with damped updating [24] to approximate the Hessian. Hence, the minimization is performed using a quasi-Newton optimization algorithm. A backtracking line search is used to find a good step length αi . Proper initialization of the parameters is crucial since the optimization problem (6) is non-convex. Step 1 summarizes the three-step process used to obtain good initial estimates of all parameters. 5.2 Evaluation of the cost function An important part of the optimization procedure is the evaluation of the cost function in (6). This requires running an EKF using the state space model (1) to estimate the orientation of the sensor. This EKF uses the angular velocity ωt as an input to the dynamic model (1a). An estimate of the angular velocity is obtained from the gyroscope measurements yω,t which are modeled as yω,t = ωt + δω + eω,t . (15) The measurement model (1b) entails the accelerometer measurements and the magnetometer measurements. The magnetometer measurement model can be found in (13). The accelerometer measurements ya,t are modeled as ya,t = Rtbn (ant − g n ) + ea,t ≈ −Rtbn g n + ea,t , (16) where ant denotes the sensor’s acceleration in the navigation frame and g n denotes the earth’s gravity. The rotation matrix Rtbn has previously been introduced in Section 4. The state in the EKF, which represents the sensor orientation, can be parametrized in different ways. In previous work we have used a quaternion representation as a 4-dimensional state vector [1]. In this work 6 we instead use an implementation of the EKF, which is sometimes called a multiplicative EKF [25, 26, 27]. Here, a 3-dimensional state vector represents the orientation deviation from a linearization point. More details on this implementation can be found in [28]. The EKF returns the one step ahead predicted measurements {b yt|t−1 (θ)}N t=1 and their covariance N {St (θ)}t=1 which can be used to evaluate (6). The cost function needs to be evaluated for the current parameter estimates in Step 2a but also needs to be evaluated once for each component of the parameter vector θ to compute the numerical gradient. Hence, each iteration i requires running the EKF at least nθ + 1 times. Note that the actual number of evaluations can be higher since the backtracking line search algorithm used to determine αi can require a varying number of additional evaluations. Since nθ = 34, computing the numerical gradient is computationally rather expensive. However, it is possible to parallelize the computations. 5.3 The parameter vector θ As apparent from Section 4, our main interest lies in determining the calibration matrix D and the offset vector o, which can be used to correct the magnetometer measurements to obtain more accurate orientation estimates. To solve the calibration problem, however, we also estimate a number of other parameters. First, the local magnetic field mn introduced in Section 4 is in general scenarios unknown and needs to be estimated. In outdoor environments, mn is equal to the local earth magnetic field and is accurately known from geophysical studies, see e.g. [29]. In indoor environments, however, the local magnetic field can differ quite significantly from the local earth magnetic field. Because of that, we treat mn as an unknown constant. Second, the gyroscope measurements that are used to describe the change in orientation of the sensor in (1a) are corrupted by a bias δω . This bias is slowly time varying but for our relatively short experiments it can be assumed to be constant. Hence, it is treated as part of the parameter vector θ. Finally, we treat the noise covariance matrices Σω , Σa and Σm as unknown. In summary, the parameter vector θ consists of D ∈ R3×3 , (17a) 3 o∈R , n 3 m ∈ {R : 3 δω ∈ R , (17b) ||mn ||22 Σω ∈ {R 3×3 : 3×3 : Σm ∈ {R 3×3 : Σa ∈ {R = 1, mnx > 0, mny = 0}, (17c) (17d) Σω  0, Σω = ΣT ω }, T Σa  0, Σa = Σa }, Σm  0, Σm = ΣT m }, (17e) (17f) (17g) where mnx and mny denote the x- and y- component of mn , respectively. The notation Σ  0 denotes the assumption that the matrix Σ is positive semi-definite. Although (17c) and (17e) – (17g) suggest that constrained optimization is needed, it is possible to circumvent this via suitable reparametrizations. The covariance matrices can be parametrized in terms of their Cholesky factorization, leading to only 6 parameters for each 3 × 3 covariance matrix. The local magnetic field can be parametrized using only one parameter as in (8). Note that in our implementation we prefer to use the representation (8b) for the ML problem (6). Although this latter parametrization does not account for the constraint mnx > 0, this is of no concern due to proper initialization. The procedure to obtain good initial estimates of all parameters is the topic of the next section. 6 Finding good initial estimates Since the optimization problem is non-convex, the parameter vector θ introduced in Section 5 needs proper initialization. An initial estimate θb0 is obtained using a three-step method. As a first step, the gyroscope bias δω and the noise covariances of the inertial sensors, Σω , Σa , and of the magnetometer, Σm , are initialized. This is done using a short batch of stationary data. Alternatively, they can be initialized based on prior sensor knowledge. As a second step, described in Section 6.1, an ellipsoid fitting problem is solved using the magnetometer data. This maps the ellipsoid of data to a sphere but can not determine the rotation of the sphere. The rotation of the sphere is determined in a third step of the initialization procedure. This step also determines an initial estimate of the normalized local magnetic field mn . 7 6.1 Ellipsoid fitting Using the definition of the normalized local magnetic field mn , we would expect all calibrated magnetometer measurements to lie on the unit sphere, kmn k22 − 1 = kRtbn mn k22 − 1 = kD−1 (ym,t − o − em,t ) k22 − 1 = 0. (18) In practice, the measurements are corrupted by noise and the equality (18) does not hold exactly. The ellipsoid fitting problem can therefore be written as T ym,t Aym,t + bT ym,t + c ≈ 0, (19) with A , D−T D−1 , T b , −2o D T c,o D −T −T D (20a) D −1 −1 , (20b) o. (20c) Assuming that the matrix A is positive definite, this can be recognized as the definition of an ellipsoid with parameters A, b and c (see e.g. [9]). We can rewrite (19) as a linear relation of the parameters as M ξ ≈ 0, (21) with  ym,1 ⊗ ym,1 ym,2 ⊗ ym,2 .. .   M =  ym,N ⊗ ym,N ym,1 ym,2 .. .  1 1  ..  , . ym,N 1   vec A ξ =  b , c (22) where ⊗ denotes the Kronecker product and vec denotes the vectorization operator. This problem has infinitely many solutions and without constraining the length of the vector ξ, the trivial solution ξ = 0 would be obtained. A possible approach to solve the ellipsoid fitting problem is to make use of a singular value decomposition [9, 2]. This approach inherently poses a length constraint on the vector ξ, assuming that its norm is equal to 1. It does, however, not guarantee positive definiteness of the matrix A. Although positive definiteness of A is not guaranteed, there are only very few practical scenarios in which the estimated matrix A will not be positive definite. A non-positive definite matrix A can for instance be obtained in cases of very limited rotation of the sensor. The problem of allowing a non-positive definite matrix A can be circumvented by instead solving the ellipsoid fitting problem as a semidefinite program [30, 31]   vec A min 21 kM  b  k22 , A,b,c (23) c s.t. Tr A = 1, 3×3 A ∈ S++ , 3×3 where S++ denotes the set of 3 × 3 positive definite symmetric matrices. By constraining the trace of the matrix A, (23) avoids the trivial solution of ξ = 0. The problem (23) is a convex optimization problem and therefore has a globally optimal solution and does not require an accurate initial guess of the parameter vector ξ. The optimization problem can easily be formulated and efficiently solved using freely available software packages like YALMIP [32] or CVX [33]. Initial estimates of the calibration matrix D and the offset vector o can be obtained from the estimated b bb, b A, c as β=  1 bT b−1b b− 4b A e 0T D e0 = βA b−1 , D b−1bb, ob0 = 1 A 2 8 −1 b c , (24a) (24b) (24c) where ob0 denotes the initial estimate of the offset vector o. From (24b) it is not possible to uniquely determine the initial estimate of the calibration matrix D. We determine an initial estimate of the e 0 . However, any calibration matrix D using a Cholesky decomposition, leading to a lower triangular D T e D0 U where U U = I3 will also fulfill (24b). As discussed in Assumption 1 in Section 4, we assume that the sensor axes of the inertial sensors and the magnetometers are related by a rotation, implying that b 0 can therefore be defined in we restrict the matrix U to be a rotation matrix. The initial estimate D e terms of D0 as b0 = D e 0 RD . D (25) The unknown rotation matrix RD will be determined in Section 6.2. 6.2 Determine misalignment of the inertial and magnetometer sensor axes The third step of the initial estimation aims at determining the misalignment between the inertial and the magnetometer sensor axes. It also determines an initial estimate of the normalized local magnetic field m b n0 . These estimates are obtained by combining the magnetometer measurements with the inertial sensor measurements. The approach is based on the fact that the inner product of two vectors is invariant T under rotation. The two vectors considered here are mn and the vertical v n = 0 0 1 . Hence, it is assumed that the inner product of the vertical vtb in the body frame b, vtb = Rtbn v n , (26a) and the normalized local magnetic field mbt in the body frame, T e −1 mbt = RD D0 (ym,t − ob0 ) , (26b) is constant. The matrix RD in (26b) denotes the rotation needed to align the inertial and magnetometer sensor axes. The rotation matrix Rtnb in (26a) is a rotation matrix representation of the orientation estimate at time t obtained from an EKF. This EKF is similar to the one described in Section 5.2. It does not use the magnetometer measurements, since they have not properly been calibrated yet and can therefore not result in accurate heading estimates. However, to determine the vertical vtb , only the sensor’s inclination is of concern, which can be determined using the inertial measurements only. The inner product between mn and v n is equal to mnz (see also (8a)). Since this inner product is invariant under rotation, we can formulate the following minimization problem 1 2 minn RD ,mz,0 N X t=1 T T e −1 kmnz,0 − (v n ) Rtnb RD D0 (ym,t − ob0 ) k22 , RD ∈ SO(3). s.t. (27) The rotation matrix RD can be parametrized using an orientation deviation from a linearization point similar to the approach described in Section 5.2. Hence, (27) can be solved as an unconstrained optimization problem. Based on these results and (25) we obtain the following initial estimates b0 = D e 0R bD , D q 2 m b n0 = 1− m b nz,0 (28a) 0 m b nz,0 T . (28b) Hence, we have obtained an initial estimate θb0 of the entire parameter vector θ as introduced in Section 5. 7 7.1 Experimental results Experimental setup Experiments have been performed using two commercially available inertial measurements units (IMUs), an Xsens MTi-100 [34] and a Trivisio Colibri Wireless IMU [35]. The experimental setup of both experiments can be found in Figure 2. The experiment with the Xsens IMU was performed outdoors 9 Figure 2: Top: experimental setup where a calibration experiment is performed outdoors. An Xsens MTi100 IMU (orange box) together with a magnetic disturbance is placed in an aluminum block. Bottom: experimental setup using a Trivisio Colibri Wireless IMU (black box). A phone is used as a source of magnetic disturbance. To avoid saturation of the magnetometer, the phone is not attached directly to the IMU. to ensure a homogeneous local magnetic field. The experiment with the Trivisio IMU was performed indoors. However, the experiment was performed relatively far away from any magnetic materials such that the local magnetic field is as homogenous as possible. The Xsens IMU was placed in an aluminum block with right angles which can be used to rotate the sensor 90◦ to verify the heading results. For both sensors, inertial and magnetometer measurements were collected at 100 Hz. 7.2 Calibration results For calibration, the IMU needs to be slowly rotated such that the assumption of zero acceleration is reasonably valid. This leads to an ellipsoid of magnetometer data as depicted in red in Figs. 1 and 3. Note that for plotting purposes the data has been downsampled to 1 Hz. To emphasize the deviation of the norm from 1, the norm of the magnetometer data is depicted in red in Figure 4 for both experiments. b and offset vector ob are For the experiment with the Xsens IMU, the following calibration matrix D found     0.74 −0.13 0.01 1.36 b = −0.12 0.68 0.01 , ob =  1.22  D (29) −0.03 0.43 1.00 −0.94 using Algorithm 1. Applying the calibration result to the magnetometer data leads to the unit sphere of data in blue in Figure 1. The norm of the magnetometer data after calibration can indeed be seen to lie around 1, as depicted in blue in Figure 4. −1/2 As a measure of the calibration quality, we analyze the normalized residuals St (yt − ybt|t−1 ) after 6 calibration from the EKF. For each time t, this is a vector in R . In the case of correctly calibrated parameters that sufficiently model the magnetic disturbances, we expect the stacked normalized residuals −1/2 6N {St (yt − ybt|t−1 )}N to be normally distributed with zero mean and standard deviation 1. The t=1 ∈ R histogram and a fitted Gaussian distribution can be found in Figure 5a. The residuals resemble a N (0, 1) distribution except for the large peak around zero and – not visible in the plot – a small amount of outliers outside of the plotting interval. This small amount of outliers is due to the fact that there are a few 10 1 0 −1 −1 0 1 Norm magnetic field Trivisio IMU Norm magnetic field Xsens IMU Figure 3: Calibration results from the experiment with the Trivisio IMU. The ellipsoid of magnetometer data (red) lies on a unit sphere after calibration (blue). 2.5 2 Before calibration After calibration 1.5 1 0 20 40 60 80 100 Time [s] 120 140 Before calibration After calibration 1.4 1.2 1 0.8 0 20 40 60 80 Time [s] 100 120 Figure 4: Norm of the magnetic field measurements before (red) and after (blue) calibration for (top) the experiment with the Xsens IMU and for (bottom) the experiment with the Trivisio IMU. 11 −4 −2 0 2 4 −4 (a) Xsens IMU, estimation data −2 0 2 4 (b) Xsens IMU, validation data −1/2 Figure 5: Histogram of the normalized residuals St (yt − ybt|t−1 ) from the EKF after calibration for the estimation data set (left) and for a validation data set (right) for the experiments performed with the Xsens IMU. A Gaussian distribution (red) is fitted to the data. measurement outliers in the accelerometer data. Large accelerations can for instance be measured when the setup is accidentally bumped into something and violate our assumption that the acceleration of the sensor is approximately zero. We believe that the peak around zero is due to the fact that the algorithm compensates for the presence of the large residuals. To analyze if the calibration is also valid for a different (validation) data set with the same experimental setup, the calibrated parameters have been used on a second data set. Figures of the ellipsoid of magnetometer data and the sphere of calibrated magnetometer data are not included since they look very similar to Figs. 1 and 4. The residuals after calibration of this validation data set can be found in Figure 5b. The fact that these residuals look very similar to the ones for the original data suggests that the calibration parameters obtained are also valid for this validation data set. The Trivisio IMU outputs the magnetometer data in microtesla. Since our algorithm scales the b and offset vector ob from Algorithm 1 are in calibrated measurements to a unit norm, the obtained D this case of much larger magnitude,     −19.77 61.74 0.59 0.09 b = −1.01 60.74 0.23  , ob =  −1.68  . (30) D −6.98 −0.39 0.06 60.80 The sphere of calibrated data and its norm can be found in blue in Figs. 3 and 4. Note that for plotting purposes, the magnetometer data before calibration is scaled such that its mean lies around b and ob are scaled accordingly to plot the red ellipsoid in Figure 3. The normalized 1. The obtained D −1/2 residuals St (yt − ybt|t−1 ) of the EKF using both the estimation and a validation data set are depicted in Figure 6. For this data set, the accelerometer data does not contain any outliers and the residuals resemble a N (0, 1) distribution fairly well. From these results we can conclude that Algorithm 1 gives good magnetometer calibration results for experimental data from two different commercially available IMUs. A good fit of the ellipsoid of data to a sphere is obtained and the algorithm seems to give good estimates analyzed in terms of its normalized residuals. Since magnetometer calibration is generally done to obtain improved heading estimates, it is important to also interpret the quality of the calibration in terms of the resulting heading estimates. In Section 7.3 this will be done based on experimental results. The heading performance will also be analyzed based on simulations in Section 8. 7.3 Heading estimation An important goal of magnetometer calibration is to facilitate good heading estimates. To check the quality of the heading estimates after calibration, the block in which the Xsens IMU was placed (shown in Figure 2) is rotated around all axes. This block has right angles and it can therefore be placed in 24 orientations that differ from each other by 90 degrees. The experiment was conducted in Enschede, the Netherlands. The dip angle δ at this location is approximately 67◦ [29]. Hence, we expect the calibrated magnetometer measurements to resemble rotations of the normalized magnetic field mn = 12 −4 −2 0 2 4 −4 (a) Trivisio IMU, estimation data −2 0 2 4 (b) Trivisio IMU, validation data −1/2 Figure 6: Histogram of the normalized residuals St (yt − ybt|t−1 ) from the EKF after calibration for the estimation data set (left) and for a validation data set (right) for the experiments performed with the Trivisio IMU. A Gaussian distribution (red) is fitted to the data. T 0.39 0 −0.92 (see also (7) and (8b)). The calibrated magnetometer data from the experiment is shown in Figure 7 and consists of the following stationary time periods: z-axis up During the period 0−105s, the magnetometer is flat with its z−axis pointing upwards. Hence, the z-axis (red) of the magnetometer measures the vertical component of the local magnetic field mnz . During this period, the sensor is rotated by 90◦ around the z-axis into 4 different orientations and subsequently back to its initial orientation. This results in the 5 steps for measurements in the x- (blue) and y-axis (green) of the magnetometer. z-axis down A similar rotation sequence is performed with the block upside down at 110 − 195s, resulting in a similar pattern for measurements in the x- and y-axis of the magnetometer. During this time period, the z-axis of the magnetometer measures −mnz instead. x-axis up The procedure is repeated with the x-axis of the sensor pointing upwards during the period 200 − 255s, rotating around the x-axis into 4 different orientations and back to the initial position. This results in the 5 steps for measurements in the y- and z-axis of the magnetometer. x-axis down A similar rotation sequence is performed with the x-axis pointing downwards at 265 − 325 seconds. y-axis down Placing the sensor with the y-axis downwards and rotating around the y-axis results in the data at 350 − 430 seconds. The rotation results in the 5 steps for measurements in the x- and z-axis of the magnetometer. y-axis up A similar rotation sequence is performed with the y-axis pointing upwards at 460 − 520 seconds. Since the experimental setup was not placed exactly vertical, it is not possible to compare the absolute orientations. However, it is possible to compare the difference in orientation which is known to be 90◦ due to the properties of the block in which the sensor was placed. To exclude the effect of measurement noise, for each of the stationary periods in Figure 7, 500 samples of magnetometer and accelerometer data are selected. Their mean values are used to estimate the orientation of the sensor. Here, the accelerometer data is used to estimate the inclination. The heading is estimated from the horizontal component of the magnetometer data. This procedure makes use of the fact that the orientation of the sensor can be determined from two linearly independent vectors in the navigation frame – the gravity and the direction of the magnetic north – and in the body frame – the mean accelerometer and magnetometer data. It is referred to as the TRIAD algorithm [36]. Table 1 reports the deviation from 90◦ between two subsequent rotations. Note that the metal object causing the magnetic disturbance as shown in Figure 2 physically prevents the setup from being properly placed in all orientations around the y-axis. Rotation around the y-axis with the y-axis pointing upwards has therefore not been included in Table 1. Our experiment investigates both the heading errors and the improvement of the heading estimates over the ones obtained after the initial calibration, i.e. Step 1 in Algorithm 1. In Table 1 we therefore 13 Calibrated magnetometer measurements 1 0.5 0 −0.5 −1 0 100 200 300 Time [s] 400 500 Figure 7: Calibrated magnetometer data of an experiment rotating the sensor into 24 different sensor orientations where the blue, green and red lines represent the data from the x-, y- and z-axis of the magnetometer, respectively. Table 1: Difference in estimated heading between two subsequent rotations around the sensor axes using calibrated magnetometer data. The values represent the deviation in degrees from 90◦ . Included are both the results using the ML estimates from Algorithm 1 and the results using initial estimates from Step 1 in the algorithm. z-axis z up x-axis z down x up y-axis x down y down ML init ML init ML init ML init ML init 0.11 0.22 0.46 0.30 0.36 0.90 1.52 0.94 0.69 2.48 1.53 1.92 1.34 4.36 3.57 2.40 0.22 0.07 0.97 0.29 0.16 0.20 0.94 0.59 0.86 1.57 0.61 1.78 1.01 1.45 0.71 1.70 0.18 0.29 0.20 0.50 1.57 0.76 0.78 0.45 b 0 (28a) and ob0 (24c) and the include both the heading errors using the initial parameter estimates D b and ob (29) obtained using Algorithm 1. As can be seen, heading errors using ML parameter estimates D the deviation from 90◦ is small, indicating that good heading estimates are obtained after calibration. Also, the heading estimates using the initial parameter estimates are already fairly good. The mean error is reduced from 1.28◦ for the initial estimate to 0.76◦ for the ML estimate. The maximum error is reduced from 4.36◦ for the initial estimate to 2.48◦ for the ML estimate. Note that the results of the ML estimate from Algorithm 1 are slightly better than the results previously reported by [1]. This can be attributed to the fact that we now use orientation error states instead of the quaternion states in the EKF (see Section 5.2). This results in slightly better estimates, but also in a smoother convergence of the optimization problem. The quality of the heading estimates is studied further in Section 8 based on a simulation study. 8 Simulated heading accuracy Magnetometer calibration is typically performed to improve the heading estimates. It is, however, difficult to check the heading accuracy experimentally. In Section 7.3, for instance, we are limited to doing the heading validation on a different data set and we have a limited number of available data points. To get more insight into the orientation accuracy that is gained by executing all of Algorithm 1, compared to just its initialization phase (Step 1 in the algorithm), we engage in a simulation study. In this study we focus on the root mean square (RMS) heading error for different simulated sensor qualities (in terms of the noise covariances and the gyroscope bias) and different magnetic field disturbances (in terms of different values for the calibration matrix D and offset vector o). In our simulation study, we assume that the local magnetic field is equal to that in Linköping, Sweden. The calibration matrix D, the offset vector o and the sensor properties in terms of the gyroscope bias 14 Table 2: Settings used in the Monte Carlo simulation. Ddiag Dskew Drot o D11 , D22 , D33 ∼ U(0.5, 1.5) ζ, η, ρ ∼ U(−30◦ , 30◦ ) ψ, θ, φ ∼ U(−10◦ , 10◦ ) o1 , o2 , o3 ∼ U(−1, 1) δω Σω Σa Σm δω,1 , δω,2 , δω,3 ∼ U(−1, 1) Σω,1 , Σω,2 , Σω,3 ∼ U(10−3 , 10−2 ) Σa,1 , Σa,2 , Σa,3 ∼ U(10−3 , 10−1 ) Σm,1 , Σm,2 , Σm,3 ∼ U(10−3 , 10−1 ) and noise covariances are all sampled from a uniform distribution. The parameters of the distributions from which the sensor properties are sampled are chosen as physically reasonable values as considered from the authors’ experience. The noise covariance matrices Σω , Σa and Σm are assumed to be diagonal with three different values on the diagonal. The calibration matrix D is assumed to consist of three parts, D = Ddiag Dskew Drot , (31) where Ddiag is a diagonal matrix with elements D11 , D22 , D33 and Drot is a rotation matrix around the angles ψ, θ, φ. The matrix Dskew models the non-orthogonality of the magnetometer axes as   1 0 0 , cos ζ 0 Dskew =  sin ζ (32) − sin η cos η sin ρ cos η cos ρ where the angles ζ, η, ρ represent the different non-orthogonality angles. The exact simulation conditions are summarized in Table 2. The simulated data consists of 100 samples of stationary data and subsequently 300 samples for rotation around all three axes. It is assumed that the rotation is exactly around the origin of the accelerometer triad, resulting in zero acceleration during the rotation. The first 100 samples are used to obtain an initial estimate of the gyroscope bias δbω,0 by computing the mean of the stationary gyroscope b ω,0 , Σ b a,0 and Σ b m,0 are initialized based on the covariance of these samples. The covariance matrices Σ b ω,0 , Σ b a,0 , Σ b m,0 and first 100 samples. The initial estimate then consists of these initial estimates δbω,0 , Σ b 0 (28a), the initial offset vector ob0 (24c) and the initial estimate of the the initial calibration matrix D local magnetic field mn0 (28b). To study the heading accuracy, the EKF as described in Section 5.2 is run with both the initial parameter values θb0 and their ML values θbML . The orientation errors ∆qt , encoded as a unit quaternion are computed using c nb ∆qt = qbtnb qref,t , (33) where denotes a quaternion multiplication and the superscript c denotes the quaternion conjugate (see e.g. [27]). It is computed from the orientation qbtnb estimated by the EKF and the ground truth nb orientation qref,t . Computing the orientation errors in this way is equivalent to subtracting Euler angles in the case of small angles. However, it avoids subtraction problems due to ambiguities in the Euler angles representation. To interpret the orientation errors ∆qt , they are converted to Euler angles. We focus our analysis on the heading error, i.e. on the third component of the Euler angles. The RMS of the heading error is plotted for 150 Monte Carlo simulations in Figure 8. As can be seen, the heading root mean square error (RMSE) using the estimate of the calibration parameters from Algorithm 1 is consistently small. The heading RMSE based on the initialization phase in Step 1 of the algorithm, however, has a significantly larger spread. This clearly shows that orientation accuracy can be gained by executing all of Algorithm 1. Note that in all simulations, analysis of the norm of the calibrated magnetometer measurements as done in Figure 4 does not indicate that the ML estimate is to be preferred over the estimate from the initialization phase. Hence, analysis of the norm of the calibrated magnetometer measurements does not seem to be a sufficient analysis to determine the quality of the calibration in the case when the calibration is performed to improve the heading estimates. 15 40 10 30 20 5 10 0 0 1 2 Heading error [deg] (a) Initial parameter estimate 20 40 Heading error [deg] (b) ML parameter estimate Figure 8: Histogram of the heading RMSE using the ML parameter estimate from Algorithm 1 (left, blue) and the initial parameter estimate from Step 1 in the algorithm (right, red). Note the different scales in the two plots. 9 Conclusions We have developed a practical algorithm to calibrate a magnetometer using inertial sensors. It calibrates the magnetometer for the presence of magnetic disturbances, for magnetometer sensor errors and for misalignment between the inertial and magnetometer sensor axes. The problem is formulated as an ML problem. The algorithm is shown to perform well on real data collected with two different commercially available inertial measurement units. In future work the approach can be extended to include GPS measurements. In that case it is not necessary to assume that the acceleration is zero. The algorithm can hence be applied to a wider range of problems, like for instance the flight test example discussed in [2]. The computational cost of the algorithm would, however, increase, since to facilitate the inclusion of the GPS measurements, the state vector in the EKF needs to be extended. Another interesting direction for future work would be to investigate ways of reducing the computational cost of the algorithm. The computational cost of the initialization steps is very small but actually solving the ML problem in Step 2 of Algorithm 1 is computationally expensive. The algorithm both needs quite a large number of iterations and each iteration is fairly expensive due to the computation of the numerical gradients. Interesting lines of future work would either explore different optimization methods or different ways to obtain gradient estimates. Finally, it would be interesting to extend the work to online estimation of calibration parameters. This would allow for a slowly time-varying magnetic field and online processing of the data. 10 Acknowledgements This work is supported by CADICS, a Linnaeus Center, and by the project Probabilistic modeling of dynamical systems (Contract number: 621-2013-5524), both funded by the Swedish Research Council (VR), and by MC Impulse, a European Commission, FP7 research project. The authors would like to thank Laurens Slot, Dr. Henk Luinge and Dr. Jeroen Hol from Xsens Technologies and Dr. Gustaf Hendeby from Linköping University for their support in collecting the data sets and for interesting discussions. The authors would also like to thank the reviewers for their constructive comments. 16 References [1] M. Kok and T. B. Schön. Maximum likelihood calibration of a magnetometer using inertial sensors. In Proceedings of the 19th World Congress of the International Federation of Automatic Control, pages 92–97, Cape Town, South Africa, August 2014. [2] M. Kok, J. D. Hol, T. B. Schön, F. Gustafsson, and H. Luinge. Calibration of a magnetometer in combination with inertial sensors. In Proceedings of the 15th International Conference on Information Fusion, pages 787 – 793, Singapore, July 2012. [3] N. Bowditch. The American Practical Navigator. United States government, 2002. [4] Roberto Alonso and Malcolm D. Shuster. Complete linear attitude-independent magnetometer calibration. The Journal of the Astronautical Sciences, 50(4):477–490, 2002. [5] G. M. Lerner. Scalar checking. In James R. Wertz, editor, Spacecraft attitude determination and control, pages 328–334. D. Reidel Publishing Company, Dordrecht, Holland, 1978. [6] J. F. Vasconcelos, G. Elkaim, C. Silvestre, P. Oliveira, and B. Cardeira. Geometric approach to strapdown magnetometer calibration in sensor frame. IEEE Transactions on Aerospace and Electronic Systems, 47(2):1293–1306, April 2011. [7] V. Renaudin, M. H. Afzal, and G. Lachapelle. Complete triaxis magnetometer calibration in the magnetic domain. Journal of Sensors, 2010. [8] D. Gebre-Egziabher, G. H. Elkaim, J. D. Powell, and B. W. Parkinson. Calibration of strapdown magnetometers in magnetic field domain. Journal of Aerospace Engineering, 19(2):87–102, April 2006. [9] W. Gander, G. H. Golub, and R. Strebel. Least-squares fitting of circles and ellipses. BIT Numerical Mathematics, 34(4):558–578, 1994. [10] I. Markovsky, A. Kukush, and S. Van Huffel. Consistent least squares fitting of ellipsoids. Numerische Mathematik, 98(1):177–194, 2004. [11] Y. Wu and W. Shi. On calibration of three-axis magnetometer. IEEE Sensors Journal, 15(11), November 2015. [12] X. Li and Z. Li. A new calibration method for tri-axial field sensors in strap-down navigation systems. Measurement Science and Technology, 23(10), October 2012. [13] S. Salehi, N. Mostofi, and G. Bleser. A practical in-field magnetometer calibration method for IMUs. In Proceedings of the IROS Workshop on Cognitive Assistive Systems: Closing the Action-Perception Loop, pages 39–44, Vila Moura, Portugal, October 2012. [14] S. Bonnet, C. Bassompierre, C. Godin, S. Lesecq, and A. Barraud. Calibration methods for inertial and magnetic sensors. Sensors and Actuators A: Physical, 156(2):302–311, 2009. [15] G. Troni and L. L. Whitcomb. Adaptive estimation of measurement bias in three-dimensional field sensors with angular-rate sensors: Theory and comparative experimental evaluation. In Proceedings of Robotics: Science and Systems (RSS), Berlin, Germany, June 2013. [16] L. Ljung. System Identification, Theory for the User. Prentice Hall PTR, 2nd edition, 1999. [17] T. B. Schön, A. Wills, and B. Ninness. System identification of nonlinear state-space models. Automatica, 47(1):39–49, 2011. [18] F. Lindsten and T. B. Schön. Backward simulation methods for Monte Carlo statistical inference. Foundations and Trends in Machine Learning, 6(1):1–143, 2013. [19] F. Gustafsson. Statistical Sensor Fusion. Studentlitteratur, 2012. [20] K. J. Åström. Maximum likelihood and prediction error methods. Automatica, 16(5):551–574, 1980. 17 [21] M. Segal and E. Weinstein. A new method for evaluating the log-likelihood gradient, the Hessian, and the Fisher information matrix for linear dynamic systems. IEEE Transactions on Information Theory, 35(3):682–687, 1989. [22] M. Kok, J. Dahlin, T. B. Schön, and A. Wills. Newton-based maximum likelihood estimation in nonlinear state space models. In Proceedings of the 17th IFAC Symposium on System Identification, pages 969 – 974, Beijing, China, October 2015. [23] J. Kokkala, A. Solin, and S. Särkkä. Sigma-point filtering and smoothing based parameter estimation in nonlinear dynamic systems. Pre-print, 2015. arXiv:1504.06173. [24] J. Nocedal and S. J. Wright. Numerical Optimization. Springer Series in Operations Research, 2nd edition, 2006. [25] F. L. Markley. Attitude error representations for Kalman filtering. Journal of guidance, control, and dynamics, 26(2):311–317, 2003. [26] J. L. Crassidis, F. Landis Markley, and Y. Cheng. A survey of nonlinear attitude estimation methods. Journal of Guidance, Control, and Dynamics, 30(1):12–28, 2007. [27] J. D. Hol. Sensor Fusion and Calibration of Inertial Sensors, Vision, Ultra-Wideband and GPS. Dissertation no. 1368, Linköping University, Linköping, Sweden, June 2011. [28] M. Kok. Probabilistic modeling for positioning applications using inertial sensors. Licentiate’s thesis no. 1656, Linköping University, Linköping, Sweden, June 2014. [29] National Centers for Environmental Information. https://www.ngdc.noaa.gov/geomag/geomag. shtml, Accessed on October 2, 2015. [30] G. Calafiore. Approximation of n-dimensional data using spherical and ellipsoidal primitives. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 32(2):269–278, 2002. [31] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004. [32] J. Löfberg. YALMIP: A toolbox for modeling and optimization in MATLAB. In Proceedings of the IEEE International Symposium on Computer Aided Control Systems Design, 2004, pages 284 – 289, Taipei, Taiwan, September 2004. [33] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version 2.0 beta. http://cvxr.com/cvx, September 2013. [34] Xsens Technologies B. V. http://www.xsens.com, Accessed on January 7, 2016. [35] Trivisio Prototyping GmbH. http://www.trivisio.com, Accessed on January 7, 2016. [36] M. D. Shuster and S. D. Oh. Three-axis attitude determination from vector observations. Journal of Guidance, Control, and Dynamics, 4(1):70–77, 1981. 18
3cs.SY
Generalized Binary Search For Split-Neighborly Problems Stephen Mussmann Stanford University arXiv:1802.09751v1 [cs.AI] 27 Feb 2018 Abstract In sequential hypothesis testing, Generalized Binary Search (GBS) greedily chooses the test with the highest information gain at each step. It is known that GBS obtains the gold standard query cost of O(log n) for problems satisfying the k-neighborly condition, which requires any two tests to be connected by a sequence of tests where neighboring tests disagree on at most k hypotheses. In this paper, we introduce a weaker condition, splitneighborly, which requires that for the set of hypotheses two neighbors disagree on, any subset is splittable by some test. For four problems that are not k-neighborly for any constant k, we prove that they are splitneighborly, which allows us to obtain the optimal O(log n) worst-case query cost. 1 Introduction Sequential hypothesis testing (Young & Young, 1998) aims to find the true hypothesis from a set of hypotheses by performing tests. Some examples are gathering observations to deduce the location of a hidden object or labeling data points to infer an underlying classifier. One commonly used algorithm is Generalized Binary Search (GBS), also known as the splitting algorithm, which greedily chooses the test that most evenly splits the hypothesis version space (Garey & Graham, 1974; Nowak, 2008), or equivalently greedily chooses the test with the maximal information gain (for binary tests). Greedy information gain is surprisingly effective in practice and has become the gold standard with a variety of applications, approximations, and extensions (Settles, 2012; Chu & Ghahramani, 2005; Bellala et al., 2010; Karbasi et al., 2012; Zheng et al., 2012; Proceedings of the 21st International Conference on Artificial Intelligence and Statistics (AISTATS) 2018, Lanzarote, Spain. PMLR: Volume 84. Copyright 2018 by the author(s). Percy Liang Stanford University Jedynak et al., 2012; Luo et al., 2013; Maji et al., 2014; Sun et al., 2015). We seek to explain this performance by providing a condition under which GBS attains a query cost of O(log n), the information-theoretic optimal query cost. While there has been much work on establishing that GBS attains an average cost within a log n factor of the optimal algorithm (Guillory & Bilmes, 2009; Kosaraju et al., 1999; Dasgupta, 2004; Chakaravarthy et al., 2007; Adler & Heeringa, 2008; Chakaravarthy et al., 2009; Gupta et al., 2010), establishing the asymptotically optimal query cost is a difficult and understudied problem. The few previous works have required somewhat stringent conditions. One such condition is “sample-rich” (Naghshvar et al., 2012), which states that every subset of the hypotheses has a test that returns true on exactly those hypotheses; this requires an exponential number of tests. Another line of work (Nowak, 2009, 2011) introduced the more lenient kneighborly condition, which requires that every two tests be connected by a sequence of tests where neighboring tests disagree on at most k hypotheses. As we will show in this paper, many problems of a discrete nature do not satisfy this condition. In this paper, building on the k-neighborly condition, we introduce a new, weaker condition called 1/α-splitneighborly, which requires that if neighboring tests disagree on a set of hypotheses V , then there exists a test that splits off an α fraction of V (note that |V | could be quite large, whereas k-neighborly requires |V | ≤ k). We prove that four natural problems satisfy the 1/αsplit-neighborly condition for a constant α: pool-based linear classifiers, learning monotonic CNF formulas, discrete object localization, and discrete binary classification. Furthermore, we prove that the √ value of k in the k-neighborly analysis is at least n/2 for all four problems, which yields nearly vacuous bounds. In summary, by using 1/α-split-neighborly, we show that Generalized Binary Search achieves an asymptotically optimal query cost of O(log n) in settings where the previous k-neighborly analysis tools fail. Generalized Binary Search For Split-Neighborly Problems 1 point 10 hypotheses Algorithm 1 Active querying algorithm template Input: H, X , oracle access to h∗ , method F V =H for t = 0, 1, ... do xt ← F ({(xi , yi )}t−1 i=1 ) Query xt and obtain yt = h∗ (xt ) Update V ← {h ∈ V : h(xt ) = yt } if |V | = 1 then return h ∈ V end if end for 2 points 4 hypotheses + + + 4 points 1 hypothesis - 3 points 2 hypotheses + - + - Algorithm 2 Generalized Binary Search (F ) t−1 Input: H, X , Previous test results {(xi , yi )}i=1 V = {h ∈ H : h(xi ) = yi , 1 ≤ i ≤ t − 1} xt ← argminx∈X |Eh∈V [h(x)] − 1/2| return xt 1.1 Notation In all cases, we use log to denote log2 . For a set P s∈S f (s) S and function f , we define Es∈S [f (s)] = . |S| Similarly, forP a condition C, we define 1[C(s)] . Prs∈S [C(s)] = s∈S|S| 2 Problem statement + + Figure 1: Problem of identifying a linear classifier in a pool-based active learning setting. Each linear classifier is represented by a red line with an arrow pointing towards the positive class. Each round, we select a new test (blue point), after four rounds, we have identified the true classifier. Though we focus on the noiseless and well-specified setting, both conditions can be relaxed: Kääriäinen (2006) reduces the non-persistent noisy setting (we can repeatedly query any test) to the noiseless setting, and Nowak (2011) adapts the GBS algorithm to the misspecified setting (see Section 5 for more details). Consider a set of n hypotheses H and tests X , where each h ∈ H is a mapping from X to {0, 1}. We assume that the hypotheses H are identifiable, meaning that any two hypotheses yield different results on at least one test. Assume there is a fixed but unknown hypothesis h∗ ∈ H that we wish to identify. An active querying algorithm performs a sequence of tests; on each iteration, it uses a method F to select a test xt based on the results of previous tests and receives yt = h∗ (xt ) (see Algorithm 1). We evaluate F based on the worst-case number of queries. Generalized Binary Search (also known as the “splitting algorithm” and “maximal shrinkage”) is a wellstudied method (Garey & Graham, 1974; Nowak, 2011; Dasgupta, 2004). GBS maintains the set V ⊆ H of hypotheses consistent with test results thus far, and at each step, it chooses a test that splits the elements of V as evenly as possible. See Algorithm 2 for the pseudocode. The optimal worst-case number of queries is Ω(log n) so if GBS attains O(log n) for a problem, it is asymptotically optimal. As an illustrative example, let the set of hypotheses H be linear classifiers separating 5 data points at the vertices of a regular pentagon. (see Figure 1). In this case, |H| = 20 (only including hypotheses with both + and − labels). The tests X are data points, and the test output is an indicator for h∗ classifying that point as +. Figure 1 shows the queries generated by GBS. We see that the size of the version space |V | decreases exponentially, the hallmark of a O(log n) worst-case query cost. Later, we will prove that GBS indeed attains O(log n) worst-case query cost for this this problem class of linear classifiers on the vertices of convex polygons. 3 3.1 General analysis Splits Intuitively, GBS works well when it can find tests that split the hypothesis space into roughly equal parts. The split induced by test x is a partition of the hypotheses into {h ∈ H : h(x) = 0} and {h ∈ H : h(x) = 1}. Define the split constant of a test x for a set of hypotheses V as miny∈{0,1} Prh∈V [h(x) = y], the fraction of hypotheses in the smaller partition. Note that large split constants are preferred, and 1/2 is the maximum split constant. As we will see, both the k- Stephen Mussmann, Percy Liang both return 0 and 1 by choosing tests randomly.1 Graph of Tests Δ(𝑥′, 𝑥) Definition 3.2 (Coherence parameter). The coherence parameter is the largest c such that 𝑥 𝑥′ Δ(𝑥, 𝑥 ′ ) ∀h ∈ H : Ex∼P [h(x)] ∈ [c, 1 − c] for some probability distribution P over tests. This is a concept that will be used with our condition, 1/α-split-neighborly, as well. From these two definitions, Nowak (2011) showed the following result: Hypotheses Figure 2: A test graph on the top and the action of two neighboring tests x, x0 on the hypotheses on the bottom. The hypotheses h are represented by rectangles and the tests by circles that contain the hypotheses for which the test returns 1. For a kneighborly edge to exist between two tests, the corresponding ∆(x, x0 ) and ∆(x0 , x) must have cardinality |∆(x, x0 ) ∪ ∆(x0 , x)| ≤ k. If the resulting test graph is connected, we say the problem is k-neighborly. neighborly condition (Nowak, 2011) and our new 1/αsplit-neighborly condition imply that for any version space V , there is a test with a large split constant. 3.2 Earlier work: k-neighborly and coherence parameter Tests often have a similarity structure. As an example, for the hypothesis class of linear classifiers, nearby input points (tests) yield the same result for most hypotheses. We therefore construct a similarity graph over tests, which will provide a useful analysis tool that allows us to only reason locally on the graph. Nowak (2011) defines two tests to be similar if they disagree on at most k hypotheses. k-neighborly is the condition that such a similarity graph is connected. Definition 3.1 (k-neighborly). For any two tests, x and x0 , define ∆(x, x0 ) = {h ∈ H : h(x) = 0 ∧ h(x0 ) = 1}. Let the test graph contain undirected edges (x, x0 ) for which |∆(x, x0 )∪∆(x0 , x)| ≤ k. A problem instance is k-neighborly if the test graph is connected. See Figure 2 for an illustration of the k-neighborly condition. Intuitively, the k-neighborly condition ensures that between any two tests, we can find a path where each pair of neighbors in the path are very similar. Nowak (2011) also defines the coherence parameter, which ensures an algorithm can easily find tests that Theorem 3.1 (Nowak, 2011). If a problem has a coherence parameter c and is k-neighborly, then the 1 log(n) queries, where worst-case cost of GBS is − log(λ) 1 λ = 1 − min(c, k+2 ). For large enough c, the k-neighborly analysis yields worst-case query complexity of O(k log(n)). Later, we √ show several examples where k = Ω( n), yielding the k-neighborly analysis very loose. 3.3 Split-neighborly The k-neighborly condition is a rather strong condition since it requires tests that disagree on only k hypotheses. While this sometimes holds for problems with a continuous structure, such as linear classifiers or continuous object localization, it is often not satisfied for problems with a discrete nature. Later, in Section 4, we show a variety √of problems of a discrete nature where k is at least n/2. Motivated by these discrete problems, we will now introduce a weaker condition which we call 1/α-split-neighborly, the main contribution of this paper. In 1/α-split-neighborly, two tests are not only connected if there is a small number of hypotheses on which the tests differ, but also if any subset of the hypotheses that they differ on can be split evenly (with a split constant of at least α) by some test. Definition 3.3 (1/α-split-neighborly). Let α ∈ (0, 12 ] 2 . For any two tests, x and x0 , define ∆(x, x0 ) = {h ∈ H : h(x) = 0 ∧ h(x0 ) = 1}. Define a directed test graph to have a directed edge (x, x0 ) if for any V ⊆ ∆(x, x0 ), |V | ≤ 1 or there exists a test x ∈ X such that Eh∈V [h(x)] ∈ [α, 1 − α] A problem is 1/α-split-neighborly if the test graph is strongly connected. 1 Our definition is a simple linear transformation of the definition in Nowak (2011) to account for notational differences. 2 As a special case, we say a problem is 1-splitneighborly if the graph generated by connecting nodes where |∆(x, x0 )| ≤ 1 is strongly connected. Generalized Binary Search For Split-Neighborly Problems 𝑘-neighborly 𝑥′ 1/𝛼-split-neighborly Constant coherence GBS achieves O(log n) Δ(𝑥, 𝑥 ′ ) 𝑥 Figure 4: Relationship between the different conditions, where arrows represent logical implication. Figure 3: For the problem of identifying linear classifiers in a pool-based active learning setting, an example of two tests that are not connected in the kneighborly graph for small k but are connected in the 1/α-split-neighborly graph for small 1/α. While the size of ∆(x, x0 ) is large, we can still split any subset of ∆(x, x0 ) because of the other points in the pool. 0 to close to 1, in which case there is a good split, or there is a large jump in the split constant between two neighboring tests x and x0 , which implies that |V ∩ ∆(x, x0 )|/|V | is large. Finally, from the definition of 1/α-split-neighborly, we can find a test to have a β split constant of V ∩ ∆(x, x0 ). In summary, the splitneighborly condition and coherence condition allow us to conclude that for any subset of the hypotheses, there is a test with a β split constant. Although it can be more involved to show an edge between tests in the sense of 1/α-split-neighborly rather than k-neighborly, it is a more general condition which makes the similarity graph more connected (see Figure 3 for an example). From this lemma, we get the following theorem. The coherence, k-neighborly, and 1/α-split-neighborly conditions are preserved when we restrict the hypotheses: create a problem with same tests X 0 = X but with H0 ⊆ H. This is because the conditions all are statements involving a universal quantification over the hypotheses, or subsets thereof. Furthermore, and most importantly, constant coherence and the split-neighborly condition imply that GBS has O(log n) query cost. First, we prove the following lemma, showing that constant coherence and 1/α-split-neighborly imply that any subset of H has a test with a good split constant. Lemma 3.1. If a problem is 1/α-split-neighborly and has a coherence parameter of c, then for any V ⊆ H, |V | ≤ 1 or there exists a test x ∈ X such that Eh∈V [h(x)] ∈ [β, 1 − β] where the split constant is   1 β = min c, 1 . /α + 2 Note that for large c and small α, β ≈ α. See appendix for the full proof; we only give a sketch here. Intuitively, the coherence parameter ensures that there is a good split of V or there is both a test that mostly yields 0 and a test that mostly yields 1 (for hypotheses V ). If we examine a path of tests x between the two tests, either Eh∈V [h(x)] varies smoothly from close to Theorem 3.2. If a problem is 1/α-split-neighborly and has a coherence parameter of c, then GBS has a worst log n , where case query cost of at most − log(1−β)   1 . β = min c, 1 /α + 2 From Lemma 3.1, it is clear that after m queries, we have at most n(1 − β)m hypotheses left. Thus, the worst-case query cost (to reach one hypothesis) is log n − log(1−β) . The precise proof is in the appendix. Similarly to the k-neighborly condition, for large enough coherence, the worst-case query cost is O( α1 log n). Thus, for constant α, we get O(log n) worst-case query cost, but for α → 0, we do not. In fact, k-neighborly implies k-split-neighborly (1/αsplit-neighborly, α = 1/k). Thus, our split-neighborly condition is a generalization of k-neighborly, and comparison between our theorems shows our condition is strictly more powerful than the k-neighborly condition. See Figure 4 for a diagram. Proposition 3.1. If a problem is k-neighborly, then it is k-split-neighborly. Proof. In the case that k = 1, |∆(x, x0 )| = 1 so |V | ≤ 1 so the problem is 1-split-neighborly. Note that any set of hypotheses must have a test that distinguishes at least one of the hypotheses (otherwise the hypotheses are the same). If two points x and x0 in the k-neighborly graph have an edge between them, then |∆(x, x0 ) ∪ ∆(x0 , x)| ≤ k, which implies |V | ≤ |∆(x, x0 )| ≤ k, and thus either |V | ≤ 1 or there Stephen Mussmann, Percy Liang is a test with a 1/k split constant and thus there is an edge from x to x0 in the k-split-neighborly graph. Similarly, there is an edge from x0 to x, the k-splitneighborly graph is strongly connected, and the problem is k-split-neighborly. 4 Δ(𝑥, 𝑥 ′ ) All of the proofs have a similar structure for proving 1/α-split-neighborly. First, fix a subset of hypotheses V ⊆ ∆(x, x0 ). Then, by assuming there is no test with a good split constant α, we can leverage the structure of the problems to conclude that the size of V is small. Since any two hypotheses disagree on at least one test (identifiability of hypotheses), we can always split off one of the hypotheses for a split constant of 1/|V | which is a good split if |V | ≤ 1/α. Since there is no test with a split constant α, any test either yields 1 on the vast majority of hypotheses in V or yields 0 on the vast majority of hypotheses in V . Thus, we partition the tests into two sets, X + = {x ∈ X : Pr [h(x) = 1] > 1 − α} h∈V X − = {x ∈ X : Pr [h(x) = 1] < α} = X − X + h∈V Several of the arguments will leverage the structure of this partition and use union bound to show that the probability of a single hypothesis is high, and thus V is small. Two-dimensional linear classifiers with convex polygon data pool Suppose we have a pool of unlabeled data points and our set of hypotheses is linear classifiers in the transductive setting (we group all hypotheses with the same output on all unlabeled data points together). We examine the case of two dimensions. Problem 1 (Linear classifiers on convex polygon data pool). Let X be a set of m points x ∈ R2 such that the points are the vertices of a convex polygon. Let H be equivalence classes of linear classifiers that have the 𝑥′ 𝑥 Application of analysis In this work, we establish the 1/α-split-neighborly condition for four problems: two-dimensional linear classifiers on the vertices of a convex polygon, learning monotonic disjunctions and CNF formulas, discrete object localization (under two different conditions), and discrete linear classifiers. We show that GBS achieves O(log n) cost on these problems under conditions on H by showing that the problems are 1/α-splitneighborly and have constant coherence. Further, we show the inadequacy of the k-neighborly analysis for each of these problems. 4.1 ⌈𝑚/4⌉ ⌈𝑚/4⌉ Figure 5: An illustration of ∆(x, x0 ) for linear classifiers on a data pool forming the vertices of a convex polygon. Note that we can split any subset of ∆(x, x0 ) with a split constant of at least 1/3 because the tests are interleaved in a sequence with the hypotheses. same output on X and such that P   1 3 x∈X h(x) ∈ , . |X | 4 4 This last constraint restricts the classifiers to those with balanced labels. This ensures a good coherence parameter; otherwise, no algorithm can perform better √ than Θ(m) = Θ( n). We will now show that the k-neighborly analysis for this problem is poor. See Figure 5 for a diagram. For adjacent points x and x0 , |∆(x, x0 )| = m − 2dm/4e + 1. Further note that n = |H| = m(m − 2dm/4e + 1) 3 . √ n 0 Thus, |∆(x, x )| ≥ 2 (for m ≥ 4) and so the k for the √ k-neighborly analysis is at least 2n . However, it is clear from Figure 5 that ∆(x, x0 ) is a sequence of hypotheses with tests interleaved. Thus, we can split ∆(x, x0 ) with at least a split constant of 1/3 and to get the following proposition, Proposition 4.1. The problem of learning a linear classifier on a convex polygon data pool is 3-splitneighborly. Note that because of the constraint that the minority label is at least 1/4, the coherence parameter is at least c = 1/4. Thus, the worst case query complexity is at log n ≤ 3.2 log n. most − log(1−1/5) 4.2 Monotonic CNF formulas In this section, we examine the problem of learning monotonic CNF formulas from function evaluations. To begin, we study the case of a single disjunction, such as the following, x4 ∨ x7 ∨ x9 3 We have 2 linear classifiers for each line, but we are double counting. Generalized Binary Search For Split-Neighborly Problems A. 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 1 B. 𝑥 ∨ ⋯ 1 C. 𝑥− 𝑥+ ℎ ∈ 𝑉 ⊆ Δ(𝑥 − , 𝑥 + ) has this form 𝑍 0 0 0 0 1 1 1 1 1 1 1 1 𝑥′ D. 0 1 0 0 1 1 1 1 1 1 1 1 𝑥 (2) 0 0 1 0 1 1 1 1 1 1 1 1 𝑥 (3) 0 0 0 1 1 1 1 1 1 1 1 1 𝑥 (4) E. 𝑥1 ∨ 𝑥2 ∨ 𝑥3 ∨ 𝑥4 Permuted from 1/(m + 1) and the proof is complete. Now, if there are no tests with a good split, each test must either yield 1 or 0 for the vast majority of hypotheses in V . Thus, we can define the following two sets. X + = {x ∈ X : Pr [h(x) = 1] > 1 − q} from Strictly greater than 1− 𝑍 +1 𝑞 proportion Figure 6: A proof illustration for the disjunction problem being split-neighborly. Problem 2 (Disjunction). Let the elements of H be a disjunction over d variables without any negations where the disjunction has at most m variables. Let X be the set of length d bit assignments. First, note that h(0d ) = 0 and h(1d ) = 1 for all h ∈ H and thus the coherence parameter is c = 1/2. The kneighborly is lacking for this problem. Note,  Pm analysis |H| = i=1 di . However, the bit string 0d ∈ X dis Pm agrees with all other x ∈ X for i=1 d−1 i−1 hypotheses. √ So for m ≥ 2, d ≥ 2m, then k ≥ n. See the appendix for details. On the other hand, our split neighborly analysis achieves the optimal rate in the case where m is constant and d goes to infinity. Theorem 4.1. The single disjunction problem is (m + 1)-split-neighborly. Proof. A graphic for the proof is shown in Figure 6. We will show that there are edges between tests that differ by just one bit. This will suffice since such a graph is strongly connected. In particular, we show that the test graph has a bidirectional edge from x to x0 if ||x − x0 ||1 = 1. Let x+ be the value of x or x0 with more 1’s (and let x− be the other one). Note that from monotonicity, |∆(x+ , x− )| = 0 so there is a directed edge from x+ to x− . For the other direction, fix a subset V ⊆ ∆(x− , x+ ). Without loss of generality, let x+ and x− differ in − the first coordinate so x+ 1 = 1 and x1 = 0 and + − ∀i > 1 : xi = xi . See row A of Figure 6. Because V ⊆ ∆(x− , x+ ), all hypotheses in V include x1 in the disjunction. See row B of Figure 6. For ease of notation, let q = 1/(m+1). We will proceed by showing that if there is no test with a good split, Eh∈V [h(x)] ∈ [q, 1 − q], then |V | ≤ m + 1. Then, either |V | ≤ 1 or there is a test with split constant at least h∈V X − = {x ∈ X : Pr [h(x) = 1] < q} = X − X + h∈V 0 Let x be the the element of X − with the fewest 0’s. Since x0 ∈ X − , x01 = 0. Let Z be the other indices of the 0’s. If |Z| = 0, then |∆(x− , x+ )| = |{x1 }| = 1 so |V | ≤ 1 and we are done. Define {x(j) }j∈Z to be the test resulting from x0 and changing the j th bit to a 1. By the minimal definition of x0 , ∀j ∈ Z : x(j) ∈ X + . See rows C and D of Figure 6. We now derive a useful equation. Note that for any subset Z 0 ⊆ Z, from the definition of X + and X − and union bound, Prh∈V [h(x0 ) = 0 ∧ ∀j ∈ Z 0 : h(x(j) ) = 1] > 1 − (|Z 0 | + 1)q. From the property of disjunctions, this implies Prh∈V [h has variables at Z 0 ∪ {1}] > 1 − (|Z 0 | + 1)q. See row E of Figure 6. If |Z| ≥ m, then this means that we can choose a subset Z 0 of size m. Prh∈V [h includes m + 1 variables] > 1 − (m + 1)q = 0. This means there is a non-zero probability of a hypothesis with m + 1 variables which is impossible, since our disjunctions don’t have more than m variables. So |Z| ≤ m − 1. We are nearly done. Note that the left side of the useful equation is exactly 1/|V |. Therefore, 1/|V | > 1 − (|Z| + 1)q ≥ 1 − mq ≥ 1/(m + 1). Rearranging, we find that |V | < m + 1 and we are done. We now examine the more general monotonic CNF problem from function evaluations. An example of such a monotonic formula is: (x1 ∨ x4 ∨ x5 ) ∧ (x2 ∨ x7 ∨ x8 ). Problem 3 (Conjunction of disjunctions). Let H be a conjunction of ` m-disjunctions over d variables without any negations, and where each variable does not appear in multiple disjunctions. Let X be the set of length d bit assignments. Note that there is an isomorphism between conjunctions of disjunctions and disjunctions of conjunctions by flipping the test bits and the result bit. Additionally,√for a general setting shown in the appendix, k ≥ n which renders the k-neighborly analysis very poor. However, the split-neighborly analysis suffices, Stephen Mussmann, Percy Liang Δ(𝑥, 𝑥 ′ ) Theorem 4.2. The conjunction of disjunctions problem is (m + 1 + 3(l − 1))-split-neighborly. The proof is in the appendix. Note that the value of 1/α does not depend on the number of variables, so GBS is efficient even for very large d when m, l are constant. 4.3 Other Tests 𝑥 𝑥′ Object localization in Zd Consider the problem of object localization (Chen et al., 2015) where we try to locate an object based on spatial queries. In this work, we wish to find the location z of an object in space or in an image by asking queries of the form “Is z close to point x?”. We can discretize the space into the grid of integers and define “closeness” by as whether z − x is in some set S (e.g., an `p ball). In this way, hypotheses and tests are both indexed by vectors of integers. For concreteness, for an `p norm ball, a test x returns the result of kx − zkp ≤ `. Problem 4 (Object localization). Fix a set S ⊆ Zd representing the sensing field H ⊆ {hz }z∈Zd X = Zd hz (x) = 1[x − z ∈ S] Note that there are infinitely many integer vectors, but we can take a bounded region as the hypothesis space. Note that if the bounded region of the hypothesis space is too large, the coherence parameter would be very small, and greedy would not perform well (and any algorithm for that matter, since the algorithm would have to do a linear search). One way to make the coherence parameter c = 1/2 is to choose a x∗ and ensure that H ⊆ {hz : z − x∗ ∈ S}. For cases where S is an axis-symmetric box, or equivalently where we use a weighted `∞ norm, the problem is split-neighborly. Figure 7: For the object localization problem where S is an axis-symmetric box, the 1/α-split-neighborly graph has edges between adjacent points, for example x and x0 in this figure. Thus, ∆(x, x0 ) will be a flat box. This figure shows that for d = 2, the problem is 3-split-neighborly. In fact, the problem is 4-splitneighborly for all d. The proof is in the appendix and uses union bound with the partition of X into X − and X + . For this problem, k must be at least the √ largest, axis-aligned “shadow” which for d ≥ 2, k ≥ n. 4.4 Discrete binary linear classifier Linear classifiers are a classic type of function where the output label is y = 1[w·x+b > 0]. We consider the active learning setting where H is a set of (w, b) pairs and X are points x in the feature space. In fact, this is the problem covered previously by Nowak (2011) where w and x take continuous values. Here, we consider the setting of discrete linear classifiers, where x ∈ {0, 1}d and w ∈ {−1, 0, 1}d . Problem 5 (Discrete Binary Linear Classifier). Theorem 4.3. The object localization problem where S is an axis-symmetric box is 4-split-neighborly. H ⊆ {hb,w }b∈Z,w∈{−1,0,1}d The proof is in the appendix. See Figure 7 for some intuition. Note that the value of k for this problem is the largest cross-sectional volume of the box, which √ for d ≥ 2, k ≥ n. hb,w (x) = 1[w · x > b] For axis-symmetric, axis-convex (weaker than convex) sets S, we have a dimension-dependent bound. By an axis-convex set S, we mean that if two points in S differ in only one dimension, then all integral points between them are also in S. Theorem 4.4. If S is a bounded, axis-symmetric, axis-convex set, the object localization problem is (4d + 1)-split-neighborly. X = {0, 1}d with the following holding for all hypotheses (w(+) and w(−) are the number of positive and negative vector components for w, respectively): w(+) − b ≤ r(w(−) + b) − d 8 w(−) + b ≤ r(w(+) − b − 1) − d 8 Intuitively, w(+) −b is the maximum “overshoot” of the threshold and w(−) + b is the maximum “undershoot”. Generalized Binary Search For Split-Neighborly Problems We require that the ratio between these quantities is at most r with the addition of an additive constant. Theorem 4.5. The discrete binary linear classifier problem is max(16, 8r)-split-neighborly. The coherence parameter will be constant c if there is some data distribution such that each hypothesis yields a balanced label distribution (at least probability c of minority label). Intuitively, this is necessary since if the labels are very unbalanced, it may take many queries to even find a minority label. With this condition, GBS achieves O(log n) on the discrete binary linear classifiers problem. On the other hand, the k-neighborly analysis does not work here, as before. In the special case where d is divisible by 4, b = d/4 − 1 and there are an equal number of 1 and 0 weights (just√for making the calculation simpler), for d ≥ 4, k ≥ n. See the Appendix for more details. 5 Discussion and related work The GBS algorithm, or more generally, choosing the test that maximizes the information gain, has several approximations and variants. The greedy information gain technique was introduced in MacKay (1992) and used or extended in active learning (Jedynak et al., 2012), ranking learning (Chu & Ghahramani, 2005), comparison based search (Karbasi et al., 2012), image segmentation (Maji et al., 2014), structured prediction (Sun et al., 2015; Luo et al., 2013), group identification (Bellala et al., 2010), and graphical models (Zheng et al., 2012). As the active learning survey of Settles (2012) notes, “all of the general query frameworks we have looked at contain a popular utility function that can be viewed as an approximation to [information gain] under certain conditions.” Thus, greedy information gain is seen as the gold standard, and there has been significant work finding approximations and extensions. Our work examines the other side of GBS and tries to understand that gap between GBS and the optimal solution. A large body of literature exists on the analysis of GBS and close relatives in the noiseless and well-specified version of the sequential hypothesis testing problem, known as the optimal decision tree problem (Guillory & Bilmes, 2009; Kosaraju et al., 1999; Dasgupta, 2004; Chakaravarthy et al., 2007; Adler & Heeringa, 2008; Chakaravarthy et al., 2009; Gupta et al., 2010). These analyses, which borrow ideas from submodular analysis, yield an average cost ratio of O(log n), where the average cost ratio for a method is defined as the ratio between the expected cost of the method and the expected optimal cost (note that this is significantly worse than an average query cost of O(log n)). Furthermore, there exists a problem where GBS achieves a average cost ratio of Θ(log n/ log log n) (optimal is Θ(log n) but GBS is Θ(log2 n/ log log n)), so the general upper bound for GBS is very close to tight (Dasgupta, 2004). In our work, we show that GBS achieves a constant factor cost ratio, that is, within a constant factor of the optimal cost. In many natural settings such as linear classification, the hypothesis space is exponentially large in the dimension (i.e. n = 2O(d) ), so existing guarantees are O(d) times the optimal, which itself is O(d) for many problems. In our work, we prove in multiple settings that GBS achieves the asymptotically optimal query cost. Other works extend the noiseless and well-specified assumptions to more general frameworks. Nowak (2011) provides a way to adapt GBS to the mis-specified case with only a constant factor increase in the query complexity that ensures GBS never performs worse than randomly querying tests (the naive approach). Kääriäinen (2006) provides a reduction from the noisy case to the noiseless case. There are two different noise settings which are handled separately in the literature, persistent noise (tests are not repeatable) and nonpersistent noise (tests are repeatable). Earlier work (Nowak, 2009, 2011; Naghshvar et al., 2012) that has handled noise has addressed i.i.d. noise with repeatable tests, where the outputs of the deterministic problem are flipped with a constant probability p. In the case of non-persistent i.i.d. noise, Kääriäinen (2006) presents a technique to reduce the noisy case to the deterministic case by repeatedly querying tests and using the majority vote, so that with high probability we attain the uncorrupted test result. Thus, while our work might appear to only handle the noiseless case, it actually handles the non-persistent noise case as well. Theoretical explanations for the effectiveness of GBS are still incomplete. Although GBS always achieves a cost ratio of O(log n) (Guillory & Bilmes, 2009), in the large hypothesis space regime, this factor could be very large. Furthermore, there do exist problems for which GBS performs much worse than the optimal (Dasgupta, 2004). These examples, however, tend to be contrived. Anecdotally, from the sample problems in this paper, we found that GBS is effective for most “natural” problems. In conclusion, we have made progress on characterizing this observation by introducing the 1/α-split-neighborly condition, which provably ensures that GBS achieves the asymptotically optimal query cost. Acknowledgments This research was supported by NSF grant DGE1656518. Stephen Mussmann, Percy Liang References Adler, Micah and Heeringa, Brent. Approximating optimal binary decision trees. In Approximation, Randomization and Combinatorial Optimization. Algorithms and Techniques, pp. 1–9. Springer, 2008. Bellala, Gowtham, Bhavnani, Suresh, and Scott, Clayton. Extensions of generalized binary search to group identification and exponential costs. In Advances in Neural Information Processing Systems, pp. 154–162, 2010. Chakaravarthy, Venkatesan T, Pandit, Vinayaka, Roy, Sambuddha, Awasthi, Pranjal, and Mohania, Mukesh. Decision trees for entity identification: approximation algorithms and hardness results. In Proceedings of the twenty-sixth ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, pp. 53–62. ACM, 2007. Chakaravarthy, Venkatesan T, Pandit, Vinayaka, Roy, Sambuddha, and Sabharwal, Yogish. Approximating decision trees with multiway branches. In International Colloquium on Automata, Languages, and Programming, pp. 210–221. Springer, 2009. Chen, Yuxin, Javdani, Shervin, Karbasi, Amin, Bagnell, J Andrew, Srinivasa, Siddhartha S, and Krause, Andreas. Submodular surrogates for value of information. In AAAI, pp. 3511–3518, 2015. Chu, Wei and Ghahramani, Zoubin. Extensions of gaussian processes for ranking: semisupervised and active learning. Learning to Rank, pp. 29, 2005. Dasgupta, Sanjoy. Analysis of a greedy active learning strategy. In Advances in neural information processing systems, pp. 337–344, 2004. Garey, Michael R and Graham, Ronald L. Performance bounds on the splitting algorithm for binary testing. Acta Informatica, 3(4):347–355, 1974. Guillory, Andrew and Bilmes, Jeff. Average-case active learning with costs. In International Conference on Algorithmic Learning Theory, pp. 141–155. Springer, 2009. Gupta, Anupam, Nagarajan, Viswanath, and Ravi, R. Approximation algorithms for optimal decision trees and adaptive tsp problems. In International Colloquium on Automata, Languages, and Programming, pp. 690–701. Springer, 2010. Jedynak, Bruno, Frazier, Peter I, and Sznitman, Raphael. Twenty questions with noise: Bayes optimal policies for entropy loss. Journal of Applied Probability, 49(01):114–136, 2012. Kääriäinen, Matti. Active learning in the nonrealizable case. In International Conference on Algorithmic Learning Theory, pp. 63–77. Springer, 2006. Karbasi, Amin, Ioannidis, Stratis, et al. Comparisonbased learning with rank nets. arXiv preprint arXiv:1206.4674, 2012. Kosaraju, S Rao, Borgstrom, Ryan. lem. In Workshop tures, pp. 157–168. Przytycka, Teresa M, and On an optimal split tree probon Algorithms and Data StrucSpringer, 1999. Luo, Wenjie, Schwing, Alex, and Urtasun, Raquel. Latent structured active learning. In Advances in Neural Information Processing Systems, pp. 728–736, 2013. MacKay, David JC. Information-based objective functions for active data selection. Neural computation, 4(4):590–604, 1992. Maji, Subhransu, Hazan, Tamir, and Jaakkola, Tommi S. Active boundary annotation using random map perturbations. In AISTATS, pp. 604–613, 2014. Naghshvar, Mohammad, Javidi, Tara, and Chaudhuri, Kamalika. Noisy bayesian active learning. In Communication, Control, and Computing (Allerton), 2012 50th Annual Allerton Conference on, pp. 1626–1633. IEEE, 2012. Nowak, Robert. Generalized binary search. In Communication, Control, and Computing, 2008 46th Annual Allerton Conference on, pp. 568–574. IEEE, 2008. Nowak, Robert. Noisy generalized binary search. In Advances in neural information processing systems, pp. 1366–1374, 2009. Nowak, Robert D. The geometry of generalized binary search. IEEE Transactions on Information Theory, 57(12):7893–7906, 2011. Settles, Burr. Active learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, 6(1): 1–114, 2012. Sun, Qing, Laddha, Ankit, and Batra, Dhruv. Active learning for structured probabilistic models with histogram approximation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3612–3621, 2015. Young, Linda J and Young, Jerry H. Sequential hypothesis testing. In Statistical Ecology, pp. 153–190. Springer, 1998. Zheng, Alice X, Rish, Irina, and Beygelzimer, Alina. Efficient test selection in active diagnosis via entropy approximation. arXiv preprint arXiv:1207.1418, 2012. Generalized Binary Search For Split-Neighborly Problems 6 Appendix X 6.1 Split-neighborly proofs P (x)S(x) ∈ [β, 1 − β] x∈X Theorem 3.2. If a problem is 1/α-split-neighborly and has a coherence parameter of c, for There are two possibilities, either there exists a test x such that 1 β = min(c, 1 ) /α + 2 S(x) = Eh∈V [h(x)] ∈ [β, 1 − β] log n − log(1−β) log n most H(β) GBS has a worst case query cost of at most and GBS has an average query cost of at where H(p) is the entropy of a Bernoulli(p) random variable Proof. This theorem will follow from the next three lemmas. Lemma 3.1. If a problem is 1/α-split-neighborly and has a coherence parameter of c, then for any V ⊆ H, |V | ≤ 1 or there exists a test x ∈ X such that in which case, this is the exact conclusion statement and we are done, or that there exists no test with a split constant in [β, 1 − β]. If there exists no test with a split constant in [β, 1 − β] but the weighted combination is in [β, 1 − β], then there exists tests x and x0 such that S(x) < β and S(x0 ) > 1 − β Since the problem is 1/α-split neighborly, there exists a graph over the tests that is strongly connected. Thus, there is a path from x to x0 . Since S(x) < β and S(x0 ) > 1 − β and since ∀x00 ∈ X : S(x00 ) 6∈ [β, 1 − β], there must be an edge (x0 , x1 ) along the path where S(x0 ) < β and S(x1 ) > 1 − β. Thus, Eh∈V [h(x)] ∈ [β, 1 − β] Pr [h(x0 ) = 1] = Eh∈V [h(x0 )] < β h∈V Pr [h(x1 ) = 1] = Eh∈V [h(x1 )] > 1 − β where β is defined as above. Proof. Fix a subset V ⊆ H. Assume |V | > 1, otherwise we are done. h∈V Combining these two yields, From the assumption, we have a coherence parameter of Pr [h(x0 ) = 0 ∧ h(x1 ) = 1] > 1 − 2β h∈V Recall ∆(x0 , x1 ) = {h ∈ H : h(x0 ) = 0, h(x1 ) = 1} c≥β Pr [h ∈ ∆(x0 , x1 )] > 1 − 2β From the definition, this means that there exists a probability distribution on the tests P such that for any hypothesis h, X P (x)h(x) ∈ [β, 1 − β] h∈V |V ∩ ∆(x0 , x1 )| > 1 − 2β |V | Recall from the definition of β that 1 1/α+2 ≥ β. Thus x∈X 1/α 1 β =1 ≥ 1 − 2β ≥ 1 − 2 1 /α + 2 /α + 2 α Since this is true for all h ∈ H, this is also true for all convex combinations. Thus, Thus, Eh∈V [ X P (x)h(x)] ∈ [β, 1 − β] x∈X X |V ∩ ∆(x0 , x1 )| β > |V | α P (x)Eh∈V [h(x)] ∈ [β, 1 − β] x∈X For simplicity, define the split constant S(x) = Eh∈V [h(x)]. Thus, For brevity, define ∆ = ∆(x0 , x1 ). Since there is an edge (x0 , x1 ) in the 1/α-neighborly graph, for any subset including V ∩ ∆ ⊆ ∆, either |V ∩ ∆| ≤ 1 or there exists a test x̂ such that, Stephen Mussmann, Percy Liang Eh∈V ∩∆ [h(x̂)] ∈ [α, 1 − α] |V ∩∆(x0 ,x1 )| β >α . |V | |V ∩∆(x0 ,x1 )| β > α and |V | > |V | Lemma 6.1. If, for any V ⊆ H, |V | ≤ 1 or there exists a test x ∈ X such that First, |V ∩ ∆| = 6 0, since |V | > 1 and If |V ∩ ∆| = 1, then, β 1 so 12 ≥ |V1 | > α ≥ β. Since the hypotheses are identifiable, any pair of hypotheses yield a different result on some test, so we can always find a test with a split constant of at least |V1 | , and this implies the result of the theorem. In the other case, where |V ∩ ∆| > 1, we have all the necessary pieces and it’s just a matter of crunching the algebra. P h(x̂) |V | h∈V Eh∈V [h(x̂)] = Eh∈V [h(x)] ∈ [β, 1 − β] then GBS has a worst case query cost of at most log n 1 log( 1−β ) Proof. After m queries, there are at most max(1, (1 − β)m n) remaining hypotheses since greedy will choose a test with a split constant of at least β (a split with respect to the hypotheses without a prior) and will terminate when there is a single hypothesis. Thus, when (1 − β)m n ≤ 1, the algorithm must have terminated. log n the alRearranging, we see that when m ≥ log( 1 ) 1−β gorithm must have terminated. This means that the log n worst case query cost must be at most log( . 1 ) 1−β P h(x̂) ≥ h∈V ∩∆ |V | P β h∈V ∩∆ h(x̂) ≥ α |V ∩ ∆| Lemma 6.2. If, for any V ⊆ H, |V | ≤ 1 or there exists a test x ∈ X such that β Eh∈V ∩∆ [h(x̂)] α log n then GBS has an average query cost of at most H(β) where H(p) is the entropy of a Bernoulli(p) random variable ≥ ≥ Eh∈V [h(x)] ∈ [β, 1 − β] β α=β α Proof. Define H(p) as the entropy of a Bernoulli random variable with probability p. Additionally, P P = ≤ ≤ h∈V ∩∆ h(x̂) + h(x̂) |V | h∈V Eh∈V [h(x̂)] = P h∈V \∆ h(x̂) |V | (1 − α)|V ∩ ∆| + P h∈V \∆ f (V ) = E[average queries remaining while at subset V ] (1) We will prove by induction on increasing subsets that h(x̂) f (V ) ≤ |V | (1 − α)|V ∩ ∆| + |V | − |V ∩ ∆| |V | ≤1−α ≤1−α |V ∩ ∆| |V | β =1−β α Thus, we have that Eh∈V [h(x̂)] ∈ [β, 1 − β] log(|V |) H(β) (2) Note that the base case is that f ({h}) = 0 because we are done when there is just one hypothesis left. Note that this suffices to show that the total runtime is log(n)/H(β) because |V | = n at the beginning of the algorithm. Let A, B be a partition of V based on a test split. Without loss of generality, let |A| ≤ |B|, so |A| ≤ 1/2|V |. Based on the recursive definition of cost and there is a test with a split constant of at least β (so GBS will choose a test with a split constant of at least β), which is the conclusion of the lemma. f (V ) ≤ max A,B,|A|/|V |∈[β,1/2] |B| |A| f (A) + f (B) + 1 |V | |V | Generalized Binary Search For Split-Neighborly Problems From the induction hypothesis, ≤ max ... |A| log |A| |B| log |B| + +1 |V | H(β) |V | H(β) ≤ log |A| + |B| |V | i i=1  m  X d−1 k≥ |A| max... |V | m   X d n= i=1 log |B| + H(β) H(β) k ≥1+ i−1 m−1 X d−1 i i=1 ≤ |A| max... |V | log |A| |V | + |B| |V | log |B| |V | + H(β) + log |V | H(β) k2 − n ≥ 1 + 2 ≤ − Note that 2 Value of k Disjunctions For the disjunctions problem, for m ≥ 2, d ≥ 2m, d−1 i  ≥ d i  k2 − n ≥ 1 + ( since i ≤ m − 1 ≤ d/2. m−1 X i=1    d−1 2 d ) − i m 2   d−1 d − m−1 m    d−1 d−1 ≥ ( − d/m) m−1 m−1  Thus, we have proved the statement by induction and this suffices to show that the total runtime is at most log(n)/H(β). Proof. In the case that k = 1, |∆(x, x0 )| = 1 so |V | ≤ 1 so the problem is 1-split-neighborly. Assume k > 1. Note that any set of hypotheses must have a test that distinguishes at least one of the hypotheses (otherwise the hypotheses are the same). If two points x and x0 in the k-neighborly graph have an edge between them, then |∆(x, x0 ) ∪ ∆(x0 , x)| ≤ k, which implies |∆(x, x0 )| ≤ k, and thus either |∆(x, x0 )| ≤ 1 or there is a test with a 1/k split constant and thus there is an edge from x to x0 in the k-split-neighborly graph. By a similar argument, there is also an edge from x0 to x. Since the k-neighborly graph is connected and each edge corresponds to a bidirectional edge in the k-split-neighborly graph, the k-split-neighborly graph is strongly connected and thus the problem is k-splitneighborly.    d d − i m i=1 log(|V |) f (V ) ≤ H(β) 6.2.1 m−1 X [β, 1/2] Proposition 3.1. If a problem is k-neighborly and has a uniform prior, then the problem is k-split-neighborly.  m−1 X d − 1 d−1 +( )2 i i i=1 i=1 H(β) Note that since ∈ (the condition of the |A| max), H( |V ) ≥ H(β). Thus, the max is non-positive, | and thus, 6.2 m−1 X |A| max... −H( |V | ) + H(β) + log |V | |A|/|V |  ≥ Since m ≥ 2,  ≥  d−1 − d/2 1 ≥ d/2 − 1 ≥m−1 ≥0 Thus, k 2 − n ≥ 0 and so k ≥ 6.2.2 √ n. Monotonic CNF  d Note that n = |H| = l!1 m,m,...,m,d−lm . All of the bit strings with strictly less than l ones will be trivially connected in the k-neighborly graph, because they yield 0 on all hypotheses. However, the closest test to connect them to the rest of the graph is the bit string 1l 0d−l ∈ X , which disagrees on d−l m−1,m−1,...,m−1,d−lm ≤ k hypotheses. We examine the case where d ≥ 2ml and m ≥ 2. For the monotonic CNF formulas, recall that   1 d n = |H| = l! m, m, ..., m, d − lm Stephen Mussmann, Percy Liang  k≥  d−l m − 1, m − 1, ..., m − 1, d − lm For d ≥ 2ml and m ≥ 2, k ≥  √ n ≤ k2 n.  d−l ≤k m − 1, m − 1, ..., m − 1, d − lm = 4c 2c  3c 2 c  (4c)!c!c! (3c)!(3c)! Note that we have the common Stirling’s approximation, and   1 d n= l! m, m, ..., m, d − lm = = √ d! 1 l! (m!)l (d − lm)! 2πnn+1/2 e−n ≤ n! ≤ enn+1/2 e−n Thus, (d − l)! 1 d!(d − 2l)! (d − l)! (m − 1)!l (d − lm)! ml (d − l)!2 l!(d − 2l)!   1 d!(d − 2l)! d − l ≤k l m (d − l)!2 l e3 (4c)4c+1/2 cc+1/2 cc+1/2 e−6c n ≤ k2 2π(3c)3c+1/2 (3c)3c+1/2 e−6c √ 2e3 c(4c)4c cc cc = 6π(3c)3c (3c)3c √ e3 c44c = 3π33c 33c Since d ≥ 2ml ≥ 4l, n≤k   2l d − l ml l = Since d − l ≥ 2l(m − 1) and m ≥ 2 e3 √ 256 c c( ) 3π 729 ≤1   d−l n≤k l(m − 1)  n≤k for c ≥ 1.  d−l m − 1, m − 1, ..., m − 1, d − lm Thus, for d ≥ 4, n ≤1 k2 √ k≥ n n ≤ k2 √ k≥ n 6.2.3 Discrete Linear Classifier 6.3 Necessity of Dependencies Recall that we are in the special case where d is divisible by 4, b = d/4 − 1 and there are an equal number of 1 and 0 weights (d/2). 6.3.1 All tests with fewer than d/4 1’s will yield a result of 0 for all hypotheses. The test with the next fewest hypotheses that yield 1 will be a test with exactly d/4 1’s. Thus, k is at least the number of such hypotheses that yield 1. For arbitrary data points where the points are not the vertices of a convex polygon, the linear classifier problem is not 1/α-split-neighborly for constant α. A counter-example is shown in Figure 8. 6.3.2  Disjunctions  n= d d/2 k≥   3d/4 d/4 For simplicity, define c = d/4. Linear classifiers on convex polygon data pool The linear dependence on m for the disjunctions is necessary because of the case where d = m + 1, and |H| = d (each h ∈ H lacking one variable). In this 1 case, there are no tests with split constants of m , so the problem cannot be better than (m−2)-split-neighborly (recall coherence c = 1/2). Generalized Binary Search For Split-Neighborly Problems x− be the other one). Note that |∆(x+ , x− )| = 0 so there is a directed edge (x+ , x− ). For the other direction, fix a subset V ⊆ ∆(x− , x+ ). Without loss of generality, let x+ and x− differ in the − first coordinate so x+ 1 = 1 and x1 = 0 and ∀i > 1 : + − xi = xi . Figure 8: A counterexample that shows a non-convex data pool need not be split-neighborly. Note that we can at most split off 1 of the n hypotheses by querying one of the points from the lower half. However, the problem has coherence close to 1/2 and thus it cannot be 1/α-split-neighborly for constant α. For a proof by contradiction, the problem is not (m + 1)-split-neighborly so that |V | > 1 and there is no test x such that Eh∈V [h(x)] ∈ [q, 1 − q], where q = 1/(m + 1). Let X + = {x ∈ X : Pr [h(x) = 1] > 1 − q}, h∈V X − = {x ∈ X : Pr [h(x) = 1] < q} = X − X + . h∈V 6.3.3 Monotonic CNF For the monotonic CNF problem, the linear dependence on m is necessary because of the case where l = 1, d = m + 1, and |H| = d (each h ∈ H lacking one variable). In this case, there are no tests with split 1 , so the problem cannot be better than constants of m (m − 2)-split-neighborly (recall coherence c = 1/2). Furthermore, the linear dependence on l is necessary because of the problem where m = 1, d = l + 1, and |H| = d (each h ∈ H lacking one variable). For this problem, there are no tests with split constants of 1l , so the problem cannot be better than (l − 2)-splitneighborly. Thus, although the linear dependence on m and l is necessary, it may be possible to improve the constants. 6.3.4 Object Localization For object localization with the axis-symmetric, axisconvex set S, the dependence on d is necessary because if we use the set S = {jei : |j| ≤ l, 1 ≤ i ≤ d} and consider the set of hypotheses, {±lei : 1 ≤ i ≤ d}, the 1 but problem has no test with split constant of 2d−1 has coherence c = 1/2, so it can’t be (2d − 3)-splitneighborly. 6.4 Monotonic CNF Theorem 4.2. The Conjunction of Disjunctions problem is (m + 1 + 3(l − 1))-split-neighborly. Proof. We prove this theorem by induction on l. First, for the base case l = 1. Let x0 be the the element of X − with the fewest 0’s and let the 0’s be at indices Z (note 1 ∈ Z). If |Z| < m, then h(x0 ) = 1 for all h since the disjunctions have m variables. But since x0 ∈ X − , which is a contradiction. Define {x(j) }j∈Z to be the test resulting changing the j th bit of x0 to a 1. By the minimal definition of x0 , ∀j ∈ Z : x(j) ∈ X + . Suppose |Z| > m. Take a subset Z 0 ⊆ Z such that |Z 0 | = m + 1. Then, from the definition of X + and X − , Prh∈V [h(x0 ) = 0 ∧ ∀j ∈ Z 0 : h(x(j) ) = 1] > 1 − (m + 1)q ≥ 0, which means Prh∈V [h includes variables Z 0 ] > 0. Therefore, there is a disjunction with at least m + 1 variables, which is a contradiction. Thus, |Z| = m, so there is only one hypothesis such that h(x0 ) = 0, the hypothesis with variables at Z. So 1/|V | > 1 − q (by definition of X − ), which implies |V | = 1 since q ≤ 1/2, which is a contradiction. Thus, by contradiction, the problem with l = 1 is (m + 1)split-neighborly. For l > 1, we proceed by induction. We can define the graph as above, define X − and X + as above, and x0 and Z as above. The same argument goes through that |Z| = m. Thus, (1 − q) proportion of the hypotheses have a disjunction with variables at the indices Z. These hypotheses are simply another copy of the problem with l − 1 conjunctions and d − m variables. Since that problem has 1/2 coherence and is m + 1 + 3(l − 2)-splittable (by induction hypothesis), there exists some test with a split constant of 1 m+1+3(l−2)+2 for a total split constant on the original problem of The test graph has an edge from x to x0 if ||x−x0 ||1 = 1 (the bit strings differ in one location). Let x+ be the value of x or x0 with more 1’s (and let (1 − q) 1 1 = m + 1 + 3(l − 2) + 2 m + 1 + 3(l − 1) Stephen Mussmann, Percy Liang Thus, the problem is m + 1 + 3(l − 1)-split-neighborly by induction. 6.5 Box Object Localization Theorem 4.3. The object localization problem where S is a box is 4-split-neighborly. Notationally, refer to zh as the integer vector for the hypothesis h and zh,i to be its ith component. We begin by fixing two tests x and x0 such that ||x − x0 ||1 = 1. Without loss of generality, let x0 − x = e1 where e1 is the 1st elementary vector. Since the box is axis symmetric, there exists radii ri ≥ 0 such that x − zh ∈ S ↔ ∀i : |xi − zh,i | ≤ ri . Without loss of generality, assume x = (r1 , 0, 0, ..., 0) and x0 = (r1 + 1, 0, 0, ..., 0). Recall ∆(x, x0 ) = {h : h(x) = 0 ∧ h(x0 ) = 1}, this implies that ∆(x, x0 ) = {h : zh,1 = 0 ∧ ∀i > 1 : |zh,i | ≤ ri }. We will begin by fixing a subset V ⊆ ∆(x, x0 ). As in all the application proofs, we will start by assuming by contradiction that there is no test with a split constant in the range [q, 1 − q] where q = 1/4. We will use this contradiction to show that the size of V is small, so that there is in fact a test with a split constant q which is a contradiction. 6.5.1 Majority Element 6.5.2 Side Splits Intuitively, we will create a sequence of tests that each remove at least half of the elements with the ith component not equal to m. For each test in the sequence, the probability that the test yields 1 over the hypotheses in V must be greater that 1 − q so we can prove that there aren’t many elements that disagree with m at any component. Here we recursively define sets Si , Bi , and Ai . Si will be defined in terms of Bi and Bi will be defined in terms of Si−1 . Define S0 = V and for i > 1, Si = Si−1 − Bi . Noting that we could reflect the ith component about mi , without loss of generality, suppose that Pr [zh,i > mi ] ≥ Pr [zh,i < mi ] h∈Si h∈Si Define Bi = {h ∈ Si−1 : zh,i > mi } and Ai = {h ∈ Si−1 : zh,i < mi } Note that |Bi | ≥ |Ai |. Further, there is a test x(i) = (−r1 , ..., −ri , 0, ...0) such that h(x(i) ) = 1 ↔ h ∈ Si and thus by the contradiction assumption, Fix a dimension i. Examine the tests Xi = {jei : j = 0, .., 2ri + 1} and note that for h ∈ V ⊆ ∆(x, x0 ), h(jei ) = 1[zh,i ≥ j − ri ]. By the contradiction assumption, Eh∈V [h(jei )] 6∈ [q, 1 − q] Pr [zh,i ≥ j − ri ] 6∈ [q, 1 − q] h∈V |Si | 6∈ [q, 1 − q] |S| However, since Prh∈V [zh,i = mi ] > 1 − 2q, |Bi |/|V | < 2q. We now prove by induction that |Si |/|V | > 1 − q. The base case is that |S1 |/|V | = 1 > 1 − q. As long as q ≤ 1/4, since |Si−1 |/|V | > 1 − q and |Bi |/|V | < 2q, |Si |/|S| > 1 − 3q ≥ q (since q = 1/4)and thus by the contradiction assumption |Si |/|S| > 1 − q. Note that the Bi are disjoint because Since Prh∈V [zh,i ≥ −ri ] = 1 and Prh∈V [zh,i ≥ ri +1] = 0, there must be some integer mi such that Bi ⊆ Si = V − B1 − B2 − ... − Bi−1 Pr [zh,i ≥ mi ] > 1 − q h∈V |Sd | > (1 − q)|V | Pr [zh,i ≥ mi + 1] < q h∈V |V − which implies that Pr [zh,i = mi ] > 1 − 2q h∈V d G Bi | > (1 − q)|V | i=1 |V | − d X |Bi | > (1 − q)|V | i=1 Define thus, there exists a vector m such that there is a 1 − 2q probability that an hypothesis’ ith component matches m. q|V | > d X i=1 |Bi | Generalized Binary Search For Split-Neighborly Problems Define the set of elements M 0 ⊆ V as the points with a component not equal to m. This is the union of all Ai and Bi , |M 0 | = | d [ Ai ∪ i=1 ≤ d X d [ Bi | i=1 |Ai | + i=1 d X zh ∈ S} and since S is bounded, there exists some B ∈ Z such that ±Bei ∈ X − . Thus there exists some c1 ≤ 0, c2 ≥ 0 such that (c1 − 1)ei ∈ X − , c1 ei ∈ X + , c2 ei ∈ X + , (c2 + 1)ei ∈ X − . From the definition of X + and X − , Pr [h((c1 − 1)ei ) = 0, h(c1 ei ) = 1, ... h∈V |Bi | h(c2 ei ) = 1, h((c2 + 1)ei ) = 0] > 1 − 4q i=1 ≤2 d X Define Sl = {s−i : si = l, s ∈ S} to be the slices of S along axis i at location l. Therefore, h(jei ) = 1 ↔ zh,−i ∈ Szh,i −j . |Bi | i=1 < 2q|V | 0 Also note that |M | ≥ |V | − 1 since there can only be one element that doesn’t disagree with any element of m. Thus, Note that S−l = Sl since the shape S is axis symmetric. Combining these three facts, Pr [zh,−i ∈ S|zh,i −c1 | ∩ S|zh,i −c2 | \ ... h∈V \(S|zh,i −(c1 −1)| ∪ S|zh,i −(c2 +1)| )] > 1 − 4q |V | − 1 < 2q|V | 1 1 − 2q Since q ≤ 1/3, then this implies |V | < 3 so there is a test with a split of 1/3, which is a contradiction. So in a proof by contradiction, the problem is 4-splitneighborly. Note that for l0 > l ≥ 0, Si,l0 ⊆ Si,l because of axisconvexity. To see this, suppose there was t ∈ Si,l0 \ Si,l , then there would be three elements s(−1) , s(0) , s(1) (j) (−1) (0) (1) such that s−i = t and si = −l0 , si = l, si = l0 , which would imply s(−1) ∈ S, s(0) 6∈ S, s(1) ∈ S which contradicts axis-convexity. 6.6 Thus, in order for the set composed of slices of S in the equation above to be non-empty, |V | < Convex, axis-symmetric Shape Object Localization Theorem 4.4. If S is a bounded, axis-symmetric, axis-convex shape, the object localization problem is (4d + 1)-split-neighborly. 0 Proof. Let the test graph has an edge from x to x if ||x − x0 ||1 = 1. Fix a subset V ⊆ ∆(x, x0 ). Without loss of generality, let x0 = 0d . V ⊆ ∆(x, x0 ) ⊆ {h : h(x0 ) = 1} = {h : zh − x0 ∈ S} = {h : zh ∈ S} For a proof by contradiction, the problem is not 4d+1split-neighborly so that |V | > 1 and there is no test x such that Eh∈V [h(x)] ∈ [q, 1−q], where q = 1/(4d+1). |zh,i − c1 |, |zh,i − c2 | < |zh,i − c1 + 1|, |zh,i − c2 − 1| it must be the case that zh,i = define to be mi . So, c1 +c2 2 ∈ Z which we Pr [zh,i = mi ] > 1 − 4q h∈V Repeating this argument for all dimensions and combining, Pr [∀i : zh,i = mi ] > 1 − 4dq h∈V Let There is only one such element zh = m so X + = {x ∈ X : Pr [h(x) = 1] > 1 − q} h∈V X − = {x ∈ X : Pr [h(x) = 1] < q} = X − X + h∈V Note that x0 = 0d ∈ X + since V ⊆ {h : h(x0 ) = 1}. Fix a dimension i. Examine the set of tests {jei : j ∈ Z}. From above, 0ei ∈ X + . Further, since V ⊆ {h : 1 1 1 > 1 − 4dq = 1 − 4d = |V | 4d + 1 4d + 1 So |V | < 4d + 1 so there must be a split of at least q which is a contradiction. Stephen Mussmann, Percy Liang 6.7 Discrete Binary Linear Classifiers Theorem 4.5. The discrete binary linear classifier problem is max(16, 8r)-split-neighborly. 1 1 , 8r ) Define q = min( 16 Recall that for the Discrete Binary Linear Classifier case, we have hypotheses as a pair of vectors and threshold h = (wh , bh ) ∈ {−1, 0, 1}d × Z and tests as vectors {0, 1}d . Recall h(x) = 1[wh · x > bh ]. From the problem setting of Discrete Binary Linear Classifiers, we know that, (+) wh (−) wh −b≤ (−) r(wh (+) + b ≤ r(wh Note that h(x(0) ) = 0 implies that 0 ≤ bh . Further note that, h(x(1) ) = 1 implies that wh,1 > bh . Thus, the only possibility is that wh,1 = 1 and bh = 0. Let T+− denote the number of flips from positive to negative weights and let T−+ denote the number of flips from negative to positive. Then, the weights for the new (reduction) problem will be (+) wnew = w(+) + T−+ − T+− d + b) − 8 − b − 1) − (−) wnew = w(−) + T+− − T−+ d 8 Recall w(+) is the number of positive elements of w and w(−) is the number of negative elements. Notationally wh,i refers to the ith component of wh . 6.7.1 flip the 0’s and 1’s of the test so long as we flip the nonzero entries of wh at that same position, and change bh accordingly. We flip the bits so that x becomes x(0) and x0 becomes x(1) . 0 = bnew = b − T+− + T−+ From the last equation, b = T+− − T−+ . Thus, (+) wnew = w(+) − b (−) wnew = w(−) + b Key Lemma and its Sufficiency We will first state a lemma and then prove that it implies the problem stated stated. Since, Lemma 6.3. Define w(+) − b ≤ r(w(−) + b) − x(0) = (0, 0, ..., 0) x (1) H = {h ∈ H : h(x(0) = 0 ∧ h(x(1) ) = 1 ∧ ... (+) ∧wh (−) ... ∧ wh (+) ≤ r(wh 1 (+) (−) wnew ≤ rwnew − d 8 d − 1) − } 8 Proof of Theorem 4.5 from Lemma 6.3 We will prove Theorem 4.5 by a reduction to Lemma 6.3. To show that the problem is 1/α-split-neighborly, we need to show that for two tests with x and x0 with ||x − x0 ||1 = 1 that for any subset V ⊆ ∆(x, x0 ) = {h : h(x) = 0 ∧ h(x0 ) = 1}, that |V | ≤ 1 or there exists a test x̂ such that Pr [h(x̂) = 1] ∈ [q, 1 − q] h∈V Note that by permuting the indices of x and x0 , we can make the first index the one that is different between x and x0 . Additionally, for the remaining indices we can d 8 then, d (−) ≤ rwh − ∧ ... 8 For any subset V ⊂ H 0 , there exists a test x such that Eh∈V [h(x)] ∈ [q, 1 − q] 6.7.2 w(−) + b ≤ r(w(+) − b − 1) − = (1, 0, ..., 0) 0 d 8 1 (−) (+) wnew ≤ r(wnew − 1) − d 8 We can see that the hypothesis conditions for the original theorem imply that ∆(x(0) , x(1) ) is a subset of the (−) hypotheses that satisfy the conditions based on wnew (+) and wnew so Lemma 6.3 implies the binary linear classifier is 1/q-split-neighborly which means max(16, 8r)split-neighborly. 6.7.3 Proof of Lemma 6.3 The remainder of this is devoted to proving Lemma 6.3 We begin by fixing a subset V ⊆ H 0 . As in all the application proofs, we will start by assuming by contradiction that there is no test with a split constant in the range [q, 1 − q]. We will use this contradiction to show that the size of V is small. Generalized Binary Search For Split-Neighborly Problems Recall that bh = 0 for all hypotheses in the reduced problem and wh,1 = 1. This follows from the fact that h(x(0) ) = 0 and h(x(1) ) = 1. 6.7.5 Ratio between m(−) and m(+) Note, (+) 6.7.4 Eh∈V [wh ] = Majority Vector d X Pr[wh,i = 1] i=1 Let ei be an elementary vector with all entries 0 except for the ith entry which is 1. Lemma 6.4. There exists a vector m ∈ {−1, 0, 1}d such that ∀i : mi = 0 : Prh∈V [wh,i = mi ] ≥ 1 − 2q and ∀i : mi 6= 0 : Prh∈V [wh,i = mi ] ≥ 1 − q ≤ (q)(d − m(+) ) + (1)m(+) = qd + (1 − q)m(+) m(+) ≥ 1 (+) (Eh∈V [wh ] − qd) 1−q Further note, Proof. By the contradiction assumption, there isn’t a test with a split constant greater than q, (+) Eh∈V [wh ] = d X Pr[wh,i = 1] i=1 Pr [wh · ei > bh ] 6∈ [q, 1 − q] h∈V ≥ (0)(d − m(+) ) + (1 − q)m(+) = (1 − q)m(+) m(+) ≤ Pr [wh,i > 0] 6∈ [q, 1 − q] h∈V Pr [ws,i = 1] 6∈ [q, 1 − q] h∈V Also, by the contradiction assumption, 1 (+) Eh∈V [wh ] 1−q (−) We have similar equations for m(−) and Eh∈V [wh ] Let m̄ be the vector of m without the first component. Recall that we have Pr [wh · (e0 + ei ) > bh ] 6∈ [q, 1 − q] (+) h∈V Pr [1 + wh,i > 0] 6∈ [q, 1 − q] h∈V Pr [wh,i 6= −1] 6∈ [q, 1 − q] h∈V ∀h ∈ V : wh (−) ≤ rwh (+) 1 − d 8 (−) Eh∈V [wh ] ≤ rEh∈V [wh ] − qrd 1 1 (+) (−) Eh∈V [wh ] ≤ r (Eh∈V [wh ] − qd) 1−q 1−q Pr [wh,i = −1] 6∈ [q, 1 − q] m(+) ≤ rm(−) h∈V m̄(+) ≤ rm̄(−) Since Prh∈V [wh,i = 1] + Prh [wh,i = 0] + Prh [wh,i = 1] = 1, Also, recall, Pr [wh,i = 0] 6∈ [q, 1 − 2q] h∈V (−) ∀h ∈ V : wh Thus, each index is either mostly 1, mostly 0, or mostly −1 for elements in S (since q < 1/3). Define m ∈ {−1, 01}d such that (−) (+) ≤ r(wh (+) Eh∈V [wh ] ≤ rEh∈V [wh ] − qrd − r 1 1 r (−) (+) Eh∈V [wh ] ≤ r (Eh∈V [wh ] − qd) − 1−q 1−q 1−q mi = argmax Pr [wh,i = c] c h∈V 1 − 1) − d 8 m(−) ≤ rm(+) − r 1−q m(−) ≤ r(m(+) − 1) Note that m1 = 1. m̄(−) ≤ rm̄(+) Stephen Mussmann, Percy Liang 6.7.6 Partition Let w̄ be the vector of w without the first component. Definition 6.1. Let • X + = {x : Prh∈V [w¯h · x ≥ 1] > 1 − q} Proof. We prove this by induction on the number of 1’s in x for x ∈ X ∗ . The base case is x = 0d which is trivially in X 0 . For other x, suppose xi = 1 at a location where mi = 0. Then we know x − ei ∈ X 0 by the induction hypothesis. • X 0 = {x : Prh∈V [w¯h · x = 0] > 1 − 2q} • X − = {x : Prh∈V [w¯h · x ≤ −1] > 1 − q} P rh∈V [wh · (x − ei ) = 0] > 1 − 2q Lemma 6.5. X + , X 0 , X − is a partition of {0, 1}d−1 Proof. Since q ≤ 1/4 and the three defining events are mutually exclusive. It is clear that A+ , A0 , A− are disjoint. Next we show that every point is in at least one of the sets. Suppose a point x is in neither A+ or A− . Using the contradiction assumption on the test (0, x), Pr [wh · (0, x) > 0] 6∈ [q, 1 − q] h∈V P rh∈V [wh,i = 0] > 1 − 2q From these, P rh∈V [wh · x = 0] > 1 − 4q ≥ q for q ≤ 1/5. So x 6∈ X + ∪ X − and thus x ∈ X 0 . The only other case is where xi = xj = 1 at locations where mi = 1 and mj = −1. Then we know x − ei − ej ∈ X 0 from the induction hypothesis. Pr [w¯h · x > 0] 6∈ [q, 1 − q] h∈V P rh∈V [wh · (x − ei − ej ) = 0] > 1 − 2q Pr [w¯h · x > 0] < q P rh∈V [wh,i = 1] > 1 − q h∈V Using the contradiction assumption on the test (1, x), Pr [wh · (1, x) > 0] 6∈ [q, 1 − q] P rh∈V [wh,j = −1] > 1 − q From these, h∈V P rh∈V [wh · x = 0] > 1 − 4q ≥ q Pr [1 + w¯h · x > 0] 6∈ [q, 1 − q] h∈V Pr [w¯h · x ≥ 0] 6∈ [q, 1 − q] h∈V and similarly, x ∈ X 0 . Pr [w¯h · x < 0] < q h∈V 6.7.7 Combining these, Probability Distribution We now define a probability distribution over x ∈ X ∗ . Pr [w¯h · x = 0] = 1 − Pr [w¯h · x > 0] − Pr [w¯h · x < 0] h∈V h∈V Without loss of generality, suppose m̄(+) ≥ m̄(−) . h∈V > 1 − 2q Thus, x 6∈ X + and x 6∈ X − imply x ∈ X 0 so the three sets are a partition. Definition 6.2. Define X ∗ to be every x ∈ {0, 1}d that (m̄ = 1) · x = (m̄ = −1) · x, where (m̄ = 1) is the element-wise boolean function. • Randomly draw an injection f : {i : m̄i = −1} → {i : m̄i = 1}. • Initialize x = 0d−1 • For indices {i bernoulli(1/2). : m̄i ≤ 0}, draw xi ∼ • For {i : m̄i = −1}, set xf (i) = xi Intuitively, this means that there are as many ones of x in positions where m̄ = 1 as there are places where m̄ = −1. Lemma 6.6. X ∗ ⊆ X 0 Note that the result x ∈ X ∗ because of the pairing f , there will be a 1 where m̄i = 1 for each 1 where m̄i = −1. Generalized Binary Search For Split-Neighborly Problems 6.7.8 Set T Definition 6.3. For the probability distribution, unless ∀i : m̄i = 0 : wt,i = 0 and ∀i : m̄i = −1 : wt,i + wt,f (i) = 0, call this condition(t, f ). Q(h) = Pr ∗ [wh · x = 0] Ef [1[condition(t, f )] + ... x∈X 1 1 1 (1 − 1[condition(t, f )])] > 1 − min( , ) 2 4 2r Lemma 6.7. Let T = {h ∈ V : Q(h) > 1 − 4q}, then |V | > 5|T |. 1 1 Pr[condition(t, f )] > 1 − min( , ) f 2 r Proof. For x ∈ X ∗ , since X ∗ ⊆ X 0 , Pr [wh · x = 0] > 1 − 2q h∈V P 1[wh · x = 0] > 1 − 2q |V | P X 1[wh · x = 0] > 1 − 2q P (x) h∈V |V | x∈X 0 P P h∈V x∈X 0 P (x)1[wh · x = 0] > 1 − 2q |V | P h∈V Q(h) > 1 − 2q |V | h∈V If m̄(−) = 0, then m̄(+) = 0, and thus ∀i : m̄i = 0 : wt,i = 0 so t = 0d and |T | = 1 ≤ 3. Note that Prf [condition(t, f )] > 1/2 implies that ∀i : mi = 0 : wt,i = 0. Lemma 6.9. If there exists i, j such that m̄i = m̄j = −1, then wt,i = wt,j . Proof. Prf [condition(t, f )] > 1 2 means that Pr[wt,i = −wt,f (i) ] > 1 2 Pr[wt,j = −wt,f (j) ] > 1 2 f |T | |V | − |T | (1) + (1 − 4q) > 1 − 2q |V | |V | f so 2|T | > |V | {l : m̄l = 1 ∧ wt,l = −wt,i } > 1/2 {l : m̄l = −1} {l : m̄l = 1 ∧ wt,l = −wt,j } > 1/2 {l : m̄l = −1} Lemma 6.8. |T | ≤ 3 Proof. Recall that m̄(+) ≤ rm̄(−) and m̄(−) ≤ rm̄(+) as well 1 1 1 ) since q ≤ min( 16 , 8r ) Also 1 − 4q ≥ 1 − min( 41 , 2r 1 For any t ∈ T , Q(t) > 1 − 4q ≥ 1 − min( 41 , 2r ). Define Ber(1/2) to be a Bernoulli random variable. Thus, there is some c ∈ {−1, 0, 1} such that ∀i : m̄i = 1 : wt,i = c. Pr[condition(t, f )] > 1 1 Pr ∗ [wt · x = 0] > 1 − min( , ) x∈X 4 2r X Ef [Pr[ wt,i Ber(1/2) + ... f 1 r Pr[∀i : m̄i = −1 : wt,f (i) = −c] > 1 − f i:mi =0 1 1 (wt,i + wt,f (i) )Ber(1/2) = 0]] > 1 − min( , ) 4 2r =−1 X i:mi which is only possible if wt,i = wt,j . 1 r 1 − Pr[∃i : m̄i = −1 : wt,f (i) 6= −c] > 1 − f Pr[∃i : m̄i = −1 : wt,f (i) 6= −c] < f 1 r 1 r Note that Suppose ∃j : m̄j = 1 : wt,j 6= −c, Pr[ X wt,i Ber(1/2) + ... Pr[∃i : m̄i = −1 : f (i) = j] = i:m̄i =0 X (wt,i + wt,f (i) )Ber(1/2) = 0] ≤ i:m̄i =−1 1 2 f 1 r which is a contradiction. So ∀j : m̄j = 1 : wt,j = −c. Stephen Mussmann, Percy Liang Thus, c completely determines t. Since there are three options for c, there are three options for t, and |T | ≤ 3. Since |T | ≤ 3 and 2|T | ≥ |V |, |V | ≤ 6. Thus, there is a split of 1/6 which is a contradiction since q ≤ 18 . Thus, the lemma is proved. And thus the binary linear classifier problem is split-neighborly.
8cs.DS
1 Incorporating Prior Information in Compressive Online Robust Principal Component Analysis arXiv:1701.06852v2 [cs.IT] 27 May 2017 Huynh Van Luong, Nikos Deligiannis, Jürgen Seiler, Søren Forchhammer, and André Kaup Abstract—We consider an online version of the robust Principle Component Analysis (PCA), which arises naturally in timevarying source separations such as video foreground-background separation. This paper proposes a compressive online robust PCA with prior information for recursively separating a sequences of frames into sparse and low-rank components from a small set of measurements. In contrast to conventional batch-based PCA, which processes all the frames directly, the proposed method processes measurements taken from each frame. Moreover, this method can efficiently incorporate multiple prior information, namely previous reconstructed frames, to improve the separation and thereafter, update the prior information for the next frame. We utilize multiple prior information by solving n-`1 minimization for incorporating the previous sparse components and using incremental singular value decomposition (SVD) for exploiting the previous low-rank components. We also establish theoretical bounds on the number of measurements required to guarantee successful separation under assumptions of static or slowlychanging low-rank components. Using numerical experiments, we evaluate our bounds and the performance of the proposed algorithm. In addition, we apply the proposed algorithm to online video foreground and background separation from compressive measurements. Experimental results show that the proposed method outperforms the existing methods. Index Terms—Prior information, compressed sensing, sparse signal, n-`1 minimization, compressive measurements. where k · k∗ is the matrix nuclear norm (sum of singular values). For example, in video separation, a video sequence is separated into slowly-changing background as the low-rank term L and foreground as the sparse term S, which is the quantity of interest in this paper. However, RPCA methods [1]–[3] need to access all data frames, e.g., all frames of a video, which is batch-based and not practical. Problem. We consider a version of RPCA for the recursive separation problem of a time-varying sequence of frames that processes one frame per time instance and works on a small set of measurements to separate large-scale data. In more detail, we introduce a time instance t, at which we have observed M t = Lt + S t , where S t = [x1 x1 ... xt ] and Lt = [v 1 v 2 ... v t ], where [·] denotes a matrix and xt , v t ∈ Rn are a column component in S t and Lt , respectively. In this work, we consider compressive measurements of the observed xt + v t at a time t, i.e., we observe y t = Φ(xt + v t ), where Φ ∈ Rm×n (m < n) is a random projection [4], [5], and that Lt−1 and S t−1 have been given for time instance t − 1. Following Problem (1), we want to solve a problem at time t as follows min [Lt−1 v t ] xt ,v t I. I NTRODUCTION P RINCIPAL Component Analysis (PCA) is widely used in statistical data analysis in a number of applications. A modified version of Principal Component Analysis (PCA), namely, robust PCA (RPCA) [1]–[3], is a well-known tool to efficiently capture properties of signals of interest. A potential application of RPCA is video foreground and background separation, e.g., in video surveillance and online tracking brain activation from MRI sequences. The goal of RPCA is to decompose the high-dimensional matrix M into the sum of unknown sparse S and low-rank L components. Principal Component Pursuit (PCP) [1] recovers L and S by solving the following program: min kLk∗ + λkSk1 subject to L + S = M , L,S (1) H. V. Luong, J. Seiler, and A. Kaup are with the Chair of Multimedia Communications and Signal Processing, Friedrich-Alexander-Universität Erlangen-Nürnberg, 91058 Erlangen, Germany (e-mail: [email protected], [email protected], and [email protected]). N. Deligiannis is with the Department of Electronics and Informatics, Vrije Universiteit Brussel, 1050 Brussels, and also with iMinds, 9050 Ghent, Belgium (e-mail: [email protected]). S. Forchhammer is with the Department of Photonics Engineering, Technical University of Denmark, 2800 Lyngby, Denmark (e-mail: [email protected]). ∗ + λkxt k1 subject to y t = Φ(xt + v t ), (2) where Lt−1 = [v 1 v 2 ... v t−1 ] and S t−1 = [x1 x1 ... xt−1 ] are given. We aim at solving Problem (2) given prior information Lt−1 , S t−1 , and Φ as formulated in Sec. II. Related Works. Some previous approaches [6]–[9] were proposed for online estimating low-dimensional subspaces from randomly subsampled data. However, they have not taken the sparse component xt into account. In addition, compressive ReProCS in [10], [11] was proposed to recover the sparse component for foreground extraction with compressive measurements. Compressive ReProCS [10] considers recovering xt from y t = Φt (Axt + Bv t ) [10]. It can be noted here that Φt ∈ Rn×n and A = B ∈ Rm×n . However, the background v t is not recovered after each instance and no explicit condition on how many measurements are required for successful recovery is considered. In this work, we focus on recovering both foreground and background components from compressive measurements. Recently, some related approaches to foreground extraction in [12], [13] utilizing an adaptive rate of the compressive measurement were introduced, yet they considered the background static. Another work in [14] proposed an incremental PCP method that processes one frame at a time, however it did use information of the full frame, rather than compressive measurements. Compressive PCP [15] considered the separation 2 on compressive measurements as a batch-based method. Furthermore, the problem of reconstructing a time sequence of sparse signals with prior information is also playing an important role in the context of online RPCA [10], [11], [16]. There were several studies on sparse signal recovery with prior information from low-dimensional measurements [10]– [12], [16], [17]. The study in [17] provided a comprehensive overview of the domain, reviewing a class of recursive algorithms for recovering a time sequence of sparse signals from a small number of measurements. The studies in [10], [11], [16] used modified-CS [18] to leverage prior knowledge under the condition of slowly varying support and signal values. However, all the above can not exploit multiple prior information gleaned from multiple previously recovered frames. Contributions. We propose a compressive online RPCA with multiple prior information named, CORPCA, which exploits the information for previously recovered sparse components by utilizing RAMSIA [19] and leverages the slowlychanging characteristics of low-rank components via an incremental SVD [20]. We perform the separation in an online manner by minimizing, (i) n-`1 function [19] for the sparse part; and (ii) the rank of a matrix of the low-rank part. Thereafter, the new recovered sparse and low-rank components are updated as a new prior knowledge for the next processing instance. Furthermore, we derive the number of measurements that guarantee successful compressive separation. As such, we set a bound on the number of required measurements based on the current foreground-background and prior information. The rest of this paper is organized as follows. Our problem formulation and the proposed algorithm are presented in Sec. II. We describe performance guarantees of the proposed method including the measurement bound and the proof sketch in Sec. III. We demonstrate numerical results of the proposed method in Sec. IV. II. P ROBLEM F ORMULATION AND A LGORITHM instance t, we observe y t = Φ(xt + v t ) with y t ∈ Rm . Let Z t−1 := {z 1 , ..., z J }, which is a set of z j ∈ Rn , and B t−1 ∈ Rn×d denote prior information for xt and v t , respectively. We can form the prior information Z t−1 and B t−1 based on the reconstructed {x1 , ..., xt−1 } and {v 1 , ..., v t−1 }, which is discussed in Sec. II-B. To solve the problem in (2), we fomulate the objective function of CORPCA as n 1 min H(xt , v t |y t , Z t−1 , B t−1 ) = kΦ(xt + v t ) − y t k22 xt ,v t 2 J o X , (4) + λµ βj kWj (xt − z j )k1 + µ [B t−1 v t ] ∗ j=0 where µ > 0. It can be seen that when v t is static, Problem (4) would become Problem (3). When xt and v t are batch variables without taking the prior information, Z t−1 and B t−1 , and the projection Φ into account, Problem (4) becomes Problem (1). B. CORPCA Algorithm We now describe how to solve Problem (4) and thereafter update the prior information, Z t−1 and B t−1 , for the next instance. First, we solve Problem (4) given Z t−1 and B t−1 . Let denote f (v t , xt ) = (1/2)kΦ(xt + v t ) − y t k22 , g(xt ) = Pus J λ j=0 βj kWj (xt − z j )k1 , and h(v t ) = k[B t−1 v t ]k∗ . The solution of (4) using proximal gradient methods [2], [21] gives (k+1) (k+1) at iteration k + 1 can be iteratively and v t that xt computed via the soft thresholding operator [21] for xt and the single value thresholding operator [22] for v t :  2o n  (k) (k) (k+1) (k) 1 , vt =arg min µh(v t )+ v t− v t − ∇vt f (v t , xt ) 2 2 vt n  o  2(5) (k+1) (k) 1 (k) (k) xt =arg min µg(xt )+ xt− xt − ∇xtf (v t , xt ) . 2 2 xt (6) In this section, we formulate explicitly our problem (in Sec. II-A), following by our proposed algorithm in Sec. II-B. The proposed CORPCA algorithm is described in Algorithm 1, where as suggested in [2], the convergence of Algorithm 1 is determined by k∂H(xt , v t )|x(k+1) ,v(k+1) k22 < A. Problem Formulation 2∗10−7 k(xt is defined by: The proposed CORPCA algorithm is based on RAMSIA [19], our previously proposed agorithm that uses n-`1 minimization with adaptive weights to recover a sparse signal using multiple side information signals. The objective function of RAMSIA [19] is given by: J o n X 1 βj kWj (x−z j )k1 , (3) min H(x) = kΦx−yk22 +λ x 2 j=0 where λ > 0, βj > 0 are weights among side information signals, and Wj is a diagonal weighting matrix for each side information z j , Wj = diag(wj1 , wj2 , ..., wjn ), wherein wji > 0 is the weight in Wj at index i for a given z j with z 0 = 0. The proposed CORPCA method aims at processing one frame per time instance and incorporating prior information for both sparse components and low-rank components. At time (k+1) (k+1) , vt t t )k22 and the proximal operator Γτ g1 (·) n o 1 Γτ g1 (X) = arg min τ g1 (V ) + ||V − X||22 . 2 V (7) Prior Information Update. Updating Z t and B t is carried out after each time instance so as to be used for subsequent processing. Because of the correlated characteristics of recursive frames, we can simply update prior information Z t by the J latest previous xt , i.e., Z t := {z j = xt−J+j }Jj=1 . For prior information B t ∈ Rn×d , we may have an adaptive update, where we can adjust the value d or keep a constant number d of the columns of B t . For each time instance, we use incremental singular decomposition SVD [20], namely, incSVD(·) in Algorithm 1. When we update B t = U t Γ µk g1 (Σt )V T t , the dimension of B t is increased, 2 where B t ∈ Rn×(d+1) , after each instance. However, we need to stay at a reasonable constant number of d, thus we update 3 Algorithm 1: The proposed CORPCA algorithm. Input: y t , Φ, Z t−1 , B t−1 ; Output: xt , v t , Z t , B t ; // Initialization. (−1) (0) (−1) (0) xt = xt = 0; v t = v t = 0; ξ−1 = ξ0 = 1; µ0 = 0; µ̄ > 0; λ > 0; 0 <  < 1; k = 0; g1 (·) = k · k1 ; while not converged do // Solving the problem. (k) (k−1) (k) −1 (v t − v t ); vet (k) = v t + ξk−1 ξk (k) (k−1) ξk−1 −1 (xt − xt ); ξk (k) (k) (k) (k) ft ) = ∇xt f (v ft ) ∇vt f(vet , x et , x (k) ft ) − y t ; ΦT Φ(vet (k) + x ft x (k) (k) = xt [U t−1 + = 2 vt = Θt (:, end);  (k+1) (k) (k) ft − 12 ∇xt f (vet (k) , x ft ) ; xt = Γ µk g x 2 where Γ µk g (·) is given as in RAMSIA [19]; 2 // Computing the updated weights [19]. (k+1) n(|xti − zji | + )−1 wji = P ; n (k+1) (|xtl − zjl | + )−1 l=1 −1  (k+1) (k+1) − zj )||1 +  ||Wj (xt βj = J  −1 ; P (k+1) (k+1) ||Wl (xt − zl )||1 +  l=0 p ξk+1 = (1 + 1 + 4ξk2 )/2; µk+1 = max(µk , µ̄) k = k + 1; end // Updating prior information. (k+1) Z t := {z j = xt−J+j }Jj=1 ; B t = U t (:, 1 : d)Γ µk g1 (Σt )(1 : d, 1 : d)V t (:, 1 : d)T ; 2 // Output. (k+1) (k+1) , vt , Z t, Bt; return xt  mt ≥ 2α log PJ et kδ t k2  T V t−1 · 0T n j=0 0 1  · Ve .  βj sj J  7 X βj sj + 1. (9) + 5 j=0 2) If v t is slowly-changing, i.e, B t−1 and [B t−1 v t ] are low rank. We solve the convex problem in (4) and let x̂t and v̂ t denote its solutions. Supposing that the measurement error of v t is bounded as kΦ(v̂ t − v t )k2 < σ, we have that kx̂t − xt k2 ≤ 1−2σ√ρ with 0 < ρ < 1 provided J    α n 7 X 3 log PJ + βj sj + , ρ 5ρ j=0 2ρ j=0 βj sj (10) sj J P 2 P βj w2 in both (9) and (10), wherein where α = 2 η̂ j=0 i=1 ji n P −1 o n 1  > 0 and η̂ = min n . |xti −zji |+ mt ≥ 2 B t = U t (:, 1 : d)Γ µk g1 (Σt )(1 : d, 1 : d)V t (:, 1 : d)T . The 2 advantage of using incSVD is that the computational cost of incSVD(·) is lower than conventional SVD [14], [20] since we only compute the full SVD of the middle matrix with size (d + 1) × (d + 1), where d  n, instead of n × (d + 1). The computation of incSVD(·) is elaborated below. Our goal is to compute incSVD[B t−1 v t ], i.e., [B t−1 v t ] = n×d U t Σt V T , we take the SVD t . Given matrix B t−1 ∈ R of B t−1 to obtain B t−1 = U t−1 Σt−1 V T t−1 . Therefore, we can derive (U t , Σt , V t ) via (U t−1 , Σt−1 , V t−1 ) and v t . We recast the matrix of [B t−1 v t ] by  δt Σt−1 ]· 0T kδ t k2 VT t−1 0T Theorem III.1. The problem in (4) requires mt measurements to successfully separate xt ∈ Rn with the sparse support s0 and v t ∈ Rn given compressive measurements y t = Φ(xt + v t ), where Φ ∈ Rmt ×n , whose elements are randomly drawn from an i.i.d. Gaussian distribution, and prior information including the prior Z t−1 := {z j }Jj=1 , where kxt − z j k0 = sj with z j ∈ Rn , and the prior B t−1 ∈ Rn×d . 1) If v t is not changing in time, we can recover xt successfully from (4) provided Θt = U t Γ µk g1 (Σt )V Tt ; [B t−1 v t ] = [U t−1  III. P ERFORMANCE G UARANTEES A bound is presented for the minimum number of measurements required for successful separation by solving the problem (4). This will also serve as a bound for the proposed CORPCA method. We assume that v t is either fixed or slowlychanging incurring an measurement error that is bounded by σ, i.e., kΦ(v̂ t − v t )k2 < σ, where v̂ t is any recovered background. Let s0 denote the support of the source x and sj denote the support of each difference vector x − z j ; namely, kx − z j k0 = sj , where j ∈ {0, . . . , J}. We shall derive the bound for recovering foreground xt given either fixed v t or the error bound σ. (U t , Σt , V ht ) =  i (k) ft ) incSVD B t−1 vet (k) − 12 ∇vt f (vet (k) , x ; (k+1) δt e , Σt = Σ, e and V t = ]·U kδ t k2 0 1  , (8) where et = U T v and δ = v − U e . By taking t t t−1 t t−1 t SVD of the matrix   in between the right side of (8), we yield Σt−1 et eΣ e Ve T . Eventually, we obtain U t = = U 0T kδ t k2 j i=1 Proof: We firstly derive the bound in (9) of Theorem III.1 for which the low-rank component v t is not changed and the remaining bound is a noisy case when the low-rank component v t is slowly-changing. We drive the bounds based on Proposition VII.1 (see Appendix VII-B3) by first computing the subdifferential ∂g(x) and then the distance, dist(·), between the standard normal vector g and ∂g(x). The u ∈ ∂g(x) of g(x) is derived through the separate components of the sum PJ g(x) = j=0 βj gj (x), where gj (x) = kWj (x − z j )k1 . As PJ a result, ∂g(x) = j=0 βj ∂gj (x). Considering the distance from the standard normal vector g to the subdifferential ∂g(x) given by (36), we derive a relation between this distance and all separate distances of gj (x) as 2 dist (g, τ · ∂g(x)) ≤ J X j=0 βj dist2(g, τ · ∂gj (x)), (11) 4 where PJ j=0 βj = 1. Taking the expectation of (11) delivers Eg [dist2 (g, τ ·∂g(x))] ≤ J X βj Eg [dist2 (g, τ ·∂gj (x))]. (12) j=0 1) Distance Expectation: We now consider each component Eg [dist2(g, τ · ∂gj (x))], where gj (x) = kWj (x − z j )k1 . We firstly consider x − z j . Without loss of generality, we have assumed that the sparse vector x − z j has sj supports or kx − z j k0 = sj , i.e, taking the form x − z j = (x1 − zj1 , ..., xsj − zjsj , 0, ..., 0). We have the subdifferential u ∈ ∂gj (x) of gj (x) given by:  ui = wji sign(xi − zji ), i = 1, ..., sj (13) |ui | ≤ wji , i = sj + 1, ..., n. Let us recall the weights as proposed in RAMSIA [19] for aPspecific z j with wji = ηj /(|xi − zji | + ) and the constraint n i=1 wji = n. Taking the source x − z j to derive  ηj i = 1, ..., sj |xi −zji |+ , (14) wji = ηj i = sj + 1, ..., n.  , Pn From the constraint w = n, we can derive: ηj = P −1 i=1 ji n 1 . n |xi −zji |+ i=1 We can then compute the distance from the standard normal vector g to the subdifferential ∂gj (x) based on (36) as dist2 (g, τ · ∂gj (x)) = + sj  X gi − τ wji sign(xi − zji ) i=1 n X  max(|gi | − τ wji , 0) 2 2 (15) , i=sj +1 where max(a, 0) returns the maximum value between a ∈ R and 0. Further, after taking the expectation of (15), we obtain 2 Eg [dist (g,τ · ∂gj (x))] = sj + τ 2 sj X 2 wji i=1 r + (16) Z ∞ n 2 X −v 2 /2 (v − τ wji )e dv. π i=s +1 τ wji ψ(τ η /) ψ(x)/x decreases with x > 0. Thus τ ηjj/ Considering the last term in (18), we derive 2 Eg [dist (g, τ · ∂g(x))] ≤ J X βj sj + τ j=0  +2 n− 2 J X βj j=0 J X j=0 βj sj ψ(τ η b/) τη b/ . ≤ sj X 2 wji i=1  ψ(τ ηb/) τ ηb/ (19) . PJ Inequality (19) holds due to j=0 βj (n − sj ) = n − PJ j=0 βj sj . PJ We denote s̄ = j=0 βj sj and from (39), (19), we have sj J n X X ψ(τ ηb/) o 2 2 Ug = min s̄ + τ . (20) βj wji + 2(n − s̄) τ ≥0 τ ηb/ j=0 i=1 Inserting (40) in (20) gives sj 2 2 2 J n X X 2 e(−τ ηb /2 ) o 2 Ug ≤ min s̄+τ 2 βj wji +(n− s̄) √ . τ ≥0 τ ηb/ 2π j=0 i=1 (21) To give a bound as a function of a given source x and other parameters, we can select a parameter p τ > 0 to obtain an useful bound in (21). Setting τ = (/b η ) 2 log(n/s̄) gives sj J n 2 X X 2 s̄(1 − s̄/n) Ug ≤ s̄ + 2 log( ) 2 βj . (22) w +p s̄ ηb j=0 i=1 ji π log(n/s̄) Psj 2 PJ 2 We denote α = ηb 2 j=0 βj i=1 wji in the second term in (22). Finally, applying the inequality (41) to the last term in (22) gives n 7 (23) Ug ≤ 2α log( ) + s̄. s̄ 5 As a result, from Proposition VII.1 (in Appendix VII), we get the bound in (9) as n 7 mt ≥ 2α log( ) + s̄ + 1. (24) s̄ 5 For slowly-changing v t , we assume that the measurement error kΦ(v̂ t −v t )k2 < σ. Applying Proposition VII.1 (see Appendix VII) for the noisy case, we get the bound in (10). j We use the inequality (42) for the third term in (16) to get sj X n X ψ(τ wji ) . τ wji i=1 i=sj +1 (17) 2) Bound Derivation: Inserting (17) in (12) for all functions gj (x) gives Eg [dist2 (g, τ · ∂gj (x))] ≤ sj + τ 2 2 Eg [dist (g, τ · ∂g(x))] ≤ J X 2 wji +2 βj sj + τ j=0 J X 2 J X j=0 βj sj X i=1 2 wji (18) ψ(τ ηj /) +2 , βj (n − sj ) τ ηj / j=0 where wji = ηj / in (14) are used in the third term of (18). Let us denote ηb = min{ηj }. It can be noted that the function j IV. N UMERICAL E XPERIMENT We evaluate the performance of our Algorithm 1 employing the proposed method, the existing `1 minimization [21], and the existing `1 -`1 [23], [24] minimization methods. We also compute the established bound (10) in Theorem III.1, which serves as an estimate of number of measurements that CORPCA requires for successful recovery. We generate our data as follows. We generate the low-rank component L = U V T , where U ∈ Rn×r and V ∈ R(d+q)×r are random matrices whose entries are drawn from the standard normal distribution. We set n = 500, r = 5 (rank of L), and that d = 100 is the number of vectors for training and q = 100 is the number of testing vectors. This yields L = [x1 ... xd+q ]. We generate S = [v 1 ... v d+q ], where at time instance t − 1, xt−1 ∈ Rn is generated from the standard normal distribution with support s0 . Our purpose is to consider 5 8 8 8 Foreground 30 50 50 30 Bound mt 0 10 150 250 350 450 Measurements m 30 0 150 250 350 450 Measurements m 30 10 100 90 90 70 50 50 30 Nonzeros s0 50 100 Pr(success) [%] Nonzeros s0 70 0 10 150 250 350 450 Measurements m 100 50 Background 90 Nonzeros s0 50 50 0 10 150 250 350 450 Measurements m Background 70 Background 100 Pr(success) [%] 10 Bound m`1 Pr(success) [%] 50 70 90 Nonzeros s0 70 Foreground 100 Bound m`1 !`1 Pr(success) [%] Nonzeros s0 Nonzeros s0 90 Pr(success) [%] 100 90 70 50 50 30 50 0 10 Pr(success) [%] Foreground 0 150 250 350 450 Measurements m 150 250 350 450 Measurements m 50 70 50 50 TABLE I 50 Pr(success) [%] 50 Pr(success) [%] 50 70 Nonzeros s0 70 Pr(success) [%] Nonzeros s0 (b) `1 -`Foreground (c) `1 Foreground minimization (d) Scale 1 minimization 100 100 100 Fig. 1. The90 averages of success probabilities and bounds (a)-(c), to recover xt ∈ R500 , where the x-axis is the 90 of CORPCA with different minimizations 90 number of measurements m and the y-axis is the number of nonzeros s0 . The gray scale (d) is proportional to Pr(sussess) from pure black to white. Nonzeros s0 (a) n-`1Foreground minimization a sequence of correlated sparse vectors xt . Therefore, we S UMMARY OF RECOVERY CHARACTERISTICS 30 This could lead generate x30 30 t satisfying kxt − xt−1 k0 = s0 /2. Bound mt Bound mt GRASTA ReProCS CORPCA RPCA to kxt k0 > kx k , we set theBound mt 10s0 . To avoid a large increase of10 0 t 0 0 10 0 [1] [7] [10] constraint kxt k0 150 ∈ [s0250 , s0 +350 15], 450 whenever kxt k0150 > s0250 + 15,350 450 150 250 350 450 Online X X X Measurements xt is randomly reset to kxt k0 m = s0 by setting kx k − s Measurements m Measurements m t 0 0 Full data X X X X positions that are randomly selected to zero. In this work, we Background Background Background Foreground X X 100 test our algorithms for s0 = 10 to 90. 100 100 Compressed 90 information is initialized as follows. 90 X 90 Background X Our prior To illustrate real scenarios, 70 where we do not know the sparse 70 and low-rank 70 components, we use the batch-based RPCA 50 [1] to separate the employing n-` minimization: that, at certain 50 sparsity degrees, 50 1 50 50 training set M 0 = [x1 + v 1 ... xd + v d50 ] to obtain B 0 = we can recover the 500-dimensional data from measurements [v 1 ... v d 30 ]. In this experiment, we use three 30 (J = 3) sparse of much lower 30 dimesions (m = 150 to 300, see white areas components as prior information and we set Z 0 := {0, 0, 0}. in Fig. 1(a)). Furthermore, the measurement bound mt (red 10 0 10 01(a)) is consistent with the practical 0 We run CORPCA (Sec. the test set M = [xd+1 + line in Fig. 10 results. It is 150 250II-B) 350on450 150 250 350 450 150 250 350 450 v d+1 ... xd+q + v ]. also clear that the ` and ` -` minimization methods (see d+q Measurements m 1 1 1 m Measurements m Measurements bt versus xt in terms Figs. 1(b)+1(c)) lead to a higher number of measurements, We assess the accuracy of recovery x of the success probability, denoted as Pr(success), versus thereby illustrating the benefit of incorporating multiple side the number of measurements m. At time instance t, the information into the problem. original source xt + z t is projected onto lower-dimensional V. V IDEO F OREGROUND -BACKGROUND S EPARATION observations y t having m measurements. For a fixed m, the Pr(success) is the number of times, in which the source xt We assess our CORPCA method in the application of bt with an error kb is recovered as x xt − xt k2 /kxt k2 ≤ 10−2 , compressive video separation and compare it against existing divided by the total 50 Monte Carlo simulations and where we methods. We run all methods listed in Table I on typical test √ have set parameters  = 0.8, λ = 1/ n. In addition, we will video content [25]. In this experiment, we use d = 100 frames evaluate the obtained bound mt (10), m`1 -`1 (29), and m`1 (27) as training vectors for the proposed CORPCA as well as for [23] in the presence of noisy measurements kΦ(v̂ t −v t )k2 < σ GRASTA [7] and ReProCS [10]. in the case of slowly-changing low-rank components v t . As 1) Visual Evaluation: We first consider backgroundshown in (10) of Theorem III.1, these bounds depend on ρ. foreground video separation with full access to the video data We experimentally select ρ ∈ (0, 1) to give that our estimated (the data set M ); the visual results of the various methods bound is consistent with the practical results, here we have are illustrated in Fig. 2. It is evident that, for both the video 0.6 0.4 set ρ to values of 0.8 sequences, CORPCA delivers superior visual results than the 3 ; 3 ; 3 ; and for mt m`1-`1 , and m`1 , respectively. other methods, which suffer from less-details in the foreground The results in Fig. 1(a) show the efficiency of CORPCA and noisy background images. (c) CORPCA-n-`1 150 250 350 450 Measurements m Fig. 1. Average success probabilities and bounds for the proposed CORPCA algorithm deploying the different minimizations (a)-(c), to recover xt , v t ∈ R500 . The scale (d) is proportional to Pr(success)[%] from black to white. (e) CORPCA-`1 J n o X 1 min H(x)= kΦx − yk22 + λ βj kWj (x − z j )k1 , x 2 j=0 (f) ReProCS where λ > 0, βj >0 are weights across the side information signals, and Wj is a diagonal matrix with weights for each element in the side information signal z j ; namely, Wj = diag(wj1 , wj2 , ..., wjn ) (g) GRASTA 10 (h) Scale 150 250 350 450 Measurements m 010 (e) CORPCA-`1 Pr(success) [%] Nonzeros s0 Bound m!`1 150 250 350 450 Measurements m Fig. 1. Average success probabilities and bounds for the proposed CORPCA algorithm deploying the different minimizations (a)-(c), to recover xt , v t ∈ R500 . The scale (d) is proportional to Pr(success)[%] from black to white. (f) ReProCS 50 30 60 40 20 010 70 50 50 50 30 10 0 150 250 350 450 Measurements m (g) GRASTA 100 90 70 150 250 350 450 Measurements m (h) Scale 30 Foreground 010 90 Bound m!`1 150 250 350 450 Measurements m Fig. 1. Average success probabilities and bounds for the proposed CORPCA algorithm deploying the different minimizations (a)-(c), to recover xt , v t ∈ R500 . The scale (d) is proportional to Pr(success)[%] from black to white. 70 50 J n o X 1 min H(x)= kΦx − yk22 + λ βj kWj (x − z j )k1 , x 2 j=0 (3) where λ > 0, βj >0 are weights across the side information signals, and Wj is a diagonal matrix with weights for each element in the side information signal z j ; namely, Wj = diag(wj1 , wj2 , ..., wjn ) J n o X 1 min H(x)= kΦx − yk22 + λ βj kWj (x − z j )k1 , x 2 j=0 (3) (f) ReProCS 50 30 80 60 40 20 010 0 150 250 350 450 Measurements m (h) Scale Fig. 4. Average success probabilities for CORPCA (for xt , v t ), ReProCS (for xt ), and GRASTA (for v t ). The scale (f) is proportional to Pr(success)[%] Fig. 4. Average success probabilities for CORPCA (for xt , v t ), ReProCS (for xt ), and GRASTA (for v t ). The scale (f) is Fig. proportional to Pr(success)[%] 4. Average success probabilities for CORPCA (for xt , v t ), ReProCS (for xt ), and GRASTA (for v t ). The scale (f) is proportional to Pr(success)[%] from black to white. from black to white. from black to white. (3) (e) CORPCA-`1 Background 100 Nonzeros s0 70 50 Pr(success) [%] 30 Nonzeros s0 0 150 250 350 450 Measurements m 30 0 90 Pr(success) [%] 010 70 50 50 50 80 Nonzeros s0 20 90 Background Background 100 Pr(success) [%] 40 Foreground Nonzeros s0 30 70 Pr(success) [%] 50 100 90 90 30 150 250 350 450 Measurements m Nonzeros s0 010 50 Bound m!`1 60 50 50 10 Background 80 70 Pr(success) [%] 150 250 350 450 Measurements m 70 Nonzeros s0 10 30 90 0 150 250 350 450 Measurements m Pr(success) [%] 30 100 Nonzeros s0 70 50 50 50 30 Background Pr(success) [%] Nonzeros s0 Nonzeros s0 70 Foreground Pr(success) [%] 100 90 90 50 50 10 Background 100 90 Pr(success) [%] 0 150 250 350 450 Measurements m 70 Pr(success) [%] 10 90 Nonzeros s0 30 (d) CORPCA-`1 -`1 Foreground 100 Pr(success) [%] Nonzeros s0 50 50 Pr(success) [%] (d) CORPCA-`1 -`1 Foreground 100 70 Pr(success) [%] Nonzeros s0 (d) CORPCA-`1 -`1 Foreground 90 (c) CORPCA-n-`1 Pr(success) [%] Nonzeros s0 Nonzeros s0 (c) CORPCA-n-`1 where λ > 0, βj >0 are weights across the side information signals, and Wj is a diagonal matrix with weights for each element in the side information signal z j ; namely, Wj = diag(wj1 , wj2 , ..., wjn ) (g) GRASTA 6 Original Original CORPCA CORPCA CORPCA CORPCA CORPCA CORPCA GRASTA RPCA GRASTA RPCA GRASTA RPCA (b) Bootstrap GRASTA RPCA GRASTA RPCA (a) Bootstrap (b) Bootstrap RPCA ReProCS ReProCS ReProCS ReProCS ReProCS ReProCS GRASTA m/n==0.8 0.8 m/n m/n==0.8 0.8 m/n m/n==0.6 0.6 m/n m/n==0.6 0.6 m/n m/n = 0.8 m/n = 0.8 m/n==0.4 0.4 m/n m/n==0.4 0.4 m/n m/n = 0.6 m/n = 0.6 m/n==0.2 0.2 m/n m/n==0.2 0.2 m/n m/n = 0.4 m/n = 0.4 (a) Bootstrap m/n = 0.2 m/n = 0.2 Original Original CORPCA CORPCA CORPCA CORPCA RPCA RPCA RPCA RPCA GRASTA GRASTA GRASTA (c) Curtain GRASTA ReProCS ReProCS ReProCS ReProCS Fig. 2. (c) Curtain Background and foreground separation for the different separation methods with full data of frames Bootstrap#2210 and Curtain#2866. Fig. 2. Background and foreground separation for the different separation methods with full data of frames Bootstrap#2210 and Curtain#2866. CORPCA CORPCA (b) Curtain RPCA GRASTA ReProCS established the theoretical bounds on the number of required meaRPCA GRASTA ReProCS surements to guarantee successful recovery. Furthermore, we have defined as in [?]. For comparing defined as in [?]. CORPCA to RPCA, GRASTA, and ReProCS, Fig. 5 For illustrates the ROC results for full GRASTA, data access, m/n = 1, comparing CORPCA to RPCA, andi.e., ReProCS, Fig. RPCA,the GRASTA, and ReProCS support separation 5since illustrates ROC results for full can datanot access, i.e.,them/n = 1, for both foregrounds the compressive measince RPCA, GRASTA,and andbackgrounds ReProCS canfrom not support the separation surements. The resultsand of CORPCA are from betterthe thancompressive all the remaining for both foregrounds backgrounds meamethods, especially, it is improved for all Curtain in Fig. surements. The results of significantly CORPCA are better than the remaining 5(b). Theespecially, ROC results for CORPCA, CORPCA-n-` 1 , is depicted methods, it is significantly improved for Curtain in Fig. in Figs. andresults ?? forforBootstrap and Curtain 5(b). The ?? ROC CORPCA, CORPCA-n-` depicted 1 , issequences, respectively. CORPCA-n-` performances with 1 still maintain in Figs. ?? and ?? for Bootstrap andgood Curtain sequences, the small number of measurements, e.g.,good for performances Bootstrap with until respectively. CORPCA-n-` 1 still maintain m/nsmall = 0.6 [see Fig. ??] and for Curtain until m/n =until 0.4 the number of measurements, e.g., for Bootstrap [see Fig. Furthermore, ReProCSuntil fromm/n compressive m/n = ??]. 0.6 [see Fig. ??] we andsimulate for Curtain = 0.4 measurements for obtainingweonly foreground andfrom comprared with [see Fig. ??]. Furthermore, simulate ReProCS compressive CORPCA withfor different measurement rates as and shown in Fig. 6.with The measurements obtaining only foreground comprared ROC results fordifferent ReProCSmeasurement are quickly degradated for the CORPCA with rates as shown in compressive Fig. 6. The measurements with aare high rate m/n = 0.8.for the compressive ROC results foreven ReProCS quickly degradated measurements even with a high rate m/n = 0.8. IV. C ONCLUSION C ONCLUSION This paper proposed aIV. compressive online robust PCA algorithm established the theoretical bounds on the number of required meaFig. 2. Background and foreground separation forresults thethatdifferent separation presented numerical showed the Furthermore, efficiency of CORPCA surements to guarantee successful recovery. we have and the consistency of the theoretical presented numerical results that showedbounds. the efficiency of CORPCA methods with full data access Bootstrap#2213 and Curtain#2866. and the consistency of the theoretical bounds. Fig. 3 presents the results of CORPCA under various rates on the number of measurements m over the dimension n of the data (the size of the vectorized frame). The results show that we can recover the foreground and background even by accessing a small number of measurements; for instance, we can obtain good-quality reconstructions with only m/n = 0.6 and m/n = 0.4 for Bootstrap [see Fig. 3(a)] and Curtain (CORPCA) can process a data vector timePCA instance from This paperthat proposed a compressive online per robust algorithm [see Fig. respectively. Bootstrap requires more meacompressive measurements. efficiently incorporates (CORPCA) that can 3(b)], processCORPCA a data vector per time instance multifrom ple prior information based CORPCA on the n-`efficiently minimization problem. We compressive measurements. incorporates multisurements than dueWeto the more complex foreground ple prior information based on theCurtain n-` minimization problem. information. For comparison, we illustrate the visual results obtained with ReProCS—which, however, can only recover the foreground using compressive measurements—in Fig. 4. It is clear that the reconstructed foreground images have a poorer visual quality compared to CORPCA even at a high rate m/n = 0.81 . We consider two videos [25], Bootstrap (60×80 pixels) and Curtain (64×80 pixels) [c.f., Fig. 2], having a static and a dynamic background, respectively. 2) Quantitative Results: We evaluate quantitatively the separation performance via the receiver operating curve (ROC) metric [27]. The metrics True positives and False positives are defined as in [27]. The ROC results for CORPCA deploying different recovery algorithms, CORPCA-n-`1 , CORPCA-`1 -`1 , and CORPCA`1 , are depicted in Fig. 5 for Bootstrap and Curtain sequences. It is clear that CORPCA-`1 -`1 [see Figs. 5(b), 5(e)] and CORPCA-`1 [see Figs. 5(c), 5(f)] give worse curves compred with CORPCA-n-`1 [see Figs. 5(a), 5(d)]. Specifically, CORPCA-n-`1 gives high performances with a small number of measurements, e.g., for Bootstrap until m/n = 0.6 [see Fig. 5(a)] and for Curtain until m/n = 0.4 [see Fig. 5(d)]. Fig. 6 illustrates the ROC results when assuming full data access, i.e., m/n = 1, of CORPCA, RPCA, GRASTA, and ReProCS. The results show that CORPCA delivers higher performance than the other methods, especially for the Curtain video sequence [c.f., Fig. 6(b)]. Furthermore, we compare the foreground recovery performance of CORPCA against 0.8 m/n==0.8 m/n 0.8 m/n==0.8 m/n m/n==0.6 0.6 m/n m/n==0.6 0.6 m/n m/n = 0.8 m/n = 0.8 m/n==0.2 0.2 m/n m/n==0.2 0.2 m/n m/n==0.4 0.4 m/n m/n==0.4 0.4 m/n m/n = 0.6 m/n = 0.6 m/n = 0.4 m/n = 0.4 m/n = 0.2 m/n = 0.2 (b) Curtain Fig. 3. Compressive background and foreground separation of CORPCA with different measurement rates m/n. 0.8 m/n==0.8 m/n 0.8 m/n==0.8 m/n 0.6 m/n==0.6 m/n 0.6 m/n==0.6 m/n m/n = 0.8 0.8 m/n m/n ==0.8 m/n==0.2 0.2 m/n m/n==0.2 0.2 m/n m/n==0.4 0.4 m/n m/n==0.4 0.4 m/n m/n = 0.6 m/n 0.6 m/n= =0.6 m/n = 0.4 m/n = =0.4 m/n 0.4 m/n = 0.2 m/n 0.2 m/n== 0.2 (a) Bootstrap 1 1 1 The original test videos and the reconstructed separated sequences are available online [26]. m/n = 0.8 m/n = 0.6 m/n = 0.4 m/n = 0.2 (b) Curtain Fig. 4. Compressive foreground separation of ReProCS with different measurement rates m/n. ReProCS for different compressive measurement rates: m/n = {0.8; 0.6; 0.4; 0.2}. The ROC results in Fig. 7 show that CORPCA achieves a relatively high performance with a small number of measurements. VI. C ONCLUSION This paper proposed a compressive online robust PCA with prior information for recursive separation. The proposed CORPCA method can process one data frame per time instance from compressive measurements. CORPCA efficiently incorporates multiple prior information based on the n-`1 minimization problem. To express the performance, we established the theoretical bounds on the number of required measurements to guarantee successful recovery. The numerical results showed the efficiency of CORPCA in terms of both the practical results and the theoretical bounds. The results also revealed the advantage of incorporating prior information by employing n-`1 minimization to achieve significant improvements over that of using the existing `1 -`1 and `1 minimization methods. We also test our method on the compressive video separation application using video data. The results revealed the advantage of incorporating prior information by employing n-`1 minimization and demonstrated the superior performance 7 ROC curve of dataset -- bootstrap ROC curve of dataset -- bootstrap 1 0.4 CORPCA-n-`1 : CORPCA-n-`1 : CORPCA-n-`1 : CORPCA-n-`1 : CORPCA-n-`1 : 0.4 0.3 0.1 0 0 0.1 0.2 0.3 CORPCA-`1 -`1 : CORPCA-`1 -`1 : CORPCA-`1 -`1 : CORPCA-`1 -`1 : CORPCA-`1 -`1 : 0.2 0 0.02 (a) Bootstrap: CORPCA-n-`1 0.04 1 0.8 0.8 CORPCA-n-`1 : CORPCA-n-`1 : CORPCA-n-`1 : CORPCA-n-`1 : CORPCA-n-`1 : m=n = 1 m=n = 0:8 m=n = 0:6 m=n = 0:4 m=n = 0:2 0.2 0 0.04 0.06 0.1 0.08 0.1 0 0.12 0 0.002 0.004 0.006 0.008 ROC curve of dataset -- curtain 0.6 0.4 CORPCA-`1 -`1 : CORPCA-`1 -`1 : CORPCA-`1 -`1 : CORPCA-`1 -`1 : CORPCA-`1 -`1 : m=n = 1 m=n = 0:8 m=n = 0:6 m=n = 0:4 m=n = 0:2 0.4 0 0.005 0.01 0.015 0.02 0.025 CORPCA-`1 : CORPCA-`1 : CORPCA-`1 : CORPCA-`1 : CORPCA-`1 : 0.3 (d) Curtain: CORPCA-n-`1 0.1 0 0.03 0 0.5 1 (f) Curtain: CORPCA-`1 ROC for CORPCA deploying different sparse recovery algorithms with different measurement rates for Bootstrap. ROC curve of dataset -- curtain ROC curve of dataset -- bootstrap 1 0.8 True positives True positives 0.9 0.8 0.7 0.6 CORPCA RPCA GRASTA ReProCS 0.5 0.6 0.4 CORPCA RPCA GRASTA ReProCS 0.2 0.4 0.4 0.6 0.8 0 0.02 False positives 0.06 0.08 (b) Curtain ROC for the different separation methods with full data. ROC curve of dataset -- curtain 1 1 0.8 0.8 0.6 CORPCA-m=n = 0:8 CORPCA-m=n = 0:6 CORPCA-m=n = 0:4 CORPCA-m=n = 0:2 ReProCS-m=n = 0:8 ReProCS-m=n = 0:6 ReProCS-m=n = 0:4 ReProCS-m=n = 0:2 0.4 0.2 0 True positives True positives ROC curve of dataset -- bootstrap CORPCA-m=n = 0:8 CORPCA-m=n = 0:6 CORPCA-m=n = 0:4 CORPCA-m=n = 0:2 ReProCS-m=n = 0:8 ReProCS-m=n = 0:6 ReProCS-m=n = 0:4 ReProCS-m=n = 0:2 0.6 0.4 0.2 0 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 False positives (a) Bootstrap Fig. 7. 0.04 False positives (a) Bootstrap Fig. 6. ROC for CORPCA and ReProCS with compressive measurements. 0 1.5 False positives (e) Curtain: CORPCA-`1 -`1 1 0.02 0.04 m=n = 1 m=n = 0:8 m=n = 0:6 m=n = 0:4 m=n = 0:2 0.2 False positives 0.2 0.012 0.014 (c) Bootstrap: CORPCA-`1 0.6 False positives 0 0.01 False positives 0 0.12 m=n = 1 m=n = 0:8 m=n = 0:6 m=n = 0:4 m=n = 0:2 0.5 0.2 0.02 0.08 ROC curve of dataset -- curtain True positives True positives 0.06 CORPCA-`1 : CORPCA-`1 : CORPCA-`1 : CORPCA-`1 : CORPCA-`1 : 0.1 (b) Bootstrap: CORPCA-`1 -`1 ROC curve of dataset -- curtain 0 0.2 False positives 1 0.4 0.3 0 0.4 False positives 0.6 m=n = 1 m=n = 0:8 m=n = 0:6 m=n = 0:4 m=n = 0:2 True positives 0.2 m=n = 1 m=n = 0:8 m=n = 0:6 m=n = 0:4 m=n = 0:2 True positives 0.6 0.4 0.5 True positives True positives 0.5 0.6 0.8 Fig. 5. ROC curve of dataset -- bootstrap 0.7 0.06 0.08 False positives (b) Curtain 0.1 0.12 2 #10 -3 8 improvement offered by CORPCA compared to the existing methods. VII. A PPENDIX A. Prior Art in Compressed Sensing The problem of recursive signal recovery employs Compressed Sensing (CS) [4], [5] to recover sparse components from low-rank components [12], [17], which is considered as a dynamic CS. Therefore, we review some related CS results before stating our contributions. Let x ∈ Rn denote a highdimensional sparse vector. The source x can be reduced by sampling via a linear projection at the encoder [4]. We denote a random measurement matrix by Φ ∈ Rm×n (with m  n), whose elements are sampled from an i.i.d. Gaussian distribution. Thus, we get a measurement vector y = Φx, with y ∈ Rm . At the decoder, x can be recovered by solving the Basis Pursuit problem [4], [5]: min kxk1 subject to y = Φx, (25) x Pn where kxkp := ( i=1 |xi |p )1/p is the `p -norm of x wherein xi is an element of x. Problem (25) becomes an instance of finding a general solution: min{H(x) = f (x) + g(x)}, x (26) where f := Rn → R is a smooth convex function and g := Rn → R is a continuous convex function, which is possibly non-smooth. Problem (25) emerges from Problem (26) when setting g(x) = λkxk1 and f (x) = 12 kΦx − yk22 , with Lipschitz constant L∇f [21]. We consider how many measurements are required for successful recovery. The classical `1 minimization problem in CS [4], [5], [28] requires m`1 measurements [23], [24], [29] for successful reconstruction, bounded as n 7 m`1 ≥ 2s0 log (27) + s0 + 1, s0 5 where s0 := kxk0 = |{i : xi 6= 0}| denotes the number of nonzero elements in x as the support of x, with |.| denoting the cardinality of a set and k · k0 being the `0 -pseudo-norm. Furthermore, the `1 -`1 minimization approach [23], [24], [30] reconstructs x given a signal z ∈ Rn as side information by solving the following problem: o n1 min kΦx − yk22 + λ(kxk1 + kx − zk1 ) . (28) x 2 The bound on the number of measurements required by Problem (28) to successfully reconstruct x depends on the quality of the side information signal z as [23], [24], [30]   7 n ξ m`1 -`1 ≥ 2h log + s0 + + 1, (29) s0 + ξ/2 5 2 where ξ : = |{i : zi 6= xi = 0}| − |{i : zi = xi 6= 0}| h̄ : = |{i : xi > 0, xi > zi } ∪ {i : xi < 0, xi < zi }|, wherein xi , zi are corresponding elements of x, z. It has been shown that Problem (28) improves over Problem (25) provided that the side information has good enough quality [23], [24]. B. Background on Measurement Condition We restate some key definitions and conditions in convex optimization [29], [31], which are used in the derivation of the measurement bounds. A convex cone C ⊂ Rn is a convex set that satisfies C = τ C, ∀τ ≥ 0 [31]. For the cone C ⊂ Rn , a polar cone C ◦ is the set of outward normals of C, defined by C ◦ := {u ∈ Rn : uT x ≤ 0, ∀x ∈ C}. (31) A descent cone [31, Definition 2.7] D(g, x), alias tangent cone [29], of a convex function g := Rn → R at a point x ∈ Rn —at which g is not increasing—is defined as [ D(g, x) := {y ∈ Rn : g(x + τ y) ≤ g(x)}, (32) τ ≥0 S where denotes the union operator. Recently, a new summary parameter called the statistical dimension δ(C) of cone C [31] is introduced to estimate the convex cone [31, Theorem 4.3]. The statistical dimension can be expressed in terms of the polar cone C ◦ as [31, Proposition 3.1]  δ(C) := Eg dist2 (g, C ◦ )]. (33) 1) Gaussian Width: The Gaussian width [29] is a summary parameter for convex cones; it is used to measure the aperture of a convex cone. For a convex cone C ⊂ Rn , considering a subset C ∩ Sn−1 where Sn−1 ⊂ Rn is a unit sphere, the Gaussian width [29, Definition 3.1] is defined as ω(C) := Eg [ sup g T u]. (34) u∈C∩Sn−1 where g ∼ N(0, In ) is a vector of n independent, zero-mean, and unit-variance Gaussian random variables and Eg [·] denotes the expectation with respect to g. The Gaussian width [29, Proposition 3.6] can further be bounded as  ω(C) ≤ Eg dist(g, C ◦ )], (35) where dist(g, C ◦ ) denotes the Euclidean distance of g with respect to the set C ◦ , which is in turn defined as dist(g, C ◦ ) := min{kg − uk2 : u ∈ C ◦ }. u (36) 2) Measurement Condition: An optimality condition [29, Proposition 2.1], [31, Fact 2.8] for linear inverse problems states that x0 is the unique solution of (26) if and only if D(g, x0 ) ∩ null(Φ) = {0}, (37) where null(Φ) := {x ∈ Rn : Φx = 0} is the null space of Φ. We consider the number of measurements m required to successfully reconstruct a given signal x0 ∈ Rn . Corollary 3.3 in [29] states that, given a measurement vector y = Φx0 , x0 is the unique solution of (26) with probability √ at least 1 − exp(− 12 ( m − ω(D(g, x0 )))2 ) provided that m ≥ ω 2 (D(g, x0 )) + 1. 9 3) Bound on the Measurement Condition: Let us consider that the subdifferential ∂g [32] of a convex function g at a point x ∈ Rn is given by ∂g := {u ∈ Rn : g(y) ≥ g(x) + uT (y − x) for all y ∈ Rn }. From (33), (35), and [31, Proposition 4.1], we obtain an upper bound on δ(D(g, x)) as   ω 2 (D(g, x0 )) ≤ δ(D(g, x)) ≤ min Eg dist2 (g, τ · ∂g(x)) . τ ≥0 (38) We conclude the following proposition for determine the measurement bound. Proposition VII.1. [29, Corrollary 3.3, Proposition 3.6] We have the successful recovery x ∈ Rn in (26) that we have observed measurements y = Φx, where Φ ∈ Rm×n , provided the condition m ≥ Ug + 1. For the noisy measurements ky − Φxk2 < σ, we have that kx − x̂k2 < 1−2σ√ρ provided the U +3/2 condition m ≥ g ρ , where x̂ is any solution in (26) and 0 < ρ < 1. The quantity Ug is calculated given a convex norm function g := Rn → R h i Ug = min Eg dist2 (g, τ · ∂g(x)) , (39) τ ≥0 where ∂g(x) is the subdifferential [32] of g(x) at a point x. C. Some supported results Recall that the probability density of the normal distribution N(0, 1) with zero-mean and unit variance ψ(x) is given by √ 2 ψ(x) := (1/ 2π)e−x /2 . (40) We also consider the following inequality [24]: 2 (1 − x−1 ) 1 p ≤ , ≤√ 5 2π π log(x) (41) for all x > 1. Moreover, adhering to the formulation in [24], we use the following inequality with x > 0 in our derivations: Z ∞ 2 1 ψ(x) √ (v − x)2 e−v /2 dv ≤ . (42) x 2π x R EFERENCES [1] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component analysis?” J. ACM, vol. 58, no. 3, pp. 11:1–11:37, Jun. 2011. [2] J. Wright, A. Ganesh, S. Rao, Y. Peng, and Y. Ma, “Robust principal component analysis: Exact recovery of corrupted low-rank matrices via convex optimization,” in Advances in Neural Information Processing Systems 22, 2009. [3] Z. Lin, M. Chen, and Y. Ma, “The augmented lagrange multiplier method for exact recovery of corrupted low-rank matrices,” University of Illinois at Urbana-Champaign, USA, Tech. Rep., 2009. [4] E. Candès and T. Tao, “Near-optimal signal recovery from random projections: Universal encoding strategies?” IEEE Trans. Inf. Theory, vol. 52, no. 12, pp. 5406–5425, Apr. 2006. [5] D. Donoho, “For most large underdetermined systems of linear equations the minimal `1 -norm solution is also the sparsest solution,” Communications on Pure and Applied Math, vol. 59, no. 6, pp. 797–829, 2006. [6] L. Balzano, R. Nowak, and B. Recht, “Online identification and tracking of subspaces from highly incomplete information,” in Proceedings of the Allerton Conference on Communication, Control and Computing, Sep. 2010. [7] J. He, L. Balzano, and A. Szlam, “Incremental gradient on the grassmannian for online foreground and background separation in subsampled video,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, June 2012. [8] J. Feng, H. Xu, and S. Yan, “Online robust pca via stochastic optimization,” in Advances in Neural Information Processing Systems 26, 2013. [9] H. Mansour and X. Jiang, “A robust online subspace estimation and tracking algorithm,” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), April 2015. [10] H. Guo, C. Qiu, and N. Vaswani, “An online algorithm for separating sparse and low-dimensional signal sequences from their sum,” IEEE Trans. Signal Process., vol. 62, no. 16, pp. 4284–4297, 2014. [11] C. Qiu, N. Vaswani, B. Lois, and L. Hogben, “Recursive robust PCA or recursive sparse recovery in large but structured noise,” IEEE Trans. Inf. Theory, vol. 60, no. 8, pp. 5007–5039, 2014. [12] J. F. Mota, N. Deligiannis, A. C. Sankaranarayanan, V. Cevher, and M. R. Rodrigues, “Adaptive-rate reconstruction of time-varying signals with application in compressive foreground extraction,” IEEE Trans. Signal Process., vol. 64, no. 14, pp. 3651–3666, 2016. [13] G. Warnell, S. Bhattacharya, R. Chellappa, and T. Basar, “Adaptiverate compressive sensing using side information,” IEEE Trans. Image Process., vol. 24, no. 11, pp. 3846–3857, 2015. [14] P. Rodriguez and B. Wohlberg, “Incremental principal component pursuit for video background modeling,” Journal of Mathematical Imaging and Vision, vol. 55, no. 1, pp. 1–18, 2016. [15] K. M. John Wright, Arvind Ganesh and Y. Ma, “Compressive principal component pursuit,” Information and Inference, vol. 2, no. 1, pp. 32–68, 2013. [16] C. Qiu and N. Vaswani, “Support-predicted modified-cs for recursive robust principal components’ pursuit,” in IEEE Int. Symposium on Information Theory, St. Petersburg, Russia, Jul. 2011. [17] N. Vaswani and J. Zhan, “Recursive recovery of sparse signal sequences from compressive measurements: A review,” IEEE Trans. Signal Process., vol. 64, no. 13, pp. 3523–3549, 2016. [18] N. Vaswani and W. Lu, “Modified-cs: Modifying compressive sensing for problems with partially known support,” IEEE Trans. Signal Process., vol. 58, no. 9, pp. 4595–4607, Sep. 2010. [19] H. V. Luong, J. Seiler, A. Kaup, and S. Forchhammer, “Sparse signal reconstruction with multiple side information using adaptive weights for multiview sources,” in IEEE Int. Conf. on Image Process., Phoenix, Arizona, Sep. 2016. [20] M. Brand, “Incremental singular value decomposition of uncertain data with missing values,” in European Conference on Computer Vision, 2002. [21] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM Journal on Imaging Sciences, vol. 2(1), pp. 183–202, 2009. [22] J.-F. Cai, E. J. Candès, and Z. Shen, “A singular value thresholding algorithm for matrix completion,” SIAM J. on Optimization, vol. 20, no. 4, pp. 1956–1982, Mar. 2010. [23] J. F. Mota, N. Deligiannis, and M. R. Rodrigues, “Compressed sensing with side information: Geometrical interpretation and performance bounds,” in IEEE Global Conf. on Signal and Information Processing, Austin, Texas, USA, Dec. 2014. [24] ——, “Compressed sensing with prior information: Optimal strategies, geometry, and bounds,” ArXiv e-prints, Aug. 2014. [25] L. Li, W. Huang, I. Y.-H. Gu, and Q. Tian, “Statistical modeling of complex backgrounds for foreground object detection,” IEEE Trans. Image Process., vol. 13, no. 11, pp. 1459–1472, 2004. [26] The matlab code of the CORPCA algorithm, the test sequences, and the corresponding outcomes. [Online]. Available: https://github.com/huynhlvd/corpca [27] M. Dikmen, S. F. Tsai, and T. S. Huang, “Base selection in estimating sparse foreground in video,” in 16th IEEE International Conference on Image Processing, Nov 2009. [28] D. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289–1306, Apr. 2006. [29] V. Chandrasekaran, B. Recht, P. A. Parrilo, and A. S. Willsky, “The convex geometry of linear inverse problems,” Foundations of Computational Mathematics, vol. 12, no. 6, pp. 805–849, 2012. [30] J. F. Mota, N. Deligiannis, A. Sankaranarayanan, V. Cevher, and M. R. Rodrigues, “Dynamic sparse state estimation using `1 -`1 minimization: Adaptive-rate measurement bounds, algorithms and applications,” in IEEE Int. Conf. on Acoustics, Speech and Signal Processing, Brisbane, Australia, Apr. 2015. [31] D. Amelunxen, M. Lotz, M. B. McCoy, and J. A. Tropp, “Living on the edge: phase transitions in convex programs with random data,” Information and Inference, vol. 3, no. 3, pp. 224–294, 2014. [32] J.-B. Hiriart-Urruty and C. Lemaréchal, Fundamentals of Convex Analysis. Springer, 2004.
2cs.AI
CRYSTALLIZING THE HYPOPLACTIC MONOID: FROM QUASI-KASHIWARA OPERATORS TO THE ROBINSON–SCHENSTED–KNUTH-TYPE CORRESPONDENCE FOR QUASI-RIBBON TABLEAUX arXiv:1601.06390v2 [math.CO] 8 Sep 2016 ALAN J. CAIN AND ANTÓNIO MALHEIRO Abstract. Crystal graphs, in the sense of Kashiwara, carry a natural monoid structure given by identifying words labelling vertices that appear in the same position of isomorphic components of the crystal. In the particular case of the crystal graph for the q-analogue of the special linear Lie algebra sln , this monoid is the celebrated plactic monoid, whose elements can be identified with Young tableaux. The crystal graph and the so-called Kashiwara operators interact beautifully with the combinatorics of Young tableaux and with the Robinson–Schensted–Knuth correspondence and so provide powerful combinatorial tools to work with them. This paper constructs an analogous ‘quasi-crystal’ structure for the hypoplactic monoid, whose elements can be identified with quasi-ribbon tableaux and whose connection with the theory of quasi-symmetric functions echoes the connection of the plactic monoid with the theory of symmetric functions. This quasicrystal structure and the associated quasi-Kashiwara operators are shown to interact just as neatly with the combinatorics of quasiribbon tableaux and with the hypoplactic version of the Robinson– Schensted–Knuth correspondence. A study is then made of the interaction of the crystal graph of the plactic monoid and the quasicrystal graph for the hypoplactic monoid. Finally, the quasi-crystal structure is applied to prove some new results about the hypoplactic monoid. Contents 1. Introduction 2. Preliminaries and notation 2.1. Alphabets and words 2 6 6 2010 Mathematics Subject Classification. Primary 05E15; Secondary 05E05, 20M05. Key words and phrases. hypoplactic, quasi-ribbon tableau, Robinson– Schensted–Knuth correspondence, Kashiwara operator, crystal graph. The first author was supported by an Investigador FCT fellowship (IF/01622/2013/CP1161/CT0001). For both authors, this work was partially supported by the Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) through the project UID/MAT/00297/2013 (Centro de Matemática e Aplicações). 1 2 ALAN J. CAIN AND ANTÓNIO MALHEIRO 2.2. Compositions and partitions 2.3. Weight 3. Crystals and the plactic monoid 3.1. Young tableaux and insertion 3.2. Kashiwara operators and the crystal graph 3.3. Computing the Kashiwara operators 3.4. Properties of the crystal graph 4. Quasi-ribbon tableaux and insertion 5. Quasi-Kashiwara operators and the quasi-crystal graph 6. The quasi-crystal graph and the hypoplactic monoid 7. The quasi-crystal graph and the Robinson–Schensted–Knuth correspondence 8. Structure of the quasi-crystal graph 8.1. Sizes of classes and quasi-crystals 8.2. Interaction of the crystal and quasi-crystal graphs 8.3. The Schützenberger involution 8.4. Characterizing quasi-crystal graphs? 9. Applications 9.1. Counting factorizations 9.2. Conjugacy 9.3. Satisfying an identity References 7 8 8 8 12 14 15 16 20 27 33 34 34 40 49 50 50 50 51 51 52 1. Introduction A crystal basis, in the sense of Kashiwara [Kas91, Kas90], is (informally) a basis for a representation of a suitable algebra on which the generators have a particularly neat action. It gives rise, via tensor products, to the crystal graph, which carries a natural monoid structure given by identifying words labelling vertices that appear in the same position of isomorphic components. The ubiquitous plactic monoid, whose elements can be viewed as semistandard Young tableaux, and which appears in such diverse contexts as symmetric functions [Mac95], representation theory and algebraic combinatorics [Ful97, Lot02], Kostka–Foulkes polynomials [LS81, LS78], Schubert polynomials [LS85, LS90], and musical theory [Jed11], arises in this way from the crystal basis for the q-analogue of the special linear Lie algebra sln . The crystal graph and the associated Kashiwara operators interact beautifully with the combinatorics of Young tableaux and with the Robinson–Schensted–Knuth correspondence and so provide powerful combinatorial tools to work with them. This paper is dedicated to constructing an analogue of this crystal structure for the monoid of quasi-ribbon tableaux: the so-called hypoplactic monoid. To explain this aim in more detail, and in particular CRYSTALLIZING THE HYPOPLACTIC MONOID 3 to describe the properties such an analogue should enjoy, it is necessary to briefly recapitulate some of the theory of crystals and of Young tableaux. The plactic monoid of rank n (where n ∈ N) arises by factoring the free monoid A∗n over the ordered alphabet An = {1 < . . . < n} by a relation ≡plac , which can be defined in various ways. Using Schensted’s algorithm [Sch61], which was originally intended to find longest increasing and decreasing subsequences of a given sequence, one can compute a (semistandard) Young tableau P(w) from a word w ∈ A∗n and so define ≡plac as relating those words that yield the same Young tableau. Knuth made a study of correspondences between Young tableaux and non-negative integer matrices and gave defining relations for the plactic monoid [Knu70]; the relation ≡plac can be viewed as the congruence generated by these defining relations. Lascoux & Schützenberger [LS81] began the systematic study of the plactic monoid, and, as remarked above, connections have emerged with myriad areas of mathematics, which is one of the reasons Schützenberger proclaimed it ‘one of the most fundamental monoids in algebra’ [Sch97]. Of particular interest for us is how it arises from the crystal basis for the q-analogue of the special linear Lie algebra sln (that is, the type An+1 simple Lie algebra), which links it to Kashiwara’s theory of crystal bases [KN94]. Isomorphisms between connected components of the crystal graph correspond to the relation ≡plac . Viewed on a purely combinatorial level, the Kashiwara operators and crystal graph are important tools for working with the plactic monoid. (Indeed, in this context they are sometimes called ‘coplactic’ operators [Lot02, ch. 5], being in a sense ‘orthogonal’ to ≡plac .) Similarly, crystal theory can also be used to analyse the analogous ‘plactic monoids’ that index representations of the q-analogues of symplectic Lie algebras spn (the type Cn simple Lie algebra), special orthogonal Lie algebras of odd and even rank so2n+1 and so2n (the type Bn and Dn simple Lie algebras), and the exceptional simple Lie algebra G2 (see [KS04, Lec02, Lec03] and the survey [Lec07]). The present authors and Gray applied this crystal theory to construct finite complete rewriting systems and biautomatic structures for all these plactic monoids [CGM]; Hage independently constructed a finite complete rewriting system for the plactic monoid of type Cn [Hag15]. As is described in detail later in the paper, the crystal structure meshes neatly with the Robinson–Schensted–Knuth correspondence. This correspondence is a bijection w ↔ (P, Q) where w is a word over An , and P is a semistandard Young tableau with entries in An and Q is a standard Young tableau of the same shape. (The semistandard Young tableau P is the tableau P(w) computed by Schensted’s algorithm; the standard tableau Q can be computed in parallel.) Essentially, the standard Young tableau Q corresponds to the connected component 4 ALAN J. CAIN AND ANTÓNIO MALHEIRO of the crystal graph in which the word w lies, and the semistandard Young tableau P corresponds to the position of w in that component. By holding Q fixed and varying P over semistandard tableaux of the same shape, one obtains all words in a given connected component. Consequently, all words in a given connected component correspond to tableaux of the same shape. In summary, there are three equivalent approaches to the plactic monoid: P1. Generators and relations: the plactic monoid is defined by the presentation An Rplac , where  Rplac = (acb, cab) : a ≤ b < c  ∪ (bac, bca) : a < b ≤ c . Equivalently, ≡plac is the congruence on A∗n generated by Rplac . P2. Tableaux and insertion: the relation ≡plac is defined by u ≡plac v if and only if P(u) = P(v), where P(·) is the Young tableau computed using the Schensted insertion algorithm (see Algorithm 3.2 below). P3. Crystals: the relation ≡plac is defined by u ≡plac v if and only if there is a crystal isomorphism between connected components of the crystal graph that maps u onto v. The defining relations in Rplac (known as the Knuth relations) are the reverse of the ones given in [CGM]. This is because, in the context of crystal bases, the convention for tensor products gives rise to a ‘plactic monoid’ that is actually anti-isomorphic to the usual notion of plactic monoid. Since this paper is mainly concerned with combinatorics, rather than representation theory, it follows Shimozono [Shi05] in using the convention that is compatible with the usual notions of Young tableaux and the Robinson–Schensted–Knuth correspondence. Another important aspect of the plactic monoid is its connection to the theory of symmetric polynomials. The Schur polynomials with n indeterminates, which are the irreducible polynomial characters of the general linear group GLn (C), are indexed by shapes of Young tableaux with entries in An , and they form a Z-basis for the ring of symmetric polynomials in n indeterminates. The plactic monoid was applied to give the first rigorous proof of the Littlewood–Richardson rule (see [LR34] and [Gre07, Appendix]), which is a combinatorial rule for expressing a product of two Schur polynomials as a linear combination of Schur polynomials. In recent years, there has emerged a substantial theory of noncommutative symmetric functions and quasi-symmetric functions; see, for example, [GKL+ 95, KT97, KT99]. Of particular interest for this paper is the notion of quasi-ribbon polynomials, which form a basis for the ring of quasi-symmetric polynomials, just as the Schur polynomials CRYSTALLIZING THE HYPOPLACTIC MONOID 5 form a basis for the ring of symmetric polynomials. The quasi-ribbon polynomials are indexed by the so-called quasi-ribbon tableaux. These quasi-ribbon tableaux have an insertion algorithm and an associated monoid called the hypoplactic monoid, which was first studied in depth by Novelli [Nov00]. The hypoplactic monoid of rank n arises by factoring the free monoid A∗n by a relation ≡hypo , which, like ≡plac , can be defined in various ways. Using the insertion algorithm one can compute a quasi-ribbon tableau from a word and so define ≡hypo as relating those words that yield the same quasi-ribbon tableau. Alternatively, the relation ≡hypo can be viewed as the congruence generated by certain defining relations. Thus there are two equivalent approaches to the hypoplactic monoid: H1. Generators and relations: the hypoplactic monoid is defined by the presentation An Rhypo , where Rhypo = Rplac  ∪ (cadb, acbd) : a ≤ b < c ≤ d  ∪ (bdac, dbca) : a < b ≤ c < d . Equivalently, ≡hypo is the congruence on the free monoid A∗n generated by Rhypo . H2. Tableaux and insertion: the relation ≡hypo is defined by u ≡hypo v if and only if QR(u) = QR(v), where QR(·) is the quasi-ribbon tableau computed using the Krob–Thibon insertion algorithm (see Algorithm 4.3 below). Krob & Thibon [KT97] proved the equivalence of H1 and H2, which are the direct analogues of P1 and P2. Owing to the previous success in detaching crystal basis theory from its representation-theoretic foundation and using it as a combinatorial tool for working with Young tableaux and plactic monoids, it seems worthwhile to try to find an analogue of P3 for the hypoplactic monoid. Such an analogue should have the following form: H3. Quasi-crystals: the relation ≡hypo is defined by u ≡hypo v if and only if there is a quasi-crystal isomorphism between connected components of the quasi-crystal graph that maps u onto v. The aim of this paper is to define quasi-Kashiwara operators on a purely combinatorial level, and so construct a quasi-crystal graph as required for H3. As will be shown, the interaction of the combinatorics of quasi-ribbon tableaux with this quasi-crystal graph will be a close analogue of the interaction of the combinatorics of Young tableaux with the crystal graph (and is, in the authors’ view, just as elegant). In fact, a related notion of ‘quasi-crystal’ is found in Krob & Thibon [KT99]; see also [Hiv00]. However, the Krob–Thibon quasi-crystal describes the restriction to quasi-ribbon tableaux (or more precisely words corresponding to quasi-ribbon tableaux) of the action of the 6 ALAN J. CAIN AND ANTÓNIO MALHEIRO usual Kashiwara operators: it does not apply to all words and so does not give rise to isomorphisms that can be used to define the relation ≡hypo . The paper is organized as follows: Section 2 sets up notation and discusses some preliminaries relating to words, partitions, compositions, and the notion of weight. Section 3 reviews, without proof, the basic theory of Young tableaux, the plactic monoid, Kashiwara operators, and the crystal graph; the aim is to gather the elegant properties of the crystal structure for the plactic monoid that should be mirrored in the quasi-crystal structure for the hypoplactic monoid. These properties will also be used in the study of the interactions of the crystal and quasi-crystal graphs. Section 4 recalls the definitions of quasiribbon tableaux and the hypoplactic monoid. Section 5 states the definition of the quasi-Kashiwara operators and the quasi-crystal graph, and shows that isomorphisms between connected components of this quasi-crystal graph give rise to a congruence on the free monoid. Section 6 proves that the corresponding factor monoid is the hypoplactic monoid. En route, some of the properties of the quasi-crystal graph are established. Section 7 studies how the quasi-crystal graph interacts with the hypoplactic version of the Robinson–Schensted–Knuth correspondence. Section 8 systematically studies the structure of the quasi-crystal graph. It turns out to be a subgraph of the crystal graph for the plactic monoid, and the interplay of the subgraph and graph has some very neat properties. Finally, the quasi-crystal structure is applied to prove some new results about the hypoplactic monoid in Section 9, including an analogy of the hook-length formula. 2. Preliminaries and notation 2.1. Alphabets and words. Recall that for any alphabet X, the free monoid (that is, the set of all words, including the empty word) on the alphabet X is denoted X ∗ . The empty word is denoted ε. For any u ∈ X ∗ , the length of u is denoted |u|, and, for any x ∈ X, the number of times the symbol x appears in u is denoted |u|x . Suppose u = u1 · · · uk ∈ X ∗ (where uh ∈ X). For any i ≤ j, the word ui · · · uj is a factor of u. For any i1 , . . . , im ∈ {1, . . . , k} such that i1 < i2 < . . . < im , the word ui1 ui2 · · · uim is a subsequence of u. Note that factors must be made up of consecutive letters, whereas subsequences may be made up of non-consecutive letters. [To minimize potential confusion, the term ‘subword’ is not used in this paper, since it tends to be synonymous with ‘factor’ in semigroup theory, but with ‘subsequence’ in combinatorics on words.] For further background on the free monoid, see [How95]; for presentations, see [Hig92, Ruš95]. Throughout this paper, A will be the set of natural numbers viewed as an infinite ordered alphabet: A = {1 < 2 < 3 < . . .}. Further, n CRYSTALLIZING THE HYPOPLACTIC MONOID 7 will be a natural number and An will be the set of the first n natural numbers viewed as an ordered alphabet: An = {1 < 2 < . . . < n}. A word u ∈ A∗ is standard if it contains each symbol in {1, . . . , |u|} exactly once. Let u = u1 · · · uk be a standard word. The word u is identified with the permutation h 7→ uh , and u−1 denotes the inverse of this permutation (which is also identifed with a standard word).  Further, the descent set of the standard word u is D(u) = h ∈ {1, . . . , |u| − 1} : uh > uh+1 . Let u ∈ A∗ . The standardization of u, denoted std(u), is the standard word obtained by the following process: read u from left to right and, for each a ∈ A, attach a subscript h to the h-th appearance of a. Symbols with attached subscripts are ordered by  ai < bj ⇐⇒ (a < b) ∨ (a = b) ∧ (i < j) . Replace each symbol with an attached subscript in u with the corresponding symbol of the same rank from A. The resulting word is std(u). For example: w=2 4 3 2 4 5 5 6 5 21 41 31 22 42 51 52 61 53 std(w) = 1 4 3 2 5 6 7 9 8 For further background relating to standard words and standardizaton, see [Nov00, § 2]. 2.2. Compositions and partitions. A weak composition α is a finite sequence (α1 , . . . , αm ) with terms in N ∪ {0}. The terms αh up to the last non-zero term are the parts of α. The length of α, denoted `(α), is the number of its parts. The weight of α, denoted |α|, is the sum of its parts (or, equivalently, of its terms): |α| = α1 + · · · + αm . For example, if α = (3, 0, 4, 1, 0), then `(α) = 4 and |α| = 8. Identify weak compositions whose parts are the same (that is, that differ only in a tail of terms 0). For example, (3, 1, 5, 2) is identified with (3, 1, 5, 2, 0) and (3, 1, 5, 2, 0, 0, 0). (Note that this identification does not create ambiguity in the notions of parts and weight.) A composition is a weak composition whose parts are all in N. For a composition α = (α1 , . . . , α`(α) ), define D(α) = {α1 , α1 + α2 , . . . , α1 + . . . + α`(α−)1 }. For a standard word u ∈ A∗ , define DC(u) to be the unique composition of weight |u| such that D(DC(u)) = D(u), where D(u) is as defined in Subsection 2.1. For example, if u = 143256798, then D(u) = {2, 3, 8} and so DC(u) = (2, 1, 5, 1). A partition λ is a non-increasing finite sequence (λ1 , . . . , λm ) with terms in N. The terms λh are the parts of λ. The length of λ, denoted `(λ), is the number of its parts. The weight of λ, denoted |λ|, is the sum of its parts: |λ| = λ1 + · · · + λm . For example, if λ = (5, 3, 2, 2), then `(λ) = 4 and |λ| = 12. 8 ALAN J. CAIN AND ANTÓNIO MALHEIRO 2.3. Weight. The weight function wt is, informally, the function that counts the number of times each symbol appears in a word. More formally, wt is defined by  wt : A∗ → (N ∪ {0})A u 7→ |u|1 , |u|2 , |u|3 , . . . . Clearly, wt(·) has an infinite tail of components 0, so only the prefix up to the last non-zero term is considered; thus wt(·) is a weak composition. For example, wt(542164325224) = (1, 4, 1, 3, 2, 1). See [Shi05, § 2.1] for a discussion of the basic properties of weight functions. Weights are compared using the following order: (2.1) (α1 , α2 , . . .) ≤ (β1 , β2 , . . .) k k X  X ⇐⇒ (∀k ∈ N) αi ≤ βi . i=1 i=1 When |α| = |β|, this is the dominance order of partitions [Sta05, § 7.2]. When wt(u) < wt(v), one says that v has higher weight than u (and u has lower weight than v). For example, wt(542164325224) = (1, 4, 1, 3, 2, 1) < (5, 3, 2, 2, 0, 0) = wt(432143212111); that is, 432143212111 has higher weight than 542164325224. For the purposes of this paper, it will generally not be necessary to compare weights using (2.1); the important fact will be how the Kashiwara operators affect weight. 3. Crystals and the plactic monoid This section recalls in detail the three approaches to the plactic monoid discussed in the introduction, and discusses further the very elegant interaction of the crystal structure with the combinatorics of Young tableaux and in particular with the Robinson–Schensted–Knuth correspondence. The aim is to lay out the various properties one would hope for in the quasi-crystal structure for the hypoplactic monoid. 3.1. Young tableaux and insertion. The Young diagram of shape λ, where λ is a partition, is a grid of boxes, with λh boxes in the hth row, for h = 1, . . . , `(λ), with rows left-aligned. For example, the Young diagram of shape (5, 3, 2, 2) is A Young diagram of shape (1, 1, . . . , 1) is said to be a column diagram or to have column shape. Note that, in this paper, Young diagrams are CRYSTALLIZING THE HYPOPLACTIC MONOID 9 top-left aligned, with longer rows at the top and the parts of the partition specifying row lengths from top to bottom. There is an alternative convention of bottom-left aligned Young diagrams, where longer rows are at the bottom. A Young tableau is a Young diagram that is filled with symbols from A so that the entries in each row are non-decreasing from left to right, and the entries in each column are increasing from top to bottom. For example, a Young tableau of shape (5, 3, 2, 2) is (3.1) 1 2 4 5 2 2 2 4 3 5 4 6 A Young tableau of shape (1, 1, . . . , 1) is a column. (That is, a column is a Young tableau of column shape.) A standard Young tableau of shape λ is a Young diagram that is filled with symbols from {1, . . . , |λ|}, with each symbol appearing exactly once, so that the entries in each row are increasing from left to right, and the entries in each column are increasing from top to bottom. For example, a standard Young tableau of shape (5, 3, 2, 2) is 1 2 5 8 3 6 7 10 4 12 9 11 A tabloid is an array formed by concatenating columns, filled with symbols from A so that the entries in each column are increasing from top to bottom. (Notice that there is no restriction on the relative heights of the columns; nor is there a condition on the order of entries in a row.) An example of a tabloid is (3.2) 2 1 4 1 2 5 3 2 4 4 6 5 Note that a tableau is a special kind of tabloid. The shape of a tabloid cannot in general be expressed using a partition. The column reading C(T ) of a tabloid T is the word in A∗ obtained by proceeding through the columns, from leftmost to rightmost, and reading each column from bottom to top. For example, the column readings of the tableau (3.1) and the tabloid (3.2) are respectively 5421 6432 52 2 4 and 52 6431 4 5421 2 (where the spaces are simply for 10 ALAN J. CAIN AND ANTÓNIO MALHEIRO clarity, to show readings of individual columns), as illustrated below: 1 2 4 5 2 2 2 4 3 5 4 6 2 1 4 1 2 5 3 2 4 4 6 5 Let w ∈ A∗ , and let w(1) · · · w(m) be the factorization of w into maximal decreasing factors. Let Toid(w) be the tabloid whose h-th column has height |w(h) | and is filled with the symbols of w(h) , for h = 1, . . . , m. Then C(Toid(w)) = w. (Note that the notion of reading described here is the one normally used in the study of Young tableaux and the plactic monoid, and is the opposite of the ‘Japanese reading’ used in the theory of crystals. Throughout the paper, definitions follow the convention compatible with the reading defined here. The resulting differences from the usual practices in crystal theory will be explicitly noted.) If w is the column reading of some Young tableau T , it is called a tableau word. Note that not all words in A∗ are tableau words. For example, 343 is not a tableau word, since 3 3 and 3 3 4 are the 4 only tableaux containing the correct symbols, and neither of these has column reading 343. The word w is a tableau word if and only if Toid(w) is a tableau. The plactic monoid arises from an algorithm that computes a Young tableau P(w) from a word w ∈ A∗ . Algorithm 3.1 (Schensted’s algorithm). Input: A Young tableau T and a symbol a ∈ A. Output: A Young tableau T ← a. Method: (1) If a is greater than or equal to every entry in the topmost row of T , add a as an entry at the rightmost end of T and output the resulting tableau. (2) Otherwise, let z be the leftmost entry in the top row of T that is strictly greater than a. Replace z by a in the topmost row and recursively insert z into the tableau formed by the rows of T below the topmost. (Note that the recursion may end with an insertion into an ‘empty row’ below the existing rows of T .) Using an iterative form of this algorithm, one can start from a word a1 · · · ak (where ai ∈ A) and compute a Young tableau P(a1 · · · ak ). Essentially, one simply starts with the empty tableau and inserts the symbols a1 , . . . , ak in order: However, the algorithm described below also computes a standard Young tableau Q(a1 · · · ak ) that is used in CRYSTALLIZING THE HYPOPLACTIC MONOID 11 the Robinson–Schensted–Knuth correspondence, which will shortly be described: Algorithm 3.2. Input: A word a1 · · · ak , where ai ∈ A. Output: A Young tableau P(a1 · · · ak ) and a standard Young tableau Q(a1 · · · ak ). Method: Start with an empty Young tableau P0 and an empty standard Young tableau Q0 . For each i = 1, . . . , k, insert the symbol ai into Pi−1 as per Algorithm 3.1; let Pi be the resulting Young tableau. Add a cell filled with i to the standard tableaux Qi−1 in the same place as the unique cell that lies in Pi but not in Pi−1 ; let Qi be the resulting standard Young tableau. Output Pk for P(a1 · · · ak ) and Qk as Q(a1 · · · ak ). For example, the sequence of pairs (Pi , Qi ) produced during the application of Algorithm 4.3 to the word 2213 is ( , ),  2 , 1 ,  2 2 , 1 2 , Therefore P(4323) =  1 2 1 2 , 2 3    1 2 3 1 2 4 , , . 2 3 1 2 3 1 2 4 and Q(4323) = . It is straight2 3 forward to see that the map w 7→ P (w), Q(w) is a bijection between words in A∗ and pairs consisting of a Young tableau over A and a standard Young tableau of the same shape. This is the celebrated Robinson–Schensted–Knuth correspondence (see, for example, [Ful97, Ch. 4] or [Sta05, § 7.11]. The possible definition (approach P2 in the introduction) of the relation ≡plac using tableaux and insertion is the following: u ≡plac v ⇐⇒ P(u) = P(v). Using this as a definition, it follows that ≡plac is a congruence on A∗ [Knu70]. The factor monoid A∗ /≡plac is the plactic monoid and is denoted plac. The relation ≡plac is the plactic congruence on A∗ . The congruence ≡plac naturally restricts to a congruence on A∗n , and the factor monoid A∗n /≡plac is the plactic monoid of rank n and is denoted placn . If w is a tableau word, then w = C(P(w)) and Toid(w) = P(w). Thus the tableau words in A∗ form a cross-section (or set of normal forms) for plac, and the tableau words in A∗n form a cross-section for placn . 12 ALAN J. CAIN AND ANTÓNIO MALHEIRO 3.2. Kashiwara operators and the crystal graph. The following discussion of crystals and Kashiwara operators is restricted to the context of placn . For a more general introduction to crystal bases, see [CGM]. The Kashiwara operators ẽi and f˜i , where i ∈ {1, . . . , n − 1}, are partially-defined operators on A∗n . In representation-theoretic terms, the operators ẽi and f˜i act on a tensor product by acting on a single tensor factor [HK02, § 4.4]. This action can be described in a combinatorial way using the the so-called signature or bracketing rule. This paper describes the action directly using this rule, since the the analogous quasi-Kashiwara operators are defined by modifying this rule. The definitions of ẽi and f˜i start from the crystal basis for placn , which will form a connected component of the crystal graph: (3.3) 1 1 2 2 3 3 ... n−2 n−1 n−1 n Each operator f˜i is defined so that it ‘moves’ a symbol a forwards along a directed edge labelled by i whenever such an edge starts at a, and each operator ẽi is defined so that it ‘moves’ a symbol a backwards along a directed edge labelled by i whenever such an edge ends at a: a i f˜i (a) and ẽi (a) i a Using the crystal basis given above, one sees that ẽi (i + 1) = i, f˜i (i) = i + 1, ẽi (x) is undefined for x 6= i + 1; f˜i (x) is undefined for x 6= i. The definition is extended to A∗n \ An by the recursion ( ẽi (u) v if ˜i (u) > φ̃i (v); ẽi (uv) = u ẽi (v) if ˜i (u) ≤ φ̃i (v), ( f˜i (u) v if ˜i (u) ≥ φ̃i (v); f˜i (uv) = u f˜i (v) if ˜i (u) < φ̃i (v), where ˜i and φ̃i are auxiliary maps defined by  ˜i (w) = max k ∈ N ∪ {0} : ẽi · · · ẽi (w) is defined ; | {z } k times  ˜ φ̃i (w) = max k ∈ N ∪ {0} : fi · · · f˜i (w) is defined . | {z } k times Notice that this definition is not circular: the definitions of ẽi and f˜i depend, via ˜i and φ̃i , only on ẽi and f˜i applied to strictly shorter words; the recursion terminates with ẽi and f˜i applied to single letters from the alphabet An , which was defined using the crystal basis (3.3). (Note that this definition is in a sense the mirror image of [KN94, Theorem 1.14], CRYSTALLIZING THE HYPOPLACTIC MONOID length 0 length 1 ε 1 z}|{ z}|{ z 1 2 length 2 }| 11 1 12 21 2 1 2 3 13 13 { 2 22 31 1 2 1 23 32 2 z 33 length 3 }| 111 1 112 2 1 113 122 1 2 1 123 222 2 1 2 133 223 1 2 233 211 1 2 212 { 121 1 2 311 221 131 2 1 2 1 213 312 231 132 2 1 2 1 322 331 313 1 2 323 321 232 1 2 332 2 333 Figure 1. Part of the crystal graph for plac3 . Note that each connected component consists of words of the same length. In particular, the empty word ε is an isolated vertex, and the words of length 1 form a single connected component, which is the crystal basis for plac3 . The two connected components whose highest-weight words are 211 and 121 are isomorphic. However, the components consisting of the isolated vertex 321 is not. because of the choice of definition for readings of tableaux used in this paper. Thus the definition of ẽi and f˜i is the same as [Shi05, p. 8].) Although it is not immediate from the definition, the operators ẽi and f˜i are well-defined. Furthermore, ẽi and f˜i are mutually inverse whenever they are defined, in the sense that if ẽi (w) is defined, then w = f˜i (ẽi (w)), and if f˜i (w) is defined, then w = ẽi (f˜i (w)). The crystal graph for placn , denoted Γ(placn ), is the directed labelled graph with vertex set A∗n and, for w, w0 ∈ A∗n , an edge from w to w0 labelled by i if and only if w0 = f˜i (w) (or, equivalently, w = ẽi (w0 )). Figure 1 shows part of the crystal graph Γ(plac3 ). 14 ALAN J. CAIN AND ANTÓNIO MALHEIRO Since the operators ẽi and f˜i preserve lengths of words, and since there are finitely many words in A∗n of each length, each connected component in the crystal graph must be finite. For any w ∈ A∗n , let Γ(placn , w) denote the connected component of Γ(placn ) that contains the vertex w. Notice that the crystal basis (3.3) is the connected component Γ(placn , 1). A crystal isomorphism between two connected components is a weightpreserving labelled digraph isomorphism. That is, a map θ : Γ(placn , w) → Γ(placn , w0 ) is a crystal isomorphism if it has the following properties: • θ is bijective; • wt(θ(u)) = wt(u) for all u ∈ Γ(placn , w); i v if and only • for all u, v ∈ Γ(placn , w), there is an edge u i θ(v) if there is an edge θ(u) . The possible definition (approach P3 in the introduction) of the relation ≡plac using the crystal graph Γ(placn ) is the following: two words in A∗n are related by ≡plac if and only if they lie in the same place in isomorphic connected components. More formally, w ≡plac w0 if and only if there is a crystal isomorphism θ : Γ(placn , w) → Γ(placn , w0 ) such that θ(w) = w0 (see, for example, [Lec07]). For example, 2213 ≡plac 2231, and these words appear in the same position in two of the connected components shown in Figure 2. (This figure also illustrates other properties that will be discussed shortly.) 3.3. Computing the Kashiwara operators. The recursive definition of the Kashiwara operators ẽi and f˜i given above is not particularly convenient for practical computation. The following method, outlined in [KN94], is more useful: Let i ∈ {1, . . . , n − 1} and let w ∈ A∗n . Form a new word in {+, −}∗ by replacing each letter i of w by the symbol +, each letter i + 1 by the symbol −, and every other symbol with the empty word, keeping a record of the original letter replaced by each symbol. Then delete factors −+ until no such factors remain: the resulting word is +φ̃i (w) −˜i (w) , and is denoted by ρi (w). Note that factors +− are not deleted. (The method given in [KN94] involved deleting factors +−; again, this difference is a consequence of the choice of convention for reading tableaux.) If ˜i (w) = 0, then ẽi (w) is undefined. If ˜i (w) > 0 then one obtains ẽi (w) by taking the letter i + 1 which was replaced by the leftmost − of ρi (w) and changing it to i. If φ̃i (w) = 0, then f˜i (w) is undefined. If φ̃i (w) > 0 then one obtains f˜i (w) by taking the letter i which was replaced by the rightmost + of ρi (w) and changing it to i + 1. For a purely combinatorial proof that this method of computation is correct, see [CGM, Proposition 2.1]. (Note that [CGM] uses the tableaux-reading convention from representation theory, so that the result must be reflected to fit the convention used here.) CRYSTALLIZING THE HYPOPLACTIC MONOID Γ(plac3 , 2111) z }| { 111 2 Γ(plac3 , 1211) z }| { 1211 1 2 15 Γ(plac3 , 1121) z }| { 1121 1 2 112 2 111 3 122 2 113 2 112 3 2212 1213 1312 123 2 113 3 122 3 2213 1313 1322 133 2 123 3 222 3 2313 1323 2322 2331 1332 2232 133 3 223 3 3313 2323 3331 2332 1212 1311 1 2 2 1 2 2 | {z 1 3 4 2 233 3 } 1 1 2 1 2 2 1 2 {z 1 2 4 3 3323 } 1 2231 1331 1232 2 1 2 1 2 1 2 | 1 2221 1231 1132 1 1 2 1 2 1221 1131 1 2 1 2 | {z 3332 1 2 3 4 } Figure 2. Three isomorphic components of the crystal graph for plac3 . In the component containing column readings of tableaux, the tableaux themselves are shown instead of words. The standard tableaux below each component is Q(w) for all words w in that component. 3.4. Properties of the crystal graph. In the crystal graph Γ(placn ), the length of the longest path consisting of edges labelled by i ∈ {1, . . . , n − 1} that ends at w ∈ A∗n is ˜i (w). The length of the longest path consisting of edges labelled by i that starts at w ∈ A∗n is φ̃i (w). The operators ẽi and f˜i respectively increase and decrease weight whenever they are defined, in the sense that if ẽi (w) is defined, then wt(ẽi (w)) > wt(w), and if f˜i (w) is defined, then wt(f˜i (w)) < wt(w). This is because ẽi replaces a symbol i + 1 with i whenever it is defined, which corresponds to decrementing the i + 1-th component and incrementing the i-th component of the weight, which results in an increase with respect to the order (2.1). Similarly, f˜i replaces a symbol i with i + 1 whenever it is defined. For this reason, the ẽi and f˜i are respectively known as the Kashiwara raising and lowering operators. Every connected component in Γ(placn ) contains a unique highestweight vertex: a vertex whose weight is higher than all other vertices in that component. This means that no Kashiwara raising operator ẽi is defined on this vertex. See [Shi05, § 2.4.2] for proofs and background. (The existence, but not the uniqueness, of a highest-weight vertex is a consequence of the finiteness of connected components.) 16 ALAN J. CAIN AND ANTÓNIO MALHEIRO Whenever they are defined, the operators ẽi and f˜i preserve the property of being a tableau word and the shape of the corresponding tableau [KN94]. Furthermore, all the tableau words corresponding to tableaux of a given shape with entries in An lie in the same connected component. As shown in Figure 2, the left-hand component Γ(plac3 , 2111) is made up of all the tableau words corresponding to tableaux of shape (3, 1) with entries in A3 . Each connected component in Γ(placn ) corresponds to exactly one standard tableau, in the sense that Q(w) = Q(u) if and only if u and w lies in the same connected component of Γ(placn ). In terms of the bijection w 7→ (P(w), Q(w)) of the Robinson–Schensted–Knuth correspondence, specifying Q(w) locates the particular connected component Γ(placn , w), and specifying P(w) locates the word w within that component. Highest-weight words in Γ(placn ), and in particular highest-weight tableau words, admit a useful characterization as Yamanouchi words. A word w1 · · · wm ∈ A∗n (where wi ∈ An ) is a Yamanouchi word if, for every j = 1, . . . , m, the weight of the suffix wj · · · wm is a nonincreasing sequence (that is, a partition). Thus 1231 is not a Yamanouchi word, since wt(31) = (1, 0, 1), but (1121) is a Yamanouchi word, since wt(1121) = (3, 1, 0); wt(121) = (2, 1, 0); wt(21) = (1, 1, 0), and wt(1) = (1, 0, 0). A word is highest-weight if and only if it is a Yamanouchi word. See [Lot02, Ch. 5] for further background. Highest-weight tableau words also have a neat characterization: a tableau word is highest-weight if and only if its weight is equal to the shape of the corresponding tableau. That is, a tableau word whose corresponding tableau has shape λ is highest-weight if and only if, for each i ∈ An , the number of symbols i it contains is λi . It follows that a tableau whose reading is a highest-weight word must contain only symbols i on its i-th row, for all i ∈ {1, . . . , `(λ)}. For example, the tableau of shape (5, 3, 2, 2) whose reading is a highest-weight word is: 1 2 3 4 1 1 1 1 2 2 3 4 See [KN94] for further background. 4. Quasi-ribbon tableaux and insertion This section gathers the relevant definitions and background on quasiribbon tableaux, the analogue of the Robinson–Schensted–Knuth correspondence, and the hypoplactic monoid. For further background, see [KT97, Nov00]. Let α = (α1 , . . . , αm ) and β = (β1 , . . . , βp ) be compositions with |α| = |β|. Then β is coarser than α, denoted β  α, if each partial CRYSTALLIZING THE HYPOPLACTIC MONOID 17 sum β1 +. . .+βp0 (for p0 < p) is equal to some partial sum α1 +. . .+αm0 for some m0 < m. (Essentially, β is coarser than α if it can be formed from α by ‘merging’ consecutive parts.) Thus (11)  (3, 8)  (3, 6, 2)  (3, 1, 5, 2). A ribbon diagram of shape α, where α is a composition, is an array of boxes, with αh boxes in the h-th row, for h = 1, . . . , `(α) and counting rows from top to bottom, aligned so that the leftmost cell in each row is below the rightmost cell of the previous row. For example, the ribbon diagram of shape (3, 1, 5, 2) is: row 1 (4.1) row 2 row 3 row 4 Notice that a ribbon diagram cannot contain a 2 × 2 subarray (that is, of the form ). In a ribbon diagram of shape α, the number of rows is `(α) and the number of boxes is |α|. A quasi-ribbon tableau of shape α, where α is a composition, is a ribbon diagram of shape α filled with symbols from A such that the entries in every row are non-decreasing from left to right and the entries in every column are strictly increasing from top to bottom. An example of a quasi-ribbon tableau is (4.2) 1 2 2 3 4 4 5 5 5 6 7 Note the following immediate consequences of the definition of a quasiribbon tableau: (1) for each a ∈ A, the symbols a in a quasi-ribbon tableau all appear in the same row, which must be the j-th row for some j ≤ a; (2) the h-th row of a quasi-ribbon tableau cannot contain symbols from {1, . . . , h − 1}. A quasi-ribbon tabloid is a ribbon diagram filled with symbols from A such that the entries in every column are strictly increasing from top to bottom. (Notice that there is no restriction on the order of entries in a row.) An example of a quasi-ribbon tabloid is (4.3) 1 5 2 3 6 2 4 5 4 5 7 Note that a quasi-ribbon tableau is a special kind of quasi-ribbon tabloid. A recording ribbon of shape α, where α is a composition, is a ribbon diagram of shape α filled with symbols from {1, . . . , |α|}, with each 18 ALAN J. CAIN AND ANTÓNIO MALHEIRO symbol appearing exactly once, such that the entries in every row are increasing from left to right, and entries in every column are increasing from bottom to top. (Note that the condition on the order of entries in rows is the same as in quasi-ribbon tableau, but the condition on the order of entries in columns is the opposite of that in quasi-ribbon tableau.) An example of a recording ribbon of shape (3, 1, 5, 2) is (4.4) 1 2 9 8 3 4 6 7 11 5 10 The column reading C(T ) of a quasi-ribbon tabloid T is the word in A obtained by proceeding through the columns, from leftmost to rightmost, and reading each column from bottom to top. For example, the column reading of the quasi-ribbon tableau (4.2) and the quasi-ribbon tabloid (4.3) are respectively 1 2 432 4 5 5 65 7 and 1 5 632 2 4 5 54 7 (where the spaces are simply for clarity, to show readings of individual columns), as illustrated below: ∗ 1 2 2 3 4 4 5 5 5 6 7 1 5 2 3 6 2 4 5 4 5 7 Let w ∈ A∗ , and let w(1) · · · w(m) be the factorization of w into maximal decreasing factors. Let QRoid(w) be the quasi-ribbon tabloid whose h-th column has height |w(h) | and is filled with the symbols of w(h) , for h = 1, . . . , m. (So each maximal decreasing factor of w corresponds to a column of QRoid(w).) Then C(QRoid(w)) = w. If w is the column reading of some quasi-ribbon tableau T , it is called a quasi-ribbon word. It is easy to see that the word w is a quasi-ribbon word if and only if QRoid(w) is a quasi-ribbon tableau. For example 433 is not a quasi-ribbon word, since QRoid(433) = 3 is the only 4 3 quasi-ribbon tabloid whose column reading is 433. Proposition 4.1 ([Nov00, Proposition 3.4]). A word u ∈ A∗ is a quasi-ribbon word if and only if std(u) is a quasi-ribbon word. The following algorithm gives a method for inserting a symbol into a quasi-ribbon tableau. It is due to Krob & Thibon, but is stated here in a slightly modified form: Algorithm 4.2 ([KT97, § 7.2]). Input: A quasi-ribbon tableau T and a symbol a ∈ A. Output: A quasi-ribbon tableau T ← a. CRYSTALLIZING THE HYPOPLACTIC MONOID 19 Method: If there is no entry in T that is less than or equal to a, output the quasi-ribbon tableau obtained by creating a new entry a and attaching (by its top-left-most entry) the quasi-ribbon tableau T to the bottom of a. If there is no entry in T that is greater than a, output the word obtained by creating a new entry a and attaching (by its bottom-rightmost entry) the quasi-ribbon tableau T to the left of a. Otherwise, let x and z be the adjacent entries of the quasi-ribbon tableau T such that x ≤ a < z. (Equivalently, let x be the right-most and bottom-most entry of T that is less than or equal to a, and let z be the left-most and top-most entry that is greater than a. Note that x and z could be either horizontally or vertically adjacent.) Take the part of T from the top left down to and including x, put a new entry a to the right of x and attach the remaining part of T (from z onwards to the bottom right) to the bottom of the new entry a, as illustrated here: x z x z ←a= ←a= x a z x a z [where x and z are vertically adjacent] [where x and z are horizontally adjacent] Output the resulting quasi-ribbon tableau. Using an iterative form of this algorithm, one can start from a word a1 · · · ak (where ai ∈ A) and compute a quasi-ribbon tableau QR(a1 · · · ak ). Essentially, one simply starts with the empty quasiribbon tableau and inserts the symbols a1 , a2 , . . . , ak in order. However, the algorithm described below also computes a recording ribbon RR(a1 · · · ak ), which will be used later in discussing an analogue of the Robinson–Schensted–Knuth correspondence. Algorithm 4.3 ([KT97, § 7.2]). Input: A word a1 · · · ak , where ai ∈ A. Output: A quasi-ribbon tableau QR(a1 · · · ak ) and a recording ribbon RR(a1 · · · ak ). Method: Start with the empty quasi-ribbon tableau Q0 and an empty recording ribbon R0 . For each i = 1, . . . , k, insert the symbol ai into Qi−1 as per Algorithm 4.3; let Qi be the resulting quasi-ribbon tableau. Build the recording ribbon Ri , which has the same shape as Qi , by adding an entry i into Ri−1 at the same place as ai was inserted into Qi−1 . Output Qk for QR(a1 · · · ak ) and Rk as RR(a1 · · · ak ). 20 ALAN J. CAIN AND ANTÓNIO MALHEIRO For example, the sequence of pairs (Qi , Ri ) produced during the application of Algorithm 4.3 to the word 4323 is       2 3 2 3  3 2 ( , ), 4 , 1 , , ,  3 , 2 ,  3 3 , 2 4 . 4 1 2 4 1 4 1 3 Therefore QR(4323) = 3 3 and RR(4323) = 2 4 . It is straightfor4 1 ward to see that the map u 7→ (QR(u), RR(u)) is a bijection between words in A∗ and pairs consisting of a quasi-ribbon tableau and a recording ribbon [KT97, § 7.2] of the same shape; this is an analogue of the Robinson–Schensted–Knuth correspondence. For instace, if QR(u) is (4.2) and RR(u) is (4.4), then u = 12446553275. Recall the definition of DC( ) from Subsection 2.2. Proposition 4.4 ([Nov00, Theorems 4.12 & 4.16]). For any word u ∈ A∗ , the shape of QR(u) (and of RR(u)) is DC(std(u)−1 ). The possible definition (approach H2 in the introduction) of the relation ≡hypo using tableaux and insertion is the following: u ≡hypo v ⇐⇒ QR(u) = QR(v). Using this as a definition, it follows that ≡hypo is a congruence on A∗ [Nov00], which is known as the hypoplactic congruence on A∗ . The factor monoid A∗ /≡hypo is the hypoplactic monoid and is denoted hypo. The congruence ≡hypo naturally restricts to a congruence on A∗n , and the factor monoid A∗n /≡hypo is the hypoplactic monoid of rank n and is denoted hypon . As noted above, if w is a quasi-ribbon word, then w = C(QR(w)). Thus the quasi-ribbon words in A∗ form a cross-section (or set of normal forms) for hypo, and the quasi-ribbon words in A∗n form a crosssection for hypon . This paper also uses the following equivalent characterization of ≡hypo [Nov00, Theorem 4.18]: u ≡hypo v ⇐⇒ (4.5) DC(std(u)−1 ) = DC(std(v)−1 ) ∧ wt(u) = wt(v). 5. Quasi-Kashiwara operators and the quasi-crystal graph This section defines the quasi-Kashiwara operators and the quasicrystal graph, and shows that isomorphisms between components of this graph give rise to a monoid. The following section will prove that this monoid is in fact the hypoplactic monoid. Let i ∈ {1, . . . , n − 1}. Let u ∈ A∗n . The word u has an i-inversion if it contains a symbol i + 1 to the left of a symbol i. Equivalently, u CRYSTALLIZING THE HYPOPLACTIC MONOID 21 has an i-inversion if it contains a subsequence (i + 1)i. (Recall from Subsection 2.1 that a subsequence may be made up of non-consecutive letters.) If the word u does not have an i-inversion, it is said to be i-inversion-free. Define the quasi-Kashiwara operators ëi and f¨i on A∗n as follows: Let u ∈ A∗n . • If u has an i-inverstion, both ëi (u) and f¨i (u) are undefined. • If u is i-inversion-free, but u contains at least one symbol i + 1, then ëi (u) is the word obtained from u by replacing the leftmost symbol i + 1 by i; if u contains no symbol i + 1, then ëi (u) is undefined. • If u is i-inversion-free, but u contains at least one symbol i, then f¨i (u) is the word obtained from u by replacing the rightmost symbol i by i + 1; if u contains no symbol i, then f¨i (u) is undefined. For example, ë2 (3123) is undefined since 3123 has a 2-inversion; f¨2 (3131) is undefined since 3131 is 2-inversion free but does not contains a symbol 2; f¨1 (3113) = 3123. Define  ¨i (u) = max k ∈ N ∪ {0} : ëi · · · ëi (u) is defined ; | {z } k times  ¨ φ̈i (u) = max k ∈ N ∪ {0} : fi · · · f¨i (u) is defined . | {z } k times An immediate consequence of the definitions of ëi and f¨i is that: • If u has an i-inverstion, then ¨i (u) = φ̈i (u) = 0; • If u is i-inversion-free, then every symbol i lies to the left of every symbol i + 1 in u, and so ¨i (u) = |u|i+1 and φ̈i (u) = |u|i . Remark 5.1. It is worth noting how the quasi-Kashiwara operators ëi and f¨i relate to the standard Kashiwara operators ẽi and f˜i as defined in Subsection 3.2. As discussed in Subsection 3.3, one computes the action of ẽi and f˜i on a word u ∈ A∗n by replacing each symbol i with +, each symbol i + 1 with −, every other symbol with the empty word, and then iteratively deleting factors −+ until a word of the form +p −q remains, whose left-most symbol − and right-most symbol + (if they exist) indicate the symbols in u changed by ẽi and f˜i respectively. The deletion of factors −+ corresponds to rewriting to normal form a word representing an element of the bicyclic monoid h +, − | (−+, ε) i. To compute the action of the quasi-Kashiwara operators ëi and f¨i defined 22 ALAN J. CAIN AND ANTÓNIO MALHEIRO above on a word u ∈ A∗n , replace the symbols in the same way as before, but now rewrite to normal form as a word representing an element of the monoid h +, − | (−+, 0) i, where 0 is a multiplicative zero. Any word that contains a symbol − to the left of a symbol + will be rewritten to 0 and so ëi and f¨i will be undefined in this case. If the word does not contain a symbol − to the left of a symbol +, then it is of the form +p −q , and the left-most symbol − and the rightmost symbol + indictate the symbols in u changed by ëi and f¨i . In essence, one obtains the required analogies of the Kashiwara operators by replacing the bicyclic monoid h +, − | (−+, ε) i, where −+ rewrites to the identity, with the monoid h +, − | (−+, 0) i, where −+ rewrites to the zero. The action of the quasi-Kashiwara operators is essentially a restriction of the action of the Kashiwara operators: Proposition 5.2. Let u ∈ A∗n . If ëi (u) is defined, so is ẽi (u), and ẽi (u) = ëi (u). If f¨i (u) is defined, so is f˜i (u), and f˜i (u) = f¨i (u). Proof. Let u ∈ A∗n . Suppose the quasi-Kashiwara operator ëi is defined on u. Then u contains at least one symbol i + 1 but is i-inversion-free, so that every symbol i lies to the left of every symbol i + 1 in u. So when one computes the action of ẽi , replacing every symbol i with the symbol + and every symbol i + 1 with the symbol − leads immediately to the word +φ̃i (u) −˜i (u) (that is, there are no factors −+ to delete). Hence ˜i (u) > 0, and so the Kashiwara operator ẽi is defined on u. Furthermore, ẽi acts by changing the symbol i + 1 that contributed the leftmost − to i, and, since there was no deletion of factors −+, this symbol must be the leftmost symbol i + 1 in u. Thus ẽi (u) = ëi (u). Similarly, if the quasi-Kashiwara operator f¨i is defined on u, so is the Kashiwara operator f˜i , and f˜i (u) = f¨i (u).  The original definition of the Kashiwara operators ẽi and f˜i in Subsection 3.2 was recursive: whether the action on uv recurses to the action on u or on v depends on the maximum number of times each operator can be applied to u and v separately. It seems difficult to give a similar recursive definition for the quasi-Kashiwara operators ëi and f¨i defined here: if, for example, both operators can be applied zero times to u, this may mean that u does not contain symbols i or i + 1, in which case the operators may still be defined on uv, or it may mean that u contains a symbol i + 1 to the left of a symbol i, in which case the operators are certainly not defined on uv. This concludes the discussion contrasting the standard Kashiwara operators with the quasi-Kashiwara operators defined here. The aim now is to use the operators ëi and f¨i to build the quasi-crystal graph and to establish some of its properties. CRYSTALLIZING THE HYPOPLACTIC MONOID 23 Lemma 5.3. For all i ∈ {1, . . . , n − 1}, the operators ëi and f¨i are mutually inverse, in the sense that if ëi (u) is defined, u = f¨i (ëi (u)), and if f¨i (u) is defined, u = ëi (f¨i (u)). Proof. Let u ∈ A∗n . Suppose that ëi (u) is defined. Then u contains at least one symbol i + 1 but is i-inversion-free, so that every symbol i + 1 is to the right of every symbol i. Since ëi (u) is obtained from u by replacing the left-most symbol i+1 by i (which becomes the right-most symbol i), every symbol i + 1 is to the right of every symbol i in the word ëi (u) and so ëi (u) is i-inversion-free, and ëi (u) contains at least one symbol i. Thus f¨i (ëi (u)) is defined, and is obtained from ëi (u) by replacing the right-most symbol i by i + 1, which produces u. Hence if ëi (u) is defined, u = f¨i (ëi (u)). Similar reasoning shows that if f¨i (u) is defined, u = ëi (f¨i (u)).  The operators ëi and f¨i respectively increase and decrease weight whenever they are defined, in the sense that if ëi (u) is defined, then wt(ëi (u)) > wt(u), and if f¨i (u) is defined, then wt(f¨i (u)) < wt(u). This is because ëi replaces a symbol i+1 with i whenever it is defined, which corresponds to decrementing the i + 1-th component and incrementing the i-th component of the weight, which results in an increase with respect to the order (2.1). Similarly, f¨i replaces a symbol i with i + 1 whenever it is defined. For this reason, the ëi and f¨i are respectively called the quasi-Kashiwara raising and lowering operators. The quasi-crystal graph Γn is the labelled directed graph with vertex set A∗n and, for all u ∈ A∗n and i ∈ {1, . . . , n − 1}, an edge from u to u0 labelled by i if and only if f¨i (u) = u0 (or, equivalently by Lemma 5.3, u = ëi (u0 )). Part of Γ4 is shown in Figure 3. (The notation Γn will be discarded in favour of Γ(hypon ) after it has been shown that the relationship between Γn and hypon is analogous to that between Γ(placn ) and placn .) For any u ∈ A∗n , let Γn (u) be the connected component of Γn containing the vertex u. Notice that every vertex of Γn has at most one incoming and at most one outgoing edge with a given label. A quasicrystal isomorphism between two connected components is a weightpreserving labelled digraph isomorphism. (This parallels the definition of a crystal isomorphism in Subsection 3.2.) Define a relation ∼ on the free monoid A∗n as follows: u ∼ v if and only if there is a quasi-crystal isomorphism θ : Γn (u) → Γn (v) such that θ(u) = v. That is, u ∼ v if and only if u and v are in the same position in isomorphic connected components of Γn . For example, 1324 ∼ 3142, since these words are in the same position in their connected components, as can be seen in Figure 3. The rest of this section is dedicated to proving that the relation ∼ is a congruence on A∗n . The proofs of this result and the necessary lemmata 24 ALAN J. CAIN AND ANTÓNIO MALHEIRO 1212 2121 2 2 1213 2 2131 3 1313 1 1214 3 1323 1 3 1 2434 2 3434 3 1424 1 2424 2 1 2 1434 1 2141 3 3132 3 1324 2324 3 2 1 3 3131 1314 2323 3 2 1 1414 1 3141 3 3232 3 1 3 3 4141 1 4142 1 4242 2 1 3142 3242 3 2 2 4143 1 4243 2 4343 Figure 3. The isomorphic components Γ4 (1212) and Γ4 (2121) of the quasi-crystal graph Γ4 . parallel the purely combinatorial proofs by the present authors and Gray [CGM, § 2.4] that isomorphisms of crystal graphs give rise to congruences. Lemma 5.4. Let u, v, u0 , v 0 ∈ A∗n . Suppose u ∼ v and u0 ∼ v 0 , and let θ : Γn (u) → Γn (v) and θ0 : Γn (u0 ) → Γn (v 0 ) be quasi-crystal isomorphisms such that θ(u) = v and θ0 (u0 ) = v 0 . Let i ∈ {1, . . . , n − 1}. Then: (1) ëi (uu0 ) is defined if and only if ëi (vv 0 ) is defined. If both are defined, exactly one of the following statements holds: (a) ëi (uu0 ) = uëi (u0 ) and ëi (vv 0 ) = vëi (v 0 ); (b) ëi (uu0 ) = ëi (u)u0 and ëi (vv 0 ) = ëi (v)v 0 . CRYSTALLIZING THE HYPOPLACTIC MONOID 25 (2) f¨i (uu0 ) is defined if and only if f¨i (vv 0 ) is defined. If both are defined, exactly one of the following statements holds: (a) f¨i (uu0 ) = uf¨i (u0 ) and f¨i (vv 0 ) = v f¨i (v 0 ); (b) f¨i (uu0 ) = f¨i (u)u0 and f¨i (vv 0 ) = f¨i (v)v 0 . Proof. Suppose ëi (uu0 ) is defined. Then uu0 is i-inversion-free, and contains at least one symbol i + 1. Hence both u and u0 are i-inversionfree, and at least one of u and u0 contains a symbol i + 1. Since θ and θ0 are crystal isomorphisms, wt(u) = wt(v) and wt(u0 ) = wt(v 0 ) (and thus u and v contain the same number of each symbol, and u0 and v 0 contain the same number of each symbol). Consider separately two cases depending on whether u contains a symbol i + 1: • Suppose that u does not contain a symbol i + 1 (and hence v does not contain a symbol i + 1). Then u0 must contain a symbol i + 1 and indeed the left-most symbol i + 1 of uu0 must lie in u0 . So ëi (u0 ) is defined and ëi (uu0 ) = uëi (u0 ). Thus there is an edge labelled by i ending at u0 . Since θ0 is a quasi-crystal isomorphism, there is an edge labelled by i ending at v 0 . Hence ëi (v 0 ) is defined, and so v 0 is i-inversion-free. Since v does not contain any symbol i + 1, it follows that vv 0 is i-inversion-free. Hence ëi (vv 0 ) is defined and, since the left-most symbol i + 1 of vv 0 lies in v 0 , it also holds that ëi (vv 0 ) = vëi (v 0 ). • Suppose that u contains a symbol i + 1 (and hence v contains a symbol i + 1). Then u0 cannot contain a symbol i, and the left-most symbol i + 1 of uu0 must lie in u. Then ëi (u) is defined and ëi (uu0 ) = ëi (u)u0 . So there is an edge labelled by i ending at u. Since θ is a quasi-crystal isomorphism, there is an edge labelled by i ending at v. Hence ëi (v) is defined, and so v is iinversion-free. Since v 0 does not contain any symbol i, it follows that vv 0 is i-inversion-free. Hence ëi (vv 0 ) is defined and, since the left-most symbol i + 1 in vv 0 lies in v, it also holds that ëi (vv 0 ) = ëi (v)v 0 . This proves the forward implications of the three statements relating to ëi in part (1). Interchanging u and u0 with v and v 0 proves the reverse implications. The statements for f¨i in part (2) follow similarly.  Lemma 5.5. Let u, v, u0 , v 0 ∈ A∗n . Suppose u ∼ v and u0 ∼ v 0 , and let θ : Γn (u) → Γn (v) and θ0 : Γn (u0 ) → Γn (v 0 ) be quasi-crystal  isomor0 0 0 phisms such that θ(u) = v and θ (u ) = v . Let g̈i1 , . . . , g̈ir ∈ ëi , f¨i : i ∈ {1, . . . , n − 1} . Then: (1) g̈i1 · · · g̈ir (uu0 ) is defined if and only if g̈i1 · · · g̈ir (vv 0 ) is defined. (2) When both g̈i1 · · · g̈ir (uu0 ) and g̈i1 · · · g̈ir (vv 0 ) are defined, the sequence g̈i1 , . . . , g̈ir partitions into two subsequences g̈j1 , . . . , g̈js 26 ALAN J. CAIN AND ANTÓNIO MALHEIRO and g̈k1 , . . . , g̈kt such that g̈i1 · · · g̈ir (uu0 ) = g̈j1 · · · g̈js (u)g̈k1 · · · g̈kt (u0 ), g̈i1 · · · g̈ir (vv 0 ) = g̈j1 · · · g̈js (v)g̈k1 · · · g̈kt (v 0 ); where  θ g̈j1 · · · g̈js (u) = g̈j1 · · · g̈js (v),  θ0 g̈k1 · · · g̈kt (u0 ) = g̈k1 · · · g̈kt (v 0 ). Proof. This result follows by iterated application of Lemma 5.4, with the last two equalities holding because θ and θ0 are quasi-crystal isomorphisms with θ(u) = v and θ0 (u0 ) = v 0 .  Proposition 5.6. The relation ∼ is a congruence on the free monoid A∗n . Proof. It is clear from the definition that ∼ is an equivalence relation; it thus remains to prove that ∼ is compatible with multiplication in A∗n . Suppose u ∼ v and u0 ∼ v 0 . Then there exist quasi-crystal isomorphisms θ : Γn (u) → Γn (v) and θ : Γn (u0 ) → Γn (v 0 ) such that θ(u) = v and θ0 (u0 ) = v 0 . Define a map Θ :Γn (uu0 ) → Γn (vv 0 ) as follows. For w ∈ Γn (uu0 ), choose g̈i1 , . . . , g̈ir ∈ ëi , f¨i : i ∈ {1, . . . , n−1} such that g̈i1 · · · g̈ir (uu0 ) = w; such a sequence exists because w lies in the connected component Γn (uu0 ). Define Θ(w) to be g̈i1 · · · g̈ir (vv 0 ); note that this is defined by Lemma 5.5(1).  It is necessary to prove that Θ is well-defined. Suppose0 that g̈ı̂1 , . . . , g̈ı̂m ∈ ëi , f¨i : i ∈ {1, . . . , n − 1} is such that g̈ı̂1 · · · g̈ı̂m (uu ) = w, and let z = g̈ı̂1 · · · g̈ı̂m (vv 0 ). Note that by Lemma 5.5(2), • the sequence g̈i1 , . . . , g̈ir partitions into two subsequences g̈j1 , . . . , g̈js and g̈k1 , . . . , g̈kt such that (5.1) g̈i1 · · · g̈ir (uu0 ) = g̈j1 · · · g̈js (u)g̈k1 · · · g̈kt (u0 ); • the sequence g̈ı̂1 , . . . , g̈ı̂m partitions into two subsequences g̈̂1 , . . . , g̈̂p and g̈k̂1 , . . . , g̈k̂q such that (5.2) g̈ı̂1 · · · g̈ı̂m (uu0 ) = g̈̂1 · · · g̈̂p (u)g̈k̂1 · · · g̈k̂q (u0 ). Since both g̈i1 · · · g̈ir (uu0 ) and g̈ı̂1 · · · g̈ı̂m (uu0 ) equal w, and since g̈j1 · · · g̈js (u) and g̈̂1 · · · g̈̂p (u) have length |u|, it follows that (5.3) Then g̈j1 · · · g̈js (u) = g̈̂1 · · · g̈̂p (u); g̈k1 · · · g̈kt (u0 ) = g̈k̂1 · · · g̈k̂q (u0 ). Θ(w) = g̈i1 · · · g̈ir (vv 0 ) = g̈j1 · · · g̈js (v)g̈k1 · · · g̈kt (v 0 ) [by definition] [by (5.1)] CRYSTALLIZING THE HYPOPLACTIC MONOID   = θ g̈j1 · · · g̈js (u) θ0 g̈k1 · · · g̈kt (u0 )   = θ g̈̂1 · · · g̈̂p (u) θ0 g̈k̂1 · · · g̈k̂q (u0 ) = g̈̂1 · · · g̈̂p (v)g̈k̂1 · · · g̈k̂q (v 0 ) = g̈ı̂1 · · · g̈ı̂m (vv 0 ) = z. 27 [by (5.3)] [by (5.2)] So Θ is well-defined. By Lemma 5.5(1), Θ and its inverse preserve labelled edges. To see that Θ preserves weight, proceed as follows. Since θ and θ0 are quasicrystal isomorphisms, wt(u) = wt(v) and wt(u0 ) = wt(v 0 ). Hence wt(uu0 ) = wt(vv 0 ). Therefore if g̈i1 · · · g̈ir (uu0 ) and g̈i1 · · · g̈ir (vv 0 ) are both defined, then both sequences of operators partition as above, and so  wt(w) = wt g̈i1 · · · g̈ir (uu0 )  = wt g̈j1 · · · g̈js (u)g̈k1 · · · g̈kt (u0 )   = wt g̈j1 · · · g̈js (u) + wt g̈k1 · · · g̈kt (u0 )   = wt θ(g̈j1 · · · g̈js (u)) + wt θ0 (g̈k1 · · · g̈kt (u0 )) [since θ and θ0 preserve weights]   = wt g̈j1 · · · g̈js (θ(u)) + wt g̈k1 · · · g̈kt (θ0 (u0 )) [since θ and θ0 are quasi-crystal isomorphisms]   = wt g̈j1 · · · g̈js (v) + wt g̈k1 · · · g̈kt (v 0 )  = wt g̈j1 · · · g̈js (v)g̈k1 · · · g̈kt (v 0 )  = wt g̈i1 · · · g̈ir (vv 0 ) = wt(Θ(w)). Thus Θ is a quasi-crystal isomorphism. Hence uu0 ∼ vv 0 . Therefore the relation ∼ is a congruence.  6. The quasi-crystal graph and the hypoplactic monoid Since ∼ is a congruence, it makes sense to define the factor monoid Hn = A∗n /∼. The aim is now to show that Hn is the hypoplactic monoid hypon by showing that ∼ is equal to the relation ≡hypo on A∗n as defined in Section 4 using quasi-ribbon tableaux and Algorithm 4.3. Some of the lemmata in this section are more complicated than necessary for this aim, because they will be used in future sections to prove other results. Proposition 6.1. Let u ∈ A∗n and i ∈ {1, . . . , n − 1}. (1) If ëi (u) is defined, then std(ëi (u)) = std(u). (2) If f¨i (u) is defined, then std(f¨i (u)) = std(u). 28 ALAN J. CAIN AND ANTÓNIO MALHEIRO Proof. Suppose ëi (u) is defined, and that u contains σ symbols i and τ symbols i + 1. Then during the computation of std(u), these symbols i and i + 1 become i1 , . . . , iσ and (i + 1)1 , . . . , (i + 1)τ when subscripts are attached. Since ëi (u) is defined, u is i-inversion-free and so every symbol i in u is to the left of every symbol i + 1, and ëi (u) is obtained from u by replacing the leftmost symbol i + 1 by i. Thus, during the computation of std(ëi (u)), the symbols i and i + 1 become i1 , . . . , iσ+1 and (i+1)1 , . . . , (i+1)τ −1 . Thus, symbols i1 , . . . , iσ , (i+1)1 , . . . , (i+1)τ and i1 , . . . , iσ+1 , (i+1)1 , . . . , (i+1)τ −1 are replaced by the same symbols of the same rank in A. Hence std(ëi (u)) = std(u). This proves part (1); similar reasoning proves part (2).  Corollary 6.2. If u ∈ A∗n is a quasi-ribbon word, then every word in Γn (u) is quasi-ribbon word, and all of the corresponding quasi-ribbon tableaux have the same shape as QR(u). Proof. Let w ∈ Γn (u). By Proposition 6.1, std(w) = std(u). Thus by Proposition 4.1, w is a quasi-ribbon word, and by Proposition 4.4, QR(w) has the same shape as QR(u).  Lemma 6.3. Let i ∈ {1, . . . , n − 1}. Let u ∈ A∗n be a quasi-ribbon word. Then: (1) ëi (u) is defined if and only if QR(u) contains some symbol i + 1 but does not contain i + 1 below i in the same column; (2) f¨i (u) is defined if and only if QR(u) contains some symbol i but does not contain i + 1 below i in the same column. Proof. By the definition of the reading of a quasi-ribbon tableau, and the fact that the rows of quasi-ribbon tableaux are non-decreasing from left to right, the word u has an i-inversion if and only if i+1 and i appear in the same column of QR(u). Thus part (1) follows immediately from the definition of ëi . Part (2) follows by similar reasoning for f¨i .  Proposition 6.4. Let α be a composition. (1) The set of quasi-ribbon words corresponding to quasi-ribbon tableaux of shape α forms a single connected component of Γn . (2) In this connected component, there is a unique highest-weight word, which corresponds to the quasi-ribbon tableau of shape α whose j-th row consists entirely of symbols j, for j = 1, . . . , `(α). Proof. Let w be the quasi-ribbon word such that QR(w) has shape α and has j-th row full of symbols j, for each j = 1, . . . , `(α). Clearly ëi (w) cannot be defined for any i ≥ `(α), since there are no symbols i+1 in the word w. Furthermore, for i = 1, . . . , `(α)−1, the right-most entry i in the i-th row of T lies immediately above the left-most entry i + 1 in the i + 1-th row. Thus, by Lemma 6.3, ëi (w) is not defined. Hence w is highest-weight. (Note that it is still necessary to show that w is the unique highest-weight word in its connected component of Γn .) CRYSTALLIZING THE HYPOPLACTIC MONOID 29 Let u be a quasi-ribbon word such that QR(u) has shape α but has the property that the j-th row does not consist entirely of symbols j, for at least one j ∈ {1, . . . , `(α)}. Thus there is some symbol i that appears in the h-th row, where h < i. Without loss of generality, assume i is minimal. Clearly i > 1 since h ≥ 1. Consider two cases: • The left-most symbol in the h-th row is i. Then, by the minimality of i, the symbol i − 1 cannot appear in u, for it would have to appear on the h − 1-row of QR(u). Thus u does not contain a symbol i − 1; hence ëi−1 (u) is defined. • The left-most symbol in the h-th row is not i. Then every symbol i − 1 in QR(u) must appear in a column strictly to the left of the left-most symbol i. Hence, by Lemma 6.3, ëi−1 (u) is defined. Thus some quasi-Kashiwara operator ëi can be applied to u and so u is not a highest-weight word. This implies that w is the unique highest-weight quasi-ribbon word that has shape α. Furthermore, by Corollary 6.2, applying ëi raises the weight of a word but maintains the property of being a quasiribbon word and the shape α of its corresponding quasi-ribbon tableau. Therefore some sequence of operators ëi must transform the word u to the word w. Thus the set of quasi-ribbon words whose corresponding quasi-ribbon tableau have shape α forms a connected component.  Thus, for example, the highest-weight quasi-ribbon word corresponding to a quasi-ribbon tableau of shape (3, 1, 5, 2) is 11321333434, corresponding to the quasi-ribbon tableau 1 1 1 2 3 3 3 3 3 4 4 See also Figures 3 and 4, where the connected component Γ4 (1212), shown on the left, consists entirely of quasi-ribbon words corresponding to quasi-ribbon tableaux of shape (2, 2), and its unique highest-weight word is 1212. Note that Proposition 6.4(2) only establishes the existence of unique highest-weight words in connected components consisting of quasiribbon words, and not in every component. Corollary 6.5. Let α be a composition, and let w be a quasi-ribbon word such that QR(w) has shape α. Then w is a highest-weight word if and only if wt(w) = α. Proof. Suppose w is a highest-weight word. By Proposition 6.4(2), the j-th row of QR(w) consists entirely of symbols j, for j = 1, . . . , `(α). Thus w contains exactly αj symbols j for each j, and so wt(w) = α. 30 ALAN J. CAIN AND ANTÓNIO MALHEIRO 1 1 2 2 2121 2 2 1 1 2 3 2 2131 3 1 1 3 3 1 1 1 2 4 3 1 2 3 3 1 3 1 2 3 4 4 2 3 3 4 4 3 1 2 4 4 1 2 2 4 4 2 1 2 1 3 4 4 1 2141 3 3132 3 1 2 3 4 2 2 3 4 3 2 1 3 3131 1 1 3 4 2 2 3 3 3 2 1 1 1 4 4 1 3141 3 3232 3 1 3 3 4141 1 4142 1 4242 2 1 3142 3242 3 2 2 4143 1 4243 2 4343 Figure 4. The isomorphic components Γ4 (1212) (left) and Γ4 (2121) (right) of the quasi-crystal graph Γ4 , with elements of Γ4 (1212) drawn as quasi-ribbon tableau instead of written as words. The component Γ4 (1212) consists of all quasi-ribbon words whose quasi-ribbon tableaux have shape (2, 2). None of the words in Γ4 (2112) is a quasi-ribbon word. Now suppose that wt(w) = α. In a quasi-ribbon tableau, a symbol j can only appear in rows 1 to j. Hence the α1 symbols 1 in QR(w) must appear in row 1, which has length α1 , so this row is full of symbols 1. The α2 symbols 2 must appear in rows 1 and 2, but row 1 is full of symbols 1, so row 2, which has length α2 , is full of symbols 2. Continuing in this way, row j, which has length αj , is full of symbols j. By Proposition 6.4(2), w is a highest-weight word.  CRYSTALLIZING THE HYPOPLACTIC MONOID 31 Corollary 6.6. Let u, v ∈ A∗n be highest-weight quasi-ribbon words. If wt(u) = wt(v), then u = v. Proof. By Corollary 6.5, since u and v are highest-weight, QR(u) has shape wt(u) and QR(v) has shape wt(v). Since these weights are equal, QR(u) and QR(v) have the same shape. By the characterization of highest-weight words in Proposition 6.4(2), it follows that u = v.  Proposition 6.7. There is at most one quasi-ribbon word in each ∼class. Proof. Suppose u, v ∈ A∗n are quasi-ribbon words with u ∼ v. By Proposition 6.4, there is a sequence of operators ëi1 , . . . , ëik such that u0 = ëi1 · · · ëik (u) is highest-weight. Since u ∼ v implies that the words u and v are at the same location in the isomorphic components Γn (u) and Γn (v), the word v 0 = ëi1 · · · ëik (v) is also highest-weight. The isomorphism from Γn (u) to Γn (v) preserves weight, and so wt(u) = wt(v). Thus u = v by Corollary 6.6.  Proposition 6.7 shows that each element of Hn = A∗n /∼ has at most one representative as a quasi-ribbon word. The aim is now to show that if two words are ≡hypo -related, then they are ∼-related, which will establish a one-to-one correspondence between the elements of Hn and quasi-ribbon tableaux: Lemma 6.8. Let u ∈ A∗n . Then u and C(QR(u)) have exactly the same labelled edges incident to them in Γn . Proof. The aim is to prove that ëi (u) is defined if and only if ëi (C(QR(u))) is defined. Similar reasoning shows that f¨i (u) is defined if and only if f¨i (C(QR(u))) is defined. First, note that since u ≡hypo C(QR(u)), it follows from (4.5) that wt(u) = wt(C(QR(u))). Suppose ëi (u) is defined. Then u contains a symbol i+1 but is i-inversion-free. Hence every symbol i is to the left of every symbol i + 1 in u. Therefore Algorithm 4.3 appends the symbols i + 1 to the right of any symbols i. Hence by Lemma 6.3, ëi (C(QR(u))) is defined. On the other hand, suppose that ëi (C(QR(u))) is defined. Then by Lemma 6.3, C(QR(u)) contains a symbol i+1, but no symbol i immediately above i + 1 in a column. Hence the computation of QR(u) using Algorithm 4.3 cannot involve inserting a symbol i later than a symbol i + 1. That is, u is i-inversion-free. Since QR(u) contains a symbol i + 1, so does u, and thus ëi (u) is defined.  Lemma 6.9. Let u ∈ A∗n and i ∈ {1, . . . , n − 1}. (1) Suppose that ëi (u) is defined. Then ëi (u) ≡hypo ëi (C(QR(u))) and so ëi (C(QR(u))) = C(QR(ëi (u))). (2) Suppose that f¨i (u) is defined. Then f¨i (u) ≡hypo f¨i (C(QR(u))) and so f¨i (C(QR(u))) = C(QR(f¨i (u))). 32 ALAN J. CAIN AND ANTÓNIO MALHEIRO Proof. Suppose that ëi (u) is defined. Then it follows from Lemma 6.8 that ëi (C(QR(u))) is also defined. Now, u ≡hypo C(QR(u)) by the definition of ≡hypo . Thus it follows from (4.5) that wt(u) = wt(C(QR(u))) and so wt(ëi (u)) = wt(ëi (C(QR(u)))) since ëi replaces one symbol i + 1 by a symbol i in both words. Further, it again follows from (4.5) that DC(std(u)−1 ) = DC(std(C(QR(u)))−1 ). Hence, since ëi preserves standardizations by Proposition 6.1, it follows that DC(std(ëi (u))−1 ) = DC(std(ëi (C(QR(u))))−1 ). Combining this with the equality of weights and using (4.5) again shows that ëi (u) ≡hypo ëi (C(QR(u))). By Corollary 6.2, ëi (C(QR(u))) is a quasi-ribbon word. Since it is ≡hypo -related to ëi (u), it follows that ëi (C(QR(u))) = C(QR(ëi (u))). This completes the proof of part (1); similar reasoning proves part (2).  Proposition 6.10. Let u, v ∈ A∗n . Then u ≡hypo v =⇒ u ∼ v. Proof. By Lemmata 6.8 and 6.9, the map Θ : Γn (u) → Γn (C(QR(u))) with Θ(w) = C(QR(w)) (for w ∈ Γn (u)) is a quasi-crystal isomorphism, and so u ∼ C(QR(u)). Similarly, v ∼ C(QR(v)). It follows from u ≡hypo v that QR(u) = QR(v). Since ∼ is transitive, u ∼ v.  Let u ∈ A∗n . By Proposition 6.10, QR(u) is a quasi-ribbon word that is ∼-related to u. By Proposition 6.7, QR(u) is the unique quasi-ribbon word that is ∼-related to u. Thus the following result has been proven: Theorem 6.11. Let u, v ∈ A∗n . Then u ≡hypo v ⇐⇒ u ∼ v. Corollary 6.12. hypon = A∗n /≡hypo ' A∗n /∼ = Hn . In light of this, henceforth the quasi-crystal graph Γn is denoted Γ(hypon ), and the connected component Γn (w) is denoted Γ(hypon , w) Before moving on to study how the quasi-crystal graph interacts with the hypoplactic version of the Robinson–Schensted–Knuth correspondence, it is necessary to prove one more fundamental property of the quasi-crystal graph. Proposition 6.4(2) showed that the connected components comprising quasi-ribbon words contain unique highest-weight words. The same holds for all connected components: Proposition 6.13. In every connected component in Γ(hypon ) there is a unique highest-weight word. Proof. Let u ∈ A∗n . Since u ∼ C(QR(u)), the connected component Γ(hypon , u) is isomorphic to Γ(hypon , C(QR(u))). By Proposition 6.4(2), there is a unique highest-weight word in Γ(hypon , C(QR(u))). Consequently, Γ(hypon , u) contains a unique highest-weight word.  In the crystal graph of the plactic monoid (that is the plactic monoid of type An ), the highest-weight words are characterized combinatorially as follows. Recall from Subsection 3.4 that a Yamanouchi word is a word w in A∗n such for any suffix v of w, it holds that |v|1 ≥ |v|2 ≥ CRYSTALLIZING THE HYPOPLACTIC MONOID 33 . . . ≥ |v|n . The highest-weight words in the crystal graph of the plactic monoid are precisely the Yamanouchi words [Lot02, § 5.5]. Yamanouchi words do not characterize highest-weight words in connected components of the quasi-crystal graph Γ(hypon ). For example, the highest-weight word in Γ(hypon , 2112) is 2112, which has the suffix v = 2 that does not satisfy |v|1 ≥ |v|2 . Let max(u) denote the largest symbol in the word u. Proposition 6.14. A word u ∈ A∗n is highest-weight in a component of Γ(hypon ) if and only if it contains all symbols in {1, . . . , max(u)}, with the condition that it has an i-inversion for all i ∈ {1, . . . , max(u) − 1}. Proof. These are precisely the words for which all operators ëi are undefined.  In the crystal graph Γ(placn ), suffixes of highest-weight words are also highest-weight; this is an immediate consequence of the definition of a Yamanouchi word. This does not hold in Γ(hypon ): the highestweight word 2112 has the suffix 2, which is not highest-weight. 7. The quasi-crystal graph and the Robinson–Schensted–Knuth correspondence The previous sections constructed the quasi-crystal graph Γ(hypon ) and showed that the hypoplactic congruence ≡hypo corresponds to quasicrystal isomorphisms (that is, weight-preserving labelled digraph isomorphisms) between its connected components, just as the plactic congruence ≡plac corresponds to crystal isomorphisms (that is, weightpreserving labelled digraph isomorphisms) between connected components of the crystal graph Γ(placn ). The following result shows that the interaction of the quasi-crystal graph and the hypoplactic analogue of the Robinson–Schensted–Knuth correspondence exactly parallels the very elegant interaction of the crystal graph and the usual Robinson– Schensted–Knuth correspondence. Just as the connected components of the crystal graph Γ(placn ) are indexed by standard Young tableaux, the connected components of the quasi-crystal graph Γ(hypon ) are indexed by recording ribbons: Theorem 7.1. Let u, v ∈ A∗n . The words u and v lie in the same connected component of Γ(hypon ) if and only if RR(u) = RR(v). Proof. Suppose that u and v lie in the same connected component of Γ(hypon ). Note first that |u| = |v|. Let u = u1 · · · uk and v = v1 · · · vk , where uh , vh ∈ An . By Proposition 6.1, std(u) = std(v). Therefore std(u1 · · · uh ) = std(v1 · · · vh ) for all h (this is immediate from the definition of standardization [Nov00, Lemma 2.2]). Thus QR(u1 · · · uh ) and QR(v1 · · · vh ) both have shape DC(std(u1 · · · uh )−1 ) for all h. The sequence of these shapes determines where new symbols are inserted 34 ALAN J. CAIN AND ANTÓNIO MALHEIRO during the computation of QR(u) and QR(v) by Algorithm 4.3, and so RR(u) = RR(v). Now suppose RR(u) = RR(v). Let û and v̂ be the highest-weight words in Γ(hypon , u) and Γ(hypon , v), respectively. By the forward implications, RR(û) = RR(u) = RR(v) = RR(v̂). Note that û ∼ C(QR(û)) and v̂ ∼ C(QR(v̂)), and C(QR(û)) and C(QR(v̂)) are highestweight quasi-ribbon words. Furthermore, the quasi-ribbon tableaux QR(û) and QR(v̂) have the same shape, since they both have the same shape as RR(û) = RR(v̂). By Corollary 6.5, wt(C(QR(û))) = wt(C(QR(v̂))); thus C(QR(û)) = C(QR(û)) by Corollary 6.6. Since QR(û) = QR(û) and RR(û) = RR(v̂), it follows that û = v̂ by the quasi-ribbon tableau version of the Robinson–Schensted–Knuth correspondence (see the discussion in Section 4). Hence Γ(hypon , u) = Γ(hypon , û) = Γ(hypon , v̂) = Γ(hypon , v). This completes the proof of the reverse implication.  8. Structure of the quasi-crystal graph The aim of this section is to study the structure of quasi-crystal graph and how it interacts with usual crystal graph. Many of these results are illustrated in the components of Γ(hypo4 ) and Γ(plac4 ) shown in Figures 5 to 8. These figures show both the actions of quasi-Kashiwara operators (indicated by solid lines) and other actions of Kashiwara operators (indicated by dotted lines). (Recall from Proposition 5.2 that the action of the quasi-Kashiwara operators is a restriction of the action of the Kashiwara operators.) 8.1. Sizes of classes and quasi-crystals. As previously noted, by fixing a Young tableau P and varying Q over all standard Young tableaux of the same shape as P , one obtains the plactic class corresponding to P . Notice that one obtains as a corollary that the size of this class depends on the shape of P , not on the entries of P . The so-called hook-length formula gives the number of standard Young tableaux of a given shape (see [FRT54] and [Ful97, § 4.3]) and thus of the size of a plactic class corresponding to an element of that shape. This section uses the quasi-crystal graph to give an analogue of the hook-length formula for quasi-ribbon tableaux, in the sense of stating a formula for the size of hypoplactic classes corresponding to quasiribbon tableau of a given shape. Novelli [Nov00, Theorem 5.1] proves that for any compositions α = (α1 , . . . , αh ) and γ,   X α1 + . . . + αh (8.1) #(β, γ) = , α 1 , . . . , αh βα where #(β, γ) is the size of the hypoplactic class corresponding to the quasi-ribbon tableau of shape β and content γ. (Recall that β  α denotes that β is a coarser composition than α; see Section 4.) Since CRYSTALLIZING THE HYPOPLACTIC MONOID 1111 Legend 1112 QuasiKashiwara operators 1122 1113 1222 1123 1114 z}|{ 1 2 3 2222 1223 1133 1124 2223 1233 1224 1134 2233 1333 1234 2224 1144 35 4321 2333 2234 1334 1244 3333 2334 2244 1344 3334 2344 1444 3344 2444 3444 4444 Figure 5. The connected components Γ(hypo4 , 1111) and Γ(hypo4 , 4321) of the quasi-crystal graph Γ(hypo4 ). These are also the connected components Γ(plac4 , 1111) and Γ(plac4 , 4321) of the crystal graph Γ(plac4 ): there is no vertex in these components where some ẽi or f˜i is defined but the corresponding ëi or f¨i is not. These two components are not isomorphic to any other components of either Γ(hypo4 ) or Γ(plac4 ). The other components of Γ(hypon ) and Γ(placn ) whose vertices are length-4 words are shown in Figures 6 and 7. the size of the hypoplactic class corresponding to a quasi-ribbon tableau of shape (α1 ) is always 1, the formula (8.1) allows one to iteratively compute the size of an arbitrary hypoplactic class. However, it is not immediately clear from (8.1) that the size of a hypoplactic class is only dependent on the shape of the quasi-ribbon tableau, not on its content: Proposition 8.1. Hypoplactic classes corresponding to quasi-ribbon tableaux of the same shape all have the same size. Proof. Let u, v ∈ A∗n be such that QR(u) and QR(v) have the same shape. Then C(QR(u)) and C(QR(v)) lie in the same connected component by Proposition 6.4(1). Thus there is a sequence g̈i1 , . . . , g̈ir of operators ëi and f¨i such that g̈i1 · · · g̈ir (C(QR(u))) = C(QR(v)). ¨ Each is a bijection between ≡hypo -classes, and  ëi and fi, when defined,   so C(QR(u)) ≡hypo and C(QR(v)) ≡hypo have the same size. Since u ∼ C(QR(u)) and v ∼ C(QR(v)), it follows that [u]≡hypo and [v]≡hypo have the same size.  36 ALAN J. CAIN AND ANTÓNIO MALHEIRO 2111 to Γ(hypo4 , 3212) (See Fig. ??) 3111 2112 3112 4111 2113 2122 2131 3122 4112 3113 2114 2123 3131 2141 2132 2133 2124 3132 3141 2142 3222 4122 4113 3123 3114 2121 4222 4123 3223 3124 4114 3133 2134 3232 3142 4141 2143 4223 4133 3224 4124 3233 3134 2144 3242 4142 3143 4233 4224 4134 3234 3144 4242 4143 3243 4333 4234 3244 4144 4243 4334 4244 4343 1211 4344 1311 1212 1312 1411 1213 2212 2311 1322 1412 1313 1214 2213 3311 2411 2312 2313 2214 3312 3411 2412 2322 1422 1413 1323 1314 2211 2422 1423 2323 1324 1414 3313 2314 3322 3412 4411 2413 2423 1433 2324 1424 3323 3314 2414 3422 4412 3413 2433 2424 1434 3324 3414 4422 4413 3423 3433 2434 3424 4414 4423 3434 4424 4433 1121 4434 1131 1221 1132 1141 1231 2221 1232 1142 1331 1241 2231 2232 1242 1143 1332 1341 2331 2241 2242 1243 2332 1342 1441 3331 2341 2243 1343 2342 1442 3332 3341 2441 2343 2442 1443 3342 3441 3343 2443 3442 4441 3443 4442 Legend QuasiKashiwara operators z}|{ 1 2 3 Kashiwara operators z}|{ 1 2 3 4443 Figure 6. Thirteen components of Γ(hypo4 ), which together lie inside five components of Γ(placn ). Dotted lines show actions of operators ẽi and f˜i that are not also actions of operators ëi and f¨i . Each of these components of Γ(hypo4 ) is isomorphic to at least one other component; grey lines link isomorphic components. Other components are shown in Figures 5 and 7. CRYSTALLIZING THE HYPOPLACTIC MONOID to Γ(hypo4 , 2132) (See Fig. ??) Legend 4211 3212 4311 4212 3213 QuasiKashiwara operators 4312 4213 3214 4322 4313 4214 4323 4314 3211 3121 4324 4121 3221 4131 4221 3231 4132 4231 3241 4232 4331 4241 4332 4341 1321 37 z}|{ 1 2 3 Kashiwara operators z}|{ 1 2 3 4342 1421 2321 1431 2421 3321 1432 2431 3421 2432 3431 4421 3432 4431 4432 Figure 7. Nine components of Γ(hypo4 ), which together lie inside three of the connected components Γ(placn ). Dotted lines show actions of operators ẽi and f˜i that are not also actions of operators ëi and f¨i . Each of these components of Γ(hypo4 ) is isomorphic to at least one other component; grey lines link isomorphic components. The other components whose vertices are length-4 words are shown in Figures 5 and 6. The formula for the size of hypoplactic classes is also straightforward when one uses the quasi-crystal graph: Theorem 8.2. The size of any hypoplactic class in A∗n whose quasiribbon tableau has shape α is (8.2) X    (−1)`(α)−`(β)   βα 0 |β| β1 , · · · , β`(β)  if `(α) ≤ n, otherwise. 38 ALAN J. CAIN AND ANTÓNIO MALHEIRO Proof. Note first that in a quasi-ribbon tableau, any symbol in An must lie in the first n rows. Thus if `(α) > n, then there is no quasiribbon tableau of shape α with entries in An and so the corresponding hypoplactic class if empty. So assume henceforth that `(α) ≤ n. Let T be a quasi-ribbon tableau of shape α; the aim is to describe the cardinality of the set UT = { u ∈ A∗n : QR(u) = T }. Since the operators ëi and f¨i are bijections between hypoplactic classes, assume without loss of generality that C(T ) is highest-weight. By Corollary 6.5, wt(C(T )) = α. Since all words in a hypoplactic class have the same weight, every word in UT has weight α (and thus contains exactly αi symbols i, for each i ∈ An ). Furthermore, all words in UT are highestweight and so have i-inversions for each i ∈ {1, . . . , `(α)}. Thus UT consists of exactly the words of weight α, but that do not have the property of containing all symbols i to the left of all symbols i + 1 for some i ∈ {1, . . . , `(α) − 1}. For any weak composition γ with `(γ) ≤ n, the number of words in A∗n with weight γ is   |γ| . γ1 , · · · , γ`(γ) Consider a word u ∈ A∗n with wt(u) = γ that is i-inversion-free. Then in u, every symbol i lies to the left of every symbol i + 1. Replacing each symbol j + 1 by j for each j ≥ i yields a word u0 with weight γ 0 = (γ1 , . . . , γi−1 , γi + γi+1 , γi+2 . . . , γn ); note that γ 0  γ and `(γ 0 ) = `(γ)−1. On the other hand, starting from u0 and replacing each symbol j by j + 1 for j ≥ i + 1 and replacing the rightmost γi+1 symbols i by i+1 yields u. Thus there is a one-to-one correspondence between words of weight γ that are i-inversion-free and words of weight γ 0 . Iterating this argument shows that there is a one-to-one correspondence between words of weight γ that are i-inversion-free for i ∈ I ⊆ {1, . . . , n − 1} and words of weight γI for a (uniquely determined) γI  γ. Hence, by the inclusion–exclusion principle, the number of such words that are i-inversion-free for all i ∈ {1, . . . , n − 1} is given by (8.2).  For example, let α = (2, 1, 1, 2). By Theorem 8.2, the size of the hypoplactic class whose quasi-ribbon tableau has shape α is 6! 6! 6! 6! − − − 2!1!1!2! 3!1!2! 2!2!2! 2!1!3! 6! 6! 6! + + + 4!2! 3!3! 2!4! 6! − = 19. 6! CRYSTALLIZING THE HYPOPLACTIC MONOID 39 The quasi-ribbon word 143214 corresponds to a quasi-ribbon tableau with shape α, and the hypoplactic class containing u is {143214, 413214, 431214, 432114, 143241, 413241, 431241, 432141, 143421, 413421, 431421, 432411, 144321, 414321, 434121, 434211, 441321, 443121, 443211}, which contains 19 elements, as expected. For the sake of completeness, this section closes with a formula for the number of quasi-ribbon tableaux of a given shape, which allows one to compute the size of a connected component of Γ(hypon ). Notice that the proof of this does not depend on applying the crystal structure. Theorem 8.3. The number of quasi-ribbon tableaux of shape α and symbols from An is ( n+|α|−`(α) n−`(α) 0  if `(α) ≤ n if `(α) > n. Proof. Consider a ribbon diagram of shape α, where the boundaries between the cells, including the left boundary of the first cell and the right boundary of the last cell, are indexed by the numbers 0, 1, . . . , |α|. Notice that D(α) be the set of indices of boundaries between vertically adjacent cells. For example, for shape α = (4, 4, 2, 3), the indices are as follows: 0 1 2 3 6 7 9 4 5 12 13 8 10 11 In this case, D(α) = {4, 8, 10}. A filling of such a quasi-ribbon tableau by symbols from An is weakly increasing from upper left to lower right, and is specified exactly by listing n − 1 boundaries between adjacent cells where the increase from i to i + 1 occurs in this filling. Note that such a list may contain repeated entries (and is thus formally a multiset), indicating that the difference between the entries in the cells incident on this boundary 40 ALAN J. CAIN AND ANTÓNIO MALHEIRO differ by more than 1. For example, for n = 9 the filling 0 1 2 3 2 4 4 4 6 7 4 5 5 6 7 9 12 13 5 8 8 8 10 9 9 9 11 corresponds to the multiset h0, 1, 1, 4, 6, 7, 8, 10i. Note that this multiset has length n − 1 = 8 and contains the entry 0, indicating the presence of 2 in the first cell, and a repeated entry 1, indicating the jump from 2 to 4 at this boundary. However, some of the entries in this multiset are forced by the shape of the tableau: there must be increases at boundaries 4, 8, and 10, which are between vertically adjacent cells. There is thus a one-to-one correspondence between multisets with n − 1 elements drawn from {0, . . . , |α|} and that contain D(α) (which indicates the position of the ‘forced’ increases) and fillings of a tableau of shape α. Since |D(α)| = `(α)−1, the number of such multisets is 0 if `(α) > n, and is otherwise the number of multisets with (n−1)−(`(α)− 1) elements drawn from {0, . . . , |α|}, which is (|α|+1)+((n−1)−(`(α)−1))−1 (n−1)−(`(α)−1) by the standard formula forthe number of multisets [Sta12, § 1.2], which simplifies to n+|α|−`(α) .  n−`(α) 8.2. Interaction of the crystal and quasi-crystal graphs. This section examines the interactions of the crystal graph Γ(placn ) and quasi-crystal graph Γ(hypon ). The first, and most fundamental, observation, is how connected components in Γ(placn ) are made up of connected components in Γ(hypon ): Proposition 8.4. The vertex set of every connected component of Γ(placn ) is a union of vertex sets of connected components of Γ(hypon ). Proof. By Proposition 5.2, any edge in Γ(hypon ) (whose edges indicate the action of the quasi-Kashiwara operators ëi and f¨i ) is also an edge in Γ(placn ) (whose edges indicate the action of the Kashiwara operators ẽi and f˜i ). Hence every connected component in Γ(hypon ) lies entirely within a connected component of Γ(placn ); the result follows immediately.  For example, as shown in Figure 6, the connected component Γ(plac4 , 2111) is made up of the three connected components Γ(hypo4 , 2111), Γ(hypo4 , 2112), and Γ(hypo4 , 2122). CRYSTALLIZING THE HYPOPLACTIC MONOID 41 Proposition 8.5. Let Θ : Γ(placn , u) → Γ(placn , Θ(u)) be a crystal isomorphism. Then Θ restricts to a quasi-crystal isomorphism from Γ(hypon , w) to Γ(hypon , Θ(w)), for all w ∈ Γ(placn , u). Proof. Let Θ : Γ(placn , u) → Γ(placn , Θ(u)) be a crystal isomorphism and let w ∈ Γ(placn , u). Suppose the Kashiwara operator f˜i is defined on w but the quasi-Kashiwara operator f¨i is not defined on w. Since f˜i is defined, w must contain at least one symbol i. Thus, since f¨i is undefined, w must have an i-inversion. Since Θ is a crystal isomorphism, w ≡plac Θ(w). The defining relations in Rplac preserve the property of having i-inversions (for if a defining relation in Rplac commutes a symbol i and i + 1, then a = i and c = i + 1 in this defining relation, and so b ∈ {i, i + 1}, and so the applied relation is either (i(i + 1)i, (i + 1)ii) or ((i + 1)i(i + 1), (i + 1)(i + 1)i), and both sides of these relations have i-inversions). Hence Θ(w) has an i-inversion. Hence f¨i is not defined on w. A symmetric argument shows that if f˜i is defined on Θ(w) but f¨i is not, then f¨i is not defined on w. Hence the isomorphism Θ maps edges corresponding to actions of quasi-Kashiwara operators in Γ(placn , u) to edges corresponding to actions of quasi-Kashiwara operators in Γ(placn , Θ(u)), and vice versa, and so restricts to a quasicrystal isomorphism from Γ(hypon , w) to Γ(hypon , Θ(w)) for all w ∈ Γ(placn , u).  Notice that it is possible for a single connected component of Γ(placn ) to contain distinct isomorphic connected components of Γ(hypon ). For example, as shown in Figure 8, the connected component Γ(plac4 , 321211) contains the isomorphic connected components Γ(hypo4 , 321213) and Γ(hypo4 , 321312). Note also that Γ(plac4 , 321211), contains the isomorphic one-vertex connected components Γ(hypo4 , 421323) and Γ(hypo4 , 321423). (There are other one-vertex components of Γ(hypo4 ) inside Γ(plac4 , 321211), but these have different weights and so there are no quasi-crystal isomorphisms between them.) The quasi-ribbon tableau QR(421323) has shape (1, 2, 2, 1); thus, by Theorem 8.2, the hypoplactic class containing 421323 has size 6! 6! 6! 6! − − − 1!2!2!1! 3!2!1! 1!4!1! 1!2!3! 6! 6! 6! + + + 5!1! 3!3! 1!5! 6! − = 61. 6! Since there are an odd number of components that are isomorphic to Γ(hypo4 , 421323), it follows that there must be at least one component of Γ(plac4 ) that contains an odd number of these components. 42 ALAN J. CAIN AND ANTÓNIO MALHEIRO 321211 Legend 421211 321311 321212 421311 321411 421212 321312 321213 QuasiKashiwara operators 431311 421312 421411 421213 321412 321214 321322 321313 431312 421322 431411 421313 421412 421214 321422 321314 321413 321323 z}|{ 1 2 3 431322 421323 431412 431313 421413 421314 421422 321414 321423 321324 Kashiwara operators 432322 421324 431422 431323 431413 431314 421423 421414 321433 321424 z}|{ 432422 432323 431423 431324 421433 431414 321434 421424 432423 432324 431433 431424 421434 432433 432424 431434 1 2 3 432434 Figure 8. The connected component Γ(plac4 , 321211), drawn so that the connected components of Γ(hypo4 ) are arranged vertically. Notice that it contains the isomorphic connected components Γ(hypo4 , 321213) and Γ(hypo4 , 321312), and also the isomorphic one-vertex connected components Γ(hypo4 , 421323) and Γ(hypo4 , 321423). Thus different components of Γ(plac4 ) may contain different numbers of isomorphic components of Γ(hypo4 ). Corollary 8.6. Let u, v ∈ A∗n be such that u ≡plac v but u 6= v, so that there is a non-trivial crystal isormorphism Θ : Γ(placn , u) → Γ(placn , v) with Θ(u) = v. Let s ∈ Γ(placn , u) and t ∈ Γ(placn , v) be quasi-ribbon words. Then Θ does not map Γ(hypon , s) to Γ(hypon , t). More succinctly, quasi-ribbon word components of Γ(hypon ) cannot lie in the same places in distinct isomorphic components of Γ(placn ). Proof. Without loss of generality, assume s and t are highest-weight in Γ(hypon , s) and Γ(hypon , t) respectively. Suppose, with the aim of obtaining a contradiction, that Θ maps Γ(hypon , s) to Γ(hypon , t). Then Θ(s) = t, and so s ≡hypo t. Since s and t are quasi-ribbon words, s = t by Proposition 6.4 and so Θ is trivial, which is a contradiction.  Corollary 6.2 showed that the quasi-Kashiwara operators preserve shapes of quasi-ribbon tableau. In fact, quasi-Kashiwara operators and, more generally, Kashiwara operators, preseve shapes of quasi-ribbon tabloids (see Section 4 for the definitions of quasi-ribbon tabloids): CRYSTALLIZING THE HYPOPLACTIC MONOID 43 Proposition 8.7. Let i ∈ {1, . . . , n − 1}. Let u ∈ A∗n . (1) If the Kashiwara operator ẽi is defined on u, then QRoid(ẽi (u)) and QRoid(u) have the same shape. (2) If the Kashiwara operator f˜i is defined on u, then QRoid(f˜i (u)) and QRoid(u) have the same shape. Proof. Let u ∈ A∗n and let u = u(1) · · · u(m) be the factorization of u into maximal decreasing factors (which are entries of the columns of QRoid(u)). Suppose that the Kashiwara operator ẽi is defined on u, and that the application of ẽi to u replaces the (necessarily unique) symbol i + 1 in u(k) by a symbol i; let û(k) be the result of this replacement. Then u(k) cannot contain a symbol i, for if it did, then during the computation of the action of ẽi as described in Subsection 3.3, the symbols i + 1 and i in u(k) (which would be adjacent since u(k) is strictly decreasing) would have been replaced by − and + and so would have been deleted, and so ẽi would not act on this symbol i + 1. Hence û(k) is also a decreasing word. Furthermore, the first symbol of u(k+1) is greater than or equal to the last symbol of u(k) , and so is certainly greater than or equal to the last symbol of û(k) since ẽi can only decrease a symbol. Similarly, the first symbol of u(k) is greater than or equal to the last symbol of u(k−1) . If u(k) does not start with the symbol i + 1, the first symbol of û(k) is greater than or equal to the last symbol of u(k−1) . So assume u(k) starts with the symbol i + 1; since the factorization is into maximal decreasing factors, u(k−1) ends with a symbol that is less than or equal to i + 1. If u(k−1) ends with a symbol that is strictly less than i + 1, then the first symbol of û(k) is greater than or equal to the last symbol of u(k−1) . So assume u(k−1) ends with the symbol i + 1. Then during the computation of the action of ẽi as described in Subsection 3.3, the adjacent symbols i + 1 at the end of u(k−1) and at the start of u(k) are both replaced by symbols −, and neither of these symbols are removed by deletion of factors −+, since ẽi acts on the symbol i + 1 at the start of u(k) . But this contradicts the fact that ẽi acts on the symbol replaced by the leftmost −. Thus this case cannot arise, and so one of the previous possibilities must have held true. Combining the last three paragraphs shows that the factorization of ẽi (u) into maximal decreasing factors is ẽi (u) = u(1) · · · u(k−1) û(k) u(k+1) · · · u(m) . This proves part (1). Similar reasoning for f˜i proves part (2).  Proposition 8.8. A connected component of Γ(placn ) contains at most one quasi-ribbon word component of Γ(hypon ). Proof. Suppose the connected component Γ(placn , u) contains connected components Γ(hypon , w) and Γ(hypon , w0 ) that both consist of quasiribbon words. Without loss of generality, assume that w is highest- 44 ALAN J. CAIN AND ANTÓNIO MALHEIRO 1 21 2 2 1 21 3 2 1 31 3 1 1 21 4 3 1 32 3 1 3 3 2 42 4 2 2 43 4 2 1 2 1 31 4 3 2 32 3 2 32 4 1 3 1 3 1 32 4 1 1 1 3 1 42 4 3 1 22 4 2 1 41 4 1 1 22 3 1 23 4 2 1 33 4 3 1 2 2 2 1 43 4 33 4 3 2 3 43 4 Figure 9. The component Γ(plac4 , 2121), with words drawn as quasi-ribbon tabloids of the same shape. Only the component Γ(hypo4 , 2132) consists of quasi-ribbon words, which thus appear here as quasi-ribbon tableaux. weight in Γ(hypon , w) and w0 has highest-weight in Γ(hypon , w0 ). Since w and w0 are in the connected component Γ(placn , u), the quasi-ribbon tableaux QR(w) and QR(w0 ) have the same shape by Proposition 8.7. Hence by Corollary 6.5, wt(w) = wt(w0 ) and so w = w0 by Corollary 6.6. Thus Γ(hypon , w) = Γ(hypon , w0 ).  It is possible that a connected component of Γ(placn ) contains no quasi-ribbon word components of Γ(hypon ). For example, as can be seen in Figure 6, Γ(plac4 , 2211) contains the connected components Γ(hypo4 , 2211) and Γ(hypo4 , 2312), and neither 2211 nor 2312 is a quasi-ribbon word. Thus the next aim is to characterize those connected components of Γ(placn ) that contain a (necessarily unique) quasi-ribbon word component of Γ(hypon ). In order to do this, it CRYSTALLIZING THE HYPOPLACTIC MONOID 45 is useful to discuss a shortcut that allows one to calculate quickly the Young tableau P(w) obtained when w is a quasi-ribbon word. The slide up–slide left algorithm takes a filled quasi-ribbon diagram D and produces a filled Young diagram as follows: Start from the quasiribbon diagram D. Slide all the columns upwards until the topmost entry of each is on row 1. Now slide all the symbols leftwards along their rows until the leftmost entry in each row is in the first column and there are no gaps in each row. As will be shown in Proposition 8.9(1), applying the slide up–slide left algorithm to a quasi-ribbon tableau QR(u) gives the Young tableau P(u), as in the following example: QR(1325436768) = 1 2 3 3 4 5 6 6 7 8 1 2 3 3 4 6 6 8 5 7 1 2 3 6 6 8 3 4 7 5 1 2 3 6 6 8 3 4 7 5 1 2 3 6 6 8 = P(1325436768). 3 4 7 5 Similarly, as will be shown in Proposition 8.9(2), applying the slide up–slide left algorithm to a quasi-ribbon tableau of the same shape α as QR(u), filled with entries 1, . . . , `(α), gives the standard Young tableau Q(u), as in the following example: 1 2 3 4 5 6 7 8 9 10 1 2 4 7 8 10 = Q(1325436768). 3 5 9 6 Proposition 8.9. Let T be a quasi-ribbon tableau of shape α. (1) Applying the slide up–slide left algorithm to T yields the Young tableau P(C(T )). 46 ALAN J. CAIN AND ANTÓNIO MALHEIRO (2) Applying the slide up–slide left algorithm to the (unique) quasiribbon tableau of shape α filled with entries 1, . . . , |α| yields the standard Young tableau Q(C(T )). Proof. Let U be the unique quasi-ribbon tableau of shape α filled with entries 1, . . . , |α|. The proof is by induction on the number of columns m in a ribbon diagram of shape α. Suppose m = 1. Then applying the slide up–slide left algorithm to T and U yields T and U , respectively (viewed as filled Young diagrams). Since T satisifies the condition for being a Young tableau, P(C(T )) = T . Furthermore, U is also the unique standard Young tableau with a single column of the same shape as T , and so must be Q(C(T )). Now suppose m > 1 and that the result holds for all quasi-ribbon tableau T 0 with fewer than m columns. In particular, it holds for the quasi-ribbon tableau T 0 formed by the first m − 1 columns of T . In particular, by applying the slide up–slide left algorithm to T 0 , one obtains P(C(T 0 )). If the m-th column of T contains entries a1 , . . . , ak (listed from bottom to top, so that a1 > . . . > ak ), then applying the slide up–slide left algorithm to T yields the filled Young diagram obtained by applying it to T 0 (which yields P(C(T 0 ))) and then adding the symbol ah at the rightmost end of the h-th row. Since each symbol a1 , . . . , ak is greater or equal to than every symbol in T 0 and thus in P(C(T 0 )), using Algorithm 4.2 to insert the symbols a1 , . . . , ak into P(C(T 0 )) does not involve bumping any symbols in P(C(T 0 )). That is, the symbols a1 , . . . , ak (which form a strictly decreasing sequence) bump each other up the rightmost edge of P(C(T 0 )), as in the following example 0 P(C(T )) a2 a1 a 3 ← a4 = 0 P(C(T )) a4 a3 a2 a1 Furthermore, applying the slide up–slide left algorithm to the unique quasi-ribbon tableau of the same shape as T 0 filled with entries 1, . . . , (|α|− k) yields the standard Young tableau Q(C(T 0 )). Thus applying the slide up–slide left algorithm to the unique quasi-ribbon tableau of shape α filled with entries 1, . . . , |α| yields the filled Young diagram obtained by applying it to Q(C(T 0 )) and adding |α| − k + h to the end of the h-th row for h = 1, . . . , k. By the above analysis of the behaviour of Algorithm 4.2, this is the standard Young tableau Q(C(T )).  Proposition 8.10. Let Q be a standard Young tableau. There is at most one quasi-ribbon tableau T such that Q can be obtained by applying the slide up–slide left algorithm to T . Proof. Suppose that T and T 0 are quasi-ribbon tableaux such that applying the slide up–slide left algorithm to T and T 0 yields Q. Then CRYSTALLIZING THE HYPOPLACTIC MONOID 47 by Proposition 8.9(1–2), P(C(T )) = Q(C(T )) = Q = P(C(T 0 )) = Q(C(T 0 )). Thus by the Robinson–Schensted–Knuth correspondence, C(T ) = C(T 0 ) and so T = T 0 .  Proposition 8.11. A connected component Γ(placn , w) contains a quasiribbon word component if and only if Q(w) can be obtained by applying the slide up–slide left algorithm to some quasi-ribbon tableau T of shape α and entries 1, . . . , |α| and such that `(α) ≤ n. Proof. Suppose Γ(placn , w) contains a quasi-ribbon word component. Let u ∈ Γ(placn , w) be a quasi-ribbon word. Let α be the shape of QR(u). Note that QR(u) has at most n rows, so `(α) ≤ n. By Proposition 8.9(2), applying the slide up–slide left algorithm to the unique quasi-ribbon tableau of shape α filled with entries 1, . . . , |α| yields the standard Young tableau Q(u). Since u and w are in the same connected component of Γ(placn ), the standard Young tableaux Q(w) and Q(u) are equal. On the other hand, suppose that Q(w) can be obtained by applying the slide up–slide left algorithm to some quasi-ribbon tableau T of shape α and entries 1, . . . , |α| and such that `(α) ≤ n. Let U be the quasi-ribbon tableau of shape α and entries in An such that C(U ) is highest-weight in its component of Γ(hypon ); note that U exists since `(α) ≤ n. Then by Proposition 8.9(2), Q(C(U )) = Q(w) since U has shape α. Hence C(U ) ∈ Γ(placn , w).  Define a permutation σ of {1, . . . , n} to be interval-reversing if there is some composition α = (α1 , . . . , αk ) with |α| = n such that for all h = 0, . . . , k, the permutation σ preserves the interval {α1 + . . . + αh + 1, . . . , α1 + . . . + αh+1 } and reverses the order of its elements. (For h = 0, this interval is {1, . . . , α1 }.) Thus, for example,   1 2 3 4 5 6 7 8 1 5 4 3 2 8 7 6 is an interval-reversing permutation of {1, . . . , 8}, where the appropriate composition α is (1, 4, 3). It is clear that there is only one choice for α. It is well-known that if w is a standard word (and thus a permutation), then the Robinson–Schensted–Knuth correspondence associates w with (P(w), Q(w)) and w−1 with (Q(w), P(w)). That is, Q(w) = P(w−1 ) and P(w) = Q(w−1 ). Thus involutions are associated to pairs (Q, Q), where Q is a standard Young tableau. Proposition 8.12. A connected component Γ(placn , w) contains a quasiribbon word component if and only if the Robinson–Schensted–Knuth correspondence associates (Q(w), Q(w)) with an interval-reversing involution. 48 ALAN J. CAIN AND ANTÓNIO MALHEIRO Proof. Suppose Γ(placn , w) contains a quasi-ribbon word. By Proposition 8.11, Q(w) can be obtained by applying the slide up–slide left algorithm to some quasi-ribbon tableau T of shape β and entries 1, . . . , |β| and such that `(β) ≤ n. By Proposition 8.9(1–2), P(C(T )) = Q(C(T )) = Q(w). So the Robinson–Schensted–Knuth correspondence associates (Q(w), Q(w)) wih C(T ). Let α = (α1 , . . . , αk ) be such that αi is the length of the i-th column of T . Then by the definition of T and its column reading, C(T ) is an interval-reversing permutation where the appropriate composition is α. On the other hand, suppose the Robinson–Schensted–Knuth correspondence associates (Q(w), Q(w)) with an interval-reversing involution u, where the appropriate composition is α = (α1 , . . . , αk ). Note that Q(w) = P(u) and Q(w) = Q(u). Let T be the quasi-ribbon tabloid QRoid(u). Then αi is the length of the i-th column of T , which is filled with the image of the interval {α1 + . . . + αi + 1, . . . , α1 + . . . + αi+1 } since u is an interval-reversing permutation. Since the elements in each interval are less than the elements in the next, the rows of T are increasing from left to right and so T is a quasi-ribbon tableau that is filled with the symbols 1, . . . , |u|. Note that u = C(T ), and so Q(w) = Q(u) = Q(C(T )). So by Proposition 8.9, Q(w) can be obtained from T by applying the slide up–slide left algorithm. Hence Γ(placn , w) contains a quasi-ribbon word by Proposition 8.11.  Corollary 8.13. Let w ∈ A∗n and let λ = (λ1 , . . . , λ`(λ) ) be the shape of P(w). The number of connected components of Γ(placn ) that are isomorphic to Γ(placn , w) and contain quasi-ribbon word components of Γ(hypon ) is   λ1  if |λ| − λ1 + 1 ≤ n, λ1 − λ2 , . . . , λ`(λ)−1 − λ`(λ) , λ`(λ)  0 otherwise. Proof. By Proposition 8.11, the number of such components is equal to the number of standard Young tableaux of shape λ that can be obtained by applying the slide up–slide left algorithm to some quasiribbon tableau of some shape α and entries 1, . . . , |α| and such that `(α) ≤ n. By the definition of a quasi-ribbon diagram, the number of rows `(α) of such a diagram is equal to the total number of boxes minus the number of columns plus 1 (since adjacent pairs of columns overlap in exactly one row). Since the slide up–slide left algorithm preserves the number of columns and the total number of boxes, this number of rows is |λ| − λ1 + 1. Hence if |λ| − λ1 + 1 > n, there is no standard Young tableau that can be obtained in such a way. So suppose henceforth that |λ| − λ1 + 1 ≤ n. For any composition α, there is exactly one quasi-ribbon tableau of shape α filled with symbols 1, . . . , |α|, and each of them yields (under the slide up–slide left algorithm) a different standard Young tableau by CRYSTALLIZING THE HYPOPLACTIC MONOID 49 Proposition 8.11. Furthermore, there is a one-to-one correspondence between compositions α and shapes of tabloids: one simply takes a ribbon diagram of shape α and applies the ‘slide up’ part of the slide up–slide left algorithm. Thus, to obtain the number of such quasiribbon tableaux that give the correct shape on applying the slide up– slide left algorithm, it suffices to count the number of tabloid shapes that have λj cells on the j-th row, for each j = 1, . . . , `(λ). This is equal to the number of tabloid shapes that have λ1 cells on the first row and λj − λj+1 columns ending on the j-th row, which in turn is equal to the number of words of length λ1 containing λj − λj+1 symbols j for each j < `(λ) and λ`(λ) symbols `(λ).  Note that the number of components specified by Corollary 8.13 is dependent only on the shape of P(w), not on wt(w) or w itself. This is what one would expect in working with a class of isomorphic components in Γ(placn ). Corollary 8.14. Let w ∈ A∗n , and let λ = (λ1 , . . . , λm ) be the shape of P(w). Suppose |λ| − λ1 + 1 ≤ n. Then Γ(placn , w) contains at least   λ1 λ1 − λ2 , . . . , λ`(λ)−1 − λ`(λ) , λ`(λ) connected components of Γ(hypon ). Proof. Corollary 8.13 gives the number of components isomorphic to Γ(placn , w) that contain quasi-ribbon word components. Each one of these quasi-ribbon word components is isomorphic to a connected component of Γ(hypon ) inside Γ(placn , w) by Proposition 8.5, and to distinct such connected components of Γ(hypon ) by Corollary 8.6. Hence the number of connected components of Γ(hypon ) inside Γ(placn , w) must be at least the number given in Corollary 8.13 in the case where |λ| − λ1 + 1 ≤ n.  8.3. The Schützenberger involution. The Schützenberger involution is the map ] : A∗n → A∗n that sends a ∈ An to n − a + 1 and is extended to A∗n by (a1 · · · ak )] 7→ ak ] · · · a1 ] . That is, given a word, one obtains its image under ] by reversing the word and replacing each symbol a by n − a + 1. It is well-known that the Schützenberger involution reverses the order of weights, in the sense that if u has higher weight than v, then v ] has higher weight than u] . Proposition 8.15. In Γ(hypon ), the Schützenberger involution maps connected components to connected components. If there is an edge from u to v labelled by i, then there is an edge from v ] to u] labelled by n − i. Proof. Clearly the second statement implies the first. So suppose that f¨i (u) = v. Then u contains at least one symbol i and, every symbol i is 50 ALAN J. CAIN AND ANTÓNIO MALHEIRO to the left of every symbol i + 1, and v is obtained from u by replacing the rightmost symbol i by i + 1. By the definition of ] , in the word u] , every symbol n − i + 1 is to the right of every symbol n − i. Hence ën−i (u] ) is defined, is equal to the word obtained from u] by replacing the leftmost symbol n − i + 1 by n − i, which is v ] . Hence there is an edge from v ] to u] labelled by n − i.  8.4. Characterizing quasi-crystal graphs? Stembridge [Ste03] gives a set of axioms that characterize connected components of crystal graphs. These axioms specify ‘local’ conditions that the graph must satisfy. It is natural to ask whether there is an analogous characterization for quasi-crystal graphs: Question 8.16. Is there a local characterization of quasi-crystal graphs? However, the Stembridge axioms are connected with the underlying representation theory, in the sense that they refer to whether the arrow labels correspond to orthogonal roots of the algebra. Since the quasi-crystal graphs are defined on a purely combinatorial level, any characterization of them must also be on a purely combinatorial level. 9. Applications 9.1. Counting factorizations. One interpretation of the Littlewood– Richardson rule [Ful97, ch. 5] is that the Littlewood–Richardson coefficients cνλµ give the number of different factorizations of an element of plac corresponding to a Young tableau of shape ν into elements corresponding to a tableau of shape λ and a tableau of shape µ. In particular, it shows that the number of such factorizations is independent of the content of the tableau. The quasi-crystal structure yields a similar result for hypo. Theorem 9.1. The number of distinct factorizations of an element of the hypoplactic monoid corresponding to a quasi-ribbon tableau of shape γ into elements that correspond to tableau of shape α and β is dependent only of γ, α, and β, and not on the content of the element. Proof. Let γ, α, β be compositions. Let w ∈ A∗n be a quasi-ribbon word such that γ is the shape of QR(w). Let  w = (u, v) : u, v ∈ A∗n are quasi-ribbon words, Sα,β QR(u) has shape α, QR(v) has shape β, w Sα,β w ≡hypo uv . So is a complete list of factorizations of w into elements whose corresponding tableaux have shapes α and β. Let i ∈ {1, . . . , n − 1} w and suppose ëi (w) is defined. Pick some pair (u, v) ∈ Sα,β . Then w ≡hypo uv and so ëi (uv) is defined, and ëi (w) = ëi (uv). By Lemma CRYSTALLIZING THE HYPOPLACTIC MONOID 51 5.4, either ëi (uv) = uëi (v), or ëi (uv) = ëi (u)v. In the former case, (u, ëi (v)) ∈ Sëi (w) , and in the latter case (ëi (u), v) ∈ Sëi (w) , since ëi preserves being a quasi-ribbon word and the shape of the corresponding quasi-ribbon tableau by Corollary 6.2. So ëi induces an injective map ëi (w) w from Sα,β to Sα,β . It follows that f¨i induces the inverse map from ë (w) ë (w) Si to S w . Hence |S w | = |S i |. Similarly, if f¨i (w) is defined, α,β α,β α,β w |Sα,β | f¨i (w) |Sα,β |. α,β then = Since all the quasi-ribbon words whose tableaux have shape γ lie in the same connected component of Γ(hypon ), it w follows that |Sα,β | is dependent only on γ, not on w.  9.2. Conjugacy. There are several possibile generalizations of conjugacy from groups to monoids. One possible definition, introduced by Otto [Ott84], is o-conjugacy, defined on a monoid M by (9.1) x ∼o y ⇐⇒ (∃g, h ∈ M )(xg = gy ∧ hx = yh). The relation ∼o is an equivalence relation. Another approach is primary conjugacy or p-conjugacy, defined on a monoid M by x ∼p y ⇐⇒ (∃u, v ∈ M )(x = uv ∧ y = vu). However, ∼p is reflexive and symmetric, but not transitive; hence it is sensible to follow Kudryavtseva & Mazorchuk [KM09, KM07] in working with its transitive closure ∼∗p . It is easy to show that ∼p ⊆ ∼∗p ⊆ ∼o . In some circumstances, equality holds. For instance, ∼∗p = ∼o in plac, and two elements of plac are related by ∼∗p and ∼o if and only if they have the same weight [CM13, Theorem 17]. Since hypo is a quotient of plac, the same result holds in hypo, but the machinery of quasi-crystals gives a very quick proof that two elements with the same weight are ∼o -related: Proposition 9.2. Let u, v ∈ hypon be such that wt(u) = wt(v). Then u ∼o v. Proof. Let g = n(n−1) · · · 21. Then ug is highest-weight by Proposition 6.14, because it has an i-inversion for all i ∈ {1, . . . , n − 1}, and thus no ëi is defined. Similarly gv is highest-weight. Furthermore, wt(ug) = wt(u)+wt(g) = wt(v)+wt(g) = wt(gv). Hence ug ≡hypo gv. Similarly, gu ≡hypo vg and so u ∼o v.  9.3. Satisfying an identity. Another application of the quasi-crystal structure is to prove that the hypoplactic monoid satisfies an identity. It is known that plac1 , plac2 , and plac3 satisfy identities, but whether placn satisfies an identity (perhaps dependent on n) is an important open problem [KO15]. Theorem 9.3. The hypoplactic monoid satisfies the identity xyxy = yxyx. 52 ALAN J. CAIN AND ANTÓNIO MALHEIRO Proof. Let x, y ∈ A∗n . The idea is to raise the products xyxy and yxyx to highest-weight using the same sequence of operators ëi , deduce that the corresponding highest-weight words are equal, and so conclude that xyxy ≡hypo yxyx. More formally, the proof proceeds by reverse induction on the weight of xyxy. The base case of the induction is when xyxy is highest-weight. Thus ëi (xyxy) is undefined for all i ∈ {1, . . . , n − 1}. So xyxy must have an i-inversion for all i ∈ {1, . . . , max(xyxy) − 1}. The symbols i + 1 and i may each lie in x or y, but in any case, there is an i-inversion in yxyx. Hence ëi (yxyx) is undefined for all i ∈ {1, . . . , n − 1}. So yxyx is also highest-weight. Clearly wt(xyxy) = wt(yxyx), and so QR(xyxy) = QR(yxyx) by Corollary 6.6. Hence xyxy ≡hypo yxyx. For the induction step, suppose xyxy is not highest-weight, and that 0 0 0 0 x y x y ≡hypo y 0 x0 y 0 x0 for all x0 , y 0 ∈ A∗n such that x0 y 0 x0 y 0 has higher weight than xyxy. Then ëi (xyxy) is defined for some i ∈ {1, . . . , n − 1}. Neither x nor y contains a symbol i, since otherwise there would be an i-inversion in xyxy. So ¨i (xyxy) = ¨i (yxyx) = |xyxy|i+1 = |yxyx|i+1 = 2|x|i+1 +2|y|i+1 = 2¨i (x)+2¨i (y), and ¨i (xyxy) applications ¨ (xyxy) of i change every symbol i + 1 in xyxy to i. That is, ëi i (xyxy) and ¨i (yxyx) 0 0 0 0 ëi (yxyx) are both defined and are equal to x y x y and y 0 x0 y 0 x0 ¨ (x) ¨ (y) respectively, where x0 = ëi i (x) and y 0 = ëi i (y). Since x0 y 0 x0 y 0 has higher weight than xyxy, it follows by the induction hypothesis that x0 y 0 x0 y 0 ≡hypo y 0 x0 y 0 x0 . Hence ¨ (xyxy) 0 0 0 0 ¨ (xyxy) 0 0 0 0 xyxy = f¨i i (x y x y ) ≡hypo f¨i i (y x y x ) = yxyx. This completes the induction step and thus the proof.  References [CGM] A. J. Cain, R. D. Gray, & A. Malheiro. ‘Crystal monoids & crystal bases: rewriting systems and biautomatic structures for plactic monoids of types An , Bn , Cn , Dn , and G2 ’. arXiv: 1412.7040. [CM13] C. Choffrut & R. Mercaş. ‘The lexicographic cross-section of the plactic monoid is regular’. In Combinatorics on words, vol. 8079 of Lecture Notes in Comput. Sci., pp. 83–94. Springer, Heidelberg, 2013. doi: 10.1007/978-3-642-40579-2 11. [FRT54] J. S. Frame, G. d. B. Robinson, & R. M. Thrall. ‘The hook graphs of the symmetric groups’. Canadian J. Math., 6 (1954), pp. 316–324. doi: 10.4153/CJM-1954-030-1. [Ful97] W. Fulton. Young Tableaux, vol. 35 of London Mathematical Society Student Texts. Cambridge University Press, Cambridge, 1997. With applications to representation theory and geometry. [GKL+ 95] I. M. Gelfand, D. Krob, A. Lascoux, B. Leclerc, V. S. Retakh, & J. Y. Thibon. ‘Noncommutative symmetric functions’. Adv. Math., 112, no. 2 (1995), pp. 218–348. doi: 10.1006/aima.1995.1032. [Gre07] J. A. Green. Polynomial representations of GLn , vol. 830 of Lecture Notes in Mathematics. Springer, Berlin, augmented edition, 2007. With CRYSTALLIZING THE HYPOPLACTIC MONOID [Hag15] [Hig92] [Hiv00] [HK02] [How95] [Jed11] [Kas90] [Kas91] [KM07] [KM09] [KN94] [Knu70] [KO15] [KS04] [KT97] [KT99] 53 an appendix on Schensted correspondence and Littelmann paths by K. Erdmann, Green and M. Schocker. N. Hage. ‘Finite convergent presentation of plactic monoid for type C’. Internat. J. Algebra Comput., 25, no. 8 (2015), pp. 1239–1263. doi: 10.1142/S0218196715500393. P. M. Higgins. Techniques of Semigroup Theory. Oxford Science Publications. Clarendon Press, Oxford University Press, New York, 1992. F. Hivert. ‘Hecke algebras, difference operators, and quasi-symmetric functions’. Adv. Math., 155, no. 2 (2000), pp. 181–238. doi: 10.1006/aima.1999.1901. J. Hong & S. J. Kang. Introduction to quantum groups and crystal bases, vol. 42 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2002. doi: 10.1090/gsm/042. J. M. Howie. Fundamentals of Semigroup Theory, vol. 12 of London Mathematical Society Monographs (New Series). Clarendon Press, Oxford University Press, New York, 1995. F. Jedrzejewski. ‘Plactic classification of modes’. In Mathematics and computation in music, vol. 6726 of Lecture Notes in Comput. Sci., pp. 350–353. Springer, Heidelberg, 2011. doi: 10.1007/978-3-642-215902 31. M. Kashiwara. ‘Crystalizing the q-analogue of universal enveloping algebras’. Comm. Math. Phys., 133, no. 2 (1990), pp. 249–260. url: http://projecteuclid.org/euclid.cmp/1104201397. M. Kashiwara. ‘On crystal bases of the q-analogue of universal enveloping algebras’. Duke Math. J., 63, no. 2 (1991), pp. 465–516. doi: 10.1215/S0012-7094-91-06321-0. G. Kudryavtseva & V. Mazorchuk. ‘On conjugation in some transformation and Brauer-type semigroups’. Publ. Math. Debrecen, 70, no. 1-2 (2007), pp. 19–43. G. Kudryavtseva & V. Mazorchuk. ‘On three approaches to conjugacy in semigroups’. Semigroup Forum, 78, no. 1 (2009), pp. 14–20. doi: 10.1007/s00233-008-9047-7. M. Kashiwara & T. Nakashima. ‘Crystal graphs for representations of the q-analogue of classical Lie algebras’. J. Algebra, 165, no. 2 (1994), pp. 295–345. doi: 10.1006/jabr.1994.1114. D. E. Knuth. ‘Permutations, matrices, and generalized Young tableaux’. Pacific J. Math., 34 (1970), pp. 709–727. url: projecteuclid.org/ euclid.pjm/1102971948. L. Kubat & J. Okniński. ‘Identities of the plactic monoid’. Semigroup Forum, 90, no. 1 (2015), pp. 100–112. doi: 10.1007/s00233-014-9609-9. J. A. Kim & D. U. Shin. ‘Insertion scheme for the classical lie algebras’. Comm. Algebra, 32, no. 8 (2004), pp. 3139–3167. doi: 10.1081/AGB120039283. D. Krob & J. Y. Thibon. ‘Noncommutative symmetric functions IV: Quantum linear groups and Hecke algebras at q = 0’. J. Algebraic Combin., 6, no. 4 (1997), pp. 339–376. doi: 10.1023/A:1008673127310. D. Krob & J. Y. Thibon. ‘Noncommutative symmetric functions. V. A degenerate version of Uq (glN )’. Internat. J. Algebra Comput., 9, no. 3-4 (1999), pp. 405–430. Dedicated to the memory of Marcel-Paul Schützenberger. doi: 10.1142/S0218196799000254. 54 [Lec02] [Lec03] [Lec07] [Lot02] [LR34] [LS78] [LS81] [LS85] [LS90] [Mac95] [Nov00] [Ott84] [Ruš95] [Sch61] [Sch97] [Shi05] [Sta05] ALAN J. CAIN AND ANTÓNIO MALHEIRO C. Lecouvey. ‘Schensted-type correspondence, plactic monoid, and jeu de taquin for type Cn ’. J. Algebra, 247, no. 2 (2002), pp. 295–331. doi: 10.1006/jabr.2001.8905. C. Lecouvey. ‘Schensted-type correspondences and plactic monoids for types Bn and Dn ’. J. Algebraic Combin., 18, no. 2 (2003), pp. 99–133. doi: 10.1023/A:1025154930381. C. Lecouvey. ‘Combinatorics of crystal graphs for the root systems of types An , Bn , Cn , Dn and G2 ’. In Combinatorial aspect of integrable systems, vol. 17 of Mathematical Society of Japan Memoirs, pp. 11–41. Math. Soc. Japan, Tokyo, 2007. doi: doi:10.2969/msjmemoirs/01701C020. M. Lothaire. Algebraic Combinatorics on Words, vol. 90 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, 2002. D. Littlewood & A. Richardson. ‘Group characters and algebra’. Philos. Trans. Roy. Soc. London Ser. A, 233 (1934), pp. 99–141. url: http: //www.jstor.org/stable/91293. A. Lascoux & M. P. Schützenberger. ‘Sur une conjecture de H. O. Foulkes’. C. R. Acad. Sci. Paris Sér. A-B, 286, no. 7 (1978), pp. A323– A324. A. Lascoux & M. P. Schützenberger. ‘Le monoı̈de plaxique’. In Noncommutative structures in algebra and geometric combinatorics (Naples, 1978), vol. 109 of Quad. “Ricerca Sci.”, pp. 129–156. CNR, Rome, 1981. A. Lascoux & M. P. Schützenberger. ‘Schubert polynomials and the Littlewood-Richardson rule’. Lett. Math. Phys., 10, no. 2-3 (1985), pp. 111–124. doi: 10.1007/BF00398147. A. Lascoux & M. P. Schützenberger. ‘Tableaux and noncommutative Schubert polynomials’. Funct. Anal. Appl., , no. 23 (1990), pp. 223–225. doi: 10.1007/BF01079531. I. Macdonald. Symmetric functions and Hall polynomials. Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, second edition, 1995. With contributions by A. Zelevinsky, Oxford Science Publications. J. C. Novelli. ‘On the hypoplactic monoid’. Discrete Math., 217, no. 1-3 (2000), pp. 315–336. Formal power series and algebraic combinatorics (Vienna, 1997). doi: 10.1016/S0012-365X(99)00270-8. F. Otto. ‘Conjugacy in monoids with a special Church-Rosser presentation is decidable’. Semigroup Forum, 29, no. 1-2 (1984), pp. 223–240. doi: 10.1007/BF02573327. N. Ruškuc. Semigroup Presentations. Ph.D. Thesis, University of St Andrews, 1995. url: hdl.handle.net/10023/2821. C. Schensted. ‘Longest increasing and decreasing subsequences’. Canad. J. Math., 13 (1961), pp. 179–191. doi: 10.4153/CJM-1961-015-3. M. P. Schützenberger. ‘Pour le monoı̈de plaxique’. Math. Inform. Sci. Humaines, , no. 140 (1997), pp. 5–10. url: www.ehess.fr/revue-msh/ pdf/N140R764.pdf. M. Shimozono. ‘Crystals for dummies’. Notes, 2005. url: www.aimath. org/WWN/kostka/crysdumb.pdf. R. P. Stanley. Enumerative Combinatorics (Vol.2). Cambridge University Press, 2005. CRYSTALLIZING THE HYPOPLACTIC MONOID [Sta12] [Ste03] 55 R. P. Stanley. Enumerative Combinatorics (Vol. 1), vol. 49 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, second edition, 2012. J. R. Stembridge. ‘A local characterization of simply-laced crystals’. Trans. Amer. Math. Soc., 355, no. 12 (2003), pp. 4807–4823 (electronic). doi: 10.1090/S0002-9947-03-03042-3. Centro de Matemática e Aplicações, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal E-mail address: [email protected] Centro de Matemática e Aplicações, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal Departamento de Matemática, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal E-mail address: [email protected]
4math.GR
BCOL RESEARCH REPORT 15.03 arXiv:1705.05920v1 [math.OC] 16 May 2017 Industrial Engineering & Operations Research University of California, Berkeley, CA 94720–1777 PATH COVER AND PATH PACK INEQUALITIES FOR THE CAPACITATED FIXED-CHARGE NETWORK FLOW PROBLEM ALPER ATAMTÜRK, BIRCE TEZEL AND SIMGE KÜÇÜKYAVUZ Abstract. Capacitated fixed-charge network flows are used to model a variety of problems in telecommunication, facility location, production planning and supply chain management. In this paper, we investigate capacitated path substructures and derive strong and easy-to-compute path cover and path pack inequalities. These inequalities are based on an explicit characterization of the submodular inequalities through a fast computation of parametric minimum cuts on a path, and they generalize the well-known flow cover and flow pack inequalities for the single-node relaxations of fixed-charge flow models. We provide necessary and sufficient facet conditions. Computational results demonstrate the effectiveness of the inequalities when used as cuts in a branch-and-cut algorithm. July 2015; October 2016; May 2017 1. Introduction Given a directed multigraph with demand or supply on the nodes, and capacity, fixed and variable cost of flow on the arcs, the capacitated fixed-charge network flow (CFNF) problem is to choose a subset of the arcs and route the flow on the chosen arcs while satisfying the supply, demand and capacity constraints, so that the sum of fixed and variable costs is minimized. There are numerous polyhedral studies of the fixed-charge network flow problem. In a seminal paper Wolsey (1989) introduces the so-called submodular inequalities, which subsume almost all valid inequalities known for capacitated fixed-charge networks. Although the submodular inequalities are very general, their coefficients are defined implicitly through value functions. In this paper, we give explicit valid inequalities that simultaneously make use of the path substructures of the network as well as the arc capacities. For the uncapacitated fixed-charge network flow problem, van Roy and Wolsey (1985) give flow path inequalities that are based on path substructures. Rardin and Wolsey (1993) A. Atamtürk: Department of Industrial Engineering & Operations Research, University of California, Berkeley, CA 94720. [email protected] B. Tezel: Department of Industrial Engineering & Operations Research, University of California, Berkeley, CA 94720. [email protected] S. Küçükyavuz: Department of Industrial and Systems Engineering, University of Washington, Seattle, WA 98195. [email protected] . 1 introduce a new family of dicut inequalities and show that they describe the projection of an extended multicommodity formulation onto the original variables of fixed-charge network flow problem. Ortega and Wolsey (2003) present a computational study on the performance of path and cut-set (dicut) inequalities. For the capacitated fixed-charge network flow problem, almost all known valid inequalities are based on single-node relaxations. Padberg et al. (1985), van Roy and Wolsey (1986) and Gu et al. (1999) give flow cover, generalized flow cover and lifted flow cover inequalities. Stallaert (1997) introduces the complement class of generalized flow cover inequalities and Atamtürk (2001) describes lifted flow pack inequalities. Both uncapacitated path inequalities and capacitated flow cover inequalities are highly valuable in solving a host of practical problems and are part of the suite of cutting planes implemented in modern mixed-integer programming solvers. The path structure arises naturally in network models of the lot-sizing problem. Atamtürk and Muñoz (2004) introduce valid inequalities for the capacitated lot-sizing problems with infinite inventory capacities. Atamtürk and Küçükyavuz (2005) give valid inequalities for the lot-sizing problems with finite inventory and infinite production capacities. Van Vyve (2013) introduces path-modular inequalities for the uncapacitated fixed charge transportation problems. These inequalities are derived from a value function that is neither globally submodular nor supermodular but that exhibits sub or supermodularity under certain set selections. Van Vyve and Ortega (2004) and Gade and Küçükyavuz (2011) give valid inequalities and extended formulations for uncapacitated lot-sizing with fixed charges on stocks. For uncapacitated lot-sizing with backlogging, Pochet and Wolsey (1988) and Pochet and Wolsey (1994) provide valid inequalities and Küçükyavuz and Pochet (2009) provide an explicit description of the convex hull. Contributions. In this paper we consider a generic path relaxation, with supply and/or demand nodes and capacities on incoming and outgoing arcs. By exploiting the path substructure of the network and introducing notions of path cover and path pack we provide two explicitly-described subclasses of the submodular inequalities of Wolsey (1989). The most important consequence of the explicit derivation is that the coefficients of the submodular inequalities on a path can be computed efficiently. In particular, we show that the coefficients of such an inequality can be computed by solving max-flow/min-cut problems parametrically over the path. Moreover, we show that all of these coefficients can be computed with a single linear-time algorithm. For a path with a single node, the inequalities reduce to the well-known flow cover and flow pack inequalities. Moreover, we show that the path cover and path pack inequalities dominate flow cover and flow pack inequalities for the corresponding single node relaxation of a path obtained by merging the path into a single node. We give necessary and sufficient facet-defining conditions. Finally, we report on computational experiments demonstrating the effectiveness of the proposed inequalities when used as cuts in a branch-and-cut algorithm. Outline. The remainder of this paper is organized as follows: In Section 2, we describe the capacitated fixed-charge flow problem on a path, its formulation and the assumptions we make. In Section 3, we review the submodular inequalities, discuss their computation on a path, and introduce two explicit subclasses: path cover inequalities and path pack 2 inequalities. In Section 4, we analyze sufficient and necessary facet-defining conditions. In Section 5, we present computational experiments showing the effectiveness of the path cover and path pack inequalities compared to other network inequalities. 2. Capacitated fixed-charge network flow on a path Let G = (N 0 , A) be a directed multigraph with nodes N 0 and arcs A. Let sN and tN be the source and the sink nodes of G. Let N := N 0 \ {sN , tN }. Without loss of generality, we label N := {1, . . . , n} such that a directed forward path arc exists from node i to node i + 1 and a directed backward path arc exists from node i + 1 to node i for each node i = 1, . . . , n − 1 (see Figure 1 for an illustration). In Remarks 1 and 2, we discuss how to obtain a “path” graph G from a more general directed multigraph. Let E + = {(i, j) ∈ A : i = sN , j ∈ N } and E − = {(i, j) ∈ A : i ∈ N, j = tN }. Moreover, let us partition the sets E + and E − such that Ek+ = {(i, j) ∈ A : i ∈ / N, j = k} − + and Ek = {(i, j) ∈ A : i = k, j ∈ / N } for k ∈ N . We refer to the arcs in E and E − as non-path arcs. Finally, let E := E + ∪ E − be the set of all non-path arcs. For convenience, we generalize this set notation scheme. Given an arbitrary subset of non-path arcs Y ⊆ E, let Yj+ = Y ∩ Ej+ and Yj− = Y ∩ Ej− . Remark 1. Given a directed multigraph G̃ = (Ñ , Ã) with nodes Ñ , arcs à and a path that passes through nodes N , we can construct G as described above by letting E + = {(i, j) ∈ à : i ∈ Ñ \ N, j ∈ N } and E − = {(i, j) ∈ à : i ∈ N, j ∈ Ñ \ N } and letting all the arcs in E + be the outgoing arcs from a dummy source sN and all the arcs in E − to be incoming to a dummy sink tN . Remark 2. If there is an arc t = (i, j) from node i ∈ N to j ∈ N , where |i − j| > 1, then we construct a relaxation by removing arc t, and replacing it with two arcs t− ∈ Ei− and t+ ∈ Ej+ . If there are multiple arcs from node i to node j, one can repeat the same procedure. Throughout the paper, we use the following notation: Let [k, j] = {k, k +1, . . . , j} if k ≤ j P P Pj and ∅ otherwise, c(S) = t∈S ct , y(S) = t∈S yt , (a)+ = max{0, a} and dkj = t=k dt if j ≥ k and 0 otherwise. Moreover, let dim(A) denote the dimension of a polyhedron A and conv(S) be the convex hull of a set S. The capacitated fixed-charge network flow problem on a path can be formulated as a mixed-integer optimization problem. Let dj be the demand at node j ∈ N . We call a node j ∈ N a demand node if dj ≥ 0 and a supply node if dj < 0. Let the flow on forward path arc (j, j + 1) be represented by ij with an upper bound uj for j ∈ N \ {n}. Similarly, let the flow on backward path arc (j + 1, j) be represented by rj with an upper bound bj for j ∈ N \ {n}. Let yt be the amount of flow on arc t ∈ E with an upper bound ct . Define binary variable xt to be 1 if yt > 0, and zero otherwise for all t ∈ E. An arc t is closed if xt = 0 and open if xt = 1. Moreover, let ft be the fixed cost and pt be the unit flow cost of arc t. Similarly, let hj and gj be the costs of unit flow, on forward and backward arcs (j, j + 1) and (j + 1, j) respectively for j ∈ N \ {n}. Then, the problem is formulated as X X min (ft xt + pt yt ) + (hj ij + gj rj ) (1a) t∈E j∈N 3 sN cti b1 1 d1 b2 ... 3 2 u1 d2 bn−1 b3 u2 d3 n un−1 u3 dn ctk tN Figure 1. Fixed-charge network representation of a path. s. t. ij−1 − rj−1 + y(Ej+ ) − y(Ej− ) − ij + rj = dj , 0 ≤ yt ≤ ct xt , (F1) t ∈ E, j ∈ N, (1b) (1c) 0 ≤ ij ≤ uj , j ∈ N, (1d) 0 ≤ rj ≤ bj , j ∈ N, (1e) xt ∈ {0, 1}, t ∈ E, (1f) i0 = in = r0 = rn = 0. (1g) Let P be the set of feasible solutions of (F1). Figure 1 shows an example network representation of (F1). Throughout we make the following assumptions on (F1): The set Pt = {(x, y, i, r) ∈ P : xt = 0} = 6 ∅ for all t ∈ E, ct > 0, uj > 0 and bj > 0 for all t ∈ E and j ∈ N , ct ≤ d1n + c(E − ) for all t ∈ E + , ct ≤ bj−1 + uj + (dj )+ + c(Ej− ), for all j ∈ N, t ∈ Ej+ , ct ≤ bj + uj−1 + (−dj )+ + c(Ej+ ) for all j ∈ N, t ∈ Ej− .  Assumptions (A.1)–(A.2) ensure that dim conv(P) = 2|E| + |N | − 2. If (A.1) does not hold for some t ∈ E, then xt = 1 for all points in P. Similarly, if (A.2) does not hold, the flow on such an arc can be fixed to zero. Finally, assumptions (A.3)–(A.5) are without loss of generality. An upper bound on yt can be obtained directly from the flow balance equalities (1b) by using the upper and lower bounds of the other flow variables that appear in the same constraint. As a result, the flow values on arcs t ∈ E cannot exceed the capacities implied by (A.3)–(A.5). Next, we review the submodular inequalities introduced by Wolsey (1989) that are valid for any capacitated fixed-charge network flow problem. Furthermore, using the path structure, we provide an O(|E| + |N |) time algorithm to compute their coefficients explicitly. (A.1) (A.2) (A.3) (A.4) (A.5) 4 3. Submodular inequalities on paths Let S + ⊆ E + and L− ⊆ E − . Wolsey (1989) shows that the value function of the following optimization problem is submodular: X v(S + , L− ) = max at yt (2a) t∈E s. t. (F2) ij−1 − rj−1 + y(Ej+ ) − y(Ej− ) − ij + rj ≤ dj , j ∈ N, (2b) 0 ≤ ij ≤ uj , j ∈ N, (2c) 0 ≤ rj ≤ bj , j ∈ N, (2d) 0 ≤ yt ≤ ct , t ∈ E, (2e) i0 = in = r0 = rn = 0, yt = 0, + + (2f) − t ∈ (E \ S ) ∪ L , (2g) where at ∈ {0, 1} for t ∈ E + and at ∈ {0, −1} for t ∈ E − . The set of feasible solutions of (F2) is represented by Q. We call the sets S + and L− that are used in the definition of v(S + , L− ) the objective sets. For ease of notation, we also represent the objective sets as C := S + ∪ L− . Following this notation, let v(C) := v(S + , L− ), v(C \ {t}) = v(S + \ {t}, L− ) for t ∈ S + and v(C \ {t}) = v(S + , L− \ {t}) for t ∈ L− . Similarly, let v(C ∪ {t}) = v(S + ∪ {t}, L− ), for t ∈ S + and v(C ∪ {t}) = v(S + , L− ∪ {t}) for t ∈ L− . Moreover, let ρt (C) = v(C ∪ {t}) − v(C) be the marginal contribution of adding an arc t to C with respect to the value function v. Wolsey (1989) shows that the following inequalities are valid for P: X X X a t yt + ρt (C \ {t})(1 − x̄t ) ≤ v(C) + ρt (∅)x̄t , (3) t∈E X t∈E at yt + t∈C X t∈E\C ρt (E \ {t})(1 − x̄t ) ≤ v(C) + t∈C X ρt (C)x̄t , (4) t∈E\C where the variable x̄t is defined as ( x̄t = xt , t ∈ E+ 1 − xt , t ∈ E−. In fact, inequalities (3) and (4) are also valid for fixed-charge network flow formulations where the flow balance constraints (1b) are replaced with constraints (2b). However, in this paper, we focus on formulations with flow balance equalities (1b). We refer to submodular inequalities (3) and (4) derived for path structures as path inequalities. In this paper, we consider sets S + and L− such that (F2) is feasible for all objective sets C and C \ {t} for all t ∈ C. 5 3.1. Equivalence to the maximum flow problem. Define sets K + and K − such that the coefficients of the objective function (2a) are:  +   1, t ∈ K at = −1, t ∈ K −   0, otherwise, (5) where S + ⊆ K + ⊆ E + and K − ⊆ E − \ L− . We refer to the sets K + and K − as coefficient sets. Let the set of arcs with zero coefficients in (2a) be represented by K̄ + = E + \ K + and K̄ − = E − \ K − . Given a selection of coefficients as described in (5), we claim that (F2) can be transformed to a maximum flow problem. We first show this result assuming dj ≥ 0 for all j ∈ N . Then, in Appendix A, we show that the nonnegativity of demand is without loss of generality for the derivation of the inequalities. Proposition 1. Let S + ⊆ E + and L− ⊆ E − be the objective sets in (F2) and let Y be the nonempty set of optimal solutions of (F2). If dj ≥ 0 for all j ∈ N , then there exists at least one optimal solution (y∗ , r∗ , i∗ ) ∈ Y such that yt∗ = 0 for t ∈ K̄ + ∪ K − ∪ L− . Proof. Observe that yt∗ = 0 for all t ∈ E + \S + , due to constraints (2g). Since K̄ + ⊆ E + \S + , yt∗ = 0, for t ∈ K̄ + from feasibility of (F2). Similarly, yt∗ = 0 for all t ∈ L− by constraints (2g). Now suppose that, yt∗ =  > 0 for some t ∈ Kj− (i.e., at = −1 for arc t in (F2)). Let the slack value at constraint (2b) for node j be   ∗ sj = dj − i∗j−1 − rj−1 + y ∗ (Ej+ ) − y ∗ (Ej− \ {t}) − yt∗ − i∗j + rj∗ . If sj ≥ , then decreasing yt∗ by  both improves the objective function value and conserves the feasibility of flow balance inequality (2b) for node j, since sj −  ≥ 0. If sj < , then decreasing yt∗ by  violates flow balance inequality since sj −  < 0. In this case, there must exist a simple directed path P from either the source node sN or a node k ∈ N \ {j} to node j where all arcs have at least a flow of ( − sj ). This is guaranteed because, sj <  implies that, without the outgoing arc t, there is more incoming flow to node j than outgoing. Then, notice that decreasing the flow on arc t and all arcs in path P by  − sj conserves feasibility. Moreover, the objective function value either remains the same or increases, because decreasing yt by  − sj increases the objective function value by  − sj and the decreasing the flow on arcs in P decreases it by at most  − sj . At the end of this transformation, the slack value sj does not change, however; the flow at arc t is now yt∗ = sj which is equivalent to the first case that is discussed above. As a result, we obtain a new solution to (F2) where yt∗ = 0 and the objective value is at least as large.  Proposition 2. If dj ≥ 0 for all j ∈ N , then (F2) is equivalent to a maximum flow problem from source sN to sink tN on graph G. Proof. At the optimal solution of problem (F2) with objective set (S + , L− ), the decision variables yt , for t ∈ (E + \ S + ) ∪ K − ∪ L− can be assumed to be zero due to Proposition 1 and constraints (2g). Then, these variables can be dropped from (F2) since the value v(S + , L− ) does not depend them and formulation (F2) reduces to  v(S + , L− ) = max y(S + ) : ij−1 − rj−1 + y(Sj+ ) − y(K̄j− ) − ij + rj ≤ dj , j ∈ N, 6 (2c) − (2f) . (6) Now, we reformulate (6) by representing the left hand side of the flow balance constraint by a new nonnegative decision variable zj that has an upper bound of dj for each j ∈ N :  max y(S + ) : ij−1 − rj−1 + y(Sj+ ) − y(K̄j− ) − ij + rj = zj , j ∈ N, 0 ≤ zj ≤ dj , j ∈ N, (2c) − (2f) . The formulation above is equivalent the maximum flow formulation from the source node sN to the sink node tN for the path structures we are considering in this paper.  Under the assumption that dj ≥ 0 for all j ∈ N , Proposition 1 and Proposition 2 together show that the optimal objective function value v(S + , L− ) can be computed by solving a maximum flow problem from source sN to sink tN . We generalize this result in Appendix A for node sets N such that dj < 0 for some j ∈ N . As a result, obtaining the explicit coefficients of submodular inequalities (3) and (4) reduces to solving |E| + 1 maximum flow problems. For a general underlying graph, solving |E| + 1 maximum flow problems would take O(|E|2 |N |) time (e.g., see King et al. (1994)), where |E| and |N | are the number of arcs and nodes, respectively. In the following subsection, by utilizing the equivalence of maximum flow and minimum cuts and the path structure, we show that all coefficients of (3) and (4) can be obtained in O(|E| + |N |) time using dynamic programming. 3.2. Computing the coefficients of the submodular inequalities. Throughout the paper, we use minimum cut arguments to find the explicit coefficients of inequalities (3) and (4). Figure 2a illustrates an example where N = [1, 5], E + = [1, 5], E − = [6, 10] and in Figure 2b, we give an example of an sN − tN cut for S + = {2, 4, 5}, L− = {10} and K̄ − = {7, 9, 10}. The dashed line in Figure 2b represents a cut that corresponds to the partition {sN , 2, 5} and {tN , 1, 3, 4} with a value of b1 + d2 + c7 + u2 + c4 + b4 + d5 . Moreover, we say that a cut passes below node j if j is in the source partition and passes above node j if j is in the sink partition. Let αju and αjd be the minimum value of a cut on nodes [1, j] that passes above and below node j, respectively. Similarly, let βju and βjd be the minimum values of cuts on nodes [j, n] that passes above and below node j respectively. Finally, let S − = E − \ (K − ∪ L− ), where K − is defined in (5). Recall that S + and L− are the given objective sets. Given the notation introduced above, all of the arcs in sets S − and L− have a coefficient zero in (F2). Therefore, dropping an arc from L− is equivalent to adding that arc to S − . We compute {u,d} {u,d} αj by a forward recursion and βj by a backward recursion: d u αju = min{αj−1 + uj−1 , αj−1 } + c(Sj+ ) αjd = d u min{αj−1 , αj−1 + bj−1 } + dj + c(Sj− ), (7) (8) where α0u = α0d = 0 and u d βju = min{βj+1 , βj+1 + bj } + c(Sj+ ) u d βjd = min{βj+1 + uj , βj+1 } + dj + c(Sj− ), 7 (9) (10) c2 c1 u1 1 u2 2 d1 u3 3 c7 d2 c5 u4 4 b3 b2 b1 c6 c4 c3 c8 d3 5 b4 c9 d4 c10 d5 (a) A path graph with E + = [1, 5] and E − = [6, 10]. c2 1 u1 c4 2 3 c7 d2 u3 4 b3 b2 b1 d1 u2 c5 d3 u4 5 b4 c9 d4 d5 (b) An sN − tN cut for set S + = {2, 4, 5}, L− = {10} and K̄ − = {7, 9, 10}. Figure 2. An example of an sN − tN cut. u d where βn+1 = βn+1 = 0. u Let mj and mdj be the values of minimum cuts for nodes [1, n] that pass above and below node j, respectively. Notice that muj = αju + βju − c(Sj+ ) (11) mdj = αjd + βjd − dj − c(Sj− ). (12) and For convenience, let mj := min{muj , mdj }. Notice that mj is the minimum of the minimum cut values that passes above and below node j. Since the minimum cut corresponding to v(C) has to pass either above or below node j, mj is equal to v(C) for all j ∈ N . As a result, the minimum cut (or maximum flow) value for the objective set C = S + ∪ L− is v(C) = m1 = · · · = mn . (13) Proposition 3. All values mj , for j ∈ N , can be computed in O(|E| + |N |) time. Obtaining the explicit coefficients of inequalities (3) and (4) also requires finding v(C \{t}) for t ∈ C and v(C ∪ {t}) for t ∈ / C in addition to v(C). It is important to note that we do not need to solve the recursions above repeatedly. Once the values muj and mdj are obtained for the set C, the marginals ρt (C \ {t}) and ρt (C) can be found in O(1) time for each t ∈ E. We use the following observation while providing the marginal values ρt (C) and ρt (C \{t}) in closed form. Observation 1. Let c ≥ 0 and d := (b − a)+ , then, 1. min{a + c, b} − min{a, b} = min{c, d}, 2. min{a, b} − min{a, b − c} = (c − d)+ . 8 In the remainder of this section, we give a linear-time algorithm to compute the coefficients ρt for inequalities (3) and (4) explicitly for paths. Coefficients of inequality (3): Path cover inequalities. Let S + and L− be the objective sets in (F2) and S − ⊆ E − \ L− . We select the coefficient sets in (5) as K + = S + and K − = E − \ (L− ∪ S − ) to obtain the explicit form of inequality (3). As a result, the set definition of S − = E − \ (K − ∪ L− ) is conserved. Definition 1. Let the coefficient sets in (5) be selected as above and (S + , L− ) be the objective set. The set (S + , S − ) is called a path cover for the node set N if v(S + , L− ) = d1n + c(S − ). For inequality (3), we assume that the set (S + , S − ) is a path cover for N . Then, by definition, v(C) = m1 = · · · = mn = d1n + c(S − ). After obtaining the values muj and mdj for a node j ∈ N using recursions in (7)-(10), it is trivial to find the minimum cut value after dropping an arc t from Sj+ : t ∈ Sj+ , v(C \ {t}) = min{muj − ct , mdj }, j ∈ N. Similarly, dropping an arc t ∈ L− j results in the minimum cut value: t ∈ L− j , v(C \ {t}) = min{muj , mdj + ct }, j ∈ N. Using Observation 1, we obtain the marginal values ρt (C \ {t}) = (ct − λj )+ , t ∈ S+ ρt (C \ {t}) = min{λj , ct }, t ∈ L− and where λj = (muj − mdj )+ , j ∈ N. On the other hand, all the coefficients ρt (∅) = 0 for arcs t ∈ E \ C. First, notice that, for t ∈ E + \ S + , v({t}) = 0, because the coefficient at = 0 for t ∈ E + \ S + . Furthermore, v({t}) = 0 for t ∈ E − \ L− , since all incoming arcs would be closed for an objective set (∅, {t}). As a result, inequality (3) for the objective set (S + , L− ) can be written as X X y(S + ) + (ct − λj )+ (1 − xt ) ≤ d1n + c(S − ) + min{ct , λj }xt + y(E − \ (L− ∪ S − )). t∈L− t∈S + (14) We refer to inequalities (14) as path cover inequalities. Remark 3. Observe that for a path consisting of a single node N = {j} with demand d := dj > 0, the path cover inequalities (14) reduce to the flow cover inequalities (Padberg et al., 1985, van Roy and Wolsey, 1986). Suppose that the path consists of a single node N = {j} with demand d := dj > 0. Let S + ⊆ E + and S − ⊆ E − . The set (S + , S − ) is a flow cover if λ := c(S + ) − d − c(S − ) > 0 and the resulting path cover inequality X y(S + ) + (ct − λ)+ (1 − xt ) ≤ d + c(S − ) + λx(L− ) + y(E − \ L− ) (15) t∈S + 9 is a flow cover inequality. Proposition 4. Let (S + , S − ) be a path cover for the node set N . The path cover inequality for node set N is at least as strong as the flow cover inequality for the single node relaxation obtained by merging the nodes in N . Proof. Flow cover and path cover inequalities differ in the coefficients of variables xt for t ∈ S + and t ∈ L− . Therefore, we compare the values λj , j ∈ N of path cover inequalities (14) to the value λ of flow cover inequalities (15) and show that λj ≤ λ, for all j ∈ N . The merging of node set N in graph G is equivalent to relaxing the values uj and bj to be infinite for j ∈ [1, n − 1]. As a result, the value of the minimum cut that goes above a the merged node is m̄u = c(S + ) and the value of the minimum cut that goes below the merged node is m̄d = d1n + c(S − ). Now, observe that the recursions in (7)-(10) imply that the minimum cut values for the original graph G are smaller: muj = αju + βju − c(Sj+ ) ≤ c(S + ) = m̄u and mdj = αjd + βjd − dj − c(Sj− ) ≤ d1n + c(S − ) = m̄d for all j ∈ N . Recall that the coefficient for the flow cover inequality is λ = (m̄u − m̄d )+ and the coefficients for path cover inequality are λj = (muj − mdj )+ for j ∈ N . The fact that (S + , S − ) is a path cover implies that mdj = d1n + c(S − ) for all j ∈ N . Since m̄d = mdj and muj ≤ m̄u for all j ∈ N , we observe that λj ≤ λ for all j ∈ N . Consequently, the path cover inequality (14) is at least as strong as the flow cover inequality (15).  c1 = 15 c2 = 35 u1 = 10 1 u2 = 10 2 b1 = 15 d1 = 10 c3 = 30 c4 = 10 u3 = 20 3 b2 = 15 d2 = 10 4 b3 = 10 d3 = 5 d4 = 15 Figure 3. A lot-sizing instance with backlogging. Example 1. Consider the lot-sizing instance in Figure 3 where N = [1, 4], S + = {2, 3}, L− = ∅. Observe that mu1 = 45, md1 = 40, mu2 = 65, md2 = 40, mu3 = 60, md3 = 40, and mu4 = 45, md4 = 40. Then, λ1 = 5, λ2 = 25, λ3 = 20, and λ4 = 5 leading to coefficients 10 and 10 for (1 − x2 ) and (1 − x3 ), respectively. Furthermore, the maximum flow values are v(C) = 40, v(C \ {2}) = 30, and v(C \ {3}) = 30. Then, the resulting path cover inequality (14) is y2 + y3 + 10(1 − x2 ) + 10(1 − x3 ) ≤ 40, (16) and it is facet-defining for conv(P) as will be shown in Section 4. Now, consider the relaxation obtained by merging the nodes in [1, 4] into a single node with incoming arcs {1, 2, 3, 4} and demand d = 40. As a result, the flow cover inequalities can be applied to the merged 10 node set. The excess value for the set S + = {2, 3} is λ = c(S + ) − d = 25. Then, the resulting flow cover inequality (15) is y2 + y3 + 10(1 − x2 ) + 5(1 − x3 ) ≤ 40, and it is weaker than the path cover inequality (16).  Coefficients of inequality (4): Path pack inequalities. Let S + and L− be the objective sets in (F2) and let S − ⊆ E − \ L− . We select the coefficient sets in (5) as K + = E + and K − = E − \ (S − ∪ L− ) to obtain the explicit form of inequality (4). As a result, the set definition of S − = E − \ (K − ∪ L− ) is conserved. Definition 2. Let the coefficients in (5) be selected as above and (S + , L− ) be the objective set. The set (S + , S − ) is called a path pack for node set N if v(S + , L− ) = c(S + ). For inequality (4), we assume that the set (S + , S − ) is a path pack for N and L− = ∅ for simplicity. Now, we need to compute the values of v(C), v(E), v(E \ {t}) for t ∈ C and v(C ∪ {t}) for t ∈ E \ C. The value of v(C ∪ {t}) can be obtained using the values muj and mdj that are given by recursions (7)-(10). Then, t ∈ Ej+ \ Sj+ , v(C ∪ {t}) = min{muj + ct , mdj }, j∈N and t ∈ Sj− , v(C ∪ {t}) = min{muj , mdj + ct }, j ∈ N. Then, using Observation 1, we compute the marginal values ρt (C) = min{ct , µj }, t ∈ Ej+ \ Sj+ and ρt (C) = (ct − µj )+ , t ∈ Sj− , µj = (mdj − muj )+ , j ∈ N. where Next, we compute the values v(E) and v(E \{t}) for t ∈ C. The feasibility of (F1) implies that (E + , ∅) is a path cover for N . By Assumption (A.1), (E + \ {t}, ∅) is also a path cover for N for each t ∈ S + . Then v(E) = v(E \ {t}) = d1n and ρt (E \ {t}) = 0, t ∈ S + ∪ L− . Then, inequality (4) can be explicitly written as X X y(S + ) + (yt − min{ct , µj }xt ) + (ct − µj )+ (1 − xt ) ≤ c(S + ) + y(E − \ S − ). (17) t∈E + \S + t∈S − We refer to inequalities (17) as path pack inequalities. Remark 4. Observe that for a path consisting of a single node N = {j} with demand d := dj > 0, the path pack inequalities (17), reduce to the flow pack inequalities ((Atamtürk, 2001)). Let (S + , S − ) be a flow pack and µ := d−c(S + )+c(S − ) > 0. Moreover, the maximum flow that can be sent through S + for demand d and arcs in S − is c(S + ). Then, the value 11 function v(S + ) = c(S + ) and the resulting path pack inequality X X (ct − µ)+ (1 − xt ) ≤ c(S + ) + y(E − \ S − ) (18) y(S + ) + (yt − min{ct , µ}xt ) + t∈E + \S + t∈S − is equivalent to the flow pack inequality. Proposition 5. Let (S + , S − ) be a path pack for the node set N . The path pack inequality for N is at least as strong as the flow pack inequality for the single node relaxation obtained by merging the nodes in N . Proof. The proof is similar to that of Proposition 4. Flow pack and path pack inequalities only differ in the coefficients of variables xt for t ∈ E + \ S + and t ∈ S − . Therefore, we compare the values µj , j ∈ N of path pack inequalities (17) to the value µ of flow pack inequalities (18) and show that µj ≤ µ for all j ∈ N . For the single node relaxation, the values of the minimum cuts that pass above and below the merged node are m̄u = c(S + ) and m̄d = d1n + c(S − ), respectively. The recursions in (7)-(10) imply that muj = αju + βju − c(Sj+ ) ≤ c(S + ) = m̄u and mdj = αjd + βjd − dj − c(Sj− ) ≤ d1n + c(S − ) = m̄d . The coefficient for flow pack inequality is µ = (m̄d − m̄u )+ and for path pack inequality µj = (mdj − muj )+ . Since (S + , S − ) is a path pack, the minimum cut passes above all nodes in N and muj = c(S + ) for all j ∈ N . As a result, muj = m̄u for all j ∈ N and mdj ≤ m̄d . Then, observe that the values µj ≤ µ, j ∈ N.  Example 1 (continued). Recall the lot-sizing instance with backlogging given in Figure 3. Let the node set N = [1, 4] with E − = ∅ and S + = {3}. Then, mu1 = 30, md1 = 40, mu2 = 30, md2 = 40, mu3 = 30, md3 = 30, mu4 = 30, md4 = 30, leading to µ1 = 10, µ2 = 10, µ3 = 0 and µ4 = 0. Moreover, the maximum flow values are v(C) = 30, v(C ∪ {1}) = 40, v(C ∪ {2}) = 40, v(C ∪ {4}) = 30, v(E) = 40, and v(E \ {3}) = 40. Then the resulting path pack inequality (17) is y1 + y2 + y3 + y4 ≤ 30 + 10x1 + 10x2 (19) and it is facet-defining for conv(P) as will be shown in Section 4. Now, suppose that the nodes in [1, 4] are merged into a single node with incoming arcs {1, 2, 3, 4} and demand d = 40. For the same set S + , we get µ = 40 − 30 = 10. Then, the corresponding flow pack inequality (18) is y1 + y2 + y3 + y4 ≤ 30 + 10x1 + 10x2 + 10x4 , which is weaker than the path pack inequality (19).  Proposition 6. If |E + \S + | ≤ 1 and S − = ∅, then inequalities (14) and (17) are equivalent. Proof. If E + \ S + = ∅ and S − = ∅, then it is easy to see that the coefficients of inequality (17) are the same as (14). Moreover, if |E + \ S + | = 1 (and wlog E + \ S + = {j}), then the 12 resulting inequality (17) is y(E + ) − y(E − ) ≤ v(C) + ρj (C)xj  = v(C) + v(C ∪ {j}) − v(C) xj = v(C ∪ {j}) − ρj (C)(1 − xj ), which is equivalent to path cover inequality (14) with the objective set (E + , ∅).  4. The strength of the path cover and pack inequalities The capacities of the forward and the backward path arcs play an important role in finding the coefficients of the path cover and pack inequalities (14) and (17). Recall that K + and K − are the coefficient sets in (5), (S + , L− ) is the objective set for (F2) and S − = E − \ (K − ∪ L− ). Definition 3. A node j ∈ N is called backward independent for set (S + , S − ) if d αju = αj−1 + uj−1 + c(Sj+ ), or u αjd = αj−1 + bj−1 + dj + c(Sj− ). Definition 4. A node j ∈ N is called forward independent for set (S + , S − ) if d βju = βj+1 + bj + c(Sj+ ), or u βjd = βj+1 + uj + dj + c(Sj− ). Intuitively, backward independence of node j ∈ N implies that the minimum cut either passes through the forward path arc (j − 1, j) or through the backward path arc (j, j − 1). Similarly, forward independence of node j ∈ N implies that the minimum cut either passes through the forward path arc (j, j + 1) or through the backward path arc (j + 1, j). In Lemmas 7 and 8 below, we further explain how forward and backward independence affect + − the coefficients of path cover and pack inequalities. First, let Sjk = ∪ki=j Si+ , Sjk = ∪ki=j Si− − k and L− jk = ∪i=j Li if j ≤ k, and ∅ otherwise. Lemma 7. If a node j ∈ N is backward independent for set (S + , S − ), then the values λj + − and µj do not depend on the sets S1j−1 , S1j−1 and the value d1j−1 . d Proof. If a node j is backward independent, then either αju = αj−1 + uj−1 + c(Sj+ ) or − + d u u d αj = αj−1 + bj−1 + dj + c(Sj ). If αj = αj−1 + uj−1 + c(Sj ), then the equality in (7) d u d implies αj−1 + uj−1 ≤ αj−1 . As a result, the equality in (8) gives αjd = αj−1 + dj + c(Sj− ). Following the definitions in (11)–(12), the difference wj := muj − mdj is βju − βjd + uj−1 which only depends on sets Sk+ and Sk− for k ∈ [j, n], the value djn and the capacity of the forward path arc (j − 1, j). u u d If αjd = αj−1 +bj−1 +dj +c(Sj− ), then the equality in (8) implies αj−1 +bj−1 ≤ αj−1 . As a + u u u result, the equality in (7) gives αj = αj−1 + c(Sj ). Then, the difference wj = βj − βjd − bj−1 13 which only depends on sets Sk+ and Sk− for k ∈ [j, n], the value djn and the capacity of the backward path arc (j, j − 1). Since the values λj and µj are defined as (wj )+ and (−wj )+ respectively, the result follows.  Remark 5. Let wj := muj − mdj . If a node j ∈ N is backward independent for a set (S + , S − ), d + uj−1 + c(Sj+ ), then then we observe the following: (1) If αju = αj−1 wj = βju − βjd + uj−1 , u + bj−1 + dj + c(Sj− ), then and (2) if αjd = αj−1 wj = βju − βjd − bj−1 . Lemma 8. If a node j ∈ N is forward independent for set (S + , S − ), then the values λj + − and µj do not depend on the sets Sj+1n , Sj+1n and the value dj+1n . d d Proof. The forward independence implies either βju = βj+1 + bj + c(Sj+ ) and βjd = βj+1 + − + − u u d u dj + c(Sj ) or βj = βj+1 + c(Sj ) and βj = βj+1 + uj + dj + c(Sj ). Then, the difference wj = muj − mdj is either αju + αjd + bj or αju + αjd − uj and in both cases, it is independent of the sets Sk+ , Sk− for k ∈ [1, j − 1] and the value d1j−1 .  Remark 6. Let wj := muj − mdj . If a node j ∈ N is forward independent for a set (S + , S − ), d then we observe the following: (1) If βju = βj+1 + bj + c(Sj+ ), then wj = αju − αjd + bj , u and (2) if βjd = βj+1 + uj + dj + c(Sj− ), then wj = αju − αjd − uj . Corollary 9. If a node j ∈ N is backward independent for set (S + , S − ), then the values + − λk and µk for k ∈ [j, n] are also independent of the sets S1j−1 , S1j−1 and the value d1j−1 . + − Similarly, if a node j ∈ N is forward independent for set (S , S ), then the values λk and + − µk for k ∈ [1, j] are also independent of the sets Sj+1n , Sj+1n and the value dj+1n . Proof. The proof follows from recursions in (7)–(10). If a node j is backward independent, u d u d we write αj+1 and αj+1 in terms of αj−1 and αj−1 and observe that the difference wj+1 = + u d u d mj+1 − mj+1 does not depend on αj−1 nor αj−1 which implies independence of sets S1j−1 , − S1j−1 and the value d1n . We can repeat the same argument for wj , j ∈ [j + 2, n] to show independence. u d We show the same result for forward independence by writing βj−1 and βj−1 in terms of u d u d βj+1 and βj+1 , we observe that wj does not depend on βj+1 nor βj+1 . Then, it is clear that + − wj−1 is also independent of the sets Sj+1n , Sj+1n and the value dj+1n . We can repeat the same argument for wj , j ∈ [1, j − 1] to show independence.  Proving the necessary facet conditions frequently requires a partition of the node set N into two disjoint sets. Suppose, N is partitioned into N1 = [1, j − 1] and N2 = [j, n] for some j ∈ N . Let EN 1 and EN 2 be the set of non-path arcs associated with node sets N1 and N2 . We consider the forward and backward path arcs (j − 1, j) and (j, j − 1) to be in the set of non-path arcs EN 1 and EN 2 since the node j − 1 ∈ Ni and j ∈ / Ni for i = 1, 2. In 14 + + − − + + particular, EN 1 := (j, j − 1) ∪ E1j−1 , EN 1 := (j − 1, j) ∪ E1j−1 and EN 2 := (j − 1, j) ∪ Ejn , − − + − + − ` ` EN 2 := (j, j − 1) ∪ Ejn , where Ek` and Ek` are defined as ∪i=k Ei and ∪i=k Ei if k ≤ ` respectively, and as the empty set otherwise. Since the path arcs for N do not have associated fixed-charge variables, one can assume that there exists auxiliary binary variables x̃k = 1 for + + k ∈ {(j − 1, j), (j, j − 1)}. Moreover, we partition the sets S + , S − and L− into SN 1 ⊇ S1j−1 , − − − − + + − − − − SN 1 ⊇ S1j−1 , LN 1 := L1j−1 and SN 2 ⊇ Sjn , SN 2 ⊇ Sjn , LN 2 := Ljn . Then, let v1 and v2 be the value functions defined in (F2) for the node sets N1 and N2 and the objective + − + − u d u d sets (SN 1 , LN 1 ) and (SN 2 , LN 2 ). Moreover, let αj , αj , βj and βj be defined for j ∈ N in + − − − recursions (7)-(10) for the set (S , S ) and recall that S = E \ (K − ∪ L− ). Lemma 10. Let (S + , L− ) be the objective set for the node set for N = [1, n]. If αju = + d u αj−1 + uj−1 + c(Sj+ ) or βj−1 = βjd + bj−1 + c(Sj−1 ), then + − + − v(S + , L− ) = v1 (SN 1 , LN 1 ) + v2 (SN 2 , LN 2 ), + + + where N1 = [1, j − 1], N2 = [j, n] and the arc sets are SN 1 = (j, j − 1) ∪ S1j−1 , SN 2 = + − − − − (j − 1, j) ∪ Sjn , SN 1 = S1j−1 , SN 2 = Sjn . Proof. See Appendix B.1.  Lemma 11. Let (S + , L− ) be the objective set for the node set for N = [1, n]. If αjd = − u d αj−1 + bj−1 + dj−1 + c(Sj− ) or βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ), then + − + − v(S + , L− ) = v1 (SN 1 , LN 1 ) + v2 (SN 2 , LN 2 ), + + + + − where N1 = [1, j − 1], N2 = [j, n] and the arc sets are SN 1 = S1j−1 , SN 2 = Sjn , SN 1 = − − − (j − 1, j) ∪ S1j−1 , SN 2 = (j, j − 1) ∪ Sjn . Proof. See Appendix B.2.  Lemma 12. Let (S + , L− ) be the objective set for the node set for N = [1, n]. If αju = − d d αj−1 + uj−1 + c(Sj+ ) and βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ), then − + − + v(S + , L− ) = v1 (SN 1 , LN 1 ) + v2 (SN 2 , LN 2 ), + + + + where N1 = [1, j − 1], N2 = [j, n] and the arc sets are SN 1 = S1j−1 , SN 2 = (j − 1, j) ∪ Sjn , − − − − SN 1 = S1j−1 , SN 2 = (j, j − 1) ∪ Sjn . Proof. See Appendix B.3.  Lemma 13. Let (S + , L− ) be the objective set for the node set for N = [1, n]. If αjd = + u u αj−1 + bj−1 + dj + c(Sj− ) and βj−1 = βjd + bj−1 + c(Sj−1 ), then + − + − v(S + , L− ) = v1 (SN 1 , LN 1 ) + v2 (SN 2 , LN 2 ), + + + + where N1 = [1, j − 1], N2 = [j, n] and the arc sets are SN 1 = (j, j − 1) ∪ S1j−1 , SN 2 = Sjn , − − − − SN 1 = (j − 1, j) ∪ S1j−1 , SN 2 = Sjn . Proof. See Appendix B.4.  In the remainder of this section, we give necessary and sufficient conditions for path cover and pack inequalities (14) and (17) to be facet-defining for the convex hull of P. 15 Theorem 14. Let N = [1, n], and dj ≥ 0 for all j ∈ N . If L− = ∅ and the set (S + , S − ) is a path cover for N , then the following conditions are necessary for path cover inequality (14) to be facet-defining for conv(P): (i) ρt (C \ {t}) < ct , for all t ∈ C, (ii) maxt∈S + ρt (C \ {t}) > 0, (iii) if a node j ∈ [2, n] is forward independent for set (S + , S − ), then node j − 1 is not backward independent for set (S + , S − ), (iv) if a node j ∈ [1, n − 1] is backward independent for set (S + , S − ), then node j + 1 is not forward independent for set (S + , S − ), (v) if maxt∈S + (ct − λi )+ = 0 for i = p, . . . , n for some p ∈ [2, n], then the node p − 1 is i not forward independent for (S + , S − ), (vi) if maxt∈S + (ct − λi )+ = 0 for i = 1, . . . , q for some q ∈ [1, n − 1], then the node q + 1 i is not backward independent for (S + , S − ). Proof. (i) If for some t0 ∈ S + , ρt0 (C \ {t0 }) ≥ ct0 , then the path cover inequality with the objective set S + \ {t0 } summed with yt0 ≤ ct0 xt0 results in an inequality at least as strong. Rewriting the path cover inequality using the objective set S + , we obtain X (yt + ρt (S + \ {t})(1 − xt )) + yt0 ≤ v(S + ) − ρt0 (S + \ {t0 })(1 − xt0 ) + y(E − \ S − ) t∈S + \{t0 } = v(S + \ {t0 }) + ρt0 (S + \ {t0 })xt0 + y(E − \ S − ). Now, consider summing the path cover inequality for the objective set S + \ {t0 } X (yt + ρt (S + \ {t, t0 })(1 − xt )) ≤ v(S \ {t0 }) + y(E − \ S − ), t∈S + \{t0 } and yt0 ≤ ct0 xt0 . The resulting inequality dominates inequality (3) because ρt (S + \ {t}) ≤ ρt (S + \ {t, t0 }), from the submodularity of the set function v. If the assumption of L− = ∅ is dropped, this condition extends for arcs t ∈ L− as ρt (C \ {t}) > −ct with a similar proof. (ii) If L− = ∅ and maxt∈S + ρt (C \ {t}) = 0, then summing flow balance equalities (1b) for all nodes j ∈ N gives an inequality at least as strong. (iii) Suppose a node j is forward independent for (S + , S − ) and the node j − 1 is backward independent for (S + , S − ) for some j ∈ [2, n]. Lemmas 10–13 show that the nodes N and the arcs C = S + ∪ L− can be partitioned into N1 = [1, j − 1], N2 = [j, n] and C1 , C2 such that the sum of the minimum cut values for N1 , N2 is equal to the minimum cut for N . From Remarks 5 and 6 and Corollary 9, it is easy to see that λi for i ∈ N will not change by the partition procedures described in Lemmas 10–13. We examine the four cases for node j − 1 to be forward independent and node j to be backward independent for the set (S + , S − ). + d u (a) Suppose αju = αj−1 + uj−1 + c(Sj+ ) and βj−1 = βjd + bj−1 + c(Sj−1 ). Consider + + the partition procedure described in Lemma 10, where SN 1 = (j, j − 1) ∪ S1j−1 , + + − − − − SN = (j − 1, j) ∪ S , S = S , S = S . Then, the path cover inequalities jn 1j−1 jn 2 N1 N2 16 for nodes N1 and N2 rj−1 + j−1 X X j−1 X  + yt + (ct − λi )+ (1 − xt ) ≤ v1 (SN ) + y(Ei− \ Si− ) + ij−1 1 i=1 t∈S + i i=1 and ij−1 + n X X n X  + yt + (ct − λi )+ (1 − xt ) ≤ v2 (SN ) + y(Ei− \ Si− ) + rj−1 2 i=j t∈S + i i=j summed gives n X X  yt + (ct − λi )+ (1 − xt ) ≤ v(S + ) + y(E − \ S − ), i=1 t∈S + i which is the path cover inequality for N with the objective set S + . − u d (b) Suppose αjd = αj−1 + bj−1 + dj−1 + c(Sj− ) and βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ). + + + + Consider the partition described in Lemma 11, where SN 1 = S1j−1 , SN 2 = Sjn , − − − − SN 1 = (j − 1, j) ∪ S1j−1 , SN 2 = (j, j − 1) ∪ Sjn . The path cover inequalities for nodes N1 and N2 j−1 X X j−1 X  + yt + (ct − λi )+ (1 − xt ) ≤ v1 (SN ) + y(Ei− \ Si− ) 1 i=1 t∈S + i i=1 and n X X n X  + yt + (ct − λi ) (1 − xt ) ≤ v2 (SN 2 ) + y(Ei− \ Si− ). + i=j t∈S + i i=j summed gives the path cover inequality for nodes N and arcs C. − d d (c) Suppose αju = αj−1 +uj−1 +c(Sj+ ) and βj−1 = βju +uj−1 +dj−1 +c(Sj−1 ). Consider + + + + the partition described in Lemma 12, where SN 1 = S1j−1 , SN 2 = (j − 1, j) ∪ Sjn , − − − − SN 1 = S1j−1 , SN 2 = (j, j − 1) ∪ Sjn . The path cover inequalities for nodes N1 and N2 j−1 X X j−1 X  + yt + (ct − λi )+ (1 − xt ) ≤ v1 (SN ) + y(Ei− \ Si− ) + ij−1 1 i=1 t∈S + i i=1 and ij−1 + n X X n X  + yt + (ct − λi )+ (1 − xt ) ≤ v2 (SN ) + y(Ei− \ Si− ). 2 i=j t∈S + i i=j summed gives the path cover inequality for nodes N and arcs C. + u u (d) Suppose αjd = αj−1 + bj−1 + dj + c(Sj− ) and βj−1 = βjd + bj−1 + c(Sj−1 ). Consider + + + + the partition described in Lemma 13, where SN 1 = (j, j − 1) ∪ S1j−1 , SN 2 = Sjn , − − − − SN 1 = (j − 1, j) ∪ S1j−1 , SN 2 = Sjn . The path cover inequalities for nodes N1 and 17 N2 rj−1 + j−1 X X j−1 X  + yt + (ct − λi )+ (1 − xt ) ≤ v1 (SN ) + y(Ei− \ Si− ) 1 i=1 t∈S + i and n X X i=1 n X  + yt + (ct − λi )+ (1 − xt ) ≤ v2 (SN ) + y(Ei− \ Si− ) + rj−1 . 2 i=j t∈S + i i=j summed gives the path cover inequality for nodes N and arcs C. (iv) The same argument for condition (iii) above also proves the desired result here. (v) Suppose (ct − λi )+ = 0 for all t ∈ Si+ and i ∈ [p, n] and the node p − 1 is forward independent for some p ∈ [2, n]. Then, we partition the node set N = [1, n] into + u N1 = [1, p − 1] and N2 = [p, n]. We follow Lemma 10 if βp−1 = βpd + bp−1 + c(Sp−1 ) − + − + d u and follow Lemma 11 if βp−1 = βp + up−1 + dp−1 + c(Sp−1 ) to define SN 1 , SN 1 , SN 2 − and SN 2 . Remark 6 along with the partition procedure described in Lemma 10 or 11 implies that λi will remain unchanged for i ∈ N1 . The path cover inequality for nodes N and arcs C is y(S + ) + p−1 X X (ct − λi )+ (1 − xt ) ≤ v(S + ) + y(E − \ S − ). i=1 t∈S + i + u If βp−1 = βjd + bp−1 + c(Sp−1 ), then the path cover inequality for nodes N1 and arcs + − SN 1 , SN 1 described in Lemma 10 is rp−1 + p−1 X X p−1 X  + yt + (ct − λi )+ (1 − xt ) ≤ v(SN ) + y(Ei− \ Si− ) + ip−1 . 1 i=1 t∈S + i i=1 Moreover, let m̄up and m̄dp be the values of minimum cut that goes above and below + − node p for the node set N2 and arcs SN 2 , SN 2 and observe that m̄up = βpu + up−1 and m̄dp = βpd . Then, comparing the difference λ̄p := (m̄up − m̄dp )+ = (βpu − βpd + up−1 )+ to λp = (mup − mdp )+ = (βpu − βpd + αpu − αpd + c(Sp+ ) − dp − c(Sp− ))+ , we observe that λ̄p ≥ λp since αpu − αpd + c(Sp+ ) − dp − c(Sp− ) ≤ up−1 from (7)–(8). Since (ct − λp )+ = 0, then (ct − λ̄p )+ = 0 as well. Using the same technique, it is easy to observe that λ̄i ≥ λi + for i ∈ [p + 1, n] as well. As a result, the path cover inequality for N2 with sets SN 2, − SN 2 is ip−1 + n X y(Si+ ) ≤ + v(SN 2) + i=p n X y(Ei− \ Si− ) + rp−1 . i=p + − The path cover inequalities and for N2 , SN 2 , SN 2 summed gives the path cover inequality for N , S , S . − d Similarly, if βp−1 = βju + up−1 + dp−1 + c(Sp−1 ), the proof follows very similar to the previous argument using Lemma 11. Letting m̄up and m̄dp be the values of minimum + for N1 , SN 1, + − − SN 1 18 + − cut that goes above and below node p for the node set N2 and arcs SN 2 , SN 2 , we get m̄up = βpu and m̄dp + bp−1 = βpd under this case. Now, notice that that αpu − αpd + c(Sp+ ) − dp − c(Sp− ) ≥ −bp−1 from (7)–(8), which leads to λ̄p ≥ λp . Then the proof follows same as above. (vi) The proof is similar to that of the necessary condition (v). We use Lemmas 12 and 13 and Remark 6 to partition the node set N and arcs S + , S − into node sets N1 = [1, q] + − + − and N2 = [q + 1, n] for q ∈ [2, n] and arcs SN 1 , SN 1 and SN 2 , SN 2 . Next, we check the values of minimum cut that goes above and below node q for the node set N1 and arcs + − u d + − SN 1 , SN 1 . Then, observing −uq ≤ βq − αq + c(Sq ) − dq − c(Sq ) ≤ bq from (9)–(10), it + is easy to show that the coefficients xt for t ∈ SN 1 are equal to zero in the path cover + − inequality for node set N1 . As a result, the path cover inequalities for N1 , SN 1 , SN 1 + − + − and for N2 , SN 2 , SN 2 summed gives the path cover inequality for N , S , S .  Remark 7. If the node set N consists of a single node, then the conditions (i) and (ii) of Theorem 14 reduce to the sufficient facet conditions of flow cover inequalities given in (Padberg et al., 1985, Theorem 4) and (van Roy and Wolsey, 1986, Theorem 6). In this setting, conditions (iii)–(vi) are no longer relevant. Theorem 15. Let N = [1, n], E − = ∅, dj > 0 and |Ej+ | = 1, for all j ∈ N and let the set S + be a path cover. The necessary conditions in Theorem 14 along with (i) (ct − λj )+ > 0 for all t ∈ Sj+ , j ∈ N , (ii) (ct − λj )+ < c(E + \ S + ) for all t ∈ Sj+ , j ∈ N are sufficient for path cover inequality (14) to be facet-defining for conv(P).  Proof. Recall that dim conv(P) = 2|E| + n − 2. In this proof, we provide 2|E| + n − 2 affinely independent points that lie on the face F ( ) X F = (x, y, i, r) ∈ P : y(S + ) + (ct − λj )+ (1 − xt ) = d1n . t∈S + First, we provide Algorithm 1 which outputs an initial feasible solution (x̄, ȳ, ī, r̄), where all the arcs in S + have non-zero flow. Let d¯j be the effective demand on node j, that is, the sum of dj and the minimal amount of flow that needs to be sent from the arcs in Sj+ to ensure v(S + ) = d1n . In Algorithm 1, we perform a backward pass and a forward pass on the nodes in N . This procedure is carried out to obtain the minimal amounts of flow on the forward and backward path arcs to satisfy the demands. For each node j ∈ N , these minimal outgoing flow values added to the demand dj give the effective demand d¯j . Algorithm 1 ensures that at most one of the path arcs (j − 1, j) and (j, j − 1) have non-zero flow for all j ∈ [2, n]. Moreover, note that sufficient condition (i) ensures that all the arcs in S + have nonzero flow. Moreover, for at least one node i ∈ N , it is guaranteed that c(Si+ ) > d¯i . Otherwise, ρt (C) = ct for all t ∈ S + which contradicts the necessary condition (i). Necessary conditions (iii) and (iv) ensure that īj < uj and r̄j < bj for all 19 Algorithm 1 Initialization: Let d¯j = dj for j ∈ N for j = (n − 1)nto 1 do + o , Let ∆ = min uj , d¯j+1 − c(S + ) j+1 d¯j = d¯j + ∆, d¯j+1 = d¯j+1 − ∆, īj = ∆. end for for j = 2 to n do + + Let ∆ = d¯j−1 − c(Sj−1 ) , d¯j = d¯j + ∆, d¯j−1 − ∆ r̄j−1 = ∆ − min{∆, īj−1 } īj−1 = īj−1 − min{∆, īj−1 } end for ȳj = d¯j , for all j ∈ S + . x̄j = 1 if j ∈ S + , 0 otherwise. ȳj = x̄j = 0, for all j ∈ E − . j = 1, . . . , n − 1. Let e := arg max{c(Si+ ) − d¯i } i∈N be the node with the largest excess capacity. Also let 1j be the unit vector with 1 at position j. Next, we give 2|S + | affinely independent points represented by w̄t = (x̄t , ȳt , īt , r̄t ) and t w̃ = (x̃t , ỹt , ĩt , r̃t ) for t ∈ S + : (i) Select w̄e = (x̄, ȳ, ī, r̄) given by Algorithm 1. Let ε > 0 be a sufficiently small value. We define w̄t for e 6= t ∈ S + as ȳt = ȳe + ε1e − ε1t , x̄t = x̄e . If t < e, then īt = īe and r̄jt = r̄je for j < t and for t ≥ e, r̄jt = r̄je + ε for t ≤ j < e. (ii) In this class of affinely independent solutions, we close the arcs in S + one at a time and P open all the arcs in E + \ S + : x̃t = x̄ − 1t + j∈E + \S + 1j . Next, we send an additional ȳt − (ct − λj )+ amount of flow from the arcs in S + \ {t}. This is a feasible operation because v(C \ {t}) = d1n − (ct − λj )+ . Let (y∗ , i∗ , r∗ ) be the optimal solution of (F2) corresponding to v(S + \{t}). Then let, ỹjt = yj∗ for j ∈ S + \{t}. Since v(C \{t}) < d1n , additional flow must be sent through nodes in E + \S + to satisfy flow balance equations (1b). This is also a feasible operation, because of assumption (A.1). Then, the forward and backward path flows ĩt and r̃t are calculated using the flow balance equations. In the next set of solutions, we give 2|E + \S + |−1 affinely independent points represented by ŵt = (x̂t , ŷt , ît , r̂t ) and w̌t = (x̌t , y̌t , ǐt , řt ) for t ∈ E + \ S + . (iii) Starting with solution w̄e , we open arcs in E + \ S + , one by one. ŷt = ȳe , x̂t = x̄e + 1t , ît = īe , r̂t = r̄e . (iv) If |E + \ S + | ≥ 2, then we can send a sufficiently small ε > 0 amount of flow from arc t ∈ E + \ S + to t 6= k ∈ E + \ S + . Let this set of affinely independent points be represented by w̌t for t ∈ E + \ S + . While generating w̌t , we start with the solution w̃e , 20 where the non-path arc in Se+ is closed. The feasibility of this operation is guaranteed by the sufficiency conditions (ii) and necessary conditions (iii) and (iv). e (a) If ỹte = ct , then there exists at least one arc t 6= m ∈ E + \S + such that 0 ≤ ỹm < cm + + e due to sufficiency assumption (ii), then for each t ∈ E \ S such that ỹt = ct , let Pm−1 y̌t = ỹe − ε1t + ε1m , x̌t = x̃e . If t < m, then ǐt = ĩe and řt = r̃e + ε i=t 1i . If P t−1 t > m, then ǐt = ĩe + ε i=m 1i and řt = r̃e . e e (b) If ỹt < ct and there exists at least one arc t 6= m ∈ E + \ S + such that ỹm = 0, then the same point described in (a) is feasible. e (c) If ỹte < ct and there exists at least one arc t 6= m ∈ E + \ S + such that ỹm = cm , t e then, we send ε amount of flow from t to m, y̌ = ỹ + ε1t − ε1m , x̌t = x̃e . Pm−1 If t < m, then ǐt = ĩe + ε i=t 1i and řt = r̃e . If t > m, then ǐt = ĩe and P t−1 řt = r̃e + ε i=m 1i . Finally, we give n − 1 points that perturb the flow on the forward path arcs (j, j + 1) for j = 1, . . . , n − 1 represented by w̆j = (x̆j , y̆j , ĭj , r̆j ). Let k = min{i ∈ N : Si+ 6= ∅} and ` = max{i ∈ N : Si+ 6= ∅}. The solution given by Algorithm 1 guarantees īj < uj and r̄j < bj for j = 1, . . . , n − 1 due to necessary conditions (iii) and (iv). (v) For j = 1, . . . , n − 1, we send an additional ε amount of flow from the forward path arc (j, j + 1) and the backward path arc (j + 1, j). Formally, the solution w̆j can be obtained by: y̆j = ȳe , x̆j = x̄e , ĭj = īe + ε1j and r̆j = r̄e + ε1j .  Next, we identify conditions under which path pack inequality (17) is facet-defining for conv(P). Theorem 16. Let N = [1, n], dj ≥ 0 for all j ∈ N , let the set (S + , S − ) be a path pack and L− = ∅. The following conditions are necessary for path pack inequality (17) to be facet-defining for conv(P): (i) ρj (S + ) < cj , for all j ∈ E + \ S + , (ii) maxt∈S − ρt (C) > 0, (iii) if a node j ∈ [2, n] is forward independent for set (S + , S − ), then node j − 1 is not backward independent for set (S + , S − ), (iv) if a node j ∈ [1, n − 1] is backward independent for set (S + , S − ), then node j + 1 is not forward independent for set (S + , S − ), (v) if maxt∈E + \S + ρt (C) = 0 and maxt∈S − ρt (C) = 0 for i = p, . . . , n for some p ∈ [2, n], i i i then the node p − 1 is not forward independent for (S + , S − ), (vi) if maxt∈E + \S + ρt (C) = 0 and maxt∈S − ρt (C) = 0 for i = 1, . . . , q for some q ∈ i i i [1, n − 1], then the node q + 1 is not backward independent for (S + , S − ). Proof. (i) Suppose that for some k ∈ E + \ S + , ρk (S + ) = ck . Then, recall the implicit form of path pack inequality (17) is X X y(E + \ {k}) + yk + ρt (S + )(1 − xt ) ≤ v(S + ) + ρt (S + )xt + ck xk + y(E − \ S − ). t∈S − 21 k6=t∈E + \S + Now, if we select ak = 0 in (F2), then the coefficients of xk and yk become zero and summing the path cover inequality X X ρt (S + )xt + y(E − \ S − ). ρt (S + )(1 − xt ) ≤ v(S + ) + y(E + \ {k}) + t∈S − k6=t∈E + \S + with yk ≤ ck xk gives the first path cover inequality. (ii) Suppose that ρj (S + ) = 0 for all j ∈ S − . Then the path pack inequality is X  y(E + ) ≤ v(S + ) + ρt (S + )xt + y E − \ (L− ∪ S − ) , t∈E + \S + where L− = ∅. If an arc j is dropped from S − and added to L− , then v(S + ) = v(S + ∪ {j}) since ρj (S + ) = 0 for j ∈ S − . Consequently, the path pack inequality with S − = S − \ {j} and L− = {j} X X  y(E + ) + ρt (S + ∪ {j})(1 − xt ) ≤ v(S + ) + ρt (S + ∪ {j})xt + y E − \ (L− ∪ S − ) . t∈S − t∈E + \S + But since 0 ≤ ρt (S + ∪ {j}) ≤ ρt (S + ) from submodularity of v and ρt (S + ) = 0 for all t ∈ S − , we observe that the path pack inequality above reduces to X  y(E + ) ≤ v(S + ) + ρt (S + ∪ {j})xt + y E − \ (L− ∪ S − ) t∈E + \S + and it is at least as strong as the first pack inequality for S + , S − and L− = ∅. (iii)–(iv) We repeat the same argument of the proof of condition (iii) of Theorem 14. Suppose a node j is forward independent for (S + , S − ) and the node j − 1 is backward independent for (S + , S − ) for some j ∈ [2, n]. Lemmas 10–13 show that the nodes N and the arcs C = S + ∪ L− can be partitioned into N1 = [1, j − 1], N2 = [j, n] and C1 , C2 such that the sum of the minimum cut values for N1 , N2 is equal to the minimum cut for N . From Remarks 5 and 6 and Corollary 9, it is easy to see that µi for i ∈ N will not change by the partition procedures described in Lemmas 10–13. We examine the four cases for node j − 1 to be forward independent and node j to be backward independent for the set (S + , S − ). For ease of notation, let Q+ jk := k X X (yt − min{µi , ct }xt ) i=j t∈E + \S + i i and Q− jk := k X X i=j (ct − µi )+ (1 − xt ) t∈Si− for j ≤ k and j ∈ N , k ∈ N (and zero if j > k), where the values µi are the coefficients that appear in the path pack inequality (17). As a result, the path pack inequality can be written as − − − y(S + ) + Q+ 1n ≤ v(C) + Q1n + y(E \ S ). d αj−1 + uj−1 + c(Sj+ ) u (a) Suppose αju = and βj−1 = the partition procedure described in Lemma 10, 22 (20) + βjd + bj−1 + c(Sj−1 ). Consider + + where SN = (j, j − 1) ∪ S1j−1 , 1 + + − − − − SN 2 = (j − 1, j) ∪ Sjn , SN 1 = S1j−1 , SN 2 = Sjn . Then, the path pack inequalities for nodes N1 is + − − − rj−1 + y(S1j−1 ) + Q+ 1j−1 ≤ v1 (C1 ) + Q1j−1 + y(E1j−1 \ S1j−1 )f + ij−1 . (21) Similarly, the path pack inequality for N2 is + − − − ij−1 + y(Sjn ) + Q+ jn ≤ v2 (C2 ) + Qjn + y(Ejn \ Sjn ) + rj−1 . (22) Inequalities (21)–(22) summed gives the path pack inequality (20). − u d (b) Suppose αjd = αj−1 + bj−1 + dj + c(Sj− ) and βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ). + + + + Consider the partition described in Lemma 11, where SN 1 = S1j−1 , SN 2 = Sjn , − − − − SN 1 = (j − 1, j) ∪ S1j−1 , SN 2 = (j, j − 1) ∪ Sjn . The submodular inequality (4) for nodes N1 where the objective coefficients of (F2) are selected as at = 1 + − − for t ∈ E1j−1 , at = 0 for t = (j, j − 1), at = −1 for t ∈ EN 1 \ SN 1 and at = 0 − for t ∈ SN 1 is X + − − − y(S1j−1 )+ kt (1 − xt ) + Q+ (23) 1j−1 ≤ v1 (C1 ) − Q1j−1 + y(E1j−1 \ S1j−1 ), + t∈SN 1 + where kt for t ∈ SN 1 are some nonnegative coefficients. Similarly, the submodular inequality (4) for nodes N2 , where the objective coefficients of (F2) + are selected as at = 1 for t ∈ Ejn , at = 0 for t = (j − 1, j), at = −1 for − − − t ∈ EN 2 \ SN 2 and at = 0 for t ∈ SN 2 is X + − − − y(Sjn )+ kt (1 − xt ) + Q+ (24) jn ≤ v2 (C2 ) − Qjn + y(Ejn \ Sjn ), + t∈SN 2 + where kt for t ∈ SN 2 are some nonnegative coefficients. The sum of inequalities (23)–(24) is at least as strong as the path pack inequality (20). − d d (c) Suppose αju = αj−1 + uj−1 + c(Sj+ ) and βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ). + + + Consider the partition described in Lemma 12, where SN 1 = S1j−1 , SN 2 = + − − − − (j −1, j)∪Sjn , SN 1 = S1j−1 , SN 2 = (j, j −1)∪Sjn . The submodular inequality (4) for nodes N1 where the objective coefficients of (F2) are selected as at = 1 + − − for t ∈ E1j−1 , at = 0 for t = (j, j − 1), at = −1 for t ∈ EN 1 \ SN 1 and at = 0 − for t ∈ SN 1 is X + − − − y(S1j−1 )+ kt (1 − xt ) + Q+ (25) 1j−1 ≤ v1 (C1 ) − Q1j−1 + y(E1j−1 \ S1j−1 ) + ij−1 , + t∈SN 1 + where kt for t ∈ SN 1 are some nonnegative coefficients. The path pack inequality for N2 is + − − − ij−1 + y(Sjn ) + Q+ jn ≤ v2 (C2 ) + Qjn + y(Ejn \ Sjn ). (26) The sum of inequalities (25)–(26) is at least as strong as inequality (20). + u u (d) Suppose αjd = αj−1 + bj−1 + dj + c(Sj− ) and βj−1 = βjd + bj−1 + c(Sj−1 ). + + Consider the partition described in Lemma 13, where SN 1 = (j, j − 1) ∪ S1j−1 , + + − − − − SN 2 = Sjn , SN 1 = (j − 1, j) ∪ S1j−1 , SN 2 = Sjn . The path pack inequalities 23 for nodes N1 is + − − − rj−1 + y(S1j−1 ) + Q+ 1j−1 ≤ v1 (C1 ) + Q1j−1 + y(E1j−1 \ S1j−1 ). (27) The submodular inequality (4) for nodes N2 where the objective coefficients of + (F2) are selected as at = 1 for t ∈ Ejn , at = 0 for t = (j − 1, j), at = −1 for − − − t ∈ EN 2 \ SN 2 and at = 0 for t ∈ SN 2 is X + − − − y(Sjn )+ kt (1 − xt ) + Q+ (28) jn ≤ v2 (C2 ) − Qjn + y(Ejn \ Sjn ) + rj−1 , + t∈SN 2 + where kt for t ∈ SN 2 are some nonnegative coefficients. The sum of inequalities (27)–(28) is at least as strong as the path pack inequality (20). (v) Suppose (ct − µi )+ = 0 for all t ∈ Si− and i ∈ [p, n] and the node p − 1 is forward independent. Then, we partition the node set N = [1, n] into N1 = [1, p − 1] and + u N2 = [p, n]. We follow Lemma 10 if βp−1 = βpd + bp−1 + c(Sp−1 ) and follow Lemma − + − + − d u 11 if βp−1 = βp + up−1 + dp−1 + c(Sp−1 ) to define SN 1 , SN 1 , SN 2 and SN 2 . Remark 6 along with the partition procedure described in Lemma 10 or 11 implies that µi will remain unchanged for i ∈ N1 . + u If βp−1 = βjd + bp−1 + c(Sp−1 ), then the coefficients µi of the path pack inequality + − for nodes N1 and arcs SN 1 , SN 1 described in Lemma 10 is the same as the coefficients of the path pack inequality for nodes N and arcs S + , S − . Moreover, let m̄up and m̄dp be the values of minimum cut that goes above and below node p for the node + − set N2 and arcs SN 2 , SN 2 and observe that m̄up = βpu + up−1 and m̄dp = βpd . Then, comparing the difference µ̄p := (m̄dp − m̄up )+ = (βpd − βpu − up−1 )+ to µp = (mdp − mup )+ = (βpd − βpu + αpd − αpu − c(Sp+ ) + dp + c(Sp− ))+ , we observe that µ̄p ≥ µp since αpd − αpu − c(Sp+ ) + dp + c(Sp− ) ≥ −up−1 from (7)–(8). Since (ct − µp )+ = 0, then (ct − µ̄p )+ = 0 as well. Using the same technique, it is easy to observe that µ̄i ≥ µi for i ∈ [p + 1, n] as well. As a result, the path pack inequality for N2 with + − + sets SN 2 , SN 2 , summed with the path pack inequality for nodes N1 and arcs SN 1 , − + − SN 1 give the path pack inequality for nodes N and arc S , S . − d u Similarly, if βp−1 = βj + up−1 + dp−1 + c(Sp−1 ), the proof follows very similar to the previous argument using Lemma 11. Letting m̄up and m̄dp be the values of + minimum cut that goes above and below node p for the node set N2 and arcs SN 2, − SN 2 , we get m̄up = βpu and m̄dp + bp−1 = βpd under this case. Now, notice that that αpd − αpu − c(Sp+ ) + dp + c(Sp− ) ≤ bp−1 from (7)–(8), which leads to µ̄p ≥ µp . Then the proof follows same as above. (vi) The proof is similar to that of the necessary condition (v) above. We use Lemmas 12–13 and Remark 6 to partition the node set N and arcs S + , S − into node sets + − + − N1 = [1, q] and N2 = [q + 1, n] and arcs SN 1 , SN 1 and SN 2 , SN 2 . Next, we check the values of minimum cut that goes above and below node q for the node set N1 + − d u + − and arcs SN 1 , SN 1 . Then, observing −bq ≤ βq − αq − c(Sq ) + dq + c(Sq ) ≤ uq from − + + (9)–(10), it is easy to see that the coefficients of xt for t ∈ SN 1 and t ∈ EN 1 \ SN 1 24 are equal to zero in the path pack inequality for node set N1 . As a result, the path + − + − pack inequalities for N1 , SN 1 , SN 1 and for N2 , SN 2 , SN 2 summed gives the path + − pack inequality for N , S , S .  Remark 8. If the node set N consists of a single node, then the conditions (i) and (ii) of Theorem 16 reduce to the necessary and sufficient facet conditions of flow pack inequalities given in (Atamtürk, 2001, Proposition 1). In this setting, conditions (iii)–(vi) are no longer relevant. Theorem 17. Let N = [1, n], E − = ∅, dj > 0 and |Ej+ | = 1, for all j ∈ N and let the objective set S + be a path pack for N . The necessary conditions in Theorem 16 along with (i) for each j ∈ E + \ S + , either S + ∪ {j} is a path cover for N or ρj (S + ) = 0, (ii) for each t ∈ S + , there exists jt ∈ E + \ S + such that S + \ {t} ∪ {jt } is a path cover for N, (iii) for each j ∈ [1, n − 1], there exists kj ∈ E + \ S + such that the set S + ∪ {kj } is a path cover and neither node j is backward independent nor node j +1 is forward independent for the set S + ∪ {kj } are sufficient for path pack inequality (17) to be facet-defining for conv(P). Proof. We provide 2|E| + n − 2 affinely independent points that lie on the face:     X F = (x, y, i, r) ∈ P : y(S + ) + (yt − min{µj , ct }xt ) = c(S + ) .   + + t∈E \S Let (y∗ , i∗ , r∗ ) ∈ Q be an optimal solution to (F2). Since S + is a path pack and E − = ∅, v(S + ) = c(S + ). Then, notice that yt∗ = ct for all t ∈ S + . Moreover, let e be the arc with largest capacity in S + , ε > 0 be a sufficiently small value and 1j be the unit vector with 1 at position j. First, we give 2|E + \ S + | affinely independent points represented by z̄ t = (x̄t , ȳt , īt , r̄t ) and z̃ t = (x̃t , ỹt , ĩt , r̃t ) for t ∈ E + \ S + . (i) Let t ∈ E + \ S + , where S + ∪ {t} is a path cover for N . The solution z̄ t has arcs in S + ∪ {t} open, x̄tj = 1 for j ∈ S + ∪ {t}, 0 otherwise, ȳjt = yj∗ for j ∈ S + and ȳtt = ρt (S + ), 0 otherwise. The forward and backward path arc flow values ītj and r̄jt can then be calculated using flow balance equalities (1b) where at most one of them can be nonzero for each j ∈ N . Sufficiency condition (i) guarantees the feasibility of z̄ t . (ii) Let t ∈ E + \ S + , where ρt (S + ) = 0 and let t 6= ` ∈ E + \ S + , where S + ∪ {`} is a path cover for N . The solution z̄ t has arcs in S + ∪{t, `} open, x̄tj = 1 for j ∈ S + ∪{t, `}, and 0 otherwise, ȳjt = yj∗ for j ∈ S + , ȳtt = 0, ȳ`t = ρ` (S + ), and 0 otherwise. The forward and backward path arc flow values ītj and r̄jt can then be calculated using flow balance equalities (1b) where at most one of them can be nonzero for each j ∈ N . Sufficiency condition (i) guarantees the feasibility of z̄ t . 25 (iii) The necessary condition (i) ensures that ρt (S + ) < ct , therefore ȳtt < ct . In solution z̃ t , starting with z̄ t , we send a flow of ε from arc t ∈ E + \ S + to e ∈ S + . Let Pt−1 ỹt = ȳt + ε1t − ε1e and x̃t = x̄t . If e < t, then r̃t = r̄t + ε i=e 1i , ĩt = īt and if e > t, Pe−1 then ĩt = īt + ε i=t 1i , r̃t = r̄t . Next, we give 2|S + | − 1 affinely independent feasible points ẑ t and ž t corresponding to t ∈ S + that are on the face F . Let k be the arc in E + \ S + with largest capacity. (iv) In the feasible solutions ẑ t for e 6= t ∈ S + , we open arcs in S + ∪ {k} and send an ε flow from arc k to arc t. Let ŷt = ȳk + ε1k − ε1t and x̂t = x̄k . If t < k, then Pk−1 Pt−1 r̂t = r̄k + ε i=t 1i , ît = īk and if t > k, then ît = īk + ε i=k 1i , r̂t = r̄k . (v) In the solutions ž t for t ∈ S + , we close arc t and open arc jt ∈ E + \ S + that is introduced in the sufficient condition (ii). Then, x̌tj = 1 if j ∈ S + \ {t} and if j = jt and x̌tj = 0 otherwise. From sufficient condition (ii), there exists y̌jt values that satisfy the flow balance equalities (1b). Moreover, these y̌jt values satisfy inequality (17) at equality since both S + ∪ {jt } and S + \ {t} ∪ {jt } are path covers for N . Then, the forward and backward path arc flows are found using flow balance equalities where at most one of ǐtj and řjt are nonzero for each j ∈ N . Finally, we give n−1 points z̆ j corresponding to forward and backward path arcs connecting nodes j and j + 1. (vi) In the solution set z̆ j for j = 1, . . . , n − 1, starting with solution z̄ kj , where kj is introduced in the sufficient condition (iii), we send a flow of ε from both forward path arc (j − 1, j) and backward path arc (j, j − 1). Since the sufficiency condition (iii) k k ensures that r̄j j < bj and īj j < uj , the operation is feasible. Let y̆j = ȳkj , x̆j = x̄kj , ĭj = īkj + ε1j and r̆j = r̄kj + ε1j .  5. Computational study We test the effectiveness of path cover and path pack inequalities (14) and (17) by embedding them in a branch-and-cut framework. The experiments are ran on a Linux workstation with 2.93 GHz Intel R CoreTM i7 CPU and 8 GB of RAM with 1 hour time limit and 1 GB memory limit. The branch-and-cut algorithm is implemented in C++ using IBM’s Concert Technology of CPLEX (version 12.5). The experiments are ran with one hour limit on elapsed time and 1 GB limit on memory usage. The number of threads is set to one and the dynamic search is disabled. We also turn off heuristics and preprocessing as the purpose is to see the impact of the inequalities by themselves. Instance generation. We use a capacitated lot-sizing model with backlogging, where constraints (1b) reduce to: ij−1 − rj−1 + yj − ij + rj = dj , j ∈ N. Let n be the total number of time periods and f be the ratio of the fixed cost to the variable cost associated with a non-path arc. The parameter c controls how large the non-path arc capacities are with respect to average demand. All parameters are generated from a discrete uniform distribution. The demand for each node is drawn from the range [0, 30] 26 ¯ 1.25 × c × d], ¯ where d¯ and non-path arc capacities are drawn from the range [0.75 × c × d, is the average demand over all time periods. Forward and backward path arc capacities are ¯ 2.0 × d] ¯ and [0.3 × d, ¯ 0.8 × d], ¯ respectively. The variable costs pt , ht drawn from [1.0 × d, and gt are drawn from the ranges [1, 10], [1, 10] and [1, 20] respectively. Finally, fixed costs ft are set equal to f × pt . Using these parameters, we generate five random instances for each combination of n ∈ {50, 100, 150}, f ∈ {100, 200, 500, 1000} and c ∈ {2, 5, 10}. Finding violated inequalities. Given a feasible solution (x∗ , y∗ , i∗ , r∗ ) to a linear programming (LP) relaxation of (F1), the separation problem aims to find sets S + and L− that maximize the difference X X y ∗ (S + ) − y ∗ (E − \ L− ) + (ct − λj )+ (1 − x∗t ) − (min{λj , ct })x∗t − d1n − c(S − ) t∈L− t∈S + for path cover inequality (14) and sets S + and S − that maximize X X y ∗ (S + ) − y ∗ (E − \ S − ) − min{ct , µj }x∗t + (ct − µj )+ (1 − xt ) − c(S + ) t∈S − t∈E + \S + for path pack inequality (17). We use the knapsack relaxation based heuristic separation strategy described in (Wolsey and Nemhauser, 1999, pg. 500) for flow cover inequalities to choose the objective set S + with a knapsack capacity d1n . Using S + , we obtain the values λj and µj for each j ∈ N and let S − = ∅ for path cover and path pack inequalities (14) and (17). For path cover inequalities (14), we add an arc t ∈ E − to L− if λj x∗t < yt∗ and λj < ct . We repeat the separation process for all subsets [k, `] ⊆ [1, n]. Results. We report multiple performance measures. Let zINIT be the objective function value of the initial LP relaxation and zROOT be the objective function value of the LP relaxation after all the valid inequalities added. Moreover, let zUB be the objective function value of the best feasible solution found within time/memory limit among all experiments for ROOT INIT , root gap= 100 × zUB −z . We compute an instance. Let init gap= 100 × zUBz−z zUB UB the improvement of the relaxation due to adding valid inequalities as gap imp= 100 × init gap−root gap −zLB , . We also measure the optimality gap at termination as end gap = zUBzUB init gap where zLB is the value of the best lower bound given by CPLEX. We report the average number of valid inequalities added at the root node under column cuts, average elapsed time in seconds under time, average number of branch-and-bound nodes explored under nodes. If there are instances that are not solved to optimality within the time/memory limit, we report the average end gap and the number of unsolved instances under unslvd next to time results. All numbers except initial gap, end gap and time are rounded to the nearest integers. In Tables 1, 2 and 3, we present the performance with the path cover (14) and path pack (17) inequalities under columns spi. To understand how the forward and backward path arc capacities affect the computational performance, we also apply them to the single node relaxations obtained by merging a path into a single node, where the capacities of forward and backward path arcs within a path are considered to be infinite. In this case, the path inequalities reduce to the flow cover and flow pack inequalities. These results are presented under columns mspi. 27 Table 1. Effect of path size on the performance. p≤5 p=1 n f c init gap imp cuts gap imp p ≤ 0.5 × n cuts gap imp gap (m)spi (m)spi spi mspi spi mspi spi mspi cuts p≤n gap imp cuts spi mspi spi mspi spi mspi 2 14.8 5 44.3 10 58.3 34% 56% 60% 21 52 54 87% 52% 212 106 97% 99% 69% 303 148 99% 96% 70% 277 147 99% 2 14.5 200 5 49.8 10 66.3 32% 43% 38% 22 44 47 81% 57% 257 133 96% 61% 1965 241 96% 61% 2387 241 99% 57% 378 162 100% 57% 1264 171 100% 57% 1420 171 98% 50% 392 169 99% 51% 1235 197 99% 51% 1283 197 2 19.1 500 5 54.4 10 73.0 23% 35% 34% 19 36 43 77% 48% 266 128 90% 49% 2286 222 90% 49% 3249 222 99% 49% 522 185 100% 49% 1981 205 100% 49% 2074 205 99% 40% 498 196 99% 40% 1336 236 99% 40% 1445 236 2 14.6 1000 5 59.7 10 76.9 18% 31% 30% 15 36 41 73% 39% 264 99 83% 40% 2821 211 83% 40% 3918 212 98% 45% 487 201 100% 45% 2077 239 100% 45% 2329 239 100% 36% 529 215 100% 37% 1935 268 100% 37% 2149 268 Average: 45.5 36% 36 92% 51% 365 157 97% 52% 1609 206 97% 52% 1894 206 100 50 52% 1164 158 97% 52% 1233 158 69% 664 151 99% 69% 664 151 70% 574 167 99% 70% 574 167 In Table 1, we focus on the impact of path size on the gap improvement of the path cover and path pack inequalities for instances with n = 50. In the columns under p = 1, we obtain the same results for both mspi and spi since the paths are singleton nodes.. We present these results under (m)spi. In columns p ≤ q, we add valid inequalities for paths of size 1, . . . , q and observe that as the path size increases, the gap improvement of the path inequalities increase rapidly. On average 97% of the initial gap is closed as longer paths are used. On the other hand, flow cover and pack inequalities from merged paths reduce about half of the initial gap. These results underline the importance of exploiting path arc capacities on strengthening the formulations. We also observe that the increase in gap improvement diminishes as path size grows. We choose a conservative maximum path size limit of 0.75 × n for the experiments reported in Tables 2, 3 and 4. In Table 2, we investigate the computational performance of path cover and path pack inequalities independently. We present the results for path cover inequalities under columns titled cov, for path pack inequalities under pac and for both of them under the columns titled spi. On average, path cover and path pack inequalities independently close the gap by 63% and 53%, respectively. However, when used together, the gap improvement is 96%, which shows that the two classes of inequalities complement each other very well. In Table 3, we present other performance measures as well for instances with 50, 100, and 150 nodes. We observe that the forward and backward path arc capacities have a large impact on the performance level of the path cover and pack inequalities. Compared to flow cover and pack inequalities added from merged paths, path cover and path pack inequalities reduce the number of nodes and solution times by orders of magnitude. This is mainly due to better integrality gap improvement (50% vs 95% on average). 28 In Table 4, we examine the incremental effect of path cover and path pack inequalities over the fixed-charge network cuts of CPLEX, namely flow cover, flow path and multicommodity flow cuts. Under cpx, we present the performance of flow cover, flow path and multi-commodity flow cuts added by CPLEX and under cpx spi, we add path cover and path pack inequalities addition to these cuts. We observe that with the addition of path cover and pack inequalities, the gap improvement increases from 86% to 95%. The number of branch and bound nodes explored is reduced about 900 times. Moreover, with path cover and path pack inequalities the average elapsed time is reduced to almost half and the total number of unsolved instances reduces from 13 to 6 out of 180 instances. Tables 1, 2, 3 and 4 show that submodular path inequalities are quite effective in tackling lot-sizing problems with finite arc capacities. When added to the LP relaxation, they improve the optimality gap by 95% and the number of branch and bound nodes explored decreases by a factor of 1000. In conclusion, our computational experiments indicate that the use of path cover and path pack inequalities is beneficial in improving the performance of the branch-and-cut algorithms. Table 2. Effect of path cover (cov) and path pack (pac) inequalities when used separately and together (spi). gap imp nodes cuts time c init gap cov 2 100 5 10 14.8 44.3 58.3 62% 18% 75% 37% 77% 39% 96% 97% 93% 273 6258 7 759 319 12366 9 357 213 29400 63 290 13 1151 0.3 0.6 25 435 0.1 1.0 11 386 0.1 2.1 0.2 0.1 0.1 2 200 5 10 50 14.5 49.8 66.3 73% 34% 92% 67% 38% 100% 61% 48% 97% 148 3268 18 1593 576 11525 3 736 226 8799 14 619 41 2469 0.7 0.3 36 1022 0.4 0.9 32 739 0.2 0.7 0.6 0.1 0.1 2 500 5 10 19.1 54.4 73.0 57% 57% 56% 75% 59% 65% 92% 635 1825 19 1587 316 2577 1.6 0.3 99% 348 363 1 902 148 1164 0.3 0.1 97% 8410 5284 11 727 67 698 3.0 0.5 1.0 0.1 0.1 2 1000 5 10 14.6 59.7 76.9 61% 65% 90% 278 258 60 1362 427 2094 0.8 0.3 59% 81% 100% 1063 208 2 1673 364 1792 1.3 0.1 51% 77% 99% 3791 1452 5 1202 155 1032 2.1 0.2 0.9 0.1 0.1 45.5 63% 53% 96% 1357 6751 18 984 136 1297 0.9 0.6 0.3 n f Average: pac spi cov 29 pac spi cov pac spi cov pac spi Table 3. Comparison of path inequalities applied to paths (spi) versus applied to merged paths (mspi). gapimp n 50 100 150 f c nodes init spi mspi spi gap cuts mspi spi mspi time (endgap:unslvd) spi mspi 2 14.8 96% 52% 100 5 44.3 97% 69% 10 58.3 93% 70% 7 9 63 430 1151 553 435 468 386 195 146 160 0.2 0.1 0.1 0.2 0.1 0.1 2 14.5 92% 59% 200 5 49.8 100% 57% 10 66.3 97% 53% 18 3 14 330 2469 1112 1022 615 739 226 176 173 0.6 0.1 0.1 0.2 0.3 0.2 2 19.1 92% 43% 500 5 54.4 99% 48% 10 73.0 97% 48% 19 1 11 2041 2577 705 1164 5659 698 238 214 248 1.0 0.1 0.1 0.7 0.3 1.4 2 14.6 90% 45% 1000 5 59.7 100% 50% 10 76.9 99% 40% 60 2 5 612 2094 2265 1792 9199 1032 301 241 314 0.9 0.1 0.1 0.4 0.7 2.3 2 13.9 95% 65% 39 100 5 42.2 98% 70% 19 10 57.8 94% 59% 230 7073 3114 20897 1337 395277 1298 410 297 346 1.3 0.2 0.4 3.2 4.8 88.2 2 16.1 89% 56% 290 151860 6919 200 5 47.6 99% 55% 7 455192 2355 10 65.7 95% 54% 104 4130780 1872 478 331 399 11.0 0.3 0.5 58.4 126.1 962.3 (1.1:1) 2 17.5 84% 36% 1047 956745 11743 500 5 53.9 99% 41% 4 332041 3874 10 72.9 96% 42% 34 1175647 1495 475 444 474 47.7 0.4 0.3 390.9 115.5 352.5 2 17.9 91% 41% 173 57147 10919 1000 5 58.5 100% 45% 1 284979 3261 10 75.7 97% 36% 88 3158262 2358 570 501 657 21.3 23.0 0.3 92.8 0.5 1047.0 (0.7:1) 2 13.2 94% 64% 336 163242 5159 100 5 44.8 99% 65% 17 3024118 2087 10 56.9 95% 65% 404 7254052 1492 704 431 476 11.3 107.6 0.5 929.6 0.9 2087.3 (0.7:1) 2 14.7 92% 53% 519 2772494 12636 200 5 48.1 99% 55% 15 3802938 2462 10 65.2 95% 50% 330 9377122 2047 744 508 567 27.2 1390.6 (0.1:1) 0.6 1483.0 (1.2:2) 0.9 3585.9 (8.2:5) 2 19.3 86% 33% 7927 7619674 22275 500 5 54.4 100% 45% 7 7873043 4927 10 72.3 97% 41% 250 10219548 2678 792 1087.3 3165.6 (4.0:4) 641 0.8 2813.6 (4.3:3) 713 1.2 3422.8 (11.0:5) 2 19.6 88% 34% 2824 7316675 33729 1000 5 57.5 100% 39% 2 9661586 6710 10 75.8 96% 37% 99 9910056 3981 724 804.8 3260.3 (2.5:3) 709 0.8 3578.9 (9.7:5) 829 1.2 3412.3 (15.2:5) Average: 45.2 95% 50% 416 2504012 4619 440 30 56.3 903.0 (1.6:36) Table 4. Effectiveness of the path inequalities when used together with CPLEX’s network cuts. gapimp n f c nodes init cpx spi cpx cpx spi gap time (endgap:unslvd) cpx cpx spi cpx 2 100 5 10 13.9 42.2 57.8 96% 99% 99% 85% 97% 93% 35 5 10 1715 75 2970 1.0 0.2 0.3 0.5 0.1 0.6 2 200 5 10 16.1 47.6 65.7 90% 99% 97% 79% 95% 89% 288 7 61 9039 52 3186 6.6 0.3 0.4 2.1 0.1 0.7 2 500 5 10 17.5 53.9 72.9 85% 99% 98% 63% 94% 89% 1232 6 11 455068 92 4621 57.3 0.4 0.4 95.2 0.1 0.9 2 1000 5 10 17.9 58.5 75.7 91% 100% 97% 76% 93% 85% 173 1 117 18109 156 5297 22.2 0.3 0.7 3.6 0.1 1.0 2 100 5 10 13.2 44.8 56.9 94% 100% 99% 86% 97% 92% 365 5 16 60956 119 15929 9.7 0.4 0.5 19.0 0.1 3.9 2 200 5 10 14.7 48.1 65.2 92% 99% 97% 80% 96% 91% 954 11 181 216436 284 3992 44.9 0.5 0.9 66.7 0.2 1.2 2 500 5 10 19.3 54.4 72.3 86% 100% 97% 69% 94% 88% 7647 4943603 1049.9 5 5434 0.8 141 141211 1.4 2 1000 5 10 19.6 57.5 75.8 88% 100% 96% 71% 90% 89% 3051 2788993 917.4 (0.2:1) 3 4322 0.8 196 10588 2.5 619.4 (0.4:1) 1.2 2.8 2 100 5 10 14.1 42.7 57.5 94% 100% 99% 82% 97% 93% 1623 8 26 384.0 0.1 13.8 2 200 5 10 16.3 48.0 65.0 89% 99% 98% 78% 95% 90% 4279 5634851 259.9 13 1310 0.9 128 163145 1.2 2 500 5 10 16.3 54.5 72.0 88% 99% 96% 72% 93% 90% 8083 6805861 1226.3 (0.3:1) 2137.6 (0.7:3) 7 6606 1.6 2.2 376 900152 3.2 302.4 2 1000 5 10 18.0 57.9 75.6 82% 100% 96% 63% 94% 84% 13906 9894589 3000.5 (1.2:4) 2835.9 (3.0:5) 4 1977 3.4 0.8 704 6127929 15.0 1785.0 (1.8:2) Average: 45.0 95% 86% 1213 1087194 185.1 (0.0:6) 320.2 (0.2:13) 100 150 200 31 864841 213 45263 32.2 0.5 0.7 1215.1 (0.2:1) 1.6 35.9 1940.4 (0.1:1) 0.5 52.3 Acknowledgements A. Atamtürk and Birce Tezel are supported, in part, the National Science Foundation grant #0970180 and by grant FA9550-10-1-0168 from the Office of the Assistant Secretary of Defense for Research and Engineering. Simge Küçükyavuz is supported, in part, by the National Science Foundation grant #1055668. The authors are also thankful to the Associate Editor and anonymous referees for their constructive feedback that improved the paper substantially. References Atamtürk, A. (2001). Flow pack facets of the single node fixed–charge flow polytope. Operations Research Letters, 29:107–114. Atamtürk, A. and Küçükyavuz, S. (2005). Lot sizing with inventory bounds and fixed costs. Operations Research, 53:711 – 730. Atamtürk, A. and Muñoz, J. C. (2004). A study of the lot–sizing polytope. Mathematical Programming, 99:43–65. Gade, D. and Küçükyavuz, S. (2011). A note on lot-sizing with fixed charges on stocks: the convex hull. Discrete Optimization, 8:385–392. Gu, Z., Nemhauser, G. L., and Savelsbergh, M. W. P. (1999). Lifted flow cover inequalities for mixed 0–1 integer programs. Mathematical Programming, 85:439–467. King, V., Rao, S., and Tarjan, R. (1994). A faster deterministic maximum flow algorithm. Journal of Algorithms, 17(3):447–474. Küçükyavuz, S. and Pochet, Y. (2009). Uncapacitated lot sizing with backlogging: the convex hull. Mathematical Programming, 118(1):151–175. Ortega, F. and Wolsey, L. A. (2003). A branch-and-cut algorithm for the single-commodity, uncapacitated, fixed-charge network flow problem. Networks, 41:143–158. Padberg, M. W., van Roy, T. J., and Wolsey, L. A. (1985). Valid linear inequalities for fixed charge problems. Operations Research, 33:842–861. Pochet, Y. and Wolsey, L. A. (1988). Lot-size models with backlogging: Strong reformulations and cutting planes. Mathematical Programming, 40(1-3):317–335. Pochet, Y. and Wolsey, L. A. (1994). Polyhedra for lot-sizing with Wagner–Whitin costs. Mathematical Programming, 67(1-3):297–323. Rardin, R. L. and Wolsey, L. A. (1993). Valid inequalities and projecting the multicommodity extended formulation for uncapacitated fixed charge network flow problems. European Journal of Operational Research, 71:95–109. Stallaert, J. I. A. (1997). The complementary class of generalized flow cover inequalities. Discrete Applied Mathematics, 77:73–80. van Roy, T. J. and Wolsey, L. A. (1985). Valid inequalities and separation for uncapacitated fixed charge networks. Operations Research Letters, 4:105–112. van Roy, T. J. and Wolsey, L. A. (1986). Valid inequalities for mixed 0–1 programs. Discrete Applied Mathematics, 14:199–213. Van Vyve, M. (2013). Fixed-charge transportation on a path: optimization, LP formulations and separation. Mathematical Programming, 142:371–395. 32 Van Vyve, M. and Ortega, F. (2004). Lot-sizing with fixed charges on stocks: the convex hull. Discrete Optimization, 1:189–203. Wolsey, L. A. (1989). Submodularity and valid inequalities in capacitated fixed charge networks. Operations Research Letters, 8:119–124. Wolsey, L. A. and Nemhauser, G. L. (1999). Integer and Combinatorial Optimization. Wiley Series in Discrete Mathematics and Optimization. Wiley. Appendix A. Equivalency of (F2) to the maximum flow problem In Section 3, we showed the maximum flow equivalency of v(S + , L− ) under the assumption that dj ≥ 0 for all j ∈ N . In this section, we generalize the equivalency for the paths where dj < 0 for some j ∈ N . Observation 2. If dj < 0 for some j ∈ N , one can represent the supply amount as a dummy arc incoming to node j (i.e., added to Ej+ ) with a fixed flow and capacity of −dj and set the modified demand of node j to be dj = 0. Given the node set N with at least one supply node, let T (N ) be the transformed path using Observation 2. Transformation T ensures that the dummy supply arcs are always open. As a result, they are always in the set S + . We refer to the additional constraints that fix the flow to the supply value on dummy supply arcs as fixed-flow constraints. Notice that, v(S + , L− ) computed for T (N ) does not take fixed-flow constraints into account. In the next proposition, for a path structure, we show that there exists at least one optimal solution to (F2) such that the fixed-flow constraints are satisfied. Proposition 18. Suppose that dj < 0 for some j ∈ N . If (F2) for the node set N is feasible, then it has at least one optimal solution that satisfies the fixed-flow constraints. Proof. We need to show that v(S + , L− ) has an optimal solution where the flow at the dummy supply arcs is equal to the supply values. The transformation T makes Proposition 1 applicable to the modified path T (N ). Let Y be the set of optimal solutions of (F2). Then, there exists a solution (y∗ , i∗ , r∗ ) ∈ Y where yt∗ = 0 for t ∈ E − \ (S − ∪ L− ). Let p ∈ Sj+ represent the index of the dummy supply arc with cp = −dj . If yp∗ < cp , then satisfying the fixed-flow constraints require pushing flow through the arcs in E − \ L− . We use Algorithm 2 to construct an optimal solution with yp∗ = cp . Note that each arc in Ek− \ L− k for k ∈ N appear in (F2) with the same coefficients, therefore we merge these outgoing arcs into one P in Algorithm 2. We represent the merged flow and capacity by Ȳk− = t∈E − \(S − ∪L− ) yt∗ k k k  and C̄k = c Ek− \ (Sk− ∪ L− k ) for k ∈ N .  Proposition 18 shows that, under the presence of supply nodes, transformation T both captures the graph’s structure and does not affect (F1)’s validity. As a result, Propositions 1 and 2 become relevant to the transformed path and submodular path inequalities (14) and (17) are also valid for paths where dj < 0 for some j ∈ N . 33 Algorithm 2 J : Set of supply nodes in N where the nodes are sorted with respect to their order in N . (y∗ , i∗ , r∗ ) ∈ Y: yt∗ = 0 for all t ∈ E − . for q ∈ J do Let p be the dummy supply arc in Sq+ ∆ = cp − yp∗ for j = q to n do Ȳj− = Ȳj− + min{C̄j − Ȳj− , ∆} ∆ = ∆ − min{C̄j − Ȳj− , ∆} i∗j = i∗j + ∆ if i∗j > uj then ∆ = i∗j − uj i∗j = uj Let k := j break inner loop end if end for if ∆ > 0 then for j = k to 1 do Ȳj− = Ȳj− + min{C̄j − Ȳj− , ∆} ∆ = ∆ − min{C̄j − Ȳj− , ∆} rj∗ = rj∗ + ∆ if rj∗ > bj then ∆ = rj∗ − bj break inner loop end if end for end if if ∆ > 0 then (F2) is infeasible for the node set N . end if end for Appendix B. Proofs + − B.1. Proof of Lemma 10. Recall that C = S + ∪ L− and let C1 = SN 1 ∪ LN 1 and + − C2 = SN 2 ∪ LN 2 . In (13), we showed that the value of the minimum cut is v(C) = mi = min{αiu + βiu − c(Si+ ), αid + βid − di − c(Si− )} for all i ∈ N . For node set N1 and the arc set C1 , the value of the minimum cut is u d v1 (C1 ) = min{αj−1 + bj−1 , αj−1 }. {u,d} This is because of three observations: (1) the values αi for i ∈ [1, j − 2] are the same + for the node sets N1 and N , (2) for the arc set C1 the set Sj−1 now includes the backward path arc (j, j − 1) and (3) node j − 1 is the last node of the first path. Similarly, for node 34 set N2 and the arc set C2 , the value of the minimum cut is v2 (C2 ) = min{βju + uj−1 , βjd }. {u,d} For nodes N2 and the arc set C2 , (1) the values βi for i ∈ [j + 1, n] are the same for the node sets N2 and N , (2) for the arc set C2 the set Sj+ now includes the forward path arc (j − 1, j) and (3) node j is the first node of the second path. d d Now, if αju = αj−1 +uj−1 +c(Sj+ ), then αjd = αj−1 +dj +c(Sj− ) from equations in (7)–(8). d Then, rewriting v(C) = mj and v1 (C1 ) in terms of αj−1 : d v(C) = αj−1 + min{βju + uj−1 , βjd } and d v1 (C1 ) = αj−1 . As a result, the values v1 (C1 ) and v2 (C2 ) summed gives the value v(C) under the assumption for the value of αju . + − u d Similarly, if βj−1 = βjd + bj−1 + c(Sj−1 ), then βj−1 = βjd + dj−1 + c(Sj−1 ) from equations in (9)–(10). Then, rewriting v(C) = mj−1 and v2 (C2 ) in terms of βjd : u d v(C) = βjd + min{αj−1 + bj−1 , αj−1 } and v2 (C2 ) = βjd . As a result, the values v1 (C1 ) and v2 (C2 ) summed gives the value v(C) under the assumption u for the value of βj−1 . B.2. Proof of Lemma 11. The proof follows closely to that of Lemma 10. Let C = + − + − S + ∪ L− , C1 = SN 1 ∪ LN 1 and C2 = SN 2 ∪ LN 2 . For node set N1 and the arc set C1 , the value of the minimum cut is u d v1 (C1 ) = min{αj−1 , αj−1 + uj−1 }, − − where uj−1 is added because c(SN 1 ) = c(S1j−1 ) + uj−1 . Similarly, for node set N2 and the arc set C2 , the value of the minimum cut is v2 (C2 ) = min{βju , βjd + bj−1 }, − − where bj−1 is added because c(SN 2 ) = c(Sjn ) + bj−1 . u u Now, if αjd = αj−1 + bj−1 + dj−1 + c(Sj− ), then αju = αj−1 + c(Sj+ ) from equations in u (7)–(8). Then, rewriting v(C) = mj and v1 (C1 ) in terms of αj−1 : u v(C) = αj−1 + min{βju , βjd + bj−1 } and u v1 (C1 ) = αj−1 . As a result, the values v1 (C1 ) and v2 (C2 ) summed gives the value v(C) under the assumption for the value of αjd . − + d u Similarly, if βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ), then βj−1 = βju + c(Sj−1 ) from equations u in (9)–(10). Then, rewriting v(C) = mj−1 and v2 (C2 ) in terms of βj : u d v(C) = βju + min{αj−1 , αj−1 + uj−1 } 35 and v2 (C2 ) = βju . As a result, the values v1 (C1 ) and v2 (C2 ) summed gives the value v(C) under the assumption d for the value of βj−1 . B.3. Proof of Lemma 12. The proof follows closely to that of Lemmas 10 and 11. Let − + − + C = S + ∪ L− , C1 = SN 1 ∪ LN 1 and C2 = SN 2 ∪ LN 2 . For node set N1 and the arc set C1 , the value of the minimum cut is u d v1 (C1 ) = min{αj−1 , αj−1 } and for node set N2 and the arc set C2 , the value of the minimum cut is v2 (C2 ) = min{βju + uj−1 , βjd + bj−1 }. − d d Now, if αju = αj−1 + uj−1 + c(Sj+ ) and βj−1 = βju + uj−1 + dj−1 + c(Sj−1 ), then αjd = − + u u + dj + c(Sj ) and βj−1 = βj + c(Sj ). Then, rewriting v(C) = mj , v1 (C1 ) and v2 (C2 ): d αj−1 d d v(C) = αj−1 + min{uj−1 + βju , βjd } = αj−1 + uj−1 + βju , d v1 (C1 ) = αj−1 and v2 (C2 ) = βju + uj−1 . As a result, the values v1 (C1 ) and v2 (C2 ) summed gives the value v(C) under the assumption d for the values of αju and βj−1 . B.4. Proof of Lemma 13. The proof follows closely to that of Lemmas 10 and 11. Let + − + − C = S + ∪ L− , C1 = SN 1 ∪ LN 1 and C2 = SN 2 ∪ LN 2 . For node set N1 and the arc set C1 , the value of the minimum cut is u d v1 (C1 ) = min{αj−1 + bj−1 , αj−1 + uj−1 } and for node set N2 and the arc set C2 , the value of the minimum cut is v2 (C2 ) = min{βju , βjd }. + u u Now, if αjd = αj−1 + bj−1 + dj + c(Sj− ) and βj−1 = βjd + bj−1 + c(Sj−1 ), then αju = u d αj−1 + c(Sj+ ) and βj−1 = βjd + dj + c(Sj− ). Then, rewriting v(C) = mj , v1 (C1 ) and v2 (C2 ): u u v(C) = αj−1 + min{βju , βjd + bj−1 } = αj−1 + βjd + bj−1 , u v1 (C1 ) = αj−1 + bj−1 and v2 (C2 ) = βjd . As a result, the values v1 (C1 ) and v2 (C2 ) summed gives the value v(C) under the assumption u for the values of αjd and βj−1 . 36
8cs.DS
arXiv:1609.06288v2 [math.GR] 15 Nov 2017 NO POSITIVE CONE IN A FREE PRODUCT IS REGULAR SUSAN HERMILLER AND ZORAN ŠUNIĆ Abstract. We show that there exists no left order on the free product of two nontrivial, finitely generated, left-orderable groups such that the corresponding positive cone is represented by a regular language. Since there are orders on free groups of rank at least two with positive cone languages that are context-free (in fact, 1-counter languages), our result provides a bound on the language complexity of positive cones in free products that is the best possible within the Chomsky hierarchy. It also provides a strengthening of a result by Cristóbal Rivas which states that the positive cone in a free product of nontrivial, finitely generated, left-orderable groups cannot be finitely generated as a semigroup. As another illustration of our method, we show that the language of all geodesics (with respect to the natural generating set) that represent positive elements in a graph product of groups defined by a graph of diameter at least 3 cannot be regular. 1. Introduction 1.1. Basic definitions, background, and notation. A total left order on a group G is a total order ≤ on G compatible with left multiplication; that is, g ≤ g′ implies hg ≤ hg′ for all g, g′ , h ∈ G. Throughout this paper we assume that all group orders are total left orders (and we will not use the adjectives total or left again). A group G is called orderable if G admits an order. Saying that (G, ≤) is an ordered group means that we consider the group G along with some specific order ≤ on G. The positive cone of an ordered group (G, ≤) is the set G+ = {g ∈ G | e < g} of positive elements with respect to ≤. The positive cone is a subsemigroup of G and G is partitioned as G = G+ ⊔ {e} ⊔ G− , where e is the identity element of G and G− = (G+ )−1 is the set of negative elements in G. Let X be a finite set of symbols and φ : X ∗ → G a surjective homomorphism from the free monoid X ∗ to the group G. We sometimes suppress φ from the notation and denote φ(w) by w and φ(L) by L. A language over X is a subset L ⊂ X ∗ . Let λ denote the empty word. The class of regular languages over X is the closure of the finite languages under 2010 Mathematics Subject Classification. 20E06, 06F15, 68Q45. Key words and phrases. left-orderable groups, free products, graph products, regular languages. 1 2 SUSAN HERMILLER AND ZORAN ŠUNIĆ the operations of finite union, finite intersection, complementation, concatenation, and Kleene star. Regular languages are the languages accepted by finite state automata. The reverse of any regular language is regular, and for any finite set Y and monoid homomorphism α : X ∗ → Y ∗ the image α(L) of any regular language L over X is a regular language over Y . We say that a language L ⊆ X ∗ represents a subset S of G if φ(L) = S. A subset S of G represented by a regular language is often called a rational subset of G. A language over X that represents the positive cone G+ of an ordered group (G, ≤) is called a positive cone language for (G, ≤) over X (and with respect to φ). 1.2. Regular positive cones. We are interested in the question of which finitely generated groups admit a representation of a positive cone by a regular language. Several examples of such groups have been shown to exist. As a first example, note that there are orders on Z2 that have a regular positive cone language. Indeed, let X = hx, y, x−1 , y −1 i. The regular language {xn y m | n ≥ 1, m ∈ Z} ∪ {y m | m ≥ 1} represents the positive cone of the “lexicographic order” on Z2 = hx, yi. Both the Dehornoy order [Deh94] and the Dubrovina-Dubrovin order [DD01] on the braid group Bn , with n ≥ 3, admit positive cone languages that are regular. Moreover, the positive cone for the latter is finitely generated as a semigroup. Rourke and Wiest [RW00] provided regular positive cone languages for certain orders on mapping class groups of compact surfaces with a finite number of punctures and non-empty boundary. On the negative side, Calegari [Cal03] showed that no hyperbolic 3manifold group admits a positive cone language that is simultaneously regular and geodesic. His argument applies only to languages consisting of (quasi-)geodesics because it uses the fact that hyperbolic 3-manifold groups contain quasi-geodesically embedded copies of the free group of rank 2. Note that the issue is subtle, since groups admitting regular positive cone languages may contain free subgroups. Indeed, braid groups along with the Dehornoy orders and the Dubrovina-Dubrovin orders provide such examples. 1.3. The main result. It is well known that the free product of two orderable groups is orderable [Vin49]. Rivas [Riv12] showed that the space of all orders on a free product G = A ∗ B, where A and B are nontrivial, finitely generated, orderable groups, is uncountable and has the structure of a Cantor set. Navas [Nav10] pointed out that, in the case of the free group Fk of finite rank k, k ≥ 2, the same conclusion follows from the earlier work of McCleary [McC85], and also provided another proof in this case. In our NO POSITIVE CONE IN A FREE PRODUCT IS REGULAR 3 main result we show that, despite such an abundance of available orders, a positive cone language in a free product is never regular. Theorem 1. Let A and B be two nontrivial, finitely generated, orderable groups. There exists no order on G = A ∗ B such that its positive cone is represented by a regular language (for any finite alphabet X, any homomorphism φ : X ∗ → G, and any choice of a language representing the positive cone). The orders defined on free groups Fk in [Šun13a, Šun13b] have context-free positive cone languages. More specifically, the sets of freely reduced words representing the elements in their positive cones are 1-counter languages (the stack of the push-down automaton uses a one-letter alphabet). In light of Theorem 1, it follows that the orders on Fk from [Šun13a, Šun13b] are, at least from the language theoretic point of view, among the simplest ones possible in the context of free products. 1.4. Other results and remarks. The main result immediately provides the following corollary. Corollary 2. Let A and B be two nontrivial, finitely generated, orderable groups. There exists no order on G = A ∗ B such that its positive cone is finitely generated as a semigroup. This corollary was already established by Rivas [Riv12] by using a very different approach. In particular, Linnell [Lin06], [Nav10, Proposition 1.8] observed that if a positive cone on a finitely generated group G is a finitely generated semigroup, then the corresponding order is isolated in the space of all orders. Since the space of orders on G = A ∗ B has no isolated points [Riv12], no order on G has a finitely generated positive cone. This is an excellent approach, but it seems that, in general, it is not easy to establish that there are no isolated points in the space of orders of some group (including the case of the free group). Also, the “no isolated points” approach is not helpful in establishing stronger results in the spirit of Theorem 1. Namely, the space of orders on Z2 is a Cantor set, but, as we already mentioned, there are orders on Z2 that have a regular positive cone language. Another approach to Corollary 2 is provided by Kielak [Kie15] who proved that a finitely generated group with infinitely many ends cannot be a fraction group of a proper, finitely generated subsemigroup. The proof of Theorem 1 is based on the following lemma. Lemma 3. Let (G, ≤) be a finitely generated ordered group and let L ⊆ X ∗ be a language representing the positive cone G+ . Denote by P ref (L) the language of prefixes of the words in L. If there exists a set of positive elements T in G such that T is unbounded (with respect to ≤) from above and T −1 ⊆ P ref (L), 4 SUSAN HERMILLER AND ZORAN ŠUNIĆ then L is not regular. In our next result we consider graph products of groups which contain subgroups that are free products of pairs of vertex subgroups. (In the following we use the convention that edges of the defining graph represent commuting relations). The graph product construction preserves many properties of groups; for example, Chiswell [Chi12] has shown that graph products preserve orderability, and Loeffler, Meier and Worthington [LMW02] have shown that the graph product preserves regularity of the set of all geodesics, using the union of the vertex group generating sets. One important family of graph products is the class of right-angled Artin groups, which are graph products of cyclic groups of infinite order; in their case both the language of geodesics and the language of all geodesics representing elements in the positive cone are regular for each vertex group. In order to illustrate that the applicability of Lemma 3 goes beyond free products, we modify the proof of Theorem 1 to show that regularity of the geodesic positive cone languages of orders on the vertex groups cannot be preserved in the graph product for graphs of diameter at least 3, when using the same union of the vertex group generating sets. Theorem 4. Let Γ be a finite, simple graph of diameter at least 3 and V its vertex set. For each v ∈ V let Gv be a nontrivial, orderable group with generating set Yv for a finite set Yv , and let X = ⊔v∈V (Yv ⊔ Yv−1 ). Let GΓ be the associated graph product group, and suppose that φ : X ∗ → GΓ has the property that φ(x−1 ) = φ(x)−1 for all x ∈ X. Let Geo be the language of all geodesics for GΓ with respect to X. There exists no order ≤ on GΓ such that the positive cone language Geo+ = { w ∈ Geo | w represents a positive element in GΓ } is regular. Remark 5. Let X = Y ⊔Y −1 , for some finite set Y , and let φ : X ∗ → G be a homomorphism such that φ(x−1 ) = φ(x)−1 for all x ∈ X. Let L+ ⊆ X ∗ be a regular language representing the positive cone in the ordered group (G, ≤). Since the homomorphic image of the reverse of a regular language is regular, the language (L+ )−1 of formal inverses of the words in L+ is also regular, and hence the union L = L+ ⊔ {λ} ⊔ (L+ )−1 is a regular language representing the entire group G (i.e., φ(L) = G). Thus, every regular language L+ over X representing the positive cone G+ is induced from some regular language L representing G, by L+ = { w ∈ L | w represents a positive element in G }. Using this viewpoint, Theorem 4 implies, in particular, that no regular positive cone language can be induced from the regular language of all geodesics in a right-angled Artin group on a graph Γ, when the diameter of the graph Γ is at least 3. NO POSITIVE CONE IN A FREE PRODUCT IS REGULAR 5 2. Proofs We first observe that the question of representability of a subset of a group by a regular language is independent of the choice of the alphabet X and the homomorphism φ : X ∗ → G. Lemma 6. Let X and Y be two finite alphabets and φX : X ∗ → G and φY : Y ∗ → G be two surjective homomorphisms to the finitely generated group G. Let S be a subset of G represented by a regular language L ⊆ X ∗ . Then there exists a regular language L′ ⊆ Y ∗ representing S. Proof. For every letter x ∈ X, choose a word wx over Y such that φY (wx ) = φX (x). Define a monoid homomorphism α : X ∗ → Y ∗ by α(x) = wx for all x ∈ X. The homomorphic image α(L) is a regular language over Y representing S.  Because of the independence on the alphabet, we can always work with a group alphabet X = Y ⊔ Y −1 , for some finite set Y , and φ : X ∗ → G with the property that φ(x−1 ) = φ(x)−1 for all x ∈ X. Moreover, by Benois’ Lemma [Ben69] (see [BS10] for an exposition), we can always make a further simplification and assume that a regular language representing a set of elements of G lies within the set R(X) of freely reduced group words over X. Lemma 7 (Benois’ Lemma). Let X = Y ⊔Y −1 be a group alphabet, for some finite set Y , and let φ : X ∗ → G be a surjective homomorphism to a group G such that φ(x−1 ) = φ(x)−1 for all x ∈ X. Let L be a regular language over X that represents S in G. Then there exists a regular language L′ ⊆ R(X) representing S. Moreover, one such language is the language of freely reduced words of the words in L. Proof of Lemma 3. By way of contradiction, assume that L is regular and let A be an automaton on k states accepting L. Consider the ball B of radius k − 1 in G with respect to X. Since T is unbounded from above, there exists an element t ∈ T that is greater (with respect to the order ≤) than every element in B. Then t−1 ∈ T −1 , and so there exists a word w in P ref (L) representing t−1 . Now w ∈ P ref (L) implies that there exists a word u such that wu ∈ L. Since the word wu is accepted by the automaton A on k states, there must be a word v of length at most k − 1, such that the word wv is also accepted by the automaton. Therefore e < wv = t−1 v, which implies that t < v. This is impossible since v is an element in B.  Our proof of Theorem 1 also relies on the structure of words representing elements in a free product. Let A and B be nontrivial groups. For each g ∈ A ∗ B, the reduced factorization of g is the unique expression of g in the form g = a1 b1 · · · am bm where a1 ∈ A, ai ∈ A \ {e} for all i ≥ 2, bi ∈ B \ {e} for all i ≤ m − 1, and bm ∈ B. The nontrivial elements ai , bi in this factorization are the syllables of g. Similarly for each word w over a 6 SUSAN HERMILLER AND ZORAN ŠUNIĆ partitioned alphabet ZA ⊔ ZB , the ZA /ZB factorization of w is the unique ∗ , u ∈ Z ∗ \ {λ} for expression of the form w = u1 v1 · · · un vn where u1 ∈ ZA i A ∗ ∗ all i ≥ 2, vi ∈ ZB \ {λ} for all i ≤ n − 1, and vn ∈ ZB . Lemma 8. Let A and B be nontrivial groups with generating sets YA and YB , respectively, where YA and YB are finite sets, and let G = A ∗ B and X = YA ⊔ YB ⊔ YA−1 ⊔ YB−1 . Suppose that φ : X ∗ → G has the property that φ(x−1 ) = φ(x)−1 for all x ∈ X. If w ∈ X ∗ and φ(w) has a reduced factorization φ(w) = a1 b1 · · · am bm , then for all i the elements of A ∗ B with reduced factorization a1 b1 · · · ai b̂i with b̂i ∈ {bi , e} are represented by prefixes of the word w. Proof. We prove this for the element g ′ = a1 b1 · · · ai bi with bi 6= e; the case that b̂i = e is nearly identical. Let ZA = YA ⊔ YA−1 and ZB = YB ⊔ YB−1 , and write the ZA /ZB factorization of w as w = u1 v1 · · · un vn . Another reduced factorization of φ(w) can be obtained from the product (u1 )(v1 ) · · · (un )(vn ) by finitely many applications of the following operation: Remove a factor uj or vj that is the trivial element, and replace the resulting product of two contiguous factors (uj )(uk ) from the same factor group by a single factor (uj uk ) (and similarly, replace (vj )(vk ) by (vj vk )). Since φ(w) only admits one reduced factorization, the result of applying this operation until no further reduction can occur is the product a1 b1 · · · am bm . If m = n, then no instance of this operation can occur, and uj = aj and vj = bj for all j. In this case the prefix w′ = u1 v1 · · · ui vi of w satisfies φ(w′ ) = g ′ . Otherwise we have m > n, and for at least one index k a trivial element uk or vk is removed and the resulting contiguous factors from the same factor group are combined. The syllables of φ(w) obtained from this process can be written as aj = (uj1 · · · ujℓj ) and bj = (vj1′ · · · vj ′ ′ ) where ℓj , ℓ′j ≥ 1 and ℓ j j1 < · · · < jℓj ≤ j1′ < · · · < jℓ′ ′ . In this case the prefix w′ = u1 v1 · · · ui′ ′ vi′ ′ j of w satisfies φ(w′ ) = g ′ . ℓ i ℓ i  Proof of Theorem 1. By way of contradiction, assume that G admits an order ≤ with a regular positive cone language L. In light of Lemma 6, we may assume that L ⊆ X ∗ , where X = YA ⊔ YB ⊔ YA−1 ⊔ YB−1 , YA and YB are finite sets, YA generates A and YB generates B, and φ : X ∗ → G has the property that φ(x−1 ) = φ(x)−1 for all x ∈ X. In light of Lemma 3 it is sufficient to prove that G ⊆ P ref (L). Let g be an element of G. Since G is the free product of two nontrivial groups, there exists a letter x ∈ X such that the syllable length of gx is strictly greater than the syllable length of g. The elements gxg −1 and gx−1 g−1 form a pair of nontrivial, mutually inverse elements in G. Therefore, at least one of them is positive. Without loss of generality, assume that gxg −1 is positive. There exists a word w ∈ L representing gxg −1 . By our NO POSITIVE CONE IN A FREE PRODUCT IS REGULAR 7 choice of x, the reduced factorization of gxg −1 is the product (i.e., concatenation) of the reduced factorization of g, x, and the reduced factorization of g −1 , and so Lemma 8 shows that there exists a prefix w′ of w that represents g. Since w′ ∈ P ref (L) and w′ = g, we see that g ∈ P ref (L). Thus, G ⊆ P ref (L).  Proof of Theorem 4. Let t and u be vertices of Γ at distance at least 3 in Γ. Note that this distance constraint implies that there is no vertex v of Γ that is adjacent to both t and u. Lemma 3 shows that it is sufficient to prove that Geo ⊆ P ref (Geo+ ). For each vertex v of the graph Γ, let Geov denote the set of geodesic words in the vertex group Gv over the generators Yv±1 , and let Zv be the union of the sets Yv′ over all vertices v ′ adjacent to v in Γ. Define a monoid homomorphism πv : X ∗ → (Yv±1 ∪ {$})∗ , where $ denotes a letter not in X, by defining  ±  a if a ∈ Yv πv (a) := $ if a ∈ X \ (Yv± ∪ Zv± )   λ if a ∈ Zv± . That is, the map πv records all of the occurrences of generators of Gv , as well as all of the occurrences (replaced by $) of generators of vertex groups that do not commute with Gv and hence do not allow letters from Gv on either side to interact. A word w over X is a geodesic for GΓ with respect to X if and only if for all v ∈ V , πv (w) ∈ Geov ($Geov )∗ (see, for example, [CH14, Proposition 3.3]). Now let w be any word in Geo. If πt (w) ends with a letter a in Yt±1 , then we can write w = w1 aw2 such that no letter of Yu±1 lies in w2 , and so the word πu (w) = πu (w1 )$πu (w2 ) ends with a nonempty string in $∗ . Similarly if πu (w) ends with a letter in Yu±1 then πt (w) ends with $. By swapping the roles of t and u if necessary, we may assume that πt (w) either ends with $ or is the empty word. Let x ∈ Yt ∩ Geot and y ∈ Yu ∩ Geou (that is, neither x nor y is e), and let w̃ ∈ X ∗ be the word w̃ = wxyx−1 w−1 . Then πv (w̃) = πv (w)πv (xyx−1 )πv (w)−1 where πv (w)−1 is the formal inverse of πv (w) in (Yv±1 ∪ {$})∗ , with $−1 = $, and satisfies πv (w)−1 ∈ Geov ($Geov )∗ . If v ∈ / {t, u}, then since v cannot be adjacent to both t and u we have πv (xyx−1 ) = $i for some i ∈ {1, 2, 3}. For the case that v = t, the word πt (w) ends with $, the word πt (w)−1 begins with $, and πt (xyx−1 ) = x$x−1 ∈ Geot $Geot . And in the case that v = u we have πu (xyx−1 ) = $y$ ∈ $Geou $. Hence for all vertices v of Γ, the image πv (w̃) lies in Geov ($Geov )∗ , and so w̃ = wxyx−1 w−1 ∈ Geo. By symmetry, the word wxy −1 x−1 w−1 also is in Geo. Since wxyx−1 w−1 and wxy −1 x−1 w−1 represent a pair of nontrivial, mutually inverse elements, one of them represents a positive element, which shows that w ∈ P ref (Geo+ ). Thus, Geo ⊆ P ref (Geo+ ).  8 SUSAN HERMILLER AND ZORAN ŠUNIĆ Acknowledgments. We thank the referee for their knowledgeable input. The work of the first author was supported by the National Science Foundation under Grant No. DMS-1313559. References [Ben69] Michèle Benois. Parties rationnelles du groupe libre. C. R. Acad. Sci. Paris Sér. A-B, 269:A1188–A1190, 1969. [BS10] Laurent Bartholdi and Pedro V. Silva. Rational subsets of groups. Available at https://arxiv.org/abs/1012.1532, 2010. [Cal03] Danny Calegari. Problems in foliations and laminations of 3-manifolds. In Topology and geometry of manifolds (Athens, GA, 2001), volume 71 of Proc. Sympos. Pure Math., pages 297–335. Amer. Math. Soc., Providence, RI, 2003. [Chi12] I. M. Chiswell. Ordering graph products of groups. Internat. J. Algebra Comput., 22(4):1250037, 14, 2012. [CH14] Laura Ciobanu and Susan Hermiller. Conjugacy growth series and languages in groups. Trans. Amer. Math. Soc., 366(5):2803–2825, 2014. [Deh94] Patrick Dehornoy. Braid groups and left distributive operations. Trans. Amer. Math. Soc., 345(1):115–150, 1994. [DD01] T. V. Dubrovina and N. I. Dubrovin. On braid groups. Mat. Sb., 192(5):53–64, 2001. [Kie15] Dawid Kielak. Groups with infinitely many ends are not fraction groups. Groups Geom. Dyn., 9(1):317–323, 2015. [Lin06] Peter Linnell. The topology on the space of left orderings of a group. http://arxiv.org/abs/math/0607470, 2006. [LMW02] Joseph Loeffler, John Meier, and James Worthington. Graph products and Cannon pairs. Internat. J. Algebra Comput., 12(6):747–754, 2002. [McC85] Stephen H. McCleary. Free lattice-ordered groups represented as o-2 transitive l-permutation groups. Trans. Amer. Math. Soc., 290(1):69–79, 1985. [Nav10] Andrés Navas. On the dynamics of (left) orderable groups. Ann. Inst. Fourier (Grenoble), 60(5):1685–1740, 2010. [Riv12] Cristóbal Rivas. Left-orderings on free products of groups. J. Algebra, 350:318– 329, 2012. [RW00] Colin Rourke and Bert Wiest. Order automatic mapping class groups. Pacific J. Math., 194(1):209–227, 2000. [Šun13a] Zoran Šunić. Explicit left orders on free groups extending the lexicographic order on free monoids. C. R. Math. Acad. Sci. Paris, 351(13-14):507–511, 2013. [Šun13b] Zoran Šunić. Orders on free groups induced by oriented words. Available at http://arxiv.org/abs/1309.6070, 2013. [Vin49] A. A. Vinogradov. On the free product of ordered groups. Mat. Sbornik N.S., 25(67):163–168, 1949. Department of Mathematics, University of Nebraska, Lincoln, NE 685880130, USA E-mail address: [email protected] Department of Mathematics, Hofstra University, Hempstead, NY 11549, USA E-mail address: [email protected]
4math.GR
arXiv:1603.03372v1 [cs.SY] 10 Mar 2016 Output Regulation for Systems on Matrix Lie-group Simone de Marco ∗1 , Lorenzo Marconi †1 , Tarek Hamel‡2 and Robert Mahony§3 1 CASY-DEI, Universitá di Bologna, Bologna, 40133, Italy 2 I3S, University of Nice, Sophia Antipolis, France 3 Centre of Excellence in Robotic Vision, Research School of Engineering, Australian National University, Canberra, Australia Abstract This paper deals with the problem of output regulation for systems defined on matrix Lie-Groups. Reference trajectories to be tracked are supposed to be generated by an exosystem, defined on the same Lie-Group of the controlled system, and only partial relative error measurements are supposed to be available. These measurements are assumed to be invariant and associated to a group action on a homogeneous space of the state space. In the spirit of the internal model principle the proposed control structure embeds a copy of the exosystem kinematic. This control problem is motivated by many real applications fields in aerospace, robotics, projective geometry, to name a few, in which systems are defined on matrix Lie-groups and references in the associated homogenous spaces. 1 Introduction Regulating the output of a system in such a way to achieve asymptotic tracking of a reference trajectory is a central topic in control. Among the different approaches proposed so far in the related literature, the output regulation undoubtedly plays a central role. The main peculiarity of output regulation is the fact of considering references to be tracked as belonging to a family of trajectories generated as solutions of an autonomous system (typically referred to as ∗ [email protected][email protected][email protected] § [email protected] exosystem). Tackling this problem in error feedback contexts typically leads to solutions in which the regulator embeds a copy of the exosystem properly updated by means of error measurements. The problem has been intensively studied in the linear context in the mid seventies by Davison, Francis and Wonham [8],[7],[6], and then extended to a quite general nonlinear context by Isidori and Byrnes [10]. In both the linear and nonlinear framework the developed theory has led to the celebrated internal model principle. Assumptions of the aforementioned pioneering works were weakened recently [5], [21] thanks to the observation that the problem of output regulation can be cast as the problem of nonlinear observers design. In fact, the many tools available in literature for designing nonlinear observers, such as high-gain and adaptive nonlinear observers, have been used in the design of internal model-based regulators by thus widening the class of systems that can dealt with. It is worth noting, however, that most of the frameworks considered so far for output regulation deal with systems and exosytems defined on real state space and not much efforts have been done to extend the results of output regulation to systems and exosystems whose states live in more general manifolds. Regarding control and observation of systems defined in more general manifolds, the synthesis of nonlinear observers for invariant systems on Lie-groups has recently played an important role. Many physical systems, such as aerial vehicles, mobile robotic vehicles, robotic manipulators, can be described by geometric models with symmetries. Symmetric structures reflect the fact that the behavior of a symmetric system in one point in the space is independent from the choice of a particular set of configuration coordinates. That is, the laws of motion of a symmetric system are invariant under a change of the configuration space. Preserving such a symmetry is undoubtedly a key point in the observer design. Aghannan and Rouchon [1] first had pointed out the main role of invariance in the observer design for mechanical systems with symmetries. More recent works, based on the aforementioned paper, take advantage of the symmetry of left invariant systems on Lie-groups to define invariant error coordinates in order to build an invariant observer [2], [3], [9], [23]. Lageman et. al [13], [14] exploit the invariance properties of the system in order to have autonomous invariant error dynamics. This paper presents an attempt to extend the idea of internal model-based control to systems defined on more general state spaces, precisely systems defined on matrix Lie-Groups. This control problem is motivated by a wide range of real world applications in which both the controlled system and the exosystem are modeled on matrix Lie-groups. For example, the attitude control problem for satellite [4], in which the system and the exosystem are defined on the group of rigid-body rotations SO(3) or the control [11] of VTOL (Vertical Takeoff and Landing), whose kinematic equations are described by the group of rigid-body translation SE(3). The relevance of this control problem is motivated also by projective geometry problem such as image homographies. This geometry can be modeled by the special linear group SL(3) [16], [18] and in this context the control problem is to continuously warp a reference image of a reference scene in such a way the “controlled” homography matches an image sequence taken by a moving camera. In this paper, we consider reference trajectories to be tracked modelled on the same Lie-group of the controlled system. The exosystem (the reference system) is assumed to be right invariant while the controlled system is assumed to be left invariant. We assume that only partial relative error information are available for measurements. The measurements are assumed to be associated with a left group action on a homogeneous space of the state space. The paper extends preliminary results presented in [17] where the problem of output regulation for systems defined on matrix Lie-Group was tackled under the assumption that the exosystem has constant velocity. Here we remove the limitations of [17] by considering a generic exosystem structure. The main result of the paper is to propose a general structure of the regulator that depends only on relative measures in homogenous space such that the closed-loop system tracks homogenous references generated by the exosystem with a certain domain of attraction. This is achieved by means of a regulator embedding a copy of the exosystem fed by a non-linear function of the measured output along with a stabilising control action. Going further, a control design based on back-stepping techniques, in the special case of systems defined on the special orthogonal group SO(3), is presented for the fully actuated dynamic model. The paper has six sections and it’s organised as follow. Section 2 presents mathematical preliminaries, notation and the problem formulation. Section 3 presents the main result of the paper, a theorem for kinematic systems on matrix Lie-Groups with invariant relative error measurements and its proof. In Section 4 a stability analysis is provided for the specific case of systems defined on the special orthogonal group extending local results of Section 3 to almost global results. In the same section is also proposed a regulator design, based on backstepping techniques, for fully actuated dynamic systems whose kinematic state space is posed on the Lie-Group of orthogonal rotations SO(3). An illustrative example with simulations is shown in Section 5. 2 2.1 Notation and Problem Formulation Notation and basic facts Let G be a general matrix Lie-group. For X ∈ G, the group inverse element is denoted by X −1 , and I is the identity element of G. Let g be the Lie-algebra associated to the Lie-group G. For X ∈ G and U ∈ g, the AdX U denotes the adjoint operator, namely the mapping AdX : G × g → g defined as AdX U := XU X −1 , with X ∈ G, U ∈ g . Let tr (A) note that for any A, B ∈ Rn×n ,  denote the trace operator and ⊤ n×n tr A B defines an inner product on R . For any A ∈ Rn×n and B ∈ Rn×n  tr A⊤ B = vec(A)⊤ vec(B) 2 where vec(A) ∈ Rn is the column vector obtained by the concatenation of columns of the matrix A as follows vec(A) = [a1,1 , ..., an,1 , a1,2 , ..., an,2 , ..., a1,n , ..., an,n ]⊤ . For all A ∈ Rn×n let Pg (A) denote the orthogonal projection of A onto g with respect to the trace inner product. For any U ∈ g and any A ∈ Rn×n , one has   tr U ⊤ A = tr U ⊤ Pg (A) . We denote by mrp (matrix representation) the mapping mrp : Rn → g that maps the vector v ∈ Rn in an element of the algebra. Let vrp (vector representation) denote the inverse of mrp operator, namely vrp(mrp(v)) = vrp(U ) = v, Uv , v U ∈ Rn ∈ g. With Ω ∈ R3 we denote by Ω× the matrix representation of Ω in the Lie-algebra so(3) associated to the Lie-group SO(3). Namely, if Ω = [Ω1 , Ω2 , Ω3 ]⊤ , then   0 −Ω3 Ω2 0 −Ω1  . Ω× =  Ω3 −Ω2 Ω1 0 Let U ∈ g be an n × n matrix and let Uv ∈ Rk with n2 ≥ k. Then, it is straightforward to verify that vec(U ) is a linear combination of the vectorial 2 representation vrp(U ) and there exists a duplication matrix D ∈ Rn ×k such that vec(U ) = DUv . With U ∈ g and V ∈ g elements of the same Lie-algebra one has  tr U ⊤ V = vrp⊤ (U )Qg vrp(V ) where Qg = DT D and D is the duplication matrix defined above. 2.2 Problem formulation In this paper we consider the left invariant kinematic system Ẋ = XU, X(0) ∈ G (1) X ∈ G the state of the system and U ∈ g the control input. We consider a right invariant exosystem on G of the form Ẋd = ◦ Ud Xd ◦ Ud = mrp(Cw) ẇ = Sw (2a) (2b) (2c) where Xd ∈ G and ◦ Ud ∈ g are n × n matrices, vrp(◦ Ud ) ∈ Rκ ,w ∈ Rm , C ∈ Rκ×m , and S ∈ Rm×m with m ≥ κ and S = −S ⊤ . The input w models exogenus signals that represent velocity references to be tracked. Note that U ∈ g in (1) is associated to a body-fixed velocity while ◦ Ud ∈ g of the exosystem (2) is associated with an inertial velocity input. Consider a linear left group action1 of G on Rn , (X, y) 7→ Xy and reference vectors in the form yi = X −1 Xd ẙi , i = 1, 2, ..., ν where ẙi are known constant reference vectors, elements of the homogenous space associated to the Lie-group X. Let E = X −1 Xd (3) denotes the state error of the system as an element of the group G. Let Xr ∈ G be a constant “reference” element of the group. Xr can be applied to ẙi to generate constant reference vectors yir = Xr ẙi . Define an “error vector” ei by ei = yir − yi = yir − Er yir , where Er = EXr−1 . (4) The control problem considered is the design of a feedback control action U as a function of X and yi , in such a way the error ei converges to zero with a certain domain of attraction. To have a physical intuition of those reference and error vectors we address the specific case of SE(2). In particular (see Figure 1), we consider two mobile robots modeled as kinematic systems. The first one, the exosystem, is moving along a certain trajectory while the controlled vehicle measures its relative position with respect to the exosystem. In this context ẙi is a known point in the frame associated to the exosystem mobile robot (for example the coordinate of ẙ1 = [1, 0, 1]⊤) in its own body-fixed frame with an offset given by Xr . The known vector Xd ẙ1 is the inertial representation of ẙ1 , while Er ẙ1 expresses the vector ẙ1 in the frame associated to the actual robot. In this context the control objective is to drive the actual mobile robot with the velocity as control input in order to track the exosystem (the desired one) in such a way the error y1r − Er y1r converges to zero with a certain domain of attraction. The following assumptions will be used in the next sections and they are instrumental for stability analysis. Assumption 1. There are sufficient independent measurements yi with i = 1, . . . , ν such that ν ℓ(E) := ν 1X 2 1X r keki = ky − Er yir k2 2 i=1 2 i=1 i (5) 1 In the paper a linear left group action is considered since it is the natural action for the Lie-groups SO(n) and SE(n). All results presented, however, hold also for a generic left group action. e xo -system system Figure 1: Reference vectors and error vectors in SE(2).The Inertial reference frame is represented with dotted lines. is locally positive definite in Er ∈ G around the identity matrix Er = I. Assumption 2. There exist a compact set Wd ⊂ G × g which is invariant for (2). 3 A General Output Regulation Result The main goal of this section is to present a general structure of the regulator that solves the problem of output regulation formulated above. Theorem 1. Consider system (1) along with exosystem (2). Let the controller be given by the control law U = AdX −1 ∆ − kp ν X P ei (yir − ei )⊤ i=1 ∆ = mrp(Cδ) δ̇ = Sδ + C ⊤ Q vrp(β) ν X P(X −⊤ ei (yir − ei )⊤ X ⊤ ) β = −kI  (6a) (6b) (6c) (6d) i=1 with kp and kI some positive gains. If Assumptions 1 and 2 hold then the compact set S = {(X, δ, (Xd , w)) ∈ G × Rm × Wd : X −1Xd = Xr , δ = w} is locally asymptotically stable for the closed-loop system. Furthermore (X, δ, (Xd , w)) ∈ S ⇒ ei = 0 ∀i = 1, . . . , ν . (7) Note that in the specific case of exosystem constant velocity, namely S = 0, C = I, the control law proposed in (6) is reduced to the PI control action presented in [17] Theorem 3.1. Remark 1. Local properties of Theorem 1 can be extended to global ones exploiting the particular structure of the Lie Group considered. We will show in the next section that, for the specific case of systems posed on the special orthogonal group SO(3), the control law in (6) achieve almost global stability of the error system dynamics. In the special orthogonal group of rotation there is a topological obstruction that deny the possibility to have global results with a smooth control action [22]. To overcame this topological constraint an hybrid control law is needed [20]. Proof. It’s straightforward to verify that condition (7) follows from the definition of the set S and of ei ’s. In what follow we prove that S is locally asymptotically stable. Consider the following candidate Lyapunov function ν L= 1X 1 ⊤ kei k2 + w̃ w̃ 2 i=1 2kI {z } | | {z } L L1 (8) 2 where w̃ = w − δ, which, by assumption, is positive definite around Er = I. Taking the derivatives along the solution of (1) and (2) one obtains L̇ = L˙1 + L˙2 . (9) The first element of the equation above can be written as ν ν X 1X d r r L˙1 = kyi − Er yir k2 = − e⊤ i Ėr yi . 2 i=1 dt i=1 (10) Bearing in mind the definition of Er in (3), it turns out that the time derivative of Er is given by Ėr d (X −1 Xd Xr−1 ) dt = (X −1 (−Ẋ)X −1 )Xd Xr−1 + X −1 Ẋd Xr−1 = −X −1 XU X −1 Xd Xr−1 − X −1◦ Ud Xd Xr−1 = −(U X −1 − X −1◦ Ud XX −1 )Xd Xr−1 = −(U − AdX −1 ◦ Ud )Er = (11) and, substituting E˙r into (10), one obtains L˙1 = = ν P i=1 ν P i=1 ◦ r e⊤ i (U − AdX −1 Ud )Er yi r ˜ e⊤ i (U − AdX −1 ∆ − AdX −1 ∆)Er yi (12) where ˜ = ◦ Ud − ∆ . ∆ Now let’s focus on the i’th term of (12). It follows       ˜ − AdX −1 ∆ Er y r ˜ − AdX −1 ∆ Er y r = tr e⊤ U − AdX −1 ∆ −1 ∆ U − Ad e⊤ X i i i i     ⊤  ⊤ ⊤  ⊤ ˜ ⊤ P X −⊤ Er yir e⊤ = tr (U − AdX −1 ∆) P Er yir e⊤ −tr ∆ X i i where in the last equation above it has been introduced the projection P associated with the Lie-algebra. Recalling L̇1 one has L˙1 = ν X i=1   ⊤  ⊤ tr (U − AdX −1 ∆) P Er yir e⊤ i − ν X i=1     ˜ ⊤ P X −⊤ Er y r e⊤ ⊤ X ⊤ . (13) tr ∆ i i The second term of (9) can be written as 1 1 L˙2 = w̃⊤ w̃˙ = w̃⊤ (Sw − δ̇) kI kI and, substituting δ̇ from (6c) into the above equation, one obtains 1 1 1 ⊤ w̃ (Sw − δ̇) = w̃⊤ S w̃ − w̃⊤ C ⊤ Q vrp(β). kI kI kI Bearing in mind that S is a skew-symmetric matrix and hence w̃⊤ S w̃ = 0, one gets 1 ⊤ 1 w̃ S w̃ − w̃⊤ C ⊤ Q vrp(β) kI kI 1 = − (C w̃)⊤ Q vrp(β) kI 1 ˜ vrp(β) = − vrp⊤ (∆)Q kI 1 ˜ ⊤ β) . = − tr(∆ kI and, recalling the expression of L̇2 and the “innovation term” β (6d), one has L˙2 = ν X i=1     ˜ ⊤ P X −⊤ Er y r e⊤ ⊤ X ⊤ . tr ∆ i i (14) Finally, substituting (13) and (14) into (9) and introducing the expression of U (6a), it yields L̇ = = =   ⊤  ⊤ tr (U − AdX −1 ∆) P Er yir e⊤ i i=1   ⊤   ν    P ⊤ r ⊤ r ⊤ ⊤ tr −kp P Er yi ei P Er yi ei i=1   ν ⊤ 2 P P Er yir e⊤ . −kp i ν P i=1 Since L is positive definite in the error state and since the exosystem state (Xd , w) lies in a compact set, it follows that the whole state is globally bounded and solutions exist for all time. From this, using similar LaSalle arguments to Theorem 3.1 in [17], it is possible to conclude that the set S is locally asymptotically stable. 4 The Special Case of SO(3) In this section we consider the specific case of systems defined in SO(3). By changing the notation used in the previous sections to a more classical notation we denote by R ∈ SO(3) the rotation matrix of the system and by Ω ∈ R3 its angular velocity. In this section we consider the case in which the kynematic system (1) is completed with the dynamics equation of motion Ṙ = RΩ× J Ω̇ = −Ω ∧ JΩ + Γ. (15a) (15b) Thus modelling a rigid body with control input Γ ∈ R3 and J the inertia matrix. In this context the rotation velocity Ω is a state component of the system. In this framework the exosystem is described by ◦ Ṙd Ωd× ẇ = ◦ Ωd× Rd = (Cw)× = Sw (16) where Rd is the desired orientation and ◦ Ωd is the desired angular velocity in the inertial reference frame. In order to deal with the new control input Γ, we split the problem into two sub-problems following the backstepping paradigm [12]. In the first sub-problem, Section 4.1, we consider Ω as control input instead of a state component of the system and we solve the output regulation problem providing the stability analysis. In the second one, Section 4.2, we consider the control input Ω ≡ Ωc as virtual velocity reference. The virtual control input Ωc , then, is backstepped in order to get the real torque input Γ. 4.1 Angular velocity Ω× as control input In this context it turns out that the control law (6) reads as (with Re := Er ) Ωc× = AdR⊤ ∆× + ∆× = (Cδ)× ν kp X (ei ∧ yir )× 2 i=1 δ̇ = Sδ + C ⊤ Qso(3) β ν kI X β= R (ei ∧ yir ) 2 i=1 (17a) (17b) (17c) (17d) where Ωc is the desired virtual input mentioned earlier. We continue the analysis under the following assumption. Assumption 3. There are two or more linear independent measurements yi (i = 1, . . . , ν; ν ≥ 2) and the symmetric matrix ν kp X r r⊤ y y Y = 2 i=1 i i has three distinct eigenvalues. Note that Assumption 3 specializes Assumption 1 in the specific case of systems posed on SO(3). Indeed, it is straightforward to verify that two is the minimum number of measurements such that the cost function (5) is positive definite around the identity element of the group. Theorem 1 ensures the local attractiveness of the compact set S in (18), exploiting the structure of the special orthogonal group one can shown further properties of the set, namely almost global stability and local exponential stability. The extensions of the local properties of Theorem 1 in the special case of SO(3) are stated in the following Proposition. Proposition 1. Consider the system (15a) along with exosystem (16) and let the controller be given by (17). Let Assumption 3 holds. Then the set S = {(R, δ, (Rd , w)) ∈ SO(3) × Rm × (SO(3) × Rm ) : R⊤Rd = Xr , δ = w} (18) is almost globally asymptotically and locally exponentially stable for the closedloop system. Note that Assumption 2 is no more needed in Proposition 1 since the special orthogonal group SO(3) is a compact manifold. Proof. It is straightforward to verify that, under Assumption 3, the Lyapunov function (8) is positive definite and L(0, 0) = 0. In what follows we proceed by steps showing that: 1. The dynamic of the group error for the closed-loop system has only four ∗ isolated equilibrium points (Re , ∆) = (Rej , 0), j = 1, . . . , 4. ∗ 2. The equilibrium point (Re1 , ∆) = (I3 , 0) is locally exponentially stable. ∗ 3. The three equilibria with (Rej , ∆) 6= (I3 , 0), j = 2, . . . , 4 are unstable. Consider the dynamic of the group error for the closed-loop system ! ν X k p ˜× + Ṙe = AdR⊤ ∆ (Re yir ∧ yir )× Re . 2 i=1 ˜ = 0, which implies As consequence of Theorem 1 for Ṙe = 0 one has that ∆ 0 = = ν kp P (R∗ y r ∧ yir ) 2 i=1 e i ν ν kp P kp P ⊤ ⊤ ⊤ ⊤ Re∗ yir yir − y r y r Re∗ = Re∗ Y − Y Re∗ . 2 i=1 2 i=1 i i (19) ⊤ Proceeding like in the proof of Theorem 5.1 in [19] one has that Re∗ Y = Y Re∗ implies either Re∗ = I3 or tr(Re∗ ) = −1. As consequence, Re∗ is a symmetric matrix and there are only four possible values of Re∗ that satisfy eq. (19), namely  ∗ R = I3    e1 ∗ ⊤ ⊤ Re2 = u1 u⊤ 1 − u2 u2 − u3 u3 ∗ ⊤ ⊤ R = −u1 u1 + u2 u2 − u3 u⊤  3   e3 ∗ ⊤ ⊤ Re4 = −u1 u⊤ 1 − u2 u2 + u3 u3 where u1 , u2 , u3 are the eigenvectors of Y associated to the eigenvalues λ1 , λ2 , λ3 ordered in ascending order. This, in turn, proves step 1. Hereafter we prove item 2 (namely local exponential stability of the set S). Without loss of generality consider Xr = I3 , C = I3 . The dynamics of the error and velocity error for the closed-loop system can be written as   ν ˜ + kp P (Re y r ∧ y r ) Re Ṙe = Re Rd⊤ ∆ i i 2 i=1 ×  ν P ˜˙ × = ˜ + kI Rd Re⊤ (Re y r ∧ y r ) ∆ S∆ i i i=1 and, denoting ˜ = R∆ ˜, ∆ × Ṙ = −RS where R ∈ SO(3), one gets   ν P ˜ + kp Ṙe = Re Rd⊤ R⊤ ∆ (Re yir ∧ yir ) Re 2 i=1 × ν P ˙ ⊤ ˜ = k ∆ RR R (R y r ∧ y r ) . I d i=1 e e i i (20) ˜ of equation (20) around the Consider a first order approximation of (Re , ∆) ∗ equilibrium point (Re1 , 0). It suffices to show that the origin of the linearized system is uniformly exponentially stable in order to prove that the equilibrium ∗ (Re1 , 0) is locally exponentially stable. To this purpose consider the approximation Re = I + x× and ∆ = θ, with x, θ ∈ R3 . By neglecting high order terms one obtains      ẋ A(t) B(t)⊤ x = (21) −C(t) 0 θ θ̇ where A(t) ν kp P yi× yi× , B(t) := RRd , 2 i=1 ν P −kI RRd yi× yi× . := C(t) := i=1 The proof follows the results derived from Theorem 1 in [15] which establish sufficient conditions for the uniform exponential stability of the origin of a linear time-varying system having the standard form in (21). As Pνa matter⊤ of fact, the and Q = constant symmetric positive definite matrices P = kI i=1 yi× yi× kI−1 kp P 2 satisfy the conditions PB ⊤ = C ⊤ and −Q = A⊤ P + PA + Ṗ in Assumption 2 of [15]. Furthermore, it is straightforward to verify that the term B is uniformly persistently exciting, namely for any positive constant ε there exists T > 0 such that Z t+T Z t+T ⊤ RRd Rd⊤ R⊤ dτ = T I3 > εI3 . B(τ )B(τ ) dτ = t t Moreover from Theorem 1 one has that |B(t)| and ∂B(t) remain bounded for all ∂t time. Hence all conditions of Theorem 1 in [15] are satisfied, thus the claim ∗ in item 2 holds. In order to prove that the others three equilibria (Rej , 0), j ∈ {2, 3, 4}, are unstable, it suffices to prove that the origin of the linearized time-varying system is unstable. To this purpose consider the first order ap∗ proximation Re = Rej (I3 + x× ), ∆ = θ with x, y ∈ R3 around an equilibrium ∗ point (Rej , 0). The linearization of (20) is given by ẋj = Υj xj + Rd⊤ R⊤ θ θ̇j = 2 (22) kI RRd Υj xj kp ν kp P ∗ R∗ yi× Rej yi× , with j = 2, . . . , 4. 2 i=1 ej The proof follows results derived from the Chataev’s theorem [12]. In particular, consider the functions where Υj := Vj (xj , θj ) = 1 kI ⊤ ⊤ xj Υj xj − |θj |2 , 2kp 4 j = 2, . . . , 4 and, for an arbitrarily small radius r > 0, define the set Uj,r := {(xj , θj )⊤ | Vj (xj , θj ) > 0, |xj , θj | < r}. Following the Chateav’s theorem, we show now that Υj is not singular, at least one of its eigenvalues is positive and that Uj,r is non-empty for each j ∈ {2, 3, 4} and r > 0. As a matter of fact, consider the characteristic polynomial of the matrix Υj . One has ∗ det(Y Rej − kp ν P ∗ yi⊤ Rej yi I3 − λ̄I3 )  det(Rq λq Rq⊤ Rq R̄j Rq⊤ − tr λq R̄j I3 − λI3 )  det(λq R̄j − tr λq R̄j I3 − λI3 )  det(λq R̄j − tr λq R̄j I3 − λ̄I3 ). det(Υj − λ̄I3 ) = = = = i=1 (23) ∗ where we have used the fact that the matrices Y and Rej can be decomposed ⊤ ⊤ ∗ as Y = Rq λq Rq and Rej = Rq R̄j Rq with λq = diag(λ1 , λ2 , λ3 ) and R̄2 = diag(1, −1, −1), R̄3 = diag(−1, 1, −1), R̄4 = diag(−1, −1, 1). From equation (23) one gets λ(Υ2 ) = λ(Υ3 ) = λ(Υ4 ) = [λ2 + λ3 ; λ3 − λ1 ; λ2 − λ1 ]⊤ [λ3 − λ2 ; λ3 + λ1 ; λ1 − λ2 ]⊤ [λ2 − λ3 ; λ1 − λ3 ; λ1 + λ2 ]⊤ . Since, by Assumption 3, the eigenvalues λ1 , λ2 , λ3 are distinct and 0 ≤ λ1 < λ2 < λ3 it is straightforward to verify that the matrix Υj is not singular and at least one of its eigenvalues is positive. As consequence the set Uj,r is non-empty. Consider now the derivatives of Vj along the trajectories of the system. One has V̇j (xj , θj ) = kI ⊤ ⊤ x Υ Υj xj . kp j j Since the matrix Υj is full rank it is straightforward to verify that Υ⊤ j Υj > 0, with j = 2, . . . , 4. As consequence V̇j is positive for each (x, θ) ∈ Uj,r . From this we conclude that all the conditions of Chateav’s Theorem are satisfied and thus the origin of the system (22) is unstable for j ∈ {2, 3, 4}. This completes the proof. 4.2 Backstepping procedure In order to deal with the new control input Γ starting with the virtual input Ωc , a backstepping procedure is developed. Define Ω̃ as Ω̃ = Ω − Ωc . Consider (6b), (6c) completed with the following control law ˙ +2 Γ = Ω× JΩ − JΩ× R⊤ ∆ + JR⊤ ∆ ν X (ei ∧ yir ) (24a) i=1 + Jkp ν X r yi× (yir − ei )× (Ω̃ + α) − kD Ω̃ i=1 kI β= 2 ν   X r kp R (yir − ei )× yi× J ⊤ Ω̃ + R (ei ∧ yir ) i=1 ! (24b) P with α = kp νi=1 0.5 (ei ∧yir ) and kD a positive arbitrary gain. By backstepping Ω̃ it turns out that control law (24) solves the stabilization problem as stated in the following proposition. Proposition 2. Consider system (15a), (15b) along with exosystem (16) and let the controller be given by (24), (6b), (6c). Let Assumption 3 holds.Then the set S1 = {((R, Rd ), (δ, w), (Ω× , ◦ Ωd× )) : ∈ SO(3)2 × R2m × so(3)2 : R⊤Rd = Xr , δ = w , Ω = ◦ Ωd } is almost globally asymptotically stable and locally exponentially stable for the closed-loop system. Proof. Consider the following function 1 Lbs = L + Ω̃⊤ J Ω̃ 2 where L is the Lyapunov function defined in (8). It is straightforward to verify that under assumption 3 the function Lbs is definite positive respect to the compact set S1 . Differentiating Lbs along the solutions of the closed-loop system and bearing in mind the expression of Ωc , one has L̇bs   ν 1 kp P 2 (Re yir ∧ yir )× + w̃⊤ Sw − δ̇ 4 i=1 kI  ν P 1 (Re yir ∧ yir )× + tr Ω̃⊤ 2  × i=1    ν 1 ⊤ c . ˜ ⊤ AdR P (Re y r ∧ y r ) − tr ∆ −Ω JΩ + Γ − J Ω̇ + Ω̃ × × i i × 2 i=1 = − Recalling the expression of Ωc× in vectorial form and differentiating it, one obtains   ν ˙ + kp P y r (Re y r ) Ω̃ + α Ω̇c = −Ω× R⊤ ∆ + R⊤ ∆ i× i × 2 i=1 ν kp P ˜ − y r (Re yir )× R⊤ ∆. 2 i=1 i× Substituting the expression of Ω̇c in the Lyapunov function and recalling the expression of Γ (24a) and δ̇ (6c), one has L̇bs ν kp P 2 (Re yir ∧ yir )× − kD Ω̃2 4 i=1 ! ν ν P P 2 ˜ ⊤ kp R (Re y r ) y r J ⊤ Ω̃ − R (Re y r ∧ y r ) − β . +∆ i i i × i× kI i=1 i=1 = − Introducing the expression of β (24b) in the above expression one obtains L̇bs = − ν kp X (Re yir ∧ yir )× 4 i=1 2 − kD Ω̃2 . It follows that the compact set S1 is stable in the sense of Lyapunov and that Ω̃ converges to zero. The proof can be completed using similar arguments to Proposition 1. 5 Simulation Results As illustrative example we consider two rigid bodies posed on SO(3). The first one, the simulated system, is modeled as fully actuated dynamic system (15b) while the second one, the simulated exosystem, is modeled as kinematic system (16). The reference directions considered in the simulations are ẙ1 = [1, 0, 0]⊤ and ẙ2 = [0, 1, 0]⊤ . Initial states of the simulated system are chosen as R(0) = I3 , Ω(0) = 0 while the controller gains are chosen to be kp = kd = 2 and kI = 0.4. The inertia matrix in the body-fixed frame of the system is that of an non-axisymmetric rigid body J = diag(2, 1.5, 1)[kg m2 ]. The desired velocity in the inertial frame was chosen as ◦ Ωd = (cos(t), 2 cos(5t), 3 cos(7t))[rad s−1 ]. The initial yaw, pitch and roll of the simulated exosystem are chosen respectively as 180◦, 45◦ and 45◦ . Figure 2 and Figure 3 show the evolution of tr (I − Re ) and the relative velocity error Ω̃ . Plots show that, in steady state, the velocity of the simulated system Ω converges to Ωc and the relative error Re converges to the identity element of the group. t r( I − R e ) 4 3 2 1 0 0 5 10 15 20 25 t - - -Ω̃- 4 [rad/s] 3 2 1 0 0 5 10 15 20 25 t Figure 2: Evolution of the relative Error Re and velocity error Ω̃ in SO(3). Case of perfect knowledge of the Inertia matrix. −12 6.5 x 10 t r( I − R e ) 6 5.5 5 4.5 260 262 264 266 268 t 270 272 274 276 −6 1.5 x 10 - - -Ω̃- [rad/s] 1 0.5 0 260 262 264 266 268 t 270 272 274 276 Figure 3: Evolution of the relative Error Re and velocity error Ω̃ in SO(3) after 260 seconds. Case of perfect knowledge of the Inertia matrix. A second simulation was run, considering a slight unknown variation in the inertia tensor, to have a numerical hint on the robustness properties of the proposed control law. The inertia matrix of the implemented control law is chosen as Jnom = diag(2, 1.5, 1)[kg m2 ] while the real value is Jreal = diag(2.2, 1.3, 1.1)[kg m2 ]. Figure 4 and Figure 5 show that only practical regulation can be achieved in presence of uncertainties in the inertia tensor. t r( I − R e ) 4 3 2 1 0 0 5 10 15 20 25 t - - -Ω̃- 4 [rad/s] 3 2 1 0 0 5 10 15 20 25 t Figure 4: Evolution of the relative Error Re and velocity error Ω̃ in SO(3). Case of uncertainties in the Inertia. 0.02 t r( I − R e ) 0.015 0.01 0.005 0 260 262 264 266 268 t 270 272 274 276 0.8 - - -Ω̃- [rad/s] 0.6 0.4 0.2 0 260 262 264 266 268 t 270 272 274 276 Figure 5: Evolution of the relative Error Re and velocity error Ω̃ in SO(3) after 260 second. Case of uncertainties in the Inertia. 6 Conclusion In this paper, the problem of output regulation for systems defined on matrix Lie-Groups with invariant measurements was considered. The proposed control law structure embeds a copy of the exosystem kinematic updated by means of error measurements. A rigorous stability analysis was provided for both the general case and the particular case of systems posed on the special orthogonal group SO(3). Future works will consider the problem of robustness with respect to uncertainties in the parameters (for instance, in the inertia tensor) and external unknown disturbances. Note, in fact, that the control law proposed in (24) assumes a perfect knowledge of them. Acknowledgement This research was supported by integrated project SHERPA (G.A. 600958) supported by the European Community under the 7th Framework Programme, and by the Australian Research Council and MadJInnovation through the Linkage grant LP110200768. References [1] N. Aghannan and P. Rouchon. An intrinsic observer for a class of lagrangian systems. IEEE Trans. on Automatic Control, 48(6):936–945, 2003. [2] S. Bonnabel, P. Martin, and P. Rouchon. Symmetry-preserving observers. IEEE Trans. on Automatic Control, 53(11):2514–2526, 2008. [3] S. Bonnabel, P. Martin, and P. Rouchon. Non-linear symmetry-preserving observers on lie groups. IEEE Trans. on Automatic Control, 54(7):1709– 1713, 2009. [4] F. Bullo. Nonlinear Control of Mechanical Systems: A Riemannian Geometry Approach. PhD thesis, California Istitute of Technology, 1999. [5] C. I. Byrnes and A. Isidori. Nonlinear internal model for output regulation. IEEE Trans. on Automatic Control, 49(12):2244–2247, 2004. [6] E.J. Davison. The robust control of a servomechanism problem for linear time-invariant multivariable systems. IEEE Trans. on Automatic Control, 1976. [7] B.A. Francis. The linear multivariable regulator problem. SIAM J. Contr. Optimiz., 1977. [8] B.A. Francis and W.M. Wonham. The internal model principle of control theory. Automatica, 1976. [9] Minh-Duc Hua, Mohammad Zamani, Jochen Trumpf, Robert Mahony, and Tarek Hamel. Observer design on the special euclidean group SE(3). Decision and Control and European Control Conference, 2011. [10] A. Isidori and C.I. Byrnes. Output regulation for nonlinear systems. IEEE Trans. on Automatic Control, 1990. [11] A. Isidori, L. Marconi, and A. Serrani. Robust Autonomous Guidance. An internal model approach. Springer, 2003. [12] H. Khalil. Non linear systems. Prentice-Hall, 1996. [13] C. Lageman, J. Trumpf, and R. Mahony. Observers for systems with invariat outputs. Proceedings of the European Control Conference, pages 4587–4592, 2009. [14] C. Lageman, J. Trumpf, and R. Mahony. Gradient-like observers for invariant dynamics on lie group. IEEE Trans. on Automatic Control, 55(2):367– 377, 2010. [15] A. Loria and E. Panteley. Uniform exponential stability of linear timevarying system: revisited. Systems & Control Letters, 2002. [16] Y. Ma, S.Soatto, J. Kosecka, and S. Sastry. An Invitation to 3-D Vision: From Images to Geometric Models. SpringerVerlag, 2003. [17] R. Mahony, T. Hamel, and L. Marconi. Adding an integrator for output regulation of systems with matrix lie-group states. American Control Conference, 2015. [18] R. Mahony, T. Hamel, P. Morin, and E. Malis. Non-linear complementary filters on the special linear group. International Journal of Control, 2012. [19] R. Mahony, T. Hamel, and J. M. Pflimlin. Non-linear complementary filters on the special orthogonal group. International Journal of Control, 2012. [20] C. Mayhew, R. Sanfelice, and A. Teel. Quaternion-based hybrid control for robust global attitude tracking. IEEE Trans. on Automatic Control, 56(11):2555–2566, 2011. [21] F. Delli Priscoli, L. Marconi, and A. Isidori. A new approach to adaptive nonlinear regulation. SIAM J. Contr. Optimiz., 2006. [22] D. S. Bernstein S. P. Bath. A topological obstruction to continuous global stabilization of rotational motion and the unwinding phenomenon. Systems & Control Letters, 39:63–70, 2000. [23] J. Trumpf, R. Mahony, T. Hamel, and C. Lageman. Analysis of non-linear attitude observers for time-varying reference measurements. IEEE Trans. on Automatic Control, pages 2789–2800, 2012.
3cs.SY
Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code∗ Guillermo Vigueras Manuel Carro† IMDEA Software Institute Campus de Montegancedo 28223 Pozuelo de Alarcón, Madrid, Spain {guillermo.vigueras,manuel.carro}@imdea.org Salvador Tamarit Julio Mariño Universidad Politécnica de Madrid Campus de Montegancedo 28660 Boadilla del Monte, Madrid, Spain {salvador.tamarit,julio.marino}@upm.es The current trends in next-generation exascale systems go towards integrating a wide range of specialized (co-)processors into traditional supercomputers. Due to the efficiency of heterogeneous systems in terms of Watts and FLOPS per surface unit, opening the access of heterogeneous platforms to a wider range of users is an important problem to be tackled. However, heterogeneous platforms limit the portability of the applications and increase development complexity due to the programming skills required. Program transformation can help make programming heterogeneous systems easier by defining a step-wise transformation process that translates a given initial code into a semantically equivalent final code, but adapted to a specific platform. Program transformation systems require the definition of efficient transformation strategies to tackle the combinatorial problem that emerges due to the large set of transformations applicable at each step of the process. In this paper we propose a machine learning-based approach to learn heuristics to define program transformation strategies. Our approach proposes a novel combination of reinforcement learning and classification methods to efficiently tackle the problems inherent to this type of systems. Preliminary results demonstrate the suitability of this approach. Keywords: Program Transformation, Machine Learning, Heterogeneous Systems 1 Introduction The introduction of multi-core processors marked the end to an era dominated by a constant increase of clock-rate, following Moore’s law. Today, even the strategy to increase performance through higher core-counts is facing physical limitations. In addition, the performance increase results in power consumption issues which are currently one of the main limitations for the transition from petascale systems to next-generation exascale systems. Hence, the metric for measuring performance improvement of computational platforms has changed from absolute FLOPS numbers to FLOPS per Watt and FLOPS per surface unit of hardware. Thus, rather than having CPU-based supercomputers a growing trend goes towards integrating a wide range of specialized (co-)processors into traditional supercomputers. These specialized architectures can outperform general purpose platforms while requiring lower energy consumption and less real estate. A study of energy consumption in data centers [8] has revealed the impact that efficient computer platforms can have in the worldwide energy consumption. This study analyzed historical energy consumption data for the periods 2000-2005 and 2005-2010. The study results showed that in 2005 data centers used 0.97% of the total electricity consumption of the world, while in 2010 this value increased to ∗ Work partially funded by EU FP7-ICT-2013.3.4 project 610686 POLCA, Comunidad de Madrid project S2013/ICE-2731 N-Greens Software, and MINECO Projects TIN2012-39391-C04-03 / TIN2012-39391-C04-04 (StrongSoft), TIN2013-44742C4-1-R (CAVI-ROSE), and TIN2015-67522-C3-1-R (TRACES). † Manuel Carro is also affiliated to the Universidad Politcnica de Madrid A. Villanueva (Ed.): PROLE 2016 EPTCS 237, 2017, pp. 52–67, doi:10.4204/EPTCS.237.4 c G. Vigueras, M. Carro, S. Tamarit & J. Mariño, This work is licensed under the Creative Commons Attribution License. G. Vigueras, M. Carro, S. Tamarit & J. Mariño, 53 Listing 1: Initial code. // ABSTR: [3, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 3, 2, 4, 0] int dead_rows = K / 2; int dead_cols = K / 2; int normal_factor = K * K; for (r = 0; r < N - K + 1; r++) { for (c = 0; c < N - K + 1; c++) { sum = 0; for (i = 0; i < K; i++) { for (j = 0; j < K; j++) { sum += input_image[r+i][c+j] * kernel[i][j]; } } output_image[r+dead_rows][c+dead_cols] = (sum / normal_factor); } } 1.3% instead of the expected 2.2%, according to the trends from the period 2000-2005. The study explained the reduction from 2.2% down to 1.3% as a consequence of the increasing use of energyefficient computational platforms, showing the economic impact of using greener architectures. However, integrating different specialized devices increases the degree of heterogeneity of the system. On the other hand, compilers typically optimize code for only one (homogeneous) destination infrastructure at a time, requiring the developer to manually adapt the program in order to be executed on hybrid architectures. As a result, the high degree of heterogeneity limits programmability of such systems to a few experts, and significantly reduces portability of the application onto different resource infrastructures. Opening the access of heterogeneous platforms to a wider spectrum of users is an important issue to be tackled. In past decades, different scientific areas showed an enormous advance and development thanks to the use of massive parallel architectures to implement in-silico models. The use of “virtual labs” through computational platforms allowed a better understanding of the physical phenomena under study, investigate a much wider range of solutions, and drastically reduce the cost with respect to performing real experiments. The computational requirements of scientific and industrial applications are pushing a driving the development of exascale systems. For that reason, new programming models, tools and compilers are required in order to open the access of exascale systems to a wider programming community, enabling translational research to enhance human welfare. In order to ease the programmability of heterogeneous platforms, we have proposed [17, 16] a rulebased transformation system which can help developers to convert architecture-neutral code into code which can be deployed onto different specific architectures. While the mentioned transformation system is capable of performing sound program transformations, it lacked an automatic method to guide the transformation process. This paper describes a machine learning-based approach to learn transformation strategies from existing examples and to guide that system by using these strategies. Rule-based transformation systems pose different problems, like the explosion in the number of states to be explored, arising from the application of transformation rules in an arbitrary order or the definition of a stop criteria for the transformation system. For the latter we propose the use of classification trees 54 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code and for the former we propose a novel approach based on reinforcement learning. Both approaches use code abstractions to effectively capture relevant features of a given code. In order to illustrate the approach, Listing 1 shows code performing a 2D convolution with its associated abstraction at the top of the listing (see Section 4 for a description of the code abstraction). The code in Listing 1 is well suited to be parallelized, by adding OpenMP pragmas, for a multi-core CPU. However, different code transformations, altering different code abstractions features, are required to ease the translation to different target platforms like GPUs or FPGAs. For example, by coalescing the two outer loops, obtaining a linear iteration space, or by transforming the data layout of 2D arrays into 1D arrays, we obtain a sequential code easier to map onto the two platforms mentioned before (see Section 5.3). The code features to be transformed are represented in the abstraction vector in Listing 1 through the first element (maximum loop depth), the twelfth element (number of non-1D arrays) and the fourteenth element (total number of loops). Our approach learns different sequences of code transformations that change the relevant abstraction features and obtain, for the example in Listing 1, a code performing a 2D convolution within a for loop with a linear iteration space, using 1D arrays. We have performed a preliminary evaluation of the approach obtaining promising results that demonstrate the suitability of the approach for this type of transformation systems. The rest of the paper is organized as follows: Section 2 reviews previous work in the field of program transformation systems in general and previous approaches using machine learning techniques. Section 3 describes the toolchain where our machine learning-based approach is integrated. Section 4 describes the abstraction of code defined in order to apply machine learning methods. Section 5 describes the methods used to learn program transformation strategies. Section 6 presents some preliminary results and, finally, Section 7 summarizes the conclusions and proposes future work. 2 State of the Art Rule-based program transformation systems support the formulation of basic code transformation steps as generic rules and arrange their automatic application. This scheme offers the flexibility of splitting complex code transformations into small steps, admitting efficient implementations that can scale to large programs. By adding more rules, the transformation system can also increase its capabilities in a scalable way. Rule-based systems also allow to decouple the definition of transformation steps (i.e. rule applications) from the strategies followed to apply the rules. This decoupling provides more flexibility to try different transformation strategies and select the best one according to the purpose of the system [2, 14]. Rule-based transformation has been used before to generate code for different computational platforms. The transformation of C-like programs and its compilation into synchronous architectures, like FPGAs, and asynchronous platforms, like multi-core CPUs, has been addressed before [3]. However, the input language of this approach (Handel-C) is meant to specify synchronous systems, thus limiting its applicability to this type of systems. A completely different approach is to use linear algebra to transform the mathematical specification of concrete scientific algorithms [6, 5, 4]. Here, the starting point is a mathematical formula and, once the formula is transformed, code is generated for the resulting expression. However, a reasonable acceleration over hand-tuned code happens only for algorithms within that concrete domain, and applying the ideas to other contexts does not seem straightforward. Machine learning techniques have been already used in the field of compilation and program transformation [9, 12, 1]. All these approaches share the same principles as they obtain an abstract representation of the input programs in order to apply machine learning methods. Nevertheless, previous approaches target some specific architectures limiting the applicability of the approach and making it not suitable for heterogeneous platforms. In our approach, we obtain program abstractions in order to enable the machine G. Vigueras, M. Carro, S. Tamarit & J. Mariño, 55 Ready code Translated code GPGPU (OpenCL) Initial code OpenMP MPI FPGA (MaxJ) Transformation Structure, S2S, properties Translation Glue code, calls to libraries Figure 1: Stages of the transformation tool. learning-guided transformation and compilation of a program written in C for heterogeneous systems. Additionally, none of the previous works have explored the use of reinforcement learning methods [7] in the field of program transformation and compilation. 3 Program Transformation Toolchain for Heterogeneous systems In this work we propose the automatic learning of transformation strategies for a rule-based transformation toolchain [17, 16]. This type of systems offer a higher flexibility over compilation infrastructures where the code transformations and optimizations are embedded in the compilation engine [15]. The rule-based transformation toolchain mentioned before [17, 16] defines a process consisting of two main stages (Fig. 1). In the one, input C code is transformed into semantically equivalent C code which is better suited for a given platform. It basically reshapes the input code taking into account syntactic/semantic restrictions of the compiler/programming models of each target platform (e.g. if the target compiler does not accept general if-then-else statements, the input code will be transformed to remove them if possible, without modifying the program semantics). In this phase, transformation rules, written in a C-like DSL and working at the abstract syntax tree (AST) level, rewrite input code which matches a pattern and fulfills some semantic conditions, into resulting code which is expressed as a template in the rewriting rule. Since at each step several rules can generally be applied at several code locations, this first phase defines a search space exploration problem where the states correspond to the different configurations of code obtained from the application of rules, and the transitions among states correspond to the applied rules. The usually large number of applicable rules and code locations where they can be applied causes a combinatorial explosion of the number of states. The second phase of the toolchain consists of a translation process which generates code for a given platform taking as input the code generated in the first phase. This platform-specific code is fed to a platform-specific compiler. This second phase poses another problem: determine when code being subject to the transformation process is ready for the translation phase, i.e. when the transformation process can stop. Since some platforms might require a long compilation process (taking up to tens of hours for some FPGA compilers, for example, for moderate-sized programs), the toolchain should translate and compile only code that is likely to be successfully compiled on a given target platform, rather than trying to compile every single code obtained during the transformation phase. Additionally, when considering the first and second phases together, the intermediate code states obtained in the transformation 56 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code sequences of the first phase might not improve monotonically non-functional metrics (e.g. execution time, energy consumption, etc.), but still produce code with better performance in the second phase. Given the nature of the problems associated with the transformation toolchain, we propose the use of different machine learning techniques to effectively guide the transformation process. Concretely, we have used reinforcement learning methods to learn transformation strategies that take into account the non-monotonic behavior of the transformation sequences, and we have used classification methods to learn appropriate states that can be translated and compiled in the target platforms. Machine learning techniques require descriptions of the problem domain; therefore, we compute an abstract representation of the input programs and the intermediate code generated during the transformation process. 4 Mapping Code to Abstraction As mentioned in before, machine learning operates on descriptions of the problem domain (C code, in our case). We use code abstractions as descriptions; these should be rich enough to reflect the changes performed by the transformation rules at the AST level. For that reason, we include in the abstractions quantitative descriptions involving features like AST patterns, control flow, data layout, data dependencies, etc. The current code abstraction consists of a vector of features (described later) which capture traits which are modified by the transformation rules and also reflect code patterns that match the syntactic/semantic restrictions of target compiler/programming models. These abstractions are generated through a code analysis tool that parses the AST to extract the features mentioned before, implementing a function A : Code → Abstraction that maps codes to abstractions. The tool has been implemented in Python using the pycparser1 module and can extract feature information either by directly analyzing the code or by parsing code annotations provided by the user or by third-party analyzers. These match those supported by the transformation engine [17]: • Maximum nested loop depth: Depth of nested for loops. A value of 0 means that no nested loops are present, 1 means that there are two nested loops, and so on. • Number of function calls: Number of function calls present in the analyzed code. • Number of array writes shifted in bodies of for loops: This feature counts the number of array accesses with positive offset according to the following pattern: a for loop which performs multiple write accesses to the same array in the same iteration, some (but not all) of which use a positive offset with respect to the loop iteration variable. Listing 2 illustrates this feature. It would take a value of 1 for the code on the left, since for each iteration there are multiple write accesses to v[] and one of them has positive offset (because of the index i+1). For the code on the right, this feature takes a value 0, since for each iteration only position i+1 is written to: there is no position written to with a non-positive offset. Detecting writes with this characteristic is necessary since this pattern would fail to compile directly for a FPGA. • Irregular loops: This feature takes a binary value and informs whether there are loops containing or not statements that alter the regular flow of a structured program (e.g., there are break or continue statements). 1 https://github.com/eliben/pycparser G. Vigueras, M. Carro, S. Tamarit & J. Mariño, Listing 2: Array writes shifted. for(i=1;i<N;i+=2) { v[i] = v[i-1]; v[i+1] = v[i-1]*i; } for(i=0;i<N-1;i++) { aux = i*i; v[i+1] = aux; } Listing 4: Loop with schedule pattern. #pragma stml loop_schedule for(j=0; j<M; j++) { w[j] = 0; for(i=0;i<N;i++) w[j] += v[j*N+i]; } 57 Listing 3: Static loop limits. for(j=0;j<N;j++) { for(i=0;i<size(v);i++) update(v,i); clean(v); } Listing 5: Aux. variable array index. aux = 0; for(j=0; j<N; j++) { w[j] = v[aux]; aux++; } • Global variables: This binary feature states if any global variable is written or not within the fragment of code of interest. • if statements: This feature counts the number of if statements within the piece of code analyzed. • Static limits of for loops: This feature captures whether some for loop in the analyzed code has a non-static iteration limit, i.e. if it can change between executions of the loop. Listing 3 shows an example where the upper limit of the innermost loop is non-static, since it might change for each iteration of the outer loop if elements of v are removed by function clean. • Iteration independent loop: This feature counts the number of for loops in the analyzed code which do not have carried dependencies across iterations. This information is obtained from annotations provided by the user or by external code analyzers. • Any for loop with loop schedule: This feature detects patterns where two nested loops are used to iterate over an array split in fragments. This code pattern is detected through an annotation that can be inserted either by the user or by the transformation tool after applying a loop scheduling transformation. Listing 4 shows an example of loop schedule where array v is accessed in chunks of size N. • Number of loop-invariant variables: This feature quantifies the number of variables that are assigned outside a for loop and are not modified inside it. • Number of loop hoisted variable modifications: This feature is used to count the number of variables that are assigned values outside a for loop and are modified within the loop. • Number of non-1D array: This feature counts the number of arrays in the code with two or more dimensions. • Number of auxiliary variables to access arrays: This feature counts the number of auxiliary variables used to index an array. If a variable used as array index is not one of the iteration variables of the for loops that iterate over the array, then it is considered as an auxiliary index variable. Listing 5 shows a simple example where the aux variable is used to index the array v instead of using the iteration variable j of the for loop. 58 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code • Total number of for loops: This feature counts the total number of for loops, either nested or non-nested. For example, this feature takes the value 2 for the code in Listing 4. • Non-normalized for loops: This feature counts the number of non-normalized for loops. A for loop is considered normalized iff its iteration step is 1. The features described above were sufficient to obtain preliminary results for a set of use cases (Section 6). However, we plan to increase the vector of features as the set of rules and use cases grows. 5 Automatic Learning of Transformation Heuristics As mentioned before, two outstanding problems faced by the transformation engine are dealing with efficiently finding transformation sequences in a very large (even infinite) state space with a non-monotonic fitness function and deciding when to stop searching, since there is no a-priori general way to determine that an optimum (in the sense of the most efficient possible code) state has been reached. Our approach uses classification trees to solve the latter and reinforcement learning to solve the former. We will describe our approaches in the next sections. 5.1 Classification Trees In machine learning and statistics, classification is the problem of identifying the category to which a new observation belongs among a set of pre-defined categories. Classification is done on the basis of a training set of data containing observations for which it is known to which category they belong [10]. Different approaches can be used to classify. We have decided to start evaluating the adequacy of classification trees for our problem since it performs feature selection without complex data preparation. A classification tree is a simple representation to classify examples according to a set of input features. The input features should have finite discrete domains and there is a single target variable called the classification feature. Each element of the domain of the target variable is called a class. In a classification tree each internal (non-leaf) node is labeled with an input feature and each leaf node is labeled with a class or a probability distribution over the classes. Thus, a tree can be learned by splitting the source data set into subsets based on values of input features. This process is recursively repeated on each derived subset. The recursion is completed when the subset of data in a node has the same value for the target variable, or when splitting no longer improves the predictions. The source data typically comes in records of the form (x,Y ) = ([x1 , x2 , x3 , ..., xk ],Y ) The dependent variable, Y , is the target variable that the classification tree generalizes in order to be able to classify new observations. The vector x is composed of the input features xi , used for the classification. In this context, the source input data for our problem is composed of the vectors of the abstractions described in Section 4, i.e., k = 15 in our case. The domain of the target variable can take values among the four different final platforms we currently support: FPGA, GPU, Shared-Memory CPU (SM-CPU) and Distributed-Memory CPU (DM-CPU). Since a given code and its associated abstraction might be well suited for more than one platform, we compute the number of classes for our target variable as 2n − 1, where n is the number of target platforms. Thus, we obtain 24 − 1 = 15 classes. The classes obtained for the target variable make it possible to define the final states of the transformation stage of the toolchain described in Section 3, which will also serve to define the final states for the reinforcement learning algorithm that is described next. The classification-based learning described G. Vigueras, M. Carro, S. Tamarit & J. Mariño, 59 in this section has been implemented using the Python library Scikit-learn [11]. This library implements several machine learning algorithms, provides good support and ample documentation, and is widely used in the scientific community. 5.2 Reinforcement Learning Reinforcement learning [10] is an area of machine learning concerned with how software agents ought to take actions in an environment to maximize some notion of cumulative reward. A reinforcement learning agent interacts with its environment in discrete time steps. At each time t, the agent receives an observation ot which typically includes the reward rt . It then chooses an action at from the set of available actions, which is subsequently sent to the environment. The environment moves from the current state st to a new state st+1 providing the reward rt+1 associated with the transition (st , at , st+1 ). The goal of a reinforcement learning agent is to collect as much reward as possible. According to the previous description, reinforcement learning seems well suited to represent the optimization process of a programmer or a compiler, since it typically consists of iteratively improving an initial program in discrete steps, where code changes correspond to actions and code versions obtained during the optimization process correspond to states. Moreover, code is typically evaluated after every change, often according to some non-functional properties such as execution time, memory consumption speedup factor, . . . The result of these evaluations can be easily translated into rewards and penalties that support the learning process. The result of the learning process of the agent is a state-action table (Fig. 3) which will eventually contain values for each combination (s, a) of states and actions. These values are scores for the expected profit to be obtained from applying action a to state s. This table is initially filled in with a default value and is iteratively updated following a learning process which we briefly describe below. The process of reinforcement learning is based on a set of predetermined transformation sequences which are assumed to be models to learn from. Each sequence S is composed of a set of states S = s0 , s1 , . . . , st−1 , st , st+1 , . . . , sl and the actions which transform one state into the next one. The final state of each transformation sequence has a different reward value which in our case is related with the performance of the final code corresponding to state sl in the sequence (better performance gives higher rewards). The training phase of reinforcement learning consists of an iterative, stochastic process in which states (i.e. code abstractions) from the training set are randomly selected. For each state s a learning episode is started by selecting the action a with the highest value in Q for that s and moving to a new state s0 according to transition (s, a, s0 ). From state s0 the same transition process is repeated to advance in the learning episode until a final state is reached or a given number of episode steps is performed. When the episode terminates, the values in Q corresponding to the states and actions of the visited sequence are updated according to (1), where Qinit (st , at ) is the initial value of Q for state st and action at . Note that st (resp. at ) is not the t-th state in some fixed ordering of states, but the t-th state in the temporal ordering of states in the sequence used to learn.   Q(st , at ) + α · (rt+1 + γ · Q(st+1 , at+1 ) − Q(st , at ) ) if st not f inal (1) Q(st , at ) =  Qinit (st , at ) otherwise The final states in (1) are defined based on the classification described in Section 5.1. Two additional parameters appear in (1): the learning rate α, 0 < α ≤ 1, and the discount factor γ, 0 < γ ≤ 1. The learning rate determines to what extent the newly acquired information will override the old information. A factor 60 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code SM : Abstraction → State Q : State × Action → R AM : Rule → Action RS : Code → Rule RS(c) = arg maxru∈Rule Q(SM(A(c)), AM(ru)) Figure 2: RL function definitions. of 0 will make the agent not learn anything, while a factor of 1 would make the agent consider only the most recent information. The discount factor implements the concept of delayed reward by determining the importance of future rewards. A factor of 0 will make the agent opportunistic by considering only current rewards, while a factor close to 1 will make it strive for long-term rewards. If the discount factor reaches or exceeds 1, the values in Q may diverge [10]. In order to use the reinforcement learning state-action in our setting, we need to define some mappings. The abstraction of a concrete piece of code is provided by function A (described in Section 4). Code abstractions and transformation rules must be mapped to states and actions, respectively, in order to index the state-action table. This mapping is done through functions SM and AM (Fig. 2). Based on the mapping of abstractions and rules defined, the reinforcement learning state-action table of (1) can also be modeled as a function Q (see Fig. 2). Using functions A, SM, AM and Q, the strategy of the transformation toolchain for selecting rules at each transformation step can be modeled with function RS, defined in Fig. 2. This function takes as input a given code c and selects the transformation rule r associated to action AM(r) that maximizes the value provided by Q for the state SM(A(c)) associated to input code c. Thus, the rule selection strategy is modeled by function RS defined in Fig. 2. The operator arg max in function RS may return, by definition, the empty set, a singleton, or a set containing multiple elements. However, in our problem, parameters α and γ as well as the reward values rt+1 appearing in (1) can be tuned to ensure that a single rule is returned, thus avoiding a nondeterministic RS function. Section 6 gives further details on how we selected their values. The relationships among the code, its abstraction, the rules, and the contents of the state-action c0 c1 c2 c3 c4 c5 .. . c6 c7 c8 Actions = Rules Code abstraction } Concrete code r0 r1 s0 q0,0 q0,1 s1 q1,0 q1,1 A C0 SM C1 SM .. . Cm SM sm rn ··· ··· q0,n q1,n .. . .. . .. . .. . q0,m q1,m ··· ··· qm,n } States ≈ Code Abstractions Figure 3: State-Action table for code, code abstraction, and rules. G. Vigueras, M. Carro, S. Tamarit & J. Mariño, 61 matrix are depicted in Fig. 3. Table Q is used as follows: for a concrete code ck we find its abstraction Ci = A(ck ). Let us assume i = 0. From the row i corresponding to Ci in matrix Q we obtain the column j with the highest value qi, j (in our example, q0,1 , in blue and boldface). Column j corresponds to rule R j , which is expected to give the most promising code when applied to a code state whose abstraction is Ci (in our case it would be R1 ). Rule R j would be applied to ck to give c0 . If c0 corresponds to a final state, the procedure finishes. Otherwise, we repeat the procedure taking c0 as input and finding again a rule to transform c0 . In order to implement the reinforcement learning-based module in charge of learning heuristics for program transformation we have used the Python library PyBrain [13]. This library adopts a modular structure separating in classes the different concepts present in reinforcement learning, such as the environment, the observations and rewards, the actions, etc. This modularity allowed us to extend the different classes and ease their adaptation to our problem. The PyBrain library also provides flexibility to configure the different parameters of the reinforcement learning algorithm. 5.3 A Simple Example We will use a 2D convolution kernel as an example to show the resulting state-action table obtained after learning from a simple transformation sequence with five states and two transformation rules. The first rule (R0 ) considered in the example transforms a non-1D array into a 1D array. The second rule (R1 ) performs a collapse of two nested for loops producing a single loop. The initial, intermediate and final codes obtained from the application of these rules is described below. Listing 6 through Listing 9 highlights changes in code using this style to indicate the portion of the code that will be changed after rule application and code highlighted using this style indicates the code resulting after applying a transformation rule. Listing 1 shows the initial code and the associated vector of features, as a code comment, according to the description in Section 4. Listing 6 shows the result of applying rule R0 to the code in Listing 1. The array input image is transformed into a 1-D array and the vector of features associated to code is changed accordingly. Listing 7 shows the result of applying again rule R0 to the code in Listing 6. The array kernel is then transformed into a 1-D array and the feature vector is updated. Listing 8 shows the result of applying rule R0 to the code in Listing 7. Again, the array output image is transformed into a 1-D array and the vector of features updated. Listing 9 shows the result of applying rule R1 to the code in Listing 8. The two outermost loops are collapsed into one for loop, but keeping an iteration space with the same number of iterations. Note that the code abstraction reflects the change, since the number of loops has decreased by one. Fig. 4 shows a table with the resulting values of the state-action table (Q) for the transformation sequence described before. Fig. 4 has as many rows as states obtained from the evaluation of SM(A(Ci )) for each code Ci , where C0 is the initial code and C4 is the final code classified as ready-code for an FPGA. Fig. 4 shows the learned sequence composed of four steps: three consecutive applications of rule R0 and one application of rule R1 . The values in Q for this sequence are highlighted in blue, and they decrease from the state SM(A(C3 )), with the highest reward, down to the initial state SM(A(C0 )). This decay behavior is caused by the discount factor (γ) introduced in (1). It should be noted that the values in Q for the final states are not updated by the recursive expression in (1). Thus, the final state SM(A(C4 )) keeps the initial value of 1. It should be noted that numerical values in table shown in Fig. 4 are obtained as the result of an iterative and stochastic learning process as defined in (1) and using parameter values described in Section 6. The example used in this section shows the transformation of a piece of C code. However, the fact 62 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code Listing 6: Transformation step 1. Listing 7: Transformation step 2. // ABSTR: [3, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 2, 2, 4, 0] // ABSTR: [3, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 2, 4, 0] int dead_rows = K / 2; int dead_cols = K / 2; int dead_rows = K / 2; int dead_cols = K / 2; int normal_factor = K * K; int normal_factor = K * K; for (r = 0; r < N - K + 1; r++) { for (r = 0; r < N - K + 1; r++) { for (c = 0; c < N - K + 1; c++) { for (c = 0; c < N - K + 1; c++) { sum = 0; sum = 0; for (i = 0; i < K; i++) { for (i = 0; i < K; i++) { for (j = 0; j < K; j++) { for (j = 0; j < K; j++) { sum += sum += input_image[(r+i)*(N - K + 1) + input_image[(r+i)*(N - K + 1) + (c+j)] (c+j)] * kernel[i*K+j]; } * kernel[i][j]; } } output_image[r+dead rows][c+dead cols] = } (sum / normal_factor); output_image[r+dead_rows][c+dead_cols] = } (sum / normal_factor); } } } Listing 8: Transformation step 3. Listing 9: Transformation step 4. // ABSTR: [3, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 2, 4, 0] // ABSTR: [2, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 2, 3, 0] int dead_rows = K / 2; int dead_cols = K / 2; int dead_rows = K / 2; int dead_cols = K / 2; int normal_factor = K * K; int normal_factor = K * K; for(r = 0; r < N - K + 1; r++) { for(c = 0; c < N - K + 1; c++) { sum = 0; for (i = 0; i < K; i++) { for (j = 0; j < K; j++) { sum += input_image[(r+i)*(N - K + 1) + (c+j)] * kernel[i*K+j]; } } output_image[(r+dead rows)*(N-K+1) + for (z = 0; z < (N - K + 1)*(N - K + 1); z++) { int r = (z / (N - K + 1)); int c = (z % (N - K + 1)); sum = 0; for (i = 0; i < K; i++) { for (j = 0; j < K; j++) { sum += input_image[(r+i)*(N - K + 1) + (c+j)] * kernel[i*K+j]; } } output_image[(r+dead_rows)*(N - K + 1) + (c+dead_cols)] = (sum / normal_factor); } (c+dead cols)] = (sum / normal_factor); } } G. Vigueras, M. Carro, S. Tamarit & J. Mariño, SM(A(C0 )) SM(A(C1 )) SM(A(C2 )) SM(A(C3 )) SM(A(C4 )) AM(R0 ) 17.03718317 17.25327145 17.51541052 16.72942327 1. 63 AM(R1 ) 16.21544456 16.80486418 16.7189079 17.78007298 1. RS(Ci ) R0 R0 R0 R1 - Figure 4: Values learned for Q table. Figure 5: Execution times for transformation sequences (on a CPU). Figure 6: Execution times for OpenCL versions (on a GPU). that the machine learning methods used work on program abstractions makes the approach generic and suitable for other imperative languages (e.g., FORTRAN, which is also widely used in scientific computing). The application of the approach to other languages would require changes to the tool described in Section 4 in order to account for some specific syntactical patterns of a particular programming language. Nevertheless, most of the abstraction features identified and described in Section 4 are also applicable to other imperative languages since they capture common aspects like control flow, data layout, data dependencies, etc. 6 Results We will present now some results obtained using machine learning-based transformation strategies. We will first show evidence to support our claim, made in Section 5.2, regarding the non-monotonic behavior of non-functional properties for transformation sequences and, second, we will evaluate the applicability of reinforcement learning to learn from these non-monotonic sequences. Non-monotonicity has two sides: on one side while transforming (sequential) code to a shape better suited to some final platform, the efficiency observed in a CPU for a good sequence can increase and decrease. On the other side, the final code which extracts the best performance on a CPU is not necessarily the most efficient on the final platform. In order to show this behavior we have identified four different transformation sequences for a use case which performs image compression using the discrete cosine transform. Each of the identified sequences finally produces C code with a shape adequate to be mechanically translated to OpenCL and executed on a GPU. We have measured the average execution time of 30 runs for each intermediate state of each sequence. It can be seen (Fig. 5) that these execution 64 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code times do not change monotonically. As expected, we also obtain different sequential codes ready to be translated to OpenCL and with different execution times (still on a CPU). As a second step, we have measured the performance of the OpenCL version generated from the final state of each sequence. Fig. 6 shows the execution times of the original sequential program and OpenCL versions as well as the speedup of each OpenCL version with respect to the sequential program (on top of each bar). The OpenCL code obtained from sequence 4 is the fastest one, with a 2.53× speedup factor w.r.t. the initial code on a CPU. However, looking at Fig. 5, sequence 4 was not the fastest one on the CPU —it was in fact the second slowest one. Looking together at Fig. 5 and 6, it is easy to see that there is no clear correlation between the non-functional property measured (execution time) of each final sequential code and the performance of the corresponding OpenCL version. Based on these results it can be concluded that an effective method must be used to discover and learn the uncorrelated relation between final sequential codes and parallel versions for a range of target platforms. We have decided to base our approach on reinforcement learning, since it is driven by final performance measurements rather than on intermediate values which can lead to sub-optimal results. We have also performed a preliminary evaluation of reinforcement learning as a technique to learn and then guide a rule-based program transformation tool. We have selected four use cases and identified different transformation sequences leading to code that can be mechanically translated to OpenCL and executed on a GPU. These four cases and their transformation sequences have been used as training set. The cases in the training set are the image compression program (compress), mentioned before in Fig. 5 and 6, an image filter that splits the different color channels of a given RGB input image in different images (rgbFilter), the detection of edges in an image using a Sobel filter (edgeDetect) and another one performing image segmentation given a threshold value (threshold). Once the training set is defined, the reinforcement learning process requires tuning the two parameters in (1): the learning rate (α) and the discount factor (γ). For this purpose, an empirical study was performed in which parameter values leading to transformation sequences providing the fastest OpenCL versions were selected. As an outcome, a value of 0.5 was used for α and 0.6 for γ. Also, reward values have been chosen in order to give a higher reinforcement to those sequences leading to final codes with better performance. In our example, the reward values used for the best sequences of each use case application are substantially higher, with a ratio of 1:100, with respect to the rest of transformation sequences. After training, three different use case applications were used as prediction set. We have selected applications that share code patterns with the examples in the prediction set. This matches the idea that sets of transformation rules can very naturally depend on the application domain, and therefore it makes sense that the training sets also belong to the same domain. The applications in the prediction set were mechanically transformed according to the previously learned sequences and finally translated by hand into OpenCL. Independently, OpenCL versions of the initial C code were manually written to provide a comparison with what a programmer would do. The evaluation of the reinforcement learning approach was made by comparing the performance of the hand-coded versions with that of the versions mechanically generated. Fig. 7 shows the results obtained for the training and prediction sets in terms of speedup of the OpenCL versions w.r.t. the original sequential program. By comparing the results for the training and the prediction sets we can assess the behavior of the transformation system after the learning phase. Looking at results in Fig. 7 we can see that the transformation sequences leads to code that provide acceleration factors comparable to those of the manually coded versions. Although this preliminary evaluation is based on a small sample, it shows that our approach seems promising in tackling the problem of deciding strategies for rule-based program transformation systems. G. Vigueras, M. Carro, S. Tamarit & J. Mariño, 65 Figure 7: Speed up for train and prediction sets. The results in this section show a preliminary, but very positive, evaluation for the generation of code for GPU platforms. We believe that the same approach can be followed in order to target different platforms which are present in heterogeneous systems. In this way, a separate state-action table can be used for learning adequate transformation sequences for each target platform and rewards would be assigned based on the performance of the final code generated for each platform. 7 Conclusions and Future Work We have proposed a machine learning-based approach to learn heuristics for guiding the code transformation process of a rule-based program transformation system [17, 16]. This type of systems pose a number of problems, such as state explosion, arising from the application of transformation rules in an arbitrary order or the definition of a stop criteria. For the latter we have proposed the use of classification trees and for the former we have proposed a novel approach based on reinforcement learning. We have also performed a preliminary evaluation of the approach, which provided promising results that demonstrate the suitability of the approach for this type of transformation systems. As future work, we plan to continue expanding the set of use cases for training the different machine learning techniques used. As a consequence, we expect to enrich the code features identified to obtain program abstractions in order to capture new code patterns. New transformation sequences for the learning set can be automatically synthesized by swapping commutative rules in existing sequences, doing a sort of inverse of the partial order reduction used in model checking algorithms. While this may enrich the training set, it may also introduce unwanted noise, and therefore we need to evaluate this possibility carefully. We also expect that the increase in the training set will result in better prediction outcomes, but it will also increase the complexity of efficiently using the learning techniques. In the case of reinforcement learning, having more learning cases results in a bigger state space. For that reason we foresee using common methods to reduce the number of states like clustering or principal component analysis techniques. We also plan to explore other features of reinforcement learning like the use of different learning rates for different states or transformation sequences in order to learn and converge faster towards transformed codes providing final versions with better performance. We also plan to define multi-objective reinforcement learning rewards based on different non-functional properties like energy consumption, resource usage, etc. and even a combination of them. This future line would make it possible to define transformation strategies that try to optimize different aspects and generate, for example, a final code among the fastest ones that consumes the least amount of energy. 66 Towards Automatic Learning of Heuristics for Mechanical Transformations of Procedural Code Also, the introduction of weights into the multi-objective rewards would offer programmers the flexibility to select which non-functional property or set of properties they want to focus on for generating the final code. References [1] F. Agakov, E. Bonilla, J. Cavazos, B. Franke, G. Fursin, M. F. P. O’Boyle, J. Thomson, M. Toussaint & C. K. I. Williams (2006): Using Machine Learning to Focus Iterative Optimization. In: Proceedings of the International Symposium on Code Generation and Optimization, CGO ’06, IEEE Computer Society, Washington, DC, USA, pp. 295–305, doi:10.1109/CGO.2006.37. [2] Otto Skrove Bagge, Karl Trygve Kalleberg, Eelco Visser & Magne Haveraaen (2003): Design of the CodeBoost Transformation System for Domain-Specific Optimisation of C++ Programs. In: Third International Workshop on Source Code Analysis and Manipulation (SCAM 2003), IEEE, pp. 65–75, doi:10.1109/SCAM.2003.1238032. [3] Ashley Brown, Wayne Luk & Paul Kelly (2005): Optimising Transformations for Hardware Compilation. Technical Report, Imperial College London. Available at http://www.ashleybrown.co.uk/wordpress/ wp-content/uploads/2010/09/cobble-cml-report.pdf. [4] Edoardo Di Napoli, Diego Fabregat-Traver, Gregorio Quintana-Orti & Paolo Bientinesi (2014): Towards an Efficient Use of the BLAS Library for Multilinear Tensor Contractions. Applied Mathematics and Computation 235, pp. 454–468, doi:10.1016/j.amc.2014.02.051. [5] Diego Fabregat-Traver & Paolo Bientinesi (2013): Application-Tailored Linear Algebra Algorithms: A Search-Based Approach. International Journal of High Performance Computing Applications (IJHPCA) 27(4), pp. 425 – 438, doi:10.1177/1094342013494428. [6] Franz Franchetti, Yevgen Voronenko & Markus Püschel (2006): FFT program generation for shared memory: SMP and multicore. In: Proceedings of the ACM/IEEE SC2006 Conference on High Performance Networking and Computing, November 11-17, 2006, Tampa, FL, USA, p. 115, doi:10.1145/1188455.1188575. [7] Leslie Pack Kaelbling, Michael L. Littman & Andrew P. Moore (1996): Reinforcement Learning: A Survey. Journal of Artificial Intelligence Research 4, pp. 237–285, doi:10.1613/jair.301. [8] Jonathan G Koomey (2008): Worldwide Electricity Used in Data Centers. Environmental Research Letters 3(3), p. 034008, doi:10.1088/1748-9326/3/3/034008. [9] Giovanni Mariani, Gianluca Palermo, Roel Meeuws, Vlad Mihai Sima, Cristina Silvano & Koen Bertels (2014): DRuiD: Designing reconfigurable architectures with decision-making support. In: 19th Asia and South Pacific Design Automation Conference, ASP-DAC 2014, Singapore, January 20-23, 2014, pp. 213– 218, doi:10.1109/ASPDAC.2014.6742892. [10] Stephen Marsland (2009): Machine Learning: An Algorithmic Perspective, 1st edition. Chapman & Hall/CRC, doi:10.1111/j.1751-5823.2010.00118 11.x. [11] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot & E. Duchesnay (2011): Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12, pp. 2825– 2830. [12] Gennady Pekhimenko & AngelaDemke Brown (2010): Efficient Program Compilation Through Machine Learning Techniques. In Ken Naono, Keita Teranishi, John Cavazos & Reiji Suda, editors: Software Automatic Tuning, Springer New York, pp. 335–351, doi:10.1007/978-1-4419-6935-4 19. [13] Tom Schaul, Justin Bayer, Daan Wierstra, Yi Sun, Martin Felder, Frank Sehnke, Thomas Rückstieß & Jürgen Schmidhuber (2010): PyBrain. Journal of Machine Learning Research, doi:10.1145/1756006.1756030. G. Vigueras, M. Carro, S. Tamarit & J. Mariño, 67 [14] Sibylle Schupp, Douglas Gregor, David Musser & Shin-Ming Liu (2002): Semantic and behavioral library transformations. Information and Software Technology 44(13), pp. 797–810, doi:10.1016/S09505849(02)00122-2. [15] Richard M. Stallman & GCC Developer Community (2009): Using The Gnu Compiler Collection: A Gnu Manual For Gcc Version 4.3.3. CreateSpace, Paramount, CA. [16] Salvador Tamarit, Julio Mariño, Guillermo Vigueras & Manuel Carro (2016): Towards a Semantics-Aware Code Transformation Toolchain for Heterogeneous Systems. In Alicia Villanueva, editor: Proceedings of XIV Jornadas sobre Programación y Lenguajes (PROLE 2016), pp. 17–32. Available at http://hdl.handle. net/11705/PROLE/2016/014. [17] Salvador Tamarit, Guillermo Vigueras, Manuel Carro & Julio Mariño (2015): A Haskell Implementation of a Rule-Based Program Transformation for C Programs. In Enrico Pontelli & Tran Cao Son, editors: International Symposium on Practical Aspects of Declarative Languages, LNCS 9131, Springer-Verlag, pp. 105–114, doi:10.1007/978-3-319-19686-2 8.
6cs.PL
THE GARDEN OF EDEN THEOREM: OLD AND NEW TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT arXiv:1707.08898v2 [math.DS] 29 Sep 2017 Abstract. We review topics in the theory of cellular automata and dynamical systems that are related to the Moore-Myhill Garden of Eden theorem. Contents 1. Introduction 2. Configuration spaces and shifts 2.1. Notation 2.2. Configurations spaces 2.3. Group actions 2.4. Shifts 2.5. Patterns 2.6. Subshifts 3. Cellular automata 3.1. Definition 3.2. The Curtis-Hedlund-Lyndon theorem 3.3. Operations on cellular automata 3.4. Surjectivity of cellular automata, GOE configurations, and GOE patterns 3.5. Pre-injectivity of cellular automata and mutually erasable patterns 4. The Garden of Eden theorem for Zd 5. The Garden of Eden theorem for general amenable groups 5.1. Amenability 5.2. Entropy 5.3. Tilings 5.4. The Garden of Eden theorem for amenable groups 6. Failure of the Garden of Eden theorem for nonamenable groups 7. The Garden of Eden theorem for subshifts 7.1. Strongly irreducible subshifts 7.2. The Moore and the Myhill properties for subshifts 8. Garden of Eden theorems for other dynamical systems 8.1. Dynamical systems 2 3 3 4 4 5 5 6 7 7 9 10 10 11 12 15 15 19 22 24 26 27 27 28 30 30 Date: October 2, 2017. 2010 Mathematics Subject Classification. 37B15, 37B10, 37B40, 37C29, 37D20, 43A07, 68Q80. Key words and phrases. Garden of Eden theorem, cellular automaton, mutually erasable patterns, dynamical system, homoclinicity, pre-injectivity, surjunctivity, amenability, soficity. 1 2 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT 8.2. Expansiveness 8.3. Homoclinicity 8.4. The Moore and the Myhill properties for dynamical systems 8.5. Anosov diffeomorphisms 8.6. Principal algebraic dynamical systems 9. Some Additional topics 9.1. Infinite alphabets and uncountable groups 9.2. Linear cellular automata 9.3. Gromov’s Garden of Eden theorem 9.4. Cellular automata over homogeneous sets References Added in proof Additional references 31 32 32 33 34 36 36 37 39 40 40 42 43 1. Introduction In the beginning, the Garden of Eden theorem, also known as the Moore-Myhill theorem, is a result in the theory of cellular automata which states that a cellular automaton is surjective if and only if it satisfies a weak form of injectivity, called pre-injectivity. The theorem was obtained by Moore and Myhill in the early 1960s for cellular automata with finite alphabet over the groups Zd . The fact that surjectivity implies pre-injectivity for such cellular automata was first proved by Moore in [39], and Myhill [43] established the converse implication shortly after. The proofs of Moore and Myhill appeared in two separate papers both published in 1963. The biblical terminology used to designate the Moore-Myhill theorem comes from the fact that configurations that are not in the image of a cellular automaton are called Garden of Eden configurations because, when considering the sequence of consecutive iterates of the cellular automaton applied to the set of configurations, they can only occur at time 0. Surjectivity of a cellular automaton is equivalent to absence of Garden of Eden configurations. In 1988 [49, Question 1], Schupp asked whether the class of groups for which the Garden of Eden theorem remains valid is precisely the class of virtually nilpotent groups. By a celebrated result of Gromov [29], a finitely generated group is virtually nilpotent if and only if it has polynomial growth. In 1993, Machı̀ and Mignosi [36] proved that the Garden of Eden theorem is still valid over any finitely generated group with subexponential growth. As Grigorchuk [28], answering a longstanding open question raised by Milnor [38], gave examples of groups whose growth lies strictly between polynomial and exponential, it follows that the class of finitely generated groups satisfying the Garden of Eden theorem is larger than the class of finitely generated virtually nilpotent groups. Actually, it is even larger than the class of finitely generated groups with subexponential growth. Indeed, Machı̀, Scarabotti, and the first author [20] proved in 1999 that every amenable group satisfies the Garden of Eden theorem and it is a well known fact that there are finitely generated amenable groups, such as the solvable Baumslag-Solitar group BS(1, 2) = ha, b : aba−1 = b2 i, that are amenable THE GARDEN OF EDEN THEOREM: OLD AND NEW 3 and have exponential growth. It was finally shown that the class of groups that satisfy the Garden of Eden theorem is precisely the class of amenable groups. This is a consequence of recent results of Bartholdi [2], Bartholdi and Kielak [3], who showed that none of the implications of the Garden of Eden theorem holds when the group is nonamenable. In [30, Section 8], Gromov made an important contribution to the subject by providing a deep analysis of the role played by entropy in the proof of the Garden of Eden theorem and indicating new directions for extending it in many other interesting settings. He mentioned in particular [30, p 195] the possibility of proving an analogue of the Garden of Eden theorem for a suitable class of hyperbolic dynamical systems. Some results in that direction were subsequently obtained by the authors in [17], [16], and [18]. In particular, a version of the Garden of Eden theorem was established for Anosov diffeomorphisms on tori in [17] and for principal expansive algebraic actions of countable abelian groups in [18]. The present article is intended as a reasonably self-contained survey on the classical Garden of Eden theorem and some of its generalizations. Almost all results presented here have already appeared in the literature elsewhere but we sometimes give complete proofs when we feel they might be helpful to the reader. The general theory of cellular automata over groups is developed in our monograph [12]. The present survey is a kind of complement to our book since for instance cellular automata between subshifts are not considered in [12] while they are treated here. The paper is organized as follows. Configuration spaces and shifts are presented in Section 2. Cellular automata are introduced in Section 3. Section 4 contains the proof of the Garden of Eden theorem in the case G = Zd following Moore and Myhill. The proof of the Garden of Eden theorem in the case of an arbitrary countable amenable group is given in Section 5. Examples of cellular automata that do not satisfy the Garden of Eden theorem for groups containing nonabelian free subgroups are described in Section 6. We also discuss the results of Bartholdi and Kielak mentioned above, which, together with the Garden of Eden theorem, lead to characterizations of amenability in terms of cellular automata. Extensions of the Garden of Eden theorem to certain classes of subshifts are reviewed in Section 7. In Section 8, we present versions of the Garden of Eden theorem we obtained for certain classes of dynamical systems. The final section briefly discusses some additional topics and provides references for further readings. 2. Configuration spaces and shifts 2.1. Notation. We use the symbol Z to denote the set of integers {. . . , −2, −1, 0, 1, 2, . . . }. The symbol N denotes the set of nonnegative integers {0, 1, 2, . . . }. The cardinality of a finite set X is written |X|. We use multiplicative notation for groups except for abelian groups such as Zd = Z × Z × · · · × Z | {z } d times for which we generally prefer additive notation. 4 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Let G be a group. We denote the identity element of G by 1G . If A, B are subsets of G and g ∈ G, we write AB := {ab : a ∈ A, b ∈ B}, A−1 := {a−1 : a ∈ A}, gA := {g}A and Ag := A{g}. A subset A ⊂ G is said to be symmetric if it satisfies A = A−1 .” 2.2. Configurations spaces. Let U be a countable set, called the universe, and A a finite set, called the alphabet. Depending on the context, the elements of A are called letters, or symbols, or states, or colors. As usual, we denote by AU the set consisting of all maps x : U → A. An element of AU is called a configuration of the universe U. Thus, a configuration is a way of attaching a letter of the alphabet to each element of the universe. If x ∈ AU is a configuration and V ⊂ U, we shall write x|V for the restriction of x to V, i.e., the element x|V ∈ AV defined by x|V (v) = x(v) for all v ∈ V. If X ⊂ AU , we shall write (2.1) XV := {x|V : x ∈ X} ⊂ AV . Two configurations x, y ∈ AU are said to be almost equal if they coincide outside of a finite set, i.e., there is a finite subset Ω ⊂ U such that x|U \Ω = y|U \Ω . Being almost equal clearly defines an equivalence relation on AU . We equip the configuration set AU with its prodiscrete topology, that is, the Q product topology obtained by taking the discrete topology on each factor A of AU = u∈U A. A neighborhood base of a configuration x ∈ AU is given by the sets (2.2) V (x, Ω) = V (x, Ω, U, A) := {y ∈ AU : x|Ω = y|Ω }, where Ω runs over all finite subsets of U. In this topology, two configurations are “close” if they coincide on a “large” finite subset of the universe. Every finite discrete topological space is compact, totally disconnected, and metrizable. As a product of compact (resp. totally disconnected) topological spaces is itself compact (resp. totally disconnected) and a countable product of metrizable spaces is itself metrizable, it follows that AU is a compact totally disconnected metrizable space. Note that AU is homeomorphic to the Cantor set as soon as A contains more than one element and U is infinite. 2.3. Group actions. An action of a group G on a set X is a map α : G × X → X satisfying α(g1, α(g2 , x)) = α(g1 g2 , x) and α(1G , x) = x for all g1 , g2 ∈ G and x ∈ X. In the sequel, if α is an action of a group G on a set X, we shall simply write gx instead of α(g, x), if there is no risk of confusion. Suppose that a group G acts on a set X. The orbit of a point x ∈ X is the subset Gx ⊂ X defined by Gx := {gx : g ∈ G}. A point x ∈ X is called periodic if its orbit is finite. A subset Y ⊂ X is called invariant if Gy ⊂ Y for all y ∈ Y . One says that Y ⊂ X is fixed by G if gy = y for all g ∈ G and y ∈ Y . Suppose now that a group G acts on two sets X and Y . A map f : X → Y is called equivariant if f (gx) = gf (x) for all g ∈ G and x ∈ X. Let X be a topological space. An action of a group G on X is called continuous if the map x 7→ gx is continuous on X for each g ∈ G. Note that if G acts continuously on X then, for each g ∈ G, the map x 7→ gx is a homeomorphism of X with inverse x 7→ g −1 x. THE GARDEN OF EDEN THEOREM: OLD AND NEW 5 2.4. Shifts. From now on, our universe will be a countable group. So let G be a countable group and A a finite set. Given an element g ∈ G and a configuration x ∈ AG , we define the configuration gx ∈ AG by gx := x ◦ Lg−1 , where Lg : G → G is the left-multiplication by g. Thus gx(h) = x(g −1 h) for all h ∈ G. Observe that, for all g1 , g2 ∈ G and x ∈ AG , g1 (g2 x) = x ◦ Lg2−1 ◦ Lg1−1 = x ◦ Lg2−1 g1−1 = x ◦ L(g1 g2 )−1 = (g1 g2 )x, and Therefore the map 1G x = x ◦ L1G = x ◦ IdG = x. G × AG → AG (g, x) 7→ gx defines an action of G on AG . This action is called the G-shift, or simply the shift, on AG . Observe that if two configurations x, y ∈ AG coincide on a subset Ω ⊂ G, then, for every g ∈ G, the configurations gx and gy coincide on gΩ. As the sets V (x, Ω) defined by (2.2) form a base of neighborhoods of x ∈ AG when Ω runs over all finite subsets of G, we deduce that the map x 7→ gx is continuous on AG for each g ∈ G. Thus, the shift action of G on AG is continuous. 2.5. Patterns. A pattern is a map p : Ω → A, where Ω is a finite subset of G. If p : Ω → A is a pattern, we say that Ω is the support of p and write Ω = supp(p). Let P(G, A) denote the set of all patterns. There is a natural action of the group G on P(G, A) defined as follows. Given g ∈ G and a pattern p ∈ P(G, A) with support Ω, we define the pattern gp ∈ P(G, A) as being the pattern with support gΩ such that gp(h) = p(g −1h) for all h ∈ gΩ. It is easy to check that this defines an action of G on P(G, A), i.e., g1 (g2 p) = (g1 g2 )p and 1G p = p for all g1 , g2 ∈ G and p ∈ P(G, A). Observe that supp(gp) = g supp(p) for all g ∈ G and p ∈ P(G, A). Note also that if p is the restriction of a configuration x ∈ AG to a finite subset Ω ⊂ G, then gp is the restriction of the configuration gx to gΩ. Example 2.1. Take G = Z and let A be a finite set. Denote by A⋆ the set of words on the alphabet A. We recall that A⋆ is the free monoid based on A and that any element w ∈ A⋆ can be uniquely written in the form w = a1 a2 · · · an , where ai ∈ A for 1 ≤ i ≤ n and n ∈ N is the length of the word w. The monoid operation on A⋆ is the concatenation of words and the identity element is the empty word, that is, the unique word with length 0. Now let us fix some finite interval Ω ⊂ Z of cardinality n, say Ω = {m, m + 1, . . . , m + n − 1} with m ∈ Z and n ∈ N. Then one can associate to each pattern p : Ω → A the word w = p(m)p(m + 1) · · · p(m + n − 1) ∈ A⋆ . 6 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT This yields a one-to-one correspondence between the patterns supported by Ω and the words of length n on the alphabet A. This is frequently used to identify each pattern supported by Ω with the corresponding word. 2.6. Subshifts. A subshift is a subset X ⊂ AG that is invariant under the G-shift and closed for the prodiscrete topology on AG . Example 2.2. Take G = Z and A = {0, 1}. Then the subset X ⊂ AG , consisting of all x : Z → {0, 1} such that (x(n), x(n + 1)) 6= (1, 1) for all n ∈ Z, is a subshift. This subshift is called the golden mean subshift. Example 2.3. Take G = Zd and A = {0, 1}. Then the subset X ⊂ AG , consisting of all x ∈ AG such that (x(g), x(g + ei )) 6= (1, 1) for all g ∈ G, where (ei )1≤i≤d is the canonical basis of Zd , is a subshift. This subshift is called the hard-ball model. For d = 1, the hard-ball model is the golden mean subshift of the previous example. Example 2.4. Take G = Z and A = {0, 1}. Then the subset X ⊂ AG , consisting of all bi-infinite sequences x : Z → {0, 1} such that there is always an even number of 0s between two 1s, is a subshift. This subshift is called the even subshift. Example 2.5. Take G = Z2 and A = {0, 1} = Z/2Z (the integers modulo 2). Then the subset X ⊂ AG , consisting of all x : Z2 → {0, 1} such that x(m, n) + x(m + 1, n) + x(m, n + 1) = 0 for all (m, n) ∈ Z2 , is a subshift. This subshift is called the Ledrappier subshift. Remark 2.6. Every intersection of subshifts and every finite union of subshifts X ⊂ AG is itself a subshift. Therefore the subshifts X ⊂ AG are the closed subsets of a topology on AG . This topology is coarser (it has less open sets) than the prodiscrete topology on AG . It is not Hausdorff as soon as G is not trivial and A has more than one element. Given a (possibly infinite) subset of patterns P ⊂ P(G, A), it is easy to see that the subset X(P ) ⊂ AG defined by X(P ) := {x ∈ AG such that (gx)|supp(p) 6= p for all g ∈ G and p ∈ P } is a subshift. Conversely, let X ⊂ AG be a subshift. One says that a pattern p ∈ P(G, A) appears in X if p ∈ Xsupp(p) , i.e., if there is a configuration x ∈ X such that x|supp(p) = p. Then one easily checks that X = X(P ) for P := {p ∈ P(G, A) such that p does not appear in X}. One says that a subshift X ⊂ AG is of finite type if there exists a finite subset P ⊂ P(G, A) such that X = X(P ). The hard-ball models (and hence in particular the golden mean subshift) and the Ledrappier subshifts are examples of subshifts of finite type. On the other hand, the even subshift is not of finite type. THE GARDEN OF EDEN THEOREM: OLD AND NEW 7 3. Cellular automata 3.1. Definition. Let G be a countable group and let A, B be finite sets. Suppose that X ⊂ AG and Y ⊂ B G are two subshifts. Definition 3.1. One says that a map τ : X → Y is a cellular automaton if there exist a finite subset S ⊂ G and a map µ : AS → B such that (3.1) τ (x)(g) = µ((g −1x)|S ) for all x ∈ X and g ∈ G, where we recall that (g −1 x)|S denotes the restriction of the configuration g −1 x ∈ X to S. Such a set S is called a memory set and µ is called a local defining map for τ . It immediately follows from this definition that a map τ : X → Y is a cellular automaton if and only if τ extends to a cellular automaton τe : AG → B G . Observe also that if S is a memory set for a cellular automaton τ : X → Y and g ∈ G, then Formula (3.1) implies that the value taken by the configuration τ (x) at g only depends on the restriction of x to gS. Finally note that if S is a memory set for a cellular automaton τ , then any finite subset of G containing S is also a memory set for τ . Consequently, the memory set of a cellular automaton is not unique in general. However, every cellular automaton admits a unique memory set with minimal cardinality (this follows from the fact that if S1 and S2 are memory sets then so is S1 ∩ S2 ). Example 3.2. Take G = Z and A = {0, 1} = Z/2Z. Then the map τ : AG → AG , defined by τ (x)(n) := x(n + 1) + x(n) G for all x ∈ A and n ∈ Z, is a cellular automaton admitting S := {0, 1} ⊂ Z as a memory set and µ : AS → A given by µ(p) := p(0) + p(1) S for all p ∈ A , as a local defining map. Using the representation of patterns with support S by words of length 2 on the alphabet A (cf. Example 2.1), the map µ is given by µ(00) = µ(11) = 0 and µ(01) = µ(10) = 1. Example 3.3 (Majority vote). Take G = Z and A = {0, 1}. Then the map τ : AG → AG , defined by ( 0 if x(n − 1) + x(n) + x(n + 1) ≤ 1 τ (x)(n) := 1 otherwise for all x ∈ AG and n ∈ Z, is a cellular automaton admitting S := {−1, 0, 1} ⊂ Z as a memory set and µ : AS → A given by µ(000) = µ(001) = µ(010) = µ(100) = 0 and µ(011) = µ(101) = µ(110) = µ(111) = 1 as local defining map. The cellular automaton τ is called the majority vote cellular automaton. 8 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Remark 3.4. A cellular automaton τ : AG → AG , where G = Z, A = {0, 1}, admitting S = {−1, 0, 1} as a memory set is called an elementary cellular automaton. Each one of these cellular automata is uniquely determined by its local defining map µ : AS → A, so that there are exactly 28 = 256 elementary cellular automata. They are numbered from 0 to 255 according to a notation that was introduced by Wolfram (cf. [56]). To obtain the number n of an elementary cellular automaton τ , one proceeds as follows. One first lists all eight patterns p ∈ AS in increasing order from 000 to 111. The number n is the integer whose expansion in base 2 is a8 a7 . . . a1 , where ak is the value taken by the local defining map of τ at the k-th pattern in the list. One also says that τ is Rule n. For instance, the elementary cellular automaton described in Example 3.2 is Rule 102 while the one described in Example 3.3 is Rule 232. Example 3.5. Let G be a countable group, A a finite set, and X ⊂ AG a subshift. Then the identity map IdX : X → X is a cellular automaton with memory set S = {1G } and local defining map µ = IdA : AS = A{1G } = A → A. Example 3.6. Let G be a countable group, A a finite set, and X ⊂ AG a subshift. Let s ∈ G and denote by Rs the right-multiplication by s, that is, the map Rs : G → G defined by Rs (h) := hs for all s ∈ G. Then the subset Y ⊂ AG defined by Y := {x ◦ Rs such that x ∈ X} is a subshift. Moreover, the map τ : X → Y , defined by τ (x) := x ◦ Rs for all x ∈ X, is a cellular automaton with memory set S = {s} and local defining map µ = IdA : AS = A → A. Observe that if s is in the center of G, then X = Y and τ : X → X is the shift map x 7→ s−1 x. Example 3.7. Take G = Z and A = {0, 1}. Let X ⊂ AG and Y ⊂ AG denote respectively the golden mean subshift and the even subshift. For x ∈ X, define τ (x) ∈ AG by ( 0 if (x(n), x(n + 1)) = (0, 1) or (1, 0) τ (x)(n) := 1 if (x(n), x(n + 1)) = (0, 0) for all n ∈ Z. It is easy to see that τ (x) ∈ Y for all x ∈ X. The map τ : X → Y is a cellular automaton admitting S := {0, 1} ⊂ Z as a memory set and the map µ : AS → A, defined by µ(00) = µ(11) = 1 and µ(01) = µ(10) = 0. Note that the map µ′ : AS → A, defined by µ′ (00) = 1 and µ′ (01) = µ′ (10) = µ′ (11) = 0 is also a local defining map for τ . Thus, τ is the restriction to X of both Rule 153 and Rule 17. THE GARDEN OF EDEN THEOREM: OLD AND NEW 9 3.2. The Curtis-Hedlund-Lyndon theorem. The definition of a cellular automaton given in the previous subsection is a local one. For τ to be a cellular automaton, it requires the existence of a rule, commuting with the shift, that allows one to evaluate the value taken by τ (x) at g ∈ G by applying the rule to the restriction of x to a certain finite set, namely the left-translate by g of a memory set of the automaton. The following result, known as the Curtis-Lyndon-Hedlund theorem (see [31]), yields a global characterization of cellular automata involving only the shift actions and the prodiscrete topology on the configuration spaces. Theorem 3.8. Let G be a countable group and let A, B be finite sets. Let τ : X → Y be a map from a subshift X ⊂ AG into a subshift Y ⊂ B G . Then the following conditions are equivalent: (a) τ is a cellular automaton; (b) τ is equivariant (with respect to the shift actions of G) and continuous (with respect to the prodiscrete topologies). Proof. Suppose first that τ : X → X is a cellular automaton. Let S ⊂ G be a memory set and µ : AS → B a local defining map for τ . For all g, h ∈ G and x ∈ X, we have that τ (gx)(h) = µ((h−1 gx)|S ) −1 (by Formula (3.1)) −1 = µ(((g h) x)|S ) = τ (x)(g −1 h) = gτ (x)(h). Thus τ (gx) = gτ (x) for all g ∈ G and x ∈ X. This shows that τ is equivariant. Now let Ω be a finite subset of G. Recall that Formula (3.1) implies that if two configurations x, y ∈ X coincide on gS for some g ∈ G, then τ (x)(g) = τ (y)(g). Therefore, if the configurations x and y coincide on the finite set ΩS, then τ (x) and τ (y) coincide on Ω. It follows that τ (X ∩ V (x, ΩS, G, A)) ⊂ V (τ (x), Ω, G, B). This implies that τ is continuous. Thus (a) implies (b). Conversely, suppose now that the map τ : X → Y is equivariant and continuous. Let us show that τ is a cellular automaton. As the map ϕ : X → B defined by ϕ(x) := τ (x)(1G ) is continuous, we can find, for each x ∈ X, a finite subset Ωx ⊂ G such that if y ∈ X ∩ V (x, Ωx , G, A), then τ (y)(1G ) = τ (x)(1G ). The sets X ∩ V (x, Ωx , G, A) form an open cover of X. As X is compact, there is a finite subset F ⊂ X such that the sets V (x, Ωx , G, A), x ∈ F , cover X. Let us set S = ∪x∈F Ωx and suppose that two configurations y, z ∈ X coincide on S. Let x0 ∈ F be such that y ∈ V (x0 , Ωx0 , G, A), that is, y|Ωx0 = x0 |Ωx0 . As Ωx0 ⊂ S, we have that y|Ωx0 = z|Ωx0 and therefore τ (y)(1G ) = τ (x0 )(1G ) = τ (z)(1G ). We deduce that there exists a map µ : AS → B such that τ (x)(1G ) = µ(x|S ) for all x ∈ X. 10 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Now, for all x ∈ X and g ∈ G, we have that τ (x)(g) = (g −1 τ (x))(1G ) = τ (g −1 x)(1G ) (by definition of the shift action on B G ) (since τ is equivariant) −1 = µ((g x)|S ). This shows that τ is a cellular automaton with memory set S and local defining map µ. Thus (b) implies (a).  3.3. Operations on cellular automata. Proposition 3.9. Let G be a countable group and les A, B, C be finite sets. Suppose that X ⊂ AG , Y ⊂ B G , Z ⊂ C G are subshifts and that τ : X → Y , σ : Y → Z are cellular automata. Then the composite map σ ◦ τ : X → Z is a cellular automaton. Proof. This is an immediate consequence of the characterization of cellular automata given by the Curtis-Hedlund-Lyndon theorem (cf. Theorem 3.8) since the composite of two equivariant (resp. continuous) maps is itself equivariant (resp. continuous).  Remark 3.10. If we fix the countable group G, we deduce from Proposition 3.9 and Example 3.5 that the subshifts X ⊂ AG , with A finite, are the objects of a concrete category CG in which the set of morphisms from X ∈ CG to Y ∈ CG consist of all cellular automata τ : X → Y (cf. [15, Section 3.2]). In this category, the endomorphisms of X ∈ CG consist of all cellular automata τ : X → X and they form a monoid for the composition of maps. Proposition 3.11. Let G be a countable group and let A, B be finite sets. Suppose that X ⊂ AG , Y ⊂ B G are subshifts and that τ : X → Y is a bijective cellular automaton. Then the inverse map τ −1 : Y → X is a cellular automaton. Proof. This is again an immediate consequence of the Curtis-Hedlund-Lyndon theorem since the inverse of a bijective equivariant map is an equivariant map and the inverse of a bijective continuous map between compact Hausdorff spaces is continuous.  3.4. Surjectivity of cellular automata, GOE configurations, and GOE patterns. In what follows, we keep the notation introduced for defining cellular automata. Let τ : X → Y be a cellular automaton. A configuration y ∈ Y is called a Garden of Eden configuration for τ , briefly a GOE configuration, if it does not belong to the image of τ , i.e., there is no x ∈ X such that y = τ (x). One says that a pattern p ∈ P(G, B) is a Garden of Eden pattern for τ , briefly a GOE pattern, if the pattern p appears in the subshift Y but not in the subshift τ (X), i.e., there exists y ∈ Y such that p = y|supp(p) but there is no x ∈ X such that p = τ (x)|supp(p) . Note that the set of GOE configurations (resp. of GOE patterns) is an invariant subset of Y (resp. of P(G, B)). Observe also that if p ∈ P(G, A) is a GOE pattern for the cellular automaton τ : X → Y , then every configuration y ∈ Y such that y|supp(p) = p is a GOE configuration for τ . THE GARDEN OF EDEN THEOREM: OLD AND NEW 11 Example 3.12. It is easy to check that the pattern with support Ω := {0, 1, 2, 3, 4} associated with the word 01001 is a GOE pattern for the majority vote cellular automaton described in Example 3.3. Proposition 3.13. Let τ : X → Y be a cellular automaton. Then the following conditions are equivalent: (a) τ is surjective; (b) τ admits no GOE configurations; (c) τ admits no GOE patterns. Proof. The equivalence of (a) and (b) as well as the implication (b) =⇒ (c) are trivial. The implication (c) =⇒ (b) easily follows from the compactness of τ (X).  3.5. Pre-injectivity of cellular automata and mutually erasable patterns. Let G be a countable group and let A and B be finite sets. Recall that two configurations x, y ∈ AG are called almost equal if they coincide outside of a finite subset of G. Definition 3.14. Let X ⊂ AG and Y ⊂ B G be subshifts. One says that a cellular automaton τ : X → Y is pre-injective if there are no distinct configurations x1 , x2 ∈ X that are almost equal and satisfy τ (x1 ) = τ (x2 ). A pair of configurations (x1 , x2 ) ∈ X × X is called a diamond if x1 and x2 are distinct, almost equal, and have the same image under τ (cf. [32, Definition 8.1.15]). Thus, preinjectivity is equivalent to the absence of diamonds. If (x1 , x2 ) is a diamond, the nonempty finite subset {g ∈ G : x1 (g) 6= x2 (g)} ⊂ G is called the support of the diamond (x1 , x2 ). Every injective cellular automaton is clearly pre-injective. The converse is false, as shown by the following examples. Example 3.15. Take G = Z, A = {0, 1} = Z/2Z, and consider the cellular automaton τ : AG → AG described in Example 3.2 (Rule 102 in Wolfram’s notation). Then τ is preinjective. Indeed, it is clear that if two configurations x, x′ ∈ AG coincide on Z ∩ (−∞, n0 ] for some n0 ∈ Z and satisfy τ (x) = τ (x′ ) then x = x′ . However, τ is not injective since the two constant configurations have the same image. Example 3.16. Consider the cellular automaton τ : X → Y from the golden mean subshift to the even subshift described in Example 3.7. It is easy to see that τ is pre-injective by an argument similar to the one used in the previous example. However, τ is not injective since the two sequences in X with exact period 2 have the same image under τ , namely the constant sequence with only 0s. Let Ω ⊂ G be a finite set and p1 , p2 ∈ XΩ two patterns with support Ω appearing in X. One says that the patterns p1 and p2 are mutually erasable with respect to τ , briefly ME, provided the following hold: 12 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT (MEP-1) the set Xp1 ,p2 := {(x1 , x2 ) ∈ X × X : x1 |Ω = p1 , x2 |Ω = p2 and x1 |G\Ω = x2 |G\Ω } is nonempty; (MEP-2) for all (x1 , x2 ) ∈ Xp1 ,p2 one has τ (x1 ) = τ (x2 ). Note that “being ME” is an equivalence relation on XΩ . This equivalence relation is not trivial in general. Example 3.17. The patterns with support Ω := {0, 1, 2} associated with the words 00000 and 00100 are ME patterns for the majority vote cellular automaton described in Example 3.3. Observe that if the patterns p1 and p2 are ME, then so are gp1 and gp2 for all g ∈ G. Proposition 3.18. Let τ : X → Y be a cellular automaton. Then the following conditions are equivalent: (a) τ is pre-injective; (b) τ admits no distinct ME patterns. Proof. Suppose first that τ admits two distinct ME patterns p1 and p2 . Let Ω denote their common support. Take (x1 , x2 ) ∈ Xp1 ,p2 . Then the configurations x1 and x2 are almost equal since they coincide outside of Ω. Moreover, they satisfy x1 6= x2 and τ (x1 ) = τ (x2 ). Therefore (x1 , x2 ) is a diamond for τ . It follows that τ is not pre-injective. Suppose now that τ is not pre-injective and let us show that τ admits two distinct ME patterns. By definition, τ admits a diamond (x1 , x2 ) ∈ X × X. Let ∆ denote the support of this diamond and let S ⊂ G be a memory set for τ with 1G ∈ S. Consider the set Ω := ∆S −1 S. We claim that the patterns p1 := x1 |Ω and p2 := x2 |Ω are distinct ME patterns. First observe that p1 6= p2 since ∅ 6= ∆ ⊂ Ω. Moreover, Xp1 ,p2 is nonempty since, by construction, (x1 , x2 ) ∈ Xp1 ,p2 . To complete the proof, we only need to show the following: if (y1 , y2) ∈ Xp1 ,p2 then τ (y1 ) = τ (y2 ). Let g ∈ G. Suppose first that g ∈ G \ ∆S −1 . Then gS ∩ ∆ = ∅. Since y1 and y2 coincide on G \ ∆, we deduce that (3.2) τ (y1 )(g) = τ (y2 )(g) for all g ∈ G \ ∆S −1 . Suppose now that g ∈ ∆S −1 . Then gS ⊂ ∆S −1 S = Ω. As y1 |Ω = p1 = x1 |Ω (resp. y2 |Ω = p2 = x2 |Ω ), by construction, we deduce that (3.3) τ (y1 )(g) = τ (x1 )(g) = τ (x2 )(g) = τ (y2 )(g) for all g ∈ ∆S −1 . From (3.2) and (3.3) we deduce that τ (y1 ) = τ (y2 ).  4. The Garden of Eden theorem for Zd In this section, we present a proof of the Garden of Eden theorem of Moore and Myhill for cellular automata over the group Zd . This is a particular case of the Garden of Eden theorem for amenable groups that will be established in the next section. d d Theorem 4.1. Let A be a finite set, d ≥ 1 an integer, and τ : AZ → AZ a cellular automaton. Then the following conditions are equivalent: THE GARDEN OF EDEN THEOREM: OLD AND NEW 13 (a) τ is surjective; (b) τ is pre-injective. The implication (a) =⇒ (b) is due to Moore [39] and the converse to Myhill [43]. We shall prove the contraposite of each implication. Before undertaking the proof of Theorem 4.1, let us first introduce some notation and establish some preliminary results. Let S ⊂ Zd be a memory set for τ . Since any finite subset of Zd containing a memory set for τ is itself a memory set for τ , it is not restrictive to suppose that S = {0, ±1, . . . , ±r}d for some integer r ≥ 1. Let us set, for each integer m ≥ 2r, Ωm := {0, 1, . . . , m − 1}d and d Ω+ m := {−r, −r + 1, . . . , m + r − 1} d Ω− m := {r, r + 1, . . . , m − r − 1} , so that (4.1) |Ωm | = md , d − d |Ω+ m | = (m + 2r) , and |Ωm | = (m − 2r) . d d − Observe that if two configurations x1 , x2 ∈ AZ coincide on Ωm (resp. Ω+ m , resp. Z \ Ωm ) d then τ (x1 ) and τ (x2 ) coincide on Ω− m (resp. Ωm , resp. Z \ Ωm ). Also, let us set, for all integers k, n ≥ 1 Tnk := {t = (t1 k, t2 k, . . . , td k) ∈ Zd : 0 ≤ tj ≤ n − 1}, and observe that the nd cubes t + Ωk , for t running over Tnk , form a partition of the cube Ωnk . Finally, we introduce the following additional notation. Given a finite subset Ω ⊂ G and p, q ∈ AΩ we write p ∼ q if and only if the patterns p and q are ME for τ . As usual, we denote by AΩ / ∼ the quotient set of AΩ by ∼, i.e., the set of all ME-equivalence classes of patterns supported by Ω. In the proof of Theorem 4.1 we shall make use of the following elementary result (here d one should think of a := |A| as to the cardinality of the alphabet set and a(nk) (resp. d a(nk−2r) , for nk ≥ 2r) as to the number of all patterns supported by Ωnk (resp. Ω− nk ), cf. (4.1). Lemma 4.2. Let a, k, d, r be positive integers with a ≥ 2. n0 (a, k, d, r) ∈ N such that  nd  d d k < a(nk−2r) (4.2) a −1 Then there exists n0 = for all n ≥ n0 . Proof. Taking logarithms to base a, Inequality (4.2) is equivalent to d    d 2r k (4.3) loga a − 1 < k − . n 14 Since TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT d   d  d  2r k d k , = k = lim k − loga a − 1 < loga a n→∞ n we deduce that there exists n0 ∈ N such that (4.3) and therefore (4.2) are satisfied for all n ≥ n0 .  Proof of Theorem 4.1. We can assume that a := |A| ≥ 2. Suppose first that τ is not pre-injective. Then, by Proposition 3.18, τ admits two distinct ME patterns, say p1 and p2 . Denote by Ω ⊂ Zd their common support. Since, for all t ∈ Zd , the patterns tp1 and tp2 (with support t+Ω) are also distinct and ME, and any finite subset of Zd containing the support of two distinct ME patterns is itself the support of two distinct ME patterns, we may assume that Ω = Ωk for some integer k ≥ 2r. As the patterns tp1 and tp2 are ME, we have that d |At+Ωk / ∼ | ≤ |At+Ωk | − 1 = ak − 1 (4.4) for all t ∈ Zd . Now observe that two patterns with support Ωnk are ME if their restrictions to t + Ωk are ME for every t ∈ Tnk . Using (4.4), we deduce that Y d d |AΩnk / ∼ | ≤ |At+Ωk / ∼ | ≤ (ak − 1)n . t∈Tnk Taking n ≥ n0 (a, k, d, r), we then get d d d d − τ (AZ )Ω− ≤ |AΩnk / ∼ | ≤ (ak − 1)n < a(nk−2r) = |AΩnk |. nk d − This implies that τ (AZ )Ω− $ AΩnk , so that there must exist a GOE pattern for τ with nk support Ω− nk . Consequently, τ is not surjective. This shows that (a) =⇒ (b). Let us now turn to the proof of the converse implication. Suppose that τ is not surjective. Then, by Proposition 3.13, there exists a GOE pattern p for τ . Since tp is a GOE pattern for every t ∈ Zd , and any finite subset of Zd containing the support of a GOE pattern is itself the support of a GOE pattern, we can assume that p is supported by the cube Ωk for some integer k ≥ 2r. Decompose again Ωnk into the nd translates t + Ωk , with t ∈ Tnk , and observe that tp ∈ At+Ωk is GOE for every t ∈ Tnk . Any pattern q ∈ AΩnk which is not GOE satisfies that q|t+Ωk is not GOE for every t ∈ Tnk . As a consequence, we have that Y d d d d τ (AZ )Ωnk ≤ (4.5) τ (AZ )t+Ωk ≤ (ak − 1)n . t∈Tnk Let us fix now some element a0 ∈ A and consider the set X consisting of all configurations d x ∈ AZ that satisfy x(g) = a0 for all g ∈ Zd \ Ω− nk . THE GARDEN OF EDEN THEOREM: OLD AND NEW 15 Observe that if x1 and x2 are in X, then τ (x1 ) and τ (x2 ) coincide on Zd \ Ωnk . It follows that |τ (X)| = |τ (X)Ωnk | . On the other hand, taking n ≥ n0 (a, k, d, r) and using (4.5), we get d d d d − |τ (X)Ωnk | ≤ τ (AZ )Ωnk ≤ (ak − 1)n < a(nk−2r) = |AΩkn | = |X| and hence |τ (X)| < |X|. By the pigeon-hole principle, this implies that there exist two distinct configurations x1 , x2 ∈ X such that τ (x1 ) = τ (x2 ). As all configurations in X are almost equal, we deduce that τ is not pre-injective. This shows that (b) =⇒ (a).  Remark 4.3. The proof of the implication (a) =⇒ (b) shows that if τ admits two distinct ME patterns supported by a cube of side k ≥ 2r, then a cube of side nk − 2r, with n ≥ n0 (a, k, d, r), must support a GOE pattern. Conversely, a small addition to the proof of the implication (b) =⇒ (a) yields that if τ admits a GOE pattern supported by a cube of side k ≥ 2r, then a cube of side kn + 2r, with n ≥ n0 (a, k, d, r), supports two distinct d ME patterns. Indeed, the proof shows the existence of two configurations x1 , x2 ∈ AZ that coincide outside of Ω− nk and satisfy τ (x1 ) = τ (x2 ). It then follows from the proof of the implication (b) =⇒ (a) in Proposition 3.18 that the set + (Ω− nk + (−S)) + S = Ωnk + S = Ωnk supports two distinct ME patterns. 5. The Garden of Eden theorem for general amenable groups 5.1. Amenability. (cf. [27], [47], [12, Chapter 4], [21, Chapter 9]) Definition 5.1. A countable group G is called amenable if there exists a sequence (Fn )n∈N of nonempty finite subsets of G such that |Fn \ Fn g| = 0 for all g ∈ G. (5.1) lim n→∞ |Fn | Such a sequence is called a Følner sequence for G. Note that if A and B are finite sets with the same cardinality, then |A \ B| = |B \ A| and |A △ B| = |A \ B| + |B \ A| = 2|A \ B|, where △ denotes symmetric difference of sets. As |F g| = |F | for every finite subset F ⊂ G and any g ∈ G, it follows that Condition (5.1) is equivalent to each of the following conditions: |Fn g \ Fn | (5.2) lim = 0 for all g ∈ G, n→∞ |Fn | or |Fn △ Fn g| (5.3) lim = 0 for all g ∈ G. n→∞ |Fn | 16 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Example 5.2. All finite groups are amenable. Indeed, if G is a finite group, then the constant sequence, defined by Fn := G for all n ∈ N, is a Følner sequence for G since Fn \ Fn g = ∅ for every g ∈ G. Example 5.3. The free abelian groups of finite rank Zd , d ≥ 1, are also amenable. As a Følner sequence for Zd , one can take for instance the sequence of cubes (5.4) Fn := {x ∈ Zd : kxk∞ ≤ n} = {0, ±1, . . . , ±n}d , where kxk∞ := max1≤i≤d |xi | for all x = (x1 , . . . , xd ) ∈ Zd is the sup-norm. To see this, observe that |Fn | = (2n + 1)d (5.5) and, by the triangle inequality, Since Fn + g ⊂ Fn+kgk∞ for all g ∈ Zd . Fn ⊂ Fn+kgk∞ , this implies (5.6) |(Fn + g) \ Fn | ≤ (2n + kgk∞ + 1)d − (2n + 1)d . As the right-hand side of (5.6) is a polynomial of degree d−1 in n while |Fn | is a polynomial of degree d in n by (5.5), we conclude that |(Fn + g) \ Fn | = 0, n→∞ |Fn | lim which is (5.2) in additive notation. Let G be a finitely generated group. If S ⊂ G is a finite symmetric generating subset, the Cayley graph of G with respect to S is the graph G(G, S) whose set of vertices is G and two vertices g, h ∈ G are joined by an edge if and only if h = gs for some s ∈ S. Equip the set of vertices of G(G, S) with its graph metric and consider the ball Bn ⊂ G of radius n centered at 1G . It is easy to see that the sequence of positive integers (|Bn |)n∈N is submultiplicative. Thus the limit p (5.7) γ(G, S) := lim n |Bn | n→∞ exists and satisfies 1 ≤ γ(G, S) < ∞. One says that the group G has subexponential growth if γ(G, S) = 1 and exponential growth if γ(G, S) > 1. The fact that G has subexponential (resp. exponential) growth does not depend on the choice of the finite generating subset S ⊂ G although the value of γ(G, S) does. Example 5.4. The groups Zd have subexponential growth. Indeed, if e1 , . . . , ed is the canonical basis of Zd , and we take S := {±e1 , . . . , ±ed }, then the graphPdistance between two vertices g and h of G(Zd , S) is kg − hk1 , where we write kxk1 := 1≤i≤d |xi | for the 1-norm of x = (x1 , . . . , xd ) ∈ Zd . We then have Bn = {g ∈ Zd : kgk1 ≤ n} ⊂ {0, ±1, . . . , ±n}d THE GARDEN OF EDEN THEOREM: OLD AND NEW 17 and hence |Bn | ≤ (2n + 1)d . This implies that Zd has subexponential growth. Here it can be checked that the sequence (Bn )n∈N is also a Følner sequence for Zd . When G is an arbitrary finitely generated group of subexponential growth and S is a finite symmetric generating set for G, it can be shown that one can always extract a Følner sequence from the sequence (Bn )n∈N . Consequently, every finitely generated group with subexponential growth is amenable. Example 5.5. A (nonabelian) free group on two generators has exponential growth and is not amenable. Indeed, let G be a free group based on two generators a and b. Consider the finite symmetric generating subset S ⊂ G defined by S := {a, b, a−1 , b−1 }. Then every element g ∈ G can be uniquely written in reduced form, i.e., in the form (5.8) g = s1 s2 . . . sn , where n ≥ 0, si ∈ S for all 1 ≤ i ≤ n, and si+1 6= s−1 for all 1 ≤ i ≤ n − 1. The integer i ℓS (g) := n is called the length of g with respect to the generators a and b. It is equal to the distance from g to 1G in the Cayley graph G(G, S). We deduce that |Bn | = 4 · 3n−1 for all n ≥ 1 so that √ n γ(G, S) = lim 4 · 3n−1 = 3 > 1. n→∞ This shows that G has exponential growth. Now suppose by contradiction that (Fn )n∈N is a Følner sequence for G and choose some positive real number ε < 1/2. Since the sequence (Fn )n∈N is Følner, it follows from (5.1) that there exists an integer N ≥ 0 such that the set F := FN satisfies (5.9) |F \ F s| ≤ ε|F | for all s ∈ S. Denote, for each s ∈ S, by Gs the subset of G consisting of all elements g 6= 1G whose reduced form ends with the letter s−1 . The four sets Gs , s ∈ S, are pairwise disjoint so that X (5.10) |F ∩ Gs | ≤ |F |. s∈S On the other hand, for each s ∈ S, we have that (5.11) |F | = |F \ Gs | + |F ∩ Gs | = |(F \ Gs )s| + |F ∩ Gs |. We now observe that (G \ Gs )s ⊂ Gs−1 so that (F \ Gs )s ⊂ (F s \ F ) ∪ (F ∩ Gs−1 ) 18 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT and hence |(F \ Gs )s| ≤ |F s \ F | + |F ∩ Gs−1 | = |F \ F s| + |F ∩ Gs−1 | ≤ ε|F | + |F ∩ Gs−1 | (by (5.9)). By using (5.11), we deduce that |F | ≤ ε|F | + |F ∩ Gs−1 | + |F ∩ Gs | for all s ∈ S. After summing up over all s ∈ S, this yields X 4|F | ≤ 4ε|F | + (|F ∩ Gs−1 | + |F ∩ Gs |) s∈S = 4ε|F | + 2 X s∈S |F ∩ Gs |. Finally, combining with (5.10), we obtain 4|F | ≤ 4ε|F | + 2|F | and hence |F | ≤ 2ε|F |, which is a contradiction since F 6= ∅ and ε < 1/2. This proves that G is not amenable. The class of amenable groups is closed under the operations of taking subgroups, quotients, extensions (this means that if 1 → H → G → K → 1 is an exact sequence with both H and K amenable, so is G), and inductive limits. Consequently, all locally finite groups, all abelian groups and, more generally, all solvable groups are amenable. On the other hand every group containing a free subgroup on two generators is nonamenable. This implies for instance that all nonabelian free groups and the groups SL(n, Z), n ≥ 2, are nonamenable. However, there are groups containing no free subgroups on two generators that are nonamenable. The first examples of such a group was given in [45] where Ol’šanskiı̆ constructed a nonamenable monster group in which every proper subgroup is cyclic. Let us note that there are finitely generated groups of exponential growth that are amenable. For example, the Baumslag-Solitar group BS(1, 2), i.e., the group with presentation ha, b : aba−1 = b2 i, and the lamplighter group, i.e., the wreath product (Z/2Z) ≀ Z, have exponential growth but are both solvable and hence amenable (cf. [23]). The original definition of amenability that was given by von Neumann [44] in 1929 is that a group G is amenable if there exists a finitely additive invariant probability measure defined on the set of all subsets of G. A key observation due to Day [22] is that this is equivalent to the existence of an invariant mean on the Banach space ℓ∞ (G) of bounded real-valued functions on G. It is also in [22] that the term amenable occured for the first time (see [46, p. 137]). The fact that nonabelian free groups are not amenable is related to the Hausdorff-Banach-Tarski paradox which actually was the motivation of von Neumann for introducing the notion of amenability. THE GARDEN OF EDEN THEOREM: OLD AND NEW 19 5.2. Entropy. Let G be countable group, A a finite set, and X a subset of AG (not necessarily a subshift). Given a finite subset Ω ⊂ G, recall (cf. (2.1)) that XΩ := {x|Ω : x ∈ X} ⊂ AΩ . Suppose now that the group G is amenable and fix a Følner sequence F = (Fn )n∈N for G. The entropy of X (with respect to F ) is defined by (5.12) hF (X) := lim sup n→∞ log |XFn | . |Fn | Since XF ⊂ AF and hence log |XF | ≤ |F | · log |A| for every finite subset F ⊂ G, we always have hF (X) ≤ hF (AG ) = log |A|. Example 5.6. Take G = Z and A = {0, 1}. Let us compute the entropy of the golden mean subshift X ⊂ AG (cf. Example 2.2) with respect to the Følner sequence F = (Fn )n∈N , where Fn := {0, 1, . . . , n}. We observe that un := |XFn | satisfies u0 = 2, u1 = 3, and un+2 = un+1 + un for all n ≥ 2. Thus, the sequence (un )n∈N is a Fibonacci sequence and, by Binet’s formula,  1 un = √ ϕn+3 − (1 − ϕ)n+3 , 5 √ where ϕ := (1 + 5)/2 is the golden mean (this is the origin of the name of this subshift). It follows that log un = log ϕ. hF (X) = lim sup n→∞ n + 1 Example 5.7. Take again G = Z and A = {0, 1}, and let us compute now the entropy of the even subshift X ⊂ AG (cf. Example 2.4) with respect to the Følner sequence F = (Fn )n∈N , where Fn := {0, 1, . . . , n}. We observe that un := |XFn | satisfies u0 = 2, u1 = 4, and un+2 = 1 + un+1 + un for all n ≥ 2. Thus, the sequence (vn )n∈N , defined by vn := 1 + un for all n ∈ N, is a Fibonacci sequence. As v0 = 3 and v1 = 5, by applying again Binet’s formula, we get  1 un = −1 + vn = −1 + √ ϕn+5 − (1 − ϕ)n+5 , 5 where ϕ is the golden mean. It follows that hF (X) = lim sup n→∞ log un = log ϕ. n+1 Thus, the even subshift has the same entropy as the golden mean subshift with respect to F. Example 5.8. Take G = Z2 and A = Z/2Z. Let us compute the entropy of the Ledrappier subshift X ⊂ AG (cf. Example 2.5) with respect to the Følner sequence F = (Fn )n∈N , where Fn := {0, 1, . . . , n}2 . We observe that, for each x ∈ X, the pattern x|Fn is entirely 20 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT determined by x|Hn , where Hn ⊂ Z2 is the horizontal interval Hn := {0, 1, . . . , 2n} × {0}. Therefore, un := |XFn | satisfies log un ≤ |Hn | · log |A| = (2n + 1) log 2. This gives us hF (X) = lim sup n→∞ log un = 0. (n + 1)2 Remark 5.9. It is a deep result due to Ornstein and Weiss [46] that, when G is a countable amenable group, A a finite set, and X ⊂ AG a subshift, then the lim sup in (5.12) is actually a true limit and does not depend on the particular choice of the Følner sequence F for G. However, we shall not need it in the sequel. An important property of cellular automata that we shall use in the proof of the Garden of Eden theorem below is that they cannot increase entropy. More precisely, we have the following result. Proposition 5.10. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N and A, B finite sets. Suppose that τ : AG → B G is a cellular automaton and X is a subset of AG . Then one has hF (τ (X)) ≤ hF (X). For the proof, we shall use the following general property of Følner sequences. Lemma 5.11. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N and let S be a finite subset of G. Then one has (5.13) lim n→∞ |Fn S \ Fn | = 0. |Fn | Proof. Observe that Fn S \ Fn = so that |Fn S \ Fn | = Thus we get for all n ∈ N. As [ s∈S [ s∈S (Fn s \ Fn ) (Fn s \ Fn ) ≤ X s∈S |Fn s \ Fn |. |Fn S \ Fn | X |Fn s \ Fn | ≤ |Fn | |Fn | s∈S |Fn s \ Fn | =0 n→∞ |Fn | lim for each s ∈ S by (5.2), this gives us (5.13).  THE GARDEN OF EDEN THEOREM: OLD AND NEW 21 Proof of Proposition 5.10. Let us set Y := τ (X) and let S ⊂ G be a memory set for τ with 1G ∈ S. Recall that it immediately follows from (3.1) that if two configurations coincide on gS for some g ∈ G then their images by τ take the same value at g. We deduce that (5.14) |YΩ | ≤ |XΩS | for every finite subset Ω ⊂ G. Now observe that XΩS ⊂ XΩ × XΩS\Ω ⊂ XΩ × AΩS\Ω , so that we get log |YΩ | ≤ log |XΩ | + |ΩS \ Ω| · log |A|. After replacing Ω by Fn and dividing both sides by |Fn |, this inequality becomes (5.15) As log |YFn | log |XFn | |Fn S \ Fn | ≤ + · log |A|. |Fn | |Fn | |Fn | |Fn S \ Fn | =0 n→∞ |Fn | lim by Lemma 5.11, taking the limsup in (5.15) finally gives the required inequality hF (Y ) ≤ hF (X).  Corollary 5.12. Let G be a countable amenable group and let A, B be finite sets with |A| < |B|. Then there exists no surjective cellular automaton τ : AG → B G . Proof. This is an immediate consequence of Proposition 5.10 since hF (AG ) = log |A| < log |B| = hF (B G ).  The following example (cf. [46, p. 138]) shows that Corollary 5.12 becomes false if the amenability hypothesis is removed. Example 5.13. Let G be the free group on two generators a and b. Take A := Z/2Z and B := Z/2Z × Z/2Z, so that |A| = 2 and |B| = 4. Consider the map τ : AG → B G defined by τ (x)(g) = (x(g) + x(ga), x(g) + x(gb)) for all x ∈ AG and g ∈ G. Observe that τ is a cellular automaton with memory set S = {1G , a, b} and local defining map µ : AG → B given by µ(p) = (p(1G ) + p(a), p(1G ) + p(b)) for all p ∈ AS . It is easy to check that τ is surjective. Note that AG and B G are totally disconnected compact abelian topological groups and τ is a continuous group morphism whose kernel consists of the two constant configurations in AG . 22 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT 5.3. Tilings. Let G be a group. Given a finite subset E ⊂ G, let us say that a subset T ⊂ G is an E-tiling of G provided the sets tE, t ∈ T , are pairwise disjoint and there exists a finite subset E ′ ⊂ G such that the sets tE ′ , t ∈ T , cover G. Example 5.14. Take G = Zd and E = {0, ±1, ±2, . . . , ±m}d for some m ∈ N, then T := ((2m + 1)Z)d ⊂ Zd is an E-tiling (here one can take E ′ = E). Given any nonempty finite subset E of a group G, we can use Zorn’s lemma to prove that there always exists an E-tiling T ⊂ G. Indeed, consider the set S(E) consisting of all subsets S ⊂ G such that the sets sE, s ∈ S, are pairwise disjoint. We first observe that S(E) is nonempty since {1G } ∈ S(E). On the other hand, S(E) is inductive with respect to set inclusion since if S ′ ⊂ S(E) is a chain, then M := ∪S∈S ′ S belongs to S(E) and is an upper bound for S ′ . By Zorn’s lemma, there exists a maximal element T ∈ S(E). As T ∈ S(E), the sets tE, t ∈ T , are pairwise disjoint. Now, given any g ∈ G, we can find, by maximality of T , an element t = t(g) ∈ T such that gE ∩ tE 6= ∅ and hence g ∈ tEE −1 . It follows that the sets tEE −1 , t ∈ T , cover G. Since the set E ′ := EE −1 is finite, this shows that T is an E-tiling of G. For the proof of the Garden of Eden theorem in the next subsection, we shall use some technical results about tilings in amenable groups. Lemma 5.15. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N . Let E ⊂ G be a nonempty finite subset and T ⊂ G an E-tiling. Define, for n ∈ N, the subset Tn ⊂ T by Tn := {t ∈ T : tE ⊂ Fn }. Then there exist a constant α = α(F , T ) > 0 and n0 ∈ N such that (5.16) |Tn | ≥ α|Fn | for all n ≥ n0 . Proof. Since T is a E-tiling, there exists a finite subset E ′ ⊂ G such that the sets tE ′ , t ∈ T , cover G. After replacing E ′ by E ′ ∪ E, if necessary, we may assume that E ⊂ E ′ . Define, for n ∈ N, Tn+ := {t ∈ T : tE ′ ∩ Fn 6= ∅}. Clearly Tn ⊂ Tn+ . As the sets tE ′ , t ∈ Tn+ , cover Fn , we have |Fn | ≤ |Tn+ | · |E ′ | so that |Tn+ | 1 ≥ ′ |Fn | |E | (5.17) for all n ∈ N. Now observe that Tn+ =T∩ [ g∈E ′ Fn g −1 ! and Tn = T ∩ \ h∈E −1 Fn h ! , THE GARDEN OF EDEN THEOREM: OLD AND NEW so that Tn+ \ Tn = T ∩ ⊂ [ g∈E ′ g∈E ′ g∈E ′ , h∈E We deduce that |Tn+ \ Tn | ≤ X g∈E ′ , h∈E As Fn g −1 \ Fn g −1 \ [ = [ \ \ Fn h−1 h∈E 23 ! Fn h−1 h∈E (Fn g −1 \ Fn h−1 ). |Fn g −1 \ Fn h−1 | = X g∈E ′ , h∈E |Fn \ Fn h−1 g|. |Fn \ Fn h−1 g| =0 n→∞ |Fn | for all g ∈ E ′ and h ∈ E by (5.1), it follows that lim |Tn+ | − |Tn | |T + \ Tn | = n →0 |Fn | |Fn | 1 , we deduce that there exists n0 ∈ N such as n → ∞. Using (5.17) and taking ε := 2|E ′| that |T + | |T + | − |Tn | 1 |Tn | = n − n ≥ ′ − ε = α, |Fn | |Fn | |Fn | |E | 1 , for all n ≥ n0 .  where α := 2|E ′ | Proposition 5.16. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N and let A be a finite set. Let X ⊂ AG be a subset and suppose there exist a nonempty finite subset E ⊂ G and an E-tiling T ⊂ G such that XtE $ AtE for all t ∈ T . Then hF (X) < log |A|. Proof. Let us set, as above, Tn := {t ∈ T : tE ⊂ Fn } and write [ tE, Fn∗ := Fn \ for all n ∈ N. Observe that S t∈Tn t∈Tn tE ⊂ Fn so that Y ∗ XtE XFn ⊂ AFn × t∈Tn and (5.18) |Fn | = |Fn∗ | + |Tn | · |E|. 24 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT It follows that log |XFn | ≤ |Fn∗ | · log |A| + X t∈Tn X log |XtE |  ≤ |Fn∗ | · log |A| + = |Fn∗ | · log |A| + |Tn | · log(|A||E| − 1) t∈Tn log |AtE | − 1 = |Fn∗ | · log |A| + |Tn | · |E| · log |A| + |Tn | · log(1 − |A|−|E| ) = |Fn | · log |A| + |Tn | · log(1 − |A|−|E| ), where the last equality follows from (5.18). Setting c := − log(1 − |A|−|E|) > 0, we deduce that log |XFn | hF (X) = lim sup ≤ log |A| − cα < log |A|, |Fn | n→∞ where α = α(F , T ) is as in (5.16).  Corollary 5.17. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N and let A be a finite set. Let X ⊂ AG be a subshift and suppose that there exists a nonempty finite subset E ⊂ G such that XE $ AE . Then one has hF (X) < log |A|. Proof. If T is an E-tiling of G, we deduce from the shift-invariance of X that XtE $ AtE for all t ∈ T , so that Proposition 5.16 applies.  5.4. The Garden of Eden theorem for amenable groups. The following result is due to Machı̀, Scarabotti, and the first author [20]. Since the groups Zd are all amenable, it extends Theorem 4.1. Theorem 5.18. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N and A a finite set. Suppose that τ : AG → AG is a cellular automaton. Then the following conditions are equivalent: (a) τ is surjective; (b) hF (τ (AG )) = log |A|; (c) τ is pre-injective. Proof. The implication (a) =⇒ (b) is obvious since hF (AG ) = log |A|. In order to show the converse implication, let us suppose that τ is not surjective, that is, the image subshift X := τ (AG ) is such that X $ AG . Since X is closed in AG , there exists a finite subset E ⊂ G such that XE $ AE . By applying Corollary 5.17, we deduce that hF (X) < log |A|. This shows (b) =⇒ (a). Let S ⊂ G be a memory set for τ such that 1G ∈ S. Let us show (b) =⇒ (c). Suppose that τ is not pre-injective. By virtue of Proposition 3.18, we can find a nonempty finite subset Ω ⊂ G and two patterns p1 , p2 ∈ AΩ that are mutually erasable for τ . Let E := ΩS −1 S. Observe that Ω ⊂ E since 1G ∈ S. Let T ⊂ G be an E-tiling of G. Consider the subset Z ⊂ AG defined by Z := {z ∈ AG : z|tΩ 6= tp1 for all t ∈ T }. THE GARDEN OF EDEN THEOREM: OLD AND NEW 25 Observe that ZtE $ AtE for all t ∈ T . By using Proposition 5.10 and Proposition 5.16, we deduce that hF (τ (Z)) ≤ hF (Z) < log |A|. We claim that τ (Z) = τ (AG ). Let x ∈ AG . Let Tx := {t ∈ T : x|tΩ = tp1 } and define z ∈ Z by setting, for all g ∈ G, ( tp2 (g) if g ∈ tΩ for some t ∈ Tx z(g) := x(g) otherwise. Let us check that τ (z) = τ (x). Let g ∈ G. If g ∈ / ∪t∈Tx tΩS −1 , then gS ∩ tΩ = ∅ for all t ∈ Tx and therefore z|gS = x|gS , so that τ (z)(g) = τ (x)(g). Suppose now that g ∈ tΩS −1 for some (unique) t = t(g) ∈ Tx and consider the configuration y ∈ AG defined by setting, for all h ∈ G, ( tp2 (h) if h ∈ tΩ y(h) := x(h) otherwise. Observe that x|G\tΩ = y|G\tΩ . Since the patterns x|tΩ = tp1 and y|tΩ = tp2 are mutually erasable, we deduce that τ (y) = τ (x). Moreover, as gS ⊂ tΩS −1 S = tE, we have z|gS = y|gS , and therefore τ (z)(g) = τ (y)(g) = τ (x)(g). This shows that τ (z) = τ (x), and the claim follows. We conclude that hF (τ (AG )) = hF (τ (Z)) < log |A|. This shows the implication (b) =⇒ (c). Finally, let us show (c) =⇒ (b). Let us set as above X := τ (AG ) and suppose that hF (X) < log |A|. As 1G ∈ S, we have Fn ⊂ Fn S −1 so that XFn S −1 ⊂ XFn × AFn S −1 \F n , for all n ∈ N. We deduce that log |XFn | |Fn S −1 \ Fn | log |XFnS −1 | ≤ + log |A|. (5.19) |Fn | |Fn | |Fn | As |Fn S −1 \ Fn | =0 n→∞ |Fn | by Lemma 5.11, we deduce from (5.19) that lim lim sup n→∞ log |XFnS −1 | log |XFn | ≤ lim sup = hF (X) < log |A|. |Fn | |Fn | n→∞ Consequently, we can find n0 ∈ N such that, (5.20) |XFn0 S −1 | < |A||Fn0 | . Fix a0 ∈ A and consider the subset Z ⊂ AG defined by Z := {z ∈ AG : z(g) = a0 for all g ∈ G \ Fn0 }. Note that |Z| = |A||Fn0 | . Let z1 , z2 ∈ Z. If g ∈ G \ Fn0 S −1 , then z1 and z2 coincide on gS ⊂ G\Fn0 so that τ (z1 )(g) = τ (z2 )(g). Therefore τ (z1 ) and τ (z2 ) coincide on G\Fn0 S −1 . This implies that |τ (Z)| ≤ |XFn0 S −1 |. Using (5.20), we deduce that |τ (Z)| < |Z|. By the pigeon-hole principle, there exist two distinct elements z1 , z2 ∈ Z such that τ (z1 ) = τ (z2 ). 26 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT As all elements in Z are almost equal (they coincide outside of the finite set Fn0 ), we conclude that τ is not pre-injective.  6. Failure of the Garden of Eden theorem for nonamenable groups Let us say that a countable group G has the Moore property if every surjective cellular automaton τ : AG → AG with finite alphabet A over G is pre-injective and that it has the Myhill property if every pre-injective cellular automaton τ : AG → AG with finite alphabet A over G is surjective. Also let us say that a countable group G satisfies the Garden of Eden theorem if G has both the Moore and the Myhill properties. Theorem 5.18 tells us that every countable amenable group satisfies the Garden of Eden theorem. The examples below, essentially due to Muller [42] (see also [36, Section 6], [20, Section 6], [12, Chapter 5]), show that neither the Moore nor the Myhill property holds for countable groups containing nonabelian free subgroups. Example 6.1. Let G be a countable group and suppose that G contains two elements a and b generating a nonabelian free subgroup H ⊂ G. Take A = {0, 1} and let S := {a, a−1 , b, b−1 }. Consider the cellular automaton τ : AG → AG with memory set {1G } ∪ S defined by ( 0 if x(g) + x(ga) + x(ga−1 ) + x(gb) + x(gb−1 ) ≤ 2 := τ (x)(g) 1 otherwise for all x ∈ AG and g ∈ G. The pair of configurations (x1 , x2 ) ∈ AG × AG , defined by x1 (g) = 0 for all g ∈ G, and x2 (g) = 0 for all g ∈ G \ {1G } and x2 (1G ) = 1, is a diamond for τ . Therefore τ is not pre-injective. However, τ is surjective. To see this, let y ∈ AG . Let us show that there exists x ∈ AG such that τ (x) = y. Let R ⊂ G be a complete set of representatives of the left cosets of H in G. We define x as follows. Every element g ∈ G can be uniquely written in the form g = rh with r ∈ R and h ∈ H. If g ∈ R, i.e., h = 1G , we set x(g) := 0. Otherwise, we set x(g) := y(rh− ), where h− is the predecessor of h in H, i.e., the unique element h− ∈ H such that ℓS (h− ) = ℓS (h) − 1 and h = h− s for some s ∈ S (here ℓS (·) denotes the length of the reduced form for elements of H, see Example 5.5). One easily checks that τ (x) = y. This shows that τ is surjective. Thus the Moore implication fails to hold for groups containing nonabelian free subgroups. Example 6.2. Let G be a countable group and suppose that G contains two elements a and b generating a nonabelian free subgroup H ⊂ G. Let A = Z/2Z × Z/2Z be the Klein four-group and consider the group endomorphisms p and q of A respectively defined by p(α, β) = (α, 0) and q(α, β) = (β, 0) for all (α, β) ∈ A. Let τ : AG → AG be the cellular automaton with memory set S := {a, a−1 , b, b−1 } defined by τ (x)(g) = p(x(ga)) + p(x(ga−1 ) + q(x(gb)) + q(x(gb−1 )) for all x ∈ AG and g ∈ G. The image of τ is contained in (Z/2Z × {0})G . Therefore τ is not surjective. We claim that τ is pre-injective. As τ is a group endomorphism of THE GARDEN OF EDEN THEOREM: OLD AND NEW 27 AG , it suffices to show that there is no configuration with finite support in the kernel of τ . Assume on the contrary that there is an element x ∈ AG with nonempty finite support Ω := {g ∈ G : x(g) 6= 0A } ⊂ G such that τ (x) = 0. Let R ⊂ G be a complete set of representatives of the left cosets of H in G. Let us set Ωr := Ω ∩ rH for all r ∈ R. Then Ω is the disjoint union of the sets Ωr , r ∈ R. Let r ∈ R such that Ωr 6= ∅ and consider an element g = rh ∈ Ωr with h ∈ H at maximal distance from the identity in the Cayley graph of (H, S) (i.e., with ℓS (h) maximal). We have that x(g) = (α, β) 6= (0, 0) = 0A . Suppose first that α 6= 0. We can find s ∈ {a, a−1 } such that ℓS (hs) = ℓS (h) + 1. For all t ∈ S \ {s−1 }, we have that ℓS (hst) = ℓ(h) + 2 and hence x(gst) = 0A by maximality. It follows that τ (x)(gs) = p(x(g)) = (α, 0) 6= 0A , which contradicts the fact that x is in the kernel of τ . Suppose now that α = 0. Then β 6= 0. We take now s ∈ {b, b−1 } such that ℓS (hs) = ℓS (h) + 1. By an argument similar to the one that we used in the first case, we get τ (x)(gs) = q(x(g)) = (β, 0) 6= 0A , so that we arrive at a contradiction also in this case. Thus τ is pre-injective. This shows that the Myhill implication fails to hold for groups containing nonabelian free subgroups. As mentioned in Subsection 5.1, there are nonamenable countable groups containing no nonabelian free subgroups. However, Bartholdi [2] (see [12, Chapter 5]) proved that the Moore property fails to hold for all nonamenable countable groups. Recently, Bartholdi and Kielak [3] also proved that the Myhill property fails to hold for all nonamenable countable groups. Combining these results with the Garden of Eden theorem for amenable groups (Theorem 5.18), this yields the following characterization of amenability in terms of cellular automata. Theorem 6.3. Let G be a countable group. Then the following conditions are equivalent: (a) G is amenable; (b) G has the Moore property; (c) G has the Myhill property; (d) G satisfies the Garden of Eden theorem. 7. The Garden of Eden theorem for subshifts 7.1. Strongly irreducible subshifts. Let G be a countable group and A a finite set. A subshift X ⊂ AG is called strongly irreducible if there is a finite subset ∆ ⊂ G satisfying the following property: if Ω1 and Ω2 are finite subsets of G such that Ω1 ∆ does not meet Ω2 , then, given any two configurations x1 , x2 ∈ X, there exists a configuration x ∈ X which coincides with x1 on Ω1 and with x2 on Ω2 . Example 7.1. The full shift AG is strongly irreducible (one can take ∆ = {1G }). Example 7.2. The even subshift X ⊂ {0, 1}Z , described in Example 2.4, is strongly irreducible (one can take ∆ = {−2, −1, 0, 1, 2}). 28 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Example 7.3. The hard-ball model, described in Example 2.3, is strongly irreducible (one can take ∆ = {0, ±e1 , . . . , ±ed }). In particular (d = 1), the golden mean subshift is strongly irreducible. Example 7.4. The Ledrappier subshift, described in Example 2.5, is not strongly irreducible. Fiorenzi [25, Theorem 4.7] obtained the following extension of Theorem 5.18. Theorem 7.5. Let G be a countable amenable group with Følner sequence F = (Fn )n∈N and A, B finite sets. Suppose that X ⊂ AG is a strongly irreducible subshift of finite type and Y ⊂ B G is a strongly irreducible subshift with hF (X) = hF (Y ) and that τ : X → Y is a cellular automaton. Then the following conditions are equivalent: (a) τ is surjective; (b) hF (τ (X)) = hF (Y ); (c) τ is pre-injective. Example 7.6. The cellular automaton τ : X → Y from the golden mean subshift to the even subshift described in Example 3.7 satisfies all the hypotheses in the previous theorem. As τ is pre-injective (cf. Example 3.16), we deduce that τ is surjective. Note that here one might also easily obtain surjectivity of τ by a direct argument. 7.2. The Moore and the Myhill properties for subshifts. Let G be a countable group, A a finite set, and X ⊂ AG a subshift. One says that the subshift X has the Moore property if every surjective cellular automaton τ : X → X is pre-injective and that it has the Myhill property if every pre-injective cellular automaton τ : X → X is surjective. One says that X has the Moore-Myhill property or that it satisfies the Garden of Eden theorem if it has both the Moore and the Myhill properties. From Theorem 7.5, we immediately deduce the following. Corollary 7.7. Let G be a countable amenable group and A a finite set. Then every strongly irreducible subshift of finite type X ⊂ AG has the Moore-Myhill property. Example 7.8. Let G = Zd and A = {0, 1}. Consider the hard-ball model X ⊂ AG described in Example 2.3. As Zd is amenable and X is both strongly irreducible and of finite type, we deduce from Corollary 7.7 that X has the Moore-Myhill property. In particular (d = 1), the golden mean subshift has the Moore-Myhill property. Example 7.9 (Fiorenzi). Let A = {0, 1} and let X ⊂ AZ be the even subshift (cf. Example 2.4). Consider the cellular automaton σ : AZ → AZ with memory set S = {0, 1, 2, 3, 4} and local defining map µ : AS → A given by ( 1 if y(0)y(1)y(2) ∈ {000, 111} or y(0)y(1)y(2)y(3)y(4) = 00100 µ(y) = 0 otherwise. Then one has σ(X) ⊂ X, and the cellular automaton τ := σ|X : X → X is not pre-injective. Indeed, the configurations x1 , x2 ∈ X defined by x1 = · · · 0 · · · 00(100)100 · · · 0 · · · THE GARDEN OF EDEN THEOREM: OLD AND NEW 29 and satisfy x2 = · · · 0 · · · 00(011)100 · · · 0 · · · τ (x1 ) = · · · 1 · · · 11(100)10011 · · · 1 · · · = τ (x2 ). Observe, alternatively, that the patterns p, q with support Ω := {0, 1, . . . , 12} defined by ( ( 1 if n = 6, 9 1 if n = 7, 8, 9 p(n) = and q(n) = 0 otherwise 0 otherwise, for all n ∈ Ω, are ME. From a case-by-case analysis, one can show that τ is surjective. It follows that X does not have the Moore property. We refer to [24, Section 3] and [19, Counterexample 2.18] for more details. As the even subshift is strongly irreducible and Z is amenable, the previous example shows that the hypothesis that X is of finite type cannot be removed from Corollary 7.7. However, we have the following (cf. [14]). Theorem 7.10. Let G be a countable amenable group and A a finite set. Then every strongly irreducible subshift X ⊂ AG has the Myhill property. Example 7.11. The even subshift X ⊂ {0, 1}Z has the Myhill property since it is strongly irreducible and Z is amenable. Example 7.12. Let A = {0, 1}. Let x0 , x1 ∈ AZ denote the two constant configurations respectively defined by x0 (n) = 0 and x1 (n) = 1 for all n ∈ Z. Note that X = {x0 , x1 } is a subshift of finite type. The map τ : X → X given by τ (x0 ) = τ (x1 ) = x0 is a cellular automaton which is pre-injective but not surjective. It follows that X does not have the Myhill property. This very simple example shows that the hypothesis that X is strongly irreducible cannot be removed neither from Corollary 7.7 nor from Theorem 7.10. Note that X has the Moore property since X is finite, so that every surjective self-mapping of X is injective and therefore pre-injective. Example 7.13 (Fiorenzi). Let A = {0, 1, 2} and let X ⊂ AZ be the subshift of finite type consisting of all x ∈ AZ such that x(n)x(n + 1) ∈ / {01, 02} for all n ∈ Z. Thus a configuration x : Z → A is in X if and only if one of the following conditions is satisfied: (i) x(n) = 0 for all n ∈ Z; (ii) x(n) 6= 0 for all n ∈ Z; (iii) there exists n0 ∈ N such that x(n) ∈ {1, 2} for all n ≤ n0 and x(n) = 0 for all n > n0 . Consider the cellular automaton σ : AZ → AZ with memory set S = {0, 1} and local defining map ( y(0) if y(1) 6= 0 µ(y) = 0 otherwise. 30 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Observe that σ(x) = x if x ∈ X is of type (i) or (ii) while, if x ∈ X is of type (iii), then σ(x) is obtained from x by replacing its rightest nonzero term by 0. We deduce that σ(X) ⊂ X and that the cellular automaton τ := σ|X : X → X is surjective but not pre-injective (see [24, Counterexample 4.27]). It follows that X does not have the Moore property. It turns out that X does not have the Myhill property either. Indeed, consider now the cellular automaton σ ′ : AZ → AZ with memory set S ′ = {−1, 0} and local defining map ( y(0) if y(−1)y(0) ∈ / {10, 20} µ′ (y) = y(−1) otherwise. Observe that σ ′ (x) = x if x ∈ X is of type (i) or (ii) while, if x ∈ X is of type (iii), then σ ′ (x) is obtained from x by repeating on its right its rightest nonzero term. We deduce that σ ′ (X) ⊂ X and that the cellular automaton τ ′ := σ ′ |X : X → X is injective and hence pre-injective. However, τ ′ is not surjective (observe for instance that the pattern p ∈ A{−1,0,1} defined by p(−1)p(0)p(1) = 120 is a Garden of Eden pattern for τ ′ ). Let A be a finite set and X ⊂ AZ a subshift. One says that a word u ∈ A⋆ of length n appears in X if there is a configuration x ∈ X and m ∈ Z such that u = x(m)x(m + 1) · · · x(m + n − 1). The subset L(X) ⊂ A⋆ consisting of all words that appear in X is called the language of X. One says that the subshift X is irreducible if given any two words u, v ∈ L(X) there exists a word w ∈ L(X) such that uwv ∈ L(X). Clearly every strongly irreducible subshift X ⊂ AZ is irreducible. The converse is false as shown by the following example. Example 7.14. Let A = {0, 1} and consider the subshift X ⊂ AZ consisting of the two configurations x ∈ AZ that satisfy x(n) 6= x(n + 1) for all n ∈ Z. It is clear that X is irreducible but not strongly irreducible. Observe that X is of finite type. The following result is an immediate consequence of [32, Theorem 8.1.16] (cf. [24, Corollary 2.19]). Theorem 7.15. Let A be a finite set. Then every irreducible subshift of finite type X ⊂ AZ has the Moore-Myhill property. 8. Garden of Eden theorems for other dynamical systems 8.1. Dynamical systems. By a dynamical system, we mean a triple (X, G, α), where X is a compact metrizable space, G is a countable group, and α is a continuous action of G on X. The space X is called the phase space of the dynamical system. If there is no risk of confusion, we shall write (X, G), or even sometimes simply X, instead of (X, G, α). Example 8.1. Let G be a countable group and A a compact metrizable topological space (e.g. a finite set with its discrete topology). Equip AG = {x : G → A} with the product topology. The shift action σ of G on AG is the action defined by σ(g, x) = gx, where (gx)(h) = x(g −1 h) for all x ∈ AG and g, h ∈ G. Then (AG , G, σ) is a dynamical system. THE GARDEN OF EDEN THEOREM: OLD AND NEW 31 Example 8.2. If (X, G, α) is a dynamical system and Y ⊂ X is a closed α-invariant subset, then (Y, G, α|Y ), where α|Y denotes the action of G on Y induced by restriction of α, is a dynamical system. In particular, if G is a countable group, A a finite set, and X ⊂ AG a subshift, then (X, G, σ|X ) is a dynamical system. Example 8.3. Let f : X → X be a homeomorphism of a compact metrizable space X. The dynamical system generated by f is the dynamical system (X, Z, αf ), where αf is the action of Z on X given by αf (n, f ) := f n (x) for all n ∈ Z and x ∈ X. We shall also write (X, f ) to denote the dynamical system generated by f . Remark 8.4. If we fix the countable group G, the dynamical systems (X, G) are the objects of a concrete category DG in which the morphisms from an object X ∈ DG to another object Y ∈ DG consist of all equivariant continuous maps τ : X → Y . It follows from the Curtis-Hedlund-Lyndon theorem (cf. Theorem 3.8) that the category CG described in Remark 3.10 is a full subcategory of the category DG . e G) be two dynamical systems. Let (X, G) and (X, e G) are topologically conjugate if One says that the dynamical systems (X, G) and (X, they are isomorphic objects in the category DG , i.e., if there exists an equivariant homeoe → X. morphism h : X e G) if there exists an equivariant continuous One says that (X, G) is a factor of (X, e → X. Such a map θ is then called a factor map. A factor map surjective map θ : X e → X is said to be finite-to-one if the pre-image set θ−1 (x) is finite for each x ∈ X. A θ: X finite-to-one factor map is said to be uniformly bounded-to-one if there is an integer K ≥ 1 such that |θ−1 (x)| ≤ K for all x ∈ X. 8.2. Expansiveness. One says that a dynamical system (X, G) is expansive if there exists a neighborhood W ⊂ X × X of the diagonal ∆X := {(x, x) : x ∈ X} ⊂ X × X such that, for every pair of distinct points x, y ∈ X, there exists an element g = g(x, y) ∈ G such that (gx, gy) ∈ / W . Such a set W is then called an expansiveness neighborhood of the diagonal. If d is a metric on X compatible with the topology, the fact that (X, G) is expansive is equivalent to the existence of a constant δ > 0 such that, for every pair of distinct points x, y ∈ X, there exists an element g = g(x, y) ∈ G such that d(gx, gy) ≥ δ. Example 8.5. Let G be a countable group and A a finite set. Then the G-shift on AG is expansive. Indeed, it is clear that W := {(x, y) ∈ AG × AG : x(1G ) = y(1G )} is an expansiveness neighborhood of ∆AG . Example 8.6. If (X, G) is an expansive dynamical system and Y ⊂ X is a closed invariant subset, then (Y, G) is expansive. Indeed, if W is an expansiveness neighborhood of ∆X , then W ∩(Y ×Y ) is an expansiveness neighborhood of ∆Y . In particular, if G is a countable 32 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT group, A a finite set, and X ⊂ AG a subshift, then the dynamical system (X, G, σ|X ) is expansive. 8.3. Homoclinicity. Let (X, G) be a dynamical system. Two points x, y ∈ X are called homoclinic with respect to the action of G on X, ore more briefly, G-homoclinic, if for any neighborhood W ⊂ X × X of the diagonal ∆X , there is a finite set F = F (W, x, y) ⊂ G such that (gx, gy) ∈ W for all g ∈ G \ F . If d is a metric on X that is compatible with the topology, two points x, y ∈ X are homoclinic if and only if lim d(gx, gy) = 0, g→∞ where ∞ is the point at infinity in the one-point compactification of the discrete group G. This means that, for every ε > 0, there is a finite subset F = F (ε, d, x, y) ⊂ G such that d(gx, gy) < ε for all g ∈ G \ F . Homoclinicity clearly defines an equivalence relation on X (transitivity follows from the triangle inequality). The equivalence classes of this relation are called the G-homoclinicity classes of X. Definition 8.7. Let (X, G) be a dynamical system and Y a set. One says that a map τ : X → Y is pre-injective if its restriction to each G-homoclinicity class is injective. Example 8.8. Let G be a countable group and A a finite set. Two configurations x, y ∈ AG are homoclinic with respect to the shift action of G on AG if and only if they are almost equal (see e.g. [16, Proposition 2.5]). Indeed, first observe that the sets WΩ := {(x, y) ∈ AG × AG : x|Ω = y|Ω }, where Ω runs over all finite subsets of G, form a neighborhood base of the diagonal ∆AG ⊂ AG × AG (this immediately follows from the definition of the product topology). Now, if x, y ∈ AG are almost equal, then the set D ⊂ G consisting of all g ∈ G such that x(g) 6= y(g) is finite, so that ΩD −1 is also finite for every finite subset Ω ⊂ G. As (gx, gy) ∈ WΩ for every g ∈ G\ΩD −1 , this implies that x and y are homoclinic. Conversely, suppose that x, y ∈ AG are homoclinic. Then there exists a finite subset F ⊂ G such that (gx, gy) ∈ W{1G } for all g ∈ G \ F . This implies that x(g) = y(g) for all g ∈ G \ F −1 , so that x and y are almost equal. Example 8.9. Let (X, G, α) be a dynamical system and Y ⊂ X a closed invariant subset. Denote as above by α|Y the restriction of α to Y . Then two points x, y ∈ Y are homoclinic with respect to α|Y if and only if they are homoclinic with respect to α. In particular, if G is a countable group, A a finite set, and X ⊂ AG a subshift, then two configurations x, y ∈ X are homoclinic with respect to σ|X if and only if they are almost equal. It follows that the definition of pre-injectivity for cellular automata between subshifts given in Definition 3.14 agrees with the one given in Definition 8.7 above. 8.4. The Moore and the Myhill properties for dynamical systems. Let (X, G) be a dynamical system. An endomorphism of (X, G) is a continuous equivariant map τ : X → X. THE GARDEN OF EDEN THEOREM: OLD AND NEW 33 One says that the dynamical system (X, G) has the Moore property if every surjective endomorphism of (X, G) is pre-injective and that it has the Myhill property if every preinjective endomorphism of (X, G) is surjective. One says that (X, G) has the Moore-Myhill property or that it satisfies the Garden of Eden theorem if it has both the Moore and the Myhill properties. Observe that all these properties are invariants of topological conjugacy in the sense that if the dynamical systems (X, G) and (Y, G) are topologically conjugate then (X, G) has the Moore (resp. the Myhill, resp. the Moore-Myhill) property if and only if (Y, G) has the Moore (resp. the Myhill, resp. the Moore-Myhill) property. Remark 8.10. In the particular case when (X, G) is a subshift, it immediately follows from Theorem 3.8 and Example 8.9 that these definitions are equivalent to the ones given in Subsection 7.2. 8.5. Anosov diffeomorphisms. Let f : M → M be a diffeomorphism of a compact smooth manifold M. One says that f is an Anosov diffeomorphism (see e.g. [51], [7], [50]) if the tangent bundle T M of M continuously splits as a direct sum T M = Es ⊕ Eu of two df -invariant subbundles Es and Eu such that, with respect to some (or equivalently any) Riemannian metric on M, the differential df is exponentially contracting on Es and exponentially expanding on Eu , i.e., there are constants C > 0 and 0 < λ < 1 such that (i) kdf n (v)k ≤ Cλn kvk, (ii) kdf −n (w)k ≤ Cλn kwk for all x ∈ M, v ∈ Es (x), w ∈ Eu (x), and n ≥ 0. Example 8.11 (Arnold’s cat). Consider the matrix   0 1 A= 1 1 and the diffeomorphism f of the 2-torus T2 = R2 /Z2 = R/Z × R/Z given by     x2 x1 f (x) := Ax = for all x = ∈ T2 . x1 + x2 x2 The dynamical system (T2 , f ) is known as Arnold’s cat. The diffeomorphism √ f is Anosov. 1 1+ 5 Indeed, the eigenvalues of A are λ1 = − , and λ2 = ϕ, where ϕ := is the golden ϕ 2 mean. As −1 < λ1 < 0 and 1 < λ2 , it follows that df = A is exponentially contracting in the direction of the eigenline associated with λ1 and uniformly expanding in the direction of the eigenline associated with λ2 . Example 8.12 (Hyperbolic toral automorphism). More generally, if A ∈ GLn (Z) has no eigenvalue on the unit circle, then the diffeomorphism f of the n-torus Tn = Rn /Zn , defined by f (x) = Ax for all x ∈ Tn , is Anosov. Such a diffeomorphism is called a hyperbolic toral automorphism. In [17, Theorem 1.1], we obtained the following result. 34 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Theorem 8.13. Let f be an Anosov diffeomorphism of the n-dimensional torus Tn . Then the dynamical system (Tn , f ) has the Moore-Myhill property. The proof given in [17] uses two classical results. The first one is the Franks-Manning theorem [26], [37] which states that (Tn , f ) is topologically conjugate to a hyperbolic toral automorphism. The second one is a result of Walters [55] which says that every endomorphism of a hyperbolic toral automorphism is affine. We do not know if the dynamical system (M, f ) has the Moore-Myhill property whenever f is an Anosov diffeomorphism of a compact smooth manifold M. However, we have obtained in [16, Theorem 1.1] the following result. Theorem 8.14. Let X be a compact metrizable space equipped with a continuous action of a countable amenable group G. Suppose that the dynamical system (X, G) is expansive e ⊂ AG , and a uniformly and that there exist a finite set A, a strongly irreducible subshift X e → X. Then the dynamical system (X, G) has the Myhill bounded-to-one factor map θ : X property. A homeomorphism f of a topological space X is called topologically mixing if, given any two nonempty open subsets U, V ⊂ X, there exists an integer N ≥ 0 such that f n (U) ∩ V 6= ∅ for all n ∈ Z that satisfy |n| ≥ N. By the classical work of Bowen (cf. [6, Theorem 28 and Proposition 30] and [5, Proposition 10]), dynamical systems generated by topologically mixing Anosov diffeomorphisms satisfy the hypotheses of Theorem 8.14. As a consequence (cf. [17, Corollary 4.4]), we get the following partial extension of Theorem 8.13. Corollary 8.15. Let f be a topologically mixing Anosov diffeomorphism of a compact smooth manifold M. Then the dynamical system (M, f ) has the Myhill property. Remark 8.16. All known examples of Anosov diffeomorphisms are topologically mixing. Also, all compact smooth manifolds that are known to admit Anosov diffeomorphisms are infra-nilmanifolds. We recal that a nilmanifold is a manifold of the form N/Γ, where N is a simply-connected nilpotent Lie group and Γ is a discrete cocompact subgroup of N and that a infra-nilmanifold is a manifold that is finitely covered by some nilmanifold. 8.6. Principal algebraic dynamical systems. An algebraic dynamical system is a dynamical system of the form (X, G), where X is a compact metrizable abelian topological group and G is a countable group acting on X by continuous group morphisms. Note that if (X, G) is an algebraic dynamical system, then, for each g ∈ G, the map x 7→ gx is a continuous group automorphism of G with inverse x 7→ g −1 x. Example 8.17. Let G be a countable group and A a compact metrizable topological group (for example a finite discrete abelian group, or the n-dimensional torus Tn , or the infinite-dimensional torus TN , or the group Zp of p-adic integers for some prime p). Then the G-shift (AG , G) is an algebraic dynamical system. Example 8.18. Let X be a compact metrizable abelian group and f : X → X a continuous group automorphism (for example X = Tn and f ∈ GLn (Z)). Then the dynamical system (X, f ) generated by f is an algebraic dynamical system. THE GARDEN OF EDEN THEOREM: OLD AND NEW 35 Let (X, G) be an algebraic dynamical system. If d is a metric on X that is compatible with the topology then a point x ∈ X is homoclinic to 0X if and only if one has lim d(gx, 0X ) = 0. g→∞ The set ∆(X, G) consisting of all points of X that are homoclinic to 0X is an G-invariant additive subgroup of X, called the homoclinic group of (X, G) (cf. [33]). Two points x, y ∈ X are homoclinic if and only if x − y ∈ ∆(X, G). It follows that the set of homoclinicity classes of (X, G) can be identified with the quotient group X/∆(X, G). b of X. We recall that if L is a locally compact Consider now the Pontryagin dual X b are the characters of L, i.e., the abelian group, the elements of its Pontryagin dual L b is continuous group morphisms χ : L → T, where T := R/Z, and that the topology on L the topology of uniform convergence on compact subsets. (see e.g. [41]). As the abelian b is a discrete countable abelian group. There is also group X is compact and metrizable, X b a natural dual action of G on X defined by gχ(x) := χ(g −1 x) b and x ∈ X. Note that χ 7→ gχ is a group automorphism of X b for for all g ∈ G, χ ∈ X, each g ∈ G. We recall that the integral group ring Z[G] of G consists of all formal series X r= rg g, g∈G where rg ∈ Z for all g ∈ G and rg = 0 for all but finitely many g ∈ G, and the operations on Z[G] are defined by the formulas X r+s= (rg + sg )g, g∈G rs = X rg 1 sg 2 g 1 g 2 g1 ,g2 ∈G for all r= X g∈G rg g, s= X g∈G sg g ∈ Z[G]. b extends to a left Z[G]-module structure on X. b By linearity, the action of G on X Conversely, if M is a countable left Z[G]-module and we equip M with its discrete c is a compact metrizable abelian group. The left topology, then its Pontryagin dual M Z[G]-module structure on M induces by restriction an action of G on M, and, by dualizing, c by continuous group morphisms, so that (M c, G) is an algebraic we get an action of G on M dynamical system. Using the fact that every locally compact abelian group is isomorphic to its bidual, one shows that Pontryagin duality yields a one-to-one correspondence between algebraic 36 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT dynamical systems with acting group G and countable left Z[G]-modules (see [48], [34], [35]). Let f ∈ Z[G] and consider the cyclic left Z[G]-module Mf := Z[G]/Z[G]f obtained by quotienting the ring Z[G] by the principal left ideal generated by f . The algebraic dynamical system associated by Pontryagin duality with Mf is denoted by (Xf , G) and is called the principal algebraic dynamical system associated with f . In [18], we obtained the following result. Theorem 8.19. Let G be a countable abelian group and f ∈ Z[G]. Suppose that the principal algebraic dynamical system (Xf , G) associated with f is expansive and that Xf is connected. Then the dynamical system (Xf , G) satisfies the Moore-Myhill property. The proof uses two main ingredients. The first one is a result of Lind and Schmidt [33] which asserts that the homoclinic group of (Xf , G), equipped with the discrete topology and the induced action of G, is conjugate to the Pontryagin dual Z[G]/Z[G]f of Xf . The second is a rigidity result of Bhattacharya [4] according to which every endomorphism of (Xf , G) is affine (see [18] for more details). −1 In the case G = Zd , the group ring Z[G] can be identified with the ring Z[u1 , u−1 1 , . . . , ud , ud ] of Laurent polynomials with integral coefficients on d commuting indeterminates. Example 8.20. For G = Z and f = u2 − u − 1 ∈ Z[u, u−1 ] = Z[G], one can check that the associated principal algebraic dynamical system (Xf , Z) is topologically conjugate to Arnold’s cat on T2 (see e.g. [48, Example 2.18.(2)]). Thus, we recover from Theorem 8.19 that Arnold’s cat satisfies the Moore-Myhill property. 9. Some Additional topics 9.1. Infinite alphabets and uncountable groups. The notion of a subshift and that of a cellular automaton between subshifts can be extended to the case where the alphabet sets are infinite and the group is not countable. More specifically, let G be a (possibly uncountable) group and A a (possibly infinite) set. The prodiscrete topology on AG is the Q product topology obtained by taking the discrete topology on each factor A of AG = g∈G A. The prodiscrete topology on AG is not metrizable as soon as G is uncountable and A contains more than one element. However, this topology is induced by the prodiscrete uniform structure on AG , that is, the product uniform structure on AG obtained by taking the discrete uniform structure on each factor A of AG (see [12, Appendix B] for more details). A subset X ⊂ AG is called a subshift if X is invariant under the shift action and closed for the prodiscrete topology. Let G be a group and let A, B be sets. Suppose that X ⊂ AG and Y ⊂ B G are two subshifts. One defines cellular automata between X and Y exactly as in Definition 3.1. Every cellular automaton τ : X → Y is continuous with respect to the topologies on X and Y induced by the prodiscrete topologies on AG and B G . The converse is false in general [10, Section 4], [12, Example 1.8.2]. However, the Curtis-Hedlund-Lyndon theorem (cf. Theorem 3.8) admits the following generalization [10, Theorem 1.1], [12, Theorem 1.8.1]: a map τ : X → Y is a cellular automaton if and only if it is equivariant (with respect to the THE GARDEN OF EDEN THEOREM: OLD AND NEW 37 G-shift actions) and uniformly continuous (for the uniform structures on X and Y induced by the prodiscrete uniform structures on AG and B G ). One can extend the notion of amenability defined only for countable groups in Section 5.1 by declaring that a general group G is amenable if all of its finitely generated subgroups are amenable. This extension makes sense since every finitely generated group is countable and every subgroup of a countable amenable group is itself amenable. The Garden of Eden theorem (cf. Theorem 5.18) remains valid in this more general setting: if G is a (possibly uncountable) amenable group, A a finite set, and τ : AG → AG a cellular automaton, then τ : AG → AG is surjective if and only if it is pre-injective. The proof can be reduced to the case when the group G is finitely generated (and hence countable) by using the operations of restriction and induction for cellular automata (see [11] and [12, Section 1.7]). One can also directly follow the proof given above for Theorem 5.18 by replacing the Følner sequence by a Følner net (see [12, Theorem 5.8.1]). 9.2. Linear cellular automata. Let G be a group and let K be a field. Let A be a finite-dimensional vector space over K and set d = dimK (A). Observe that A is infinite as soon as the field K is infinite (e.g. K = R) and d 6= 0. Taking A as an alphabet, the configuration set AG inherits a natural product vector space structure. The support of a configuration x ∈ AG is the subset supp(x) := {g ∈ G : x(g) 6= 0} ⊂ G. Thus, x ∈ AG has finite support if and only if it is almost equal to the constant zero-configuration. We denote by A[G] ⊂ AG the vector subspace consisting of all configurations with finite support. Recall that an involutive K-algebra is a K-algebra equipped with an involution that is a K-algebra anti-automorphism. The vector space K[G] has a natural additional structure of an involutive K-algebra. The multiplication on K[G] is the convolution product defined by X X α(g1)β(g2 ) = α(h)β(h−1g) (αβ)(g) := g1 ,g2 ∈G: g1 g2 =g h∈G for all α, β ∈ K[G] and g ∈ G, and the involution is the map α 7→ α∗ given by α∗ (g) := α(g −1) for all α ∈ K[G] and g ∈ G. This involutive K-algebra is called the group algebra of the group G with coefficients in K. Note that the group G embeds as a subgroup of the group of invertible elements of K[G] via the map g 7→ δg , where δg ∈ K[G] is defined by δg (g) = 1 and δg (h) = 0 for all h ∈ G with h 6= g, and that G ⊂ K[G] is a base for the vector space K[G]. The G-shift action on AG is clearly K-linear, so that it yields a left K[G]-module structure on AG . Observe that A[G] is a submodule of AG . A linear cellular automaton over the group G and the alphabet A is a cellular automaton τ : AG → AG that is K-linear with respect to the vector space structure on AG (if S is any memory set for τ and µ : AS → A is the associated local defining map, this is equivalent to requiring that µ is K-linear). 38 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT Let us denote by LCA(G; A) the vector space consisting of all linear cellular automata τ : AG → AG . Let τ ∈ LCA(G; A). Note that A[G] is stable under τ . Indeed, if S ⊂ G is a memory set for τ , then supp(τ (x)) ⊂ supp(x)S −1 for all x ∈ AG ) (see [12, Proposition 8.2.3]). Moreover, τ is pre-injective if and only if τ |A[G] : A[G] → A[G] is injective (cf. [12, Proposition 8.2.5]). Observe also that τ is a K[G]-module endomorphism of AG and hence of A[G]. The vector space LCA(G; A) has a natural structure of a K-algebra with the composition of maps as the multiplicative operation. Furthermore, the restriction map τ 7→ τ |A[G] yields a K-algebra isomorphism from LCA(G; A) onto EndK[G] (A[G]), the endomorphism Kalgebra of the K[G]-module A[G] (cf. [12, Theorem 8.7.6]). It turns out that A[G] is a free K[G]-module with rank d. Actually, if (ei )1≤i≤d is a base for the vector space A, then the family (xi )1≤i≤d , where xi ∈ A[G] is the configuration defined by xi (1G ) = ei and xi (g) = 0 for g 6= 1G , is a free base for the K[G]-module A[G] (cf. [12, Proposition 8.7.3]). One deduces that EndK[G](A[G]) is isomorphic, as a K-algebra, to the K-algebra Matd (K[G]) of d × d matrices with coefficients in the group algebra K[G]. It follows that LCA(G; A) and Matd (K[G]) are isomorphic as K-algebras (cf. [12, Corollary 8.7.8]). For instance, the map Φ : Matd (K[G]) → LCA(G; A), sending each matrix α = (αij )1≤i,j≤d ∈ Matd (K[G]) to the unique linear cellular automaton τ ∈ LCA(G; A) such that X ∗ τ (xi ) = αji xj 1≤j≤d for all 1 ≤ i ≤ d, is a K-algebra isomorphism. The adjoint of a matrix α ∈ Matd (K[G]) is the matrix α∗ ∈ Matd (K[G]) given by ∗ ∈ K[G] for all 1 ≤ i, j ≤ d. The involution α 7→ α∗ makes Matd (K[G]) (α∗ )ij := αji into an involutive K-algebra. Let us transport this involution to LCA(G; A) via Φ. Thus, LCA(G; A) becomes an involutive K-algebra with involution τ 7→ τ ∗ satisfying X τ ∗ (xi ) = αij xj 1≤j≤d for all 1 ≤ i ≤ d and τ = Φ(α) ∈ LCA(G; A). Note that this involution on LCA(G; A) depends on the choice of a base for A. Consider now the non-degenerate K-bilinear symmetric map A × A → K defined by X X X a·b= ai bi for all a = ai ei and b = bi ei , with ai , bi ∈ K for 1 ≤ i ≤ d. 1≤i≤d 1≤i≤d 1≤i≤d Then the K-bilinear map A[G] × AG → K, defined by X hx, yi := x(g) · y(g) g∈G for all x ∈ A[G] and y ∈ AG , is non-degenerate in both arguments. Given a linear cellular automaton τ : AG → AG , Bartholdi [1] (see also [52]) observed that hτ (x), yi = hx, τ ∗ (y)i THE GARDEN OF EDEN THEOREM: OLD AND NEW 39 for all x ∈ A[G] and y ∈ AG , and used this to show that τ is pre-injective (resp. surjective) if and only if τ ∗ is surjective (resp. pre-injective). In [8] (see also [12, Theorem 8.9.6]) a linear version of the Garden of Eden theorem is proved, namely that if G is amenable and τ ∈ LCA(G; A), then τ is surjective if and only if it is pre-injective. Let G be a nonamenable group. In [3, Theorem 1.1] Bartholdi showed that there exists a finite field K (in [1] he actually observed that the field K can be arbitrary), a finite dimensional vector space A over K, and a pre-injective linear cellular automaton τ ∈ LCA(G; A) which is not surjective. As a consequence ([1, Corollary]), the cellular automaton τ ∗ ∈ LCA(G; A) is surjective but not pre-injective. These two facts, in combination with the linear version of the Garden of Eden theorem in [8], yield a characterization of group amenability in terms of linear cellular automata. The linear version of the Garden of Eden theorem has been extended in [13] to linear cellular automata τ : X → X with X ⊂ AG a strongly irreducible linear subshift of finite type, and in [9] to the case when the alphabet A is a semi-simple left-module of finite length over a (possibly noncommutative) ring. 9.3. Gromov’s Garden of Eden theorem. In [30, Subsection 8.F’], Gromov proved a Garden of Eden type theorem generalizing Theorem 5.18 under several aspects. First of all, the alphabet set A is only assumed to be countable, not necessarily finite. In addition, the universe is (the vertex set V of) a connected simplicial graph G = (V, E) of bounded degree with a natural homogeneity condition (to admit a dense pseudogroup of partial isometries). The classical case corresponds to G = C(G, S) being the Cayley graph of a finitely generated group G with respect to a finite and symmetric generating subset S ⊂ G \ {1G }. The dense pseudogroup of partial isometries is, in this particular case, given by partial left-multiplication by group elements. In this more general setting, the category corresponding to that of cellular automata consists now of the following: • stable spaces, i.e. (stable) projective limits of locally-finite projective systems (XΩ ) of A-valued maps on (subsets of) V with a suitable finiteness and irreducibility condition (bounded propagation) and admitting a dense holonomy (corresponding to shift-invariance in the classical case) as objects, and • maps of bounded propagation (this condition corresponds to continuity) admitting a dense holonomy (this corresponds to G-equivariance), as morphisms. The notion of a Følner sequence and of amenability for simplicial graphs, together with the corresponding notion of entropy (for the above-mentioned spaces of A-valued maps), carry verbatim from the group theoretical framework. All this said, Gromov’s theorem states as follows. Let G = (V.E) be an amenable simplicial connected graph of bounded degree admitting a dense pseudogroup of partial isometries and let A be a finite or countably infinite alphabet set. Suppose that X, Y ⊂ AV are stable spaces of bounded propagation with the same entropy. Let τ : X → Y be a a map of bounded propagation admitting a dense holonomy. Then τ is surjective if and only if it is pre-injective. 40 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT In [19, Lemma 3.11] it is shown that a stable space of bounded propagation is strongly irreducible (cf. Section 7) and of finite type (cf. Section 2). However, as shown in [19, Counterexample 3.13], the converse fails to hold: strong irreducibility and finite type conditions do not imply, in general, bounded propagation. As a consequence, the following theorem (cf. [19, Theorem B]) improves on Gromov’s theorem. Let G = (V.E) be an amenable simplicial connected graph of bounded degree admitting a dense pseudogroup of partial isometries and let A be a finite or countably infinite alphabet set. Suppose that X, Y ⊂ AV are strongly irreducible stable spaces of finite type with the same entropy. Let τ : X → Y be a a map of bounded propagation admitting a dense holonomy. Then τ is surjective if and only if it is pre-injective. Note that this last result also covers Theorem 7.5. 9.4. Cellular automata over homogeneous sets. Cellular automata where the universe is a set endowed with a transitive group action have been investigated by Moriceau [40]. Versions of the Curtis-Hedlund-Lyndon theorem and of the Garden of Eden theorem in this more general setting have been obtained by Wacker [53], [54]. Acknowledgements. We thank Laurent Bartholdi for valuable comments and remarks. References [1] L. Bartholdi, Linear cellular automata, duality and sofic groups, arXiv:1612.06117. [2] L. Bartholdi, Gardens of Eden and amenability on cellular automata, J. Eur. Math. Soc. (JEMS), 12 (2010), pp. 241–248. [3] L. Bartholdi and D. Kielak, Amenability of groups is characterized by Myhill’s theorem, arXiv:1605.09133. [4] S. Bhattacharya, Orbit equivalence and topological conjugacy of affine actions on compact abelian groups, Monatsh. Math., 129 (2000), pp. 89–96. [5] R. Bowen, Markov partitions and minimal sets for Axiom A diffeomorphisms, Amer. J. Math., 92 (1970), pp. 907–918. [6] , Markov partitions for Axiom A diffeomorphisms, Amer. J. Math., 92 (1970), pp. 725–747. [7] M. Brin and G. Stuck, Introduction to dynamical systems, Cambridge University Press, Cambridge, 2002. , The Garden of Eden theorem for linear cellular automata, Ergodic Theory Dynam. Systems, [8] 26 (2006), pp. 53–68. , Amenability and linear cellular automata over semisimple modules of finite length, Comm. [9] Algebra, 36 (2008), pp. 1320–1335. [10] , A generalization of the Curtis-Hedlund theorem, Theoret. Comput. Sci., 400 (2008), pp. 225– 229. , Induction and restriction of cellular automata, Ergodic Theory Dynam. Systems, 29 (2009), [11] pp. 371–380. , Cellular automata and groups, Springer Monographs in Mathematics, Springer-Verlag, Berlin, [12] 2010. , A Garden of Eden theorem for linear subshifts, Ergodic Theory Dynam. Systems, 32 (2012), [13] pp. 81–102. [14] , The Myhill property for strongly irreducible subshifts over amenable groups, Monatsh. Math., 165 (2012), pp. 155–172. THE GARDEN OF EDEN THEOREM: OLD AND NEW [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] 41 , Surjunctivity and reversibility of cellular automata over concrete categories, in Trends in harmonic analysis, vol. 3 of Springer INdAM Ser., Springer, Milan, 2013, pp. 91–133. , Expansive actions of countable amenable groups, homoclinic pairs, and the Myhill property, Illinois J. Math., 59 (2015), pp. 597–621. , A garden of Eden theorem for Anosov diffeomorphisms on tori, Topology Appl., 212 (2016), pp. 49–56. T. Ceccherini-Silberstein and M. Coornaert, A Garden of Eden theorem for principal algebraic actions, preprint 2017. arXiv:1706.06548. T. Ceccherini-Silberstein, F. Fiorenzi, and F. Scarabotti, The Garden of Eden theorem for cellular automata and for symbolic dynamical systems, in Random walks and geometry, Walter de Gruyter, Berlin, 2004, pp. 73–108. T. Ceccherini-Silberstein, A. Machı̀, and F. Scarabotti, Amenable groups and cellular automata, Ann. Inst. Fourier (Grenoble), 49 (1999), pp. 673–685. M. Coornaert, Topological dimension and dynamical systems, Universitext, Springer, Cham, 2015. Translated and revised from the 2005 French original. M. M. Day, Amenable semigroups, Illinois J. Math., 1 (1957), pp. 509–544. P. de la Harpe, Topics in geometric group theory, Chicago Lectures in Mathematics, University of Chicago Press, Chicago, IL, 2000. F. Fiorenzi, The Garden of Eden theorem for sofic shifts, Pure Math. Appl., 11 (2000), pp. 471–484. , Cellular automata and strongly irreducible shifts of finite type, Theoret. Comput. Sci., 299 (2003), pp. 477–493. J. Franks, Anosov diffeomorphisms, in Global Analysis (Proc. Sympos. Pure Math., Vol. XIV, Berkeley, Calif., 1968), Amer. Math. Soc., Providence, R.I., 1970, pp. 61–93. F. P. Greenleaf, Invariant means on topological groups and their applications, Van Nostrand Mathematical Studies, No. 16, Van Nostrand Reinhold Co., New York, 1969. R. I. Grigorchuk, Degrees of growth of finitely generated groups and the theory of invariant means, Izv. Akad. Nauk SSSR Ser. Mat., 48 (1984), pp. 939–985. M. Gromov, Groups of polynomial growth and expanding maps, Inst. Hautes Études Sci. Publ. Math., (1981), pp. 53–73. , Endomorphisms of symbolic algebraic varieties, J. Eur. Math. Soc. (JEMS), 1 (1999), pp. 109– 197. G. A. Hedlund, Endomorphisms and automorphisms of the shift dynamical system, Math. Systems Theory, 3 (1969), pp. 320–375. D. Lind and B. Marcus, An introduction to symbolic dynamics and coding, Cambridge University Press, Cambridge, 1995. D. Lind and K. Schmidt, Homoclinic points of algebraic Zd -actions, J. Amer. Math. Soc., 12 (1999), pp. 953–980. , Symbolic and algebraic dynamical systems, in Handbook of dynamical systems, Vol. 1A, NorthHolland, Amsterdam, 2002, pp. 765–812. D. Lind and K. Shmidt, A survey of algebraic actions of the discrete Heisenberg group, Uspekhi Mat. Nauk, 70 (2015), pp. 77–142. A. Machı̀ and F. Mignosi, Garden of Eden configurations for cellular automata on Cayley graphs of groups, SIAM J. Discrete Math., 6 (1993), pp. 44–56. A. Manning, There are no new Anosov diffeomorphisms on tori, Amer. J. Math., 96 (1974), pp. 422– 429. J. Milnor, A note on curvature and fundamental group, J. Differential Geometry, 2 (1968), pp. 1–7. E. F. Moore, Machine models of self-reproduction, vol. 14 of Proc. Symp. Appl. Math., American Mathematical Society, Providence, 1963, pp. 17–34. S. Moriceau, Cellular automata on a G-set, J. Cell. Autom., 6 (2011), pp. 461–486. 42 TULLIO CECCHERINI-SILBERSTEIN AND MICHEL COORNAERT [41] S. A. Morris, Pontryagin duality and the structure of locally compact abelian groups, Cambridge University Press, Cambridge-New York-Melbourne, 1977. London Mathematical Society Lecture Note Series, No. 29. [42] D. E. Muller, Unpublished class notes, university of illinois, (1976). [43] J. Myhill, The converse of Moore’s Garden-of-Eden theorem, Proc. Amer. Math. Soc., 14 (1963), pp. 685–686. [44] J. von Neumann, Zur allgemeinen theorie des Masses, Fund. Math., 13 (1929), pp. 73–116. [45] A. J. Ol′ šanskiı̆, On the question of the existence of an invariant mean on a group, Uspekhi Mat. Nauk, 35 (1980), pp. 199–200. [46] D. S. Ornstein and B. Weiss, Entropy and isomorphism theorems for actions of amenable groups, J. Analyse Math., 48 (1987), pp. 1–141. [47] A. L. T. Paterson, Amenability, vol. 29 of Mathematical Surveys and Monographs, American Mathematical Society, Providence, RI, 1988. [48] K. Schmidt, Dynamical systems of algebraic origin, vol. 128 of Progress in Mathematics, Birkhäuser Verlag, Basel, 1995. [49] P. E. Schupp, Arrays, automata and groups—some interconnections, in Automata networks (ArgelèsVillage, 1986), vol. 316 of Lecture Notes in Comput. Sci., Springer, Berlin, 1988, pp. 19–28. [50] M. Shub, Global stability of dynamical systems, Springer-Verlag, New York, 1987. With the collaboration of A. Fathi and R. Langevin, Translated from the French by Joseph Christy. [51] S. Smale, Differentiable dynamical systems, Bull. Amer. Math. Soc., 73 (1967), pp. 747–817. [52] M. C. H. Tointon, Characterizations of algebraic properties of groups in terms of harmonic functions, Groups Geom. Dyn., 10 (2016), pp. 1007–1049. [53] S. Wacker, Cellular automata on group sets and the uniform Curtis-Hedlund-Lyndon theorem, in Cellular automata and discrete complex systems, vol. 9664 of Lecture Notes in Comput. Sci., Springer, [Cham], 2016, pp. 185–198. , The Garden of Eden theorem for cellular automata on group sets, in Cellular automata, vol. 9863 [54] of Lecture Notes in Comput. Sci., Springer, [Cham], 2016, pp. 66–78. [55] P. Walters, Topological conjugacy of affine transformations of tori, Trans. Amer. Math. Soc., 131 (1968), pp. 40–50. [56] S. Wolfram, A new kind of science, Wolfram Media, Inc., Champaign, IL, 2002. Added in proof Recently, after our preprint [18] had circulated, Hanfeng Li posted his paper [60] containing an impressive Garden of Eden type theorem generalizing several results mentioned above in our survey. The key notion in Li’s paper is that of specification, a strong orbit tracing property which was introduced by Rufus Bowen for Z-actions in relation to his studies on Axiom A diffeomorphisms in [57] (see also [59, Definition 21.1]) and was extended to Zd -actions by Ruelle in [61]. Several versions and generalizations of specification have appeared in the literature (see, in particular, [33, Definition 5.1] and [58, Definition 6.1]). Here is the one we need (cf. [58, Definition 6.1]). Definition 9.1. A dynamical system (X, G) has the weak specification property if for any ε > 0 there exists a nonempty symmetric finite subset ∆ ⊂ G satisfying the following property: if (Ωi )i∈I is any finite family of finite subsets of G such that ∆Ωi ∩ Ωj = ∅ for all distinct i, j ∈ I, and (xi )i∈I is any family of points in X, then there exists x ∈ X such that d(sx, sxi ) ≤ ε for all i ∈ I and s ∈ Ωi , where d is any metric compatible with the topology on X. It is straightforward (cf. [60, Proposition A.1]) to check that is G is a countable group, A is a finite alphabet set, and X ⊂ AG is a subshift, then the shift dynamical system (X, G) has the weak THE GARDEN OF EDEN THEOREM: OLD AND NEW 43 specification property if and only if it is strongly irreducible (cf. Section 7). Also, it is easy to see that the weak specification property passes to factors. Li [60, Theorem 1.1] proved the following: Theorem 9.2. Let (X, G) be a dynamical system. Suppose that the group G is amenable and that (X, G) is expansive and has the weak specification property. Then (X, G) has the Myhill property. Note that Theorem 9.2 covers Theorem 8.14, by virtue of the remarks following Definition 9.1. Recall (cf. Example 7.9) that if X ⊂ {0, 1}Z denotes the even subshift (cf. Example 2.4), then (X, Z) is expansive, has the weak specification property (since X is strongly irreducible), but does not have the Moore property. This shows that from the hypotheses of Theorem 9.2 one cannot deduce the Moore property, in general. However, Li [60, Theorem 1.2] proved that when restricting to the class of algebraic dynamical systems (cf. Section 8.6) with amenable acting group, the Moore property follows from expansiveness and weak specification: Theorem 9.3. Let (X, G) be an algebraic dynamical system. Suppose that the group G is amenable and that (X, G) is expansive and has the weak specification property. Then (X, G) has the Moore property. As an immediate consequence of Theorems 9.2 and 9.3, one deduces the following (cf. [60, Theorem 1.3]): Corollary 9.4. Let (X, G) be an algebraic dynamical system. Suppose that the group G is amenable and that (X, G) is expansive and has the weak specification property. Then (X, G) has the Moore-Myhill property. Moreover, since (cf. [60, Lemma 2.1], see also [58, Theorem 1.2]) every expansive principal algebraic action has the weak specification property, one immediately deduces (cf. [60, Theorem 1.3]): Corollary 9.5. Let (X, G) be a principal algebraic dynamical system. Suppose that the group G is amenable and that (X, G) is expansive. Then (X, G) has the Moore-Myhill property. Note that Corollary 9.5 covers Theorem 8.19 (since every abelian group is amenable and no assumption on connectedness of the phase space is required). Additional references [57] R. Bowen, Periodic points and measures for Axiom A diffeomorphisms, Trans. Amer. Math. Soc., 154 (1971), pp. 377–397. [58] N. Chung and H. Li, Homoclinic groups, IE groups, and expansive algebraic actions, Invent. Math., 199 (2015), no.3, pp. 805-858. [59] M. Denker, Ch. Grillenberger, and K. Sigmund, Ergodic theory on compact spaces. Lecture Notes in Mathematics, Vol. 527. Springer-Verlag, Berlin-New York, 1976. [60] H. Li, Garden of Eden and specification, preprint 2017. arXiv:1708.09012. [61] D. Ruelle, Statistical mechanics on a compact set with Z ν action satisfying expansiveness and specification, Trans. Amer. Math. Soc., 187 (1971), pp. 237–251. Dipartimento di Ingegneria, Università del Sannio, C.so Garibaldi 107, 82100 Benevento, Italy E-mail address: [email protected] Université de Strasbourg, CNRS, IRMA UMR 7501, F-67000 Strasbourg, France E-mail address: [email protected]
7cs.IT
Optimal Covariance Change Point Detection in High Dimension Daren Wang1 , Yi Yu2 , and Alessandro Rinaldo1 arXiv:1712.09912v1 [math.ST] 28 Dec 2017 1 Department of Statistics and Data Science, Carnegie Mellon University 2 School of Mathematics, University of Bristol Abstract We study the problem of change point detection for covariance matrices in high dimensions. We assume that {Xi }i=1,...,n is a sequence of independent, centered p-dimensional sub-Gaussian random vectors is observed whose covariance matrices are piece-wise constant. Our task is to recover with high accuracy the number and locations the of change points, which are unknown. Our generic model setting allows for all the model parameters to change with n, including the dimension p, the minimal spacing between consecutive change points, the magnitude of smallest change and the maximal operator norm of the covariance matrices of the sample points. We introduce two procedures, one based on the binary segmentation algorithm (e.g. Vostrikova, 1981) and the other on its extension known as wild binary segmentation of Fryzlewicz (2014), and demonstrate that, under suitable conditions, both are able to consistently estimate the number and locations of change points. Our second algorithm, called Wild Binary Segmentation through Independent Projection (WBSIP) is shown to be minimax optimal in the in terms of all the relevant parameters. Our minimax analysis also reveals a phase transition effect based on our generic model setting. To the best of our knowledge, this type of results has not been established elsewhere in the change point detection literature. Keywords: Change point detection; High dimensional covariance estimation; High dimensional covariance testing; Binary segmentation; Wild binary segmentation; Random projection; Minimax optimal. 1 Introduction Change point detection in time series data has a long history and can be traced back to World War II, during which the change point detection was specifically demonstrated in a sequential fashion and heavily used in quality control. Wald (1945) was published shortly after the war and formally introduced sequential probability ratio test (SPRT), which consists of simple null and alternative hypotheses and is based on sequentially calculating the cumulative sum of log-likelihood ratios. Page (1954) relaxes the statistic to the cumulative sum of weights, which are not restricted to loglikelihood ratios, namely cumulative sum (CUSUM). Since then, a tremendous amount of efforts have been dedicated to this problem (see, e.g. James et al., 1987; Siegmund and Venkatraman, 1995) with applications ranging from clinical trials to education testing, signal processing and cyber security. In the last few decades, with the advancement of technology, an increasing demand from emerging application areas, e.g. finance, genetics, neuroscience, climatology, has fostered the development of statistical theorties and methods for change point detection also in an offline fashion. Given a 1 time series data set, instead of assuming it is stationary over the whole time course, it is more robust and realistic to assume that it is stationary only within time segments, and the underlying models may change at certain time points. Formally, we assume that we observed n variables X1 . . . , Xn such that,   F1 , i = 1, . . . , t1 − 1,    F , i = t , . . . , t − 1, 2 1 2 Xi ∼ (1)  . . . ,    F , i = t K K−1 , . . . , n, where F1 , . . . , FK , K ≥ 2, are distribution functions, Fk−1 6= Fk , k = 2, . . . , K, and {t1 , . . . tK−1 } ⊂ {2, . . . , n − 1} are unknown change point locations. Testing the existence of change points and estimating the locations of the change points are of primary interest. The simplest and best studied scenario of model (1) is that {Xi = fi + εi }ni=1 is a univariate time series, and {fi }ni=1 is piecewise constant signal. Change point detection here means detecting changes in the mean of a univariate time series. An abundance of literature exists on this model and variants thereof. See Section 1.1 below for a literature review. The properties of the covariance are also of key importance in statistics, and detecting the covariance changes in a time series data set also has a long history. See Section 1.1 below for some literature review. In this paper we investigate the problem of change point detection and localization for covariance matrices. We consider what is arguably the most basic setting for the problem in (1): we observed a sequence (X1 . . . , Xn ) of independent and centered random vectors in Rd wih respective covariance matrices (Σ1 , . . . , Σn ) with that maxi kΣi kop ≤ B. We assume that Σi = Σi+1 for all coordinates i except for a few, which we refer to as the change points. We denote with ∆ the minimal distance between two consecutive change points and with κ the magnitude of the smallest change, measured by the operator norm of the difference between the covariance matrices at two consecutive change points. The parameters p, ∆, B and κ completely characterize the difficulty of change point localization problem, which, for a given n, is increasing in p and B and decreasing in ∆ and κ. In fact, for reason that will become clear later on, it will be convenient to aggregate the parameter B and κ into one parameter κ , B2 which effectively plays the role of the signal to noise ratio. Our goal to identify all the change points and to localize them, i.e. estimate their values accurately. In our analysis, we allow all the relevant parameters p, ∆ and Bκ2 to change with n and derive upper and lower bounds for the localization rate in terms of such parameters. We propse two algorithms for covariance change point localization. The first one, called BSOP (Binary Segmentation in Operator Norm), is based on an adaptation to the covariance setting of the popular binary segmentaiton (BS) algorithm (e.g. Vostrikova, 1981) for change point detection of a univariate signal. The BSOP algorithm, described below in Algorithm 1, is rather simple to implement. Under appropriate assumptions we show in Theorem 1 that BSOP can consistently estimate all the change point, with a localization rate that is sub-optimal, especially, when the dimension p is allowed to be increasing with n. Our second algorithm, called WBSIP (Wilde Binary Segmentation through Independent Projections) is significantly more refined, as it combines data splitting, random projections and the wild binary segmentation (WBS) procedure of Fryzlewicz (2014); see Algorithm 3. For all its complexity, WBSIP yields much sharper localization rates 2 than WBOP under a different, milder set of assumptions; see Theorem 2. We further obtain lower bounds on the localization rate and demonstrate that WBSIP is in fact minimax rate optimal. Interestingly, our analysis reveals  a phase transition for the problem of covariance change point localization: if ∆ = Θ B 4 pκ−2 then no estimator will be able to estimate the location of the change point at a vanishing rate; see Lemma 3. On the other hand, if ∆ = Ω B 4 pκ−2 log n then WBSIP will guarantee a localization rate of the order B 4 κ−2 log n, which, up to a log n term, is a minimax optimal rate; see Lemma 4. It appears that this type of phase transition is unique to our generic model setting. In fact, many authors (e.g. Aue et al., 2009; Avanesov and Buzun, 2016) have already studied the consistency of change point estimation in covariate structure. To the best of our knowledge, this phase transition based on the spacing parameter ∆ and the dimension p has not been established elsewhere. Overall, our results provide a complete characterization of the problem of change point localization in the covariance setting described above. 1.1 Relevant and Related Literature For the classical problem of change point detection and localization of the mean of a univariate time series, least squares estimation is a natural choice. For example, Yao and Au (1989) uses the least squares estimators to show that the distance between the estimated change points and the truth are within Op (1). Lavielle (1999) and Lavielle and Moulines (2000) consider penalized least-squares estimation and prove consistency in the presence of strong mixing and long-range dependence of the error terms εi ’s. Harchaoui and Lévy-Leduc (2010) consider the least squares criterion with a total variation. Besides least squares estimation based methods, other attempts have also been made. For instance, Davis et al. (2006) propose a model based criterion on an autoregressive model. Wang (1995) detects the change points using fast discrete wavelet transform. Harchaoui and Lvy-Leduc (2010), Qian and Jia (2012), Rojas and Wahlberg (2014) and Lin et al. (2017) consider the fused lasso (Tibshirani et al., 2005) to penalise the differences between the signals at two consecutive points to detect the change points. Frick et al. (2014) introduce a multiscale statistic based procedure, namely simultaneous multiscale change point estimator (SMUCE), to detect change points in piecewise constant signals with errors from exponential families. More recently, Li et al. (2017) extends the multiscale method to a class of misspecified signal functions which are not necessarily piecewise constants. The method of Davies and Kovac (2001) can also be used to determine the location of change points in a piece-wise constant signal. Chan and Walther (2013) study the power of the likelihood ratio test statistics to detect the presence of change points. Among all the methods, binary segmentation (BS) (e.g. Vostrikova, 1981) is ‘arguably the most widely used change-point search method’ (Killick et al., 2012). It goes through the whole time course and searches for a change point. If a change point is detected, then the whole time course is split into two, and the same procedure is conducted separately on the data sets before and after the detected change point. The procedure is carried on until no change point is detected, or the remaining time course consists of too few time points to continue the test. Venkatraman (1992) proves the consistency results of the BS method in univariate time series mean change point detection, with the number of change points allowed to increase with the number of time points. It is worth to mention a variant of BS, namely wild binary segmentation (WBS), which is proposed in Fryzlewicz (2014) and which can be viewed as a flexible moving window techniques, or a hybrid of moving window and BS. Instead of starting with the whole data set and doing binary segmentation, WBS randomly draws a collection of intervals under certain conditions, conducts 3 BS on each interval, and return the one which has the most extreme criterion value among all the intervals. Compared to BS, under certain conditions, WBS is more preferable when multiple change points are present. In the univariate time series mean change point detection problem, Venkatraman (1992) shows that in order to achieve the estimating consistency using BS algorithm, the minimum gap between two consecutive change points should be at least of order n1−β , where n is the number of time points, and 0 ≤ β < 1/8; as claimed in Fryzlewicz (2014), by using WBS algorithm, this rate can be reduced to log(n). All the literature mentioned above tackles univariate time series models, however, in the big data era, data sets are now routinely more complex and often appear to be multi- or high-dimensional, i.e. Xi ∈ Rp , where p is allowed to grow with the number of data points n. Horváth and Hušková (2012) propose a variant of the CUSUM statistic by summing up the square of the CUSUM statistic in each coordinate. Cho and Fryzlewicz (2012) transform a univariate non-stationary time series into multi-scale wavelet regime, and conduct BS at each scale in the wavelet context. Jirak (2015) allows p to tend to infinity together with n, by taking maxima statistics across panels coordinatewise. Cho and Fryzlewicz (2015) propose sparsified binary segmentation method which aggregates the CUSUM statistics across the panel by adding those which exceed a certain threshold. Cho (2015) proposes the double CUSUM statistics which, at each time point, picks the coordinate which maximizes the CUSUM statistic, and de facto transfers the high-dimensional data to a univariate CUSUM statistics sequence. Aston and Kirch (2014) introduces the asymptotic concept of highdimensional efficiency which quantifies the detection power of different statistics in this setting. Wang and Samworth (2016) study the problem of estimating the location of the change points of a multivariate piecewise-constant vector-valued function under appropriate sparsity assumptions on the number of changes. As for change point detection in more general scenarios, the SPRT procedure (Wald, 1945) can be easily used for the variance change point detection. Based on a generalized likelihood ratio statistic, Baranowski et al. (2016) tackle a range of univariate time series change point scenarios, including the variance change situations, although theoretical results are missing. Picard (1985) proposes tests on the existence of change points in terms of spectrum and variance. Inclan and Tiao (1994) develop an iterative cumulative sums of squares algorithm to detect the variance changes. Gombay et al. (1996) propose some tests on detection of possible changes in the variance of independent observations and obtained the asymptotic properties under the non-change null hypothesis. Berkes et al. (2009), among others, extend the tests and corresponding results to linear processes, as well as ARCH and GARCH processes. Aue et al. (2009) considers the problem of variance change point detection in a multivariate time series model, allowing the observations to have m-dependent structures. Note that the consistency results in Aue et al. (2009) are in the asymptotic sense that the number of time points diverges and the dimension of the time series remains fixed. Aue et al. (2009) also require the existence of good estimators of the covariance and precision matrices, and the conditions thereof are left implicit. Barigozzi et al. (2016) deal with a factor model, which is potentially of high dimension p/n = O(log2 (n)), and use the wavelet transforms to make the data possibly dependent across the timeline. Note that the model in Barigozzi et al. (2016) can be viewed as a specific covariance change point problem, where the additional structural assumption allows the dimensionality to go beyond the sample size. As for the problem of hypothesis testing for high dimensional covariance matrices, which corresponds to the problem of change point detection, the literature is also abundant. In the cases where p fixed and n → ∞, the likelihood ratio test (Anderson, 2003) has a χ2p(p+1)/2 limiting distribution 4 under H0 : Σ = I. When both n, p → ∞ and p/n → c ∈ (0, ∞), Johnstone (2001) extended Roy’s largest root test (Roy, 1957) and derived the Tracy–Widom limit of its null distribution, to name but a few. In the case where n, p → ∞ and p/n → ∞, Birke and Dette (2005) derived the asymptotic null distribution of the Ledoit–Wolf test (Ledoit and Wolf, 2002). More recently, Cai and Ma (2013) studied the testing problem in the setting p → ∞ from p a minimax point of view and derived the testable region in terms of Frobenius norm is of order p/n. Organization of the paper The rest of the paper is organized as follows. In Section 2, we propose two different methods for covariance change point detection problem. All three methods are shown to be consistent, but under different conditions, and the wild binary segmentation with random projection has the location error rate being log(n). In Section 3, we show the lower bound of the location error rate in the covariance change point detection problem is log(n), which implies that the wild binary segmentation with random projection is minimax optimal. Further discussion and future work directions can be found in Section 4. Notation For a vector v ∈ Rp and matrix Σ ∈ Rp×p , kvk and kΣkop = maxkvk=1 |v ⊤ Σv| indicate the Euclidean Rand the operator norm, respectively; for any integrable function f (·) : R 7→ R, denote kf k1 = x∈R |f (x)|dx as the ℓ1 -norm of f (·). For functions f (·) and g(·) satisfying f (n)/g(n) → a, if a ∈ (0, ∞), we denote by f (n) = O(g(n)). 2 Main Results In this paper, we study the covariance change point detection in high dimension. To be specific, we consider a centered and independent time series {Xi }ni=1 ⊂ Rp⊗n . Let {ηk }K k=1 ⊂ {1, . . . , n} be the collection of time points at which the covariance matrices of Xi ’s change. The model is formally summarized as follows with Orlicz norm k · kψ2 , which is defined in Definition 2 in Appendix A. Assumption 1. Let X1 , . . . , Xn ∈ Rp be independent sub-Gaussian random vectors such that E(Xi ) = 0 , E(Xi Xi′ ) = Σi and kXi kψ2 ≤ B < ∞ for all i. Let {ηk }K+1 k=0 ⊂ {0, . . . , n} be a collection of change points, such that η0 = 0 and ηK+1 = n and that Σηk +1 = Σηk +2 = . . . = Σηk+1 , for any k = 0, . . . , K. Assume the jump size κ = κ(n) and the spacing ∆ = ∆(n) satisfy that inf {ηk − ηk−1 } ≥ ∆ > 0, k=1,...,K+1 and kΣηk − Σηk−1 kop = κk ≥ κ > 0, for any k = 1, . . . , K + 1. Remark 1. The parameter B. The assumption that maxi kXi kψ2 ≤ B is imposed in order to control the order of magnitude of the random fluctuations of the CUSUM covariance statistics, given below in Definition 1, around its means. See Lemma 5 in the Appendix for details. At the same 5 time, the chain of inequalities (2) shows also that maxi kΣi kop ≤ 2B 2 , so that the same assumptions amounts also to a uniform upper bound on the operator norm of the covariance matrices of the date points. In this regard, B may be reminiscent of the assumption that the signal be of bounded magnitude sometimes used in the litarature on mean change point detection: see, e.g. Fryzlewicz (2014, Assumption 3.1(ii)) and Venkatraman (1992, Condition (iii) on page 11). However, while in the mean change point detection problem such boundedness assumption can be in fact removed because the optimal solution is translation invariant (see, e.g., Wang and Samworth, 2016), this is not the case in the present setting. Indeed, a constant shift in the largest eigenvalue of the population covariance matrices also implies a change in the precision with which such matrices can be estimated. Thus, it is helpful to think of B 2 as some form of variance term. Remark 2. Relationship between κ and B. The parameters κ and B are not variation independent, as they satisfy the inequality κ ≤ B 2 /4. In fact,   K n n n κ ≤ max kΣηk − Σηk−1 kop ≤ 2 max kΣi kop = 2 max sup E (v ⊤ Xi )2 ≤ 4 max kXi k2ψ2 ≤ 4B 2 , (2) i=1 k=1 i=1 v∈S p−1 i=1 where the second-to-last inequality follows from Equation (16) in Appendix A. For ease of readability, we will instead use the weaker bound κ ≤ B 2 throughout. In fact, in our analysis we will quantify the combined effect of both κ and B with their ratio Bκ2 , which we will refer to as the signal-to-noise ratio. Larger values of such quantities lead to better performance of our algorithm. It is important to notice that the signal-to-noise ratio, and the task of change-point detection itself, remains invariant with respect to any multiplictive rescaling of the data by an arbitsraty non-zero constant. In our model Assumption 1, all the relevant parameters p, ∆, K, B and κ are allowed to be functions of the sample size n, although we do make this dependence explicit in our notation for ease of readability. This generic setting allows us to study the covariance change point problem with potential high dimensional data and with growing number of change points and decreasing jump sizes. Motivated by the univariate CUSUM statistic for mean change point detection, we define the CUSUM statistic in the covariance context. Definition 1 (Covariance CUSUM). For X1 , . . . , Xn ∈ Rp , a pair of integers (s, e) such that 0 ≤ s < e − 1 < n, and any t ∈ {s + 1, . . . , e − 1}, the covariance CUSUM statistic is defined as s s t e X X e − t t−s s,e ⊤ e Xi Xi − Xi Xi⊤ . St = (e − s)(t − s) (e − s)(e − t) i=s+1 i=t+1 Its expected value is e s,e Σ t = s t X e−t Σi − (e − s)(t − s) i=s+1 s e X t−s Σi . (e − s)(e − t) i=t+1 In the rest of this section, we propose two algorithms to detect the covariance change points. Detailed algorithms and the main consistency results are presented in this section, with the proofs provided in the Appendices. The advantages of each algorithm will be discussed later in the section. 6 2.1 Consistency of the BSOP algorithm We begin our study by analyzing the performance of a direct adaptation of the binary segmentation (BS) algorithm to the matrix setting based on the distance induced by the operator norm. The resulting algorithm, which we call BSOP, is given in Algorithm 1. The BSOP procedure works as follows: given any time interval (s, e), BSOP first computes the maximal operator norm of the covariance CUSUM statistics over the time points in (s + ⌈p log n⌉, e − ⌊p log n⌋); if such maximal value exceeds a predetermined threshold τ , then BSOP will identify the location b of the maximum as a change point. The interval (s, e) is then split into two subintervals at b and the procedure is then iterated separately on each of the resulting sub-intervals (s, b) and (b, e) until an appropriate stopping condition is met. The BSOP algorithm differs from the standard BS implementation in one aspect: the maximization of the of the operator norm of the CUSUM covariance operator is carried out only over the time points in (s, e) that are away by at least p log n from the endpoints of the interval. Such restriction is needed to obtain adequate tail bounds for the operator norm of the covariance CUSUM statistics S̃ts,e given in Definition 1 and of the centered and weighted empirical covariance matrices. See Lemma 5 in Appendix A. Algorithm 1 Binary Segmentation through Operator Norm. BSOP((s, e), τ ) INPUT: {Xi }ei=s+1 ⊂ Rp⊗(e−s), τ > 0. Initial FLAG ← 0, while e − s > 2p log(n) + 1 and FLAG = 0 do a ← max⌈s+p log(n)⌉≤t≤⌊e−p log(n)⌋ Sets,e op if a ≤ τ then FLAG ← 1 else b ← arg max⌈s+p log(n)⌉≤t≤⌊e−p log(n)⌋ Sets,e op add b to the collection of estimated change points BSOP((s, b − 1), τ ) BSOP((b, e), τ ) end if end while OUTPUT: The collection of estimated change points. Remark 3 (Uniqueness of the maxima of the covariance CUSUM statistic.). The probability that the maximiser of kSets,e kop over multiple time points is non unique is zero. To analyze the performance of the BSOP algorithm we will impose the following assumption, which is, for the most part, modeled after Assumption 3.2 in Fryzlewicz (2014). Assumption 2. For a sufficiently large constant Cα > 0 and sufficient small constant cα , cκ > 0, assume that ∆κB −2 ≥ Cα nΘ , p ≤ cα n8Θ−7 / log(n), where Θ ∈ (7/8, 1]. When the parameters κ and B are fixed, the above assumption requires ∆, the minimal spacing between consecutive change point, to be of slightly smaller order than n, the size of the time series. This is precisely Assumption 3.3 in Fryzlewicz (2014) (see also Cho and Fryzlewicz, 2015). The 7 fact that ∆ cannot be too small compared to n in order for the BS algorithm to exhibit good performance is well known: see, e.g., Olshen et al. (2004). In Assumption 2, we require also the 8Θ−7 dimension p to be upper bounded by nlog n , which means that p is allowed to diverge as n → ∞. Remark 4 (Generalizing Assumption 2). In Assumption 2 we impose certain constraints on the scaling of the quantities B, κ, ∆ and p in relation to n that are captured by a single parameter Θ, whose admissible values lye in (7/8, 1]. The strict lower bound of 7/8 on Θ is determined by the calculations outlined below on (5) and (6), which are needed to ensure the existence of a non-empty range of value for the input parameter τ to the BSOP algorithm. In fact, Assumption 2 may be generalized by by allowing for different types of scaling in n of the signal-to-noise ratio κB −2 , the minimal distance ∆ between consecutive change points and the dimension p. In detail, we may require that κB −2  nΘ1 , ∆  nΘ2 and p log n  nΘ3 for a given triplet of parameters (Θ1 , Θ2 , Θ3 ) in an appropriate subset of [0, 1]3 . Such a generalization would then lead to consistency rates in n that depend on all these parameters simultaneously. However, the range of allowable values of (Θ1 , Θ2 , Θ3 ) is not a product set due to non-trivial constraints among them. We will refrain from providing details and instead rely on the simpler formulation given in Assumption 2. b Theorem 1 (Consistency of BSOP). Under Assumptions 1 and 2, let B = {η̂k }K k=1 be the collection of the estimated change points from the BSOP((0, n), τ ) algorithm, where the parameter τ satisfies p √ (3) B 2 p log(n) + 2 ǫn B 2 < τ < C1 κ∆(e − s)−1/2 , for some constant C1 ∈ (0, 1) and where ǫn = C2 B 2 κ−1 n5/2 ∆−2 for some C2 > 0. Then,  b =K P K and p p log(n)  max |ηk − η̂k | ≤ ǫn ≥ 1 − n3 9p 2n−cp , k=1,...,K (4) for some absolute constant c > 0. Remark 5. The condition (3) on the admissible values of the input parameter τ to the BSOP algorithm is well defined. Indeed, by Assumption 2, for all pairs (s, e) such that e − s > 2p log n, we have that p cα κ∆n−1/2 ≤ (1/8)κ∆(e − s)−1/2 (5) B 2 p log(n) ≤ B 2 cα n4Θ−7/2 ≤ B 2 cα nΘ n−1/2 ≤ Cα and √ 1/2 1/2 1/2 5/4+Θ−7/4 2 ǫn B 2 = 2C2 B 3 κ−1/2 n5/4 ∆−1 (p log(n))1/4 ≤ (2C2 Cα−1 c1/4 n α )Bκ ≤ (C1 /8)κ∆n−1/2 B −1 κ1/2 ≤ (C1 /8)κ∆(e − s)−1/2 , (6) where in the chain of inequalities we have used Assumption 2 repeatedly. It is also worth noting that the difference between the right hand side and the left hand side of (3) increase as Θ increases to 1. Finally we remark that in the proof of Theorem 1, we actually let C1 = 1/8, but this is an arbitrary choice and it essentially depends on the constants Cα and cα from Assumption 2. 8 Proof of Theorem 1. By induction, it suffices to consider any pair of integers s and e such that (s, e) ⊂ (0, T ) and satisfying ηr−1 ≤ s ≤ ηr ≤ . . . ≤ ηr+q ≤ e ≤ ηr+q+1 , q ≥ −1, max{min{ηr − s, s − ηr−1 }, min{ηr+q+1 − e, e − ηr+q }} ≤ ǫn , where q = −1 indicates that there is no change point contained in (s, e). It follows that, for sufficiently small cα and sufficiently large Cα , p C2 B 2 κ−1 n5/2 p log(n)∆−2 ǫn ≤ ∆/4 ∆/4 1/2 ≤ 4C2 B 2 κ−1 n5/2 cα n4Θ−7/2 Cα3 κ−3 B 6 n3Θ −3 2 −4 Θ−1 ≤ (4C2 c1/2 α Cα )(κ B )n ≤ (κ2 B −4 )nΘ−1 ≤1 where the second inequality stems from Assumption 2, the third inequality holds by choosing sufficiently small cα and sufficiently large Cα and the last inequality follows from the fact that κ ≤ B 2 . Then, for any change point should be (s, e), it is either the case that |ηp − s| ≤ ǫn , or that |ηp − s| ≥ ∆ − ǫn ≥ ∆ − ∆/4 = 3∆/4. Similar considerations apply to the other endpoint e. As a consequence, the fact that min{|ηp − e|, |ηp − s|} ≤ ǫn implies that ηp is a detected change point found in the previous induction step, while if min{ηp − s, ηp − e} ≥ 3∆/4 we can conclude that ηp ∈ (s, e) is an undetected change point. In order to complete the induction step, it suffices to show that BSOP([s, e], τ ) (i) will not find any new change point in the interval (s, e) if there is none, or if all the change points in (s, e) have been already detected and (ii) will identify a location b such that |ηp − b| ≤ ǫn if there exists at least one undetected p change point in (s, e). Set λ = B 2 p log(n). Then, the event A1 ({Xi }ni=1 , λ) defined in Equation (17) holds with probbability at least 1−2×9p n3 ncp , for some universal constant c > 0. The proof will be completed in two steps. Step 1. First we will show that on the event A1 ({Xi }ni=1 , λ), BSOP((s, e), τ ) can consistently detect or reject the existence of undetected change points within (s, e). Suppose there exists ηp ∈ (s, e) such that min{ηp − s, ηp − e} ≥ 3∆/4. Set δ = p log n. Then 3 ∆, since δ ≤ 32 p log(n) ≤ cα n8Θ−7 ≤ cα nΘ ≤ cα Cα−1 ∆B −2 κ1 ≤ 3∆/32, where the last inequality follows from Assumption 2. With this choice of δ, we apply Lemma 20 in Appendix D (where we set c1 = 3/4) and obtain that max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e kop ≥ (3/8)κ∆(e − s)−1/2 . kΣ t 9 On the event A1 ({Xi }ni=1 , λ), max t=⌈s+δ⌉,...,⌊e−δ⌋ kSets,e kop ≥ max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e kop − λ ≥ (3/8)κ∆(e − s)−1/2 − λ ≥ (1/8)κ∆(e − s)−1/2 , kΣ t (7) where the last inequality follows from (5) (in the last step we have set C1 = 1/8). If (3) holds, then, on the event A1 ({Xi }ni=1 , λ), BSOP((s, e), τ ) detects the existence of undetected change points if there are any. Next, suppose there does not exist any undetected change point within (s, e). Then, one of the following cases must occur. (a) There is no change point within (s, e); (b) there exists only one change point ηr within (s, e) and min{ηr − s, e − ηr } ≤ ǫn ; (c) there exist two change points ηr , ηr+1 within (s, e) and that max{ηr − s, e − ηr+1 } ≤ ǫn . Observe that if case (a) holds then, on the event A1 ({Xi }ni=1 , λ), we have that max t=⌈s+δ⌉,...,⌊e−δ⌋ kSets,e kop ≤ max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e kop + λ = λ < τ, kΣ t where the last inequality follows from (3). If situation (c) holds, then on, the event A1 ({Xi }ni=1 , λ), we have max t=⌈s+δ⌉,...,⌊e−δ⌋ kSets,e kop ≤ max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e kop +λ ≤ max{kΣ e s,e kop , kΣ e s,e kop }+λ ≤ 2√ǫn B 2 +λ, kΣ ηr ηr+1 t where the first inequality follows from A1 ({Xi }ni=1 , λ), the second inequality from Lemma 19 and the third inequality from Lemma 23. (Both Lemmas are in Appendix D.2.) Case (b) can be handled in a similar manner. Thus, if (3) holds, then on the event A1 ({Xi }ni=1 , λ), BSOP((s, e), τ ) has no false positives when there are no undetected change points in (s, e). Step 2. Assume now that there exists a change point ηp ∈ (s, e) such that min{ηp −s, ηp −e} ≥ 3∆/4 and let b ∈ arg max Sets,e op . t=⌈s+δ⌉,...,⌊e−δ⌋ To complete the proof it suffices to show that |b − ηk | ≤ ǫn . Let v be such that v ∈ arg max |u⊤ Sebs,e u|. kuk=1 Consider the univariate time series {Yi (v)}ni=1 and {fi (v)}ni=1 defined in (53) and (54) in Appendix D.2. By Lemma 21 there, b ∈ arg maxs≤t≤e |Yet (v)|. Next, we wish to apply Corollary 10 to the time series {Yi (v)}ei=s and {fi (v)}ei=s . Towards that end, we first need to ensure that the conditions required for that result to hold are verified. (Notice that in the statement of Corollary 10, the fi ’s are assumed to be uniformly bounded by B1 , while in this proof the fi (v)’s defined in (54) are assumed to be bounded by 2B 2 .) First, the collection of the change points of the time series {fi (v)}ei=s+1 is a subset of {ηk }K+1 k=0 ∩ (s, e). The condition (26) and the inequality 10 √ 2 δB 2 ≤ (3c1 /4)κ∆(e − s)−1/2 are straightforward consequences of Assumption 2, while (34) follows from the fact that e s,e |fets,e (v) − Yets,e (v)| ≤ Sets,e − Σ t op ≤ λ. Similarly, (33) stems from the relationships max t=⌈s+δ⌉,...,⌊e−δ⌋ |Yets,e (v)| = max t=⌈s+δ⌉,...,⌊e−δ⌋ kSets,e kop ≥ max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e kop −λ ≥ (1/8)κ∆(e−s)−1/2 kΣ t where the first inequality holds under the event A1 ({Xi }ni=1 , λ) and the second inequality is due to (7) and Assumption 2. Thus, all the assumptions of Corollary 10 are met. An application of that result yields that there exists ηk , a change point of {fi (v)}ei=s satisfying (29), such that |b − ηk | ≤ C2 λ(e − s)5/2 ∆−2 κ−1 ≤ ǫn . The proof is complete by observing that (29) implies min{ηk − s, ηk − e} ≥ 3∆/4, as discussed in the argument before Step 1. The proof of the theorem relies on a non-trivial extension of the arguments for proving consistency of the BS algorithm in one-dimensional mean change-point detection problems, as done in Venkatraman (1992) (see also Fryzlewicz, 2014). The main difficulty that prevents a direct e s,e application of those results is the fact the regions of monotonicity of the function t 7→ Σ t op are hard to derive. Instead, for each par of integers 1 ≤ s < e ≤ n with e − s > 2p log n, we study the one-dimensinal time series {(v ⊤ Xi )2 }i=1,...,n of the squared coefficients of the projeciton of the data along a one-dimensional linear subspace spanned by a distinguished unit vector v, e s,e , where which we term the shadow vector. This is simply the leading singular vector of Σ b op e s,e . As it turns out, with such a choice of the shadow vector, b = arg maxt∈(s+p log n,e−p log n) Σ t op the local maxima of CUSUM statistic applied to the corresponding one-dimensional time series coincide with the local maxima of the time series of the values of the operator norm of the CUSUM covarance statistics. As a result, for the purpose of detecting local maxima of the CUSUM covariance statistic, it is enough and in fact much simpler to study the univariate times series of the squared projections onto the appropriate ghost vector. Of course, at each iteration of the BSOP algorithm a new ghost vector and a new univarite time series is obtained and a new local maximum is found. See Appendix D.2 for further comments on the uses and intepretation of the shadow vector. Theorem 1 implies, that with high probability, the BSOP algorithm will identify all the change points and estimate their locations with an error that is bounded by ǫn  B 2 −2 5/2 p ∆ n p log n. κ Noitce that, as expected, the performance of BSOP is decreasing in the inverse of the signal-tonoise ratio parmeter Bκ2 , the inverse of the minima distance ∆ between change points and the dimension p. The above bound yields a family of rates of consistency for BSOP, depending on the scaling of each of the quantities involved in it. For example, in the simplest and most favorable 11 scenario whereby B, κ and the dimension p are constants, the bound implies a rate for change point localization of the order p ǫn  n−2Θ+3/2 log n, n which is decreasing in the Θ ∈ (7/8, 1]. In particular, when the number q of change points is also kept constant, we have that Θ = 1, yielding a localization rate of order logn n . As we will see in the next section, the dependence on the parameters B, κ and ∆ is suboptimal. The advantage of BSOP over the rate-optimal algorithm we introduce next is that BSOP only requires one input parameter, the threshold value τ . Furthermore, when the spacing parameter ∆ is comparable with n and the dimension p of the data grows slowly with respect with n, then BSOP can still deliver good consistency rates. Therefore, despite its suboptimality in general, BSOP is a simple and convenient algorithm which may serve as a competitive benchmark for other proceudres. 2.2 Consistency of the WBSIP algorithm In this section we describe and analyze the performance of a new algorithm for covariance changepoint detection, which we term WBSIP for Wild Binary Segmentation through Independent Projections. The WBSIP algorithm is a generalization of the wild binary segmentation or WBS procedure of Fryzlewicz (2014) for mean change-point detection and further exploits the properties shadow vectors. The WBSIP procedure begins by splitting the data into halves and by selecting at random a collection of M pairs of integers (s, e) such that 1 ≤ s < e ≤ n and e−s > p log n+1. In its second step, WBSIP computes, for each of the M random integer intervals previously generated, a shadow vector using one half od the data and its corresponding one-dimensional time series using the other half. The final step of the procedure is to apply the WBS algorithm over the resulting univariate time series. The details of the algorithm are given in Algorithm 2, which describes the computation of the shadow vectors by principal component methods, and Algorithm 3, which implements WBS to the resulting one dimensional time series. Remark 6. The idea of combining the WBS algorithm with sample splitting is due to Wang and Samworth (2016), who applied it to the problem of mean change-point detection in multivariate settings. Like us, Wang and Samworth (2016) employ one half of the sample to compute an optimal direction and then project the second half of the data along such direction to obtain a univariate time series which can then be handled using WBS. One of the main differences between our approach and that of Wang and Samworth (2016) is the fact that we do no require the shadow vectors to be consistent estimators of subspaces related the true covariance matrices. In particular, our analysis holds without any eigengap assumptions. In order to analyze the performance of the WBSIP procedure, we will impose the following assumption, which is significantly weaker then Assumption 2. Assumption 3. There exists a sufficiently large absolute constant C > 0 such that ∆κ2 ≥ Cp log(n)B 4 . Remark 7. We recall that all the parameters ∆, κ, p, B are allowed to depend on n. Since κ ≤ B 2 , and assuming without loss of generality that the constant C in the previous assumption is larger than 8, we further have that p log(n) ≤ ∆κ2 B −4 C −1 ≤ ∆/8, 12 √ which is used repeatedly below. In fact, in the proof we will set C = 32 2. This choice is of course arbitrary and is only made for convenience in carrying out the calculations below. We also recall that the ratio B 4 κ−2 is invariant of multiplicative scaling, i.e. if Xi′ = αXi for all i, then the corresponding ratio B 4 κ−2 stays the same. Algorithm 2 Principal Component Estimation PC({Xi }ni=1 , {(αm , βm )}M m=1 ) INPUT: {Xi }ni=1 , {(αm , βm )}M m=1 for m = 1, . . . , M do if βm − αm > 2p log(n) + 1 then dm ← arg max⌈αm +p log(n)⌉≤t≤⌊βm −p log(n)⌋ kSetαm ,βm kop um ← arg maxkvk2 =1 v ⊤ Sedαmm ,βm v| else um ← 0 end if end for OUTPUT: {um }M m=1 . Algorithm 3 Wild Binary Segmentation through Independent Projection. {(αm , βm )}M m=1 , τ, δ) INPUT: Two independent samples {Wi }ni=1 , {Xi }ni=1 , τ , δ. n M {um }M m=1 = P C({Wi }i=1 , {(αm , βm )}m=1 ) for i ∈ {s, . . . , e} do for m = 1, . . . , M do 2 Yi (um ) ← u⊤ m Xi end for end for for m = 1, . . . , M do (s′m , e′m ) ← [s, e] ∩ [αm , βm ] and (sm , em ) ← (⌈s′m + δ⌉, ⌊e′m − δ⌋) if em − sm ≥ 2 log(n) + 1 then bm ← arg maxsm +log(n)≤t≤em −log(n) |Yetsm ,em (um )| am ← Yebsmm ,em (um ) else am ← −1 end if end for m∗ ← arg maxm=1,...,M am if am∗ > τ then add bm∗ to the set of estimated change points WBSRP((s, bm∗ ), {(αm , βm )}M m=1 , τ ) WBSRP ((bm∗ + 1, e), {(αm , βm )}M m=1 , τ ) end if OUTPUT: The set of estimated change points. 13 WBSIP((s, e), Similarly to the BSOP algoithm, WBSIP also applies a slight modification to the WBS algorithm as originally proposed in Fryzlewicz (2014). When computing the shadow vectors in Algorithm 2, the search for the optimal direction onto which to project the data is restricted, for any given candidate interval, only to the time points that are at least p log n away from the endpoints of the interval. As remarked in the previous section, this ensures good tail bounds for the operator norms of the matrices involved. A second, more substantial, adaptation of WBS is used in Algorithm 3: when searching for candidate change points inside a given interval, the algorithm only consider time points that are δ-away from the endpoints of the interval, where δ is an upper bound on the localization error – the term ǫn in Theorem 2 below. The reasons for such restriction are somewhat subtle: once an estimated change point is found near a true change point, in its next iteration the algorithm can no longer look for change points in that proximity, since this will result, with high probability, in spurious detections. This phenomenon is due to the fact that the behavior of the CUSUM statistic of the projected data is not uniform around its local maxima. Therefore, after a true change point has been detected, the algorithm must scan only nearby regions of low signal-to-noise ratio – so that the probability of false positives can be adequately controlled with a proper choice of the thresholding parameter τ . Thus the need to stay away by at leat the localization error form the detected chaing points. A very similar condition is imposed in the main algorithm of Wang and Samworth (2016) and implicitly in Korkas and Fryzlewicz (2017). The value of δ is left as an input paramerer (as in Wang and Samworth, 2016), but any value between the localization error ǫn given in the statement of Theorem 2 and the minimal distance ∆ between change points will do. In the proof of we Theorem 2 we set δ ≤ 3∆/32. M Theorem 2 (Consistency of WBSIP). Let Assumptions 1 and 3 hold let {(αm , βm )}M m=1 ⊂ (0, T ) be a collection of intervals whose end points are drawn independently and uniformly from {1, . . . , T } and such that max1≤m≤M (βm − αm ) ≤ C∆ for an absolute constant C > 0. Set ǫn = C1 B 4 log(n)κ−2 , for a C1 > 0. Suppose there exist c2 , c3 > 0, sufficiently small, such that the input parameters τ and δ satisfy p √ B 2 log(n) < τ < c2 κ ∆, (8) ǫn < δ ≤ c3 ∆. b Then the collection of the estimated change points B = {η̂k }K k=1 returned by WBSIP with input , τ and δ satisfies parameters of (0, n), {(αm , βm )}M m=1 n o b = K; P K max |ηk − η̂k | ≤ ǫn k=1,...,K  2 −c ≥1 − 2n M n − n3 9p 2n−cp − exp log(n/∆) − M ∆2 /(16n2 ) for some absolute constants c > 0. Remark 8 (The relationships among the constants in Theorem 2). The choice of the constant C is essentially arbitrary but will affect the choice of the constants C1 , c2 and c3 , where c2 and c3 in particular have to be picked small enough. This dependence can be tracked in the proof but we refrain from giving further details. 14 The above theorem implies that the WBSIP algorithm can estimate the change points perfectly well, with high probability, with a localization rate upper bounded by B 4 log n ǫn  2 . n κ n The fact that the dimension p does not appear explicitly in the localization rate is an interesting, if not perhaps surprising, finding. Of course, the dimension does affect (negatively) the performance of the algorithm through Assumption 3: keeping n and ∆ fixed, a larger value of p implies a larger 4 value of Bκ2 in order for that assumption to hold. In turn, this leads to a larger bound in Theorem 2. Furthermore, the dimension p appears in the probability of the event that WBSIP fails to localize all the change points. We remark that, for the different problem of high-dimensional mean change point detection, Wang and Samworth (2016) also obtained a localization rate independent of the dimension: see Theorem 3 there. In Section 3 below we will prove that Assumption 3 is in fact essentially necessary for any algorithm to produce a vanishing localization rate. Proof. Since ǫn is the desired order of localization rate, by induction, it suffices to consider any generic (s, e) ⊂ (0, T ) that satisfies ηr−1 ≤ s ≤ ηr ≤ . . . ≤ ηr+q ≤ e ≤ ηr+q+1 , q ≥ −1, max{min{ηr − s, s − ηr−1 }, min{ηr+q+1 − e, e − ηr+q }} ≤ ǫn , where q = −1 indicates that there is no change point contained in (s, e). Note that under Assumption 3, ǫn ≤ ∆/4; it, therefore, has to be the case that for any change point ηp ∈ (0, T ), either |ηp − s| ≤ ǫn or |ηp − s| ≥ ∆ − ǫn ≥ 3∆/4. This means that min{|ηp − e|, |ηp − s|} ≤ ǫn indicates that ηp is a detected change point in the previous induction step, even if ηp ∈ (s, e). We refer to ηp ∈ [s, e] as an undetected change point if min{ηp − s, ηp − e} ≥ 3∆/4. In order to complete the induction step, it suffices to show that WBSIP((s, e), {(αm , βm )}M m=1 , τ, δ) (i) will not detect any new change point in (s, e) if all the change points in that interval have been previous detected, and (ii) will find a point b in (s, e) (in fact, in (s + δ, e− δ)) such that |ηp − b| ≤ ǫn if there exists at least one undetected change point in (s, e). Let n M {um }M m=1 = P C({Xi }i=1 , {(αm , βm )}m=1 ). n n Since the intervals {(αm , βm )}M m=1 are generated independently from {Xi }i=1 ∪ {Wi }i=1 , the rest of the argument is made under M defined in Equation (20) of Appendix A, which has no effects on the distribution of {Xi }ni=1 ∪ {Wi }ni=1 . p Step 1. Let λ1 = B 2 p log(n). In this step, we are to show that, on the event A1 ({Wi }ni=1 , λ1 ) and for some c′1 > 0, ′ ′ sup |u⊤ m (Σηk − Σηk−1 )um | ≥ c1 kΣηk − Σηk−1 kop = c1 κk for every k = 1, . . . , K + 1 (9) 1≤m≤M On the event M, for any ηk ∈ (0, n), without loss of generality, there exists αk ∈ [ηk −3∆/4, ηk −∆/2] and βk ∈ [ηk + ∆/2, ηk + 3∆/4]. Thus [αk , βk ] contains only one change point ηk . Using Lemma 20 in Appendix D and the inequality p log(n) ≤ ∆/8, we have that √ e ηαk ,βk kop ≥ (1/2)kΣη − Ση kop ∆. e αk ,βk kop = kΣ max kΣ (10) t k k−1 k t=⌈αk +δ⌉,...,⌊βk −δ⌋ 15 Let bk ∈ arg maxt=⌈αk +δ⌉,...,⌊βk −δ⌋ kSetαk ,βk kop , where Sets,e denote the covariance CUSUM statistics of {Wi }ei=s+1 at evaluated t. Since kΣηk − Σηk−1 kop = κk , by definition, e αk ,βk uk | ≥ |u⊤ Seαk ,βk uk | − λ1 |u⊤ k Σbk k bk = ≥ max t=⌈αk +δ⌉,...,⌊βk −δ⌋ max e αk ,βk kop − 2λ1 kΣ t √ − Σηk−1 kop ∆ − 2λ1 √ − Σηk−1 kop ∆ t=⌈αk +δ⌉,...,⌊βk −δ⌋ ≥ (1/2)kΣηk ≥ (1/4)kΣηk kSetαk ,βk kop − λ1 where the first and second inequalities hold on the event A1 ({Wi }ni=1 , λ1 ), the third inequality follows from (10) and the last inequality from Assumption 3. Next, observe that q t−αk  t ≤ ηk , (βk −αk )(βk −t) (βk − ηk )(Σηk − Σηk−1 ), αk ,βk e q = Σt βk −t  (βk −αk )(t−αk ) (ηk − αk )(Σηk − Σηk−1 ), t ≥ ηk . Using the above expression, for bk ≥ ηk , we have that √ e αk ,βk uk (1/4)kΣηk − Σηk−1 kop ∆ ≤ u⊤ k Σbk s βk − bk (ηk − αk ) u⊤ = k (Σηk − Σηk−1 )uk (βk − αk )(bk − αk ) s (βk − ηk )(ηk − αk ) ⊤ uk (Σηk − Σηk−1 )uk ≤ βk − αk √ ≤ 2∆ u⊤ k (Σηk − Σηk−1 )uk . √ Therefore (9) holds with c′1 = 1/(2 2). The case of bk < ηk follows from very similar calculations. Step 2. In this step, we will show that WBSIP((s, e), {(αm , βm )}M m=1 , τ, δ) will consistently detect or reject the existence of undetected change points within (s, e), provided that (9) holds and on p 2 log(n), and M, given in Equation (18) , λ ), where λ = B the two events B1 ({Xi }ni=1 , {um }M 2 m=1 2 and Equation (20) in Appendix A, respectively. Let am , bm and m∗ be defined as in WBSIP((s, e), {(αm , βm )}M m=1 , τ, δ). Denote Yi (um ) = s,e s,e ⊤ 2 ⊤ e e (um Xi ) and fi (um ) = um Σi um . Let Yt (um ) and ft (um ) be defined as in (53) and (54) of Appendix D.2 respectively. Suppose there exists a change point ηp ∈ (s, e) such that min{ηp − s, e − ηp } ≥ 3∆/4. Let δ ≤ 3∆/32. Then, on the event M, there exists an interval (αm , βm ) selected by WBSIP such that αm ∈ [ηp − 3∆/4, ηp − ∆/2] and βm ∈ [ηp + ∆/2, ηp + 3∆/4]. Then [s′m , e′m ] = [αm , βm ] ∩ [s, e], and [sm , em ] = [s′m + δ, e′m − δ] (see details of the WBSIP procedure in Algorithm 3). Moreover, we have that min{ηp −sm , em −ηp } ≥ (1/2)∆. Thus, [sm , em ] contains at most one change point of the time series {fi (um )}ni=1 . A similar calculation as the one shown in the proof of Lemma 20 gives that √ max |fetsm ,em (um )| ≥ (1/8) ∆|u⊤ m (Σηp − Σηp−1 )um |, ⌈sm +log(n)⌉≤t≤⌊em −log(n)⌋ 16 where em − sm ≤ (3/2)∆ is used in the last inequality. Therefore am = ≥ max ⌈sm +log(n)⌉≤t≤⌊em −log(n)⌋ max ⌈sm +log(n)⌉≤t≤⌊em −log(n)⌋ |Yetsm ,em (um )| |fetsm ,em (um )| − λ2 √ ≥ (1/8) ∆|u′m (Σηp − Σηp−1 )um | − λ2 , where the first inequality holds on the event B1 ({Xi }ni=1 , {um }M m=1 , λ2 ). Thus for any undetected change point ηp within (s, e), it holds that am∗ = ≥ sup am (11) √ √ sup (1/8) ∆|u′m (Σp − Σp−1 )um | − λ2 ≥ (c′1 /8)κp ∆ − λ2 (12) 1≤m≤M 1≤m≤M √ ≥ (c′1 /16)κp ∆ (13) where the second inequality follows from (9), and the last inequality from p √ λ2 = B 2 log(n) ≤ (c′1 /16)κ ∆, √ by choosing the constant C in Assumption 3 to be at least 32. Then, WBSIP((s, e), {(αm , βm )}M m=1 , τ, δ) correctly accepts the existence of undetected change points on the events (9), B1 ({Xi }ni=1 , {um }M m=1 , λ2 , δ) and M. Suppose there does not exists any undetected change points within (s, e), then for any (s′m , e′m ) = (αm , βm ) ∩ (s, e), one of the following situations must hold. (a) There is no change point within (s′m , e′m ); (b) there exists only one change point ηr within (sm , em ) and min{ηr − s′m , e′m − ηr } ≤ ǫn ; or (c) there exist two change points ηr , ηr+1 within (sm , em ) and max{ηr − s′m , e′m − ηr+1 } ≤ ǫn . Observe that if (a) holds, then, on the event B1 ({Xi }ni=1 , {um }M m=1 , λ2 ) given in Equation (18) ′ ′ in Appendix A, for (sm , em ) = (sm + δ, em − δ), we have max ⌈sm +log(n)⌉≤t≤⌊em −log(n)⌋ |Yetsm ,em (um )| ≤ max ⌈sm +log(n)⌉≤t≤⌊em −log(n)⌋ |fetsm ,em (um )| + λ2 = 0 + λ2 . If (b) or (c) holds, then since (sm , em ) = (s′m + δ, e′m − δ) and δ ≤ ǫn , it must be the case that (sm , em ) does not contain any change points. This reduces to case (a). Therefore if (8) holds, then WBSIP((s, e), {(αm , βm )}M m=1 , τ, δ) will always correctly reject the existence of undetected change points, on the event B1 ({Xi }ni=1 , {um }M m=1 , λ2 ). Step 3. Assume that there exists a change point ηp ∈ (s, e) such that min{ηp − s, ηp − e} ≥ 3∆/4. Let am , bm and m∗ be defined as in WBSIP((s, e), {(αm , βm )}M m=1 , τ ). To complete the proof it suffices to show that, on the events B1 ({Xi }ni=1 , {um }M m=1 , λ2 ) and n M B2 ({Xi }i=1 , {um }m=1 , λ2 ) given in Equation (18) and Equation (19) respectively of Appendix A, there exists a change point ηk ∈ [sm∗ , em∗ ] such that min{ηk −s, ηk −e} ≥ 3∆/4 and |bm∗ −ηk | ≤ ǫn . Consider the univariate time series {Yi (um∗ )}ni=1 and {fi (um∗ )}ni=1 defined in (53) and (54) of m∗ is a Appendix D.2. Since the collection of the change points of the time series {fi (um∗ )}ei=s m∗ 17 em∗ subset of that of {ηk }K+1 k=0 ∩ [s, e], we may apply Corollary 13 to the time series {Yi (um∗ )}i=sm∗ em∗ and {fi (um∗ )}i=sm∗ . To that end, we will need to ensure that the assumptions of Corollary 13 are verified. Let δ′ = log(n) and λ = λ2 . Observe that (48) and (49) are straightforward consequences of Assumption 3, (46) and (47) follow from the definiiton of B1 ({Xi }ni=1 , {um }M m=1 , λ2 ) and n M B2 ({Xi }i=1 , {um }m=1 , λ2 ), and that (45) follows from (13). Thus, all the conditions in Corollary 13 are met, and we therefore conclude that there exists a m∗ change point ηk , which is also a change point of {fi (v)}ei=s , satisfying m∗ min{em∗ − ηk , ηk − sm∗ } > ∆/4 (14) and |bm∗ − ηk | ≤ max{C3 λ22 κ−2 , δ′ } ≤ ǫn , where the last inequality holds because λ22 κ−2 = B 4 log(n)κ−2 ≥ log(n), which is a consequence of the inequality B 2 ≥ κ. The proof is complete with the following two observations: i) The change points of {fi (um∗ )}ei=s belong to (s, e) ∩ {ηk }K k=1 ; and ii) Equation (14) and (sm∗ , em∗ ) ⊂ (s, e) imply that min{e − ηk , ηk − s} > ∆/4 > ǫn . As discussed in the argument before Step 1, this implies that ηk must be an undetected change point of {Xi }ni=1 in the covariance structure. 3 Lower bounds In this section, we provide lower bounds for the problem of chenge point estimation with high dimensional covariance matrices. In Theorem 2 we showed that if the distribution of {Xi }ni=1 satisfies Assumption 1 and additionally the condition that ∆ ≥ CB 4 κ−2 p log(n) for sufficiently large C as given in Assumption 3, then the WBSIP algorithm can, with high probability, detect all the change points with a localization rate of the order B 4 log n . (15) κ2 n Assumption 3 might seem a bit arbitrary at first glance. However, we will show in the next result that if the class of distribution of interest allows for a spacing parameter ∆ = cB 4 κ−2 p for some sufficiently small constant c, then it is not possible to estimate the location of the change point at a vanoishing rate. This result furher implies that the WBSIP algorithm is optimal in the sense of requiring the minimax scaling for the problem parameters. To that effect, we will consider the following class of data generating distribution. Assumption 4. Let X1 , . . . , Xn ∈ Rp be independent Gaussian random vectors such that Xi ∼ Np (0, Σi ) with kΣi kop ≤ 2σ. Let {ηk }K+1 k=0 ⊂ {0, . . . , n} be a collection of change points, such that η0 = 0 and ηK+1 = n and that Σηk +1 = Σηk +2 = . . . = Σηk+1 , for any k = 1, . . . , K + 1. Assume there exists parameters κ = κ(n) and ∆ = ∆(n) such that inf {ηk − ηk−1 } ≥ ∆ > 0, k=1,...,K+1 kΣηk − Σηk−1 kop = κk ≥ κ > 0, for any k = 1, . . . , K + 1. 18 n We will denote with Pκ,∆,σ the joint distribution of the data when the above assumption is in n effect. Notice that Pκ,∆,σ satisfies Assumption 1 with B = 8σ. Below we prove that consistent 2 4 estimation of the locations of the change point requires ∆κ p B to diverge, as in Assumption 3. We recall that all the parameters of interest, ∆, κ, σ and p are allowed to change with n. The proof is based on a construction used in Cai and Ma (2013) to obtain minimax lower bounds for a class of hypothesis testing problems involving covariance matices. Lemma 3. Let {Xi }ni=1 be a time series satisfies Assumption 4 with only one change point and let n Pκ,∆,σ denote the corresponding joint distribution. Consider the class of distributions P = P(n) =  n Pκ,∆,σ  σ4 p 2 : ≤ ∆ ≤ n/3, κ ≤ σ /4 , 33κ2 Then inf sup EP (|η̂ − η|) ≥ n/6. η̂ P ∈P Remark 9 (Phase transiton for the localization rate in covariance change point detection problem). In light of Lemma 3 and Theorem 2, we conclude that, in the covariance change point detection problem, the solution undergoes a phase transition, which are able to characterize up to logarithmic factor (in n). Specifically, • if ∆ ≥ CB 4 p log(n)/κ2 for a sufficiently large constant C > 0, then it is possible to estimate the locations of the change points with a localization rate vanishing in n; • on the other hand, if ∆ = CB 4 p/κ2 for a sufficiently small constant c > 0, then the localization rate of any algorithm remains, at least in the worst case, bounded away from 0. To the best of our knowledge, this phase transition effect is new and unique to our settings. We conclude this section by showing that the localization rate that we have obtained for the WBSIP algorithm, given above in (15), is, up to a logarithmic factor, minimax optimal. n o n : ∆ < n/2, κ ≤ σ 2 /4 . Then, Lemma 4. Consider the class of distribution Q = Pκ,∆,σ inf sup EP (|η̂ − η|) ≥ cσ 4 κ−2 . η̂ P ∈Q 4 Discussion In this paper, we tackle the problem of change point detection for a time series of length n of independent p-dimensional random vectors with covariance matrices that are pirece-wise constant. We allow the dimension, as well as other parameters quantifying the difficulty of the problem, to change with n. We have devised two procedures based on existing algorithms for change point detection – binary segmentation and wild binary segmentation – and show that the localization rates they yield are consistent with those in the univariate time series mean change point detection problems. In particular we demonstrate the algorithm WBSIP, which applies wild binary segmentation to carefully chosen univariate projections of the data, produces a localization rate that is, up to a logarithmic factor, minimax optimal. 19 The model setting adopted in the paper allows for the dimension p to grow with n. However, in order for the localization rates of any procedure to vanish with n it must be the case that p is of smaller order than n. One possible future direction is to consider different high dimensional settings whereby p is permitted to grow even faster than n, with additional structural assumptions on the underlying covariance matrices. For instance, we may model the covariance matrices as spiked matrices with sparse leading eigenvectors. A common, undesirable feature of the WBSIP algorithms is the fact that, for given interval (s, e), the search of the next change point is limited to points inside the interval that are δ away from the endpoints, where δ is an inout parameter that is larger than the localization error. Such restriction, which appears also in the algorithm of Wang and Samworth (2016) for mean change point localization of high-dimensional time series, is made in order to prevent the algorithms from returning spurious change points in the proximity of a true change point. The reason for such phenomenon is subtle, and is ultimately due to the fact that the rate of decay of the expected value of the covariance CUSUM statistics around the true change points is in general not uniform, as it depends on the magnitude of the change. A possible solution to such an issue – which appears to be unavoidable – would be to design adaptive algorithm yielding local rates, one for each change point. We will pursue this line of research in future work. One key assumption used in this paper is the the time series of comprised of independent observations. This is of course a rather strong condition, which might not apply to many real-life problems. In order to handle time dependence, a natural approach is introduce mixing conditions (see, e.g. Wu, 2005) or assume the observations come from certain well-defined time series models. Further extensions that will be worth pursuing include the cases in which model is mis-specified or the observations are contaminated We leave these interesting extensions to future work but expect that many of the results derived in this manuscript will provide the theoretical underpinning for devising and studying more complicated algorithms. Acknowledgments We thank Dr Haeran Cho for her constructive comments. References Anderson, T. W. (2003). An Introduction to Multivariate Statistical Analysis. 3rd ed. Wiley Series in Probability and Statistics, John Wiley & Sons. Aston, J. A. D. and Kirch, C. (2014). Efficiency of change point tests in high dimensional settings. arXiv preprint arXiv: 1409.1771. Aue, A., Hömann, S., Horváth, L. and Reimherr, M. (2009). Break detection in the covariance structure of multivariate nonlinear time series models. The Annals of Statistics, 37 4046–4087. Avanesov, V. and Buzun, N. (2016). Change-point detection in high-dimensional covariance structure. arXiv preprint arXiv:1610.03783. Baranowski, R., Chen, Y. and Fryzlewicz, P. (2016). Narrowest-Over-Threshold detection of multiple change-points and change-point-like feature. arXiv preprint arXiv: 1609.00293. 20 Barigozzi, M., Cho, H. and Fryzlewicz, P. (2016). Simultaneous multiple change-point and factor analysis for high-dimensional time series. arXiv preprint arXiv: 1612.06928. Berkes, I., Gabrys, R., Horváth, L. and Kokoszka, P. (2009). Detecting changes in the mean of functional observations. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 71 927–946. Berthet, Q. and Rigollet, P. (2013). Optimal detection of sparse principal components in high dimension. The Annals of Statistics, 41 1780–1815. Birke, M. and Dette, H. (2005). A note on testing the covariance matrix for large dimension. Statistics and Probability Letters, 74 281–289. Cai, T. T. and Ma, Z. (2013). Optimal hypothesis testing for high dimensional covariance matrices. Bernoulli, 19 2359–2388. Chan, H. P. and Walther, G. (2013). Detection with the scan and the average likelihood ratio. Statistica Sinica, 1 409–428. Cho, H. (2015). Change-point detection in panel data via double cusum statistic. Electronic Journal of Statistics in press. Cho, H. and Fryzlewicz, P. (2012). Multiscale and multilevel technique for consistent segmentation of nonstationary time series. Statistica Sinica, 22 207–229. Cho, H. and Fryzlewicz, P. (2015). Multiple change-point detection for high-dimensional time series via Sparsified Binary Segmentation. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 77 475–507. Davies, P. L. and Kovac, A. (2001). Local extremes, runs, strings and multiresolution. Ann. Statist., 29 1–65. Davis, R. A., Lee, T. C. M. and Rodriguez-Yam, G. A. (2006). Structural break estimation for nonstationary time series models. Journal of the American Statistical Association, 101 223–239. Frick, K., Munk, A. and Sieling, H. (2014). Multiscale change point inference. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76 495–580. Fryzlewicz, P. (2014). Wild binary segmentation for multiple change-point detection. The Annals of Statistics, 42 2243–2281. Gombay, E., Horváth, L. and Hušková, M. (1996). Estimators and tests for change in variances. Statistics and Risk Modeling, 14 145–160. Harchaoui, Z. and Lévy-Leduc, C. (2010). Multiple change-point estimation with a total variation penalty. Journal of American Statistical Association, 105 1480–1493. Harchaoui, Z. and Lvy-Leduc, C. (2010). Multiple change-point estimation with a total variation penalty. Journal of the American Statistical Association, 105 1480–1493. Horváth, L. and Hušková, M. (2012). Change-point detection in panel data. Journal of Time Series Analysis, 33 631–648. 21 Inclan, C. and Tiao, G. C. (1994). Use of cumulative sums of squares for retrospective detection of changes of variance. Journal of the American Statistical Association, 89 913–923. James, B., James, K. L. and Siegmund, D. (1987). Tests for a change-point. Biometrika, 74 71–83. Jirak, M. (2015). Uniform change point tests in high dimension. The Annals of Statistics, 43 2451–2483. Johnstone, I. M. (2001). On the distribution of the largest eigenvalue in principal components analysis. The Annals of Statistics, 29 295–327. Killick, R., Fearnhead, P. and Eckley, I. A. (2012). Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107 1590–1598. Korkas, K. K. and Fryzlewicz, P. (2017). Multiple change-point detection for non-stationary time series using wild binary segmentation. Statistica Sinica, 27 287–311. Lavielle, M. (1999). Detection of multiple changes in a sequence of dependent variables. Stochastic Processes and their Applications, 83 79–102. Lavielle, M. and Moulines, E. (2000). Least-squares estimation of an unknown number of shifts in a time series. Journal of Time Series Analysis, 21 33–59. Ledoit, O. and Wolf, M. (2002). Some hypothesis tests for the covariance in principal components analysis. The Annals of Statistics, 30 1081–1102. Li, H., Guo, Q. and Munk, A. (2017). Multiscale change-point segmentation: Beyond step functions. arXiv preprint arXiv: 1708.03942. Lin, K., Sharpnack, J. L., Rinaldo, A. and Tibshirani, R. J. (2017). A sharp error analysis for the fused lasso, with application to approximate changepoint screening. In Advances in Neural Information Processing Systems 30 (I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan and R. Garnett, eds.). 6887–6896. Olshen, A. B., Venkatraman, E., Lucito, R. and Wigler, M. (2004). Circular binary segmentation for the analysis of array-based dna copy number data. Biostatistics, 5 557–572. Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41 100–115. Picard, D. (1985). Testing and estimating change-points in time series. Advances in Applied Probability, 17 841–867. Qian, J. and Jia, J. (2012). Https://arxiv.org/abs/1211.5194. On pattern recovery of the fused lasso. Tech. rep. Rojas, C. R. and Wahlberg, B. (2014). On change point detection using the fused lasso method. Tech. rep. Https://arxiv.org/abs/1401.5408. Roy, S. N. (1957). Some Aspects of Multivariate Analysis. New York: Wiley. 22 Siegmund, D. and Venkatraman, E. S. (1995). Using the generalized likelihood ratio statistic for sequential detection of a change-point. The Annals of Statistics 255–271. Tibshirani, R., Saunders, M., Rosset, S., Zhu, J. and Knight, K. (2005). Sparsity and smoothness via the fused lasso. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67 91–108. Venkatraman, E. S. (1992). Consistency results in multiple change-point problems. Ph.D. thesis, Stanford University. Vershynin, R. (2010). Introduction to the non-asymptotic analysis of random matrices. arXiv preprint arXiv:1011.3027. Vostrikova, L. (1981). Detection of the disorder in multidimensional random-processes. Doklady Akademii Nauk SSSR, 259 270–274. Wald, A. (1945). Sequential tests of statistical hypotheses. The Annals of Mathematical Statistics, 16 117–186. Wang, T. and Samworth, R. J. (2016). High-dimensional changepoint estimation via sparse projection. arXiv preprint arXiv:1606.06246. Wang, Y. (1995). Jump and sharp cusp detection by wavelets. Biometrika, 82 385–397. Wu, W. B. (2005). Nonlinear system theory: Another look at dependence. Proceedings of the National Academy of Sciences of the United States of America, 102 14150–14154. Yao, Y.-C. and Au, S.-T. (1989). Least-squares estimation of a stop function. Sankhyā: The Indian Journal of Statistics, Series A 370–381. Yu, B. (1997). Festschrift for Lucien Le Cam, vol. 423, chap. Assouad, Fano, and Le Cam. Springer Science & Business Media, 435. Appendices A Probabilistic bounds In this section, we give basic high-probability concentration bounds on the fluctuations of te covariance CUSUM statistics using the notions of sub-Gaussian and sub-Exponential random vectors. We also state some properties of the randomly selected intervals {sm , em } in the WBS algorithm, which hold with high probability. We start by introducing the definitions of sub-Gaussian and sub-Exponential random variables through Orlicz norms. See, e.g., Vershynin (2010) for more details. Definition 2. (i) A random variable X ∈ R is sub-Gaussian if kXkψ2 := sup k−1/2 {E(|X|k )}1/k < ∞. k≥1 23 A random vector X ∈ Rp is sub-Gaussian if kXkψ2 := sup kv ⊤ Xkψ2 < ∞, v∈S p−1 where S p−1 denote unit sphere in Eucliden norm in Rp . (ii) A random variable Y ∈ R is sub-Exponential if kY kψ1 = sup k−1 E(|Y |k )1/k < ∞. k≥1 A random vector Y ∈ Rp is sub-Exponential if kY kψ1 = sup kv ⊤ Y kψ1 < ∞. v∈S p−1 We note that if X ∈ R is sub-Gaussian, then X 2 is sub-Exponential, due to the easily verifiable fact that kXk2ψ2 ≤ kX 2 kψ1 ≤ 2kXk2ψ2 . (16) Recall the sample and the population versions of the covariance CUSUM statistics given in Definition 1. For λ > 0, we define the events, which depend on {Xi }ni=1 .   s,e s,e n e e ≤ λ, min{t − s, e − t} ≥ p log(n) (17) A1 ({Xi }i=1 , λ) = sup St − Σ t op 0≤s<t<e≤n and A2 ({Xi }ni=1 , λ) = ( Pe ⊤ i=s+1 (Xi Xi sup 0≤s<e≤n √ e−s − Σi ) op ≤ λ, ) e − s ≥ p log(n) . p Next, for an arbitary collection {vm }M m=1 of deterministic unit vectors in R we define the events ) ( s,e n M ⊤ es,e e )vm ≤ λ, min{t − s, e − t} ≥ log(n) B1 ({Xi }i=1 , {vm }m=1 , λ) = sup sup vm (St − Σ t 1≤m≤M 0≤s<t<e≤n (18) and B2 ({Xi }ni=1 , {vm }M m=1 , λ) = ( sup Pe ⊤ ⊤ i=s+1 vm (Xi Xi sup 1≤m≤M 0≤s<e≤n √ e−s − Σi )vm ≤ λ, e − s ≥ log(n) . Lemma 5. Suppose {Xi }ni=1 ⊂ Rp are i.i.d sub-Gaussian centered random vectors such that sup kXi kψ2 ≤ B. 1≤i≤n 24 ) (19) There exists an absolute constant c > 0 such that, p P(A1 ({Xi }ni=1 , B 2 p log(n))) ≥ 1 − 2 × 9p n3 n−cp , p P(A2 ({Xi }ni=1 , B 2 p log(n))) ≥ 1 − 2 × 9p n2 n−cp , p 2 log(n))) ≥ 1 − 2n3 M n−c , P(B1 ({Xi }ni=1 , {vm }M , B m=1 p 2 P(B2 ({Xi }ni=1 , {vm }M log(n))) ≥ 1 − 2n2 M n−c , m=1 , B for any set {vm }M m=1 of deterministic unit vectors. Proof. We first tackle A1 . For any v ∈ Rp such that kvk2 = 1, we can write e s,e)v = v (Sets,e − Σ t ⊤ where e X i=s+1 e   X ⊤ 2 ⊤ 2 ai Zi , ai (v Xi ) − E((v Xi ) ) = i=s+1 q e−t  (e−s)(t−s) q ai = t−s  (e−s)(e−t) s + 1 ≤ i ≤ t, t + 1 ≤ i ≤ e, and Zi = (v ⊤ Xi )2 − E[(v ⊤ Xi )2 ]. Since min{t − s, e − t} ≥ p log(n), we further have that e X a2i = 1, and i=s+1 p max |ai | ≤ 1/ p log(n). s+1≤i≤e Thus by Proposition 5.16 in Vershynin (2010), for any ǫ ≥ 0, ! ( )! p e X ǫ2 ǫ p log(n) ai Zi ≥ ǫ ≤ 2 exp −c min P , , K2 K i=s+1 where c > 0 is an absolute constant and K = max kZi kψ1 ≤ 2k(v ⊤ Xi )2 kψ1 ≤ 4kXi k2ψ2 ≤ 4B 2 . i Therefore,   p  e s,e v ≥ B 2 p log(n) ≤ 2n−cp . P v ⊤ Sets,e − Σ t Let N1/4 be a minimal 1/4-net (with respect to the Eucidean norm) of the unit sphere in Rp . Then, card(N1/4 ) ≤ 9p and, by a standard covering argument followed by a union bound, we arrive at the inequality p P(A1 ({Xi }ni=1 , B 2 p log(n))) ≥ 1 − 2 × 9p n3 n−cp , for a universal constant c > 0. Following the same arguments we have, p P(A2 ({Xi }ni=1 , B 2 p log(n))) ≥ 1 − 2 × 9p n2 n−cp , p 2 P(B1 ({Xi }ni=1 , {vm }M , B log(n))) ≥ 1 − 2n3 M n−c , m=1 and for some c > 0. 2 P(B2 ({Xi }ni=1 , {vm }M m=1 , B p log(n))) ≥ 1 − 2n2 M n−c , 25 M Let {sm }M m=1 , {em }m=1 be two sequences independently selected at random in [s, e], and M= K \  sm ∈ Sk , em ∈ Ek , for some m ∈ {1, . . . , M } , (20) k=1 where Sk = [ηk − 3∆/4, ηk − ∆/2] and Ek = [ηk + ∆/2, ηk + 3∆/4], k = 1, . . . , K. In the lemma below, we give a lower bound on the probability of M. Under the scaling assumed in our setting, this bound approaches 1 as n grows. Lemma 6. For the event M defined in (20), we have   ∆2 n . P(M) ≥ 1 − exp log − M ∆ 16n2 Proof. Since the number of change points are bounded by n/∆, M K Y X   P Mc ≤ 1 − P sm ∈ Sk , em ∈ Ek ≤ K(1 − ∆2 /(16n2 ))M ≤ n/∆(1 − ∆2 /(16n2 ))M . k=1 m=1 B Properties of the univariate CUSUM statistics In this section, we derive some important and useful properties of the univariate CUSUM statistic. Our results and proofs build upon the existing literature on univariate mean change point detection; see in partiucular, Venkatraman (1992) and Fryzlewicz (2014), whose notation will be used throughout. It is imprtnat however to note that we have made several non-trivial modifications of those arguments, and have made a special effort in keeping track of the changes in all the key parameters. This careful treatment eventually allows us to achieve tight upper bounds for the the localization rate implied by the WBSIP algorithm and which in turn have revealed a phase transition in the problem parameters (see Section 3). In particular, the results of this section can be used to sharpen existing analyses of the BS and WBS algorithms. B.1 Results from Venkatraman (1992) We start by introducing some notations for one dimensional change point detection and the corresponding CUSUM statistics. Let {Yi }ni=1 , {fi }ni=1 ⊂ R be two univariate, sequences. We will make the following assumptions on. Assumption 5 (Univariate mean change points). Let {ηk }K+1 k=0 ⊂ {0, . . . , n}, where η0 = 0 and ηK+1 = n, and fηk−1 +1 = fηk−1 +2 = . . . = fηk for all 1 ≤ k ≤ K + 1, Assume inf {ηk − ηk−1 } ≥ ∆ = ∆(n) > 0, k=1,...,K+1 |fηk+1 − fηk | := κk > 0, k = 1, . . . , K, sup k=1,...,K+1 |fηk | < B1 . 26 as We also have the corresponding CUSUM statistics over any generic interval [s, e] ⊂ [1, T ] defined Yets,e = fets,e = s s t X e−t Yi − (e − s)(t − s) i=s+1 e−t (e − s)(t − s) t X i=s+1 fi − s s e X t−s Yi , (e − s)(e − t) i=t+1 e X t−s fi . (e − s)(e − t) i=t+1 Throughout this Appendix B, all of our results are proven by regarding {Yi }Ti=1 and {fi }Ti=1 as two deterministic sequences. We will frequently assume that fets,e is a good approximation of Yets,e in ways that we will specify through appropriate assumptions. Observe that the function fets,e is only well defined on [s, e] ∩ Z. Our first result, which is taken from Venkatraman (1992), shows that the there exists a continuous realization of the discrete function fets,e Lemma 7. Suppose [s, e] ⊂ [1, T ] satisfies ηr−1 ≤ s ≤ ηr ≤ . . . ≤ ηr+q ≤ e ≤ ηr+q+1 , q ≥ 0. Then there exists a continuous function Fets,e : [s, e] → R such that Fers,e = fers,e for every r ∈ [s, e]∩ Z with the following additional properties. 1. |Fets,e | is maximized at the change points within [s, e]. In other words, arg max |Fets,e | ∩ {ηr , . . . , ηr+q } = 6 ∅. s≤t≤e 2. If Fets,e > 0 for some t ∈ (s, e), then Fets,e is either monotonic or decreases and then increases within each of the interval [s, ηr ], . . . , [ηr+q , e]. The proof of this lemma can be found in Lemma 2.2 and 2.3 of Venkatraman (1992). We remark that if Fets,e ≤ 0 for all t ∈ (s, e), then it suffices to consider the time series {−fi }Ti=1 and a similar result as in the second part of Lemma 7 still holds. Throughout the entire section, we always view fets,e as a continuous function and frequently invoke Lemma 7 as a basic property of the CUSUM statistics without further notice. Our next lemma is an adaptation of a result first obtained by Venkatraman (1992), which quantifies characterizes how fast the CUSUM statistics decays around a good change point. An analogous result, derive using different arguments, can be found in Proposition 21 in Wang and Samworth (2016). Lemma 8 (Venkatraman (1992) Lemma 2.6). Let [s, e] ⊂ [1, T ] be any generic interval. For some c1 , c2 > 0 and λ > 0 such that min{ηk − s, e − ηk } ≥ c1 ∆ fes,e ≥ c2 κ∆(e − s)−1/2 , ηk (21) (22) suppose there exists sufficient small c3 such that max |fets,e | − feηs,e ≤ 2λ ≤ c3 κ∆3 (e − s)−5/2 k s≤t≤e 27 (23) Then there exists an absolute constant c > 0 such that if the point d ∈ [s, e] is such that |d − ηk | ≤ c1 ∆/16, then feηs,e − feds,e > cfeηs,e |ηk − d|∆(e − s)−2 k k Remark 10. If feηs,e k < 0, and d ∈ [s, e] is such that |d − ηk | ≤ c1 /16, then by considering the sequence {−fi }ni=1 , it holds that (−feηs,e ) − (−feds,e ) > c(−feηs,e )|ηk − d|∆(e − s)−2 k k Proof. Without loss of generality, assume that d ≥ ηk . Following the argument of Venkatraman (1992) Lemma 2.6, it suffices to consider two cases: (1) ηk+1 > e, and (2) ηk+1 ≤ e. Case 1. Let El be defined as in the case 1 in Venkatraman (1992) Lemma 2.6. There exists es,e (which in the notation of Venkatraman a c > 0 such that, for every d ∈ [ηk , ηk + c1 ∆/16], feηs,e k − fd (1992) is the term El ) can be written as feηs,e |d−ηk | √ k e − ηk p ηk − s + (d − ηk ) p e−s (ηk − s + (d − ηk ))(e − ηk ) + p (ηk − s)(e − ηk − (d − ηk )) Using the inequality (e − s) ≥ 2c1 ∆, the previous expression is lower bounded by . ≥ c′ |d − ηk |feηs,e ∆(e − s)−2 . k Case 2. Let h = c1 ∆/8 and l = d − ηk ≤ h/2. Then, following closely the initial calculations for case 2 of Lemma 2.6 of Venkatraman (1992), we obtain that − feds,e ≥ E1l (1 + E2l ) + E3l , feηs,e k where E1l = p (ηk − s + l)(e − ηk − l) E2l = p and E3l = − p feηs,e k l(h − l) (ηk − s + l)(e − ηk − l) + p (ηk − s)(e − ηk ) , ((e − ηk − h) − (ηk − s))((e − ηk − h) − (ηk − s) − l)  p  p p (ηk − s + l)(e − ηk − l) + (ηk − s + h)(e − ηk − h) (ηk − s)(e − ηk ) + (ηk − s + h)(e − ηk − h) (feηs,e − feηs,e k )l k +h h s (ηk − s + h)(e − ηk − h) . (ηk − s + l)(e − ηk − l) Since h = c′′ ∆ and l ≤ h/2, Observe that E1l ≥ c′′ feηs,e |d − η|∆(e − s)−2 . k ηk − s ≤ ηk − s + l ≤ ηk − s + h ≤ 9(ηk − s)/8, e − ηk ≥ e − ηk − l ≥ e − ηk − h ≥ 7(e − ηk )/8. (24) 28 Thus E2l = p ((e − ηk − h) − (ηk − s))2 + l(h + ηk − s) − l(e − ηk )  p  p p (ηk − s + l)(e − ηk − l) + (ηk − s + h)(e − ηk − h) (ηk − s)(e − ηk ) + (ηk − s + h)(e − ηk − h) −l(e − ηk ) (ηk − s + h)(e − ηk − h) −l(e − ηk ) ≥ ≥ −1/2 (ηk − s)(7/8)(e − ηk ) ≥ where (24) is used in the second inequality and the fact that l ≤ h/2 ≤ c1 ∆/16 ≤ (ηk − s)/16 is used in the last inequality. For E3l , observe that ≤ 2λ ≤ max |fets,e| − feηs,e ≤ |feηs,e | − feηs,e feηs,e − feηs,e k k k k +h k +h s≤t≤e and that (21) and l/2 ≤ h = c1 ∆/8 imply that ηk − s ≤ ηk − s + l ≤ ηk − s + h ≤ 9(ηk − s)/8 and e − ηk ≥ e − ηk − l ≥ e − ηk − h ≥ 7(e − ηk )/8. Therefore, 2(d − ηk )λ E3l ≥ − c1 ∆/8 s (9/8)(ηk − s)(e − ηk ) (ηk − s)(7/8)(e − ηk ) 32(d − ηk )λ c1 ∆ ′′ ≥ −(c /4)fes,e (d − ηk )∆(e − s)−2 , ≥− ηk where the first inequality follows from (24) and the last inequality follows from (22) and (23), for a sufficiently small c3 . Thus, feηs,e − feds,e ≥ E1l (1 + E2l ) + E3l ≥ (c′′ /4)feηs,e |ηk − d|∆(e − s)−2 . k k The following proposition is a direct consequence of Lemma 8 and essentially characterized the localization error rate of the BS algorithm. Proposition 9. Consider any generic interval (s, e) ⊂ (0, T ) such that ηr−1 ≤ s ≤ ηr ≤ . . . ≤ ηr+q ≤ e ≤ ηr+q+1 , q ≥ 0. (25) Let b ∈ arg maxs≤t≤e |Yets,e |. Suppose for some c1 > 0 and κ > 0, max{min{ηr − s, s − ηr−1 }, min{ηr+q+1 − e, e − ηr+q }} = ǫn , where ǫn < min{(3c1 /8)2 κ2 ∆2 (e − s)−1 B1−2 , ∆/4}, 29 (26) and |Yebs,e | ≥ c1 κ∆(e − s)−1/2 . (27) Assume also that there exists sufficient small c3 > 0 such that sup |fets,e − Yets,e | = λ1 < min{c3 κ∆3 (e − s)−5/2 , s≤t≤e (c1 /4)κ∆(e − s)−1/2 }. (28) Then there exists a change point ηk ∈ [s, e] and an absolute constant C1 > 0 such that and min{e − ηk , ηk − s} > (3c1 /8)2 κ2 ∆2 (e − s)−1 B1−2 (29) |feηs,e | ≥ |Yebs,e | − λ1 ≥ max |fets,e| − 2λ1 . k (30) |ηk − b| ≤ C1 λ1 (e − s)5/2 ∆−2 κ−1 , s≤t≤e Proof. Observe that from (28), max |fets,e| ≤ max |Yets,e | + λ1 ≤ |Yebs,e | + λ1 ≤ |febs,e | + 2λ1 . s<t<e s<t<e (31) Suppose ηk ≤ b ≤ ηk+1 for some r − 1 ≤ k ≤ r + q. Observe that |febs,e | ≥ |Yebs,e | − λ1 > (3c1 /4)κ∆(e − s)−1/2 > 0, where the second inequality follows from (27) and (28). It suffices to consider the case in which febs,e > 0, since, if febs,e < 0, then the same arguments can be applied to the time series {−fi }ni=1 . From Lemma 7, fets,e is either monotonic or decreasing and then increasing on [ηk , ηk+1 ]. Thus } ≥ febs,e. , feηs,e max{feηs,e k+1 k es,e If fets,e is locally decreasing at b, then feηs,e k ≥ fb . Therefore ≥ febs,e > (3c1 /4)κ∆(e − s)−1/2 . feηs,e k (32) Step 1. We first show that (32) implies (29). For the sake of contradiction, suppose that min{e − ηk , ηk − s} ≤ ((3c1 /8)κ∆(e − s)−1/2 B1−1 )2 . Then r r (e − η )(η − s) (e − ηk )(ηk − s) k k s,e feηk ≤ B1 + B1 e−s e−s p ≤ 2 min{e − ηk , ηk − s}B1 ≤ (3c1 /4)κ∆(e − s)−1/2 . This is a contradiction to (32). Therefore (29) holds for ηk . Step 2. We now apply Lemma 8, since (22) and (23) hold in virtue of (32) and (28), respectively. Thus, we will need to prove that min{ηk − s, e − ηk } ≥ (3/4)∆. 30 For the sake of contradiction, assume that min{ηk − s, e − ηk } < (3/4)∆. Suppose ηk − s < (3/4)∆. Since ηk − ηk−1 ≥ ∆, one has s − ηk−1 ≥ ∆/4. This also means that ηk is the first change point within [s, e]. Therefore k = r in (25). By (26), min{ηr − s, s − ηr−1 } ≤ ǫ < ∆/4. Since s − ηk−1 = s − ηp−1 ≥ ∆/4, it must be the case that ηk − s = ηr − s ≤ ǫ ≤ (3c1 /8)2 κ2 ∆2 (e − s)−1 B1−2 . This is a contradiction to (29). Therefore ηk − s ≥ (3/4)∆. The argument of e − ηk ≥ (3/4)∆ can be made analogously. Step 3. By Lemma 8, if there exists ad and a sufficient large constant C1 satisfying d ∈ [ηk , ηk + C1 λ1 (e − s)5/2 ∆−2 κ−1 ], then |ηk − d|∆(e − s)−2 ≥ λ1 , − feds,e > cfeηs,e feηs,e k k where the last inequality follows from (32) and (28). For the sake of contradiction, suppose b ≥ d. Then − λ1 ≤ max |fets,e| − 2λ1 , febs,e ≤ feds,e < feηs,e k s<t<e where the first inequality follows from Lemma 7 which ensures that fets,e is decreasing on [ηp , b] and d ∈ [ηp , b]. This is a contradiction to (31). Thus b ∈ [ηk , ηk + C1 λ1 (e − s)5/2 ∆−2 κ−1 ]. (30) follows from (31) and (32). The argument for the case when febs,e is locally increasing at b is similar and therefore we omit the details. √ Corollary 10. Let δ > 0 be such that 2 δB1 ≤ (3c1 /4)κ∆(e−s)−1/2 . Let b′ = arg max⌈s+δ⌉≤t≤⌊e−δ⌋ |Yets,e |. Suppose all the assumption in Proposition 9 hold except that (27) and (28) are replaced by and −1/2 |Yebs,e ′ | ≥ c1 κ∆(e − s) sup ⌈s+δ⌉,...,⌊e−δ⌋ (33) |fets,e − Yets,e | = λ1 < min{c3 κ∆3 (e − s)−5/2 , (c1 /4)κ∆(e − s)−1/2 }, (34) respectively. Then, all the conclusions of Proposition 9 still hold for b′ . Proof. For any t ∈ [s + δ] ∪ [e − δ], p √ |fets,e | ≤ 2 min{e − t, t − s}B1 ≤ 2 δB1 ≤ (3c1 /4)κ∆(e − s)−1/2 . Let ηk and b defined as in the proof of Proposition 9. Then by (32) ηk , b ∈ ⌈s + δ⌉, . . . , ⌊e − δ⌋. Therefore b = b′ and all the conclusions of Proposition 9 still hold. 31 B.2 Results from Fryzlewicz (2014) Below, we will derive some further properties of the CUSUM statistic using the ANOVA decompositiontype of arguments first introduced by Fryzlewicz (2014), which are particularly effective at separating the noise from the signal. Some of the results below are contained in Fryzlewicz (2014), but others requires different, subtle arguments. For completeness, we include all the proofs. For a pair (s, e) of positive integers with s < e, let Wds,e be the two dimensional linear subspace of Re−s spanned by the vectors u1 = (1, . . . , 1, 0, . . . , 0)⊤ | {z } | {z } d−s u2 = (0, . . . , 0, 1, . . . , 1)⊤ . | {z } | {z } e−d d−s e−d For clarity, we will use h , i to denote the inner product of two vectors in the Euclidean space. Lemma 11. For x = (xs+1 , . . . , xe )⊤ ∈ Re−s , let Pds,e (x) be the projection of x onto Wds,e . 1. The projection Pds,e (x) satisfies Pds,e (x) e 1 X = xi + hx, ψds,e iψds,e , e−s i=s+1 where h·, ·i is the inner product in Euclidean space, and ψds,e = ((ψds,e )s , . . . , (ψds,e )e−s )⊤ with q e−d  i = s + 1, . . . , d, (e−s)(d−s) , s,e q (ψd )i = d−s − (e−s)(e−d) , i = d + 1, . . . , e, i.e. the i-th entry of Pds,e (x) satisfies ( Pds,e (x)i 2. Let x̄ = 1 e−s Pe i=s+1 xi . = 1 d−s 1 e−d Pd xj , Pj=s+1 e j=d+1 xj , i = s + 1, . . . , d, i = d + 1, . . . , e. Since hx̄, ψds,e i = 0, kx − Pds,e (x)k2 = kx − x̄k2 − hx, ψds,e i2 . Proof. The results hold following the fact that the projection matrix of subspace Wds,e is   1/(d − s) · · · 1/(d − s) 0 ··· 0   .. .. .. .. .. ..   . . . . . .     1/(d − s) · · · 1/(d − s) 0 · · · 0 s,e . PW s,e =   0 ··· 0 1/(e − d) · · · 1/(e − d)  d     .. .. .. .. .. ..   . . . . . . 0 ··· 0 1/(e − d) · · · 1/(e − d) 32 (35) For any pair d1 , d2 ∈ {s + 1, . . . , e} and f ∈ Re−s , the following two statements are equivalent: hf, ψds,e i2 ≤ hf, ψds,e i2 ⇐⇒ kf − Pds,e (f )k2 ≥ kf − Pds,e (f )k2 . 1 2 1 2 Lemma 12. Assume Assumption 1. Let [s0 , e0 ] be an interval with e0 − s0 ≤ CR ∆ and containing at lest one change point ηr such that ηr−1 ≤ s0 ≤ ηr ≤ . . . ≤ ηr+q ≤ e0 ≤ ηr+q+1 , q ≥ 0. Suppose that min{ηp′ − s0 , e0 − ηp′ } ≥ ∆/16 for some p′ and let κs,e max = max{κp : min{ηp − s0 , e0 − ηp } ≥ ∆/16}. Consider any generic [s, e] ⊂ [s0 , e0 ], satisfying min{ηr − s0 , e0 − ηr } ≥ ∆/16 for all ηr ∈ [s, e]. Let b ∈ arg maxs<t<e |Yets,e |. For some c1 > 0, λ > 0 and δ > 0, suppose that √ |Yebs,e | ≥ c1 κs,e max ∆ sup |Ye s,e − fes,e | ≤ λ s<t<e t (36) t and sup √ s1 <t<e1 1 e1 − s1 e1 X (Yt − ft ) ≤ λ for every t=s1 +1 If there exists a sufficient small c2 > 0 such that √ λ ≤ c2 κs,e max ∆ and e1 − s1 ≥ δ. δ ≤ c2 ∆, (37) (38) then there exists a change point ηk ∈ (s, e) such that min{e − ηk , ηk − s} > ∆/4 and |ηk − b| ≤ min{C3 λ2 κ−2 k , δ}. Proof. Without loss of generality, assume that febs,e > 0 and that fets,e is locally decreasing at b. Observe that there has to be a change point ηk ∈ [s, b], or otherwise febs,e > 0 implies that fets,e is decreasing, as a consequence of Lemma 18. Thus if s ≤ ηk ≤ b ≤ e, then √ √ √ s,e s,e ≥ febs,e ≥ |Yebs,e | − λ ≥ c1 κs,e feηs,e (39) max ∆ − c2 κmax ∆ ≥ (c1 /2)κmax ∆. k Observe that e − s ≤ e0 − s0 ≤ CR ∆ and that that (s, e) has to contain at least one change point or otherwise |feηs,e k | = 0 which contradicts (39). Step 1. In this step, we are to show that min{ηk − s, e − ηk } ≥ min{1, c21 }∆/16. Suppose ηk is the only change point in (s, e). So min{ηk − s, e − ηk } ≥ min{1, c21 }∆/16 must hold or otherwise it follows from Lemma 17, we have √ c1 √ c1 | < κk ∆ ≤ κs,e |feηs,e max ∆, k 4 2 which contradicts (39). 33 Suppose (s, e) contains at least two change points. Then ηk − s ≤ min{1, c21 }∆/16 implies that ηk is the first change point in [s, e]. Therefore √ √ 1 es,e 1 es,e| + c1 κr ∆ | + 2κ | f max | f | ≤ |feηs,e η − s ≤ r r η t k 4 k+1 4 s<t<e 2 √ 3 e s,e c1 s,e 1 e s,e ≤ |Yb | + λ + κmax ∆ ≤ |Yb | + λ < |Yebs,e | − λ 4 2 4 where the first inequality follows from Lemma 18, the fourth inequality follows from (36), and the last inequality holds when c2 is sufficiently small. This contradicts (39). Step 2. By Lemma 8 there exists d such that √ −1 d ∈ [ηk , ηk + λ ∆(κs,e max ) ] es,e > 2λ. For the sake of contradiction, suppose b ≥ d. Then and that feηs,e k − fd febs,e ≤ feds,e < feηs,e − 2λ ≤ max |fets,e| − 2λ ≤ max |Yets,e | + λ − 2λ = |Yebs,e | − λ, k s<t<e s<t<e that fets,e is decreasing on [ηp , b] and where the first inequality follows from Lemma 7, which ensures √ −1 d ∈ [ηp , b]. This is a contradiction to (39). Thus b ∈ [ηk , ηk + λ ∆(κs,e max ) ]. Step 3. Let f s,e = (fs+1 , . . . , fe )⊤ ∈ R(e−s) and Y s,e = (Ys+1 , . . . , Ye )⊤ ∈ R(e−s) . By the definition of b, it holds that Y s,e − Pbs,e (Y s,e ) 2 (Y s,e ) ≤ Y s,e − Pηs,e k 2 2 ≤ Y s,e − Pηs,e (f s,e) . k For the sake of contradiction, throughout the rest of this argument suppose that, for some sufficient large constant C3 to be specified, ηk + max{C3 λ2 κ−2 k , δ} < b. (This will of course imply that ηk + the bound −2 max{C3 λ2 (κs,e max ) , δ} Y s,e − Pbs,e (Y s,e ) 2 (40) < b). We will show that this leads to 2 (f s,e ) , > Y s,e − Pηs,e k (41) which is a contradiction. derive (41) from (40), we note that min{e−ηk , ηk −s} ≥ min{1, c21 }∆/16 √ Tos,e and that |b − ηk | ≤ λ ∆(κmax )−1 implies that √ −1 min{e − b, b − s} ≥ min{1, c21 }∆/16 − λ ∆(κs,e ≥ min{1, c21 }∆/32, (42) max ) where the last inequality follows from (38) and holds for an appropriately small c2 > 0. Equation (41) is in turn implied by 2hεs,e , Pb (Y s,e ) − Pηk (f (s,e) )i < kf s,e − Pb (f s,e )k2 − kf s,e − Pηk (f s,e )k2 , where εs,e = Y s,e − f s,e. By (35), the right hand side of (43) satisfied the relationships kf s,e − Pb (f s,e )k2 − kf s,e − Pηk (f s,e )k2 = hf s,e, ψηk i2 − hf s,e, ψb i2 = (fes,e)2 − (fes,e )2 ηp b | − febs,e)|feηs,e ≥ (feηs,e k k s,e 2 ≥ c|d − ηk |(fe ) ∆−1 ′ ≥ c |d − 34 ηk 2 ηk |(κs,e max ) , (43) where Lemma 8 and (39) are used in the second and third inequalities. The left hand side of (43) can in turn be rewritten as 2hεs,e , Pb (X s,e ) − Pηk (f s,e)i = 2hεs,e , Pb (X s,e ) − Pb (f s,e)i + 2hεs,e , Pb (f s,e ) − Pηk (f s,e)i. (44) The second term on the right hand side of the previous display can be decomposed as   ηk e b X X X  εs,e (Pb (f s,e)i − Pηk (f s,e )i ) + + hεs,e , Pb (f s,e ) − Pηk (f s,e )i =  i i=s+1 i=ηk +1 i=b+1 = I + II + III. In order to bound the terms I, II and III, observe that, since e − s ≤ e0 − s0 ≤ CR ∆, the interval [s, e] must contain at most CR + 1 change points. Let ηr′ −1 < s ≤ ηr′ ≤ . . . ≤ ηp′ +q′ < e ≤ ηp′ +q′ +1 . Then p′ + q ′ + 1 − r ′ ≤ CR + 1. Step 4. We can write √ I = ηk − s √ 1 ηk − s ηk X i=s+1 εs,e i ! ηk b X 1 X 1 fi − fi b−s ηk − s i=s+1 i=s+1 ! . Thus, Pk P Pk ηk b X (ηk − s)( ηi=s+1 fi + bi=ηk +1 fi ) − (b − s) ηi=s+1 fi 1 X 1 fi = fi − b−s ηk − s (b − s)(ηk − s) i=s+1 i=s+1 P Pk Pk fi + (ηk − s) bi=ηk +1 fi ) (ηk − b) ηi=s+1 (ηk − b) ηi=s+1 fi + (ηk − s)(b − ηk )fηk +1 ) = = (b − s)(ηk − s) (b − s)(ηk − s) = ηk X b − ηk b − ηk 1 fi + fηk+1 ≤ (CR + 1)κs,e − max b−s ηk − s b−s i=s+1 where Lemma 16 is used in the last inequality. It follows from Equation (37) that √ √ |b − ηk | 4 2 s,e (CR + 1)κmax ≤ |b − ηk |∆−1/2 λ(CR + 1)κs,e |I| ≤ ηk − sλ max , b−s min{1, c1 } where (42) is used in the last inequality. Step 5. For the second term II, we have that    d b e X X X p 1 1 1 εs,e   fi − fi  |II| = b − ηk  √ b−s e − ηk b − ηk i=η +1 i i=s+1 i=ηk +1 k   b e X X p 1 1 fi − fηk + fi − fηk+1  ≤ b − ηk λ  fηk − fηk+1 + b−s e − ηk i=s+1 i=ηk +1 p s,e s,e ≤ b − ηk (κs,e max + (CR + 1)κmax + (CR + 1)κmax ), 35 where the first inequality follows from (42) and (37), and the second inequality from Lemma 16. Step 6. Finally, we have that III = Therefore, |III| ≤ √ √ e−b 1 e−b e X i=b+1 ! εs,e  i 1 e − ηk e X i=ηk +1 fi − 1 e−b e X i=b+1  fi  . √ b − ηk 4 2 s,e (CR + 1)κmax ≤ |b − ηk |∆−1/2 λ(CR + 1)κs,e e − bλ max . e−b min{1, c1 } Step 7. Using the first part of Lemma 11, the first term on the right hand side of (44) can be bounded as hεs,e , Pd (X s,e ) − Pd (f s,e )i ≤ λ2 . Thus (43) holds if n o p 2 −1/2 s,e s,e 2 b − η |b − ηk |(κs,e ) ≥ C max |b − η |∆ λκ , . λκ , λ k k max max max √ Since λ ≤ c3 ∆κ, the first inequality holds. The second inequality follows from |b − ηk | ≥ −2 C3 λ2 (κk )−2 ≥ C3 λ2 (κs,e max ) , as assumed in (40). This completes the proof. Corollary 13. Let [s0 , e0 ] be a generic interval satisfying e0 − s0 ≤ CR ∆ and containing at lest one change point ηr such that ηr−1 ≤ s0 ≤ ηr ≤ . . . ≤ ηr+q ≤ e0 ≤ ηr+q+1 , κs,e max q ≥ 0. = max{κp : min{ηp −s0 , e0 −ηp } ≥ Suppose min{ηp′ −s0 , e0 −ηp′ } ≥ ∆/16 for some p′ and, let ∆/16}. Consider a generic interval (s, e) ⊂ (s0 , e0 ), satisfying min{ηp − s0 , e0 − ηp } ≥ ∆/16 for all ηp ∈ [s, e]. Let δ′ > 0 be some constant and b′ = arg max⌈s+δ′ ⌉,...,⌊e−δ′ ⌋ |Yets,e |. Suppose in addition that, for some positive constants c1 and c2 , √ s,e ∆, (45) |Yebs,e ′ | ≥ c1 κmax s,e s,e sup |Ye − fe | ≤ λ, (46) t=⌈s+δ′ ⌉,...,⌊e−δ′ ⌋ t e1 X 1 (Yt − ft )| ≤ λ, | e1 − s1 t=s +1 s1 ≤t≤e1 1 √ s,e λ ≤ c2 κmax ∆, sup and t √ for every δ′ ≤ c2 ∆. e1 − s1 ≥ δ′ , (47) (48) (49) Then there exists a change point ηk ∈ [s, e] such that min{e − ηk , ηk − s} > ∆/4 ′ |ηk − b′ | ≤ max{C3 λ2 κ−2 k , δ }. Proof. By the same proof of Corollary 10, if b is defined as in Lemma 12, then b = b′ if c2 is sufficiently small. 36 C Proofs of the Results from Section 3 e u = σ 2 Ip + κuu′ . Observe that if κ ≤ σ 2 /4, Proof of Lemma 3. For any vector u ∈ Rp , denote Σ e u kop = σ 2 + κ ≤ 2σ 2 . kΣ Step 1. p n denote the joint distribution of independent random vectors {X }n Let Pe0,u i i=1 in R such that i.i.d. e u) X1 , . . . , X∆ ∼ Np (0, Σ i.i.d. and X∆+1 , . . . , Xn ∼ Np (0, σ 2 I). n denote the joint distribution of independent random vectors {X }n p Similarly, let Pe1,u i i=1 in R with i.i.d. X1 , . . . , Xn−∆ ∼ Np (0, σ 2 I) P i.i.d. e u ). and Xn−∆+1 , . . . , Xn ∼ Np (0, Σ n √ en u∈{±1}p / p Pi,u . Let η(Pi,u ) denote the location of the n . Then since η(P en ) = ∆ and η(Pen ) = n − ∆ to the distribution Pei,u 0,u 1,u n n e e |η(P0,u ) − η(P1,u )| ≥ n/3. By Le cam’s lemma (see, e.g. Yu, 1997), Let Pein = 1 2p inf sup η̂ P ∈P n κ e ,∆,σ change point associated √ for any u ∈ {±1}p / p, EP (|η̂ − η|) ≥ (n/3)(1 − dT V (Pe0n , Pe1n )), where T V (Pe0n , Pe1n ) = 21 kPe0n − Pe1n k1 . n Let Σu = Ip + κ euu′ , where κ e = κ/σ 2 . Observe that by assumption κ e ≤ 1/4. Denote with P0,u n the joint distributions of independent samples {X }n p and P1,u i i=1 in R where i.i.d. i.i.d. X1 , . . . , X∆ ∼ Np (0, Σu ) and i.i.d. X1 , . . . , Xn−∆ ∼ Np (0, I) and X∆+1 , . . . , Xn ∼ Np (0, I) and ii.d. Xn−∆+1 , . . . , Xn ∼ Np (0, Σu ), respectively. Since total variation distance is invariant under rescaling of the covariance, then kPe0n − Pe1n k1 = kP0n − P1n k1 . Therefore inf sup η̂ P ∈P n κ e ,∆,σ EP (|η̂ − η|) ≥ (n/3)(1 − dT V (P0n , P1n )) (50) Step 2. Let x = (x1 , . . . , x∆ ), y = (x∆+1 , . . . , xn−∆ ) and z = (xn−∆+1 , . . . , xn ). Let • f0 (x) denote the joint distribution of X1 , . . . , X∆ ∼ Np (0, I) and fu (x) denote the joint i.i.d. distribution of X1 , . . . , X∆ ∼ Np (0, Σu ); i.i.d. • g0 (x) denote the joint distribution of X∆+1 , . . . , Xn−∆ ∼ Np (0, I); i.i.d. • h0 (x) denote the joint distribution of Xn−∆+1 , . . . , Xn ∼ Np (0, I) and hu (x) denote the i.i.d. joint distribution of Xn−∆+1 , . . . , Xn ∼ Np (0, Σu ). 37 Then, kP0n − P1n k1 Z Z Z 1 = 2p = = ≤ Z g0 (y)dy Z Z 1 2p Z Z 1 2p X √ u∈{±1}p / p Z Z 1 2p fu (x)g0 (y)h0 (z) − X √ u∈{±1}p / p 1 2p fu (x)h0 (z) − X √ u∈{±1}p / p 1 2p f0 (x)g0 (y)hu (z) dxdydz X √ u∈{±1}p / p X fu (x)h0 (z) − X fu (x)h0 (z) − f0 (x)h0 (z) + f0 (x)g0 (z) − √ u∈{±1}p / p √ u∈{±1}p / p X f0 (x)hu (z) dxdz 1 2p √ u∈{±1}p / p f0 (x)hu (z) dxdz 1 2p X √ u∈{±1}p / p =2kP0∆ − P1∆ k1 , i.i.d. where P0∆ is the joint distribution of X1 , . . . , X∆ ∼ Np (0, I) and P1∆ = f0 (x)hu (z) dxdz 1 2p i.i.d. P ∆ √ u∈{±1}p / p P1,u , ∆ is the joint distribution of X , . . . , X where P1,u 1 ∆ ∼ Np (0, Σu ). Thus (50) becomes inf sup η̂ P ∈P n κ e ,∆,σ EP (|η̂ − η|) ≥ (n/3)(1 − kP0∆ − P1∆ k1 ). (51) Step 3. To bound 2kP0∆ − P1∆ k1 , let P0 = Np (0, Ip ) and Pu = Np (0, Σu ). It is easy to see that χ 2 (P1∆ , P0∆ ) For any u, v ∈ {±1}p , EP ∆ 0  = EP ∆ 0  dPu∆ dPv∆ dP0∆ dP0∆ 2 1 dP1∆ −1 = p ∆ 4 dP0  =  EP0  X EP ∆ 0 u,v∈{±1}p dPu dPv dP0 dP0 ∆  dPu∆ dPv∆ dP0∆ dP0∆  − 1. = (1 − (e κu′ v)2 )−∆/2 , where the last equality follows from Lemma 14. Denote U and V to be the p dimensional Rademacher variables with U being independent of V and εp = (1′ V /p)2 . Thus χ2 (P1∆ , P0∆ ) = 1 4p X u,v∈{±1}p (1 − (e κu′ v)2 )−∆/2 − 1  ∆/2  κU ′ V /p)2 −1 = EU,V 1 − (e   ∆/2 = EV 1 − (e −1 κ1′ V /p)2  ≤ E exp(εp κ e2 ∆) − 1, 38 where √ the inequality follows from inequality (1 − t)−∆/2 ≤ exp(∆t) for any t ≤ 1/2 and that κ e ≤ 1/ 2. The Höffding’s equality, applied to Rademacher variables, gives P (εp ≥ λ) ≤ 2e−2pλ . (52) Thus Z ∞  P exp εp κ e2 ∆ ≥ u du 0 Z ∞  ≤1+ P ε ≥ log(u)/(e κ2 ∆) du   Z1 ∞ log(u)2p 2 exp − ≤1+ du κ e2 ∆ 1 2 , =1− 1 − κe2p 2∆  E exp(εp κ e ∆) = 2 where the second inequality follows from (52), and the last equality holds if κe2p 2 ∆ > 1. Thus if q 4 2p = 2pσ , then, using the well-known fact that kP0∆ − P1∆ k1 ≤ 2 χ2 (P1∆ , P0∆ ), we obtain ∆ = 33e κ2 33κ2 the bounds s q 2 ∆ ∆ = 1/2. kP0 − P1 k1 ≤ 2 χ2 (P1∆ , P0∆ ) ≤ 2 2p −1 κ e2 ∆ This and (51) give inf sup EP (|η̂ − η|) ≥ n/6. η̂ P ∈P Proof of Lemma 4. Let  n Q = Pκ,∆ : ∆ < n/2 . Using the same argument as in Step 1 of the proof of Lemma 3, it suffices to take κ = κ/σ 2 ≤ 1/4, e and consider Σu = Ip + κ euu′ . Let u to be any unit vector in Rp and δ a positive number. Let P0n denote the joint distribution of independent samples {Xi }ni=1 in Rp where iid. iid. X1 , . . . , X∆−1 ∼ Np (0, I), X∆ , . . . , Xn ∼ Np (0, Σu ) and P1n the joint distribution of independent samples {Xi }ni=1 in Rp where iid. X1 , . . . , X∆+δ ∼ Np (0, I), iid. X∆+δ+1 , . . . , Xn ∼ Np (0, Σu ). n . Then since η(P n ) = ∆ and Let η(Pin ) denote the location of the change point of distribution Pi,u 0 η(P1n ) = ∆ + δ, thus |η(P0n ) − η(P1n )| ≥ δ. By Le cam’s lemma (Yu, 1997), 1 inf sup EP (|η̂ − η|) ≥ δ(1 − dT V (P0n , P1n )) = δ(1 − kP0n − P1n k1 ) η̂ P ∈Q 2 39 Since P0n and P1n are identically distributed on X1 , . . . X∆ − 1 and on X∆+δ , . . . , Xn , kP0n − P1n k1 = kP0δ − P1δ k1 , where P0δ is the joint distribution of X1 , Xδ ∼ Np (0, I) and P1δ is the joint distribution of X1 , Xδ ∼ Np (0, Σu ). By Lemma 14, χ2 (P1δ , P0δ ) = (1 − κ e2 )−δ/2 − 1 ≤ 4e κ2 δ. if κ2 δ/2 ≤ 1/4 and δ/2 ≥ 2. Thus by taking δ = e κ−2 /4 ≥ 4 we obtain that   q 1 δ δ 2 χ (P1 , P0 ) ≥ e inf sup EP (|η̂ − η|) ≥ δ 1 − κ−2 /32 = σ 4 κ−2 /32. η̂ P ∈Q 2 Lemma 14. Let P0 = Np (0, Ip ) and Pu = Np (0, Ip + κuu′ ). Then   −1/2 dPu dPv EP0 = 1 − (κu′ v)2 . dP0 dP0 Proof. See, e.g., lemma 5.1 in Berthet and Rigollet (2013). Lemma 15. For t ≥ 2, x ≥ 0, if tx ≤ 1/4, (1 − x)−t − 1 ≤ 4tx Proof. There exists s ∈ [0, x] such that (1 − x)−t − 1 = tx + t(t + 1)x2 (1 − s)−t−2 ≤ tx + 4t2 x2 (1 − x)−t ≤ 2tx + 4t2 x2 ((1 − x)−t − 1). Thus (1 − x)−t − 1 ≤ D 2tx ≤ 4tx 1 − 4t2 x2 Properties of the covariance CUSUM statistic D.1 Properties of 1d CUSUM statistics Lemma 16. Suppose [s, e] ⊂ [1, T ] such that e − s ≤ CR ∆, and that ηr−1 ≤ s ≤ ηr ≤ . . . ≤ ηr+q ≤ e ≤ ηr+q+1 , q ≥ 0. Denote κs,e max = max{ηp − ηp−1 : r ≤ p ≤ r + q}. Then for any r − 1 ≤ p ≤ r + q, e 1 X fi − fηp ≤ CR κs,e max . e−s i=s 40 Proof. Since e − s ≤ CR ∆, the interval [s, e] contains at most CR + 1 change points. Observe that e 1 X fi − fηp e−s i=s ηr+1 ηr X 1 X = (fηr−1 − fηp ) + (fηr − fηp ) + . . . + e−s i=s i=ηr +1 e X ηr+1 ηr X 1 X s,e ≤ |p − r|κmax + |p − r − 1|κs,e max + . . . + e−s ≤ 1 e−s i=s e X (fηr+q − fηp ) i=ηr+q +1 i=ηr +1 e X i=ηr+q +1 |p − r − q − 1|κs,e max (CR + 1)κs,e max , i=s where |p1 − p2 | ≤ CR + 1 for any ηp1 , ηp2 ∈ [s, e] is used in the last inequality. Lemma 17. If ηp is the only change point in [s, e], then r q (ηp − s)(e − ηp ) s,e e κp ≤ min{ηp − s, e − ηp }κp |fηp | = e−s Lemma 18. Let [s, e] contains two or more change points such that ηr−1 ≤ s ≤ ηr ≤ . . . ≤ ηr+q ≤ e ≤ ηr+q+1 , q ≥ 1. If ηr − s ≤ c21 ∆ then √ |feηs,e | ≤ c1 |feηs,e | + 2κr ηr − s. r r+1 This can be useful in testing when there are exactly two change points with ηr − s ≤ λ2 κ−2 r , e − ηr+1 ≤ λ2 κ−2 r+1 . It is also useful to show ηr − s ≥ ∆/4 for some absolute constant c when | ≥ max |fets,e | − 2λ. |feηs,e r s≤t≤e Proof. Consider the sequence {gt }et=s+1 be such that ( fηr+1 if s + 1 ≤ t ≤ ηr , gt = ft if ηr + 1 ≤ t ≤ e. For any t ≥ ηr , feηs,e − geηs,e = r r s √ (e − s) − t (ηr − s)(fηr+1 − fηr ) ≤ ηr − sκr (e − s)(t − s) 41 Thus √ | + ηr − sκr | ≤ |e gηs,e |feηs,e r r s √ (ηr − s)(e − ηr+1 ) s,e ≤ |e gηr+1 | + ηr − sκr (ηr+1 − s)(e − ηr ) r √ c21 ∆ s,e |e gηr+1 | + ηr − sκr ≤ ∆ √ e ≤ c1 |fηs,e | + 2 ηr − sκr . r+1 where the first inequality follows from the observation that the first change point of gt in [s, e] is at ηr+1 . D.2 Properties of the covariance CUSUM statistics All of our consistency results heavily rely on the properties of population quantity of the CUSUM statistic. In the covariance change point detection problem, however, it is not trivial to analyze the e s,e properties of the function t 7→ Σ in the multiple change point case. For example, it is difficult t op e s,e to determine the regions of monotonicity of Σ as a function of t as is done in Venkatraman t op (1992, Lemma 2.2). As a remedy, we introduce the concept of shadow vector, which is defined as a maximizer of the operator norm of the CUSUM statistics in all the following results. In this way, we turn the covariance change point detection problem into a mean change point detection problem. For any v ∈ Rp with kvk = 1, let Yi (v) = (v ⊤ Xi )2 and fi (v) = v ⊤ Σi v, i = 1, . . . , n. Note that both {Yi (v)} and {fi (v)} are univariate sequences, we hence have the corresponding CUSUM statistics defined below s s t e X X e − t t−s s,e e Yi (v) − Yi (v), (53) Yt (v) = (e − s)(t − s) (e − s)(e − t) i=s+1 i=t+1 s s t e X X e − t t−s s,e fi (v) − fi (v). (54) fet (v) = (e − s)(t − s) (e − s)(e − t) i=s+1 i=t+1 The key rationale of the CUSUM based BS algorithm or any variants thereof being a powerful tool selecting the change points is that the population version of the CUSUM statistic achieving its maxima at the true change points. In Lemma 19, we show the same holds for the covariance CUSUM statistic. e s,e Lemma 19. Assume (s, e) ∩ {ηk }K k=1 6= ∅ and Assumption 1. The quantity Σt of t achieves its maxima at the true change points, i.e. arg max t=s+1,...,e−1 e s,e Σ t op op as a function ∩ {ηk }K k=1 6= ∅. Proof. For the sake of contradiction suppose that there exists t∗ ∈ (s, e) \ {ηk }K k=1 such that t∗ ∈ arg max t=s+1,...,e−1 42 e s,e Σ t op , and e s,e Σ t∗ op > e s,e Σ ηk max k: ηk ∈(s,e) op . n ′ n e s,e Let v ∈ arg maxkuk=1 |u′ Σ t∗ u|, and consider the sequence {fi (v)}i=1 = {v Σi v}i=1 . By the above display, we have e s,e fets,e Σ ∗ (v) = t∗ > op max k: ηk ∈(s,e) e s,e Σ ηk op ≥ max k: ηk ∈(s,e) feηk (v) , (55) where fets,e (v) is defined in (54). It follows from Lemma 2.2 of Venkatraman (1992), the quantities fets,e (v) is maximized at the change points of the time series {ft (v)}et=s+1 . Note that the change points of the sequence {ft (v)}et=s+1 is a subset of {ηk }K k=1 , this contradicts (55). Lemma 19 shows that the population version of the covariance CUSUM statistic is maximized at the true change points in terms of the operator norm. In Lemma 20 below, we give the lower bound of the maxima thereof. One can interpret it as the signal strength. Lemma 20. Under Assumption 1, let 0 ≤ s < ηk < e ≤ n be any interval satisfying min{ηk − s, e − ηk } ≥ c1 ∆. Then for any 0 < δ < (c1 /8)∆, max t=⌈s+δ⌉,...,⌊e−δ⌋ Proof. Let e s,e kop ≥ (c1 /2)κ∆(e − s)−1/2 . kΣ t  v ∈ arg max u⊤ Σηk − Σηk+1 u . kuk=1 Therefore we have e s,e Σ t op ≥ fets,e(v) . Since  fηk (v) − fηk+1 (v) = v ⊤ Σηk − Σηk+1 v = Σηk − Σηk+1 max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e kop ≥ kΣ t max t=⌈s+δ⌉,...,⌊e−δ⌋ op ≥ κ, |fets,e(u)| ≥ (c1 /2)κ∆(e − s)−1/2 , where the second inequality follows from the same arguments in Venkatraman (1992, Lemma 2.4) by regarding fets,e(u) as the CUSUM statistic for a univariate time series. In Lemma 21, we show that we can conform the problem of change point detection on covariance to the one on mean via shadow vectors. This translation allows us to convert a high dimensional covariance problem to a univariate mean problem. Lemma 21. Under the same assumptions as Lemma 20, let b∈ and denote a shadow vector by arg max t=⌈s+δ⌉,...,⌊e−δ⌋ Sets,e op v ∈ arg max u⊤ Sebs,e u . , kuk=1 Then b∈ arg max t=⌈s+δ⌉,...,⌊e−δ⌋ 43 Yets,e (v) . (56) Proof. It suffices to show that Yebs,e (v) = Sebs,e op ≥ max t=⌈s+δ⌉,...,⌊e−δ⌋ Sets,e op ≥ max t=⌈s+δ⌉,...,⌊e−δ⌋ Yets,e (v) . For the shadow vector v defined in (56), it is tempting to argue that we can estimate v and then use {Yi (v)}ni=1 as the new sequence to derive an upper for the rate of localization. This ideal approach, however, suffers from two non-trivial obstacles. • We don’t have any guarantee that the estimation of v is consistent. This is because estimating the first principle direction of any sample matrix in general requires a non-vanish gap between the first and the second eigenvalues of the corresponding population matrix. Since b depends on the data, without more involved additional assumption, it is difficult to show Sebs,e converges to a population quantity as n → ∞. • Suppose v given in (56) has a well defined population quantity, say E(v). The estimation of v depends on {Xi }ni=1 and therefore E(Yi (v)) 6= fi (E(v)). On one hand our knowledge of the population version of the shadow vector v is very limited; on the other hand, in Lemma 22 below we show that without estimating the population shadow vector, the CUSUM statistics and its corresponding population quantity are close enough and that the maximum of the CUSUM statistic is detectable. Lemma 22. Under the same assumptions as Lemma 20, assume max t=⌈s+δ⌉,...,⌊e−δ⌋ e s,e Sets,e − Σ t op ≤ λ, where λ1 > 0. For the shadow vector v defined in (56), then max t=⌈s+δ⌉,...,⌊e−δ⌋ Yets,e (v) ≥ (c1 /2)κ∆(e − s)−1/2 − λ. Proof. Observe that for all compatible t and v,  e s,e e s,e v ≤ Ses,e − Σ Yets,e (v) − fets,e (v) = v ⊤ Sets,e − Σ t t t op ≤ λ. In addition, for the shadow vector v, we have max t=⌈s+δ⌉,...,⌊e−δ⌋ ≥ max t=⌈s+δ⌉,...,⌊e−δ⌋ Yets,e (v) = e s,e Σ t op max t=⌈s+δ⌉,...,⌊e−δ⌋ Sets,e op − λ1 ≥ (c1 /2)κ∆(e − s)−1/2 − λ, where the last inequality follows from Lemma 20. We finally prove a simple property of covariance CUSUM statistics. p e s,e kop ≤ 2 min{e − t, t − s}B 2 for any t ∈ (s, e). Lemma 23. kΣ t Proof. Observe that r p (e − t)(s − t) 2 s,e e kΣt kop ≤ 2 B ≤ 2 min{e − t, t − s}B 2 . e−s 44
10math.ST
arXiv:1702.04436v1 [math.GR] 15 Feb 2017 THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS MUHAMMAD INAM Abstract. We show that all of the Schützenberger complexes of an Adian inverse semigroup are finite if the Schützenberger complex of every positive word is finite. This enables us to solve the word problem for certain classes of Adian inverse semigroups (and hence for the corresponding Adian semigroups and Adian groups). 1. Introduction Groups and semigroups are often presented by a pair hX|Ri,where X denotes the set of set of generators and R denotes the set of defining relations. If R = {(u, v)|u, v ∈ X + }, then the presentation hX|Ri is called a positive presentation. We consider positive presentations throughout this paper. We can construct two undirected graph corresponding to a positive presentation hX|Ri. These graphs are called the left graph and the right graph of the presentation and are denoted by LGhX|Ri and RGhX|Ri respectively. The vertices of the LGhX|Ri and the RGhX|Ri are labeled by the elements of the set X. In the LGhX|Ri, we obtain an edge by joining the vertex labeled by the prefix letter of u with the vertex labeled by the prefix letter of v for all (u, v) ∈ R. The RGhX|Ri is constructed dually by joining the vertex labeled by the suffix letter of u with the vertex labeled by suffix letter of v for all (u, v) ∈ R. If there is no cycle in the left and the right graph of a presentation then the presentation is called a cycle free presentation or an Adian presentation. These presentation were first studied by S. I. Adian [1], where is shown that the finitely presented Adian semigroups SghX|Ri embeds in the corresponding Adian GphX|Ri. Latter in [6] John H. Remmers generalized this result to any Adian presentation and proved that an Adian semigroup SghX|Ri embeds in the corresponding Adian group GphX|Ri. Unless stated otherwise, throughout this paper we consider our presentations to be Adian presentations. A semigroup S is called an inverse semigroup if for every element a ∈ S there exists a unique element b ∈ S such that aba = a and bab = b. This unique element b is called the inverse of a and denoted by a−1 . The natural partial order on the elements of an inverse semigroup S is defined as a ≤ b if and only if a = aa−1 b, for some a, b ∈ S. The congruence relation σ on S is defined as for a, b ∈ S, aσb if and only if there exists an element Date: April 11, 2018. 1 2 MUHAMMAD INAM c ∈ S such that c ≤ ab. It turns out that the σ is the minimal group congruence on S. So, S/σ is the maximal group homomorphic image of S and if S is presented by a presentation InvhX|Ri, then S/σ is isomorphic corresponding group GphX|Ri. Detailed proofs of these facts about inverse semigroups are provided in the text [3]. In order to be consistent with most of the literature about inverse semigroups, we abuse the notation of σ and we also denote the natural homomorphism from invhX|Ri to GphX|Ri by σ. An inverse semigroup S is called E-unitary if σ −1 (1) = {e|e2 = e}. S J. B. Stephen [7] introduced the notion of Schützenberger graphs to solve the word problem for inverse semigroups. If M = InvhX|Ri is an inverse semigroup then we may consider the corresponding Cayley graph Γ(M, X). The vertices of this graph are labeled by the elements of M and there exists a directed edge labeled by x ∈ X ∪ X −1 from the vertex labeled by m1 to the vertex labeled by m2 if m2 ≡ m1 x. The Cayley graph Γ(M, X) is not necessarily strongly connected, unless M happens to be a group, therefore there may not be an edge labeled by x−1 from m2 to m1 . The strongly connected components of Γ(M, X) are called the Schützenberger graphs of M . For any word u ∈ (X ∪ X −1 )∗ the strongly connected component of Γ(M, X) that contains a vertex labeled by u is the Schützenberger graph of u and it is denoted by SΓ(M, X, u). The vertices of SΓ(M, X, u) are labeled by the elements of R-class of u, i.e., Ru = {m ∈ M |mm−1 = uu−1 }, because M if x ∈ X ∪ X −1 labels an edge from a vertex labeled m ∈ M to a vertex labeled by mx, then there exists an edge labeled by x−1 from mx to m in Γ(M, X) if and only if mR(mx). There exists a natural graph morphism (not necessarily injective) from a Schützenberger graph of M to the Cayley graph of M . Meakin showed that each Schützenberger graph of M embeds in the Cayley graph of M if and only if M is E-unitary. It is useful to consider Schützenberger automaton (uu−1 , SΓ(M, X, u), u) with initial vertex uu−1 ∈ M , terminal vertex u ∈ M and set of states are the vertices of SΓ(M, X, u). The language accepted by this automaton is L(u) = {v ∈ M |v labels a path from uu−1 to u in SΓ(M, X, u)}. Here u and v are regarded as elements of both (X ∪ X −1 )∗ and M . So, L(u) is a subset of M . The following result of Stephen [7] plays a key role in solving the word problem for inverse semigroups. Theorem 1.1. Let M = InvhX|Ri and let u, v ∈ (X ∪ X −1 )∗ Then (1) L(u) = {v ∈ M |v ≥ u in the natural partial order on M }. (2) u = v⇔L(u) = L(v) ⇔ u ∈ L(v) and v ∈ L(u)⇔(uu−1 , SΓ(M, X, u), M u) and (vv −1 , SΓ(M, X, v), v) are isomorphic as birooted edge-labeled graphs. THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 3 We briefly describe the iterative procedure described by Stephen [7] for building a Schützenberger graph. Let InvhX|Ri be a presentation of an inverse monoid. Given a word u = a1 a2 ...an ∈ (X ∪ X −1 )∗ , the linear graph of u is the birooted inverse word graph (αu , Γu , βu ) consisting of the set of vertices V ((αu , Γu , βu )) = {αu , βu , γ1 , ..., γn−1 } and edges (αu , a1 , γ1 ), (γ1 , a2 , γ2 ), ..., (γn−2 , an−1 , γn−1 ), (γn−1 , an , βu ), together with the corresponding inverse edges. Let (α, Γ, β) be a birooted inverse word graph over X ∪ X −1 . The following operations may be used to obtain a new birooted inverse word graph (α0 , Γ0 , β 0 ): • Determination or folding: Let (α, Γ, β) be a birooted inverse word graph with vertices v, v1 , v2 , with v1 6= v2 , and edges (v, x, v1 ) and (v, x, v2 ) for some x ∈ X ∪ X −1 . Then we obtain a new birooted inverse word graph (α0 , Γ0 , β 0 ) by taking the quotient of (α, Γ, β) by the equivalence relation which identifies the vertices v1 and v2 and the two edges. In other words, edges with the same label coming out of a vertex are folded together to become one edge. • Elementary P-expansion: Let r = s be a relation in R and suppose that r can be read from v1 to v2 in Γ, but s cannot be read from v1 to v2 in Γ. Then we define (α0 , Γ0 , β 0 ) to be the quotient of Γ ∪ (αs , Γs , βs ) by the equivalence relation which identifies vertices v1 and αs and vertices v2 and βs . In other words. we “sew” on a linear graph for s from v1 to v2 to complete the other half of the relation r = s. An inverse word graph is deterministic if no folding can be performed and closed if it is deterministic and no elementary expansion can be performed over a presentation hX|Ri. Note that given a finite inverse word graph it is always possible to produce a determinized form of the graph, because determination reduces the number of vertices. So, the process of determination must stop after finitely many steps, We note also that the process of folding is confluent [7] . If (α1 , Γ1 , β1 ) is obtained from (α, Γ, β) by an elementary P-expansion, and (α2 , Γ2 , β2 ) is the determinized form of (α1 , Γ1 , β1 ), then we write (α, Γ, β) ⇒ (α2 , Γ2 , β2 ) and say that (α2 , Γ2 , β2 ) is obtained from (α, Γ, β) by a Pexpansion. The reflexive and transitive closure of ⇒ is denoted by ⇒∗ . For u ∈ (X ∪ X −1 )∗ , an approximate graph of (uu−1 , SΓ(u), u) is a birooted inverse word graph A = (α, Γ, β) such that u ∈ L[A] and y ≥ u for all y ∈ L[A]. Stephen showed in [7] that the linear graph of u is an approximate graph of (uu−1 , SΓ(u), u). He also proved the following: Theorem 1.2. Let u ∈ (X ∪ X −1 ) and let (α, Γ, β) be an approximate graph of (uu−1 , SΓ(u), u). If (α, Γ, β) ⇒∗ (α0 , Γ0 , β 0 ) and (α0 , Γ0 , β 0 ) is closed , then (α0 , Γ0 , β 0 ) is the Schützenberger graph of u, (uu−1 , SΓ(u), u). 4 MUHAMMAD INAM In [7], Stephen showed that the class of all birooted inverse words graphs over X ∪ X −1 is a co-complete category and that the directed system of all finite P-expansions of a linear graph of u has a direct limit. Since the directed system includes all possible P-expansions, this limit must be closed. Therefore, by 1.2, the Schützenberger graph is the direct limit. Full P- expansion (a generalization of the concept of Pexpansion): A full P-expansion of a birooted inverse word graph (α, Γ, β) is obtained in the following way: • Form the graph (α0 , Γ0 , β 0 ), which is obtained from (α, Γ, β) by performing all possible elementary P-expansions of (α, Γ, β), relative to (α, Γ, β). We emphasize that an elementary P-expansion may introduce a path labeled by one side of relation in R, but we do not perform an elementary P-expansion that could not be done to (α, Γ, β) when we do a full Pexpansion. • Find the determinized form (α1 , Γ1 , β1 ), of (α0 , Γ0 , β 0 ). The birooted inverse word graph (α1 , Γ1 , β1 ) is called the full P-expansion of (α, Γ, β). We denote this relationship by (α, Γ, β) ⇒f (α1 , Γ1 , β1 ). If (αn , Γn , βn ) is obtained from (α, Γ, β) by a sequence of full P-expansions then we denote this by (α, Γ, β) ⇒∗f (αn , Γn , βn ). We now expand the notion of Schützenberger graph to the Schützenberger complexes. The Schützenberger complexes were first defined by Steinberg in [8]. Later in [4], Steven Linblad made a small modification in Steinberg’s definition of Schützenberger complexes. In this paper, we are using Linblad’s definition of Schützenberger complexes. Let M = hX|Ri be an inverse monoid and m ∈ M . The Schüutzenberger complex SC(m) for m ∈ M is defined as follows: (1) The 1-skeleton of SC(m) is the Schützenberger graph SΓ(m). (2) For each relation (r, s) ∈ R and vertex v, if r and s can be read at v, then there is a face with boundary given by the pair of paths labeled by r and s starting from v. In similar manner, Stephen’s approximate graphs can be viewed as approximate complexes by sewing on a face each time an elementary expansion is performed, and identifying faces if a determination results in their entire boundaries being identified. 2. The word problem for Adian semigroups, Adian inverse semigroups and Adian groups The following theorem was first proved by Adian in [1] for finite presentations. Later, it was generalized by Remmers to any Adian presentation, in [6], by using a geometric approach. Theorem 2.1. An Adian semigroup SghX|Ri embeds in the corresponding Adian group GphX|Ri. THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 5 From the embedding in Theorem 2.1, we can derive the fact that every Adian semigroup embeds in an Adian inverse semigroup, as proved in the following proposition. Proposition 1. An Adian semigroup S = SghX|Ri embeds in the Adian inverse semigroup M = InvhX|Ri. Proof. Let θ : S → M be the natural homomorphism and φ : S = hX|Ri → G = GphX|Ri be the natural homomorphism. φ is an injective homomorphism by 2.1. Note that φ = σ ◦ θ. Since φ = σ ◦ θ and φ is injective, then θ must be injective.  Conjecture 2.2. (Adian, 1976) The word problem for Adian semigroups is decidable. Remark 1. The word problem for one relation Adian semigroups is decidable. This is because Magnus [5] proved that the word problem for one relator groups is decidable and by Theorem 2.1, a one relation Adian semigroup embeds in the corresponding one relator Adian group. Proposition 2. The word problem for an Adian semigroup S = SghX|Ri and an Adian group G = GphX|Ri is decidable, if: 1. the Adian inverse semigroup M = InvhX|Ri is E−unitary and 2. the word problem for the Adian inverse semigroup M is decidable. Proof. It immediately follows from (1) that, for any word u ∈ (X ∪ X −1 )∗ , u = 1 if and only if u is an idempotent in M . So, if (2) holds, then we can G check whether u is an idempotent or not by checking the equality of words u = u2 . M If (2) holds, then by Proposition 1, S embeds in M and so the word problem for S is also decidable.  The following theorem proved in [2] establishes the first part of the Proposition 2. Theorem 2.3. (Muhammad Inam, John Meakin, Robert Ruyle) Adian inverse semigroups are E−unitary. In order to solve the word problem for the Adian inverse semigroups we prove the following theorem which enables us to solve the word problem for some classes of Adian inverse semigroups. Theorem 2.4. Let M = InvhX|Ri be a finitely presented Adian inverse semigroup. Then the Schützenberger complex of w is finite for all words w ∈ (X ∪ X −1 )∗ if and only if the Schützenberger complex of w0 is finite for all positive words w0 ∈ X + . 6 MUHAMMAD INAM Remark 2. We observe that if the word problem for a finitely presented Adian Inverse semigroup InvhX|Ri is decidable then the word problem for the corresponding Adian semigroup SghX|Ri and the corresponding Adian group GphX|Ri is also decidable. Because, if the word problem for an Adian inverse semigroup InvhX|Ri is decidable, then the corresponding Adian semigroup embeds into the Adian inverse semigroup by Proposition 1. So the decidability of the word problem for Adian semigroup follows from the decidability of the word problem of Adian inverse semigroup. We know by Theorem 2.3 that Adian inverse semigroups are E-unitary. So, if we want to check whether an element w ∈ (X ∪ X −1 )∗ is equal to the identity element of an Adian group G = GphX|Ri in G, then we just need to check that whether w is an idempotent in the corresponding Adian inverse semigroup InvhX|Ri. Which can be checked immediately if the word problem for the Adian inverse semigroup is decidable. 3. Main Theorem Let InvhX|Ri be an inverse semigroup. Then for any word w ∈ (X ∪ the sequence of approximate graphs {(αn , Γn (w), βn )|n ∈ N} obtained by full P-expansion over the presentation hX|Ri, converges to the Schützenberger graph of w over the presentation hX|Ri. There exist graph homomorphisms, ψn : Γn (w) → Γn+1 (w), such that ψn (αn ) = αn+1 and ψn (βn ) = βn+1 , for all n ∈ N. If we attach to Γn (w) 2-cells corresponding to the relations in the obvious way, we obtain an approximate complex of SC(w). We use the same notation, so that {(αn , Γn (w), βn )|n ∈ N} becomes a sequence of approximate complexes that converges to SC(w). We call a 2-cell to be an n-th generation 2-cell if it occurs in (αn , Γn (w), βn ) \ (ψn−1 (αn−1 ), ψn−1 (Γn−1 (w)), ψn−1 (βn−1 )), for all n ∈ N. The following lemma is due to Steinberg, who gives two slightly deferent proofs, in [8] and [9]. The proof we present here is very similar to the proof in [9]. X −1 )∗ , Lemma 1. Let M = InvhX|Ri be an inverse semigroup and w ∈ (X ∪ X −1 )∗ . Then the Schützenberger complex of w, SC(w), is simply connected. Proof. We use induction on the sequence of finite approximate complexes obtained by full P-expansion of w. Since (α0 , Γ0 (w), β0 ) is just a tree, it is simply connected. We assume that the the finite approximate complex (αk−1 , Γk−1 (w), βk−1 ) is simply connected and show that (αk , Γk (w), βk ) is simply connected. Each time we perform an elementary P-expansion on (αk−1 , Γk−1 (w), βk−1 ) we sew on a relation (u, v) ∈ R along with a 2-cell bounded by the path labeled by uv −1 . We are exactly attaching a simply connected space along a continuous path, the result of which is again simply connected. Thus by induction, all complexes in the sequence of approximate complexes {(αn , Γn (w), βn ) : n ∈ N} are simply connected. It follows that the limit of SC(w) of this sequence is simply connected.  THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 7 Lemma 2. Let InvhX|Ri be an Adian inverse semigroup and w ∈ (X ∪ X −1 )∗ . Then the Schützenberger complex of w contains no directed cycles of 1-cells. Proof. Meakin showed that, if M = InvhX|Ri, is an E-unitary inverse semigroup, then for all words u ∈ (X ∪ X −1 )∗ , the Schützenberger graph of the word u embeds into the Cayley graph of the maximal group homomorphic image of M , GphX|Ri. Likewise, the Schützenberger complex of u embeds in the Cayley complex of the group GphX|Ri. It has been proved in [2] that Adian inverse semigroups are E-unitary. So SC(w) embeds into the Cayley complex of the group GphX|Ri. If SC(w) contains a directed cycle then the Cayley complex of GphX|Ri contains a directed cycle as well. We assume that that this directed cycle is labeled by a word x for some x ∈ X + . Then x = 1 so there exists a G Van Kampen diagram with boundary labeled by x. But this contradicts Lemma 2(ii) of [2] which shows that a Van Kampen diagram over an Adian presentation contains no directed cycles.  The set of all edges (1-cells) of a graph (complex) whose tail vertex lies on a vertex labeled by v is denoted Staro (v) and the set of all edges(1-cells) whose tip lies at a vertex v is denoted by Stari (v). Lemma 3. Let M = InvhX, Ri be an Adian inverse semigroup, w ∈ X + , v a vertex (0-cell) of an approximate complex (αn , Γn (w), βn ) and let a, b ∈ X label two distinct edges of StarΓo n+1 (w) (ψn (v)). Then there exists a path in LGhX|Ri connecting a and b. A dual statement to the above Lemma also holds for StarΓi n+1 (w) (ψn (v)). Proof. We use induction on n to prove the above statement. If n = 0, then there are the following two cases to consider. Case 1. Let a ∈ StarΓo 0 (w) (v) and b ∈ StarΓo 1 (w) (ψ0 (v))\Starψo 0 (Γ0 (w)) (ψ0 (v )). Since w ∈ X + , (α0 , Γ0 (w), β0 ) is just a linear automaton with all of its edges directed towards the vertex labeled by β0 . If there exists an R-word ar (where a ∈ X and r ∈ X ∗ ) that labels a path of Γ0 (w) from the vertex v to a vertex v 0 , we can sew on a new path labeled by the other side of the relation from the vertex v to the vertex v 0 . Note that this is the only way we can add new edges in StarΓo 0 (w) (v), because if we attach a new path to the linear automaton (α0 , Γ0 (w), β0 ) from a vertex u to a vertex u0 and some vertex of this new path gets identified with the vertex v of the linear automaton, then there will be a new edge in StarΓo 0 (w) (v). But this is impossible unless u = v because hX|Ri is an Adian presentation, so the first letters of both the R-words in a relation are different from each other. As a consequence of attaching a new path starting from the vertex v , StarΓo 1 (w) (v) contains more than one element. So, we sew on a path labeled by the other side of the relation bs (where b ∈ X and s ∈ X ∗ ) from the 8 MUHAMMAD INAM vertex v to the vertex v 0 , for some (ar, bs) ∈ R. So, there exists an edge between a and b in LGhX|Ri. Case 2. Suppose a, b ∈ StarΓo 1 (w) (ψ0 (v)) \ Starψo 0 (Γ0 (w)) (ψ0 (v)). Then, since (α0 , Γ0 (w), β0 ) is a linear automaton, there exist R-words cs1 and cs2 (where c ∈ X and s1 , s2 ∈ X ∗ ) labeling two overlapping segments of Γ0 (w) starting from the vertex labeled by v. This allows us to sew on new paths labeled by the other sides of the relations ar1 and br2 (where b ∈ X and r1 , r2 ∈ X ∗ ), respectively, both starting from the vertex labeled by v. Since, (ar1 , cs1 ), (br2 , cs2 ) ∈ R, thus there exists a path between a and b in LGhX|Ri. Now we assume that the above statement is true for n = k and we prove the above statement for n = k + 1. If both the edges labeled by a and b belong to the set StarΓo k (w) (v), then by the induction hypothesis, there exists a path in LGhX|Ri connecting a and b. So, there is nothing to prove in this case. Note that hX|Ri is an Adian presentation and therefore SC(w) contains no positively or negatively labeled directed cycles by Lemma 2. So, if a ∈ StarΓo k (w) (v) and b ∈ StarΓo k+1 (w) (ψk (v)) \ Starψo k (Γk (w)) (ψk (v)), then either there exists a relation (br, as) ∈ R (where a, b ∈ X and r, s ∈ X ∗ ), where as labels a path beginning at vertex v in the birooted inverse word graph (αk , Γk (w), βk ), or it occurs as consequence of sewing on a path labeled by an R-word starting from a vertex v 0 (where the vertex labeled by v 0 lies before the vertex labeled by v on a positively labeled path from αk to βk ), of Γk (w) and then folding edges with the same label and same initial vertices. The latter case is impossible, because if we read an R-word ct (where c ∈ X and t ∈ X ∗ ) starting from the vertex labeled by v 0 in the finite approximate complex (αk , Γk (w), βk ) and we sew on a path labeled by the other side of the same relation, du (where d ∈ X and u ∈ X ∗ ) to obtain (αk0 , Γ0k (w), βk0 ), then all the vertices of LGhX|Ri that are labeled by those letters which also label the edges of the set StarΓo k (w) (v 0 ), are connected by a path in LGhX|Ri, by our induction hypothesis. Since the edge labeled d gets identified with one of the pre-existing edges in the set StarΓo k (w) (v 0 ), therefore there exists a path between c and d in LGhX|Ri. But, we also have (ct, du) ∈ R, i.e., there exists an edge between c and d. So, there exists a closed path in LGhX|Ri. This is a contradiction. Let StarΓo k (w) (v) = {a1 , a2 , ..., am |ai ∈ X for 1 ≤ i ≤ m}, where each ai labels an edge. Then there are the following two cases to consider: Case 1. If a ∈ StarΓo k (w) (v) and b ∈ StarΓo k+1 (w) (ψk (v))\Starψo k (Γk (w)) (ψk ( v)). Then a = ai for some i and there exists an R-word as (where a ∈ X and s ∈ X ∗ ), that labels a path starting from the vertex labeled by v, which allowed us to sew on a path labeled by the other side of the same relation, for some relation of the form (br, as) ∈ R (where b ∈ X and r ∈ X ∗ ). So, there exists an edge between a and b in LGhX|Ri and a is connected to aj THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 9 for all 1 ≤ i 6= j ≤ m by induction hypothesis. Hence, b is connected with each aj for 1 ≤ j ≤ m by a path in LGhX|Ri. Case 2. If a, b ∈ StarΓo k+1 (w) (ψk (v)) \ Starψo k (Γk (w)) (ψk (v)). Then there exist R-words of the form ai s1 and aj s2 (where ai , aj ∈ X and s1 , s2 ∈ X ∗ ) for some i, j ∈ {1, 2, ..., m}, that can be read in the finite approximate complex (αk , Γk (w), βk ), starting from the vertex labeled by v, that allowed us to sew on positively labeled paths labeled by the other sides of relations of the form (ar1 , ai s1 ), (br2 , aj s2 ) ∈ R (where a, b ∈ X and r1 , r2 ∈ X ∗ ). Hence there exist an edge between a and ai and an edge between b and aj in LGhX|Ri. So, if i = j , then there exists a path between a and b in LGhX|Ri and if i 6= j, then ai and aj are distinct edges and ai is connected with aj in the LGhX|Ri by the induction hypothesis. Hence, a and b are connected by a path in LGhX|Ri.  The following Corollary of Lemma 3 is not useful for the proof of the main theorem of this paper. However, it provides some information about the structure of Adian semigroups. Corollary 1. Let S = SghX|Ri be an Adian semigroup. Then S has no idempotent element. Proof. We assume that S contains an idempotent element w for some w ∈ X + . It follows from the Proposition 1 that S embeds in the corresponding inverse semigroup M = InvhX|Ri. So w = w2 . It follows from the Theorem M 1.1 that w2 ∈ L(w). Then there exists a least positive integer n, such that w2 labels a path from the initial vertex αn to the terminal vertex βn of the approximate complex (αn , Γn (w), βn ). We assume that a ∈ X is the prefix letter of w. Then by Lemma 3 there exists a path from a to a in LGhX|Ri. So LGhX|Ri contains a closed a path. This is a contradiction.  The following lemma shows that if InvhX|Ri is an Adian inverse semigroup, then the construction of the Schützenberger complex of a positive word only involves the elementary P-expansion process and no folding at all. Lemma 4. Let M = InvhX, Ri be an Adian inverse semigroup and w ∈ X + . Then no two edges fold together in the construction of the Schützenberger complex of w. Proof. We use induction on n to show that no two edges fold together in the construction of each finite approximate complex of the sequence {(αn , Γn (w), βn )|n ∈ N}. The above statement is true for n = 0, because w ∈ X + , so (α0 , Γ0 (w), β0 ) is just a linear automaton with no two consecutive edges oppositely oriented. Hence, no two edges fold together in the construction of (α0 , Γ0 (w), β0 ). 10 MUHAMMAD INAM We assume that the above statement is true for n = k, i.e., no two edges fold together in the construction of (αk , Γk (w), βk ). We show that the above statement is also true for n = k + 1. . We apply an elementary expansion on (αk , Γk (w), βk ) to obtain (αk0 , Γ0k (w), 0 βk ) and then perform folding in (αk0 , Γ0k (w), βk0 ) to obtain (αk+1 , Γk+1 (w), βk+1 ) in the full P-expansion. So, for any vertex v of (αk , Γk (w), βk ), if we read an R-word labeling a path starting from v to a vertex v 0 , then we sew on a new path labeled by the other side of the same relation from v to v 0 and then we perform folding if possible. In order to prove the above statement we just need to show that no two edges fold together in StarΓo 0 (w) (v). The k case that no two edges fold together in StarΓi 0 (w) (v)) is dual to the previous k case. Let {ai ∈ X|0 ≤ i ≤ l} be the set of labels of edges in StarΓo k (w) (v) and let b0 , b1 , ..., bm be the labels of those edges which are in StarΓo 0 (w) (v) \ k StarΓo k (w) (v). We claim that 1. ai 6= bj for 0 ≤ i ≤ l and 0 ≤ j ≤ m, and 2. bi 6= bj for 0 ≤ i 6= j ≤ m. To establish our first claim, we assume that ai = bj for some 0 ≤ i ≤ l and 0 ≤ j ≤ m. It has already been shown in the proof Lemma 3 that if bj ∈ StarΓo 0 (w) (v) \ StarΓo k (w) (v) then there exists a relation of the form k (bj r, cs) ∈ R (where bj , c ∈ X and r, s ∈ X ∗ ) and cs labels a path from the vertex v to vertex v 0 . It follows that c ∈ {at ∈ X|0 ≤ t ≤ l}. If c = ai , then (bj r, cs) = (ai r, ai s) ∈ R. This contradicts the fact that hX|Ri is an Adian presentation. If c = ah for some h such that 0 ≤ h 6= i ≤ l, then by Lemma 3, ai and ah are connected by a path p in LGhX|Ri. We also have (bj r, cs) = (ah r, ai s) ∈ R. So, there exists an edge e between ai and ah in LGhX|Ri. If p and e represent the same path in LGhX|Ri, then the path labeled by bj r from the vertex v to the vertex v 0 already exists in (αk , Γk (w), βk ), but this contradicts our earlier assumption that the path labeled by bj r from v to v 0 did not exist in (αk , Γk (w), βk ). So, p and e represents two different co-terminal paths. Hence there exists a cycle in LGhX|Ri. This is a contradiction. To establish our second claim, we assume that bi = bj for some 0 ≤ i, j ≤ m with i 6= j. Note that the paths starting from the vertex v with initial edges labeled by bi and bj are consequences of relations of the form (bi ri , ai si ), (bj rj , aj sj ) ∈ R respectively. If ai = aj , then either both of these relations are same or they are different relations. If both of these relations are same then this case reduces to the previous case, which has already been discussed above in the proof of claim 1. If these relations are different from each other, then there exists a cycle in LGhX|Ri, which is a contradiction. So, ai 6= aj . But then by lemma 3, there exists a path connecting ai with aj in LGhX|Ri. Also, there exist edges from ai to bi and aj to bj in LGhX|Ri, corresponding to the relators (bi ri , ai si ) and (bj rj , aj sj ). This implies that there is a cycle in LGhX|Ri, which is a contradiction. THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 11  Proposition 3. Let M = InvhX|Ri be an Adian inverse semigroup and w1 ∈ X + . Then: (i) ψn : (αn , Γn (w1 ), βn ) → (αn+1 , Γn+1 (w1 ), βn+1 ) is an embedding for all n ∈ N. (ii) SC(w1 ) has exactly one source vertex α and exactly one sink vertex β, where (α, Γ(w1 ), β) is the underlying birooted graph of SC(w1 ). (iii) Every directed edge of SC(w1 ) can be extended to a positively labeled directed transversal from α to β. Proof. (i). (i) follows immediately from Lemma 4. Since no foldings occur in the construction of Γn+1 (w) from Γn (w), the images of two distinct vertices of the approximate complex (αn , Γn (w), βn ) remain distinct under the map ψn , for all n ∈ N. (ii). (ii) follows from the fact that ψn (αn ) = αn+1 = α and ψn (βn ) = βn+1 = β for all n ∈ N. We sew on a new positively labeled path (labeled by one side of a relation) to an approximate complex (α, Γn (w), β) only when we read a positively labeled segment of a path from α to β, labeled by the other side of the same relation. Each 2-cell is two sided and no folding occurs in the construction of (α, Γn+1 (w), β) from (αn , Γn (w), βn ). So α and β remain the source and sink vertices of the approximate complex (α, Γn (w), β) for all n ∈ N. Furthermore, α and β were distinct vertices of (α, Γ0 (w1 ), β) and no two vertices get identified with each other in the construction of SC(w1 ), so α and β remain distinct vertices of SC(w1 ). (iii). Let e0 be a directed edge of SC(w1 ). The complex SC(w1 ) contains only one source vertex α and one sink vertex β. So, the initial vertex of e0 is either α or it is a terminal vertex of another edge e−1 . If the initial vertex of e0 is the terminal vertex of an edge e−1 then the initial vertex vertex of e−1 is either α or the terminal vertex of an edge e−2 . Since there are no directed cycles in SC(w1 ) (by Lemma 2), we can find a sequence of edges e−n , e−(n−1) , ..., e−1 , e0 that constitutes a positively labeled directed path from α to the terminal vertex of edge e0 . Similarly, if the terminal vertex of e0 is not β then we can find an edge e1 whose initial vertex is the terminal vertex of e0 . Continuing in this way, we can find a sequence of edges e0 , e1 , ..., em that constitutes a positively labeled directed path from the edge e0 to β. Hence, the sequence of edges e−n , e−(n−1) , ..., e0 , ..., em constitutes a positively labeled directed transversal from α to β.  The word problem for S = SghX|Ri is the question of whether there is an algorithm which given any two words u, v ∈ X + , will determine whether u = v in S. 12 MUHAMMAD INAM For any two words u, v ∈ X + , u = v if and only if there exists a transition S sequence from u to v. u ≡ w0 → w1 → ... → wn ≡ v; for some n ≥ 0, where wi−1 → wi represents that wi is obtained from wi−1 by replacing one side of a relation r (that happens to be a subword of wi−1 ) with the other side s of the same relation, for some (r, s) ∈ R. The above transition sequence is called a regular derivation sequence of length n for the pair (u, v) over the presentation SghX|Ri. A semigroup diagram or S-diagram over a semigroup presentation SghX|Ri for a pair of positive words (u, v) is a finite, planar cell complex D ⊆ R2 , that satisfies the following properties: • The complex D is connected and simply connected. • Each edge (1-cell) is directed and labeled by a letter of the alphabet X. • Each region (2-cell) of D is labeled by the word rs−1 for some defining relation (r, s) ∈ R. • There is a distinguished vertex α on the boundary of D such that the boundary of D starting at α is labeled by the word uv −1 . α is a source in D (i.e. there is no edge in D with terminal vertex α). • There are no interior sources or sinks in D. In [6], Remmers proved an analogue of Van Kampen’s Lemma for semigroups to address the word problem for semigroups. Theorem 3.1. Let S = SghX|Ri be a semigroup and u, v ∈ X + . Then there exists a regular derivation sequence of length n for the pair (u, v) over the presentation SghX|Ri if and only if there is an S-diagram over the presentation SghX|Ri for the pair (u, v) having exactly n regions. Proposition 4. Let M = InvhX|Ri be an Adian inverse semigroup and w1 , w2 ∈ X + such that w1 ≤ w2 . Then: M (i) There exists a (planar) S-diagram corresponding to the pair of words (w1 , w2 ) that embeds in SC(w1 ). (ii) w1 = w2 . M Proof. (ii). If w1 ≤ w2 , then w2 ∈ L(α, Γ(w1 ), β) (Stephen, [7]). So, w2 M labels a directed transversal from the vertex α to the vertex β. Since the transversals labeled by w1 and w2 are co-terminal and SC(w1 ) is simply connected by Lemma 1, then the closed path labeled by the word w1 w2−1 is filled with finitely many 2-cells. Every 2-cell is two sided because hX|Ri is an Adian presentation. So, we can obtain a regular derivation sequence from the word w1 to the word w2 from the complex SC(w1 ), over the presentation hX|Ri in the following way. THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 13 Geometrically, we push the transversal labeled by w1 across all first generation 2-cells that were contained in the closed path labeled w1 w2−1 to obtain a new transversal labeled by u1 ∈ X + . Combinatorially, we have replaced some of the non overlapping R-words that were subwords of the word w1 by the other side of the same relations. Then we push the transversal labeled by u1 across all generation 2-cells that were contained in the closed path labeled by w1 w2−1 to obtain a new transversal labeled by u2 ∈ X + . Again, we have just replaced some of the non overlapping R-words that were subwords of the word u1 . This process eventually terminates because the closed path labeled by w1 w2−1 contains only finitely many 2-cells. So, we obtain a regular derivation w1 → u1 → u2 → ... → w2 over the presentation hX|Ri. Hence there exists an S-diagram S corresponding to this derivation sequence. Since no two edges fold together in SC(w1 ), therefore S embeds in SC(w1 ). (ii). This follows immediately form (i) and the Theorem 3.1.  Lemma 5. Let M = InvhX|Ri be an Adian inverse semigroup, let w ∈ (X ∪ X −1 )∗ and let w1 , w2 ∈ X + label two co-terminal paths in SC(w). Then there exists an S-diagram corresponding to the pair of words (w1 , w2 ) that embeds in SC(w). Proof. Since M is E-unitary, SC(w) embeds into the Cayley complex of the group G = GphX|Ri. So, the word w1 w2−1 labels a closed path in the Cayley complex. Hence w1 w2−1 = 1. So, w1 = w2 . It follows from Theorem G G 2.1 that w1 = w2 in SghX|Ri. So, there exists a regular derivation sequence D from w1 to w2 over the presentation S = SghX|Ri. The semigroup S embeds into the inverse semigroup M , by Proposition 1. So the regular derivation sequence sequence D also holds in M . The complex SC(w) is closed under elementary P-expansion and folding and therefore all the 2cells corresponding to this regular derivation sequence D already exist in SC(w) between the paths labeled by w1 and w2 . Hence the S-diagram corresponding to the regular derivation sequence D embeds in SC(w).  In general for any finite inverse semigroup presentation (X, R), and for any word w ∈ (X ∪ X −1 )∗ , there are natural birooted graph morphisms for the sequence of approximate graphs given by Stephen’s procedure for approximating the birooted Schützenberger graph (α, Γ(w), β). (α1 , Γ1 (w), β1 ) → · · · → (αn , Γn (w), βn ) → (αn+1 , Γn+1 (w), βn+1 ) → · · · → (α, Γ(w), β) It was established in Proposition 4 that if M = InvhX|Ri is an Adian inverse semigroup and w is a positive word, that is w ∈ X + , then all of the maps in the above sequence are actually embeddings. We may abuse the notation slightly in this case and for each approximate graph Γn (w) we 14 MUHAMMAD INAM denote the initial and terminal vertices simply as α and β. (α, Γ1 (w), β) ,→ · · · ,→ (α, Γn (w), β) ,→ (α, Γn+1 (w), β) ,→ · · · ,→ (α, Γ(w), β) Recall that for a positive word w ∈ X + , a transversal of an approximate Schützenberger complex (α, Γk (w), β) is defined to be any positively labeled path in Γk (w) from α to β. Before proving the main theorem of this section we introduce the following definition of n-th generation transversal of a Schützenberger complex. Definiton 1. An n-th generation transversal of the Schützenberger complex (α, Γ(w), β) is a positively labeled path from α to β that can be read in the approximate complex (α, Γn (w), β) but cannot be read in (α, Γn−1 (w), β), for some n ∈ N. When studying certain problems involving an inverse monoid given by a finite presentation (X, R), for example when considering the word problem for M = InvhX|Ri, it is natural to first ask whether it might happen to be the case that all of the Schützenberger graphs of M are finite. The following theorem, our main theorem for this paper, shows that this question can be reduced to the question of whether or not all Schützenberger graphs of positive words are finite. Theorem 3.2. Let M = InvhX|Ri be a finitely presented Adian inverse semigroup. Then the Schützenberger complex of w is finite for all words w ∈ (X ∪ X −1 )∗ if and only if the Schützenberger complex of w0 is finite for all positive words w0 ∈ X + . Idea of the proof: We assume that the Schützenberger graph of every positive word is finite and we let w be an arbitrary word, w ∈ (X ∪ X −1 )∗ . We will use induction on the number of edges in the Munn tree M T (w) to prove that, the Schützenberger complex of w is finite. The essential part of the proof involves realizing SC(w) as the limit of a sequence of finite complexes, via the procedure of Stephen’s P-expansion. We begin with a finite inverse graph (complex) S that is closed under P-expansions relative to hX|Ri. To the complex S we attach a single edge e that is labeled by some letter, say a ∈ X. The resulting complex S1 = S ∨ {e} will in general not be closed under P-expansions. In a process similar to Stephen’s full Pexpansion construction, we define a sequence of finite complexes S1 , S2 , . . . that converges in the limit to SC(w). Our theorem will be proved if we can show that this limit is in fact a finite complex. Equivalently, we must prove that the sequence S1 , S2 , . . . stabilizes after finitely many steps at some Sk . Proof. We assume that SC(w0 ) is finite for all positive words w0 ∈ X + ,. Let w be an arbitrary word, w ∈ (X ∪ X −1 )∗ . We will show that SC(w) is finite, by applying induction on the number of edges in M T (w). For the base of our induction, we suppose that M T (w) consists of only one edge, labeled say by a ∈ X. Then by our assumption about positive words, SC(w) = SC(a) is finite. THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 15 For our induction hypothesis, we assume that SC(w0 ) is finite for all words w0 ∈ (X ∪ X −1 )∗ , whose Munn tree consist of k edges. Let w ∈ (X ∪ X −1 )∗ be a word such that M T (w) consists of k + 1 edges. We will show that SC(w) is finite. Let α be an extremal vertex of M T (w) (i.e., a leaf of the tree M T (w)), and let e be the edge of M T (w) that connects α to the remainder of the M T (w). Assume that e is a positively labeled edge with initial vertex α and terminal vertex β. The case when e is negatively labeled is dual. The sub-tree obtained by removing the edge e consists of k edges; we denote this sub-tree by T for our future reference. Note that the tree T is in fact the Munn tree of some word z ∈ (X ∪ X −1 )∗ . That is, T = M T (z) and there are k edges in M T (z). By the induction hypothesis, the Schützenberger complex S generated by T , i.e., S = SC(z), is a finite complex. There exists a graph morphism φ : T → S, and so we may regard the vertex β of T as a vertex in the complex S = SC(z). We reattach the edge e to the vertex β of S and denote the resulting finite complex by S1 , (see Figure 1.) The finite complex S = SC(z) was obtained from the subtree T by sewing on relations from R, and since we reattached the edge e in S1 = S ∨ {e}, then naturally S1 may be regarded as an approximation to SC(w). While the Schützenberger complex S = SC(z) is closed under P-expansion, the complex S1 is not necessarily closed under P-expansion. In particular, it is possible that there may be one or more relations (r, s) ∈ R such that r labels a path in S1 that begins at vertex α and the other side of the relation, s, is not read in S1 at α. It is clear that the closure of S1 under P-expansion over hX|Ri is the Schützenberger complex SC(w). α e β S Figure 1. S1 = S ∨ {e} If the edge e (labeled say, by a) gets immediately identified by folding to an a-labeled edge of the finite complex S, then we are done, because in that case we would have SC(w) = S, which was assumed to be a finite complex. So, we assume that the edge e does not get immediately identified 16 MUHAMMAD INAM by folding with any of the edges in the finite complex S. We extend the edge e to all possible maximal positively labeled paths in S1 . There are only finitely many maximal positively labeled paths in S1 with initial edge e because S1 is a finite complex that has no positively labeled cycles (by Lemma 2). We assume that these paths are labeled by w1 , w2 , ..., wn , where wi ∈ X + for 1 ≤ i ≤ n. Each such wi labels a path from α to some vertex βi of S. (See Figure 2). α e β β1 βn β2 Figure 2. S1 with maximal positively labeled paths from α to βi . In order to complete S1 under elementary P-expansion and folding, we first attach SC(wi ) to the path labeled by wi in S1 , for all i, and denote the resulting finite complex by S10 . (See Figure 3.) Each complex SC(wi ) for 1 ≤ i ≤ n, is finite since each wi ∈ X + . Thus we obtain S10 by attaching to S1 finitely many complexes, each of which is finite, and so S10 is finite. The complex S10 is not necessarily determinized. That is, as a consequence of attaching the complexes SC(wi ), there may now be vertices along the paths in S10 labeled by wi at which there exist two or more edges labeled by the same letter and so we must perform foldings to obtain a determinized graph (complex). We denote the determinized form of S10 by S2 . Since S10 is finite, then so is its determinized quotient S2 . Note however, as consequence of folding S10 to S2 , that the complex S2 may not be closed under P-expansions. And so, we iterate the procedure. In general, the complex Sk may not be closed under P-expansion. That is, there may be a relation (r, s) ∈ R and a path p labeled by r between some two vertices v and v 0 in Sk such that the other side of the relation, s, does not label a path in Sk between v and v 0 . We refer to such a path p in Sk as an unsaturated path. Note that in Stephen’s procedure we would at this stage simply attach a path labeled by s between v and v 0 as one step in the P-expansion. In our THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 17 α e SC(wn) SC(w1) SC(w2) β1 βn β2 Figure 3. S10 is obtained from S1 by sewing on SC(wi ) to each maximal positive path, with label wi , that begins at α in S1 . setting it turns out, however, that we can “speed up” Stephen’s procedure. We will show that every such unsaturated path p can be extended to a positively labeled path that begins at the vertex α. Such a path can itself be extended to a maximal positively labeled path that begins at α. So, instead of merely attaching a single cell along the unsaturated path p that is labeled by r, we instead read the label, say wj , of the maximal positively labeled path beginning at α that contains p as a subpath and we attach the (finite) complex SC(wj ) along this path. Thus in this one step we are attaching not only the one cell along the unsaturated path p, but also are attaching all cells that would arise from P-expansions on the maximal path labeled by wj . Our iterative procedure for constructing a sequence {Sn } of complexes can be summarized as follows. Suppose that Sk has been constructed. We first look for all positive words wi that label a maximal positive path in Sk that starts at the vertex α and that does not label a path starting at α in 0 Sk−1 . We obtain Sk0 by attaching SC(wi ) to each such maximal positive path in Sk . Then we obtain Sk+1 by determinizing Sk0 . Thus we obtain a sequence of finite complexes {Sn } that has SC(w) as its limit. S1 → · · · → Sk −−−−−−−−−−−−−−−−→ Sk0 −−−−−−−−−−−−−→ Sk+1 → · · · For each maximal positive path with label wi that starts at α and does not ex0 ist in Sk−1 , attach SC(wi ). Fold. (Determinize.) 18 MUHAMMAD INAM A question is whether we ever reach a graph Sk in the procedure such that all maximal positive paths that can be read at α in Sk can already be read 0 in Sk−1 at α. If this happens, then the sequence of complexes will stabilize at Sk . Equivalently, we can ask whether we ever reach a complex Sk such that every maximal positive path in Sk is closed under P-expansion. (We say that a maximal positive path p is closed under P-expansion if for every relation (r, s) ∈ R, and for every subpath of p labeled by r between vertices v and v 0 , then the other side of the relation, s, already labels a path between v and v 0 in Sk ). To answer this question, we analyze how the process of folding Sk0 to Sk+1 affects positively labeled paths that already exist in Sk0 and how new positively labeled paths may be created in Sk+1 as a consequence of the folding process. We first note, for example, that when we fold S10 to S2 that each of the attached complexes SC(wi ) that we attached to S1 will be embedded in S2 after the folding process. This follows from a fact (see Stephen, [7]) about E-unitary semigroups: If M is E-unitary and w is any word, and a word wi can be read along some path of the Schützenberger graph Γ(w), then the entire Schützenberger graph Γ(wi ) will occur as an embedded subgraph of Γ(w) along that path. Since the theorem we are proving assumes that the semigroup M has an Adian presentation, we know from the main theorem of [2] that M is E-unitary. Thus, each SC(wi ) embeds in S2 . Likewise, the original graph S is actually the Schützenberger graph Γ(z), where the word z labels a path in S1 , and so we know that the original graph S also remains embedded as a subgraph of S2 . In other words, no two vertices of any one graph SC(wi ) will become identified with each other and no two vertices of the original graph S will become identified with each other in the folding process that takes S10 to S2 . Two vertices of S10 will become identified in the folding process only if one of the vertices belongs to the original S and the other vertex belongs to one of the attached SC(wi ), or if the two vertices belong to SC(wi ) and SC(wj ), with i 6= j. Everything that we just said about the process of folding S10 to S2 holds as well for the process of folding Sk0 to Sk+1 . The original complex S and each complex SC(wi ), attached at any step of the iteration, will be embedded as subcomplexes of Sk+1 . The interest is in what new paths may be formed as a result of folding Sk0 to Sk+1 . Claim: Suppose that SC(wi ) is one of the complexes that was attached to Sk to form Sk0 . Suppose, in the process of folding Sk0 → Sk+1 , that a vertex γ of SC(wi ) gets identified, as a consequence of folding, with a vertex γ 0 of the original complex S. Then we claim that every positively labeled path in SC(wi ) from γ to βi will get identified by the folding process with a path in S from the vertex γ 0 to βi . Thus, every maximal positive path p that begins at the vertex α in Sk+1 and cannot be read beginning at the vertex α in Sk0 will factor uniquely as p = p1 p2 , where p1 is a path in some SC(wi )s and p2 is a path in the original complex S. THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 19 Proof of Claim: Since γ gets identified with γ 0 through folding, we know (Stephen, [7]) that in Sk0 there is a path from γ to γ 0 labeled by a Dyck word d that we may assume is of the form d = ss−1 , where s ∈ (X ∪ X −1 )∗ . Further, we may assume that s is a reduced word. There must be a vertex δ that lies on the path labeled by wi (the intersection of SC(wi ) and S in Sk0 ) so that s labels a path in SC(wi ) from γ to δ and s also labels a path in S from γ 0 to δ. (See Figures 5 & 6.) To prove the above claim we assume that r2 labels an arbitrary positive path in SC(wi ) from γ to βi . We need to prove that this path gets identified by folding with a path in S that is labeled by r2 from γ 0 to βi . The path in SC(wi ) that is labeled by r2 can be extended to an nth generation transversal t of SC(wi ), for some n. To complete our proof, we apply induction on the generation number, n, of the transversal of t of SC(wi ). Suppose that the path labeled by r2 lies on a 1st generation transversal t of SC(wi ). We assume that t ≡ r1 r2 where r1 labels the sub-path of t from α to γ and r2 labels the sub-path of t from γ to βi . Since γ lies on a 1st generation transversal of SC(wi ), we conclude that in the Dyck word path ss−1 that goes from γ to δ to γ 0 , it must be that either s is a positive word or s−1 is a positive word. We examine the following three cases. Case 1. (See Figure 4.) Suppose that the vertex δ is actually βi , the terminal vertex of the path labeled by wi , and suppose the Dyck word labeling the path from γ to γ 0 is the word r2 r2−1 . In this case, the word r2 labels the subpath of the transversal t of SC(wi ) from γ to βi and the word r2 also labels a path from γ 0 to βi . These two paths, both labeled by r2 , which meet at the vertex βi , will fold together so that the path labeled by r2 in SC(wi ) from γ to βi gets identified with the path in S from γ 0 to βi . So in this case the statement of the claim obviously holds. Case 2. (See Figure 5.) We assume that the Dyck word path from γ to δ to γ 0 is labeled by r3 r3−1 , where r3 is a positive word, r3 ∈ X + . So, the path in SC(wi ) from γ to δ is labeled by r3 and the path in S from γ 0 to δ is labeled by r3 . The oppositely oriented paths labeled by r3 , which meet at δ, become identified with each other through folding. We assume that the sub-path from δ to βi , of the maximal path labeled by wi , is labeled by r4 ∈ X + . Obviously, the path labeled by r4 from δ to βi is in SC(wi ) and the path labeled by r3 from γ to δ is also in SC(wi ). Hence the path labeled by r3 r4 from γ to βi is in SC(wi ). The positive words r2 and r3 r4 label two coterminal paths in SC(wi ). So, by Lemma 5 an S-diagram corresponding to the pair of words (r2 , r3 r4 ) embeds in SC(wi ). This S-diagram also embeds in S, because S contains a path labeled by one side of this S-diagram, (namely, the path from γ 0 to βi in S that is labeled by r3 r4 ), and S is closed under elementary P-expansion. So, the two S-diagrams corresponding to the pair of words (r2 , r3 r4 ) get identified with each other and our claim holds in this case. That is, the path in SC(wi ) labeled by r2 from γ to βi gets identified with a path in S from γ 0 to βi . 20 MUHAMMAD INAM α e SC(wi) r1 β γ' γ r2 r2 βi Figure 4. Case 1 for the first generation transversals of the proof of the Claim α e SC(wi) r1 r3 β δ r3 γ' γ r4 r2 βi Figure 5. Case 2 for the first generation transversals of the proof of the Claim Case 3. (See Figure 6.) We assume that the Dyck word path from γ to δ to γ 0 is labeled by r3−1 r3 , where r3 is a positive word. So, the path in SC(wi ) from δ to γ is labeled by r3 and the path in S from δ to γ 0 is labeled by r3 . The oppositely oriented paths labeled by r3 , which meet at δ, become identified with each other through folding. Now we have the path in SC(wi ) labeled by r3 r2 from δ to βi in SC(wi ). Again, we let the positive word r4 be the label of the sub-path of the maximal path labeled by wi from δ to THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 21 βi . The words r3 r2 and r4 label two co-terminal paths in SC(wi ). So, by Lemma 5 an S-diagram corresponding to the pair of words (r3 r2 , r4 ) embeds in SC(wi ). This S-diagram also embeds in S, because S contains a path labeled by one side of this S-diagram (namely, the path labeled by r4 ) and S is closed under elementary P-expansion. Hence these two S-diagrams get identified with each other through folding. In particular, the path in SC(wi ) labeled by r2 from γ to βi gets identified with a path in S from γ 0 to βi , and so our claim follows in this case as well. This concludes the base case of the inductive proof of the claim. α e SC(wi) r1 r3 β δ r3 γ' γ r4 r2 βi Figure 6. Case 3 for the first generation transversals of the proof of the Claim We assume that our claim is true for all paths from γ to βi that lie along a (n − 1)-st generation transversal of SC(wi ). We prove that our claim is true for all n-th generation transversals of SC(wi ) as well. Suppose that an arbitrary path from γ to βi is labeled by the positive word s2 and suppose that this path extends to an n-th generation transversal t of SC(wi ) . Let t ≡ s1 s2 where s1 labels the sub-path of t from α to γ and s2 labels the sub-path of t from γ to βi . Again, since the vertex γ in SC(wi ) gets identified with with the vertex γ 0 in S through folding, we know that there is a Dyck word ss−1 that labels a path from γ to γ 0 . Since γ lies on an n-th generation transversal of SC(wi ), the Dyck word path ss−1 must pass through some vertex δ of SC(wi ) that lies on an (n − 1)-st generation transversal of SC(wi ). We examine the following three cases. Case 1. Suppose the vertex δ is the vertex βi and we read the Dyck word 0 0 s2 s−1 2 in Sk from γ to γ . Then we fold the oppositely oriented paths labeled by s2 . So in this case it is obvious that that path in SC(wi ) labeled by s2 gets identified with the path in S from γ 0 to βi . 22 MUHAMMAD INAM Case 2. (See Figure 7.) Suppose q denotes an (n − 1)-st generation transversal of SC(wi ) and the vertex δ lies on the transversal q. Assume also that δ has already been identified with a vertex of S as a consequence of folding along the Dyck word ss−1 . In this case (Case 2), we assume that the portion of the Dyck word path from γ to δ is labeled by a positive word s3 ∈ X + . Since δ has already been folded and identified with a vertex of 0 S, we have the Dyck word s3 s−1 3 labeling a path from γ to δ to γ in the partially folded Sk0 . The path in SC(wi ) labeled by s3 from γ to δ can be extended along the (n − 1)-st generation transversal q to the vertex βi . We assume that this path is labeled by s3 s4 ∈ X + , where s4 ∈ X + labels the sub-path of q from δ to βi . By the induction hypothesis the sub-path of q from δ to βi gets identified with a path in S. Thus we have a path in SC(wi ) labeled by s3 s4 from γ to βi and we also have a path in S labeled by s3 s4 from γ 0 to βi . The positive words s2 and s3 s4 label two co-terminal paths in SC(wi ). So by Lemma 5 an S-diagram corresponding to the pair of words (s2 , s3 s4 ) embeds in SC(wi ). This S-diagram also embeds in S, because S contains the path labeled by one side of the S-diagram (namely, s3 s4 ) and S is closed under elementary P-expansion. So the two S-diagrams corresponding to the pair of words (s2 , s3 s4 ) get identified with each other. Hence, the path in SC(wi ) labeled by s2 gets identified with a path in S, and the claim holds in this case. α e β s1 SC(wi) γ s3 s3 γ' δ s2 s4 βi Figure 7. Case 2 for the n-th generation transversal of the proof of the claim Case 3. (See Figure 8.) Suppose q denotes an (n − 1)-st generation transversal of SC(wi ) and δ is a vertex on the transversal q that has already been identified with a vertex of S. We also assume that there exists a path, labeled by s3 ∈ X + , from δ to γ such that we can read the Dyck word s−1 3 s3 THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 23 from γ to δ to γ 0 in Sk0 . We extend the path labeled by s3 from δ to γ along the n-th generation transversal t to a positively labeled path (labeled by s3 s2 ∈ X + ) from δ to βi . By the induction hypothesis the sub-path of q from the vertex δ to the vertex βi (say, labeled by s4 ∈ X + ) is in S. Hence s3 s2 and s4 label two co-terminal paths in SC(wi ) and the path labeled by s4 is also in S. By Lemma 5 an S-diagram corresponding to the pair of words (s3 s2 , s4 ) embeds in SC(wi ). This S-diagram also embeds in S, because S contains one side of this S-diagram and S is closed under elementary Pexpansion. Thus the two S-diagrams corresponding to the pair of words (s3 s2 , s4 ) get identified with each other and the claim holds in this case as well. This completes the inductive step in the proof of the claim. α e β s1 s3 SC(wi) s3 δ γ' γ s2 s4 βi Figure 8. Case 3 for the n-th generation transversal of the proof of the claim  We perform all possible foldings mentioned in the proof of above claim in S10 and denote the resulting complex by S2 . In this folding process edges of the complex S are folded with the edges of SC(wi ) for some i. This folding process may create new maximal positively labeled paths starting from α that did not exist in S10 . For example, S10 may contain a path labeled by r3−1 r5 for some r5 ∈ X + starting from the vertex βi as shown in the Figure 9. After folding the path labeled by r3 r3−1 , we can read a new positively labeled path starting from α and labeled by r1 r5 which could not be read in S10 . It is also possible that S10 may contain a path labeled by r4 r5 for some r5 ∈ X + starting from the vertex δ as shown in the following diagram. After folding the path labeled by r4−1 r4 , we can read a new positively labeled path starting from α and labeled by r1 r5 which could not be read in S10 . 24 MUHAMMAD INAM α e SC(wi) r1 α e β SC(wi) r1 fold γ' γ γ r5 r3 r2 β r3 r3 r2 r5 βi βi Figure 9. An example showing the construction of new maximal positively labeled paths starting from α as a consequence of folding of edges in S10 α e SC(wi) r1 r4 α e β SC(wi) r1 δ r4 r4 γ' γ w i' r2 fold β δ γ w i' r5 r2 r5 βi βi Figure 10. An example showing the construction of new maximal positively labeled paths starting from α as a consequence of folding of edges in S10 Now we consider all those maximal positively labeled paths in S2 that start from α and did not exist in S10 . Since there are only finitely many such paths, we assume that these paths are labeled by u1 , u2 , ..., um and for each i ∈ {1, 2, ..., m} the path labeled by ui terminates at the vertex γi . Note that each maximal path labeled by ui must terminate at a vertex γi that is in the original complex S. Then SC(ui ) for each i is finite by hypothesis. So, we attach SC(ui ) at the path labeled by ui for all i and denote the resulting complex by S20 . THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 25 If the paths labeled by ui and uj for some i 6= j are co-terminal, then by the same argument as above SC(ui ) and SC(uj ) are isomorphic to each other as edge labeled graphs. So, they get identified with each other. If for some i, SC(ui ) contains a transversal labeled by t ∈ X + such that θ(6= γi ) is the first vertex of u that gets identified with a vertex in S2 , then the entire positively labeled sub-path of t from θ to γi gets identified with a positively labeled path in S2 . This can be verified by using the same argument as above. We perform all possible foldings in S20 and denote the resulting complex by S3 . In this folding process edges of the complex S are folded with the edges of SC(ui ) for some i.This folding process may create new maximal positively labeled paths starting from α that did not exist in S20 . So, we repeat this entire process again. We keep repeating this process of sewing finite Schützenberger complexes of positive words and folding unless we reach to the point where we can not create new maximal positively labeled paths starting from α and terminating at a vertex in the complex S. This process of expansion and folding eventually terminates, because hX|Ri is a finite presentation, therefore an edge labeled by a letter can only be incident with finitely many 2-cells and there are only finitely many edges in S as S is a finite complex. By construction, the resulting complex is deterministic and closed with respect to elementary P-expansions, so it is the Schützenberger complex of w. So, the Schützenberger complex of w is finite.  The following Corollary follows immediately from the Theorem 3.2, because the Schützenberger complex, SC(w), for any ∈ (X ∪ X −1 )∗ , is finite over a presentation (X, R) if and only if the underlying graph (1-skeleton) of SC(w) is finite over the same presentation. Corollary 2. Let M = InvhX|Ri be a finitely presented Adian inverse semigroup. Then the Schützenberger graph of w is finite for all words w ∈ (X ∪ X −1 )∗ if and only if the Schützenberger graph of w0 is finite for all positive words w0 ∈ X + . 4. Some applications of Theorem 3.2 4.1. The word problem for a sub-family of Adian inverse semigroups that satisfy condition (?). Definiton 2. We say that a positive presentation hX|Ri satisfies condition (?), if it satisfies the following two conditions: (1) No proper prefix of an R-word is a suffix of itself or any other Rword. (2) No proper suffix of an R-word is a prefix of itself or any other Rword. 26 MUHAMMAD INAM If hX|Ri is a finite Adian presentation that satisfies condition (?), then the set of relations R consists of two types of relations. First, those relations which are of the form (u, xvy), where u and v are R-words and x, y ∈ X + . Second, those relations (u, v) ∈ R where neither u nor v contains an Rword as a proper sub-word. We construct a directed graph corresponding to an Adian presentation that satisfies condition (?) as follows. We call this graph the bi-sided graph of the presentation hX|Ri. The bi-sided graph of a positive presentation is defined as follows. Definiton 3. The bi-sided graph of the presentation hX|Ri is a finite, directed, edge-labeled graph, denoted by BS(X, R) satisfying • The vertex set of BS(X, R) is the set of all R-words. • To define the edge set of BS(X, R), let u, v be two R-words (where it may happen that u and v are the same R-word). There is a directed edge from the vertex u to the vertex v if any of the following three conditions holds: (1) (u, xvy) ∈ R, for some x, y ∈ X + . In this case, the directed edge from u to v is labeled by the ordered pair (x, y). (2) u ≡ xvy, for some x, y ∈ X + and u and v are distinct R-words. In this case, the edge from u to v is labeled by the ordered pair (x, y). (3) If (u, v) ∈ R is such that neither u nor v contains any R-word as a proper sub-word, then there is an edge in the bi-sided graph between u and v, pointing in both directions. This edge is labeled by (ε, ε), where ε denotes the empty word. In general, the bi-sided graph BS(X, R) of an Adian presentation may contain closed paths. Example 1. The bi-sided graph of the Adian presentation ha, b|aba = bi contains a directed closed path (cycle), namely the loop consisting of the single edge labeled by (a, a) from the vertex b to itself. (See Figure 11.) The presentation ha, b|aba = bi does not satisfy condition (?) either, because the R-word aba has the letter a as a prefix and as a suffix. Ed g e o f t y p e 1 Ed g e o f t y p e 2 (a,a) (a ,a ) b aba Figure 11. The bi-sided graph of ha, b|aba = bi contains a cycle, (the loop at vertex b). THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 27 Example 2. The bi-sided graph of the Adian presentation ha, b, c, d, e, f, g, h, i, j, k|a = f bg, a = jck, b = hci, c = dei contains an undirected closed path. (See figure 12.) The presentation ha, b, c, d, e, f, g, h, i, j, k|a = f bg, a = jck, b = hci, c = dei satisfies condition (?). Edge of type 2 Edges of type 1 (f, g) b (f, g) fbg (h, i) (j, k) c (ε, ε) a de (h, i) (j, k) hci jck Edge of type 3 Edges of type 2 Figure 12. The bi-sided graph of ha, b, c, d, e, f, g, h, i, j, k|a = f bg, a = jck, b = hci, c = dei Example 3. The bi-sided graph of the Adian presentation ha, b, c, d, e, f, g, h, i, j, k, l, m|a = f cg, b = hci, c = de, l = jm2 ki is a forest. (See figure 13). The presentation ha, b, c, d, e, f, g, h, i, j, k, l, m|a = f cg, b = hci, c = de, l = jm2 ki satisfies condition (?). Remark 3. For the remainder of this subsection we consider presentations hX|Ri such that (1) hX|Ri is an Adian presentation. (2) hX|Ri satisfies condition (?). (3) The bi-sided graph BS(X, R) is cycle-free. That is, there are no closed paths, directed or un-directed, in BS(X, R). In other words, BS(X, R) is a forest. Note that if the bi-sided graph of a presentation hX|Ri is a forest, then every R-word labels a vertex of a connected component in BS(X, R) that is a tree. For an R-word u, we let Tu denote the unrooted tree that contains the vertex u and we refer to Tu as the bi-sided tree of u. If u and v label two different vertices of the same bi-sided tree, then Tu and Tv denote the same unrooted, bi-sided tree. Definiton 4. Let u be an R-word. We say that a vertex labeled by v of Tu is accessible from u if there exists a path labeled by v in SC(u). 28 MUHAMMAD INAM Edges of type 2 Edges of type 3 fcg hci (h, i) (f, g) c (ε, ε) (ε, ε) de (f, g) a l jm 2k (h, i) b Edges of type 1 Figure 13. The bi-sided graph of ha, b, c, d, e, f, g, h, i, j, k, l, m|a = f cg, b = hci, c = de, l = jm2 ki Note that if there exists a directed edge from v1 to v2 in the bi-sided tree of u and v1 is accessible from u then v2 is also accessible from u. Because if v1 is accessible from u, then there exists a path labeled by v1 in SC(u). Since there exists a directed edge from v1 to v2 in the bi-sided tree of u therefore either (v1 , xv2 y) ∈ R or v1 ≡ xv2 y for some x, y ∈ X + . In either case there exists a path labeled by v2 in SC(u). However, if v2 is accessible from u, and there is an edge in BS(X, R) from v1 to v2 , then v1 is not necessarily accessible from u. Also, if a vertex v of Tu is not accessible from u then all of the vertices of Tu that lie after the vertex v going along any geodesic path from u to an extremal vertex of Tu are also not accessible. Lemma 6. Let M = InvhX|Ri be a finitely presented Adian inverse semigroup that satisfies condition (?) and (such that) BS(X, R) contains no closed paths. Then SC(u) is finite, for every R-word u. Proof. We start from the linear automaton (α, Γ0 (u), β) and obtain the approximate complex (α, Γ1 (u), β) by applying full P-expansion on (α, Γ0 (u), β). In the construction of (α, Γ1 (u), β), we attach a path labeled by one side of a relation whose other side can be read in (α, Γ0 (u), β). In the linear automaton (α, Γ0 (u), β) we can precisely read u and all those R-words that are proper subwords of u. Each relation of R with one side u is either of the form (u, xvy) for some x, y ∈ X + and an R-word v or of the form (u, v) where v contains no R-word as its proper subword. The first case corresponds to an edge of “type 1” in Tu and the second case corresponds to an edge of “type 3” in Tu . Any R-word u1 that is a proper subword of u corresponds to an edge of ”type 2” in Tu , where u ≡ xu1 y for some x, y ∈ X + . Thus, in THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 29 performing Stephen’s P-expansion to obtain Γ1 (u) from Γ0 (u), there is a precise correspondence between first generation transversals of SC(u) and the edges of Tu whose initial vertex is either labeled by u or labeled by an R-word that is a proper subword of u. If there exists an edge of type 1 or type 2 labeled by (x, y) for some x, y ∈ X + from a vertex labeled by an R-word v to the vertex u in Tu , then either (v, xuy) ∈ R or v ≡ xuy. We show that no generation 1 transversal contains a subpath labeled by v. In other words, we show that v is inaccessible from u. Note that we cannot read the word xuy in the linear automaton (α, Γ0 (u), β) because (α, Γ0 (u), β) contains only one path from α to β that is labeled by the word u. So, if (v, xuy) ∈ R then we cannot attach a path labeled by v to the linear automaton (α, Γ0 (u), β) and if v ≡ xuy then we cannot read a path labeled by v in (α, Γ0 (u), β) because it is longer than the path labeled by u. We obtain the approximate graphs (α, Γ2 (u), β) by applying the full Pexpansion on (α, Γ1 (u), β). We observe that the second generation transversals of SC(u) are obtained as a consequence of attaching paths labeled by those R-words which label the terminal vertices of those edges of Tu whose initial vertex is either a first generation transversal or a proper subword of an R-word that labels a first generation transversal of SC(u). If there exists an edge in BS(X, R) labeled by (x1 , y1 ) for some x1 , y1 ∈ X ∗ with initial vertex labeled by v1 and terminal vertex labeled by v2 such that x1 v2 y1 is a subword of an R-word that labels a first generation transversal, then we obtain a second generation transversal by sewing on a path labeled by v1 from the initial vertex to the terminal vertex of the of the path x1 v2 y1 . If there exists an edge of type 1 or type 2 labeled by (x1 , y1 ) with initial vertex v1 and terminal vertex v2 (i.e. either (v1 , x1 v2 y1 ) ∈ R or v1 ≡ x1 v2 y1 ) such that v2 is a subword of an R-word that labels a first generation transversal but x1 v2 y1 is not a subword of that transversal, then we cannot attach a path labeled by v1 to (α, Γ1 (u), β). So, none of the vertices of Tu that occur after the vertex v1 going along a path from the vertex u to an extremal vertex of Tu will be accessible from u in SC(u). We observe that when we apply full P-expansion on (α, Γn (u), β) for some n ∈ N, we cover some more vertices of Tu that were not covered before in the sense that we add some new transversals that contains an R-word that labels a vertex of Tu and that R-word does not label a path in (α, Γn (u), β). Since Tu is a finite tree and none of the R-word label two distinct vertices of Tu therefore the process of applying full P-expansion must terminate after a finite number of steps. Hence SC(u) is finite.  Lemma 7. Let M = InvhX|Ri be an Adian inverse semigroup. Let u be an R-word and z labels a proper suffix of a transversal p of SC(u). Then either (i) z contains an R word that also labels a subpath of p, or 30 MUHAMMAD INAM (ii) A prefix of z is a suffix of some R-word that also labels a subpath of p. Proof. If z does not contain an R-word that also labels a subpath of the transversal p then the initial vertex of the path labeled by z is not the initial vertex of any R-word that labels a subpath of p. Hence the initial vertex of z lies between a pair of vertices of p that are the initial and the terminal vertex of a subpath of p that is labeled by an R-word. Hence a prefix of z is a suffix of an R-word.  We also remark that a dual statement also holds for a prefix of a transversal of the Schützenberger complex of an R-word over an Adian presentation. Remark 4. In Lemma 7, if the presentation hX|Ri satisfies condition (?) and z happens to be a prefix of an R-word, then only (i) holds, because (ii) violates the condition (?). Similarly, in the dual statement to Lemma 7, if the presentation hX|Ri satisfies condition (?) and z happens to be a suffix of an R-word, then only (i) holds, because (ii) violates the condition (?). Theorem 4.1. Let M = InvhX|Ri be a finitely presented Adian inverse semigroup that satisfies condition (?) and BS(X, R) contains no closed path. Then SC(w) is finite, for all w ∈ (X ∪ X −1 )∗ . Proof. We just need to show that SC(w) is finite for all w ∈ X + and then the above theorem follows from Theorem 3.2. So, we assume that w ∈ X + and we construct the linear automaton of w, (α, Γ0 (w), β). It follows from condition (?), that no two distinct R-words will overlap with each other. However, an R-word can be a proper subword of another R-word. So, we can uniquely factorize w as x0 u1 x2 u2 ...un xn , where xi ∈ X ∗ and ui ’s are maximal R-words in the sense that none of the ui ’s are properly contained in another R-word that is also a subword of w. It follows from Lemma 6, that SC(ui ) is finite for all 1 ≤ i ≤ n. So, we attach SC(ui ) for all 1 ≤ i ≤ n to the corresponding paths labeled by ui ’s in (α, Γ0 (w), β) to construct SC(w) and denote the resulting complex by S1 . It follows from Lemma 4, that no two edges get identified with each other as a consequence of attaching SC(ui ) for all 1 ≤ i ≤ n to the linear automaton (α, Γ0 (w), β). If S1 is closed under elementary P-expansion, then we are done. Otherwise, we will be able to read a finite number of R-words labeling the paths of S1 where we can attach new 2-cells by sewing on paths labeled by the other sides of the corresponding relations. We assume that v1 , v2 , ..., vm are the R-words that label the paths of S1 where we can attach new 2-cells. Note that each of vi labels a vertex of Tuj for some i and j, that was inaccessible from uj earlier. Because if vi labels a path in S1 then by Lemma 7 and Remark 4 the path labeled by vi contains an R-word rj as a proper subword such that the R-word rj labels a path in SC(uj ). In other words, rj labels an accessible vertex of Tuj from the vertex THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 31 uj . Since rj is a proper subword of vi , therefore there exists an edge of type 2 in Tuj with initial vertex labeled by vi and the terminal vertex labeled by rj . So, Tvi and Tuj represent the same tree for some i and j. By Lemma 6, SC(vi ) is finite for all 1 ≤ i ≤ m and cover some more vertices of Tuj in the sense that SC(vi ) contain paths labeled by those R-word which also label some of the vertices of Tuj , for some 1 ≤ j ≤ n, that were not covered by SC(uj ). We attach SC(vi ) to the paths labeled by vi for all 1 ≤ i ≤ m in S1 and denote the resulting complex by S2 . No two edges get identified with each other in S2 as a consequence of attaching SC(vi )’s to S1 by Lemma 4. If S2 is closed under elementary P-expansion then we are done. Otherwise we repeat this process of attaching Schützenberger complexes of R-words and capturing more vertices of the trees Tui for some 1 ≤ i ≤ n. This process eventually terminates, because, each Tui is a finite tree with all the vertices labeled by distinct R-words and every R-word labels a vertex of exactly one tree. Hence, SC(w) is a finite complex.  Remark 5. Let M = InvhX|Ri be an Adian inverse semigroup that satisfies condition (?) and BS(X, R) contains no closed path then the word problem for M is decidable. It follows from the Theorem ?? that the Schützenberger complex of every word w ∈ (X ∪X −1 )∗ is finite over the presentation hX|Ri. So, for any two given words w1 , w2 ∈ (X∪X −1 )∗ , we can easily check whether w1 ∈ L(w2 ) and w2 ∈ L(w1 ) or not. 4.2. The word problem for Inverse semigroups given by the presentation ha, b|abm = bn ai. In this section we show that the word problem is decidable for the inverse semigroup given by the presentation M = Invha, b|abm = bn ai, where m, n ∈ N. The word problem for the case case m = n follows from Corollary 6.6 of [7]. So, throughout this section we assume that m > n. The case m < n follows from a dual argument. We can get an alternate proof for the case n = m by following along same lines as in the case of m < n. Lemma 8. The Schützenberger complex of a word ak bt for k, t ∈ N, over the presentation ha, b|abm = bn ai, is finite. Proof. We adopt a slightly different approach to construct SC(ak bt ). We draw edges labeled by a horizontally and edges labeled by b vertically. Then the linear automaton of ak bt , (α0 , Γ0 (ak bt ), β0 ), is shown in Figure 14. If t < m, then we cannot attach any 2-cell to (α0 , Γ0 (ak bt ), β0 ). So, the above lemma is true for this case. If t ≥ m, then t = q1 m+r1 , where q1 is the quotient and r1 is a remainder and 0 ≤ r1 < m. We can attach q1 2-cells in the first column along the vertical segment labeled by bt of (α0 , Γ0 (ak bt ), β0 ). After attaching all the 2-cells in the first column along the vertical segment labeled by bt we have created a new vertical segment labeled by bnq1 , because there are total q1 32 MUHAMMAD INAM 2-cells and each 2-cells contains exactly n edges on the newly attached side of the 2-cell. β0 br1 a bn bm a bt bn α0 a a bm a Figure 14. Construction of SC(ak bt ). If nq1 < m or k = 1 then the process of attaching new 2-cells terminates at this stage. If neither nq1 < m nor k = 1, then nq1 = q2 m + r2 , where q2 is a quotient, r2 is a remainder and 0 ≤ r2 < m. So, We can attach a column of q2 2-cells along the vertical segment labeled by bnq1 . This process of attaching columns of new 2-cells terminates after at most k steps. So, SC(ak bt ) is a finite complex.  Remark 6. In the above construction of SC(ak bt ) in Lemma 8, every new maximal vertical segment contains fewer edges than the other vertical side of the same column of 2-cells. Lemma 9. The Schützenberger complex of a word bt ak for k, t ∈ N, over the presentation ha, b|abm = bn ai, is finite. Proof. We draw edges labeled by a horizontally and edges labeled by b vertically. Then the linear automaton of bt ak , (α0 , Γ0 (bt ak ), β0 ), is shown in Figure 15. If t < n, then we cannot attach any 2-cell to (α0 , Γ0 (bt ak ), β0 ). So, the above lemma is true for this case. If t ≥ n, then t = q1 n + r1 , where q1 is the quotient and r1 is a remainder and 0 ≤ r1 < n. We can attach q1 2-cells in the first column along the vertical segment labeled by bt of (α0 , Γ0 (ak bt ), β0 ). After attaching all the 2-cells in the first column along the vertical segment labeled by bt we have created a new vertical segment labeled by bmq1 , because there are total q1 2-cells and each 2-cells contains exactly m edges on the newly attached side of the 2-cell. If k = 1 then the process of attaching new 2-cells terminates at this stage. Otherwise mq1 = q2 n + r2 , where q2 is a quotient, r2 is a remainder and THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 33 a bn a β0 bm a bt a bn bm a br1 α0 Figure 15. Construction of SC(bt ak ). 0 ≤ r2 < n. So, We can attach a column of q2 2-cells along the vertical segment labeled by bmq1 . Clearly, this process of attaching columns of new 2-cells terminates after k steps. So, SC(bt ak ) is a finite complex.  Remark 7. In the above construction of SC(bt ak ) in Lemma 9 every new maximal vertical segment contains more edges than the other vertical side of the same column of 2-cells. Theorem 4.2. For all w ∈ {a, b, a−1 , b−1 }∗ the Schützenberger complex of a word w, SC(w), over the presentation Invha, b|abm = bn ai is finite. Hence the word problem is decidable for M . Proof. Since hX|Ri is an Adian presentation, we just need to show that SC(w) for all w ∈ {a, b}+ is finite, then Theorem 4.2 follows from Theorem 3.2. If w is of the form ak or bk for some k ∈ N, then SC(w) is finite. We assume that w ≡ ak0 bt0 ak1 bt1 ...akl btl , where k0 , tl ∈ N ∪ {0} and ki , tj ∈ N for 1 ≤ i ≤ l and 0 ≤ j ≤ l − 1. We construct SC(w) by drawing the edges labeled by a horizontally and the edges labeled by b vertically. So, (α0 , Γ0 (w), β0 ) looks like the diagram shown in Figure 16. We attach SC(aki bti ) on the path labeled by aki bti of (α0 , Γ0 (w), β0 ) wherever it is possible to attach and we denote the resulting complex by S1 . No two edges get identified with each other as a consequence of attaching these finite complexes to (α0 , Γ0 (w), β0 ) by Lemma 4. As a consequence of attaching these finite complexes to the (α0 , Γ0 (w), β0 ), we have created at most l − 1 new maximal directed paths labeled by ak bt for some k, t ∈ N. 34 MUHAMMAD INAM β0 bt2 ak2 bt1 ak1 bt0 α0 ak0 Figure 16. (α0 , Γ0 (w), β0 ) We attach finite complexes of the form SC(ak bt ) to every new maximal path labeled by a word of the form ak bt for some k, t ∈ N in S1 . We denote the resulting complex by S2 . Again by Lemma 4, no two edges get identified with each other in S2 . We can read at most l−2 new maximal directed paths in S2 which are labeled by the words of the form ak bt for some k, t ∈ N. So, we repeat the process attaching finite Schützenberger complexes of the words of the form ak bt where ever it is possible to attach and denote the resulting complex by S3 . Note that this process of attaching finite Schützenberger complexes of the words of the form ak bt eventually terminates after at most l steps. We denote the resulting complex by S 0 . Now in S 0 , on the other side of the path labeled by w, we attach finite Schützenberger complexes of the words bti aki+1 at the paths labeled by bti aki+1 for 0 ≤ i ≤ l − 1, where ever it is possible to attach and denote the resulting complex by S10 . By Lemma 4 no two edges in S10 get identified with each other as a consequence of attaching these finite complexes. As a consequence of attaching these finite complexes we have created at most l − 1 new maximal paths which are labeled by the words of the form bt ak for some k, t ∈ N. So, we repeat the process of attaching finite Schützenberger complexes of the words of the form bt ak at the corresponding new paths in S10 . We denote the resulting complex by S20 . This process of attaching finite complexes of the words of the form bt ak terminates after at most l steps and we obtain a finite complex which is closed under elementary P-expansion and folding. Hence, SC(w) is a finite complex.  THE WORD PROBLEM FOR SOME CLASSES OF ADIAN INVERSE SEMIGROUPS 35 Acknowledgement The author of this paper is thankful to John Meakin and Robert Ruyle for their several useful suggestions. References [1] Adian, S. I., Defining relations and algorithmic problems for groups and semigroups, Proc. Steklov Inst. Math., No 85 (1966). [2] Inam, M., Meakin, J., Ruyle, R., A structural property of Adian inverse semigroups, Semigroup Forum, Nov. 2015. [3] Lawson, M. V., Inverse semigroups, World Scientific Co. Pte. Ltd., 1998. [4] Linblad, S. P.,Inverse monoids presented by a single relator, PhD thesis, Dept. of Math., University of Nebraska-Lincoln, Dec. 2003. [5] Magnus, W., Das Identitätsproblem für Gruppen mit einer definierenden Relation, Math Ann. 106 (1932), 295-307. [6] Remmers, J. H., On the Geometry of Semigroup Presentations, Advances in Mathematics, 36 (1980), 283-296. [7] Stephen, J. B., Presentations of inverse monoids, Journal of Pure and Applied Algebra (1990), 81-112. [8] Steinberg, B., A topological approach to inverse and regular semigroups, Pacific J. Math. 208 (2003), no. 2, 367-396. [9] Steinberg, B., A sampler of a topological approach to inverse semigroups, Semigroups, Algorithms, Automata and Languages, Word Scientific, (2002), 437 - 461. E-mail address: [email protected] Department of Mathematics, University of West Georgia, Carrollton, Georgia 30118 USA
4math.GR
POWER MANAGEMENT DURING SCAN BASED SEQUENTIAL CIRCUIT TESTING Reshma.p M.tech Scholar, Department of Electronics and Communication Engineering, Amrita Vishwa Vidyapeetham University, Coimbatore. [email protected] ABSTRACT This paper shows that not every scan cell contributes equally to the power consumption during scan based test. The transitions at some scan cells cause more toggles at the internal signal lines of a circuit than the transitions at other scan cells. Hence the transitions at these scan cells have a larger impact on the power consumption during test application. These scan cells are called power sensitive scan cells.A verilog based approach is proposed to identify a set of power sensitive scan cells. Additional hardware is added to freeze the outputs of power sensitive scan cells during scan shifting in order to reduce the shift power consumption.when multiple scan chain is incorporated along with freezing the power sensitive scan cell,over all power during testing can be reduced to a larger extend. KEYWORDS Scan ,Scan DFlip-Flop ,DFTAdvisor 1. INTRODUCTION With current advances in very large scale integrated technology , the sensitivity of today’s chips to deep submicrometer effects is increasing. Along with technology scaling, the increase in the operating frequency and the increase in the functional density of today’s digital designs has led to new challenges for designers and test engineers. Furthermore, dynamic power consumption and IR-drop due to excessive switching activity are critical challenges. As a result, power reduction techniques have been extensively studied by both industry and academia with respect to both design and test. Scan-based test remains one of the most widely accepted design-for-test techniques because it significantly improves the controllability and the observability of the circuit’s internal nodes with an insignificant area and performance overhead.Switching activity during scan-based test is often much higher than that during normal operation. There are multiple reasons for this phenomenon. First, the test vectors applied consecutively are not correlated. Second, nonfunctional states may be traversed during scan-test. Furthermore, test compaction and testing multiple cores simultaneously contribute toward high-switching activity. In addition, as patterns are shifted into and out of the scan chains, multiple changes of the flip-flop values can propagate into the combinational logic and cause massive amounts of switching. Scan-based tests might cause excessive circuit switching activity compared to a circuit’s normal operation . Higher switching activity causes higher peak supply currents and higher power dissipation. High power dissipation during test can cause many problems, which are generally addressed in terms of average power and peak power. Average power is the total distribution of power over a time period, and is calculated using the ratio of consumed energy to test time .Peak power is the highest power value at any given instant. When peak power is beyond the design limit, a chip cannot be guaranteed to function properly due to additional gate delays caused by the supply voltage drop. The power consumption within one clock cycle may not be large enough to elevate the temperature over the chip’s thermal capacity limit. To damage the chip, high power consumption must last for an enough number of clock cycles.The test power consumed during scan shifting and capture cycles is referred to as shift power and capture power, respectively. A typical scan chain in industrial designs consists of at least hundreds of scan cells, whereas the capture window only lasts one or a few clock cycles. Clearly, the average power consumption is determined by the shift power. Excessive shift power accumulation may make a good chip fail during test even if the peak capture power is low. Inserting no operation cycles between the end of scan shifting and the beginning of capture cycles can reduce the chance of rejecting good chips during test. This experimental result implies that test power reduction strategies should focus on reduction of the average shift power and the peak capture power.This paper, focus on reducing the average shift power in scan-based tests. An effective method to identify a set of power sensitive scan cells is used. By inserting additional hardware to freeze power sensitive scan cell outputs to pre-selected values, the average shift power consumption can be reduced dramatically. The paper is organized as follows. In Section 2,previous work. In Section 3,proposed method. In Section 4, experimental results on S27 is presented, followed by conclusions in Section 5. 2.The Previous work To reduce the switching activity during scan shift, automatic test pattern generation (ATPG)-based approach and 2) DFT-based approach were used. The advantage of the ATPG based solutions is that they do not modify the original design and the scan architecture,but modification is done on test vectors and there by power reduction can be obtained. DFT-based solutions require one to either partition the conventional scan chain architecture or insert additional hardware into the design .Different DFT approach used here are included in this literature survey. In Minimised power consumption for scan based Bist, extra logics are inserted to hold the outputs of all the scan cells at constant values during scan shifting. This methods not only minimize the average scan shift power, but also avoid peak power hazards during scan shifting. The main disadvantage of these approaches is the large area overhead, since additional logics are added to all the scan cells. Moreover, it may degrade circuit performance due to extra logics added between scan cell outputs and functional logics. To reduce the area overhead due to additional gates, supply gating transistors for the first-level gates at the outputs of scan cells are proposed in Low power scan design using first level supply gating. An alternative implementation to hold the scan cell outputs by using dynamic logic was proposed in Techniques for minimizing power dissipiation in scan and combinational circuit during test application.The method proposed in Inserting test points to control peak power during scan testing, inserts test points at selected scan cell outputs to keep the peak shift power at every shift cycle below a specified limit. Given a set of test patterns, logic simulation is carried out to identify the shift cycles in which peak power violations occur. Those cycles are called violating cycles. By using integer linear programming (ILP) techniques, the optimization problem is solved to select as few test points as possible such that all violating cycles can be eliminated. In Partial gating optimization for power reduction during test application, random vector simulation was used to guide partial test point selection. When simulating a random vector, the primary inputs and the pseudo primary inputs are changed to value X with prespecified probabilities, and the number of gates becoming X after the change is used as a cost function to identify the logic value assigned at the primary inputs and the pseudo primary inputs, as well as to select scan cells to be held during scan shifting. To explore several hundred thousands of scan cells in an industrial circuit, a significant number of random vectors need to be simulated in order to choose good test points. Motivated by the test point insertion approach along with multiple scan chain, scan shift power can be reduced to a larger extend. Some scan cells have a much larger impact on toggle rates at the internal signal lines than other scan cells. These scan cells are called power sensitive scan cells. Objective is to quickly identify power sensitive scan cells and their preferred frozen values during scan shifting. By freezing a small percentage of scan cells that are the most power sensitive, reduction in scan shift power can be achieved ,while minimizing the additional area overhead. Compared with the previous approaches , this approach has less area overhead and can avoid modifying scan cells at critical paths by not selecting them to freeze. This approach also provides a practical way to handle large industrial designs and since both freezing power sensitive scan cells and multiple scan chain approach is used, power can be reduced to a larger extend. 3.The Proposed Method A mentorgraphics tool called DFTAdvisor is used for Scan insertion.An S27 ISCAS benchmark circuit is taken and scan is inserted in that circuit using DFTAdvisor.By using Fastscan all the testvectors for stuck at faults are identified.By simulating the testbench obtained from Fastscan,along with the verilog code obtained from the DFTAdvisor,number of toggling are calculated.By using a counter toggling are counted and identify the point where the logic gate has to be inserted.In S27 circuit it is find that the toggling in between the first and zeroth ScanD-FF is higher.By introducing logic gates at the output of first ScanD-FF we can block the complete toggling occurring at the combinational part,when only the Flip-Flops are selected. An additional AND gate with the second input inverted is inserted between the scan cell output and the function logics it drives. During scan shifting, the scan enable signal se is asserted to be 1. It makes additional gate output become constant. During normal operations, se is de-asserted to be 0 and the additional gates do not impact normal operations. 4. Experimental Results For testing a sequential circuit, both combinational cloud testing as well as memory element testing is needed. The approach used here converts all D-flip-flop to scanDflipflop.DFTAdvisor is a synthesis tool capable of doing scan insertion. It accepts gate level net list format and generates a new net list with scan cells inserted.The circuit in Fig.1 shows ISCAS S27 benchmark circuit.when the synthesisable netlist code for it is given to DFTAdvisor,it will generate a code with scan inserted as shown in Fig.2. When the code obtained from DFTAdvisor is given to Fastscan,corresponding test vectors obtained are as shown in Fig.3.By simulating the testbench obtained from Fastscan,along with the code from DFTAdvisor, the result is as shown in Fig.4 Fig.1 ISCAS S27 Benchmark Ckt Fig.2a. Invocation of DFTAdvisor Fig.2b. Scan inserted ckt obtained from DFTVisualiser Fig.2c.Scan inserted code generated by DFTAdvisor Fig.3.Test vectors obtained from Fastscan Fig.4.Simulated result From the simulated result, it is clear that when scan enable is high there will be toggling at combinational part.But it is not desired.when scan enable is high ther will not be any toggling at combinational part,only flip-flop is involved.By using a counter,the toggligs when scan enable is high are calculated and found that toggling between Dff_1 and Dff_0 , and Dff_2 and Dff_0 are higher.By introducing logic gates at these points it is seen that overall toggling can be reduced to a larger extend.It is shown in Fig.5. Fig.5.a.Toggling after scan insertion Fig.5.b.Logic insertion between Dff_1 and Dff_0 Fig.5.c.Simulation result after gate insertion Fig.5.d.Toggling after logical insertion Area will be a trade-off for power.By simulating S27 in XILINX, it is observed that will not be that much increase in area overhead.It is shown in Fig.6.By using multiple scan chains ,it is verified that time required drastically decreases.corresponding simulation results are shown in Fig .7. Fig.6.a.Utilization of area in S27 Fig.6.b.Utilization of area in S27 after scan insertion Fig .6.c.Utilization of area in S27 after logical insertion between Dff_1 and Dff_0 Fig .7.a. 1_scan chain Fig .7.b. 2_scan chain Fig .7.c. 3_scan chain Fig .7.d. Simulation result_1 scan chain Fig .7.e. Simulation result-3 Scan chain 1 Scan chain Number of Clock required Time required 3 Scan chain 24 20 1600ns 1400ns Fig .7.f. Comparison result 5. CONCLUSION This paper presented and analyzed a method for reducing switching activity during scan shift by freezing a small subset of all flip-flops at the RTL. Large reductions in switching activity can be achieved with very low-area overhead. The amount of scan flip-flops that are going to be frozen can be decreased/increased depending on the design’s overhead budget. In comparison with previous methods, which freeze these flip-flops at the gate level, timing closure can be more easily met. When flip-flops are frozen at the gate level, individual timing analysis have to be implemented to determine whether or not each flipflop could be frozen without violating timing. By freezing all flip-flops simultaneously at the RTL,this approach allow the synthesis tool to automatically optimize for timing closure. In addition, this paper has presented a detailed analysis of the switching activity reduction that can be obtained with very few frozen flip-flops. In fact, in one case, a 79% reduction in switching activity was achieved with an area overhead of only 0.02%. This switching activity analysis considered both hazards and final circuit values.when multiple scan chain is applied along with freezing powersensitive scan cells, the test power is seen to reduced to a larger extend along with that time required for testing also get reduced. 6. REFERENCES [1] P. Girard, “Survey of low-power testing of VLSI circuits,” IEEE Design Test Comput., vol. 19, no. 3, pp. 80–90, May–Jun. 2002. [2] S. Ravi, “Power-aware test: Challenges and solutions,” in Proc. Int. Test Conf., Oct. 2007, pp. 1–10. [3] K. M. Butler, J. Saxena, A. Jain, T. Fryars, J. Lewis, and G. Hetherington, “Minimizing power consumption in scan testing: Pattern generation and DFT techniques,” in Proc. Int. Test Conf., Oct. 2004, pp. 355–364. [4] S. Wang and S. Gupta, “An automated test pattern generator for minimizing switching activity during scan testing activity,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 21, no. 8, pp. 954– 968, Aug. 2002. [5] R. Sankaralingam, R. R. Oruganti, and N. Touba, “Static compaction techniques to control scan vector power dissipation,” in Proc. VLSI TestSymp., 2000, pp. 35–40. “A Study of Wavelet Thresholding Denoising” ,Proceedings of ICSP2000. [6] X. Wen, Y. Yamashita, S. Kajihara, L.-T. Wang, K. Saluja, and K.Kinoshita, “On low-capture power test generation for scan testing,” in Proc. VLSI Test Symp., 2005, pp. 265–270. [7] X. Wen, S. Kajihara, K. Miyase, T. Suzuki, K. Saluja, L.-T. Wang,K. Abdel-Hafez, and K. Kinoshita, “A new ATPG method for efficient capture power reduction during scan testing,” in Proc. VLSI Test Symp., 2006, pp. 60–65. [8] Y.-T. Lin, M.-F. Wu, and J.-L. Huang, ―PHS-fill: A low power supply noise test pattern generation technique for at-speed scan testing [9] S. Kajihara, K. Ishida, and K. Miyase, ―Test vector modification for power reduction during scan testing,‖ in Proc. VLSI Test Symp., 2002, pp. 160–165. [10] J. Saxena, K. M. Butler, V. B. Jayaram, S. Kundu, N. V. Arvind, P. Spreeprakash, and M. Hachinger, ―A case study of IR-drop in structured at-speed testing,‖ in Proc. Int. Test Conf., Oct. 2003, pp. 1098–1104. [11] W. Li, S. M. Reddy, and I. Pomeranz, ―On reducing peak current and power during test,‖ in Proc. IEEE Comp. Soc. Annu. Symp. VLSI, May 2005, pp. 156–161. [12] S. Remersaro, X. Lin, Z. Zhang, S. M. Reddy, I. Pomeranz, and J. Rajski, ―Preferred fill: A scalable method to reduce capture power for scan based designs,‖ in Proc. Int. Test Conf., Oct. 2006, pp. 110. [13] N. Badereddine, P. Girard, S. Pravossoudovitch, C. Landrault, A. Virazel, and H.-J. Wunderlich, ―Structural-based power-aware assignment of don’t cares for peak power reduction during scan testing,‖ in Proc. IFIP Int. Conf. VLSI, Oct. 2006, pp. 403–408. [14] N. Ahmed, M. Tehranipoor, and V. Jayaram, ―Transition delay fault test pattern generation considering supply voltage noise in a SoC design,‖in Proc. Design Automat. Conf., Jun. 2007, pp. 533– 538. [15] N. Nicolici, B. M. Al-Hashimi, and A. C. Williams, ―Minimization of power dissipation during test application in full-scan sequential circuits using primary input freezing,‖ IEE Proc. Comput. Digital Tech., vol. 147, no. 5, pp. 313–322, Sep. 2000. [16] T.-C. Huang and K.-J. Lee. ―Reduction of power consumption in scanbased circuits during test application by an input control technique,‖ IEEE Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 20, no. 7, pp. 911–917, Jul. 2001. [17] V. Dabholkar, S. Chakravarty, I. Pomeranz, and S. M. Reddy, ―Techniques for minimizing power dissipation in scan and combinational circuits during test application,‖ IEEE Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 17, no. 12, pp. 1325–1333, Dec. 1998. [18] Elif Alpaslan, Yu Huang, Xijiang Lin, Wu-Tung Cheng, and Jennifer Dworak, “On reducing scan shift activity at RTL” IEEE Trans.Comput.-Aided Design Integr. Circuits Syst., vol. 29, no. 7, july.2010. [19] X. Lin and Y. Huang, “Scan shift power reduction by freezing power sensitive scan cells,” J. Electron. Testing: Theory Appl., vol. 24, no. 4, pp. 327–334, Aug. 2008.
5cs.CE
1 Optimal Pricing of User-Initiated Data-Plan Sharing in a Roaming Market arXiv:1712.06236v1 [cs.GT] 18 Dec 2017 Feng Wang, Lingjie Duan, and Jianwei Niu Abstract A smartphone user’s personal hotspot (pH) allows him to share cellular connection to another (e.g., a traveler) in the vicinity, but such sharing consumes the limited data quota in his two-part tariff plan and may lead to overage charge. This paper studies how to motivate such pH-enabled data-plan sharing between local users and travelers in the ever-growing roaming markets, and proposes pricing incentive for a data-plan buyer to reward surrounding pH sellers (if any). The pricing scheme practically takes into account the information uncertainty at the traveler side, including the random mobility and the sharing cost distribution of selfish local users who potentially share pHs. Though the pricing optimization problem is non-convex, we show that there always exists a unique optimal price to tradeoff between the successful sharing opportunity and the sharing price. We further generalize the optimal pricing to the case of heterogeneous selling pHs who have diverse data usage behaviors in the sharing cost distributions, and we show such diversity may or may not benefit the traveler. Lacking selfish pHs’ information, the traveler’s expected cost is higher than that under the complete information, but the gap diminishes as the pHs’ spatial density increases. Finally, we analyze the challenging scenario that multiple travelers overlap for demanding data-plan sharing, by resorting to a near-optimal pricing scheme. We show that a traveler suffers as the travelers’ spatial density increases. Index Terms Roaming markets, personal hotspot, data-plan sharing, pricing mechanism, information uncertainty. Part of the work was presented in Proc. IEEE GLOBECOM, Singapore, Dec. 2017 [1]. F. Wang and L. Duan are with the Engineering Systems and Design Pillar, Singapore University of Technology and Design, Singapore (e-mail: [email protected], [email protected]). L. Duan is the corresponding author. J. Niu is with the School of Computer Science and Engineering, Beihang University, China (e-mail: [email protected]) 2 I. I NTRODUCTION A. Background and Motivation Roaming is the ability of customers to use their mobile devices outside the geographical coverage area provided by their normal network operator [2]. As the penetration of smart mobile devices increases fast, the volume of the global data roaming market has grown by more than sixfold in the past five years [2]–[5]. Accordingly, the revenue of data roaming is expected to increase to US$50 billions in revenues by 2019 (see [5], [6]). Travelers may suffer a “bill shock”, and data roaming is typically expensive as compared to the domestic markets’ two-part tariff data plans for local users. To reduce the roaming cost, a traveler may enjoy data services by accessing to the public WiFi hotspots. However, free WiFi hotspots are limited in coverage and they usually concentrate on the public service areas (e.g., airports, stations, malls, and public libraries) to avoid formidably high deployment cost for full coverage. For example, even for a populous city such as Singapore, the WiFi coverage percentage is only 35% [7]. It is urgent to find economically viable approaches for providing travelers with ubiquitous wireless data services. With the recent technology advancements, cellular-enabled iPhones and Android phones can now set up personal hotspots (pHs) to share their cellular data connections with nearby wireless devices (e.g., phones, laptops, and tablets) [8]. The physical coverage of a pH ranges about hundreds of feet like WiFi and is expected to keep increasing. However, the development of this user-initiated data-plan sharing still lacks a clear business model and a local selfish user is only willing to share the pH connection with his own devices [19]. As such, it is important to propose incentive schemes for pH-enabled data-plan sharing in a broad popularity. A key problem to hinder the development of such data-plan sharing is the sharing cost. A selling user’s sharing consumes his monthly data quota in his two-part tariff plan and may lead to overage charge. After subscribing to a data-plan denoted by (Q, P0 , β), a local user is given a monthly data quota Q at a fixed lump-sum fee P0 and should pay for overage data beyond Q at a costly unit price β. In addition, such pH sharing consumes the local user’s energy in battery storage and requires the selling user to stay during the sharing period, incurring another waiting cost. To compensate for the costs, we aim to design a reward-based pricing scheme for a traveler to fairly cover the pH sharing cost. As shown in Fig. 1, a traveler T opportunistically demands data connection from nearby pHs (if any), by announcing a sharing price $p (ex ante) 3 pH 1 pH-range d $p buy data with roaming fee $C0 sharing price $p $p Local wireless network operator (WNO) … pH N pH 2 under two-part tariff plan (Q, P0, β) Traveler T broadcasts sharing price p to all possible pHs within pH range d Alternatively, traveler T buys from a local WNO at roaming cost C0 Traveler T Personal hotspot (pH) provided by a local user Fig. 1. A user-initiated data-plan sharing model with one traveler and personal hotspots (pHs). as a reward via short-range communications. The nearby pHs then respond to accept or not based on their private costs. Finally, the traveler T selects one pH for cellular connection and pays reward $p to the activated pH.1 However, this optimal pricing design under information uncertainty is challenging at the traveler side. Since the potential pHs are generally moving, the traveler does not know the exact number of pHs in the sharing area. Furthermore, how to estimate the private costs of pHs is another problem for the traveler. Only a pH knows his own sharing cost, which depends on the realized data usage in his two-part tariff plan. In general, different pHs have different data usage behaviors or cost distributions. The traveler would prefer to activate a pH with low cost distribution, yet this screening is not doable for ex ante pricing. Intuitively, one can expect that a high sharing payment will provide a good chance for pH sharing but leave a low sharing benefit to the traveler. This motivates our study in this work. B. Contributions In this paper, we first develop a model for pH-enabled data-plan sharing and then find the optimal pricing in different scenarios. Our main contributions are summarized as follows. 1 The terms “selling user” and “pH” are interchangeable in this paper. 4 • Pricing incentive for pH sharing under information uncertainty: In Section II, to provide an incentive of data sharing to nearby pHs, we design a pricing scheme for an on-demand traveler to reward pHs and reach a win-win situation. The pricing scheme practically considers the traveler’s information uncertainty about pHs, including the mobile pHs’ random locations following a Poisson point process (PPP) and their sharing cost distribution under the two-part monthly tariff plan. • Benchmark case under complete information: In Section III, to provide a performance bound and evaluate the proposed pH pricing schemes later, we investigate the social optimum benchmark case under complete information, where the pHs nearby are willing to report their locations and private costs. Then the traveler pays and activates the pH with the minimal cost. We show that the traveler’s expected cost decreases in both the pHs’ spatial density and their residual data quota. • Optimal pricing analysis for various pHs under information uncertainty: In Section IV, the optimal price is determined for the traveler to minimize his expected cost after sharing. We first consider homogeneous pHs whose statistics of monthly data usage are identical. Though the optimization problem is non-convex, we derive the globally optimal sharing price for the traveler. We show that the traveler’s final expected cost is larger as compared to the case under complete information and the gap diminishes with the pH spatial density. Furthermore, we successfully extend the optimal pricing scheme to the case of heterogeneous pHs who have different data usage statistics. The optimal price is derived by comparing their diverse cost distributions and targeting for the dominant pH type. This diversity may or may not benefit the traveler. • Optimal pricing for overlapping travelers: In Section V, we further consider the scenario that a traveler may overlap with other nearby travelers in demanding common pHs. To handle the non-tractable expression of the expected cost for travelers, a lower bound of the expected cost is resorted to obtain the near-optimal price for coordinating all travelers. We show that each traveler suffers from the increase of the travelers’ spatial density. C. Related Work It is noted that pricing incentive design has received increasingly attentions for wireless networks [9], [10]. The works in [11]–[13] surveyed some generic pricing and auctions schemes for mobile crowdsourcing. Some recent works investigated the pricing incentive in data trading 5 $p for data volume B pH 1 T Accept pH 1 T pH 1 pH 2~N T pH 2~N T … $p for data volume B pH N (a) Traveler request phase T Reject (b) pH response phase (c) pH sharing phase Fig. 2. An illustrative example of pH-enabled data-plan sharing between traveler T and N pHs ((if any). and user cooperation in the literature. For example, [14] and [15] studied the optimal pricing for static WiFi hotspots (e.g., in a cafe) according to customer types and network capacity. About mobile users’ data plan sharing, [16] and [17] studied mobile data trading among users under the central coordination of the wireless operator in the operator-controlled secondary market. [18] further investigated the pricing incentive to stimulate users’ relay cooperation for energy saving purpose. Only [19] investigated a user-initiated data-plan sharing scenario by employing pHs, where users with diverse data usage trade data plans and the wireless operator indirectly intervene with such sharing to control overage charge. However, [19] assumed complete information for mobile data trading and did not take into account user mobility and private cost in practice. Unlike prior works, this paper studies a unique pH pricing problem under information uncertainty, where the traveler only has partial information about pHs’ private locations and costs. II. S YSTEM M ODEL Consider a typical traveler T who wants to connect the Internet. He can potentially buy data from any neighboring pH within pH range d. Equally, we consider the traveler is located in the center of circle area A with radius d. The pH range d is determined by traveler T’s quality of service requirement [19]. Traveler T can access to the data service either by paying one activated pH within the area A or by paying the local wireless network operator (WNO). Denote by H the set of randomly appeared pH number in the area A and N , |H|. Let B be the volume of requested data to buy. For traveler T, we denote p and C0 as the price announced by the traveler to pHs nearby and the roaming fee charged by the local WNO, respectively. 6 The pH-enabled data-plan sharing procedure is illustrated in Fig. 2, and can be divided into three phases: • Traveler request phase: at the beginning, traveler T broadcasts the data-plan sharing request message via short range communications (e.g., bluetooth), where the message contains the offered price p for the data volume B. • pH response phase: upon receiving the request, in this phase each pH i ∈ H (if any) makes a decision to accept or not and responds to traveler T. • pH sharing phase: in this phase, if at least one positive pH response message is received, then traveler T will randomly select one of the positive pHs and pay p to establish the pH connection; otherwise, traveler T will resort to the WNO by paying roaming fee C0 for data volume B. It is worth noting that the pHs are moving in general and the private information of pHs (e.g., the number N and data-plan sharing costs) is not available for traveler T ex ante in traveler request phase. A. pHs’ models about mobility, cost and utility To capture the nature of pH mobility, we assume that the locations of pHs in target circle area A follow a two-dimensional Poisson point process (PPP) with spatial density λ as in [21]. The average number N of pHs in area A is λπd2 , and the probability mass function (PMF) of N is Pr(N = n) = (λπd2 )n exp(−λπd2 ), n = 0, 1, · · · n! (1) where Pr(X) denotes the probability of event X. Ideally, traveler T wants to attract only one pH at minimum price but may fail to attract none, depending on the cost distribution of various pHs in sharing. The data-plan sharing to traveler T will consume pH i’s cellular data amount B. He has subscribed to the identical two-part tariff plan (Q, P0 , β) from the local WNO in the long run, where Q is the monthly data quota, P0 is the fixed lump fee, and β is the unit price for overage data beyond Q. Denote by xi the estimated monthly data volume consumed by pH i. For each pH i ∈ H, we assume that xi follows the independent and identically distributed (i.i.d.) Gaussian distribution with mean µ and variance σ 2 , i.e., xi ∼ N (µ, σ 2). In practice, when each pH i ∈ H predicts his monthly data usage xi , he will inevitably attach an additive noise ∆i ∼ N (0, δ 2) according to his subjective estimation. By considering the shared volume B with traveler T, the 7 actual monthly usage is xi + ∆i + B for pH i ∈ H, which may be larger than quota Q and incur surcharge with rate β. Given xi estimation, the (additional) cost for pH i after sharing with traveler T is Ci (xi ) = E{β(xi + ∆i + B − Q)+ } − E{β(xi + ∆i − Q)+ }    0, if 0 ≤ xi ≤ Q − B    = β(xi + B − Q), if Q − B < xi < Q     βB, if xi ≥ Q, (2) where (x)+ , max(x, 0) and the expectation E{·} is taken over the estimation noise ∆i . Define φ(x) and Φ(X) as the probability density function (PDF) and the cumulative distribution function (CDF) of Gaussian variable xi , respectively. According to (2), the minimal cost of pH i ∈ K for data-plan sharing is zero if his monthly data usage xi plus the sharing data amount B is within the monthly data quota Q, while the maximal cost is βB if his monthly data usage xi is already larger than the data quota Q. Based on (2) for random xi , we obtain the CDF of the sharing cost Ci (xi ) ≤ βB in the following lemma: Lemma 2.1: For pH i ∈ H, the CDF of his additional cost Ci (xi ) incurred by traveler T’s data consumption B is given by Pr(Ci (xi ) ≤ c) = Proof: See Appendix A.   Φ( c + Q − B), β  1, if 0 ≤ c < βB (3) if c = βB. As a return, pH i will receive a reward price p from traveler T. The utility of pH i ∈ H via data-plan sharing is given by Ui (xi ) = p − Ci (xi ). (4) Reasonably, we assume that each pH i ∈ H has a reservation utility ǫ > 0 that must be guaranteed to cover the battery consumption2 and waiting cost during the access time of the traveler. Hence, pH i accepts traveler T’s sharing request only if Ui (xi ) ≥ ǫ. The traveler has to buy data from the WNO at cost C0 if the two following conditions hold: 2 How to quantify the transmission energy cost can be found in [20], where the energy consumption depends on the transmission time and the volume of transmission data. 8 • No pH exists in area A, i.e., N = 0 • The utility of each pH i in the non-empty set H is smaller than ǫ, i.e., p − Ci (xi ) < ǫ, ∀i ∈ H. The realized cost for traveler T is given as   p, if ∃i ∈ N , p − Ci (xi ) ≥ ǫ CT =  C0 , otherwise. (5) To ensure the mutual benefits of traveler T and pHs in this user-initiated data-plan sharing, the offer price p by the traveler T should satisfy ǫ ≤ p ≤ C0 , (6) where the first inequality in (6) guarantees the utility to increase of the pHs and the second inequality in (6) saves the roaming cost for traveler T. III. B ENCHMARK C ASE UNDER C OMPLETE I NFORMATION In this section, we consider the ideal benchmark case under complete information, where the set H and the private information of pH i ∈ H (including the expected monthly data usage xi and the cost Ci (xi ) in sharing) is available for traveler T. This case can happen when both traveler T and the pHs belong to a cooperative community (e.g., family members or friends) or pHs are altruistic by reporting their locations and private costs. This case serves as a performance benchmark (lower bound for the traveler), when comparing to Sections IV and V under incomplete information. Under complete information, if pH set H is non-empty, traveler T only offers a price to cover the reservation utility ǫ plus the minimum cost among all pHs, i.e., p = ǫ + min Ci (xi ). i∈H (7) Without loss of generality, we reorder the N pHs’ costs according to C1 ≤ C2 ≤ . . . ≤ CN , where Ci is the ith smallest pH sharing cost and depends on xi realization. Then (7) reduces to p = ǫ + C1 . Based on Lemma 2.1, it follows that the CDF of C1 is [24]    1 − 1 − Φ( c + Q − B) N , if 0 ≤ c < βB β Pr(C1 ≤ c) =  N (1 − Φ(Q)) , if c = βB. (8) (9) 9 and the PDF of C1 is ∂ Pr(C1 ≤ c) ∂c N −1 c N c = 1 − Φ( + Q − B) φ( + Q − B) β β β fC1 (c) = for 0 ≤ c < βB. (10) Based on (9) and (10), we have the following lemma. Lemma 3.1: Given the non-empty pH set (i.e., N ≥ 1), the expected cost of traveler T is given by E{C1 } = Z βB 0 1 − Φ( Proof: See Appendix B. N c + Q − B) dc. β (11) Lemma 3.1 shows that E{C1 } is a decreasing function of N ≥ 1. Denoted by EC ∗∗ the expected cost of traveler T under complete information. By substituting (11) into (8) and taking expectation of (8) over random N, we establish the follow theorem with the use of iterated expectation [23]. Theorem 3.1: The expected cost EC ∗∗ of traveler T under complete information is given by      Z βB √ − 2 λπd2 erfc x/β+Q−B−µ 2σ  dx, EC ∗∗ = ǫ + (C0 − ǫ − βB) exp(−λπd2 ) + exp  2 0 (12) where erfc(x) , √2 π R +∞ x 2 e−t dt is a complementary error function. Proof: See Appendix C. Note that if B + µ > Q, i.e., the mean usage µ plus the sharing data B is already exceed the data quota Q, a larger data usage variance implies that one pH has more chance to reduce the data overage and traveler T benefits from this increasing of variance σ 2 to decide a small p. On the other hand, if B + µ ≤ Q, then the pH is more likely to incur data overage with increasing variance and traveler T has to increase its price. Due to the fact that the function erfc(x) decreases with x and by checking EC ∗∗ ’s relationship with some key parameters in Theorem 3.1, we have the following proposition. Proposition 3.1: The traveler T’s benchmark (minimum) cost EC ∗∗ always increases with the sharing data amount B, the roaming fee C0 , and the overage charge unit price β for the pHs. Meanwhile, EC ∗∗ decreases with the pH density λ and the pHs’ expected leftover data amount Q − µ. 10 IV. T RAVELER ’ S O PTIMAL P RICING UNDER I NFORMATION U NCERTAINTY In this section, we aim to optimize the traveler’s pricing under pH information uncertainty to minimize his expected cost, which depends on pHs’ locations and cost distributions. We focus on a typical traveler and assume he has no overlap demand with another traveler. This is reasonable for many places less traveled, and will be generalized in Section V to include overlapped travelers. We will first consider homogeneous pHs’ monthly data usage to analyze the optimal price, and then extend to the more challenging case of heterogeneous pHs’ usage behavior. A. Pricing towards pHs of homogeneous data usage Consider a homogeneous pH case, where the monthly data usage xi ∼ N (µ, σ 2) of pH i ∈ H is i.i.d. Gaussian distributed with identical mean usage and variance among all pHs. To derive the traveler’s expected cost, we first analyze the pHs’ acceptance or not on the traveler’s request. 1) The successful probability for motivating pH data sharing: Denote by Phom (p|N) the probability for traveler T to attract a pH given the pH set H. We then have Phom (p|N) = 1 − (1 − Pr(p − Ci (xi ) ≥ ǫ))N , (13) where Pr(p − Ci (xi ) ≥ ǫ) is the probability that each pH i ∈ H accepts the data sharing request from traveler T. Note that Phom (p|N) in (13) equals zero if N = 0. Based on the traveler T’s estimated PDF of pH i’s sharing cost Ci (xi ) in (2), it follows that    0, if p < ǫ    Pr(p − Ci (xi ) ≥ ǫ) = Ω(p), if ǫ ≤ p ≤ ǫ + βB     1, if p > ǫ + βB. (14) where  p − ǫ + β(Q − B − µ)  1 √ . Ω(p) , 1 − erfc 2 2σβ (15) From (14), for any pH i ∈ H, we have the following observations: • If p < ǫ, i.e., the sharing price p is smaller than the given reservation utility, the pH i ∈ H will definitely reject the data sharing request. • If ǫ ≤ p ≤ ǫ + βB, the probability in (15) that one pH would like to share data connection increases in the price p and his expected leftover data Q − B − µ, but decreases in its unit price β for overage data. 11 • If p > βB + ǫ, i.e., the sharing price of the traveler is large enough to cover the maximal possible penalty ǫ + βB in data overage case, any pH i ∈ H will accept the data sharing request and sell data to the traveler. Considering all the possibilities of N ≥ 0, let Phom (p) be the expected probability for traveler T to successfully establish the data connection via pH, which is a function of price p offered by traveler T. By taking the expectation of Phom (p|N) over any possible variable N, we have ∞ X  (λπd2 )N exp(−λπd2 ) Phom (p) = 1 − (1 − Pr(p − Ci (xi ) ≥ ǫ))N N! N =0  = 1 − exp −λπd2 Pr(p − Ci (xi ) ≥ ǫ) . (16) By substituting (14), (16) is simplified to    0,    Phom (p) = 1 − exp (−λΩ(p)πd2 ) ,     1 − exp(−λπd2 ), if p < ǫ if ǫ ≤ p ≤ ǫ + βB (17) if p > ǫ + βB, which also has three cases as (14) and in general increases with p. 2) Traveler T’s pricing problem: Denote by EChom (p) the expected cost function for traveler T in the pH sharing service of p. Depending on the response of pHs (if any), the traveler T has either realized cost p or C0 in (6). Thus, we have EChom (p) = p × Phom (p) + C0 × (1 − Phom (p))    C0 ,    = C0 + (p − C0 ) (1 − exp (−λΩ(p)πd2 )) ,     C + (p − C )(1 − exp(−λπd2 )), 0 0 if p < ǫ if ǫ ≤ p ≤ ǫ + βB (18) if p > ǫ + βB, which decreases with leftover data (Q − B − µ) and pH density λ, and increases with pH’s overage cost β. We formulate the traveler T’s optimization problem to minimize his expected cost EChom (p) as min p s.t. EChom (p) (19a) ǫ ≤ p ≤ C0 . (19b) As shown in Fig. 3, we can prove that there is at most one unique minimal optimum, denoted by p̂, for EChom (p), and p̂ may be smaller than ǫ. Thanks to this unique-minimum feature of the objective function, we can derive the following result. 12 EChom(p) C0 0 ο ε p̂ C0 p Fig. 3. An illustration of function EChom (p) for the homogeneous pHs. Theorem 4.1: Facing the homogeneous pHs, traveler T should decide the optimal sharing price p∗0 as p∗0 = max{ǫ, p̂}, where p̂ is the unique solution to  h p − ǫ + β(Q − B − µ) i2 i h λπd2(C0 − p) √ √ 1 − exp −λπd2 Ω(p) 1 + = 0. exp − 2πσβ 2σβ Proof: See Appendix D. (20) (21) Note that in Theorem 4.1, the optimal solution p∗0 is decided to best tradeoff between the price and the sharing probability. In particular, problem (19) can be efficiently obtained by solving (21) via an one-dimension search. Similar to the benchmark case in Section III, p∗0 decreases with the pH density λ and the data quota Q for more data-plan sources, but increases with ǫ, µ, and β. If B + µ ≤ Q, i.e., the pHs’ expected data usage µ plus the sharing data amount B is within its data quota Q, then the traveler suffers from a large variance and p∗0 increases in the variance σ 2 ; otherwise, the traveler can benefit from the increase of σ 2 . B. Pricing towards pHs of heterogeneous data usage In this subsection, we extend the optimal pricing scheme to the case of the heterogeneous pHs. Specifically, we consider a set K , {1, . . . , K} of pH types according to their monthly data usage behaviors (e.g., light and heavy pH users). 13 1) The cost analysis for pHs’ data sharing: Now we extend the system model in Section II to the heterogeneous pH case. For the ease of exposition, all the pHs are again assumed to have an identical two-part tariff plan (Q, P0 , β) under the WNO. Each type of pHs in the area A follows independent PPPs with different spatial densities. Let λk and Hk be the spatial density and the appearance set of the kth pH-type, respectively. Accordingly, Nk = |Hk | is a Poisson distributed variable with mean λk πd2 . Let xk,i be the monthly data usage for pH i ∈ Hk of the kth pH-type. Again, each xk,i follows a Gaussian distribution, i.e., xk,i ∼ N (µk , σk2 ), where we √ √ reorder 0 ≤ µ1 − 2 2σ1 < . . . < µK − 2 2σK . By sharing data B, the additional cost of pH i ∈ Hk is    0,    Ck,i (xk,i ) = β(xk,i + B − Q),     βB, if 0 ≤ xk,i ≤ Q − B if Q − B ≤ xk,i ≤ Q if xk,i ≥ Q, where i = 1, . . . , Nk and k = 1, . . . , K. 2) Traveler T’s pricing problem: Under any given Hk , ∀k ∈ K, the successful probability for traveler T’s request to be accepted by pHs is3 Y Y Pr(p − Ck,i (xk,i ) < ǫ), Phet (p|{Hk }) = 1 − (22) k∈K i∈Nk which takes into account the distribution of any xk,i and Ck,i (xk,i ) in (22) by following a similar analysis as the previous Section IV.A. Similar to (15), we define   p − ǫ + β(Q − B − µk ) 1 √ , ∀k ∈ K. (23) Ωk (p) , 1 − erfc 2 2σk β √ √ If µk − 2 2σk ≤ p−ǫ + Q − B ≤ µk+1 − 2 2σk+1 for any k = 1, . . . , K − 1, it holds that β Ωj (p) = 0 for j = k + 1, . . . , K. Considering all the possible {Hk }, the successful probability for traveler T’s request to be accepted by a pH is then    0,     P Phet (p) = 1 − exp − kj=1 λj πd2 Ωj (p) ,     1 − exp − PK λ πd2 , k=1 k 3 Note that Q i∈Hk if p < ǫ √ if µk − 2 2σk ≤ p−ǫ β √ + Q − B ≤ µk+1 − 2 2σk+1 if p > ǫ + βB. Pr(p − Ck,i < ǫ) = 1 if Hk = ∅ without any pH presence to possibly share, for k = 1, . . . , K. (24) 14 EChet(p) EChet(p) Ƞ C0 İ 0 C0 p p1 0 C0 Ƞ İ p1 p p2 C0 H  E B  P2  Q  2 2V 2 (a) P1  2 2V 1 d C0  H E  Q  B d P2  2 2V 2 (b) C0  H E  Q  B t P2  2 2V 2 Fig. 4. An illustration of function EChet (p) for K = 2 types of heterogeneous pHs. As a result, traveler T’s expected cost EChet (p) is EChet (p) = p × Phet (p) + C0 × (1 − Phet (p)). (25) Note that both Phet (p) and EChet (p) are piecewise functions of p. We formulate the traveler T’s expected cost minimization problem as min p s.t. EChet (p) (26a) ǫ ≤ p ≤ C0 , (26b) which is more involved than problem (19) facing pHs with homogeneous data usage. Denote by phet the optimal price to (26). To solve (26), we first denote pk as the unique solution of k   p − ǫ + β(Q − B − µ ) 2 i h X λj πd2 (C0 − p) j √ √ = 0, exp − 1 − exp − λj πd Ωk (p) 1 + 2πσ β 2σ β j j j=1 j=1  k X 2 (27) for k = 1, . . . , K. Fig. 4 illustrates two possible cost curves of the traveler’s expected cost EChet (p) for K = 2 types of heterogenous pHs. Due to the K = 2 types of pHs in total, we show that there are now at most three local minimums for the expected cost. In particular, the following two cases are present: • Case 1: if C0 −ǫ β √ + Q − B ≤ µ1 − 2 2σ1 , then we have phet = C0 . This implies that none of pH will accept the data-plan sharing and then the traveler has to pay the roaming fee C0 in this case. 15 • √ Case 2: As shown in Fig. 4(a), if µ1 − 2 2σ1 ≤ C0 −ǫ β √ + Q − B < µ2 − 2 2σ2 , then we have EChet (p) = p + (C0 − p) exp −λ1 πd2 Ω1 (p) •  (28) for ǫ ≤ p ≤ C0 . Note that the local minimum is only decided by the light pHs in this case. √ Case 3: As shown in Fig. 4(b), if C0β−ǫ + Q − B ≥ µ2 − 2 2σ2 , then there are two local minimums for EChet (p). Note that the first local minimum is identical to that in case (i), while the second one (corresponding to a large p) relates to the two pH types. Note that the first type pHs (i.e., light pHs) are more willing to accept the data-plan sharing request from the traveler than the second type pHs (i.e., heavy pHs). When C0 is large, traveler T tries to attract even the pHs of the second type pHs in data-plan sharing, by designing a price p1 for the first type or designing a large price p2 for attract the pHs of the second type. It can √ be shown that if the sharing price satisfies that p ≤ ǫ + β(B + µ2 − Q − 2 2σ2 ), the heavy pHs will not consider data sharing. Generalizing the K = 2 pH types to general K types, we now establish the theorem on the optimal sharing price phet for traveler T to minimize his expected cost to attract heterogeneous pHs as follows. Theorem 4.2: Facing the heterogeneous K pH types, traveler T decides the optimal sharing price phet by targeting different groups of pHs: √ C −ǫ + Q − B ≤ µ − 2 2σ1 , the optimal price phet = C0 , since none of pH will accept • For 0 1 β • the data-plan sharing and traveler T has to resort to the WNO in this case. √ √ For µk − 2 2σk ≤ C0β−ǫ + Q − B ≤ µk+1 − 2 2σk+1 for k = 1, . . . , K − 1, then we have k candidates, i.e., p1 , . . . , pk , to optimal pricing depending on which candidate yields the smallest expected cost EChet(p). For example, the optimal price phet = pj implies that • traveler T targets for the first j types of pHs in data-plan sharing, where j ≤ K − 1. √ For C0β−ǫ + Q − B ≥ µK − 2 2σK , then all the K pH types are involved in data-plan sharing and we choose the optimal price phet from a total of K candidates (i.e., p1 , . . . , pK ) depending on which price candidate yields the smallest cost EChet (p), Proof: See Appendix E. Theorem 4.2 indicates that the traveler will target selective pHs and there exists a threshold √ sharing price ǫ + β(B + µk − Q − 2 2σk ) to incentivize the pHs of the kth type. When the roaming fee C0 for the traveler is smaller than the threshold, the traveler should design such a 16 reward to attract (or incentivize) the pHs of the first (k −1)th pH types. On the other hand, when the roaming fee C0 is larger than the threshold, in addition to the local optimal price determined by the pHs of the first (k − 1) types, there exists another local optimal price which is determined by the kth type pHs. V. E XTENSION OF P RICING FOR OVERLAPPED T RAVELERS In this section, we consider the coexistence of possibly multiple travelers in the area A, where multiple travelers request data-plan sharing from the common pool of pHs. On the behalf of all travelers in this pH-sharing platform, we aim to study the optimal pricing for all travelers for minimizing their total expected costs. Intuitively, as there are more travelers, we need to decide a larger uniform price in the data-plan sharing platform for attracting more pHs to share. We focus on a typical traveler T in the circle area A of radius d. He also face a random number of other M ≥ 0 travelers requesting pH connection. To capture the mobility of these travelers, we assume that M follows an independent PPP with spatial density λt > 0, and its PMF is Pr(M = m) =  (λt πd2 )m exp −λt πd2 , m = 0, 1, · · · m! (29) The sharing scheme in Fig. 2 also applies here, where each traveler offers a uniform price p decided centrally by the sharing platform to pHs and we consider each pH i with p − Ci (xi ) ≥ ǫ can reply to serve at most one traveler due to capacity limit. A. Analysis of pHs’ sharing probability For ease of exposition, we consider homogeneous pH case in this section. Denoted by Ny the random number of pHs satisfying the participation condition of p − Ci (xi ) ≥ ǫ, where Ny ≤ N. Conditioned on the presence of arbitrary Ny pHs in the vicinity, the probability for traveler T to be accepted and served by pHs is min{1, Ny /(M + 1)}. As a result, given N and M in the area of our interest, the probability for overlapping traveler T to be served successfully is   0, N =0 (30) Pmul (p|N, M) = P o n   N −Ny Ny N  N min 1, Ny , N ≥ 1, (t(p)) (1 − Ω(p)) Ny =1 M +1 Ny where Ω(p) = Pr(p − Ci (xi ) ≥ ǫ) as in (15). In the following, we focus on the reasonable range [ǫ, βB] of optimal p, and omit the two trivial cases of Pr(p − Ci (xi ) ≥ ǫ) = 0 and Pr(p − Ci (xi ) ≥ ǫ) = 1 for p < ǫ and p > βB, respectively. 17 pmul* ‫݌‬෤ p 0* C0 Ω-1(λt/λ) ε 0 λt λΩ(p0*) λ Fig. 5. An illustration of the optimal price p∗mul with respect to the nearby travelers’ density λt in approximate problem (36). Taking expectation of (30) over pH number N, we denote by Pmul (p|M) the probability for traveler T to establish the pH connection in the area A given M travelers nearby. Readily, we have ∞ X (λπd2 )N exp (−λπd2 ) N! N =0    ∞ X N X (λπd2 )N exp (−λπd2 ) Ny N (Ω(p))Ny (1 − Ω(p))N −Ny = min 1, . Ny M +1 N! N =1 N =1 Pmul (p|M) = Pmul (p|N, M) (31) y Taking further expectation of Pmul (p|M) over the random M, the successful sharing probability for traveler T is given by Pmul (p) = ∞ X M =0 Pmul (p|M) (λt πd2 )M exp (−λt πd2 ) , M! (32) which is difficult to simplify due to the “min” operation min{1, Ny /(M +1)} in each summation term. As it holds that min{1, Ny /(M + 1)} in (30) is no larger than both 1 and Ny /(M + 1), we UB1 propose the first upper bound Pmul (p) by choosing min{1, Ny /(M +1)} = 1 in each summation term in (30), i.e., UB1 Pmul (p)  ∞ X ∞ X N  X (λπd2 )N exp(−λπd2 ) (λt πd2 )M exp(−λt πd2 ) N (Ω(p))Ny (1 − Ω(p))N −Ny , N! M! Ny M =0 N =1 N =1 y 18 = 1 − exp(−λΩ(p)πd2 ), (33) which assumes a low traveler density case that pH supply is sufficient for traveler demand. On the other hand, for the high traveler density case, we also propose the second upper bound UB2 Pmul (p), by choosing min{1, Ny /(M + 1)} = Ny /(M + 1) in each summation term in (30), i.e., UB2 (p) Pmul , ∞ X ∞ X M =0 N =1 = PN Ny =1 Ny N Ny  (Ω(p))Ny (1 − Ω(p))N −Ny (λπd2 )N exp(−λπd2 ) (λt πd2 )M exp(−λt πd2 ) M +1 N! M! λΩ(p) (1 − exp(−λt πd2 )) , λt (34) which assumes that pH supply is not enough to meet the traveler demand. Note that the detailed derivations of these two upper bounds are relegated to Appendix F. These two upper bounds become identical when the traveler density λt equals the density of pHs who are willing to share. UB1 UB2 That is, Pmul (p) = Pmul (p) at λt = λΩ(p). Note that the pH sharing probability Ω(p) in (15) UB UB1 UB2 increases with p, we jointly propose a tighter upper bound of Pmul (p) = min{Pmul (p), Pmul (p)}, i.e., we approximate Pmul (p) in (32) as   1 − exp(−λΩ(p)πd2 ), if λt < λΩ(p), UB Pmul (p) =   λΩ(p) (1 − exp(−λt πd2 )), if λt ≥ λΩ(p). λt (35) B. Traveler T’s pricing for cost minimization We aim to minimize the typical traveler T’s expected cost, i.e., EC(p) = Pmul (p) × p + (1 − Pmul (p)) × C0 , which is equivalent to minimizing all travelers’ total expected cost. As Pmul (p) UB and EC(p) are not tractable, we turn to an approximate optimization problem by using Pmul (p) in (35) and a lower bound of EC(p): UB UB min Pmul (p) × p + (1 − Pmul (p)) × C0 p s.t. ǫ ≤ p ≤ C0 . (36a) (36b) UB Denote by p∗mul the optimal solution to approximate problem (36). Since Pmul (p) depends on the nearby traveler density λt , we next discuss the following three regimes for solving (36). 19 1) Low traveler density regime (0 ≤ λt ≤ λΩ(p∗0 ) with p∗0 given in Theorem 4.1): This regime includes sufficient pH supply for travelers, and is similar to Theorem 4.1 without traveler overlap. The traveler T only needs to consider the pH response without coordination with other travelers. UB UB1 In this case, Pmul (p) = Pmul (p) and p∗0 is the optimal solution to problem (36). Note that p∗mul = p∗0 is independent of λt as shown in Fig. 5. 2) Medium traveler density regime (λΩ(p∗0 ) ≤ λt ≤ λ): In this regime, the average traveler demand is more than the shared pH supply at traditional price p∗0 , and the traveler T needs to UB2 UB2 increase price p for minimizing Pmul (p) × p + (1 − Pmul (p)) × C0 until λt = λΩ(p). Lemma 5.1: Facing a medium traveler density regime (i.e., λΩ(p∗0 ) ≤ λt ≤ λ), the optimal price for (36) is p∗mul = Ω−1 (λt /λ), (37) where Ω−1 (x) is an inverse function of Ω(x) in (15). This price p∗mul decreases in pH density λ and increases with traveler density λt (see Fig. 5). Proof: See Appendix G. 3) High traveler density regime (λt > λ): In this regime, it always holds that λt > λΩ(p) due to Ω(p) < 1 and the pH supply is sufficient to meet travelers’ demand. The approximated UB2 UB2 cost for traveler T is always Pmul (p) × p + (1 − Pmul (p)) × C0 . Lemma 5.2: Facing a high traveler density regime (i.e., λt > λ), the optimal price to minimize the expected cost is obtained as p∗mul = p̃ (see Fig. 5), where p̃ is independent of λt and is the unique solution to  h p − ǫ + β(Q − B − µ) i2   p − ǫ + β(Q − B − µ)  p − C 1 0 √ √ +√ exp − = 0. (38) 1 − erfc 2 2σβ 2πσβ 2σβ By summarizing the above results, we have the following proposition. Proposition 5.1: In the overlapped traveler pricing problem in (36), the optimal price p∗mul is    p∗0 in (20), if 0 ≤ λt ≤ λΩ(p∗0 )    (39) p∗mul = Ω−1 ( λt ), if λΩ(p∗0 ) < λt ≤ λ λ     p̃ in (38), if λt > λ. 20 The expected cost of the traveler 2.5 EC(p*) w/ incomplete info., ǫ = 0.5 EC** w/ complete info., ǫ = 0.5 EC(p*) w/ incomplete info., ǫ = 0 EC** w/ complete info., ǫ=0 2 1.5 1 0.5 0 0 0.5 1 1.5 pH density, λ 2 2.5 ×10-3 Fig. 6. The traveler’s expected cost versus the homogeneous pH density λ under different reservation utility ǫ values. VI. N UMERICAL R ESULTS A. Monopoly Traveler Case In this subsection, we evaluate the proposed incentive pricing scheme in the homogeneous and the heterogeneous pH cases, as well as the benchmark cost for the social optimum under complete information. In the simulations, the pH range is set as d = 30 meters (m). Fig. 6 shows the traveler T’s expected cost versus the homogeneous pH spatial density λ under different reservation utility values, where C0 = $3, B = 0.2 GB, µ = 1.7 GB, and σ 2 = 0.01. The same two-part tariff plan (2 GB, $17, $0.013/MB) is set for all the pHs. It is corroborated that the social optimum cost EC ∗∗ under complete information is a lower bound for the proposed ones under incomplete information, and the cost gap becomes smaller as λ increases. It is observed in Fig. 6 that the traveler T’s expected cost decreases as λ increases, close to the reservation utility ǫ under either complete or incomplete information. As expected, the sharing price is below bounded by ǫ, and a larger λ implies more pHs nearby, which leads to a high probability for the traveler to incentivize pHs at a lower sharing price. Fig. 7 shows the traveler T’s expected cost versus the mean difference ∆µ between the light 21 2.4 EC(p*) for heterogenous pHs, Q = 1.8 GB EC(p*) for heterogenous pHs, Q = 2 GB The expected cost of the Traveler 2.2 EC(p*) for homogeneous pHs, Q = 1.8 GB 2 1.8 1.6 1.4 1.2 EC(p*) for homogeneous pHs, Q = 2 GB 1 0 0.2 0.4 0.6 0.8 Mean difference between the light and the heavy pHs, ∆µ 1 Fig. 7. The traveler T’s expected cost versus ∆µ under different monthly data quotas for pHs. and the heavy pHs under different monthly data quotas for all the pHs, where µ1 = µ − ∆µ/2, µ2 = µ + ∆µ/2, the light and the heavy pH densities are λ1 = λ2 = λ/2 = 2.5 × 10−4 , the reservation utility is ǫ = $0.2, and the remaining parameters are set the same as those in Fig. 6. It is observed that the expected cost is about EChet (p∗ ) = $2.2 for the homogeneous pHs at Q = 1.8 GB, while the expected cost EChet (p∗ ) for the heterogeneous pHs first decreases in ∆µ and then remains unchanged. By contrast, at Q = 2 GB, the expected cost EChom (p∗ ) is about $1.05 for the homogeneous pHs, while EChet (p∗ ) first increases in ∆µ and then remains unchanged for the heterogenous pHs. This is expected that, when pHs with a limited data quota (e.g., small Q), the traveler could benefit from the diverse of the pH data usage since the sharing cost reduces for the light pHs; on the other hand, with a large data quota, the traveler may not benefit from the diverse of the pH data usage due to the increasing of the sharing cost for the heavy pHs. It is interesting to observe in Fig. 7 that, when ∆µ becomes large, both the traveler’s expected cost with Q = 1.8 GB and that with Q = 2 GB converge to a fixed value EChet (p∗ ) = $1.58 for the heterogeneous pHs. The reason is as follows. For the very diverse heterogeneous pHs (i.e., 22 2.7 Optimal Approximate 2.6 The minimal expected cost 2.5 2.4 2.3 2.2 2.1 2 1.9 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 The density of travelers, λt 1.6 1.8 2 ×10-3 Fig. 8. The minimal expected cost EC(p) for traveler T in pH data sharing versus the nearby travelers’ density λt . when ∆µ becomes large), the traveler rewards the light pHs with a fixed price ǫ and the heavy pHs with a fixed price (ǫ + βB), respectively; thus the traveler’s expected cost is independent with the pH’s monthly data quota Q in this case. B. Overlapping Travelers Case In this subsection, we consider pH data sharing in the overlapping travelers case. Specifically, the identical two-part tariff plan (2GB, $17, $0.013/MB) and the reserved utility ǫ = $0.5 are set for all the pHs in the area A. The data usage statistics are also set to be identical for all the pHs with mean µ = 1.8 GB and variance σ 2 = 0.01. The data roaming fee is set to be C0 = $3 and the data usage amount B = 0.29 GB for each traveler. We set the pH-range and the pHs’ spatial density as d = 30 m and λ = 10−3 unit/m2, respectively. Fig. 8 shows the expected cost for traveler T in pH data sharing versus the nearby travelers’ density λt . It is observed that both the optimal and the approximate expected costs increase as λt increases. It is also observed that in Fig. 8, when 4 × 10−4 ≤ λt ≤ 10−3 , the approximate UB expected cost based on the Pmul (p) is slightly higher than the optimal one; in the small and 23 large λt regimes, the approximate approach achieves a close performance as the optimal one. It UB suggests that one can safely evaluate the minimal expected cost with Pmul (p) in the overlapping travelers case with a high traveler density. VII. C ONCLUSIONS In this paper, we investigated the optimal reward-based pricing problem in opportunistic pH data-plan sharing to reach a win-win situation. It was established that the benchmark social optimum under complete information. Taking into account the random mobility and the cost distribution of selfish pHs nearby, the optimal price was derived to minimize the traveler’s expected cost facing the homogeneous pHs, and further extended to the case of heterogeneous pHs. Lacking selfish pHs’ information and cooperation, the traveler’s expected cost is higher than that under the complete information, but the gap diminishes as the pH spatial density increases. It was shown that the traveler may or may not benefit from the diversity of pHs’ data usage behavior. As an extension, the overlapped-traveler pricing is investigated by resorting to a tractable lower bound approximation. The (near-)optimal sharing price is efficiently determined and serves as a safe approximation especially in large pH density case. We show that the optimal price increases with the density of travelers. Numerical results evaluated the effectiveness of the proposed schemes. A PPENDIX A. Proof of Lemma 2.1 First consider the case of c = βB. According to (2), the range of Ci (xi ) is [0, βB]. Hence, it always holds that Ci (xi ) ≤ βB, i.e., Pr(Ci (xi ) ≤ βB) = 1. Next, based on (2), if 0 ≤ xi ≤ Q − B, then Ci (xi ) = 0 and we have Pr(Ci (xi ) = 0) = Pr(0 ≤ xi ≤ Q − B) Z Q−B = φ(xi )dxi 0 = Φ(Q − B). (40) 24 As in (2), for Q − B < xi < Q, it holds that Ci (xi ) = β(xi + B − Q). By setting Ci (xi ) = c for 0 < c < βB, we have xi = c/β + Q − B. Hence, under given 0 < c < βB, it follows that Pr(Ci (xi ) ≤ c) = Pr xi < = Z  c +Q−B β c +Q−B β φ(xi )dxi 0 =Φ  c +Q−B . β (41) Based on (40) and (41), the probability of Ci (xi ) = 0 becomes identical to that of Ci (xi ) ≤ c for 0 < c < βB. Hence, it is immediate that Pr(Ci (xi ) ≤ c) = Φ( βc + Q − B) for 0 ≤ c < βB. Until now, Lemma 2.1 is proved. B. Proof of Lemma 3.1 Under the given pH set, it is required to obtain the expected value of ǫ + E{C1 }, where the expectation is taken over the variable C1 . With (9) and (10), we derive E{C1 } under N ≥ 1: Z βB− xfC1 (x)dx + βB × Pr(C1 = βB) E{C1 } = 0+ βB −   x N xd 1 − (1 − Φ( + Q − B)) + βB(1 − Φ(Q))N = β 0+  βB− Z βB−    x x N N − 1 − (1 − Φ( + Q − B)) dx + βB(1 − Φ(Q))N = x 1 − (1 − Φ( + Q − B)) β β + 0 0+ Z βB− x = (1 − Φ( + Q − B))N dx, (42) β 0+ Z where the third equality of (42) holds from the integration by parts. In addition, as (1 − Φ( βx + Q − B))N is continuous and smooth for x ≥ 0, it follows from the limit theorem that Z βB x E{C1 } = (1 − Φ( + Q − B))N dx. β 0 (43) C. Proof of Theorem 3.1 Based on Lemma 3.1 and considering all the possibilities of N (including N = 0), the expected cost for traveler T is given by [23] EC ∗∗ = C0 × Pr(N = 0) + ∞ X n=1 p(N) × Pr(N = n). (44) 25 For (44), with the probability of N, we have the following equality chain: Z βB ∞   (λπd2 )n X x ∗∗ −λπd2 exp(−λπd2 )dx EC = C0 e + ǫ+ (1 − Φ( + Q − B))n β n! 0 n=1 Z βB X ∞ x (λπd2 )n −λπd2 2 + ǫ(1 − exp(−λπd )) + = C0 e (1 − Φ( + Q − B))n exp(−λπd2 )dx β n! 0 n=1    Z βB  x 2 2 2 = ǫ + (C0 − ǫ) exp(−λπd ) + exp(−λπd ) exp λπd (1 − Φ( + Q − B)) − 1 dx β 0      Z βB √ λπd2 erfc x/β+Q−B−µ − 2 2σ  dx, = ǫ + (C0 − ǫ − βB) exp(−λπd2 ) + exp  (45) 2 0 where the second equality in (45) follows from switching the summarization and the integration √ operations and the last equality in (45) holds from the fact 1 − Φ(x) = erfc(x/ 2)/2. D. Proof of Theorem 4.1 First, we prove that p̂ is the unique solution of (21). Note that the first-order derivative of EChom (p) is ′ EChom (p) = 1 − exp i   h √   λπd2 erfc p−ǫ+β(Q−B−µ) − 2 2σβ 2  λπd2 (C0 − p) √ 1+ exp − 2πσβ p−ǫ β + Q − B − µ 2  √ . 2σ (46) ′ It then holds that EChom (p̂) = 0. Let p = p̂ − ∆, where 0 < ∆ ≤ p̂; i.e., 0 ≤ p < p̂. It is going ′ to show that EChom (p) < 0. To this end, from the monotonically decreasing function of erfc(x), it follows that erfc p−ǫ β +Q−B−µ √ 2σ ! > erfc p̂−ǫ β +Q−B−µ √ 2σ ! . (47) In addition, it holds that 2 !  2 !  p̂ − ǫ + β(Q − B − µ) p − ǫ + β(Q − B − µ) √ √ > (C0 − p̂) exp − (C0 − p) exp − 2σβ 2σβ (48)  h i2  x−ǫ+β(Q−B−µ) √ due to the nonnegative first-order derivative of function (C0 − x) exp − for 2σβ ′ 0 ≤ x ≤ C0 . Hence, by (47) and (48), it is follows that EChom (p) < 0 for p = p̂ − ∆. 26 On the other hand, let p = p̂ + ∆, where 0 < ∆ ≤ C0 − p̂; i.e., p̂ < p ≤ C0 . Likewise, it holds that the inequality (47) and 2 !  2 !  p̂ − ǫ + β(Q − B − µ) p − ǫ + β(Q − B − µ) √ √ < (C0 − p̂) exp − . (C0 − p) exp − 2σβ 2σβ (49) ′ ′ It then obtains that EChom (p) > 0 for p = p̂ + ∆. It is worth noting that EChom (C0 ) > 0; this ′ implies that p̂ < C0 . Until now, it proves that p̂ is the unique solution to EChom (p) = 0. Furthermore, the optimal pricing problem (19) requires the sharing price p to meet ǫ ≤ p ≤ C0 . If p̂ ≤ ǫ, then EChet (p) monotonically increases in p for ǫ ≤ p ≤ C0 ; hence, p∗0 = ǫ. On the other hand, if p̂ > ǫ, it holds that p∗0 = p̂. We then obtain that p∗0 = max(ǫ, p̂) and complete the proof of Theorem 4.1. E. Proof of Theorem 4.2 For k = 1, . . . , K, we define the following function K   X gk (p) , C0 + (p − C0 ) 1 − exp(− λk πd2 Ωk (p)) , (50) k=1 where ǫ ≤ p ≤ C0 . The first-order derivative of gk (p) with p is given by   2  p−ǫ+β(Q−B−µ) # " 2 √ k k λj πd (C0 − p) exp −  X  X 2σj β √ . gk′ (p) = 1 − exp − λj πd2 Ωj (p) 1 + 2πσ β j j=1 j=1 (51) Suppose that pk satisfies gk′ (pk ) = 0. Next we prove that pk is the unique solution to gk (p) = 0. √ )2 ), ∀k, decrease as p increases, it follows that each term (C0 − Since exp(−( p−ǫ+β(Q−B−µ) 2σj β   2  P p−ǫ+β(Q−B−µ) √ p) exp − decreases with p ≥ ǫ. Combining with exp(− kj=1 λj πd2 Ωj (p)), 2σ β j it can be verified that gk′ (p) is an increasing function with p ∈ [ǫ, C0 ]. Hence, it holds that pk is the unique solution to gk′ (p) = 0 and pk is the optimal price to minimize gk (p) for k = 1, . . . , K. √ √ If µk − 2 2σk ≤ C0β−ǫ + Q − B ≤ µk+1 − 2 2σk+1 , then feasible price for traveler T in √ problem (26) cannot be larger than ǫ + β(µk+1 + B − Q − 2 2σk+1 ) for k = 1, . . . , K − 1. In this case, there are k piecewise probability functions for traveler T in pH data-plan sharing. Note that each piecewise probability function yield an expected cost gk (p) for traveler T. Then the expected cost minimization problem with gk (p) admits a unique optimal solution pk for √ k = 1, . . . , K − 1. If C0β−ǫ + Q − B ≥ µk − 2 2σk , then traveler T attracts all the K pH types 27 and we have K candidates {p1 , . . . , pK }. By choosing the optimal one which yields the smallest expected cost for traveler T among of these candidate ones, Theorem 4.2 follows. UB1 UB2 F. Derivations of Pmul (p) and Pmul (p) We present the detailed calculation for obtaining the upper bound expressions in (33) and (34) as follows. UB1 Pmul (p)  ∞ X ∞ X N  N −Ny (λπd2 )N exp(−λπd2 ) (λ πd2 )M exp(−λ πd2 ) Ny  X N  t t 1 − Ω(p) Ω(p) = N! M! Ny M =0 N =1 N =1 y ∞ X ∞ X (λπd2)N exp(−λπd2 ) (λt πd2 )M exp(−λt πd2 ) N! M! M =0 N =1 ! ! ∞ ∞ ∞ 2 M X X (λ πd ) (λπd2 )N X ((1 − Ω(p))λπd2 )N t exp(−λπd2 ) exp(−λt πd2 ) − = N! N! M! M =0 N =1 N =1  = exp(λπd2 ) − exp((1 − Ω(p))λπd2 ) exp(−λπd2 ) = (1 − (1 − Ω(p))N ) = 1 − exp(−λΩ(p)πd2 ), (52) P 2 N 2 where the fourth equality of (52) follows from the fact that ∞ N =1 (λπd ) /N! = exp(λπd ) − P∞ P∞ 2 M 2 N 2 1, M =0 (λt πd ) /M! = N =1 ((1 − Ω(p))λπd ) /N! = exp((1 − Ω(p))λπd ) − 1, and exp(λt πd2 ). UB2 Likewise, for Pmul (p) in (34), we have UB2 Pmul (p) PN N Ny (Ω(p))Ny (1 − Ω(p))N −Ny (λπd2)N exp(−λπd2 ) (λt πd2 )M exp(−λt πd2 ) = M +1 N! M! M =0 N =1 ! ! ∞ ∞ X X (λt πd2 )M exp(−λt πd2 ) (λπd2 )N exp(−λπd2 ) = Ω(p)N N! (M + 1)! N =1 M =0 ! ! ∞ ∞ X (λπd2 )N −1 exp(−λt πd2 ) X (λt πd2 )M +1 2 2 = λΩ(p)πd exp(−λπd ) (N − 1)! λt πd2 (M + 1)! M =0 N =1 ∞ ∞ X X Ny =1 Ny  = λΩ(p)πd2 exp(−λπd2 ) exp(λπd2 ) exp(−λt πd2 ) (exp(λt πd2 ) − 1) λt πd2 λΩ(p) (1 − exp(−λt πd2 )), (53) λt P∞ 2 M +1 /(M + 1)! = where the fourth equation of (53) holds from the fact that M =0 (λt πd ) P ∞ exp(λt πd2 ) − 1 and N =1 (λπd2 )N −1 /(N − 1)! = exp(λπd2 ). = 28 G. Proof of lemma 5.1 First, we prove the function f (p) has a unique solution p̃, where p ∈ [ǫ, C0 ] and f (p) is defined as UB2 UB2 f (p) , Pmul (p) × p + (1 − Pmul (p)) × C0 = C0 + λΩ(p)(1 − exp(−λt πd2 )) (p − C0 ). λt (54) (55) The first-order derivative of f (p) is given as   h p − ǫ + β(Q − B − µ) i2  λ(1 − exp(−λt πd2 ))  1 √ (p − C0 ) . exp − f ′ (p) = Ω(p) + √ λt 2πσβ 2σβ (56) Note that f ′ (p̃) =0. Let p = p̃ − ∆ where 0 < ∆ < p̃ − ǫ. Then ǫ < p < p̃. The first-order h i2  √ derivative of exp − p−ǫ+β(Q−B−µ) (p − C0 ) of p ∈ [ǫ, C0 ] is positive, i.e., 2σβ 2 ! p − ǫ + β(Q − B − µ) p − ǫ + β(Q − B − µ) √ >0 (57) exp − + (C0 − p) σ2β 2 2σβ  h  h i2  i2  p̃1 −ǫ+β(Q−B−µ) p−ǫ+β(Q−B−µ) √ √ (p − C0 ) < exp − (p̃1 − C0 ). it follows that exp − 2σβ 2σβ  Furthermore, together with the monotonically increasing t(p) of p, we have f ′ (p) < f ′ (p̃). Likewise, letting p = p̃ + ∆ where 0 < ∆ < C0 − p̃, we have f ′ (p) > f ′ (p̃) = 0. Based on above analysis for f ′ (p) of p ∈ [ǫ, C0 ], it is verified that that p̃ is the unique solution to minimize f (p). Then, by checking f ′ (Ω−1 (λt /λ)) < 0, it follows that p̃ > Ω−1 (λt /λ). Taking into account the constraint of ǫ ≤ p ≤ Ω−1 (λt /λ), the optimal solution is thus p∗mul = Ω−1 (λt /λ). R EFERENCES [1] F. Wang, L. Duan, and J. Niu, “Pricing for opportunistic data sharing via personal hotspot,” accepted by Proc. IEEE GLOBECOM, Singapore, Dec. 2017, pp. 1–6. [2] GSMA, “Mobile SMS and data roaming explained,” Tech. Rep., 2012. [Online]. Available: https://www.gsma.com/aboutus/ wp-content/uploads/2012/03/smsdataroamingexplained.pdf [3] ITU, “International mobile roaming services: Facilitating competition and protecting users,” Tech. Rep., 2013. [4] O. Wyman, “The roaming tsunami: Disrupting the whole industry,” Tech. Rep., 2016. [5] Cisco, “Cisco visual networking index: Global mobile data traffic forecast update, 2016-2021,” White Paper, Mar. 2017. [6] Informa, “LTE roaming: Global market status and drivers for growth,” Tech. Rep., 2013. [7] Visitsingapore. “Wi-Fi and Connectivity,” Tech. Rep., 2017. [Online]. Available: http://www.visitsingapore.com/ travel-guide-tips/tourist-information/connectivity-wifi-prepaid/ [8] R. Chaturvedi and D. Hunsaker, “Personal wireless hotspot,” U.S. Patent D664443 S1, Jul. 24, 2012. 29 [9] L. Duan, T. Kubo, K. Sugiyama, J. Huang, T. Hasegawa, and J. Walrand, “Motivating smartphone cooperation in data acquisition and distributed computing,” IEEE Trans. Mobile Comput., vol. 13, no. 10, pp. 2320–2333, Oct. 2014. [10] S. Sen, C. Joe-Wong, S. Ha, and M. Chiang, “A survey of smart data pricing: Past proposals, current plans, and future trends,” ACM Computing Surveys (CSUR), vol. 46, no. 2, p. 15, 2013. [11] T. Luo, S. Kanhere, J. Huang, S. Das, and F. Wu, “Sustainable incentives for mobile crowdsensing: Auctions, lotteries, and trust and reputation systems,” IEEE Commun. Mag., vol. 55, no. 3, pp. 68–74, Mar. 2017. [12] Y. Zhang, Y. Gu, M. Pan, N. H. Tran, Z. Dawy, and Z. Han, “Multi-dimensional incentive mechanism in mobile crowdsourcing with moral hazard,” to appera in IEEE Trans. Mobile Comput., 2017. [13] Y. Zhang, C. Jiang, L. Song, M. Pan, Z. Dawy, and Z. Han, ”Incentive mechanism for mobile crowdsourcing using an optimized tournament model,” IEEE J. Sel. Areas Commun., vol. 35, No. 4, pp. 880–892, Apr. 2017. [14] J. Musacchio and J. Walrand, “WiFi access point pricing as a dynamic game,” IEEE/ACM Trans. Netw., vol. 14, no. 2, pp. 289–301, Apr. 2006. [15] L. Duan, J. Huang, and B. Shou, “Pricing for local and global Wi-Fi markets,” IEEE Trans. Mobile Comput., vol. 14, no. 5, pp. 1056–1070, May 2015. [16] J. Yu, M. H. Cheung, J. Huang, and H. V. Poor, “Mobile data trading: A behavioral economics perspective,” in Proc. IEEE WiOpt, Mumbai, India, May 2015, pp. 363–370. [17] L. Zheng, C. Joe-Wong, C. W. Tan, S. Ha, and M. Chiang, “Secondary markets for mobile data: Feasibility and benefits of traded data plans,” in Proc. IEEE INFOCOM, Hong Kong, China, Apr./May 2015, pp. 1580–1588. [18] Y. Guo, L. Duan, and R. Zhang, “Optimal pricing and load sharing for energy saving with cooperative communications,” IEEE Trans. Wireless Commun., vol. 15, no. 2, pp. 951–964, Feb. 2016. [19] X. Wang, L. Duan, and R. Zhang, “User-initiated data plan trading via a personal hotspot market,” IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7885–7898, Nov. 2016. [20] L. M. Feeny and M. Nilsson, “Investigating the energy consumption of a wireless network interface in an ad hoc networking environment,” in Proc. IEEE INFOCOM, Anchorage, USA, Apr. 2001, pp. 1548–1557. [21] J. G. Andrews, R. Ganti, M. Haenggi, N. Jindal, and S. Weber, “A primer on spatial modeling and analysis in wireless networks,” IEEE Commun. Mag., vol. 48, no. 11, pp. 156–163, Nov. 2010. [22] J. Laffont and D. Martimort, The Theory of Incentives: The Principal-Agent Model, Princetion, NJ: Princeton Univ. Press, 2001. [23] P. Fishburn, Utility Theory for Decidsion Making, New York: Wiley, 1970. [24] H. A. David, Order Statistics, New York: Wiley, 1981.
7cs.IT
Two sources of poor coverage of confidence intervals after model selection Paul Kabaila∗ and Rheanna Mainzer arXiv:1711.01739v1 [math.ST] 6 Nov 2017 Department of Mathematics and Statistics, La Trobe University, Australia ABSTRACT We compare the following two sources of poor coverage of post-model-selection confidence intervals: the preliminary data-based model selection sometimes chooses the wrong model and the data used to choose the model is re-used for the construction of the confidence interval. Keywords: Coverage probability, confidence interval, model selection * Corresponding author. Department of Mathematics and Statistics, La Trobe University, Victoria 3086, Australia. Tel.: +61 3 9479 2594; fax +61 3 9479 2466. E-mail address: [email protected]. 1 1. Introduction In applied statistics a model is often chosen using a data-based model selection procedure. Such procedures include hypothesis testing, minimizing a criteron such 2 (Sheather, 2008). Inferences as AIC or BIC, and maximizing a criterion such as Radj made using a model that has been selected using data-based model selection can be misleading (see e.g. Chatfield 1995). Suppose that the inference of interest is a confidence interval for a parameter θ, with minimum coverage 1 − α. Typically, a confidence interval for θ that is constructed using a model chosen by some form of data-based model selection procedure will have poor coverage properties (Kabaila, 2009, Leeb and Pötscher, 2005). This poor coverage probability can be attributed to two different sources. The first source is that the preliminary model selection procedure sometimes chooses the wrong model. When this happens, the incorrect model is treated as if it were the true model during the construction of the confidence interval. The second source is that the data used to choose the model is re-used for the construction of the confidence interval, without due acknowledgment. It is important to know the relative importance of these two sources of poor coverage. Miller (2002, Ch.6) considers the problems that arise for parameter estimation due to the re-use of the data used to choose a model. Hurvich and Tsai (1990) examine the coverage probability of confidence regions for the regression parameter vector in linear regression models after data-based model selection. They acknowledge the problem of re-using the same data and propose a possible solution: One possible solution ... is to actually perform model selection and inference on separate parts of the data. ... This procedure may seem wasteful, but we note that linear regression model selection can now be performed on quite small samples. ... Thus the available data set need not be split in half; instead, a small part may be used for the model selection stage and the remainder may be used for inference. Thus if the second source of poor coverage (re-use of the data) dominates the first source (possible choice of the wrong model) we could split the data into two parts: the first part could be used to choose the model and the second to construct a confidence interval. By allocating most of the data to the second part, we might hope to avoid constructing confidence intervals that are too wide. If, on the other hand, the first source of poor coverage dominates the second source, this data splitting strategy would not correct the defect of poor coverage 2 probability. Instead we would need to consider other methods to improve the coverage probability. We expect that problems caused by possibly choosing a wrong model would be ameliorated by using an appropriate data-weighted average of models in place of putting all of the weight on a model that may be wrong. In other words, if the first source of poor coverage dominates the second source then some form of frequentist model averaged confidence interval (Buckland, Burnham and Augustin, 1997, Hjort and Claeskens 2003, Fletcher and Turek, 2011, Turek and Fletcher, 2012, Efron, 2014) would be expected to greatly improve the coverage probability. In this paper we examine the extent to which each of these two sources contributes to the difference between true and nominal coverage probability. The basic argument presented in Section 2 gives us the means to identify Dwm , the deficit in coverage probability due to the possible choice of the wrong model, and Drd , the deficit in coverage probability due to re-use of the data. The coverage probability of the confidence interval, constructed after a preliminary data-based model selection, is given by 1 − α − Dwm − Drd . In Section 3, we apply this basic argument to the simple scenario that preliminary data-based choice is between two nested linear regression models. The simpler model is obtained from the full model by setting τ , a given linear combination of the regression parameters, to 0. The parameter of interest θ is a distinct given linear combination of the regression parameters. Let θb and τb denote the least squares estimators of θ and τ , respectively. We consider preliminary model selection using a t test of the null hypothesis τ = 0 against the alternative hypothesis τ 6= 0. Let n denote the dimension of the response vector and p denote the dimension of the regression parameter vector. Define the unknown parameter γ = τ /(var(b τ ))1/2 and b τb). the known correlation ρ = corr(θ, Figure 1 is for the confidence interval, with nominal coverage 0.95, constructed after a preliminary t test with size 0.1, when n−p = 40 and |ρ| ∈ {0.3, 0.6, 0.8}. The top panel shows graphs of Dwm , the deficit in coverage due to the choice of the wrong model, as a function of |γ|. The middle panel shows graphs of Drd , the deficit in coverage due to the re-use of the data, as a function of |γ|. The bottom panel shows graphs of the coverage probability of this post-model-selection confidence interval, given by 1 − α − Dwm − Drd , as a function of |γ|. We do not include the value ρ = 0 in these graphs since, to a very good approximation, Dwm = 0, Drd = 0 and the 3 Dwm for m = 40 0.30 | ρ | = 0.8 | ρ | = 0.6 | ρ | = 0.3 0.25 0.20 0.15 0.10 0.05 0.00 0 1 2 3 4 5 |γ| Drd for m = 40 0.30 | ρ | = 0.8 | ρ | = 0.6 | ρ | = 0.3 0.25 0.20 0.15 0.10 0.05 0.00 0 1 2 3 4 5 |γ| Coverage Probability (1 − α − Dwm − Drd) for m = 40 0.95 0.90 0.85 0.80 0.75 | ρ | = 0.8 | ρ | = 0.6 | ρ | = 0.3 0.70 0.65 0 1 2 3 4 5 |γ| Figure 1: This figure is for the confidence interval, with nominal coverage 0.95, constructed after a preliminary t test with size 0.1, when n − p = 40 and |ρ| ∈ {0.3, 0.6, 0.8}. The top panel shows graphs of Dwm , the deficit in coverage due to the choice of the wrong model, as a function of |γ|. The middle panel shows graphs of Drd , the deficit in coverage due to the re-use of the data, as a function of |γ|. The bottom panel shows graphs of the coverage probability of this confidence interval, given by 1 − α − Dwm − Drd , as a function of |γ|. 4 coverage probability is 0.95 for all |γ|. This figure shows that, for size 0.1 of the preliminary t test, n − p = 40 and |ρ| ∈ {0.6, 0.8}, Dwm , the deficit in coverage due to the choice of the wrong model, dominates Drd , the deficit in coverage due to the re-use of the data. An extensive numerical investigation of Dwm and Drd , covering a wide range of values of n − p, ρ, 1 − α and size of the preliminary t test, is reported in the Supplementary Material and summarised in Section 4. 2. The basic argument Suppose that we observe a random n-vector of responses Y for a general regression model, with given values of the explanatory variables. We denote this “full model” by M. Suppose that the parameter of interest is the scalar θ. Also suppose that a given preliminary data-based model selection is used to select between the full model and a specified set of submodels. Assuming that the selected model had been given to us a priori as the true model, we construct the usual confidence interval for θ, with nominal coverage 1 − α. Let K denote the resulting post-model-selection confidence interval. To identify the deficit in coverage probability due to the possible choice of the wrong model and the deficit in coverage probability due to re-use of the data, we consider the following hypothetical situation. Suppose that a genie provides an observation of Y ∗ the random n-vector of responses for a replicate (i.e. an independent realization) of the original experiment that gave rise to Y , with the same values of the explanatory variables. Also suppose that the same preliminary model selection based on Y is used. Assuming that the selected model had been given to us a priori as the true model, we construct the usual confidence interval for θ, with nominal coverage 1 − α using the genie’s data Y ∗ . Let K ∗ denote the resulting confidence interval. Observe that K and K ∗ are influenced by exactly the same, possibly wrong, choice of model based on Y . However, unlike K, the confidence interval K ∗ is not influenced by the re-use of the data for its construction. Because of this difference between K and K ∗ , we are able to separate out these two sources of poor coverage as follows. The coverage probability of K is       P θ ∈ K = (1 − α) − (1 − α) − P θ ∈ K ∗ − P θ ∈ K ∗ − P θ ∈ K = (1 − α) − Dwm − Drd , 5  where Dwm = (1 − α) − P θ ∈ K ∗ is the deficit in coverage probability due solely   to possibly choosing the wrong model and Drd = P θ ∈ K ∗ − P θ ∈ K is the deficit in coverage probability due solely to re-using the data. 3. The scenario of two nested linear regression models Consider the multiple linear regression model Y = Xβ + ε (1) where Y is a random n-vector of responses, X is a known n × p matrix with linearly independent columns, β is an unknown parameter p-vector and ε ∼ N (0, σ 2 In ), where σ 2 is an unknown positive parameter. Suppose that n > p. We refer to (1) as the full model and denote it by M. Suppose that the quantity of interest is θ = a> β, where a is a known p-vector (a 6= 0). Also suppose that the inference of interest is a 1 − α confidence interval for θ. Define the parameter τ = c> β − r, where c is a known p-vector, the number r is specified and the vectors a and c are linearly independent. Let βb denote the least squares estimator of β. Then θb = a> βb is the least squares estimator of θ and τb = c> βb − r is the least squares estimator of τ . Let m = n − p and σ b2 =  b > (Y − X β) b m. Also let vθ = var(θ)/σ b 2 , vτ = var(b b τb) (Y − X β) τ )/σ 2 , ρ = corr(θ,  1/2 and γ = τ (σ vτ ) We consider preliminary model selection using a t test of the null hypothesis H0 : τ = 0 against the alternative hypothesis HA : τ 6= 0. The test statistic is τb T = 1/2 . σ b vτ Define the quantile tm,a by P (T ≤ tm,a ) = a for T ∼ tm . Suppose that we reject H0 if |T | ≥ tm, 1−eα/2 ; otherwise we accept H0 . This test has size α e. Let Mτ =0 denote the model M for τ = 0. If we reject H0 then we choose the model Mτ =0 ; otherwise we choose the model M. 3.1 The confidence interval K The usual 1 − α confidence interval for θ based on the full model M is  1/2  I = θb ± tm, 1−α/2 vθ σ b , 6 where [a ± b] = [a − b, a + b] (b ≥ 0). Under the model Mτ =0 the usual 1 − α confidence interval for θ is "  1/2 # 2 2  m σ b + (b τ /v ) 1/2 τ 1/2 1/2 J = θb − ρ vθ (b 1 − ρ2 . τ /vτ1/2 ) ± tm+1,1−α/2 vθ m+1 Let K be the confidence interval, with nominal coverage 1 − α, for θ constructed as follows. We use the data Y to choose the model. We then construct the confidence interval for θ using the same data Y based on the assumption that this model had been given to us a priori as the true model. Therefore K is the post-model-selection confidence interval, with nominal coverage 1 − α, for θ. In other words, ( I if |T | ≥ tm,1−eα/2 K= J if |T | < tm,1−eα/2 . 3.2 The confidence interval K ∗ Now suppose that a genie provides the following replicate (i.e. an independent realization) of the original experiment that gave rise to Y , with the same values of the explanatory variables: Y ∗ = Xβ + ε∗ , where Y ∗ is a random n-vector of responses and ε∗ ∼ N (0, σ 2 In ). Note that ε and ε∗ are independent. Let βb∗ , θb∗ and τb∗ denote the least squares estimator of β, θ and τ , respectively, based on the genie’s data Y ∗ . Let (b σ ∗ )2 = (Y ∗ −X βb∗ )> (Y ∗ −X βb∗ )/m. The usual 1 − α confidence interval for θ based on M, and using only the genie’s data Y ∗ , is  1/2 ∗  b . I ∗ = θb∗ ± tm, 1−α/2 vθ σ The usual 1 − α confidence interval for θ based on Mτ =0 , and using only the genie’s data Y ∗ , is "  1/2 # ∗ 2 ∗ 2  m (b σ ) + ((b τ ) /v ) 1/2 τ 1/2 1/2 1 − ρ2 J ∗ = θb∗ − ρ vθ (b τ ∗ /vτ1/2 ) ± tm+1,1−α/2 vθ . m+1 Let K ∗ be the confidence interval, with nominal coverage 1−α, for θ constructed as follows. We use the original data Y to choose the model i.e. we use the test statistic T . We then construct the confidence interval for θ using the genie’s data Y ∗ based on the assumption that this model had been given to us a priori as the true model. In other words, ( I∗ K∗ = J∗ if |T | ≥ tm, 1−eα/2 if |T | < tm, 1−eα/2 . 7 3.3 Computationally convenient formulas for the deficits Dwm and Drd By the law of total probability, P (θ ∈ K) = P (θ ∈ J, |T | < tm, 1−eα/2 ) + P (θ ∈ I, |T | ≥ tm, 1−eα/2 ) = P (θ ∈ J, |T | < tm, 1−eα/2 ) + P (θ ∈ I) − P (θ ∈ I, |T | < tm, 1−eα/2 ) = (1 − α) + P (θ ∈ J, |T | < tm, 1−eα/2 ) − P (θ ∈ I, |T | < tm, 1−eα/2 ), since P (θ ∈ I) = 1 − α. By the law of total probability, and the independence between the original data Y and the genie’s data Y ∗ , P (θ ∈ K ∗ ) = P (θ ∈ J ∗ , |T | < tm, 1−eα/2 ) + P (θ ∈ I ∗ , |T | ≥ tm, 1−eα/2 ) = P (θ ∈ J ∗ ) P (|T | < tm, 1−eα/2 ) + P (θ ∈ I ∗ ) P (|T | ≥ tm, 1−eα/2 ) = P (θ ∈ J) P (|T | < tm, 1−eα/2 ) + P (θ ∈ I) 1 − P (|T | < tm, 1−eα/2 )   = P (θ ∈ I) + P (|T | < tm, 1−eα/2 ) P (θ ∈ J) − P (θ ∈ I)  = (1 − α) + P (|T | < tm, 1−eα/2 ) P (θ ∈ J) − (1 − α) . Therefore, the deficit in coverage probability due solely to possibly choosing the wrong model using the original data Y is, by definition, Dwm = (1 − α) − P (θ ∈ K ∗ )  = P (|T | < tm, 1−eα/2 ) (1 − α) − P (θ ∈ J) , and the deficit in coverage probability due solely to re-using the original data Y is, by definition, Drd = P (θ ∈ K ∗ ) − P (θ ∈ K)   = P θ ∈ I, |T | < tm, 1−eα/2 − P θ ∈ J, |T | < tm, 1−eα/2 + P |T | < tm, 1−eα/2    P θ ∈ J − (1 − α) . Let W = σ b/σ and note that W has the same distribution as (R/m)1/2 , where   1/2 1/2 R ∼ χ2m . Also let G = (θb − θ) (σ vθ ) and H = τb (σ vτ ). Note that W and (G, H) are independent, with  G H      0 1 ρ ∼N , . γ ρ 1 8 (2) Thus T = H/W has a noncentral t distribution with degrees of freedom m and  noncentrality parameter γ. Therefore we can easily evaluate P |T | < tm, 1−eα/2 . We now show how the other probabilities that appear in the expressions for Dwm and Drd are readily evaluated. Let φ and Φ denote the N (0, 1) pdf and cdf, respectively.   Also let fW denote the pdf of W . Define Ψ(x, u) = Φ g2 (x, u) − Φ g1 (x, u) , where  g1 (x, u) = −tm+1,1−α/2 x m + u2 m+1 1/2 + ργ (1 − ρ2 )1/2 and  g2 (x, u) = tm+1,1−α/2 x m + u2 m+1 1/2 + ργ (1 − ρ2 )1/2 . The following theorem gives computationally convenient formulas for the probabilities that we need to evaluate to find Dwm and Drd . Theorem 1. Z ∞ Z ∞ P (θ ∈ J) = Ψ(w, h/w) φ(h − γ) dh fW (w) dw, 0 −∞  P θ ∈ J, |T | < tm, 1−eα/2 is equal to Z ∞ Z 1 tm,1−eα/2 0  Ψ w, tm,1−eα/2 y φ(tm,1−eα/2 w y − γ) dy w fW (w) dw, −1  and P θ ∈ I, |T | < tm, 1−eα/2 is equal to   G H P −tm, 1−α/2 ≤ ≤ tm, 1−α/2 , −tm, 1−eα/2 ≤ ≤ tm, 1−eα/2 . W W  This probability is readily evaluated using the fact that G/W, H/W has a bivariate noncentral t distribution (Kshirsagar definition). Theorem 1 is proved in the appendix. It follows from Theorem 1 that the two deficits Dwm and Drd , and the coverage probability P (θ ∈ K) depend only on one unknown parameter γ, and four known quantities 1−α, α e, ρ and m. Using Theorem 1, it is simple to calculate Dwm , Drd and P (θ ∈ K). Our computer programs are written in R. We use the pmvt function in the R package mvtnorm to evaluate the bivariate noncentral t distribution (Kshirsagar definition). The following theorem reduces the number of combinations of ρ and γ that we need to consider in order to understand the properties of the Dwm , Drd and P (θ ∈ K) functions. 9 Theorem 2. For any given ρ, the deficits Dwm , Drd and the coverage probability P (θ ∈ K) are even functions of γ. For any given γ, the deficits Dwm , Drd and the coverage probability P (θ ∈ K) are even functions of ρ. Theorem 2 will be true if each of the probabilities appearing in the expressions for Dwm and Drd are, for any given ρ, even functions of γ and, for any given γ, even functions of ρ. Appendix C of Kabaila and Giri (2009) provides a proof that the coverage probability P (θ ∈ K) is, for any given ρ, an even function of γ and, for any given γ, an even function of ρ. It follows directly from this proof that P (θ ∈ J, |T | < tm,1−eα/2 ) and P (θ ∈ I, |T | < tm,1−eα/2 ) are, for any given ρ, even functions of γ and, for any given γ, even functions of ρ. Using the same arguments as those given in Appendix C of Kabaila and Giri (2009), it is straightfoward to show that P (θ ∈ J) and P (|T | < tm,1−eα/2 ) are, for any given ρ, even functions of γ and, for any given γ, even functions of ρ. 4. Extensive numerical investigation of the deficits Dwm and Drd In this section we summarise the extensive numerical investigation of Dwm , the deficit in coverage due to the choice of the wrong model, and Drd , the deficit in coverage due to the re-use of the data, reported in the Supplementary Material. Primarily our interest lies in the cases where the coverage probability of K, minimized over γ, i.e. minγ P (θ ∈ K), is substantially below 1 − α. In the Supplementary Material we compute minγ P (θ ∈ K) for m ∈ {1, 2, 5, 10, 40, 100}, |ρ| ∈ {0, 0.3, 0.6, 0.8}, 1 − α ∈ {0.9, 0.95, 0.98} and α e ∈ {0.02, 0.05, 0.1}. We find that Dwm dominates Drd when minγ P (θ ∈ K) is substantially below 1 − α. In other words, the main cause for minγ P (θ ∈ K) being substantially below 1 − α is that the preliminary t test sometimes chooses the wrong model. 5. Discussion We have defined and shown how to assess Dwm , the deficit in coverage probability due to the possible choice of the wrong model, and Drd , the deficit in coverage probability due to re-use of the data. The numerical results presented in Figure 1 and the Supplementary Material show that the main cause for the coverage probability, minimized over γ, being substantially below nominal is that the preliminary t test sometimes chooses the wrong model. This suggests that data-splitting will not help improve the coverage probability, but that some form of frequentist model averaging 10 will help. The model averaged tail area confidence interval of Turek and Fletcher (2012) is a promising method (Kabaila, Welsh and Abeysekera, 2016, and Kabaila, Welsh and Mainzer, 2016 and Kabaila, 2017). Acknowledgement This work was supported by an Australian Government Research Training Program Scholarship. Appendix: Proof of Theorem 1 Let Z= G − ρ(H − γ) . (1 − ρ2 )1/2 It follows from (2) that Z, H and W are independent random variables and that Z ∼ N (0, 1).  The event {θ ∈ J} = g1 (W, H/W ) ≤ Z ≤ g2 (W, H/W ) . Thus   P (θ ∈ J) = P g1 (W, H/W ) ≤ Z ≤ g2 (W, H/W ) Z ∞ Z ∞ Z ∞ = −∞ 0 Z   P g1 (w, h/w) ≤ Z ≤ g2 (w, h/w) φ(h − γ) dh fW (w) dw ∞  Ψ w, h/w φ(h − γ) dh fW (w) dw. = −∞ 0  |T | < tm, 1−eα/2 =  P θ ∈ J, |T | < tm, 1−eα/2 is equal to The event  − tm, 1−eα/2 W ≤ H ≤ tm, 1−eα/2 W . Hence   P g1 (W, H/W ) ≤ Z ≤ g2 (W, H/W ), −tm, 1−eα/2 W ≤ H ≤ tm, 1−eα/2 W Z = 0 ∞ Z tm,1−α/2 w e    P g1 (w, h/w) ≤ Z ≤ g2 (w, h/w) φ h − γ dh fW (w) dw. −tm,1−α/2 w e  Changing the variable of integration of the inner integral to y = h (tm,1−eα/2 w), this expression becomes Z ∞Z tm, 1−eα/2 0 1   Ψ w, tm, 1−eα/2 y φ tm, 1−eα/2 w y − γ dy w fW (w) dw. −1 The event {θ ∈ I} =  − tm, 1−eα/2 W ≤ G ≤ tm, 1−eα/2 W . Therefore P (θ ∈ I, |T | < tm, 1−eα/2 ) is equal to   H G P −tm, 1−α/2 ≤ ≤ tm, 1−α/2 , −tm, 1−eα/2 ≤ ≤ tm, 1−eα/2 . W W 11 References Buckland, S.T., Burnham, K.P. and Augustin, N.H., 1997. Model selection: an integral part of inference. Biometrics, 53, 603–618. Chatfield, D., 1995. Model uncertainty, data mining and statistical inference. Journal of the Royal Statistical Society, Series A, 158, 419–466. Efron, B., 2014. Estimation and accuracy after model selection. Journal of the American Statistical Association, 109, 991–1022. Fletcher, D. and Turek, D., (2011). Model-averaged profile likelihood confidence intervals. Journal of Agricultural, Biological and Environmental Statistics, 17, 38–51. Hjort, N.L. and Claeskens, G., 2003. Frequentist model average estimators. Journal of the American Statistical Association, 98, 879–899. Hurvich, C. M. and Tsai, C. L., 1990. The impact of model selection on inference in linear regression. The American Statistician, 44, 214–217. Kabaila, P., 2009. The coverage properties of confidence regions after model selection. International Statistical Review, 77, 405 – 414. Kabaila, P., 2017. On the minimum coverage probability of model averaged tail area confidence intervals. Accepted for publication in Canadian Journal of Statistics. Kabaila, P. and Giri, K., 2009. Upper bounds on the minimum coverage probability of confidence intervals in regression after model selection. Australian & New Zealand Journal of Statistics, 51, 271–287. Kabaila, P., Welsh, A.H. and Abeysekera, W. 2016. Model-averaged confidence intervals. Scandinavian Journal of Statistics, 43, 35–48. Kabaila, P., Welsh, A.H. and Mainzer, R., 2016. The performance of model averaged tail area confidence intervals. Communications in Statistics - Theory and Methods, DOI: 10.1080/03610926.2016.1242741 Leeb, H. and Pötscher, B., 2005. Model selection and inference: facts and fiction. Econometric Theory 21, 21–59. Miller, A., 2002. Subset Selection in Regression, 2nd ed. Chapman & Hall/CRC, Boca Raton FL. Sheather, S. J., 2009. A Modern Approach to Regression with R. Springer, New York. Turek, D. and Fletcher, D., 2012. Model-averaged Wald intervals. Computational Statistics and Data Analysis, 56, 2809–2815. 12
10math.ST
www.nature.com/scientificreports OPEN Received: 16 May 2017 Accepted: 20 September 2017 Published: xx xx xxxx Algorithm guided outlining of 105 pancreatic cancer liver metastases in Ultrasound Alexander Hann 1,2, Lucas Bettac1, Mark M. Haenle1, Tilmann Graeter3, Andreas W. Berger1, Jens Dreyhaupt4, Dieter Schmalstieg5, Wolfram G. Zoller2 & Jan Egger5,6 Manual segmentation of hepatic metastases in ultrasound images acquired from patients suffering from pancreatic cancer is common practice. Semiautomatic measurements promising assistance in this process are often assessed using a small number of lesions performed by examiners who already know the algorithm. In this work, we present the application of an algorithm for the segmentation of liver metastases due to pancreatic cancer using a set of 105 different images of metastases. The algorithm and the two examiners had never assessed the images before. The examiners first performed a manual segmentation and, after five weeks, a semiautomatic segmentation using the algorithm. They were satisfied in up to 90% of the cases with the semiautomatic segmentation results. Using the algorithm was significantly faster and resulted in a median Dice similarity score of over 80%. Estimation of the inter-operator variability by using the intra class correlation coefficient was good with 0.8. In conclusion, the algorithm facilitates fast and accurate segmentation of liver metastases, comparable to the current gold standard of manual segmentation. Pancreatic ductal adenocarcinoma (PDAC) has a poor prognosis, despite new developments regarding chemotherapy regimen in the last years, with a survival rate of less than 7% of all patients five years after diagnosis1. Additionally this disease is projected to become the second most common cause of cancer death in 20302. Most of the patients are diagnosed with a metastatic disease stage with no option for a surgical resection in curative intent3. New therapeutic options that were introduced in the recent years prolong survival of these patients diagnosed in a metastatic state4,5. The establishment of second line therapies even increase the duration of treatment6–9. Similarly, the second most common carcinoma arising from the pancreas, the neuroendocrine neoplasms (NEN), are also associated with a poor prognosis when diagnosed with liver metastasis10. Due to the prolonged time of these patients receiving chemotherapy, the number of staging examinations, which are done every two to three months, is also growing11. Ultrasound (US) is a cheap and fast method to visualize focal liver lesions. It has a similar sensitivity regarding focal liver lesions in direct comparison to computed tomography (CT) or magnetic resonance tomography (MRI). Ultrasound is - probably due to the mentioned benefits - more widespread than CT or MRI and thus included as the imaging modality of choice in the German guideline for follow up of patients after resection of stage I or II colon cancer12. The ESMO – ESDO clinical practice guidelines recommend US for the response evaluation of pancreatic cancer patients undergoing palliative treatment13. Due to the fact that liver metastasis have different appearances in ultrasound, the interpretation of these images has a poor inter-observer agreement14,15. The outlining and segmentation of liver lesions for evaluation of their size often results in different values due to this poor inter-observer agreement. Semiautomatic segmentation algorithms promise a faster and more accurate way to segment lesions but are often performed on a small set of lesions by examiners that often worked with the algorithm before. Still computer aided automatic delineation of masses are successfully used for different image modalities like in radiation therapy using MRI images16 or in the detection of breast nodules in breast cancer screening using ultrasound17,18. 1 Department of Internal Medicine I, Ulm University, Ulm, Germany. 2Department of Internal Medicine and Gastroenterology, Katharinenhospital, Kriegsbergstraße 60, 70174, Stuttgart, Germany. 3Department of Diagnostic and Interventional Radiology, Ulm University, Ulm, Germany. 4Institute of Epidemiology & Medical Biometry, Ulm University, Ulm, Germany. 5Institute for Computer Graphics and Vision, Graz University of Technology, Inffeldgasse 16, 8010, Graz, Austria. 6BioTechMed, Krenngasse 37/1, 8010, Graz, Austria. Correspondence and requests for materials should be addressed to A.H. (email: [email protected]) ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 1 www.nature.com/scientificreports/ Using the medical platform MevisLab19, we implemented a computer-based method to measure focal liver lesions on B-mode US images via a semiautomatic method. The fundamental method has already been successfully applied to the segmentation of pituitary adenomas20, prostate central glands (PCG)21, vertebral bodies22, aneurysms and stenosis22, and even liver ablation zones23. A detailed description of translating the basic method for the application of US images has been previously reported by the authors24. In this study however, we aim to evaluate the method by users that never worked with the algorithm before and on a much larger set of data. We retrospectively identified 105 B-mode US images of liver metastasis generated during the treatment course of 56 patients with histologically confirmed pancreatic cancer. Two examiners outlined in a first session all lesions and determined the greatest diameter. In a second session after five weeks, the algorithm was used to measure the lesions in a random order. The examiners were given up to five minutes to measure ten images in order to learn how to use the program. Afterwards, they were asked to measure the 105 metastases using the algorithm. In summary, the two examiners were satisfied with the result of the semiautomatic segmentation in 92 and 94 of the 105 cases. Regarding these measurements, the segmentation of the liver metastasis took median 17.2 and 10.2 seconds using the manual method, compared to 9.5 and 8.2 seconds using the semiautomatic method. The median differences of manual and semiautomatic segmentations were significant (p < 0.01) in both cases. Additionally, we were able to measure the overlap of the areas measured with the manual and semiautomatic method using the Dice similarity scores (DSC)25 and the Hausdorff Distances (HD)26. This resulted in a median DSC of 84% (95% CI; 82.4–85.3) for examiner 1 and 82% (95% CI; 80.3–84.3) for examiner 2. The median HD were 9 (95% CI; 8.6–10.6) and 10 (95% CI; 9.4–11.2) pixels, respectively. Materials and Methods Data Acquisition. Ultrasound examinations were performed using a multifrequency curved probe, which allows ultrasound acquisitions with a bandwith of 1 to 6 MHz (LOGIQ E9/GE Healthcare, Milwaukee, Il, USA and Toshiba Aplio 80, Otawara, Japan). Using the digital picture archive of the ultrasound unit at the Katharinenhospital Stuttgart (Germany), we retrospectively selected images of liver metastasis obtained from patients with PDAC or NEN. The images were included, if they fulfilled the following criteria: visible metastases surrounded by liver tissue with no visible marker or text overlaying the target lesion. Patient information was removed from the image, and the anonymized picture was subsequently ordered randomly. The local ethics committee (Ethik-Kommission der Landesärztekammer Baden-Württemberg, Stuttgart, Germany) provided a waiver of the requirement for informed consent for this retrospective study and allowed the publication of anonymized data. Segmentation Method. The segmentation method applied in this evaluation study belongs to the class of graph-based algorithms27 and the section follows the explanations in ref.24. Graph-based algorithms convert an image or parts of it into a graph G(V, E) that consists of nodes n ∈ V and edges e ∈ E. The nodes n ∈ V are sampled in the image joined by two virtual nodes s ∈ V and t ∈ V, which are called the source and the sink, respectively. The two virtual nodes s and t are utilized to perform an s-t-cut28 after the graph construction. In doing so, the graph is split in two parts: the foreground (which could be the liver metastasis) and the background (which could represent the remaining surrounding liver tissue). Moreover, the edges e ∈ E establish direct connections between the sampled and two virtual nodes, like an edge <vi, vj> ∈ E that connects the nodes vi, vj29. In our approach the graphs’ nodes are sampled along radial rays, which are equidistantly distributed around a fixed point in a clockwise manner (Supplementary Figure 1). After the nodes have been sampled, the so called ∞-weighted intra-edges that connect nodes along the same ray are set up (Supplementary Figure 2). In summary, the ∞-weighted intra-edges ensure that the s-t-cut affects only one of the edges that belong to the same ray, which ensures a star shaped min-cut/max-flow result30. Following, the ∞-weighted inter-edges between nodes from different rays are set up (Supplementary Figures 3 and 4). Here, a delta value Δr influences the number of possible s-t-cuts, hence influencing the smoothness of the segmentation result. In practice, the greater the delta value, the greater is also the flexibility of the segmented contour. However, for our evaluation study, we set the delta value to a fixed value of two (Δr = 2). In the ensuing step, the weighted edges between the nodes and the source/sink are set up. Therefore, an average gray value is needed and sampled on the fly around the user-defined seed point. The average gray values is used to determine the absolute differences between this average gray value and the gray values behind the sampled nodes. Last but not least, the differences between two adjacent absolute values are calculated, which finally represent the weights of the edges. Note, that the signs of the differences (negative/ positive) define if a node is bound to the source or the sink, with the exception for the very first and last nodes of every ray, which are bound with their absolute weight values to the source or sink, respectively. In fine, our graph construction replaces the pre-defined templates with fixed seed point positions used in previous works21,31,32, with a circular template centered around a user-defined seed point in the US image. Measurement. Manual and semiautomatic measurement was performed on a Lenovo Yoga 2 Pro laptop with an Intel Core i7-4500U CPU @ 2.40 GHz, 8 GB RAM and Windows 8.1, 64 bit installed. Two examiners with over 10 years-experience in performing and interpreting over 20,000 ultrasound examinations performed the measurements. The chosen images were never tested with the algorithm nor were the images known to the examiners. Additionally, the examiners had never used the algorithm before. In the first session, both examiners outlined each of the metastasis manually and measured the largest diameter (diameter a), with an additional shorter diameter 90 degree related to the first one (diameter b). Time of measurement per metastasis was recorded. In the second session, after five weeks to decrease bias due to memory effect, the examiners performed the semiautomatic measurement. The order of the metastasis during this session was randomly redistributed. A short training including additional 10 images of pancreatic cancer liver metastases images was performed for up to five minutes prior to the session to exercise the usage of the algorithm. ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 2 www.nature.com/scientificreports/ Figure 1. Overview of the semiautomatic segmentation of the liver metastasis using US-CUT. Median (mm) Q1 (mm) Q3 (mm) Min (mm) Max (mm) 20 14 27 6 115 Table 1. Length of the largest diameter of the 105 metastases. Q1 and Q3 = quartile 1 and 3. The algorithm was applied as described before24. First, a seed point was applied in the middle of the liver lesion, which could additionally be improved by dragging the seed point around (inside the lesion). The algorithm immediately presents segmentation contours around the edges of the lesion. If the examiner is satisfied, the next image was loaded. In case of dissatisfaction with the outlining, additional helper seeds were placed by the examiner to alter the morphology of the outlining (Fig. 1). Note: the placement of one or more additional helper seeds still allowed the dragging of the initial seed point in the middle of the liver lesion, only with the difference that the helper seeds remain on the lesions border, thus supporting the algorithm with this additional knowledge (location/gray values). Dissatisfaction with the result of the outlining despite additional helper seeds and time spent on segmentation were noted. We measured the time spent per metastasis and calculated the degree of area overlap, comparing manual and semiautomatic segmentations using the Dice similarity score (DSC)25. A value of 100% resembles perfect overlap of two areas. Hausdorff distance (HD) was used in addition to the DSC to measure the degree of overlap in pixels26. This distance resembles the greatest of all distances from a point in one segmentation to the closest point in the other segmentation. Additionally, we calculated the difference between the two manually measured greatest diameters (a and b) and automatically calculated diameters from the algorithm after semiautomatic segmentation in millimeter. Statistics were performed with SPSS 24. The 95% confidence interval of the median was calculated using Bootstraping. The Wilcoxon rank sum test was used to investigate differences between time spent for manual and semiautomatic segmentation. Additionally, the two-sided one sample student t-test was used to test a significant difference between manual and semiautomatic time per examiner. The intra class correlation coefficient (ICC) of the DSC (manual vs. semiautomatic) of examiner 1 and 2 was used to estimate the inter-operator variability. Results Characteristics of the selected images. Selection of retrospective images obtained during the treatment of patients with pancreatic cancer yielded 77 ultrasound images of 46 patients diagnosed with PDAC. Additionally, 28 images of 10 patients with neuroendocrine pancreatic neoplasia were identified. All images displayed liver metastases without overlaying of marker or text. The median of the maximal lesion diameter was 20 mm (Table 1). The examiners were satisfied in 92 (88%) and 94 (90%) of the 105 cases with the result using the algorithm. In seven identical cases, both examiners were not satisfied with the outlining drawn by the algorithm despite the placement of helper seeds. In the following sections, our analysis will focus on the measurements of the segmentation that were marked as “satisfied”, unless we explicitly mention comparisons to the segmentations that were marked “not satisfied.” Comparison of time spent for manual and semiautomatic segmentation. Analyzing the amount of time each examiner needed for the manual segmentation per image revealed that examiner 1 (n = 92) spent median 17.2 seconds, compared to 10.2 seconds spent by examiner 2 (n = 94). Regarding the manual segmentation, examiner 2 was faster, but both examiners needed less time, when they were using the semiautomatic algorithm. (Fig. 2) The semiautomatic segmentation was in median 8 seconds faster when performed by examiner 1 and 2.2 seconds faster when performed by examiner 2. Both differences were statistically significant using the Wilcoxon rank sum test regarding (p < 0.01) (Table 2). The mean values were 8.8 seconds and 2.0 seconds for examiner 1 and 2 respectively. These values were as well significant using the two-sided one sample student t-test (p < 0.01). Interestingly, analyzing the dataset including the segmentations marked as “not satisfied” revealed that examiner 1 had images where he spent more time on the semiautomatic segmentation than on the manual outlining. All these cases were part of the segmentations marked as “not satisfied”. This reflects the additional amount of time spent on trying to achieve a good result using the algorithm. ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 3 www.nature.com/scientificreports/ Figure 2. Comparison of time spent on manual and semiautomatic measurement. Box-and-whisker plots illustrate 92 and 94 marked as satisfied measurements by examiner 1 and 2 respectively. Manual (s) Semiautomatic (s) Median Q1 Q3 Min Max Median Q1 Q3 Min Max p* Examiner 1 (n = 92) 17.2 15.1 19.9 9.5 36.7 9.5 7.3 10.9 2.1 16.3 p < 0.01 Examiner 2 (n = 94) 10.2 8.7 11.6 7.4 18.3 8.2 7.0 9.7 2.1 19.6 p < 0.01 Table 2. Comparison time spent on measuring metastasis per image. Only semiautomatic measurements regarded as satisfied included. Q1 and Q3 = quartile 1 and 3. *The p-value is related to the median difference between manual and semiautomatic segmentation time. Comparison of manual and semiautomatic segmentation. The DSC was used to determine the degree of resemblance between two segmentation areas that are superimposed on one another (Fig. 3). We calculated the DSC for manual compared to semiautomatic segmentation for each examiner (Table 3). It revealed a median DSC of 84% (95% CI; 82.4–85.3) for examiner 1 and 82% (95% CI; 80.3–84.3) for examiner 2. Additional, the calculated median HD revealed a low median distance of 9 pixels (95% CI; 8.6–10.6) for examiner 1 and 10 pixel (95% CI; 9.4–11.2) for examiner 2. Furthermore, in order to analyze the accuracy of the semiautomatic segmentation, the examiners were asked to manually measure the largest diameter (diameter a) with an additional shorter diameter 90 degree related to the first one (diameter b). The corresponding two diameters were calculated by the algorithm automatically during the semiautomatic outlining using the resulting contour of the lesion. The difference between manual and semiautomatic segmentation revealed a median of only 2 mm (Q1-Q3; 1–4) for diameter a and 1 mm (Q1-Q3; 0–2) for diameter b measured by examiner 1. Examiner 2 had a median difference of 3 mm (Q1-Q3; 2–5) for diameter a and 2 mm (Q1-Q3; 1–3) for diameter b (Supplementary Table 1). Estimation of the inter-operator variability using the ICC of the DSC manual vs semiautomatic in segmentations marked a satisfied by both examiners (n = 88) was 0.8. Segmentations regarded as not adequate by the examiners. Examiner 1 regarded 13 segmentations and examiner 2, 11 segmentations as not good enough. Seven of these were regarded by both as inadequate. Supplementary Figure 5 presents an example of a segmentation regarded by examiner 1 as inadequate compared to the segmentation of the same image by examiner 2, who was satisfied with the result. Examiner 1 defined the echo-poor halo surrounding the metastasis as part of the lesion. Unfortunately, the outlining of the semiautomatic segmentation does not reach the manually drawn line despite the use of helper seeds. In contrast, examiner 1 regarded a thinner halo as part of the lesion and was satisfied with the semiautomatic segmentation result. A possible explanation for the result of examiner 1 is that the homogenous thick halo of this metastasis is regarded as normal liver tissue by the algorithm. Discussion In this contribution, we presented the results of a fast, interactive segmentation algorithm for pancreatic liver metastases in ultrasound images. We decided to evaluate this new method using a large number of images displaying hepatic metastases of pancreatic cancer patients, which have never before been tested with this algorithm. Additionally, we asked two examiners to participate in the evaluation, who had never used the algorithm, and gave them only up to five minutes to learn how to use it. Despite these challenges, the algorithm performed very well with a DSC of over 80% and a 95% CI starting at over 80%. The differences in the diameters of 1 to 3 mm were very small, and both examiners were satisfied with the results in nearly 90% of the cases. Finally, the semiautomatic measurement was significantly faster than the manual one in both cases. Although the median difference in the manual and semiautomatic segmentation ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 4 www.nature.com/scientificreports/ Figure 3. Example of a segmentation. Depicted are the native image in the background and two zoomed-in views of the metastasis (representing the blue box). The upper left box represents the segmentation results of examiner 1 and the upper right box the results of examiner 2. The red outlines represent the manual segmentations and the yellow outlines represent the results of the semi-automatic segmentation. Area DSC (%) Area HD (Voxel) Median 95% CI Min Max Median 95% CI Min Max Examiner 1 (n = 92) 84 82.4–85.3 63 93 9 8.6–10.6 4 31 Examiner 2 (n = 94) 82 80.3–84.3 57 94 10 9.4–11.2 4 133 Table 3. Comparison of satisfied manual and semiautomatic measurement per examiner. DSC = Dice similarity score, HD = Hausdorff distance, CI = confidence interval. time was just 8 and 2.2. seconds, these differences represent 47% and 22% of the manual segmentation time respectively. The time during segmentation represents a critical step in examination due to the high mental concentration of the examiner to carefully segment the lesion without over or underestimation of its size. A search of the literature revealed different publications in peer reviewed journals and contributions to congresses that described methods for segmentation of ultrasound images. Most of these publications discuss the difficulties of segmentation performed using ultrasound images and propose new ways to overcome difficulties like speckle noise existence33. One way is to enhance the contrast of the lesion in comparison to the surrounding tissue by using additional intravenous contrast dye34–36. This time-consuming method truly helps to delineate focal liver lesion, but harbors the risk of allergic reaction to the contrast dye and significantly increases the costs and length of the examination. The latter disadvantage is due to the three contrast phases that are usually intermittent recorded during an examination and that usually take up to 6 minutes37. In regular follow-up visits of cancer patients, only the sizes of liver metastases are assessed using native non-contrast enhanced images. One study involved 55 cases of focal liver lesions that received a contrast enhanced ultrasound (CEUS) for workup of a suspected liver lesion35. After selection of a key frame derived from the CEUS examination video, the radiologist placed a seed point that induced the segmentation process. The reference in this study was a manual segmentation of the liver lesion done by one radiologist. Although using the proposed algorithm resulted in a good segmentation result, only 11 out of 55 images were segmented by two radiologists. Additionally, it is not clear if the chosen images were directly evaluated for the first time with the proposed algorithm. Studies that proposed native, non-contrast enhanced ultrasound images often involved just a few cases to present the feasibility of the method38,39. One study that evaluated a semi-supervised segmentation algorithm using different ultrasound ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 5 www.nature.com/scientificreports/ datasets of focal liver lesions achieved very good Dice scores of over 90% compared to an expert radiologist who manually segmented the focal liver lesions40. This study involved images obtained from 15 patients. Although faster than the previously proposed algorithm by the group, the time for semi-automated segmentation ranged from 18 to 35 seconds, with a median of 27 seconds. This is slower than the time spent using our algorithm that was 10 seconds for examiner 1 and 8 seconds for examiner 2. Finally, there is some previous work from the authors that is related to this contribution, In our previous work23,41,42, we apply and evaluate a semi-automatic segmentation algorithm for radiofrequency ablation (RFA) zones via optimal s-t-cuts. The interactive graph-based approach builds upon a polyhedron to construct the graph. However, it was specifically designed for computed tomography acquisitions from patients that had RFA treatments of hepatocellular carcinomas (HCC). For an evaluation, we used twelve post-interventional CT datasets from the clinical routine and, as evaluation metric, we utilized also the Dice score. Compared with pure manual slice-by-slice expert segmentations from interventional radiologists, we were able to achieve a Dice score of about 80%, and our approach was even able to handle images containing the RFA needles still in place. RFA of liver tumors induces areas of tissue necrosis, which can be visualized reliably in contrast enhanced CT. This results in a bright border around the ablation zone, which makes an (automatic) segmentation easier and more reliable. The same applies for brain tumors, like glioblastoma multiforme in contrast-enhanced MRI acquisitions, where a bright border is surrounding a core of necrotic brain tissue32,43,44. Although this study presents encouraging results for the application of the algorithm in clinical routine, there are still some limitations that need to be addressed. We evaluated the algorithm with two examiners. Evaluating the method with more physicians would have increased statistical plausibility. Due to the missing direct access to video-output of an ultrasound machine, the study had to be performed with retrospectively collected images. This was also necessary because of the aim to include more than 100 images of different metastasis. We included only pancreatic cancer patients, because, in our clinic, those are often staged during palliative treatment using abdominal ultrasound combined with endoscopic ultrasound. There are several areas for future work, like the implementation of this tool to function on computers that receive the images directly from the ultrasound machine. It could be evaluated prospectively with images drawn from patients with different cancer entities. Additionally, we hope to implement the recognition of metastasis using a 3D ultrasound dataset. This segmentation is still very time consuming and often needs manual outlining of the same metastasis in more than four images in order to achieve a good result. The acquisition of 3D content with segmented liver metastasis might help implement such 3D ultrasound images in virtual reality environments45 that are explored using head mounted displays46. These displays allow immersion into computer-created real-time simulations. They have great potential to influence the way we plan interventions and teach about different disease entities. In summary, this work presents the use of an algorithm for segmentation of liver metastasis due to pancreatic cancer under real life conditions. We believe that implementation of such a method into daily practice will help assist the examiner without compromising accuracy. References 1. Kleeff, J. et al. Pancreatic cancer. Nat. Rev. Dis. Primer 2, 16022 (2016). 2. Rahib, L. et al. Projecting cancer incidence and deaths to 2030: the unexpected burden of thyroid, liver, and pancreas cancers in the United States. Cancer Res. 74, 2913–2921 (2014). 3. Hidalgo, M. Pancreatic Cancer. N. Engl. J. Med. 362, 1605–1617 (2010). 4. Conroy, T. et al. FOLFIRINOX versus gemcitabine for metastatic pancreatic cancer. N.Engl.J.Med. 364, 1817–1825 (2011). 5. Von Hoff, D. D. et al. Increased survival in pancreatic cancer with nab-paclitaxel plus gemcitabine. N.Engl.J.Med. 369, 1691–1703 (2013). 6. Ettrich, T. J. et al. DocOx (AIO-PK0106): a phase II trial of docetaxel and oxaliplatin as a second line systemic therapy in patients with advanced pancreatic ductal adenocarcinoma. BMC Cancer 16, 21 (2016). 7. Oettle, H. et al. Second-line oxaliplatin, folinic acid, and fluorouracil versus folinic acid and fluorouracil alone for gemcitabinerefractory pancreatic cancer: outcomes from the CONKO-003 trial. J. Clin. Oncol. Off. J. Am. Soc. Clin. Oncol. 32, 2423–2429 (2014). 8. Wang-Gillam, A. et al. Nanoliposomal irinotecan with fluorouracil and folinic acid in metastatic pancreatic cancer after previous gemcitabine-based therapy (NAPOLI-1): a global, randomised, open-label, phase 3 trial. Lancet Lond. Engl. 387, 545–557 (2016). 9. Hann, A., Bohle, W., Egger, J. & Zoller, W. G. Improvement in advanced pancreatic cancer survival with novel chemotherapeutic strategies - experience of a community based hospital. Z. Gastroenterol. 54, 1138–1142 (2016). 10. Falconi, M. et al. ENETS Consensus Guidelines Update for the Management of Patients with Functional Pancreatic Neuroendocrine Tumors and Non-Functional Pancreatic Neuroendocrine Tumors. Neuroendocrinology 103, 153–171 (2016). 11. Hann, A., Bohle, W., Egger, J. & Zoller, W. Feasibility of alternating induction and maintenance chemotherapy in pancreatic cancer. Sci. Rep. 7, 41549 (2017). 12. Pox, C. P. & Schmiegel, W. [German S3-guideline colorectal carcinoma]. Dtsch. Med. Wochenschr. 1946(138), 2545 (2013). 13. Seufferlein, T., Bachet, J. B., Van Cutsem, E. & Rougier, P. & ESMO Guidelines Working Group. Pancreatic adenocarcinoma: ESMOESDO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann. Oncol. Off. J. Eur. Soc. Med. Oncol. ESMO 23(Suppl 7), vii33–40 (2012). 14. Hohmann, J. et al. Characterisation of focal liver lesions with unenhanced and contrast enhanced low MI real time ultrasound: onsite unblinded versus off-site blinded reading. Eur. J. Radiol. 81, e317–324 (2012). 15. Harvey, C. J. & Albrecht, T. Ultrasound of focal liver lesions. Eur. Radiol. 11, 1578–1593 (2001). 16. Feng, Y. et al. A comparative study of automatic image segmentation algorithms for target tracking in MR-IGRT. J. Appl. Clin. Med. Phys. 17, 441–460 (2016). 17. Berg, W. A. et al. Detection of breast cancer with addition of annual screening ultrasound or a single screening MRI to mammography in women with elevated breast cancer risk. JAMA 307, 1394–1404 (2012). 18. Feng, Y. et al. An adaptive Fuzzy C-means method utilizing neighboring information for breast tumor segmentation in ultrasound images. Med. Phys. 44, 3752–3760 (2017). 19. Egger, J. et al. Integration of the OpenIGTLink network protocol for image-guided therapy with the medical platform MeVisLab. Int. J. Med. Robot. Comput. Assist. Surg. MRCAS 8, 282–290 (2012). 20. Egger, J., Zukić, D., Freisleben, B., Kolb, A. & Nimsky, C. Segmentation of pituitary adenoma: a graph-based method vs. a balloon inflation method. Comput. Methods Programs Biomed. 110, 268–278 (2013). 21. Egger, J. PCG-cut: graph driven segmentation of the prostate central gland. PloS One 8, e76645 (2013). ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 6 www.nature.com/scientificreports/ 22. Schwarzenberg, R., Freisleben, B., Nimsky, C. & Egger, J. Cube-cut: vertebral body segmentation in MRI-data through cubic-shaped divergences. PloS One 9, e93389 (2014). 23. Egger, J. et al. Interactive Volumetry Of Liver Ablation Zones. Sci. Rep. 5, 15373 (2015). 24. Egger, J., Schmalstieg, D., Chen, X., Zoller, W. G. & Hann, A. Interactive Outlining of Pancreatic Cancer Liver Metastases in Ultrasound Images. Sci. Rep. 7, 892 (2017). 25. Dice, L. R. Measures of the Amount of Ecologic Association Between Species. Ecology 26, 297–302 (1945). 26. Huttenlocher, D. P., Klanderman, G. A. & Rucklidge, W. J. Comparing images using the Hausdorff distance. IEEE Trans. Pattern Anal. Mach. Intell. 15, 850–863 (1993). 27. Jianbo, S. & Malik, J. Normalized cuts and image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 22, 888–905 (2000). 28. Boykov, Y. & Kolmogorov, V. An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision. IEEE Trans. Pattern Anal. Mach. Intell. 26, 1124–1137 (2004). 29. Kang, L., Xiaodong, W., Chen, D. Z. & Sonka M. Optimal Surface Segmentation in Volumetric Images-A Graph-Theoretic Approach. IEEE Trans. Pattern Anal. Mach. Intell. 28, 119–134 (2006). 30. Veksler, O. Star Shape Prior for Graph-Cut Image Segmentation. In Computer Vision – ECCV 2008 (eds. Forsyth, D., Torr, P. & Zisserman, A.) 5304, 454–467 (Springer Berlin Heidelberg, 2008). 31. Egger, J. et al. Square-Cut: A Segmentation Algorithm on the Basis of a Rectangle Shape. PLoS ONE 7, e31064 (2012). 32. Egger, J., Freisleben, B., Nimsky, C. & Kapur, T. Template-cut: a pattern-based segmentation paradigm. Sci. Rep. 2, 420 (2012). 33. Xiaohui, H, Bruce, C., Pislaru, C. & Greenleaf, J. F. A novel region growing method for segmenting ultrasound images. in 2, 1717–1720 (IEEE, 2000). 34. Bahrami, N. et al. Boundary delineation for hepatic hemangioma in ultrasound images. Conf. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. IEEE Eng. Med. Biol. Soc. Annu. Conf. 2011, 7989–7992 (2011). 35. Bakas, S. et al. Fast semi-automatic segmentation of focal liver lesions in contrast-enhanced ultrasound, based on a probabilistic model. Comput. Methods Biomech. Biomed. Eng. Imaging Vis. 1–10, doi:https://doi.org/10.1080/21681163.2015.1029642 (2015). 36. Gatos, I. et al. A new automated quantification algorithm for the detection and evaluation of focal liver lesions with contrastenhanced ultrasound: Focal liver lesions and contrast-enhanced ultrasound. Med. Phys. 42, 3948–3959 (2015). 37. Sporea, I. et al. Contrast-enhanced ultrasound (CEUS) for the evaluation of focal liver lesions - a prospective multicenter study of its usefulness in clinical practice. Ultraschall Med. Stuttg. Ger. 1980(35), 259–266 (2014). 38. Jain, N. & Kumar, V. Liver Ultrasound Image Segmentation Using Region-Difference Filters. J. Digit. Imaging, doi:https://doi. org/10.1007/s10278-016-9934-5 (2016). 39. Quan, L., Zhang, D., Yang, Y., Liu, Y. & Qin, Q. Segmentation of tumor ultrasound image via region-based Ncut method. Wuhan Univ. J. Nat. Sci. 18, 313–318 (2013). 40. Ciurte, A. et al. Semi-supervised segmentation of ultrasound images based on patch representation and continuous min cut. PloS One 9, e100972 (2014). 41. Egger, J. et al. RFA-cut: Semi-automatic segmentation of radiofrequency ablation zones with and without needles via optimal s-tcuts. Conf. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. IEEE Eng. Med. Biol. Soc. Annu. Conf. 2015, 2423–2429 (2015). 42. Egger, J. et al. Semi-automatische Segmentierung von Schädigungszonen in post-interventionellen CT-Daten. in Bildverarbeitung für die Medizin 2015 (eds. Handels, H., Deserno, T. M., Meinzer, H.-P. & Tolxdorff, T.) 281–286, doi:https://doi.org/10.1007/978-3662-46224-9_49 (Springer Berlin Heidelberg, 2015). 43. Egger, J. et al. Nugget-Cut: A Segmentation Scheme for Spherically- and Elliptically-Shaped 3D Objects. in Pattern Recognition (eds Goesele, M., Roth, S., Kuijper, A., Schiele, B. & Schindler, K.) 6376, 373–382 (Springer Berlin Heidelberg, 2010). 44. Egger, J., Lüddemann, T., Schwarzenberg, R., Freisleben, B. & Nimsky, C. Interactive-cut: Real-time feedback segmentation for translational research. Comput. Med. Imaging Graph. Off. J. Comput. Med. Imaging Soc. 38, 285–295 (2014). 45. Egger, J. et al. HTC Vive MeVisLab integration via OpenVR for medical applications. PloS One 12, e0173972 (2017). 46. Chen, X. et al. Development of a surgical navigation system based on augmented reality using an optical see-through head-mounted display. J. Biomed. Inform. 55, 124–131 (2015). Acknowledgements The work received funding from BioTechMed-Graz (“Hardware accelerated intelligent medical imaging”) and the 6th Call of the Initial Funding Program from the Research & Technology House (F&T-Haus) at the Graz University of Technology (PI: Dr. Dr. Dr. habil. Jan Egger). Alexander Hann is supported by a Baustein Grant of the University Ulm. Author Contributions Conceived and designed the experiments: A.H., J.E. Performed the experiments: A.H., L.B., M.M.H., T.G., A.B. Analyzed the data: A.H., J.E., J.D. Contributed reagents/materials/analysis tools: J.E., A.H., A.B., W.G.Z., D.S. Wrote the paper: A.H., J.E. Additional Information Supplementary information accompanies this paper at https://doi.org/10.1038/s41598-017-12925-z. Competing Interests: The authors declare that they have no competing interests. Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. © The Author(s) 2017 ScIentIfIc REPOrtS | 7: 12779 | DOI:10.1038/s41598-017-12925-z 7
1cs.CV
Proceedings of the 7th GACM Colloquium on Computational Mechanics for Young Scientists from Academia and Industry October 11-13, 2017 in Stuttgart, Germany Reduced Order Modelling for the Simulation of Quenches in Superconducting Magnets arXiv:1710.04859v1 [cs.CE] 13 Oct 2017 Sebastian Schöps1 *, Idoia Cortes Garcia1 , Michal Maciejewski2 and Bernhard Auchmann2,3 Micro Abstract This contributions discusses the simulation of magnetothermal effects in superconducting magnets as used in particle accelerators. An iterative coupling scheme using reduced order models between a magnetothermal partial differential model and an electrical lumped-element circuit is demonstrated. The multiphysics, multirate and multiscale problem requires a consistent formulation and framework to tackle the challenging transient effects occurring at both system and device level. 1 Graduate School CE, Technical University of Darmstadt, Darmstadt, Germany 2 Machine Protection and Electrical Integrity Group, CERN, Geneva, Switzerland 3 Magnet Group, Paul Scherrer Institut, Villigen, Switzerland * Corresponding author: [email protected] 1 Introduction The LHC (Large Hadron Collider) at CERN, Geneva uses superconducting magnets for focussing and bending particle beams. To ensure their superconductivity, they are cooled to very low temperatures (1.9 K). However, in the case of a malfunction the superconducting material becomes resistive and quenches. The quenched area quickly overheats, the heat propagates within the magnet and from one magnet to another causing more magnets to quench. Eventually, the whole network of magnets can be damaged. This is avoided by quench detection and protection systems, e.g. Quench heaters [3] or CLIQ (Coupling-Loss Induced Quench), [11]. Quench protection analysis can be further improved by coupling controller and circuit models [9]. 2 Magnet Models The electromagnetic field inside each magnet is described by Maxwell’s equations [8], i.e., ∇×E=− ∂B , ∂t ∇×H= ∂D +J, ∂t ∇·D=ρ, ∇·B=0 (1) on a bounded domain Ω ⊂ R3 and time interval I = (t0 , tend ] with specified boundary and initial conditions on ∂Ω and t0 . They are interlinked by the material relations D = εE , H = νB , J = σE + Js . (2) Here, E is the electric field, D the electric flux density, H the magnetic field, B the magnetic flux density, J, Js are the total, source current densities and ρ is the electric charge density. The positive (semi-)definite material tensors σ  0, ε ≻ 0 and ν ≻ 0 define the electric conductivity, the electric permittivity and the magnetic permeability, respectively. For simplicity of notation, we consider only the case of linear material laws, see e.g. [2] for a generalization. ∂ D = 0 and introducing the magnetic Assuming a magnetoquasistatic approximation with ∂t vector potential B = ∇ × a, the parabolic, semi-elliptic partial differential equation σ ∂a + ∇ × (ν∇ × a) = χim ∂t (3) is obtained, where the current density Js (r, t) = χ(r)im (t) is expressed in terms of scalar currents im . In practice, one often replaces the term σ ∂a/∂t by a homogenized model to avoid the resolution of small geometric details in the discretization, e.g. single strands within cables [1]. Following [2], equation (3) is coupled via the voltage vm to a circuit using the flux linkage Z χ(r) · a(r, t) dΩ. Φ(t) = (4) Ω The temperature in the domain Ωs can be obtained from the heat balance equation ρCp ∂T − ∇ · (k∇T ) = Ps + PJoule , ∂t (5) with suitable boundary and initial conditions, where ρ is the mass density, Cp the heat capacity, k the thermal conductivity and T the temperature. The Joule losses PJoule are defined as PJoule = qflag σ −1 kJs k2 and are activated in case of a quench by the sigmoid-type activation function qflag that depends on time, the magnetic flux density B = ∇ × a and current density Js . The power density Ps in Ωs relates the heat equation with the field solution [2]. In case of quenching, i.e., qflag > 0, the superconducting coils have a resistance with the voltage drop Z d χ ⊤ σ −1 χ dΩ. (6) vm = Φ + Rt im with Rt = qflag dt Ωs The resistance Rt inherits the dependencies on t, B and Js . As temperature increases, thermal stress is caused which leads to displacement. The problem of linear elasticity reads ∇ · (C : ǫ) = F(T, J, B) (7)  where ǫ = 21 ∇u + (∇u)⊤ denotes the strain tensor and u the displacement vector. Boundary conditions still to be defined. The right hand side F depends on the thermal stress and Lorentz forces. In turn, the computed displacement deforms the computational domain Ω. 3 Networks of Magnets Let us assume that we have Nm magnets connected within a network. Then each magnet n = 1, ..., Nm is given by the same multiphysical model which is described by the coupled system of partial differential equations (3)-(7) defined above. We address the solution of each model abstractly by the operator Ψn , such that vm,n = Ψn (im,n ) (n = 1, ..., Nm ), where the internal variables an , Tn , un are suppressed as the electrical coupling to neighboring magnets is established only via currents im,n and voltages vm,n . Let us collect all solution operators as vm = Ψ(im ) (8) with currents i⊤m = [im,1 , im,2 , ..., im,Nm ] and voltages v⊤m = [vm,1 , vm,2 , ..., vm,Nm ]. Let an electrical network consist of simple elements and described by the modified nodal analysis [7], i.e., AC CA⊤C d ϕ + AR GA⊤R ϕ + AL iL + AV iV +Am im = −AI is (t) , dt d L iL − A⊤L ϕ = 0 , dt A⊤V ϕ = vs (t) , A⊤m ϕ = vm (t) . (9) (10) (11) (12) The position of each device in the network is encoded by a sparse incidence matrix A⋆ which consists only of entries from the set {−1, 0, 1}. The corresponding branch voltage is denoted by v⋆ = A⊤⋆ ϕ and the current through the device is i⋆ . The simplest network elements are inductors, capacitors and resistors; they are determined by lumped constitutive parameters, collected in the matrices of capacitances C, inductances L and conductances G. Time dependent voltage and current sources are given as vector-valued functions vs (t) and is (t). The unknowns of the system are the scalar potentials ϕ and a few currents, i.e. iL and iv . More complicated devices can be modeled by controlled sources, see e.g. [4]. Finally, the terms enclosed in the two boxes indicate the additional (non-standard) contributions due to the field/circuit coupling via equation (8). 4 Iterative Coupling Accelerated by Reduced Models Instead of solving the large coupled system (8)-(12) monolithically, we propose an iterative scheme that fits into the framework of waveform relaxation or dynamic iteration, [5, 13]. This allows us to use a different discretization in time for each system and more importantly, even different software packages can be used. (k) Let us assume that we have an initial guess of the time-transient behavior of the voltages vm (t) for t ∈ I and k = 0. Then, in the simplest case, the following scheme can be performed (k) 1. solve the electric network equations (9)–(12) by considering the voltages vm to be known (k+1) and extract the currents im (k) 2. solve n = 1, ..., Nm electrothermal problems (3)-(6) in parallel, each on domain Ωn excited (k+1) (k+1) (k+1) . by the current im,n . Extract voltages vm,n and temperatures Tn (k) 3. solve n = 1, ..., Nm elasticity problems (7) in parallel, consider the temperature Tn to be (k+1) (k+1) . known, compute displacements um,n and deform the computational domains Ωn 4. if not converged: increase counter k = k + 1 and return to step 1. Due to the smallness of the displacements it is common practice to exclude the computation of the deformations from the iteration and solve the elasticity problem (7) only once in a postprocessing step, i.e., after the iteration of steps 1-2 and 4 has converged. Convergence for the remaining electrothermal scheme is well understood and has been discussed based on Banach’s fixed point theorem and an analysis of the contraction factor in [12], [2]. The iterative method simulates each magnet in parallel but the computational burden might still be heavy. This can be mitigated by using reduced order models to decrease the number of iterations and eventually to reduce simulation costs, [10]. Let us assume that there is a reduced order model available that approximates the multiphysical problem (3)-(7). Then, the solution of the n-th magnet can be rewritten as vm,n = Ψn (im,n ) + ∆vm,n (13) where ∆vm,n is the deviation between the reduced model and the full one. This additional reduced model allows us to define a waveform relaxation in terms of voltage differences, i.e., AC CA⊤C d (k+1) (k+1) (k+1) +Am i(k+1) + AV iV ϕ + AR GA⊤R ϕ(k+1) + AL iL = −AI is (t) , m dt d (k+1) − A⊤L ϕ(k+1) = 0 , L iL dt A⊤V ϕ(k+1) = vs (t) , (k) ) = ∆vm . A⊤m ϕ − Ψ(i(k+1) m (14) (15) (16) (17) This scheme still requires the simulation of the full model Ψ but may significantly reduced the number of evaluations. On the other hand, sophisticated reduced models Ψ may also increase the computational costs (particularly in the offline stage). However, in the simplest case the reduced model is given by an improved transmission condition in terms of the optimized Schwarz waveform relaxation [6] [2]. 5 Outlook The full contribution will show simulations of the mechanical response of a magnet during quench protection and magnet discharge. For CLIQ-protected magnets, there are two phenomena in the coil: an imbalanced current profile leading to inhomogeneous Lorentz forces and an almost homogeneous temperature profile [11]. We quantify the impact of Lorentz forces and temperature gradients during a discharge as large stresses may lead to cable degradation over time. Acknowledgements This work has been supported by the Excellence Initiative of the German Federal and State Governments and the Graduate School of CE at TU Darmstadt. References [1] L. Bortot, M. Maciejewski, M. Prioli, A. Fernandez Navarro, S. Schöps, I. Cortes Garcia, B. Auchmann, and A. Verweij. Simulation of electro-thermal transients in superconducting accelerator magnets with comsol multiphysics. In Proceedings of the European COMSOL Conference 2016, Munich, 2016. [2] I. Cortes Garcia, S. Schöps, L. Bortot, M. Maciejewski, M. Prioli, A. Fernandez Navarro, B. Auchmann, and A. Verweij. Optimized field/circuit coupling for the simulation of quenches in superconducting magnets. JMMCT, 2(1):97–104, 2017. [3] K. Dahlerup-Petersen, R. Denz, J. L. Gomez-Costa, D. Hagedorn, P. Proudlock, F. Rodrinuez-Mateos, R. Schmidt, and F. Sonnemann. The protection system for the superconducting elements of the large hadron collider at CERN. In Proceedings of the 1999 Particle Accelerator Conference, volume 5, pages 3200–3202, 1999. [4] D. Estévez Schwarz and C. Tischendorf. Structural analysis of electric circuits and consequences for MNA. Int. J. Circ. Theor. Appl., 28(2):131–162, 2000. [5] M. J. Gander. Waveform relaxation. In B. Engquist, editor, Encyclopedia of Applied and Computational Mathematics. Springer, 2012. [6] M. J. Gander and A. E. Ruehli. Optimized waveform relaxation methods for RC type circuits. IEEE Trans. Circ. Syst., 51(4):755–768, 2004. [7] C.-W. Ho, A. E. Ruehli, and P. A. Brennan. The modified nodal approach to network analysis. IEEE Trans. Circ. Syst., 22(6):504–509, 1975. [8] J. D. Jackson. Classical Electrodynamics. Wiley and Sons, New York, 3rd edition, 1998. [9] M. Maciejewski, I. Cortes Garcia, S. Schöps, B. Auchmann, L. Bortot, M. Prioli, and A. P. Verweij. Application of the waveform relaxation technique to the co-simulation of power converter controller and electrical circuit models. In MMAR 2017. IEEE, 2017. [10] M. Rathinam and L. R. Petzold. Dynamic iteration using reduced order models: A method for simulation of large scale modular systems. SIAM J. Numer. Anal., 40:1446–1474, 2002. [11] E. Ravaioli. CLIQ – A new quench protection technology for superconducting magnets. PhD thesis, University of Twente, 2015. [12] S. Schöps, H. De Gersem, and A. Bartel. A cosimulation framework for multirate timeintegration of field/circuit coupled problems. IEEE Trans. Magn., 46(8):3233–3236, 2010. [13] J. K. White, F. Odeh, A. L. Sangiovanni-Vincentelli, and A. E. Ruehli. Waveform relaxation: Theory and practice. Trans. Soc. Comput. Sim., 2(1):95–133, 1985.
5cs.CE
The Impact of Local Information on the Performance of Multiagent Systems arXiv:1710.01409v1 [cs.GT] 3 Oct 2017 Vinod Ramaswamy, Dario Paccagnan, and Jason R. Marden Abstract How does the control architecture of a multiagent system impact the achievable performance guarantees? In this paper we answer this question relative to a class of resource allocation problems, termed multiagent set covering problems. More specifically, we demonstrate how the degree of information available to each agent imposes constraints on the achievable efficiency guarantees associated with the emergent collective behavior. We measure these efficiency guarantees using the well-known metrics of price of anarchy and price of stability, which seek to bound the performance of the worst and best performing equilibria, respectively. Our first result demonstrates that when agents only know local information about the resources they can select, a system operator is forced to manage an inherent trade-off between the price of anarchy and price of stability, i.e., improving the quality of the worst equilibria comes at the expenses of the best equilibria. Our second result demonstrates that a system operator can move well beyond this price of anarchy / price of stability frontier by designing a communication protocol that provides agents with a limited amount of system-level information beyond these local quantities. Introduction There is no denying the fact that multiagent systems represent the systems of the future, with wide ranging applications in cooperative robotics [1], resource and task allocation [2], supply chain management [3], surveillance [4], among others. One concrete example is the problem of monitoring the perimeter of a wild fire, where the goal is to deploy a collection of unmanned aerial vehicles (UAV), each making independent surveillance decisions in response to local information [5]. While technological developments have paved the way Date: October 3, 2017 This work is supported by SNSF Grant #P1EZP2-172122, ONR Grant #N00014-15-1-2762 and NSF Grant #ECCS-1351866. V. Ramaswamy was with the Department of Electrical, Computer, and Energy Engineering, University of Colorado at Boulder. D. Paccagnan and J. R. Marden are with the Department of Electrical and Computer Engineering, University of California, Santa Barbara, USA, [email protected], [email protected]. D. Paccagnan is also with the Automatic Control Laboratory, Swiss Federal Institute of Technology, Zurich, Switzerland. 1 for the realization of such futuristic systems, our basic understanding of networked control architectures and of the role of information, has ultimately hindered their realization. A multiagent system can be characterized by a collection of individual subsystems, each making independent decisions in response to locally available information. At a high level, the design of a networked control architecture entails specifying two key elements: the information available to each subsystem, attained either through sensing or communication, and a decision-making mechanism that prescribes how each subsystem processes available information to take decisions. The need for distributed decision-making stems from the scale, spacial distribution, and sheer quantity of information associated with such systems which excludes the possibility for centralized decision making. For example, in the problem of monitoring the perimeter of a wild fire, each individual UAV may be required to make surveillance choices using only information regarding its own aerial view of the landscape as well as minimal information regarding the location of neighboring UAVs due to limited communication capabilities. The quality of a networked control architecture is ultimately gauged by several dimensions including the stability and efficacy of the emergent collective behavior, characteristics of the transient behavior, as well as the level of information propagated throughout the system. Two central questions associated with the design process that we seek to address in this paper are as follows: (i) What are the decision-making rules that optimize the performance of the emergent collective behavior for a given level of information? (ii) What is the value of information in networked control architectures? That is, how does informational availability translate to attainable performance guarantees for the emergent behavior through the design of appropriate decision-making mechanisms? For example, how do information limitations to the UAVs impact the quality at which the UAVs can collectively monitor the perimeter of a wild fire? Multiagent surveillance problems highlight several challenges inherent to the design of such networked control architectures [6, 7]. For this class of problems, the goal is to derive local decision-making rules for the individual subsystems (i.e., sensors) ensuring that sensors efficiently allocate themselves over a given mission space without the need for centralized coordination. A common assumption in the existing literature is that each 2 subsystem is endowed with information pertaining to its local environment as well as the surveillance choices of neighboring sensors [6]. Within these confines, researchers have established a wide range of local decision-making rules that ensure the collective behavior will settle at an allocation; however, the performance guarantees associated with this allocation have often been overlooked. Motivated by this omission, recent work in [8] characterizes the best attainable performance guarantees as a function of the underlying informational dependencies for such multi-agent surveillance problems. For systems comprised of n sensor, [8] demonstrates that any decision-making rule that relies solely on local information can at best guarantee that the resulting collective behavior is within (1/n) of the optimal system behavior. However, if agents are equipped with slightly more information regarding the state of the system, this impossibility result vanishes and one can easily design networked architectures with a limiting behavior that is guaranteed to be within (1/2) of the optimal behavior [8]. The resounding message from the result above is that informational dependencies play a fundamental role in the attainable performance guarantees associated with networked control architectures. The goal of this paper is to shed further light on the interplay between informational availability and performance guarantees for a class of set covering problems. Covering problems consist of a collection of resources, each with a respective value or worth, and a collection of agents, each capable of covering various subsets of resources. The overarching goal is to establish a networked control architecture that allows agents to reach the optimal allocation through a process where the individual agents adjust their covering choice based on available information. Of specific interest will be identifying how the information available to the individual agents impacts the attainable performance guarantees associated with the underlying architecture. We approach this problem through a game theoretic lens where we model the individual agents as players in a game and each agent is associated with a local objective function that guides its decision-making process [9]. We treat these local objective functions as our design parameter and focus our analysis on characterizing the performance guarantees associated with the resulting equilibria of the designed game. An equilibrium of a game is an outcome where each agent is satisfied with their choice given the choices of the other agents. We concentrate our analysis on two well-studied performance metrics in the game theoretic literature termed the price of anarchy and price of stability [10, 11]. Informally, the price of anarchy seeks to provide performance guarantees associated with the worst 3 performing equilibrium relative to the optimal allocation. The price of stability, on the other hand, seeks to provide similar performance guarantees when restricting attention to the best performing equilibrium. The lack of uniqueness of equilibria implies that these bounds are often quite different. Our first result concentrates on deriving the agent objective functions that optimize the quality of the resulting equilibria when individual agents are only aware of information regarding the resources they can select. For this class of informational dependencies, we show that there does not exist agent objective functions that simultaneously optimize both the price of anarchy and price of stability. Rather, there is a fundamental trade-off between these two performance measures. In other words, improving the performance of the worst equilibrium necessarily comes at the expense of the performance of the best equilibrium. In Theorem 3, we characterize this fundamental trade-off and identify the agent objective functions that reach this price of anarchy / price of stability frontier. This result is one of the first in the existing literature that demonstrates a fundamental tradeoff between the price of anarchy and price of stability. Whether or not this trade-off is a more fundamental feature of networked architectures is an open and interesting question. Our second result seeks to address the question of whether providing the agents with additional information, i.e., system-level attributes beyond information relative to the local resources, can permit the design of networked architectures with performance guarantees breaching this frontier. In Theorem 4, we answer this question in the affirmative as we identify a minimal (and easily attainable) piece of system-level information that can permit the realization of decision-making rules with performance guarantees beyond the price of anarchy / price of stability frontier provided in Theorem 3. It is important to stress that the specific make-up of this information is not necessarily important; rather, the importance of this result centers on the fact that certain attributes of the system can be exploited in networked architectures if those attributes are propagated to the agents. Minimizing the amount of information that needs to be propagated throughout the system to move beyond the frontier is clearly an important question that warrants future attention. 4 Model and Performance Metrics In this section we introduce the multiagent set covering problem and our game theoretic model for the design of local decision-making mechanisms [12]. Further, we define the objectives and performance metrics of interest, as well as provide a review of the relevant literature. Covering problems Let R = {r1 , r2 , . . . , rm } be a finite set of resources where each resource r ∈ R is associated with a value vr ≥ 0 defining its importance. We consider a covering problem where the goal is to allocate a collection of agents N = {1, . . . , n} to resources in R in order to maximize the cumulative value of the covered resources. The set of possible assignments for each agent i ∈ N is given by Ai ⊆ R and we express an allocation by the tuple a = (a1 , a2 , . . . , an ) ∈ A = A1 × · · · × An .2 The total value, or welfare, associated with an allocation a is given by X W (a) = vr . r∈∪i∈N ai The goal of the covering problem is to find an optimal allocation, i.e., an allocation aopt ∈ A such that W (aopt ) ≥ W (a) for all a ∈ A.3 We will often express an allocation a as (ai , a−i ) with the understanding that a−i = (a1 , . . . , ai−1 , ai+1 , . . . , an ) denotes the collection of choices of the agents other than agent i. A game theoretic model This paper focuses on deriving distributed mechanisms for attaining near optimal solutions to covering problem where the individual agents make independent covering choices in response to local available information. Specifically, in this section we assume that each 2 Many of the forthcoming results hold for the more general multi-selection covering problem where each agent’s action set is of the form Ai ⊆ 2R . However, in this paper we present solely the singleton selection problem for a more direct exposition. 3 Approximation algorithms for finding a near optimal solution to such covering problems, or more general submodular optimization problems, have been extensively studied in the literature [13, 14, 15, 16]. The focus of this literature is predominantly on centralized algorithms for finding near optimal allocations. By centralized, we mean a computer program that takes in the full problem specification and outputs an allocation. In contrast, our focus is on distributed solutions where there is no such central authority and each decision-making entity has incomplete information about the system as a whole. 5 agent i has information only regarding the resources that the agent can select, i.e., the resources Ai ⊆ R. Rather than directly specifying a decision-making process, here we focus on the design of local agent objective functions that adhere to these informational dependencies and will ultimately be used to guide the agents’ selection process. To that end, we consider the framework of distributed welfare games [9] where each agent is associated with a local utility or objective function Ui : Ai → R and for any allocation a = (ai , a−i ) ∈ A, the utility of agent i is Ui (ai = r, a−i ) = vr · f (|a|r ), (1) where |a|r captures the number of agents that choose resource r in the allocation a, i.e. the cardinality of the set {i ∈ N : ai = r}, and f : N → R defines the fractional benefit awarded to each agent for selecting a given resource in allocation a.4 We will refer to f as the distribution rule throughout. Note that an agent’s utility function in (1) is consistent with the local information available as it only depends on the resource that the agent selected, the number of agents that also selected this resource, as well as the distribution rule f and relevant resource value vr . We will express such a welfare sharing game by the tuple G = {N, R, {Ai }i∈N , f, {vr }r∈R } and drop the subscripts on the above sets, e.g., denote {vr }r∈R as simply {vr }, for brevity. The goal of this paper is to derive the distribution rule f that optimizes the performance of the emergent collective behavior. Here, we focus on the concept of pure Nash equilibrium as a model for this emergent collective behavior [17]. A pure Nash equilibrium, which we will henceforth refer to as just an equilibrium, is defined as an allocation ane ∈ A such that for all i ∈ N and for all ai ∈ Ai , we have ne ne Ui (ane i , a−i ) ≥ Ui (ai , a−i ). In essence, an equilibrium represents an allocation for which no single agent has a unilateral incentive to alter its covering choice given the choices of the other agents. It is important to highlight that an equilibrium might not exist in a general game G. Nevertheless, when restricting attention to the class of games with utility functions defined in (1), an equilibrium is guaranteed to exist as the resulting game is known to be a congestion game [18].5 4 The sets N and R denote the sets of natural and real numbers, respectively. There is a rich body of literature that provides distributed decision-making mechanisms that ensures the emergent collective behavior will reach an equilibrium for the class of games considered in this paper [19, 20, 21, 22, 23]. However, we will not discuss such results due to space considerations. 5 6 We will measure the efficiency of an equilibrium allocation in a game G through two commonly studied measures, termed price of anarchy and price of stability, defined as follows: W (ane ) PoA(G) = min ≤ 1, ane ∈G W (aopt ) W (ane ) ≤ 1, PoS(G) = max ane ∈G W (aopt ) where we use the notation ane ∈ G to imply an equilibrium of the game G. In words, the price of anarchy characterizes the performance of the worst equilibrium of G relative to the performance of the optimal allocation, while the price of stability focuses on the best equilibrium in the game G. Such distinction is required as equilibria are guaranteed to exists for the class of utilities considered in (1), but in general they are not unique. By definition 0 ≤ PoA(G) ≤ PoS(G) ≤ 1. Throughout, we require that a system designer commits to a distribution rule without explicit knowledge of the agent set N , resource set R, action sets {Ai }, and resource valuations {vr }. Note that once a particular distribution rule f has been chosen, this distribution rule defines a game for any realization of the parameters. The objective of the system designer is to provide desirable performance guarantees irrespective of the realization of these parameter, even if they where chosen by an adversary. To that end, let Gf denote the family of games induced by a given distribution rule f , i.e., any game G ∈ Gf is of the above form. We will measure the quality of a distribution rule f by a worst-case analysis over the set of induced games Gf , which is the natural extension of the price of anarchy and price of stability defined above, i.e., PoA(Gf ) = min PoA(G), (2) PoS(Gf ) = min PoS(G). (3) G∈Gf G∈Gf The price of anarchy PoA(Gf ) for a given distribution rule f provides a bound on the quality of any equilibrium irrespective of the agent set N , resource set R, action sets {Ai }, and resource valuations {vr }. The price of stability, on the other hand, provides similar performance guarantees when restricting attention to the best equilibrium.6 The goal of this paper is to investigate the design of distribution rules that optimize the metrics introduced in (2) and (3). 6 One motivation for studying the price of stability is the availability of distributed learning rules that lead the collective behavior to the best equilibrium, e.g., [24, 25, 26]. 7 Related work The field of utility design has been receiving significant research attention due to applicability to multiagent systems [23, 27, 28] and social influence [29]. Utility design focuses on the design of agent objective functions geared at providing some desired features in the resulting game, e.g., existence of equilibria, efficiency of equilibria, among others. While recent work in [30] has identified all design approaches that ensure equilibrium existence in local utility designs, in general the question of optimizing the efficiency of the resulting equilibria, i.e., optimizing the price of anarchy, is far less understood. Nonetheless, there are a few positive results in this domain worth reviewing. The most notable result is [12] which focuses on the same set covering problem presented in this manuscript. In [12], the author characterizes the distribution rule f that optimizes the price of anarchy; however, there is no mention of the accompanying price of stability or the uniqueness of the prescribed rule. Further, there is no characterization of the price of anarchy for alternative distribution rules. Different setups where optimizing the price of anarchy has been explored include concave cost sharing games [31], reverse carpooling games [32], among others. The bulk of the research regarding optimal utility design has concentrated on a specific class of objectives, termed budget-balanced objectives, which imposes the constraint that the sum of the agents’ objectives is equal to the system welfare for every allocation. With regards to the objective functions considered in (1), the only design that satisfies this budget-balanced constraint is 1 fes (j) = , j ∀j ≥ 1, (4) where we use the notation “es” to signify equal share. Within the confines of budgetbalanced agent objective functions, several works have identified the optimality of the Shapley value objective design with regards to the price of anarchy guarantees [33, 34, 35].7 However, the imposition of this budget-balanced constraint is unwarranted in the context of multiagent system design and its removal allows for improved performance, as shown in [12] and in this manuscript.8 7 Informally, the Shapley value objective design extends the equal share design in (4) to more general settings beyond set covering. 8 The work of [12] demonstrates how to design a distribution rule in order to optimize the efficiency as measured by the price of anarchy. Building on these results, this manuscript is geared at quantifying the efficiency as measured by the price of stability, as well as to understand the role of information within the class of covering problems. 8 An illustrative example In this section we consider a simple family of covering problems consisting of two agents, N = {1, 2}, and three resources, R = {r1 , r2 , r3 }, to illustrate the challenges that a system designer faces when designing a distribution rule f without explicit knowledge of the action sets A1 and A2 and the resource valuations v1 , v2 , and v3 . We begin by investigating the efficiency guarantees associated with the unique budgetbalanced distribution rule in (4). Consider one problem instance where A1 = {r1 }, A2 = {r1 , r2 }, and v1 = 2, v2 = 0.99, and v3 = 0. Note that agent 1 is fixed at resource r1 for this instance. The payoff matrix and system welfare for the different allocations are: Agent 2 r1 r2 1, 1 2, 0.99 Agent 1 r1 Agent 2 r1 r2 2 2.99 Agent 1 r1 Payoff Matrix System Welfare In each box of the payoff matrix, the first entry is the payoff to agent 1 and the second entry is the payoff to agent 2. Note that for this instance, there is a unique equilibrium (r1 , r1 ) which garners a total welfare of 2. The optimal total welfare is 2.99; hence the price of anarchy and price of stability satisfy PoA(Gfes ) ≤ 2/2.99 ≈ 0.67 and PoS(Gfes ) ≤ 0.67. Consider an alternative distribution rule, termed marginal contribution, that is of the form  fmc (k) = if k = 1, otherwise. 1 0 (5) Note that employing the marginal contribution distribution rule for the problem instance drawn above induces a unique equilibrium (r1 , r2 ) which is the optimal allocation. However, consider an alternative problem instance where A1 = {r1 , r2 }, A2 = {r2 , r3 }, and v1 = 1.99, v2 = 2, and v3 = 0.01. The payoff matrix and system welfare for the different allocations are: Agent 2 Agent 1 r1 r2 r2 1.99, 2 0, 0 Agent 2 r3 1.99, 0.01 2, 0.01 Agent 1 Payoff Matrix r1 r2 r2 r3 3.99 2 2 2.01 System Welfare 9 Note that for this instance, there are now two equilibria, namely (r1 , r2 ) and (r2 , r3 ), while the optimal allocation is (r1 , r2 ). Hence the price of anarchy and price of stability must satisfy PoA(Gfmc ) ≤ 2.01/3.99 ≈ 0.5 and PoS(Gfmc ) ≤ 1. These case studies suggest that fmc outperforms fes with regards to the price of stability, while fes outperforms fmc with regards to the price of anarchy. However, whether or not these conclusion hold more generally for alternative parameters sets is, at this stage, very much unclear. While specific examples provide upper bounds on the price of stability / price of anarchy, a completely different set of techniques will have to be employed to provide lower bounds on these efficiency measures. Characterizing both the lower and upper bounds associated with these efficiency measures is essential to adequately compare the performance guarantees associated with any two distribution rules.9 Results In this section we characterize the inherent tension between the price of anarchy and price of stability as design objectives. We begin our analysis by strengthening the results in [12] to attain a tight price of anarchy for any distribution rule f . We then proceed to characterize the achievable price of stability guarantees. We further highlight the implications on the price of stability associated with requiring a desired price of anarchy. Finally, we show how to move beyond the price of anarchy / price of stability frontier given in Theorem 3 if we allow the agents to condition their choice on a higher degree of system-level information. The proofs of the forthcoming theorems will be presented in the Supplementary Materials. A tight price of anarchy Recall that Gf is the set of set covering games induced by a given distribution rule f . We now introduce a subset of these games that impose constraints on the maximum 9 A natural question is whether the locality requirement specified in the agents’ utility functions in (1) is detrimental from a performance perspective. To address this question, suppose that a system designer was able to directly set each agent’s utility function as the global welfare, i.e., Ui = W , which clearly violates our locality condition. Observe that, within the last example highlighted above, Ui = W produces the same equilibria that fmc induces. Hence, the price of anarchy is at best 0.5 for the case when Ui = W . As we will see in the ensuing section, there are alternative utility designs conforming to (1) that guarantee a better price of anarchy than 0.5. Hence, while setting Ui = W might seem a natural choice to optimize the price of anarchy, there are alternative choices that are local and yield far better results. 10 number of agents that could select a single resource. More formally, let Gfk ⊂ Gf , k ≥ 1, denote the family of games where at most k agents can select any single resource, i.e. G = {N, R, {Ai }, f, {vr }} ∈ Gfk if max |a|r ≤ k. a∈A, r∈R We will refer to the games Gfk as games with cardinality k.. Our first result strengthens the findings of [12] to obtain a tight price of anarchy for any distribution rule f over each set of games Gfk , where k ≥ 1. Theorem 1 Consider any non-increasing distribution rule f and cardinality k ≥ 1. The price of anarchy associated with the induced family of games Gfk is PoA(Gfk ) = 1 , 1 + χkf (6) where χkf = max 1≤j≤k−1 {j · f (j) − f (j + 1), (k − 1) · f (k)} . (7) Further, the optimal price of anarchy is max PoA(Gfk ) = 1 − f 1 1 (k−1)(k−1)! + Pk−1 1 , (8) i=1 i! and the unique distribution rule that achieves this price of anarchy over the induced games Gfk is, for all 1 ≤ j ≤ k k fgar (j) = (j − 1)! 1 (k−1)(k−1)! 1 (k−1)(k−1)! + Pk−1 1 ! + Pk−1 i=j i! 1 i=1 i! . (9) The importance of Theorem 1 is that it allows us to explicitly compute the price of anarchy for any given distribution rule. This characterization will be crucial in the forthcoming section, where we present the trade-off between price of anarchy and price of stability. Further, this characterization allows us to optimize the price of anarchy over admissible distribution rules f and thus to obtain (9). We will commonly refer to k fgar as Gairing’s distribution rule conditioned on instances of cardinality k to reflect its origination in [12]. 11 A tight price of stability In this section we turn our attention to the other performance metric of interest, the price of stability. Our goal here parallels our objective in the preceding section as we attempt to derive a tight price of stability for any distribution rule f . The following theorem gives such a characterization. Theorem 2 Consider any non-increasing distribution rule f and cardinality k ≥ 1. The price of stability associated with the induced family of games Gfk is   1 k PoS(Gf ) = min . (10) 1≤j≤k 1 + (j − 1) · f (j) Further, the optimal price of stability is max PoS(Gfk ) = 1, f (11) and the unique distribution rule that achieves this price of stability over the induced games Gfk is the marginal contribution distribution rule fmc defined in (5). k Figure 1 illustrates the unique distribution rule fgar that optimizes the price of anarchy given in Theorem 1 and the unique distribution rule fmc that optimizes the price of stability given in Theorem 2. A trade-off between the anarchy and stability The previous two subsections highlight how the structure of f impacts both the price of anarchy and price of stability guarantees. In this section we seek to identify the compatibility (or lack thereof) of these two measures of efficiency. That is, how do requirements on the price of anarchy impact the attainable price of stability? The following theorem provides an answer to this question. Theorem 3 Consider any cardinality k ≥ 1 and desired price of anarchy α satisfying 0 ≤ α ≤ maxf PoA(Gfk ) as defined in (8). (i) For any non-increasing distribution rule f , if PoA(Gfk ) ≥ α, then the price of stability satisfies10 PoS(Gfk ) ≤ Z k (α) , 10 The expression of Z k (α) is given in (26) in the Supplementary Materials. 12 (12) 1 fgar fmc 0.8 0.6 f (j) 0.4 0.2 0 1 2 3 4 5 6 7 8 9 10 k j fmc (j) fgar f (j) 1 1 1 2 0 0.418 3 0 0.254 4 0 0.180 5 0 0.139 6 0 0.113 7 0 0.095 8 0 0.082 9 0 0.072 10 0 0.065 j Figure 1: Optimal distribution rules. This figure illustrates the unique distribution k that optimizes the price of anarchy given in Theorem 1 and the unique distrirule fgar bution rule fmc that optimizes the price of stability given in Theorem 2; k = 10. The corresponding values are presented in the table. where Z k (α) is non-increasing in α, Z k (α) = 1 for all α ≤ 0.5, and lim Z k (1 − 1/e) = 1 − 1/e. k→∞ (ii) The bound in (12) is tight, i.e., maxf :PoA(Gfk )≥α PoS(Gfk ) = Z k (α). The results of Theorem 3 are illustrated in Figure 2 where we highlight the plausible price of anarchy / price of stability region. In particular, Theorem 3-(i) establishes that there does not exist a distribution rule f that attains a price of stability and price of anarchy in the red region of the figure. Hence, there is an inherent tension between these two measures of efficiency as improving the performance of the worst equilibria necessarily comes at the expense of the best equilibria, and vice versa. The expression of Z k (α) defines this trade-off. The second claim of Theorem 3 guarantees that the bound obtained in (12) is tight, i.e., there are distribution rules that achieve price of anarchy and price of stability guarantees that lie exactly on the boundary of the red region. Adding Information The previous section highlights a fundamental tension between the price of stability and price of anarchy for the given covering problem when restricted to local agent objective 13 1 not achievable 0.8 Z(α) 0.6 (1 − 1e , 1 − 1e ) PoS 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 PoA Figure 2: Tradeoff between anarchy and stability. The figure provides an illustration of the inherent trade-off between the price of anarchy and price of stability. For simplicity, we define Z(α) = limk→∞ Z k (α) and illustrate this trade-off in the case of large k. First, note that the grey region is not achievable since 0 ≤ PoA(Gfk ) ≤ PoS(Gfk ) ≤ 1 by definition. Theorem 3 demonstrates that the red region is also not achievable. That is, there does not exist a distribution rule with joint price of anarchy and price of stability guarantees in the red region. For example, if α ≤ 1/2, then a price of stability of 1 is attainable while meeting this price of anarchy demand. However, if α = 1 − 1/e, then a price of stability of 1 is no longer attainable. In fact, the best attainable price of stability is now also 1 − 1/e. functions of the form (1). In this section, we challenge the role of locality in these fundamental trade-offs. That is, we show how to move beyond the price of anarchy / price of stability frontier given in Theorem 3 if we allow the agents to condition their choice on a higher degree of system-level information. With this goal in mind, we introduce a minimal and easily attainable piece of systemlevel information that can permit the realization of decision-making rules with efficiency guarantees beyond this frontier. To that end, for each allocation a ∈ A we define the information flow graph (V, E) where each node of the graph represents an agent and we 14 Agent i 1 2 3 4 5 Ni (a) {2, 3, 4} {3, 4} {2, 4} {2, 3} ∅ Qi (a) A1 ∪ A2 ∪ A3 ∪ A4 A2 ∪ A 3 ∪ A4 A2 ∪ A 3 ∪ A4 A2 ∪ A 3 ∪ A 4 A5 Figure 3: Example of graph (V, E). The sets Ni (a) and Qi (a) are enumerated for each agent in the table on the right. Observe that the graph (V, E) as well as the sets Ni (a) and Qi (a) are allocation dependent. construct a directed edge i → j if ai ∈ Aj for i 6= j (no self loops). Based on this allocation-dependent graph, we define for each agent i the set Ni (a) ⊆ N consisting of all the agents that can reach i through a path in the graph (V, E). Similarly, for each agent i we define Qi (a) = ∪j∈Ni (a) Aj ∪ Ai , (13) which consists of the union of Ai and all the sets of other agents that can reach i through a path in the graph. An example is shown in Figure 3. Lastly, building upon this graph we define the following two quantities: Vi (a) = xi (a) = max vr , (14) max vr . (15) r∈Ai \a−i r∈Qi (a)\a The term Vi (a) captures the highest valued resource in agent i’s choice set Ai that is not covered by any agent. If the set Ai \ a−i is empty, we set Vi (a) = 0. Similarly, the term xi (a) captures the highest-valued resource in the enlarged set Qi (a) not currently covered by any other agent. If the set Qi (a) \ a is empty, we set xi (a) = 0. We are now ready to specify the information based game. As before, we have a set of agents N and each agent has an action set Ai ⊆ R. Here, we consider a statebased distribution rule that toggles between the two extreme optimal distribution rules: Gairing’s rule and the marginal contribution rule. More formally, the distribution rule for agent i is now of the form ( fmc (|a|r ) if Vi (a) ≥ xi (a), fik (ai = r, a−i ) = (16) k fgar (|a|r ) otherwise, 15 and the corresponding utilities are given by Ui (ai = r, a−i ) = vr · fik (a), as we allow the system-level information xi (a) and Vi (a) to prescribe which distribution rule each agent k applies. We denote with fsb = {fik }i∈N the collection of distribution rules in (16) and informally refer to it as to the state-based distribution rule.11 k The next theorem demonstrates how fsb attains performance guarantees beyond the price of stability / price of anarchy frontier established in Theorem 3. k Theorem 4 Consider any cardinality k ≥ 1 and let fsb = {fik }, where fik is defined in (16). First, an equilibrium is guaranteed to exist in any game G ∈ Gfkk . Furthermore, sb the price of anarchy and price of stability associated with the induced family of games Gfkk sb is PoS(Gfkk ) =1, sb PoA(Gfkk ) = max PoA(Gfk ). sb f We remark on Theorem 4 for the case when k → ∞ for ease of exposition. Note that for this case, a consequence of attaining a price of anarchy of 1−1/e was a price of stability also of 1 − 1/e and this was achieved using Gairing’s distribution rule in (9). Using the state-based rule given in (16), a system designer can achieve the optimal price of anarchy without any consequences for the price of stability. Hence, the identified piece of systemlevel information was crucial for moving beyond the inherited performance limitations by adhering to our notion of local information. Whether alternative forms of system-level information could move us beyond these guarantees, or achieve these guarantees with less information, is an open research question. Conclusions How should a system operator design a networked architecture? The answer to this question is non-trivial and involves weighing the advantages and disadvantages associated with different design choices. In this paper we highlight one novel trade-off pertaining to the price of anarchy and price of stability in set covering problems with local information. Further, we demonstrate how a system designer can move beyond these limitations by equipping the agents with additional information about the system. Fully realizing 11 We express the distribution rule as fik (a) instead of merely fik (xi (a), Vi (a)) for brevity. 16 the potential of multiagent systems requires to pursue a more formal understanding of the inherent limitations and performance trade-offs associated with networked architectures. While this paper focused purely on two performance measures, other metrics of interest include convergence rates, robustness to adversaries, fairness, among others. In each of these settings, it is imperative that a system operator fully understands the role of information within these trade-offs. Only then, will a system operator be able to effectively balance the potential performance gains with the communication costs associated with propagating additional information through the system. References [1] M. J. B. Krieger, J.-B. Billeter, L. Keller, Ant-like task allocation and recruitment in cooperative robots, Nature 406, 992–995 (2000). [2] B. P. Gerkey, M. J. Mataric, A formal analysis and taxonomy of task allocation in multi-robot systems, The International Journal of Robotics Research 23, 939–954 (2004). [3] A. S. L. El Yasmine, B. A. Ghani, D. Trentesaux, B. Bouziane, Supply Chain Management Using Multi-Agent Systems in the Agri-Food Industry (Springer, 2014), pp. 145–155. [4] D. Vallejo, P. Remagnino, D. N. Monekosso, L. Jiménez, C. González, A Multi-agent Architecture for Multi-robot Surveillance (Springer, 2009), pp. 266–278. [5] K. Alexis, G. Nikolakopoulos, A. Tzes, L. Dritsas, Coordination of Helicopter UAVs for Aerial Forest-Fire Surveillance (Springer, 2009), pp. 169–193. [6] S. Martinez, J. Cortes, F. Bullo, Motion coordination with distributed information, Control Systems Magazine 27, 75–88 (2007). [7] F. Bullo, J. Cortés, S. Martı́nez, Distributed Control of Robotic Networks (Princeton University Press, 2009). [8] J. Marden, IEEE Transactions on Control of Network Systems , (2016). Available at http://ieeexplore.ieee.org/document/7451235. (paper published online before print). 17 [9] J. R. Marden, A. Wierman, Distributed welfare games, Operations Research 61, 155–168 (2013). [10] E. Anshelevich, et al., The price of stability for network design with fair cost allocation, in Proceedings of the 45th Annual IEEE Symposium on Foundations of Computer Science (2004), pp. 295–304. [11] E. Koutsoupias, C. Papadimitriou, Worst-Case Equilibria (Springer, 1999), pp. 404– 413. [12] M. Gairing, Covering Games: Approximation through Non-cooperation (Springer, 2009), pp. 184–195. [13] G. Nemhauser, L. Wolsey, M. Fisher, An analysis of approximations for maximizing submodular set functions – I, Mathematical Programming 14, 265–294 (1978). [14] G. Nemhauser, L. Wolsey, M. Fisher, An analysis of approximations for maximizing submodular set functions – II (Springer), pp. 73–87. [15] U. Feige, J. Vondrak, Approximation algorithms for allocation problems: Improving the factor of 1 - 1/e, in Proceedings of the 47th Annual IEEE Symposium on Foundations of Computer Science (FOCS’06) (2006), pp. 667–676. [16] A. Krause, C. Guestrin, Near-optimal observation selection using submodular functions, in Proceedings of the 22nd Conference on Artifical Intelligence (2007). [17] J. F. Nash, Equilibrium points in n-person games, Proceedings of the National Academy of Sciences 36, 48–49 (1950). [18] D. Monderer, L. S. Shapley, Potential games, Games and Economic Behavior 14, 124 – 143 (1996). [19] H. P. Young, Individual Strategy and Social Structure: An Evolutionary Theory of Institutions (Princeton University Press, 1998). [20] H. P. Young, Strategic Learning and its Limits (Oxford University Press, 2005). [21] D. Fudenberg, D. Levine, The Theory of Learning in Games (MIT Press, 1998). 18 [22] J. R. Marden, G. Arslan, J. S. Shamma, Joint strategy fictitious play with inertia for potential games, IEEE Transactions on Automatic Control 54, 208–220 (2009). [23] J. Marden, J. Shamma, Handbook of Game Theory Vol. 4 (Elsevier, 2014), pp. 861– 899. [24] J. R. Marden, J. S. Shamma, Revisiting log-linear learning: Asynchrony, completeness and a payoff-based implementation, Games and Economic Behavior 75, 788–808 (2012). [25] L. Blume, The statistical mechanics of strategic interaction, Games and Economic Behavior 5, 387–424 (1993). [26] L. Blume, The Economy as an evolving complex system II (Addison-Wesley, 1997), pp. 425–460. [27] J. R. Marden, G. Arslan, J. S. Shamma, Connections between cooperative control and potential games, IEEE Transactions on Systems, Man and Cybernetics. Part B: Cybernetics 39, 1393–1407 (2009). [28] D. Paccagnan, J. R. Marden, The risks and rewards of conditioning noncooperative designs to additional information, in Proceedings of the 55th IEEE Allerton Conference on Communication, Control, and Computing (2017). [29] P. N. Brown, J. R. Marden, Studies on robust social influence mechanisms: Incentives for efficient network routing in uncertain settings, IEEE Control Systems 37, 98–115 (2017). [30] R. Gopalakrishnan, J. R. Marden, A. Wierman, Potential games are necessary to ensure pure nash equilibria in cost sharing games, Mathematics of Operations Research 39, 1252–1296 (2014). [31] J. R. Marden, M. Philips, Optimizing the price of anarchy in concave cost sharing games, in Proceedings of the 2017 American Control Conference (2017), pp. 5237– 5242. [32] J. R. Marden, M. Effros, The price of selfishness in network coding, IEEE Transactions on Information Theory 58, 2349–2361 (2012). 19 [33] P. von Falkenhausen, T. Harks, Optimal cost sharing for resource selection games, Mathematics of Operations Research 38, 184–208 (2013). [34] P. von Falkenhausen, T. Harks, Optimal cost sharing for capacitated facility location games, European Journal of Operations Research 239, 187–198 (2014). [35] H.-L. Chen, T. Roughgarden, G. Valiant, Designing network protocols for good equilibria, SIAM Journal on Computing 39, 1799–1832 (2010). Supplementary Materials Proof of Theorem 1 The distribution rule maximizing the price of anarchy and the corresponding optimal value of (8) are derived in [12] by means of a different approach. Thus, in the following we only show the tight bound of (6). Proof: For any non increasing distribution rule, [12] proves that the price of anarchy satisfies PoA(Gfk ) ≥ 1 1+χkf . We will show that PoA(Gfk ) ≤ 1 1+χkf holds too, which will complete the proof. To do so, note that by definition χkf = max {j · f (j) − f (j + 1), (k − 1) · f (k)}. 1≤j≤k−1 Hence, we can distinguish two cases. For each of them we will construct an instance of covering game that achieves PoA(Gfk ) = 1 , 1+χkf thus completing the proof. i) χkf = j · f (j) − f (j + 1) for some 1 ≤ j ≤ k − 1. First, suppose f (j) > f (j + 1) and define γ = f (j + 1)/f (j) < 1. Consider the instance given in Figure S1, where there are j agents at each of the m + 1 levels (m > 0). Each agent is represented by a solid line and the circles at either end of the line indicate the choices available to that agent. In particular, in Level 0 there are j agents and j + 1 resources of which j valued at γ 0 f (j) and a single one valued at γ 0 . At each subsequent level k > 0, there are j agents and j resources of which j − 1 valued at γ k f (j) and a single one valued at γ k . Further, at each level k > 0 one of the agents can choose between the common element at level k or the common at level k − 1. The equilibrium allocation ane consists of all the agents 20 0 f (j) 0 f (j) Level 0 1 f (j) 1 f (j) f (j) 0 f (j) m f (j) Level m m f (j) 0 0 ... Level 1 m 1 1 m f (j) f (j) = f (j+1) f (j) Figure S1: First instance used in case i). selecting the common resource to their right in Figure S1. This produces a welfare m+1 of W (ane ) = 1 + γ + · · · + γ m = 1−γ , which in the limit of m → ∞ reduces to 1−γ W∞ (ane ) = 1 . 1−γ The optimal allocation aopt covers all resources, except one of those valued γ m f (j), yielding a total welfare of W (aopt ) = W (ane ) + W (ane )(j − 1)f (j) + f (j)(1 − γ m ), which in the limit of m → ∞ reduces to W∞ (aopt ) =W∞ (ane )(1 + (j − 1)f (j)) + f (j) =W∞ (ane )(1 + jf (j) − f (j + 1)) =W∞ (ane )(1 + χkf ) , from which the desired result follows: W∞ (ane ) W∞ (aopt ) = 1 . 1+χkf Now, suppose f (j) = f (j + 1). Consider the instance given in Figure S2, consisting of j agents and j + 1 resources, one of them valued at 1 and j − 1 valued at f (j). One equilibrium allocation ane consists of all the agents selecting the resource to their right, producing a welfare of W (ane ) = 1. On the other hand, the optimum yields W (aopt ) = (j − 1)f (j) + 1 = (jf (j) − f (j + 1) + 1) = χkf + 1, and hence W (ane ) 1 = 1+χ k. W (aopt ) f 21 0 f (j) 0 f (j) 1 m f (j) f (j) = f (j) f (j) 1 f (j) f (j) Figure S2: Second instance used in case i). ii) χkf = (k − 1)f (k). The second instance discussed in the previous case applies here as well, with j = k. This completes the proof.  Proof of Theorem 2 The distribution rule maximizing the price of stability and the corresponding optimal value of (11) are trivial to obtain, once the result in (10) is proven. Before proceeding with showing the latter, we observe that we can restrict our attention to games G ∈ Gfk where there is an optimal allocation that is disjoint, i.e., there is an optimal allocation aopt such that aopt 6= aopt for any i 6= j. The following claim, stated without proof, formalizes i j this observation. Claim 1 Let Ḡfk ⊆ Gfk be the set of games where each game G ∈ Ḡfk has an optimal allocation that is disjoint. The price of stability and price of anarchy satisfy PoA(Ḡfk ) = PoA(Gfk ), PoS(Ḡfk ) = PoS(Gfk ). Throughout the remainder of the proof, we will concentrate on games that possess 22 f an optimal allocation that is disjoint. We will continue to denote the game set as Gfk as opposed to Ḡfk for ease of presentation. We will prove the above price of stability result through a series of intermediate claims. A central piece of the forthcoming analysis is the focus on a particular sequence of allocations a0 , a1 , . . . , am derived from a series of unilateral deviations of players switching from their choice in the equilibrium allocation to their choice in the optimal allocation. We begin by observing that any game G in the class Gfk is a congestion game, and thus is a potential game as introduced in [18], with a potential function φ : A → R of the form12 φ(a) = |a|r XX vr · f (j). (17) r∈R j=1 It is well-known that an equilibrium is guaranteed to exist in any potential game [18], and one such equilibrium is the allocation that optimizes the potential function φ, i.e., ane ∈ arg maxa∈A φ(a). Focusing on this specific equilibrium, we consider a sequence of k−1 allocations taking the form a0 = ane and ak = (aopt i(k) , a−i(k) ) for all k ∈ {1, . . . , m} where i(k) is the deviating player in the k-th profile. The selection of the deviating players I = {i(1), . . . , i(m)} is chosen according to the following rules: (i) Let i(1) ∈ N be any arbitrary player. opt ne (ii) For each k ≥ 1, if aopt / ane then the sequence is terminated. i(k) = ai (1) or ai(k) ∈ (iii) Otherwise, let i(k + 1) be any agent in the set {j ∈ N : akj = aopt i(k) } and repeat. opt Claim 2 Define Q = ∪i∈I ane i and Q̄ = ∪i∈I ai . Then X X X Ui (ane ) ≥ vr f (|ane |r ) + vr . i∈I r∈Q∩Q̄ (18) r∈Q̄\Q Proof: We begin with two observations on the above sequence of allocations: (a) the sequence of allocations can continue at most n steps due to the disjointness of aopt and opt (b) if the sequence continues, it must be that for player i(k+1) ane i(k+1) = ai(k) . Observation (b) ensures us that m−1 X Ui(k+1) (ak ) − Ui(k) (ak ) = 0. k=1 12 A proof of this can be found in [18], where the potential function φ is also defined. 23 (19) Accordingly, we have that φ(a0 ) − φ(am ) = = m−1 X k=0 m−1 X φ(ak ) − φ(ak+1 ) Ui(k+1) (ak ) − Ui(k+1) (ak+1 ) k=0 0 m 0 m = Ui(1) (a ) − Ui(m) (a ) + m−1 X Ui(k+1) (ak ) − Ui(k) (ak ) k=1 = Ui(1) (a ) − Ui(m) (a ) ≥ 0 . The first and third equalities follow by rearranging the terms. The second equality can be shown using the definition of φ as in (17); the last equality follows by (19). The inequality derives from the fact that a0 = ane optimizes the potential function. Thanks to observation (b), one can show that opt Q \ Q̄ = ane i(1) \ ai(m) ne and Q̄ \ Q = aopt i(m) \ ai(1) . opt opt ne If Q \ Q̄ 6= ∅, it must be that ane i(1) 6= ai(m) so that Q \ Q̄ = ai(1) and Q̄ \ Q = ai(m) . Using opt opt ane / ane and thus the resource aopt i(1) 6= ai(m) in condition (ii), tells us that ai(m) ∈ i(m) is not chosen by anyone else in the allocation am . Thus, when Q \ Q̄ 6= ∅, X X Ui(1) (a0 ) − Ui(m) (am ) = vr f (|ane |r ) − vr ≥ 0. r∈Q\Q̄ (20) r∈Q̄\Q When Q \ Q̄ = ∅, also Q̄ \ Q = ∅ and thus the previous inequality still holds. Rearranging P the terms and adding r∈Q∩Q̄ vr f (|ane |r ) to each side gives us X X X vr f (|ane |r ) ≥ vr f (|ane |r ) + vr . (21) r∈Q r∈Q∩Q̄ r∈Q̄\Q Finally note that X i∈I Ui (ane ) ≥ X vr f (|ane |r ) r∈Q which together with (21) completes the proof.  Our second claim shows that there exist a collection of disjoint sequences that covers all players in N . We will express a sequence merely by the deviating player set I with the understanding that this set uniquely determines the sequence of allocations. 24 Claim 3 There exists a collection of deviating players I 1 , . . . , I p chosen according the process described above such that ∪k I k = N and I j ∩ I k = ∅ for any j 6= k. Proof: Suppose I 1 , I 2 , . . . , I k represent the first k sequences of deviating players. Further assume that they are all disjoint. Choose some player i ∈ N \ ∪k I k to start the (k + 1)th sequence. If no such player exists, we are done. Otherwise, construct the sequence according to the process depicted above. If the sequence terminates without selecting a player in ∪k I k , then repeat this process to generate the (k + 2)-th sequence. Otherwise, let ik+1 (j), j ≥ 2, denote the first player in the (k + 1)-th sequence contained in the set ∪k I k . Since aopt 6= aopt (for i 6= j), this player must be contained in the set ∪k ik (1), i.e., i j the first player in a previous sequence. Suppose this player is i` (1), where ` ∈ {1, . . . , k}. If this is the case, replace the `-th sequence with {ik+1 (1), . . . , ik+1 (j − 1), I ` } which is a valid sequence and disjoint from the others. Then repeat the process above to choose the (k + 1)-th sequence. Note that this process can continue at most n-steps and will always result in a collection of disjoint sequences that cover all players in N . This completes the proof.  In the following we complete the proof of Theorem 2, by means of Claims 2 and 3. Proof: We being showing a lower bound on the price of stability. Let I 1 , . . . , I p denote a collection of deviating players that satisfies Claim 3. Further, let Qk and Q̄k be defined as above for each sequence k = 1, . . . , p. Using the result (18) from Claim 2, we have X ne Ui (a ) = p X X Ui (ane ) k=1 i∈I k i∈N ≥ p X X  k=1 =   vr f (|ane |r ) + r∈Qk ∩Q̄k X X vr f (|ane |r ) + r∈aopt ∩ane vr  r∈Q̄k \Qk X vr , r∈aopt \ane where the above equality follows from the fact that Q̄i ∩ Q̄j = ∅ for any i 6= j which is due to the disjointness of aopt . Using the definition of Ui (ane ), we have X X X vr |ane |r f (|ane |r ) + vr (|ane |r − 1) f (|ane |r ) ≥ r∈ane ∩aopt r∈ane \aopt r∈aopt \ane Define γ = maxj≤n (j − 1)f (j). Working with the above expression we have X X X vr (γ + 1) + vr γ ≥ vr , r∈ane \aopt r∈ane ∩aopt 25 r∈aopt \ane vr . which gives us that (γ + 1)W (ane ) ≥ W (aopt ) which completes the lower bound. We will now provide an accompanying upper bound on the price of stability. To that end, consider a family of examples parameterized by a coefficient j ∈ {1, . . . , k}. For each j, the game consists of j agents and (j +1)-resources R = {r0 , r1 , . . . , rj } where the values of the resources are vr0 = 1 and vr1 = · · · = vrj = f (j) − ε where ε > 0 is an arbitrarily small constant, and the action set of each agent i ∈ {1, . . . , j} is Ai = {r0 , ri }. The unique equilibrium is of the form ane = (r0 , . . . , r0 ) as every agent selects resource r0 and the total welfare is W (ane ) = 1. The optimal allocation is of the form aopt = (r0 , r2 , . . . , rj ) which generates a total welfare of W (aopt ) = 1 + (j − 1)(f (j) − ε). Performing a worst case analysis over ε and j gives   1 PoS(Gf ) ≤ min 1≤j≤k 1 + (j − 1) · f (j) which completes the proof.  Proof of Theorem 3 Observe that χkf defined in (7) can be equivalently characterized as the solution to the the following optimization problem χkf = min x s. t j · f (j) − f (j + 1) ≤ x, 1 ≤ j ≤ k − 1, (k − 1) · f (k) ≤ x. This allows us to state the following corollary, which will be useful to prove Theorem 3. Corollary 1 Let Fαk , α ∈ (0, 1), denote the family of distribution rules that guarantee a price of anarchy of at least α in games with cardinality k ≥ 1, i.e., Fαk = {f : PoA(Gfk ) ≥ α}. Then, the set Fαk is uniquely defined by the set of distribution rules f that satisfy the following k-linear equations, 1 − 1, α 1 (k − 1)f (k) ≤ − 1. α j · f (j) − f (j + 1) ≤ 26 ∀j ≤ k − 1, Finally, let us define PoS(G k ; α) = max PoS(Gfk ) k f ∈Fα (22) as the best achievable price of stability given that the price of anarchy is guaranteed to exceed α, where α ∈ (0, 1). We will now state an enriched version of Theorem 3 using the notation just introduced, as this allows for a more precise proof. Theorem 5 Let k ≥ 1. The function PoS(G k ; α) satisfies the following: i) For any α ≤ 1/2, PoS(G k ; α) = 1. ii) For any α satisfying 1 2 < α ≤ maxf PoA(Gfk ), PoS(G k ; α) = 1 , 1 + ξαk where ξαk = max (j − 1)(j − 1)! 1 − j≤k  ! X j−1 1 1 −1 . α i! i=1 (23) (24) iii) Lastly, if α = maxf PoA(Gfk ) then PoS(G k ; α) = max PoA(Gfk ), f (25) and limk→∞ PoS(G k ; α) = 1 − 1e . Thus, the function Z k (α) introduced in Theorem 3 takes the form ( 1 if α ≤ 12 Z k (α) = 1 1 . if α > k 2 1+ξ (26) α Proof: i) Consider the marginal contribution distribution rule fmc , defined in (5). It is straightforward to verify that χkfmc = 1 and hence fmc ∈ Fαk for any α ≤ 1/2. Further, it is well-known that PoS(Gfkmc ) = 1, see [23]. This completes the first part of the proof. ii) We shift our attention to the price of stability. From Theorem 2 and (22) we have   1 k . (27) PoS(G ; α) = max min k j≤k 1 + (j − 1) · f (j) f ∈Fα 27 The optimization problem in (27) is equivalent to min max (j − 1) · f (j) f j≤k   1 s.t. j · f (j) − − 1 ≤ f (j + 1), α   1 (k − 1)f (k) ≤ −1 . α ∀j ≤ k − 1, We find an optimal solution by first solving for the distribution rule that satisfies the k −1 linear inequalities with equality, followed by verifying that the resulting distribution rule is non increasing. Such distribution rule can be computed recursively, and is of the form ( ! )  X j−1 1 1 f (j) = max (j − 1)! 1 − −1 ,0 . α i! i=1 Using Theorem 2 on such distribution rule, gives the desired result in (23). iii) The final result can be proven as follows: replace α in (24) with the optimal price of anarchy of (8), i.e., set α=1− 1 1 (k−1)(k−1)! + Pk−1 1 . i=1 i! After some manipulation, ξαk in (24) becomes ξαk = (1 − α) = 1 1 (k−1)(k−1)! Thanks to (23), the price of stability reads as Pk−1 1 + i=1 (k−1)(k−1)! PoS(G k ; α) = Pk−1 1 + i=0 (k−1)(k−1)! + 1 i! 1 i! Pk−1 1 . i=1 i! = max PoA(Gfk ) , f proving the claim in (25). Taking the limit k → ∞ in the previous expression, gives the final result.  Proof of Theorem 4 Proof of PoS result We begin our proof with a lemma that identifies a structure about the state based distribution rule {fik }. 28 Lemma 1 Let a be any allocation. Then for each agent i ∈ N , one of the following two statements is true. - Ui (a0i = r, a−i ) = vr · fmc (|a−i | + 1), ∀a0i ∈ Ai ; or - Ui (a0i = r, a−i ) = vr · fgar (|a−i | + 1), ∀a0i ∈ Ai . Informally, this lemma states that for a given allocation a, the state based distribution rule will either evaluate every resource at the marginal contribution distribution or every resource at the Gairing distribution rule for a given agent i. Proof: Let a be any allocation and i be any agent. Extend the definition of xi (a) in (15) as yi (a) = zi (a) = max r∈(Qi (a)\a)∩Ai max r∈(Qi (a)\a)\Ai vr , (28) vr , (29) and note that xi (a) = max{yi (a), zi (a)}. First observe that for any a0i ∈ Ai : (i) yi (a) ≤ Vi (a); (ii) zi (a) = zi (a0i , a−i ); and (iii) Vi (a) = Vi (a0i , a−i ). Accordingly, xi (a) > Vi (a) if and only if xi (a) = zi (a). Consequently, xi (a) > Vi (a) if and only if xi (a0i , a−i ) > Vi (a0i , a−i ) which completes the proof.  We will now prove that an equilibrium exists and the price of stability is 1. In particular, we will show that the optimal allocation aopt is in fact an equilibrium. Proof: Let aopt be an optimal allocation. We begin by showing that Vi (aopt ) ≥ xi (aopt ) for all i ∈ N . Suppose this was not the case, and there exists an agent i such that Vi (aopt ) < xi (aopt ). This implies that there exists a resource r ∈ Qi (aopt ) \ aopt such that vr > vr̃ = maxr0 ∈Ai :|aopt |=0 vr0 . By definition of Qi (a) and Ni (a) there exists a sequence −i of players {i0 , i1 , . . . , im } such that aopt ik ∈ Aik−1 for all k ≥ 1, and r ∈ Aim . Hence, consider a new allocation where aik = aopt ik+1 for all k ∈ {0, . . . , m − 1} and aim = r, where aj = aopt for all other agents j ∈ / {i0 , i1 , . . . , im }. The welfare of this allocation j is W (a) ≥ W (aopt ) + vr − vr̃ > W (aopt ), which contradicts the optimality of aopt . This means that every agent will be using the marginal cost distribution rule to evaluate its utility in the allocation aopt . Now, suppose aopt is not an equilibrium for sake of contradiction. This means, that opt there exists an agent i with an action ai ∈ A such that Ui (ai , aopt ). Since −i ) > Ui (a 29 agents are using the marginal contribution distribution rule, which follows from Vi (aopt ) ≥ xi (aopt ), we have opt opt W (ai , aopt ) − Ui (aopt ) + Ui (ai , aopt ), −i ) = W (a −i ) > W (a which contradicts the optimality of aopt . This completes the proof.  Informal discussion and proof of PoA result Consider a game G = (N, R, {Ai }, {fik }, {vr }) with cardinality bounded by k. Let ane be any equilibrium of the game G. A crucial part of the forthcoming analysis will center on a new game G0 derived from the original game G and the equilibrium ane , i.e., (G, ane ) −→ G0 . This new game G0 possesses the identical player set, resource set, and valuations of the resources as the game G. The difference between the games are (i) the action sets and (ii) k , as opposed to the state-based the new game G0 employs the Gairing distribution rule, fgar k distribution rule fsb . Informally, the proof proceeds in the following two steps: – Step 1: We prove that the equilibrium ane of G is also an equilibrium of G0 . Since the player set, resource set, and valuations of the resources are unchanged we have that W (ane ; G0 ) = W (ane ; G), where we write the notation W (ane ; G0 ) to mean the welfare accrued at the allocation ane in the game G0 . 0 – Step 2: We show that the optimal allocation aopt in the new game G0 is at least as good as the optimal allocation in the original game G0 , i.e., 0 W (aopt ; G0 ) ≥ W (aopt ; G). Combining the results from Step 1 and Step 2 give us W (ane ; G) W (ane ; G0 ) ≥ ≥ PoA(Gfkgar k ), 0 opt opt 0 W (a ; G) W (a ; G ) k where the last inequality follows from Theorem 1 since G0 employs fgar . 30 Construction of game G0 : We will now provide the construction of the new game G0 from the game G and the equilibrium ane . We begin with some notation that we will use to construct the new agents’ action sets in the game G0 . For each i ∈ N , let ri = ane i and define k k Hi = {r ∈ Ai : vri · fgar (|ane |ri ) < vr · fgar (1 + |ane −i |r )}, (30) to be the set of resources that will give a strictly better payoff to agent i if the agent was k k to use fgar instead of fsb . Further, denote I = {i ∈ N : Hi 6= ∅}, (31) k as the set of agents that would move from the equilibrium ane if they were to use fgar k instead of fsb . Finally, for each agent i ∈ I, let Bi = {r0 ∈ Qi (ane ) : |ane |r0 = 0} , (32) i.e. the enlarged set of resources of (13) that are not chosen by anyone at equilibrium. We are now ready to construct the new game G0 . As noted above, the agent set, resource set, and resource valuations are identical to those in G. The new action set of each agent i in game G0 is defined as ( (Ai \Hi ) ∪ Bi ∪ ∅ if i ∈ I , A0i = Ai otherwise . k . We denote this game G0 Lastly, the utility functions of the agents are derived using fgar k by the tuple G0 = (N, R, {A0i }, fgar , {vr }). Formal Proof of Step 1: Proof: In the first part of the proof, we will establish that the equilibrium allocation ane of game G is also an equilibrium allocation of game G0 . By definition of the action sets 0 {A0i }, we know that ane i ∈ Ai for all agents i ∈ N . It remains to show that for any agent i∈N 0 0 0 ne 0 0 ne Ui (ane i , a−i ; G ) ≥ Ui (ai , a−i ; G ), ∀ai ∈ Ai , (33) where we use the notation Ui (a; G0 ) to denote the utility of agent i for allocation a in the game G0 . 31 Based on the above definition of the action sets {A0i }, we only need to concentrate our attention on agents i ∈ I with choices a0i ∈ Bi , as (33) follows immediately in the other cases. Since ane is an equilibrium of game G, then by Lemma 1 we know that each agent i ∈ I employs the marginal contribution distribution rule everywhere. Accordingly, we have ne ne ne Ui (ane i , a−i ; G) = maxne vr = Vi (a ) ≥ xi (a ), r∈Ai \a−i (34) where the first equality follows from the equilibrium conditions coupled with the use of the marginal contribution distribution rule and the inequality follows from the use of the marginal contribution distribution rule and (16). ne We will conclude the proof by a case study on the potential values of Ui (ane i , a−i ; G). For brevity in the forthcoming arguments we let r = ane i . ne ne – Case (i): Suppose Ui (ane i , a−i ; G) = 0. In this case, we have that xi (a ) = 0 from ne 0 (34) which implies that any resource r0 ∈ Bi has value vr0 = 0. Hence, Ui (ane i , a−i ; G ) = 0 Ui (r0 , ane −i ; G ) = 0 and we are done. ne – Case (ii): Suppose Ui (ane i , a−i ; G) > 0. Based on the definition of the marginal cost distribution rule, this implies that |ane |r = 1, and hence k Ui (ane ; G) = vr fmc (|ane |r ) = vr fgar (|ane |r ) = Ui (ane ; G0 ), k (1) = 1. Hence, Ui (ane ; G0 ) = vr . For any which follows from the fact that fmc (1) = fgar r0 ∈ Bi , we have 0 k ne ne Ui (r0 , ane i ; G ) = vr0 fgar (1 + |a−i |r0 ) = vr0 ≤ xi (a ), where the last inequality follows for the definition of xi (ane ). Combining with (34) gives us ne 0 0 ne 0 Ui (ane i , a−i ; G ) ≥ Ui (r , a−i ; G ), which completes the proof.  Formal Proof of Step 2: We begin with a lemma that highlights a structure associated with the action sets {A0i } in the new constructed game G0 . 32 Lemma 2 If r ∈ A0i \ Ai for some agent i ∈ N and resource r ∈ R, then there exists an 0 agent j 6= i such that ane j = r and consequently r ∈ Aj . Proof: Suppose r ∈ A0i \ Ai for some agent i ∈ N and resource r ∈ R. Then, r ∈ Hi by definition of the set A0i . By Lemma 1, each agent must either be a marginal contribution agent, i.e., uses fmc at all resources, or a Gairing agent, i.e., uses fgar at all resources. Since, i ∈ I and ane is an equilibrium, agent i must be a marginal contribution agent, i.e., ne ne ne Ui (ane i = r̃, a−i ) = vr̃ · fmc (|a |r̃ ) ≥ vr · fmc (|a−i |r + 1), where the inequality follows from the equilibrium conditions. Suppose by contradiction that |ane −i |r = 0. In this case we have ne vr · fmc (|ane −i |r + 1) = vr · fgar (|a−i |r + 1). which follows from the fact that f mc (1) = f gar (1) = 1. Since fgar (k) ≥ fmc (k) for all k ≥ 1, this implies vr̃ · fgar (|ane |r̃ ) ≥ vr̃ · fmc (|ane |r̃ ) ≥ vr · fgar (|ane −i |r + 1). Hence, r ∈ / Hi leading to the contradiction. This completes the proof.  We exploit the result of Lemma 2 to prove Step 2. Proof: We conclude the proof by constructing an allocation a ∈ A0 that satisfies W (a; G0 ) = W (aopt ; G) where aopt ∈ arg maxa∈A W (a; G). We begin with an initial allocation a where for each agent i ∈ N ( aopt i a= ∅ if aopt ∈ A0i , i else , That is, we assign each agent the agent’s optimal allocation choice if it is available to them in the new action set A0i . If all agents received their optimal choice, then the proof is complete. If this is not the case, then there will be a set of uncovered resources U(a) = {r ∈ aopt : |a|r = 0} which we denote by U(a) = {r1 , . . . , rm }. We will now argue that we can construct a new allocation a0 that covers one additional resource from the set U(a), i.e., |U(a0 )| = |U(a)| − 1 and a ⊆ a0 , where we denote with |U(a)| the cardinality of U(a). To that end, consider any uncovered resource r0 ∈ U(a). By definition, there exists an agent i0 ∈ N such that aopt / A0i0 . Consequently, we have that r0 ∈ Hi0 and i0 = r0 but r0 ∈ 33 ne by Lemma 2 we know that there exists an agent i1 6= i0 such that ane i1 = r0 . Since ai1 = r0 we also have that r0 ∈ A0i1 by definition. We now analyze the following three cases: – Case 1: Suppose ai1 = ∅. Then define a new allocation a0i1 = r0 and a0j = aj for all j 6= i1 , and we are done. – Case 2: Suppose ai1 = r1 and |ane |r1 = 0, meaning that there are no agents at the resource r1 in the equilibrium allocation. Then by definition r1 ∈ Bi0 and r1 ∈ A0i0 . Define the allocation a0i0 = r1 , a0i1 = r0 , and a0j = aj for all j 6= i0 , i1 and we are done. – Case 3: Suppose ai1 = r1 and |ane |r1 > 0, meaning that there are agents at the resource r1 in the equilibrium allocation. Select any agent i2 such that ane i 2 = r1 . (i) If ai2 = ∅, then consider the allocation a0i1 = r0 , a0i2 = r1 and a0j = aj for all j 6= i1 , i2 and we are done. (ii) Otherwise, if ai2 = r2 , then let a0i1 = r0 , a0i2 = r1 , and repeat Case 2 or Case 3 depending on whether |ane |r2 = 0 or |ane |r2 > 0. Note that Case 3-(ii) can be repeated at most n iterations until an alternative case that terminates is reached. To see this, note that each time an agent is given a new choice in this process, i.e., ai → a0i 6= ai , the agent’s new choice is the agent’s equilibrium choice, i.e., a0i = ane i . Therefore, once an agent is assigned a new choice, the agent will never be reassigned in this process. Starting from a as defined above, the above process results in a new allocation a0 that satisfies |U(a0 )| = |U(a)|−1 and a ⊆ a0 . As with the allocation a, the allocation a0 satisfies maxr |a0 |r = 1 and a0 ⊆ aopt . If a0 = aopt , we are done. Otherwise, we can repeat the process depicted above to generate a new allocation a00 such that |U(a00 )| = |U(a0 )| − 1 as nowhere in the process did we rely on the fact that ai = aopt i . Repeating these arguments recursively provides the result.  34
3cs.SY
arXiv:1711.11581v2 [cs.DS] 23 Dec 2017 Outlier-robust moment-estimation via sum-of-squares Pravesh K. Kothari∗ David Steurer† December 27, 2017 Abstract We develop efficient algorithms for estimating low-degree moments of unknown distributions in the presence of adversarial outliers. The guarantees of our algorithms improve in many cases significantly over the best previous ones, obtained in recent works of [DKK+ 16, LRV16, CSV17]. We also show that the guarantees of our algorithms match information-theoretic lower-bounds for the class of distributions we consider. These improved guarantees allow us to give improved algorithms for independent component analysis and learning mixtures of Gaussians in the presence of outliers. Our algorithms are based on a standard sum-of-squares relaxation of the following conceptually-simple optimization problem: Among all distributions whose moments are bounded in the same way as for the unknown distribution, find the one that is closest in statistical distance to the empirical distribution of the adversarially-corrupted sample. ∗ Princeton University and Institute for Advanced Study Zürich. Much of this work was done while at Cornell University and the Institute for Advanced Study. Supported by a Microsoft Research Fellowship, a Alfred P. Sloan Fellowship, NSF awards, and the Simons Collaboration for Algorithms and Geometry. † ETH Contents 1 Introduction 1.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Robust identifiability of low-degree moments 1 5 10 3 Preliminaries 13 3.1 Pseudo-distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.2 Sum-of-squares proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 4 Moment estimation algorithm 16 4.1 Putting things together . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 5 Certifiably subgaussian distributions 22 6 Applications 25 6.1 Outlier-Robust Independent Component Analysis . . . . . . . . . . . . . . . . . . . . 25 6.2 Outlier-Robust Mixtures of Gaussians . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 7 Information Theoretic Lower Bounds 31 References 32 A Sum-of-squares toolkit 36 1 Introduction We consider the problem of outlier-robust parameter estimation: We are given independent draws x1 , . . . , x n from an unknown distribution D over ’d and the goal is to estimate parameters of the distribution, e.g., its mean or its covariance matrix. Furthermore, we require that the estimator is outlier-robust: even if an ε-fraction of the draws are corrupted by an adversary, the estimation error should be small. These kind of estimators have been studied extensively in statistics (under the term robust statistics) [Tuk75, MMY06, Hub11, HRRS11]. However, many robust estimators coming out of this research effort are computationally efficient only for low-dimensional distributions (because the running time is say exponential in the dimension d) [Ber06]. A recent line of research developed the first robust estimators for basic parameter estimation problems (e.g., estimating the mean and covariance matrix of a Gaussian distribution) that are computationally efficient for the high-dimensional case, i.e., the running time is only polynomial in the dimension d [DKK+ 16, LRV16, CJN17, CSV17]. Our work continues this line of research. We design efficient algorithms to estimate lowdegree moments of distributions. Our estimators succeed under significantly weaker assumptions about the unknown distribution D, even for the most basic tasks of estimating the mean and covariance matrix of D. For example, in order to estimate the mean of D (in an appropriate norm) our algorithms do not need to assume that D is Gaussian or has a covariance matrix with small spectral norm in contrast to assumptions of previous works. Similary, our algorithms for estimating covariance matrices work, unlike previous algorithms, for non-Gaussian distributions and distributions that are not (affine transformations of) product distributions. Besides these qualitative differences, our algorithms also offer quantitative improvements. In particular, for the class of distributions we consider, the guarantees of our algorithms—concretely, the asymptotic behavior of the estimation error as a function of the fraction ε of corruptions—match, for the first time in this generality, information-theoretic lower bounds. Outlier-robust method of moments Our techniques for robust estimation of mean vectors and covariance matrices extend in a natural way to higher-order moment tensors. This fact allows us to turn many non-outlier-robust algorithms in a black-box way into outlier-robust algorithms. The reason is that for many parameter estimation problems the best known algorithms in terms of provable guarantees are based on the method of moments, which means that they don’t require direct access to a sample from the distribution but instead only to its low-degree moments (e.g. [Pea94, KMV10, MV10]). (Often, a key ingredient of these algorithms in the high-dimensional setting is tensor decomposition [MR05, AFH+12, AGH+ 14, HK13b, BCMV14, BKS15, GM15, MSS16a].) If there were no outliers, we could run these kinds of algorithms on the empirical moments of the observed sample. However, in the presence of outliers, this approach fails dramatically because even a single outlier can have a huge effect on the empirical moments. Instead, we apply method-ofmoment-based algorithms on the output of our outlier-robust moment estimators. Following this strategy, we obtain new outlier-robust algorithms for independent component analysis (even if the underlying unknown linear transformation is ill-conditioned) and mixtures of spherical Gaussians 1 (even if the number of components of the mixture is large and the means are not separated). Estimation algorithms from identifiability proofs Our algorithms and their analysis follow a recent paradigm for computationally-efficient provable parameter estimation that has been developed in the context of the sum-of-squares method. We say that a parameter estimation problem satisfies identifiability if it is information-theoretically possible to recover the desired parameter from the available data (disregarding computational efficiency). The key idea of this paradigm is that a proof of identifiability can be turned into an efficient estimation algorithm if the proof is captured by a low-complexity proof system like sum-of-squares. Many estimation algorithms based on convex relaxations, in particular sum-of-squares relaxations, can be viewed as following this paradigm (e.g., compressed sensing and matrix completion [CRTV05, CR09, Gro11, Rec11]). Moreover, this paradigm has been used, with a varying degree of explicitness, in order to design a number of recent algorithms based on sum-of-squares for unsupervised learning, inverse, and estimation problems like overcomplete tensor decomposition, sparse dictionary learning, tensor completion, tensor principal component analysis [HSS15, BKS15, BM16, PS17, MSS16a]. In many of these settings, including ours, the proof of identifiability takes a particular form: given a (corrupted) sample X ⊆ ’d from a distribution D and a candidate parameter θ̂ that is close to true parameter θ of D, we want to be able to efficiently certify that the candidate parameter θ̂ is indeed close to θ. (This notion of certificate is similar to the one for NP.) Following the above paradigm, if the certification in this identifiability proof can be accomplished by a low-degree sum-of-squares proof, we can derive an efficient estimation algorithm (that computes an estimate θ̂ just given the sample X). Next we describe the form of our identifiability proofs for outlier-robust estimation. Identifiability in the presence of outliers Suppose we are given an ε-corrupted sample X  {x1 , . . . , x n } ⊆ ’d of an unknown distribution D and our goal is to estimate the mean µ ∈ ’d of D. To hope for the mean to be identifiable, we need to make some assumptions about D. Otherwise, D could for example have probability mass 1 − ε on 0 and probability mass ε on µ/ε. Then, an adversary can erase all information about the mean µ in an ε-corrupted sample from D (because only an ε fraction of the draws from D carry information about µ). Therefore, we will assume that D belongs to some class of distributions C (known to the algorithm). Furthermore, we will assume that the class C is defined by conditions on low-degree moments so that if we take a large enough (non-corrupted) sample X 0 from D, the uniform distribution over X 0 also belongs to the class C with high probability. (We describe the conditions that define C in a paragraph below.) With this setup in place, we can describe our robust identifiability proof: it consists of a (multi-)set of vectors X ′  {x1′ , . . . , x ′n } ⊆ ’d that satisfies two conditions: 1. x ′i  x i for all but an ε fraction of the indices i ∈ [n], 2. the uniform distribution over X ′ is in C. Note that given X and X ′, we can efficiently check the above conditions, assuming that the conditions on the low-degree moments that define C are efficiently checkable (which they will be). 2 Also note that the above notion of proof is complete in the following sense: if X is indeed an ε-corruption of a typical1 sample X 0 from a distribution D ∈ C, then there exists a set X ′ that satisfies the above conditions, namely the uncorrupted sample X 0 . In Section 2, we show that the above notion of proof is also sound: if X is indeed an ε-corruption of a (typical) sample X 0 from a distribution D ∈ C and X ′ satisfies the above conditions, then the Í empirical mean µ′ : n1 ni1 x ′i of the uniform distribution over X ′ is close to µ. We can rephrase this soundness as the following concise mathematical statement, which we prove in Section 2: if D and D ′ are two distributions in C that have small statistical distance, then their means are close to each other (and their higher-order moments are close to each other as well). Furthermore, the above soundness is captured by a low-degree sum-of-squares proof (using for example the sum-of-squares version of Hölder’s inequality); this fact is the basis of our efficient algorithm for outlier-robust mean estimation (see Section 4). Outlier-robustness and (certifiable) subgaussianity To motivate the aforementioned conditions we impose on the distributions, consider the following scenario: we are given an ε-corrupted sample of a distribution D over ’ and our goal is to estimate its variance σ2 up to a constant factors. To hope for the variance to be identifiable, we need to rule out for example that D outputs 0 with probability 1 − ε and a Gaussian N(0, σ2 /ε 2 ) with probability ε. Otherwise, an adversary can remove all information about σ by changing an ε fraction of the draws in a sample p from D. If we look k 1/k at the low-degree moments of this distribution D, we see that ( D(x) x ) ≈ k/ε 1−2/k ( D(x) x 2 )1/2 . So for large enough k (i.e., k ≈ log(1/ε)), the ratio between the L k norm of D and the L 2 norm of √ D exceeds that of a Gaussian by a factor of roughly 1/ ε. In order to rule out this example, we impose the following condition on the low-degree moments of D, and we show that this condition is enough to estimate the variance of a distribution D over ’ up to constant factors given an ε-corrupted sample,  D(x) (x − µ D ) k  1/k 6 √ Ck ·  2 D(x) (x − µ D )  1/2 for C > 0 and even k ∈ Ž with Ck · ε 1−2/k ≪ 1 . (1.1) Here, µ D is the mean of the distribution D. In the high-dimensional setting, a natural idea is to impose the condition Eq. (1.1) for every direction u ∈ ’d ,  1/2    1/k √ 2 k . (1.2) hx − µ D , ui hx − µ D , ui 6 Ck · D(x) D(x) In Section 2, we show that this condition is indeed enough to ensure identifiability and that it is information-theoretically possible to estimate the covariance matrix ΣD of D up to constant factors (in the Löwner order sense) assuming again that Ck · ε 1−2/k ≪ 1. Unfortunately, condition Eq. (1.2) is unlikely to be enough to guarantee an efficient estimation algorithm. The reason is that Eq. (1.2) might hold for the low-degree moments of some distribution D but every proof of this fact requires 1The sample X 0 of D should be typical in the sense that the empirical low-degree moments of the sample X 0 are close to the (population) low-degree moments of the distribution D. If the sample is large enough (polynomial in the dimension), this condition is satisfied with high probability. 3 exponential size. (This phenomenon is related to the fact that finding a vector u ∈ ’d that violates Eq. (1.2) is an NP-hard problem in general.) To overcome this source of intractability, we require that inequality Eq. (1.2) is not only true but also has a low-degree sum-of-squares proof. With this additional condition, we give a polynomialtime algorithm to estimate the covariance matrix ΣD of D up to constant factors (in the Löwner order sense) assuming again that Ck · ε 1−2/k ≪ 1. Definition 1.1 (Certifiable subgaussianity of low-degree moments). A distribution2 D over ’d with mean µ is (k, ℓ)-certifiably subgaussian with parameter C > 0 if there for every positive integer k ′ 6 k/2, there exists a degree-ℓ sum-of-squares proof3 of the degree-2k ′ polynomial inequality over the unit sphere, ∀u ∈ “d−1 . ′ D(x)  hx − µ, ui 2k 6 C · k ′ D(x) hx − µ, ui 2  k′ (1.3) In Section 5, we show that a wide range of distributions are certifiably subgaussian and that many operations on distributions preserve this property. In particular, (affine transformations of) products of scalar-valued subgaussian distributions and mixtures thereof satisfy this property. In a subsequent work, Kothari and Steinhardt [KS17] show that certifiable subgaussianity holds for distributions that satisfy a Poincaré inequality (which includes all strongly log-concave distributions). Since any valid polynomial inequality over the sphere has a sum-of-squares proof if we allow the degree to be large enough and the inequality has positive slack, we say that D is (k, ∞)-certifiably subgaussian with parameter C > 0 if the inequalities Eq. (1.3) hold for all k ′ 6 k/2. In most cases we consider, a sum-of-squares proof of degree ℓ  k is enough. In this case, we just say that D is k-certifiably subgaussian. Observe that k-certifiable subgaussianity only restricts moments up to degree k of the underlying distribution. In this sense, it is a much weaker assumption than the usual notion of subgaussianity, which imposes Gaussian-like upper bounds on all moments. Sum-of-squares and quantifier alternation Taking together the above discussion of robust identifiability proofs and certifiable subgaussianity, our approach to estimate the low-degree moments of a certifiable subgaussian distribution is the following: given an ε-corrupted sample X  {x1 , . . . , x n } ⊆ ’d from D, we want to find a (multi-)set of vectors X ′  {x1′ , . . . , x ′n } ⊆ ’d such that the uniform distributions over X and X ′ are ε-close in statistical distance and the uniform distribution over X ′ is certifiably subgaussian. It is straightforward to formulate these conditions as a system E of polynomial equations over the reals such that x1′ , . . . , x ′n are (some of the) variables (see Section 4). Our outlier-robust 2We emphasize that our notion of certifiable subgaussianity is a property only of the low-degree moments of a distribution and, in this way, significantly less restrictive then the usual notion of subgaussianity (which restricts all moments of a distribution). 3 In the special case of proving that p(u) > 0 holds for every vector u ∈ “ d−1 , a degree-ℓ sum-of-squares proof consists of a polynomial q(u) with deg q 6 ℓ − 2 and polynomials r1 (u), . . . , r t (u) with deg r τ 6 ℓ/2 such that p  q · (kuk 2 − 1) + r12 + · · · + r t2 . See Section 3 for a more general definition. 4 moment estimation proceeds by solving a standard sum-of-squares relaxation of this system E. The solution to this relaxation can be viewed as a pseudo-distribution D ′ over vectors x1′ , . . . , x ′n that satisfies the system of equations E. (This pseudo-distribution behaves in many ways like a classical probability distribution over vectors x ′1, . . . , x ′n that satisfy the equations E. See Section 3 for a definition.) Our moment estimation algorithm simply outputs the expected empirical moments ˜ D′ (x′ ,...,x′ ) 1 Ín (x ′ )⊗r with respect to the pseudo-distribution D ′. i1 i n n 1 We remark that previous work on computationally-efficiently outlier-robust estimation also used convex optimization techniques albeit in different ways. For example, Diakonikolas et al. solve an implicitly-defined convex optimization problem using a customized separation oracle [DKK+ 16]. (Their optimization problem is implicit in the sense that it is defined in terms of the uncorrupted sample which we do not observe.) An unusual feature of the aforementioned system of equations E is that it also includes variables for the sum-of-squares proof of the inequality Eq. (1.2) because we want to restrict the search to those sets X ′ such that the uniform distribution X ′ is certifiably subgaussian. It is interesting to note that in this way we can use sum-of-squares as an approach to solve ∃ ∀-problems as opposed to just the usual ∃-problems. (The current problem is an ∃ ∀-problem in the sense that we want to find X ′ such that for all vectors u the inequality Eq. (1.2) holds for the uniform distribution over X ′.) We remark that the idea of using sum-of-squares to solve problems with quantifier alternation also plays a role in control theory (where the goal is find a dynamical system together with an associated Lyapunov functions, which can be viewed as sum-of-squares proof of the fact that the dynamical system behaves nicely in an appropriate sense). However, to the best of our knowledge, this work is the first that uses this idea for the design of computationally-efficient algorithms with provable guarantees. We remark that in a concurrent and independent work, Hopkins and Li use similar ideas to learn mixtures of well-separated spherical Gaussians [HL17]. In a subsequent paper, Kothari and Steinhardt use those ideas for clustering [KS17]. 1.1 Results Without any assumptions about the underlying distribution, the best known efficient algorithms for robust mean estimation incur an estimation error that depends on the spectral norm of the √ covariance matrix Σ of the underlying distribution and is proportional to ε (where ε > 0 is the fraction of outliers) [DKK+ 17a, SCV17]. Concretely, given an ε-corrupted sample of sufficiently larger polynomial size from a distribution D, they compute an estimate µ̂ for the mean µ of D √ such that with high probability k µ̂ − µk 6 O( ε) · kΣk 1/2 . Furthermore, this bound is optimal for general distributions in the sense that up to constant factors no better bound is possible information-theoretically in terms of ε and the spectral norm of Σ. In the following theorem, we show that better bounds for the mean estimation error are possible for large classes of distributions. Concretely, we assume (certifiable) bounds on higher-order moments of the distribution (degree 4 and higher). These higher-order moment assumptions √ allow us to improve the estimation error as a function of ε (instead of a ε bound as for the unconditional mean estimation before we obtain an ε 1−1/k if we assume a bound on the degree-k 5 moments). Furthermore, we also obtain multiplicative approximations for the covariance matrix (in the Löwner order sense) regardless of the spectral norm of the covariance. (Note that our notion of certifiable subgaussianity does not restrict the covariance matrix in any way.) Theorem 1.2 (Robust mean and covariance estimation under certifiable subgaussianity). For every C > 0 and even k ∈ Ž, there exists a polynomial-time algorithm that given a (corrupted) sample S ⊆ ’d outputs a mean-estimate µ̂ ∈ ’d and a covariance-estimate Σ̂ ∈ ’d×d with the following guarantee: there exists n 0 6 (C + d)O(k) such that if S is an ε-corrupted sample with size |S| > n 0 of a k-certifiably C-subgaussian distribution D over ’d with mean µ ∈ ’d and covariance Σ ∈ ’d×d , then with high probability kµ − µ̂k 6 O(Ck)1/2 · ε 1−1/k · kΣk 1/2 kΣ−1/2 (µ − µ̂)k 6 O(Ck)1/2 · ε 1−1/k (1 − δ)Σ  Σ̂  (1 + δ)Σ for δ 6 O(Ck) · ε (1.4) (1.5) 1−2/k . (1.6) For the last two bounds, we assume in addition Ck · ε 1−2/k 6 Ω(1).4 Note that the second guarantee for the mean estimation error µ− µ̂ is stronger because kµ− µ̂k 6 kΣk 1/2 · kΣ−1/2 (µ − µ̂)k. We remark that kΣ−1/2 (µ − µ̂)k is the Mahalanobis distance between µ̂ and D. In general, our results provide information theoretically tight way to utilize higher moment information and give optimal dependence of the error on the fraction of outliers in the input sample for every (k, ℓ)-certifiably O(1)-subgaussian distribution. This, in particular, improves on pthe mean 1/2 1/2 estimation algorithm of [LRV16] by improving on their error bound of O(ε )kΣk log (d) to O(ε 3/4 kΣk 1/2 under their assumption of bounded 4th moments (whenever “certified” by SoS). Frobenius vs spectral norm for covariance-matrix estimation Previous work for robust covariance estimation [LRV16, DKK+ 17b] work with Frobenius norms for measuring the estimation error Σ − Σ̂ and obtain in this way bounds that can be stronger than ours. However, it turns out that assuming only k-certifiable subgaussianity makes it information-theoretically impossible to obtain dimension-free bounds in Frobenius norm and that we have to work with spectral norms instead. In this sense, the assumptions we make about distributions are substantially weaker compared to previous works. Concretely, [LRV16] show a bound of Σ − Σ̂ F 6 Õ(ε 1/2 )kΣk assuming a 4-th moment bound and that the distribution is an affine transformation of a product distribution. [DKK+ 17b] show a bound Σ − Σ̂ F 6 O(ε)kΣk assuming the distribution is Gaussian.5 However, even for simple k-certifiably subgaussian distributions with parameter√C 6 O(1), the information-theoretically optimal error in terms of Frobenius norm is Σ − Σ̂ F 6 O( d·ε 1−2/k )·kΣk. 4This notation means that we require Ck · ε1−2/k 6 c0 for some absolute constant c0 > 0 (that could in principle be extracted from the proof). 5 In the Gaussian case, [DKK+ 17b] establish the stronger bound Σ−1/2 (Σ − Σ̂)Σ−1/2 F 6 O(ε). This norm can be viewed as the Frobenius norm in a transformed space. Our bounds are for the spectral norm in the same transformed space. 6 Concretely, consider the mixture of N(0, I) and N(0, ε −2/k I) with weights 1 − ε and ε, respectively. Then, it’s easy to confirm that both the mixture and the standard gaussian N(0, I) are k-certifiably 2-subgaussian (using Lemma 5.3 and Lemma 5.7) and at most ε far in total variation distance. Thus, given only k-certifiable 2-subgaussianity, it is information theoretically impossible to decide which of the above two distributions generated a given ε-corrupted sample.√Finally, the difference of their covariance equals ε(ε −2/k − 1)I which has Frobenius norm Ω(ε 1−2/k ) d for any ε < 1. For this case, our algorithms guarantee the significantly stronger6 spectral-norm bound kΣ− Σ̂k 6 O(ε 1−2/k )· kΣk. Multiplicative vs. additive estimation error Another benefit of our covariance estimation algorithm is that provide a multiplicative approximation guarantee, that is, the quadratic form of the estimated covariance at any vector u is within (1 ± δ) of the quadratic form of the true covariance. This strong guarantee comes in handy, for example, in whitening or computing an isotropic transformation of the data—a widely used primitive in algorithm design. Indeed, this ability to use the estimated covariance to whiten the data is crucial in our outlier-robust algorithm for independent component analysis (see Section 6.1). The Frobenius norm error guarantees, in general, do not imply good multiplicative approximations and thus cannot be used for this application. We note that in the special case of gaussian distributions, the results of Diakonikolas et al. [DKK+ 16] allow recovering the mean and covariance in fixed polynomial time with better dependence of the error on the fraction of outliers that grows as Õ(ε)kΣk 1/2 . Our results when applied for the special case of gaussian mean and covariance estimation will require a running √ time of d O( log (1/ε)) to achieve a similar error guarantee. Their algorithm for gaussian covariance estimation also provides multiplicative error guarantees. Robust Estimation of Higher Moments Our techniques for robustly estimating mean and covariance of a distribution extend in a direct way to robustly estimating higher-order moment tensors. In order to measure the estimation error, we use a variant of the injective tensor norm (with respect to a transformed space where the distribution is isotropic), which generalizes the norm we use for the estimation error of the covariance matrix. This error bound allows us to estimate for every direction u ∈ ’d , the low-degree moments of the distribution in direction u with small error compared to the second moment in direction u. The approaches in previous works fact inherent obstacles in generalizing to the problem of estimating the higher moments with multiplicative (i.e. in every direction u) error guarantees. This type of error is in fact crucial in applications for learning latent variable models such as mixtures of Gaussians and independent component analysis. In fact, our guarantees are in some technical way stronger, which is crucial for our applications of higher-order moment estimates. Unlike spectral norms, injective norms are NP-hard to compute (even approximately, under standard complexity assumptions). For this reason, it is not clear how to make use of an injective-norm guarantee when processing moment-estimates further. Fortunately, 6 Similarly to [DKK+ 17b], we also work with norms in a transformed space (in which the distribution is isotropic) and obtain the stronger bound kΣ−1/2 (Σ − Σ̂)Σ−1/2 k 6 O(ε1−2/k ). 7 it turns out that our algorithm not only guarantees an injective-norm bound for the error but also a good certificate for this bound, in form of a low-degree sum-of-squares proof. It turns out that this kind of certificate is precisely what we need for our applications—in particular, recent tensor decomposition algorithms based on sum-of-squares [MSS16a] can tolerate errors with small injective norm if that is certified by a low-degree sum-of-squares proof. Theorem 1.3 (Robust Higher Moment Estimation). For every C > 0 and even k ∈ Ž, there exists a polynomial-time algorithm that given a (corrupted) sample S ⊆ ’d outputs a moment-estimates M̂2 ∈ k 2 ’d , . . . , M̂ k ∈ ’d with the following guarantee: there exists n 0 6 (C + d)O(k) such that if S is an ε-corrupted sample with size |S| > n 0 of a k-certifiably C-subgaussian distribution D over ’d with k 2 moment-tensors M2 ∈ ’d , . . . , M k ∈ ’d such that Ck · ε 1−2/k 6 Ω(1), then with high probability for every r 6 k/2, ∀u ∈ ’d . hM r − M̂ r , u ⊗r i 2 6 δ r · hM2 , u ⊗2i r r for δ r 6 O (Ck)r/2 · ε 1− k (1.7) Furthermore, there exist degree-k sum-of-squares proofs of the above polynomial inequalities in u. Information-theoretic optimality We show in Section 7 that the error guarantees in our robust moment-estimation algorithms are tight in their dependence on both k and ε. For example, we show that there are√two k-certifiably O(1)-subgaussian distributions with statistical distance ε but means that are Ω( kε 1−1/k ) apart. A similar statement holds for higher-order moments. The distributions are just mixtures of two one-dimensional Gaussians. Application: independent component analysis As an immediate application of our robust moment estimation algorithm, we get an algorithm for Outlier Robust Independent Component Analysis. Independent component analysis (also known as blind source separation) is a fundamental problem in signal processing, machine learning and theoretical computer science with applications to diverse areas including neuroscience. Lathauwer et. al. [DLCC07], following up on a long line of work gave algorithms for ICA based on 4th order tensor decomposition. A noise-tolerant version of this algorithm was developed in [MSS16a]. There is also a line of work in theoretical computer science on designing efficient algorithms for ICA [GVX13, VX15]. In the ICA problem, we are given a non-singular7 mixing matrix A ∈ ’d×d with condition number κ and a product distribution on ’d . The observations come from the model {Ax} that is, the observed samples are linear transformations (using the mixing matrix) of independent draws of the product random variable x. The goal is to recover columns of A up to small relative error in Euclidean norm (up to signs and permutations) from samples. It turns out that information theoretic recovery of A is possible whenever at most one source is non-gaussian. A widely used convention in this regard is the 4th moment assumption: for each i, [x 4i ] , 3 [x 2i ]. It turns out that we can assume [x 2i ]  1 without loss of generality so this condition reduces to asserting [x 4i ] , 3. 7We could also consider the case that A is rectangular and its columns are linearly independent. Essentially the same algorithm and analysis would go through in this case. We focus on the quadratic case for notational simplicity. 8 Outlier robust version of ICA was considered as an application of the outlier-robust mean and covariance estimation problems in [LRV16]. Theypsketched an algorithm with the guarantee that the relative error in the columns of A is at most ε log (d) poly(κ). 1 . Here, In particular, this guarantee is meaningful only if the fraction of outliers ε ≪ √ log (d) poly(κ) we improve upon their result by giving an outlier-robust ICA algorithm that recovers columns of A up to an error that is independent of both dimension d and the condition number κ of the mixing matrix A. Our algorithm directly follows by applying 4th order tensor decomposition. However, a crucial step in the algorithm involves “whitening” the 4th moments by using an estimate of the covariance matrix. Here, the multiplicative guarantees obtained in estimating the covariance matrix are crucial - estimates with respect to the Frobenius norm error do not give such whitening transformation in general. This whitening step essentially allows us to pretend that the mixing matrix A is well-conditioned leading to no dependence on the condition number in the error. Theorem 1.4 (Robust independent component analysis). For every C > 1 and even k ∈ Ž, there exists a polynomial-time algorithm that given a (corrupted) sample S ⊆ ’d outputs component estimates â1 , . . . , â d ∈ ’d with the following guarantees: Suppose A ∈ ’d×d is a non-singular matrix with condition number κ and columns a1 , . . . , a d ∈ ’d . Suppose x is a centered random vector with d independent √ coordinates such that every coordinate i ∈ [d] satisfies [x2i ]  1, [x4i ] − 3  γ , 0, and [xki ]1/k 6 Ck. Then, if S is an ε-corrupted sample of size |S| > n 0 from the distribution {Ax}, where n 0 6 (C + κ + d)O(k) , the component estimates satisfy with high probability max min hA−1 â i , A−1 a π(i) i 2 > 1 − δ for δ < (1 + π∈S d i∈[d] 1 |γ| ) · O(C 2 k 2 ) · ε 1−4/k . (1.8) The quantity hA−1 â i , A−1 a π(i) i is closely related to the Mahalanobis distance between â i and a π(i) with respect to the distribution {Ax} Application: learning mixtures of Gaussians As yet another immediate application of our robust moment estimation algorithm, we get an outlier-robust algorithm for learning mixtures of spherical Gaussians. Our algorithm works under the assumption that the means are linearly independent (and that the size of the sample grows with their condition number). In return, our algorithm does not require the means of the Gaussians to be well-separated. Our algorithm can be viewed as an outlier-robust version of tensor-decomposition based algorithms for mixtures of Gaussians [HK13a, BCMV14]. Theorem 1.5 (Robust estimation of mixtures of spherical Gaussians (See Theorem 6.9 for detailed error bounds)). Let D be mixtures of N(µ i , I) for i 6 q with uniform8 mixture weights. Assume that µ i s Í is Ω(1). are linearly independent and, further, assume that κ, the smallest non-zero eigenvalue of q1 i µ i µ⊤ i Given an ε-corrupted sample of size n > n 0  Ω((d log (d))k/2 /ε 2 ), for every k > 4, there’s a poly(n)d O(k) time algorithm that recovers µ̂1 , µ̂2 , . . . µ̂ q so that there’s a permutation π : [q] → [q] 8While our algorithm generalizes naturally to arbitrary mixture weights, we restrict to this situation for simplicity 9 satisfying max k( i 1Õ −1/2 µ i µ⊤ (µ̂ i − µ π(i) )k 6 O(qk)ε 1/3−1/k . i ) q i [DKK+ 16] gave an outlier-robust algorithm that learns mixtures of q Diakonikolas et. al. √ gaussians with error ≈ q ε in each of the recovered means. Their algorithm is polynomial in the dimension but has an exponential dependence on number of components q in the running time. Under the additional assumption that the means are linearly independent, our algorithm (say for k  8) recovers similar error guarantees as theirs but runs in time polynomial in both q and d. The key difference is the power of our algorithm to recover a multiplicative approximation to the 4th moment tensor which allows us to apply blackbox tensor decomposition based methods and run in fixed polynomial time [HK13b]. 2 Robust identifiability of low-degree moments In this section, we show that low-degree moments of certifiably subgaussian distributions are identifiable in the presence of outliers. As we will formalize in Section 4, these proofs of identifiability are captured by the sum-of-squares proof system at low-degree. This fact is the basis of our polynomial-time algorithms for robust moment estimation (also Section 4). We recall the basic strategy for the identifiability proofs as described in the introduction: Let D0 be a certifiably subgaussian distribution whose moments we aim to estimate. If we take a large enough sample X ⊆ ’d from D0 , then with high probability, the empirical distribution D (uniform over the sample) has approximately the same low-degree moments as the (population) D0 . In this case, the low-degree moments of D are also certifiably subgaussian9 (for essentially the same parameters C as D0 ). Our input consists of an ε-corruption Y ⊆ ’d of the sample X. In order to show that the low-degree moments of D0 (or equivalently D) are identifiable from Y, we analyze the following (a-priori inefficient) estimation procedure: 1. find a certifiably subgaussian distribution D ′ that has statistical distance at most ε from the uniform distribution over Y, 2. output the low-degree moments of D ′. We can typically find a distribution D ′ as above, because with high-probability D satisfies the conditions. What remains to prove is that no matter what distribution D ′ satisfying those conditions we choose, our moment estimates have low error. To this end, we show several concrete and quantiative instances of the following general mathematical statement: If two distributions D and D ′, whose low-degree moments are (certifiably) subgaussian, have small statistical distance, then their low-degree moments are close. Our first bound of this kind, controls the distance between the means of D and D ′ in terms of their covariances Σ and Σ′. (Later bounds will also relate Σ and Σ′ .) 9 Roughly speaking, certifiable subgaussianity is inherited by the empirical distribution from the population distribution because it is a property of the low-degree moments of distributions. We prove this claim in Section 5. 10 Lemma 2.1 (Robust identifiability of mean). Let k ∈ Ž be even. Let D, D ′ be two distributions over ’d with means µ, µ′ ∈ ’d and covariances Σ, Σ′ ∈ ’d×d . Suppose kD − D ′ k TV 6 ε < 0.9 and that D and D ′ are (k, ℓ)-certifiably subgaussian with parameter C > 0. Then, ∀u ∈ ’d . |hu, µ − µ′i| 6 δ · hu, (Σ + Σ′)ui 1/2 where δ  O √  Ck · ε 1−1/k . (2.1) In particular, kµ − µ′ k 6 δkΣ + Σ′ k 1/2 and k(Σ + Σ′ )−1/2 (µ − µ′)k 6 δ. (Here, ’d is endowed with the standard Euclidean norm and ’d×d with the corresponding operator norm.) Proof. Consider a coupling between the distributions D(x) and D ′(x ′ ) such that {x  x ′ } > 1 − ε. By Hölder’s inequality, we have for every vector u ∈ ’d ,  |hu, µ − µ′i|  | hu, x − x ′i 1x ,x′ | 6 At the same time,  ′ k hu, x − x i  1/k  6   ′ ′ hu, x − x ′i k ′ hu, x − µ − (x − µ ) + (µ − µ )i hu, x − µi k  1/k +  k  1/k hu, x ′ − µ′i k  1/k   1/k k/(k−1)  1−1/k 1x ,x′ |¨¨¨¨¨¨¨¨¨¨¨¨¨¨{z¨¨¨¨¨¨¨¨¨¨¨¨¨¨} (2.2) 6 ε 1−1/k (2.3) + |hu, µ − µ′i| (2.4) (using triangle inequality for L k -norm)  √  1/2  1/2  6 Ck · + |hu, µ − µ′i| hu, x − µi 2 + hu, x ′ − µ′i 2 (using certifiable subgaussianity) √ 6 O( Ck ) · hu, (Σ + Σ′)ui 1/2 + |hu, µ − µ′i| . (2.5) (2.6) Combining the two bounds yields   √ |hu, µ − µ′i| 6 ε 1−1/k · O( Ck ) · hu, (Σ + Σ′)ui 1/2 + |hu, µ − µ′i| . Since we that ε < 0.9, we can conclude that |hu, µ − µ′i| 6 δ · hu, (Σ + Σ′ )ui 1/2 for  assume √   δ 6 O ε 1−1/k · Ck as desired. Next we show that if two certifiably subgaussian distributions have small statistical distance, then their (raw) second moments are close in the Löwner order sense. Lemma 2.2 (Robust identifiability of second moment). Let D, D ′ be two distributions over ’d with second-moment matrices M, M ′ ∈ ’d×d . Suppose kD − D ′ k TV 6 ε < 0.9 and that D and D ′ are (k, ℓ)-certifiably subgaussian with parameters C > 0. Then,  − δ · (M + M ′)  M − M ′  δ · (M + M ′) for δ 6 O Ck · ε 1−2/k   In particular, 1 − O(δ) M ′  M  1 + O(δ) M ′ if δ < 0.9. 11  (2.7) Proof. Consider a coupling between the distributions D(x) and D ′(x ′ ) such that {x  x ′ } > 1 − ε. Then, for every vector u ∈ ’d , |hu, (M − M ′)ui|  | hu, xi 2 − hu, x ′i 2 | 6 ε 1−2/k · 6ε 1−2/k ·   hu, xi 2 − hu, x ′i 2 hu, xi k 6 ε 1−2/k · O(Ck) ·  2/k +   k/2  2/k ′ k hu, x i hu, xi 2 + (using Hölder inequality)  2/k  hu, x ′i 2  O(Ck · ε 1−2/k ) · hu, (M + M ′)ui ,  (by L k/2 triangle inequ.) (by subgaussianity)  which implies the desired bound −δ · (M + M ′)  M − M ′  δ · (M + M ′) for δ 6 O Ck · ε 1−1/k . For the third inequality, we use Lemma 5.2 (moment bounds for shifts of subgaussian distributions). For δ 6 1, we can now rearrange this bound to get the following relationship between M and ′ M, ′ ′ 1+δ 1−δ (2.8) 1+δ · M  M  1−δ · M ,   which implies the desired bound 1 − O(δ) M ′  M  1 + O(δ) M ′ for δ < 0.9.  Next, we combine the previous bounds for the first and second moments in order to establish the identifiability of the covariance matrix. Corollary 2.3 (Covariance identifiability). Under the same conditions as the previous Lemma 2.2, the covariance matrices Σ and Σ′ of the distributions D and D ′ satisfy, − δ · (Σ + Σ′)  Σ − Σ′  δ · (Σ + Σ′) for δ 6 O(Ck · ε 1−2/k ) . (2.9) As before, in the case that δ 6 0.9, the above bounds imply a strong multiplicative approximation of the covariance so that (1 − δ′) · Σ′  Σ  (1 + δ′)Σ′ for some δ′ 6 O(δ). Proof of Corollary 2.3. Let Dµ , Dµ′ be the distributions D, D ′ shifted by µ so that Dµ has expectation 0 and Dµ′ has expectation µ′ − µ. Note that Dµ − Dµ′ TV  kD − D ′ k TV 6 ε. By Lemma 2.2 (second-moment identifiability applied to Dµ , Dµ′ ), every vector u ∈ ’d satisfies, |hu, (Σ − Σ′ − (µ − µ′)(µ − µ′)T )ui| 6 δ 1 · hu, (Σ + Σ′ + (µ − µ′)(µ − µ′)T )ui . (2.10) Here, δ1 6 O(Ck · ε 1−1/k ). By Lemma 2.1 (mean identifiability), hu, (µ − µ′)(µ − µ′)T ui 6 δ 2 · hu, (Σ + Σ′)ui, where δ 2 6 O(Ck · ε 1−1/k ). Combining the inequalities gives the desired bound.  Taking together Corollary 2.3 and Lemma 2.1, we get a mean estimation error that depends only on the covariance of D (as opposed to the covariances of both D and D ′). Corollary 2.4 (Stronger mean identifiability). Under the same conditions as the previous Lemma 2.2 and the additional condition Ck · ε 1−2k 6 Ω(1), the means µ and µ′ of the distributions D and D ′ satisfy,  √ (2.11) ∀u ∈ ’d . |hu, µ − µ′i| 6 δ · hu, Σui 1/2 where δ  O Ck · ε 1−1/k , 12 where Σ ∈ ’d×d is the covariance of D. In particular, kµ − µ′ k 6 δkΣk 1/2 and kΣ−1/2 (µ − µ′)k 6 δ. (Here, ’d is endowed with the standard Euclidean norm and ’d×d with the corresponding operator norm.) Next we bound the distances of higher order moments for certifiably subgaussian distributions that are close in statistical distance. Compared to first and second order moments, distances for higher order moments are less standard. Our choice for the distance is informed by the applications of independent component analysis and learning mixtures of Gaussians. We view the order-2r moments of D and D ′ as two polynomials p and p ′ of degree 2r and we bound their difference p(u) − p ′(u) relative to the variance of the distribution in direction u. Lemma 2.5 (Robust identifiability of higher moments). Let r ∈ Ž and let D, D ′ be two distributions 2 r over ’d with second moments M2 , M2′ ∈ ’d and r-th moments M r , M r′ ∈ ’d . Suppose kD − D ′ k TV 6 ε < 0.9 and that D and D ′ are (k, ℓ)-certifiably subgaussian with parameters C > 0 and 2r 6 k. Then, for every u ∈ ’d , r |hM r − M r′ , u ⊗r i| 6 δ · hM2 + M2′ , u ⊗2 i r/2 for δ 6 O(Ck)r/2 · ε 1− k (2.12) Proof. Consider a coupling between the distributions D(x) and D ′(x ′ ) such that {x  x ′ } > 1 − ε. Then, for every vector u ∈ ’d , |hu, (M r − M r′ )ui|  | hu, xi r − hu, x ′i r | 6ε 1−r/k · 6 ε 1−r/k ·   r ′ r k/r (hu, xi − hu, x i ) hu, xi k 6 ε 1−r/k · O(C r k r ) ·  r/k  +   r/k hu, x ′i k hu, xi 2  r/2 + (using Hölder’s inequality)  r/k  (by L k/r triangle inequ.) hu, x ′i 2 r  r/2  (by subgaussianity) which implies the desired bound for δ 6 δ 6 O(Ck)r/2 · ε 1− k . For the third inequality, we rely on Lemma 5.2 (moment bounds for shifts of subgaussian distributions).  3 Preliminaries In this section, we define pseudo-distributions and sum-of-squares proofs. See the lecture notes [BS16] for more details and the appendix in [MSS16b] for proofs of the propositions appearing here. Let x  (x1 , x2 , . . . , x n ) be a tuple of n indeterminates and let ’[x] be the set of polynomials with real coefficients and indeterminates x1 , . . . , x n . We say that a polynomial p ∈ ’[x] is a sum-of-squares (sos) if there are polynomials q 1 , . . . , q r such that p  q 12 + · · · + q 2r . 3.1 Pseudo-distributions Pseudo-distributions are generalizations of probability distributions. We can represent a discrete (i.e., finitely supported) probability distribution over ’n by its probability mass function D : ’n → 13 ’ such that D > 0 and Í D(x)  1. Similarly, we can describe a pseudo-distribution by its mass function. Here, we relax the constraint D > 0 and only require that D passes certain low-degree non-negativity tests. Concretely, a level-ℓ pseudo-distribution is a finitely-supported function D : ’n → ’ such that Í Í 2 x D(x)  1 and x D(x) f (x) > 0 for every polynomial f of degree at most ℓ/2. (Here, the summations are over the support of D.) A straightforward polynomial-interpolation argument shows that every level-∞-pseudo distribution satisfies D > 0 and is thus an actual probability distribution. We define the pseudo-expectation of a function f on ’d with respect to a pseudodistribution D, denoted ˜ D(x) f (x), as x∈supp(D) ˜ D(x) f (x)  Õ x D(x) f (x) . (3.1) The degree-ℓ moment tensor of a pseudo-distribution D is the tensor D(x) (1, x1 , x2 , . . . , x n )⊗ℓ . In particular, the moment tensor has an entry corresponding to the pseudo-expectation of all monomials of degree at most ℓ in x. The set of all degree-ℓ moment tensors of probability distribution is a convex set. Similarly, the set of all degree-ℓ moment tensors of degree d pseudodistributions is also convex. Key to the algorithmic utility of pseudo-distributions is the fact that while there can be no efficient separation oracle for the convex set of all degree-ℓ moment tensors of an actual probability distribution, there’s a separation oracle running in time n O(ℓ) for the convex set of the degree-ℓ moment tensors of all level-ℓ pseudodistributions. Fact 3.1 ([Sho87, Par00, Nes00, Las01]). For any n, ℓ ∈ Ž, the following set has a n O(ℓ) -time weak separation oracle (in the sense of [GLS81]):  ˜ D(x) (1, x1 , x2 , . . . , x n )⊗d | degree-d pseudo-distribution D over ’n (3.2) . This fact, together with the equivalence of weak separation and optimization [GLS81] allows us to efficiently optimize over pseudo-distributions (approximately)—this algorithm is referred to as the sum-of-squares algorithm. The level-ℓ sum-of-squares algorithm optimizes over the space of all level-ℓ pseudo-distributions that satisfy a given set of polynomial constraints—we formally define this next. Definition 3.2 (Constrained pseudo-distributions). Let D be a level-ℓ pseudo-distribution over ’n . Let A  { f1 > 0, f2 > 0, . . . , f m > 0} be a system of m polynomial inequality constraints. We say that D satisfies the system of constraints A at degree r, denoted D r A, if for every S ⊆ [m] and every Í sum-of-squares polynomial h with deg h + i∈S max{deg f i , r}, ˜Dh· We write D Ö fi > 0 . i∈S A (without specifying the degree) if D 0 A holds. Furthermore, we say that D −n ℓ Î r A · kh k · i∈S k f i k, holds approximately if the above inequalities are satisfied up to an error of 2 where k·k denotes the Euclidean norm10 of the cofficients of a polynomial in the monomial basis. ℓ 10The choice of norm is not important here because the factor 2−n swamps the effects of choosing another norm. 14 We remark that if D is an actual (discrete) probability distribution, then we have D A if and only if D is supported on solutions to the constraints A. We say that a system A of polynomial constraints is explicitly bounded if it contains a constraint of the form {kx k 2 6 M}. The following fact is a consequence of Fact 3.1 and [GLS81], Fact 3.3 (Efficient Optimization over Pseudo-distributions). There exists an (n + m)O(ℓ) -time algorithm that, given any explicitly bounded and satisfiable system11 A of m polynomial constraints in n variables, outputs a level-ℓ pseudo-distribution that satisfies A approximately. 3.2 Sum-of-squares proofs Let f1 , f2 , . . . , f r and 1 be multivariate polynomials in x. A sum-of-squares proof that the constraints { f1 > 0, . . . , f m > 0} imply the constraint {1 > 0} consists of polynomials (p S )S⊆[m] such that Õ 1 S⊆[m] p S · Πi∈S f i . (3.3) We say that this proof has degree ℓ if for every set S ⊆ [m], the polynomial p S Πi∈S f i has degree at most ℓ. If there is a degree ℓ SoS proof that { f i > 0 | i 6 r} implies {1 > 0}, we write: { f i > 0 | i 6 r} ℓ {1 > 0} . (3.4) Sum-of-squares proofs satisfy the following inference rules. For all polynomials f , 1 : ’n → ’ and for all functions F : ’n → ’m , G : ’n → ’k , H : ’p → ’n such that each of the coordinates of the outputs are polynomials of the inputs, we have: { f > 0, 1 > 0} A , A ℓ { f + 1 > 0} A A ℓ ℓ B, B A ℓ·ℓ′ ℓ′ ℓ { f > 0}, A A ℓ+ℓ′ ℓ′ {1 > 0} { f · 1 > 0} C (transitivity) C {F > 0} {F(H) > 0} ℓ (addition and multiplication) {G > 0} ℓ·deg(H) {G(H) > 0} (substitution) . Low-degree sum-of-squares proofs are sound and complete if we take low-level pseudodistributions as models. Concretely, sum-of-squares proofs allow us to deduce properties of pseudo-distributions that satisfy some constraints. Fact 3.4 (Soundness). If D proof A r′ B, then D r·r ′+r ′ r B. A for a level-ℓ pseudo-distribution D and there exists a sum-of-squares 11Here, we assume that the bitcomplexity of the constraints in A is (n + m)O(1) . 15 If the pseudo-distribution D satisfies A only approximately, soundness continues to hold if we require an upper bound on the bit-complexity of the sum-of-squares A r ′ B (number of bits required to write down the proof). In our applications, the bit complexity of all sum of squares proofs will be n O(ℓ) (assuming that all numbers in the input have bit complexity n O(1) ). This bound suffices in order to argue about pseudo-distributions that satisfy polynomial constraints approximately. The following fact shows that every property of low-level pseudo-distributions can be derived by low-degree sum-of-squares proofs. Fact 3.5 (Completeness). Suppose d > r ′ > r and A is a collection of polynomial constraints with degree Í at most r, and A ⊢ { ni1 x 2i 6 B} for some finite B. Let {1 > 0} be a polynomial constraint. If every degree-d pseudo-distribution that satisfies D satisfies D r ′ {1 > 0}, then for every ε > 0, there is a sum-of-squares proof A d {1 > −ε}. r A also 4 Moment estimation algorithm Our robust moment estimation algorithm is a direct translation of the robust identifiability results from Section Section 2. Let Y  { y1 , . . . , yn } ⊆ ’d be an ε-corrupted sample of a distribution D over ’d . We consider the following system A : AY,ε of quadratic equations in scalar-valued variables w 1 , . . . , w n and vector-valued variables x1′ , . . . , x ′n , AY,ε :      Ín i1 w i  (1 − ε) · n    w 2i  w i    ∀i ∈ [n]. w i · (yi − x ′)  0  i ∀i ∈ [n].       (4.1) The following fact shows that the solutions to AY,ε correspond to all ε-corruptions of the (already corrupted) sample Y. In particular, if Y  { y1 , . . . , yn } is an ε-corruption of a sample X  {x1 , . . . , x n } of D, then one solution to A consists of the uncorrupted sample X so that x1′  x1 , . . . , x ′n  x n . Fact 4.1. An assignment to the variables x1′ , . . . , x ′n can be extended to a solution for AY,ε if and only if x ′i  yi holds for all but at most an ε-fraction of the indices i ∈ [n]. Note that in order to extend such an assignment, we can choose w i  1 for (1 − ε)n indices i ∈ [n] such that x i  yi and w i  0 for the remaining indices. Furthermore, we consider the following system of polynomial equations B : BC,k in vectorvalued variables x1′ , . . . , x ′n , vector-valued variables p1 , . . . , p k/2 , and matrix-valued variables Q 1 , . . . , Q k/2 . The idea is that p1 , . . . , p k/2 and Q 1 , . . . , Q k/2 are degree-ℓ sum-of-squares proofs of the inequalities Eq. (1.3) (C-subgaussianity of moments up to degree k) for the uniform distribution 16 over x1′ , . . . , x ′n . BC,k,ℓ :     ∀k ′ ∈ Ž6 k/2 .      1 n n Õ i1 hx ′i , ui 2k ′  ′  Ck · 1 n n Õ i1 hx ′i , ui 2  k′   + hp k′ , (1, u)⊗ℓ i · (1 − ku k 2)    − kQ k′ · (1, u)⊗ℓ/2 k 2     (4.2) We emphasize that in the above system of equations the vector-valued variable u  (u 1 , . . . , u d ) is not part of the solution (i.e., we are not seeking an assignment for it). Instead, each equation in BC,k,ℓ is about two polynomials with indeterminates u 1 , . . . , u d , whose coefficients are polynomials in the variables x1′ , . . . , x ′n , p1 , . . . , p k/2 , and Q 1 , . . . , Q k/2 . In particular, we could eliminate the variable u from the equations in BC,k,ℓ by explicitly writing down the expressions for the coefficients of the polynomials in u. We refrain from doing so because these expressions would be rather complicated and would obfuscate the meaning of the equations. The following fact shows that the solutions to the system BC,k correspond to certifiably subgaussian distributions. Fact 4.2. An assignment to the vector-valued variables x1′ , . . . , x ′n can be extended to a solution for BC,k,ℓ if and only if the uniform distribution over x1′ , . . . , x ′n is (k, ℓ)-certifiably subgaussian with parameter C. The following algorithm is the key ingredient of our efficient outlier-robust estimators. Algorithm 4.3 (Algorithm for moment estimation via sum-of-squares). Given: ε-corrupted sample Y  { y1 , . . . , yn } ⊆ ’d of a (k, ℓ)-certifiably 0.9C-subgaussian distribution D0 over ’d Estimate: moments M1 , . . . , M k with M r  D0 (x) x ⊗r Operation: 1. find a level-ℓ pseudo-distribution D ′ that satisfies AY,ε and BC,k,ℓ  Í  2. output moment-estimates M̂1 , . . . , M̂ k with M̂ r  ˜ D′(x′ ) 1 n (x ′ )⊗r n i1 i The following lemma is the sum-of-squares analog of Lemma 2.1. Lemma 4.4. Let D be the uniform distribution over X ⊆ ’d with mean µ ∈ ’d and covariance Σ ∈ ’d×d . Suppose D is (k, ℓ)-certifiably subgaussian with parameter C > 0 and that Ck · ε 1−2/k 6 Ω(1).12 Let Y Í be an ε-corruption of X with ε < 0.01. Let µ′ , Σ′ be the polynomials µ′(X ′)  n1 ni1 x ′i and Σ′(X ′)  Í n 1 ′ ′ T ′ d ′ ′ T ′ i1 (x i )(x i ) − (µ )(µ ) in the vector-valued variables x 1 , . . . , x n . Then, for every vector u ∈ ’ , n where δ  O √ AY,ε ∪ BC,k,ℓ  X′ ℓ  hu, µ − µ′(X ′)i k 6 δ k hu, (Σ + Σ′(X ′))ui k/2 , (4.3) Ck · ε 1−1/k . Furthermore, the bit complexity of the sum-of-squares proof is polynomial. 12 This notation means that we require Ckε1−2/k 6 γ for some absolute constant γ > 0, e.g., γ  10−10 . 17 Í Proof. Choose r1 , . . . , rn ∈ {0, 1} such that i ri  (1 − ε)n and ri (x i − yi )  0 for all i ∈ [n]. (Such a choice exists because Y is an ε-corruption of X.) Letting z i  ri · w i , we claim that w AY,ε 2 ( n Õ i1 z i > (1 − 2ε)n ) . Í wi Indeed, since {w 2i  w i } 2 {(1 − z i ) 6 (1 − w i ) + (1 − ri )}, we have A Y,ε 2 { ni1 (1 − z i ) 6 2εn} By the sum-of-squares version of Hölder’s inequality, we have for every vector u ∈ ’d , using that A Y,ε k {(1 − z i )k−1  (1 − z i )}. n w,x′ ℓ AY,ε ∪ BC,k,ℓ hu, µ − µ′i k  6  Õ 1 n  Õ i 1 n  Õ i 6 1 n i  For brevity, we will use the fact that a,b,c k AY,ε ∪ BC,k,ℓ i∈[n] c)k (a + b + w,x′ ℓ n  i∈[n]  hu, x i − x ′i i k k  Õ  1 n i (2ε) (4.4) (1 − z i )k k−1 o  k−1 . to denote the average over indices i ∈ [n]. At the same time, using 6 3k (a k + b k + c k ) (similar to Lemma A.2),  hu, x i − µ − (x ′i − µ′) + (µ − µ′)i k  Õ 1 n i k hu, x i − x ′i i k hu, x i − x ′i i k i∈[n] 6 3k hu, x i − x ′i i(1 − z i ) 6 3 Ck  hu, x i − µi k + 3k i∈[n] hu, x i − µi  2 k/2  Õ 1 n i   hu, x ′i − µ′i k + 3k hu, µ − µ′i k k + 3 Ck i∈[n] hu, x ′i − µ′i 2 (using certifiable subgaussianity) 6 O(Ck) k/2 ′ · hu, (Σ + Σ )ui Combining the two bounds yields k/2 k ′ k + 3 hu, µ − µ i . o  k/2  + 3k hu, µ − µ′i k (4.5)  hu, µ − µ′i k 6 (2ε)k−1 · O(Ck)k/2 · hu, (Σ + Σ′ )ui k/2 + 3k hu, µ − µ′i k . Since we assume that ε < 0.01, we can conclude that AY,ε ∪ BC,k,ℓ w,x′ ℓ  √  for δ 1 6 O ε 1−1/k · Ck as desired. hu, µ − µ′i k 6 δ 1k · hu, (Σ + Σ′)ui k/2 , (4.6)  18 The following lemma is the sum-of-squares analog of Lemma 2.2. Lemma 4.5 (Robust identifiability of second moment). Let D be the uniform distribution over X ⊆ ’d with second moment M ∈ ’d×d . Suppose D is (k, ℓ)-certifiably subgaussian with parameter C > 0 with k divisible by 4 and that Ck · ε 1−2/k 6 Ω(1). Let Y be an ε-corruption of X. Let M ′ be the quadratic Í polynomial n1 i x ′i x ′iT in vector-valued variables x1′ , x2′ , . . . , x ′n . Then, for every u ∈ ’d , A Y,ε ∪ BC,k,ℓ X′ ℓ   hu, (M − M ′)ui k/2 6 δ k/2 hu, M + M ′)ui k/2 Furthermore, if δ 6 0.0001, AY,ε ∪ BC,k,ℓ X′ ℓ n 1 − δ′ 6 1 hu,Mui hu, M ′ui 6 1 + δ′ for δ 6 O Ck · ε 1−2/k o  (4.7) for δ′  100δ . Furthermore, the bit complexity of the sum-of-squares proof is polynomial. Proof. By an argument analogous to the one in the proof of Lemma 4.4, we can obtain: AY,ε ∪ BC,k,ℓ  w,x′ ℓ  hu, (M − M ′)ui k/2 6 δ k/2 · hu, (M + M ′))ui k/2 , (4.8) for δ 6 O ε 1−2/k · Ck . Lemma A.4 together with (4.8) implies for some δ′ 6 100δ, A Y,ε ∪ BC,k,ℓ X′ ℓ n 1 − δ′ 6 1 hu,Mui hu, o M ′(X ′)ui 6 1 + δ′ . The following lemma is the sum-of-squares analog of Corollary 2.3.  Corollary 4.6 (Covariance estimation). Let D be the uniform distribution over X ⊆ ’d with covariance Σ ∈ ’d×d . Suppose D is (k, ℓ)-certifiably subgaussian with parameter C > 0 with k divisible by 4 and Í that Ck · ε 1−2/k 6 Ω(1). Let Y be an ε-corruption of X. Let Σ′  n1 ni1 (x ′i − µ′)(x ′i − µ′)T be the formal Í covariance of the uniform distribution over x1′ , . . . , x ′n , where µ′  n1 i x ′i . Then, for every u ∈ ’d , AY,ε ∪ BC,k,ℓ X′ ℓ 1−δ 6 ′ ′ 1 hu,Σui hu, Σ (X )ui 6 1 + δ for some δ 6 O(Ck) · ε 1−2/k . (4.9) Furthermore, the bit complexity of the sum-of-squares proof is polynomial. Proof. Let Dµ be the distributions D shifted by µ so that Dµ has expectation 0. Then, Dµ is (k, ℓ)w,X ′  w i (z i − (x i − µ)) 2 ′ ,u Í Í X ′ ′ using Lemma 5.2, for every k ′ 6 k/2, BC,k,ℓ ℓ n1 i hx ′i , ui 2k 6 (2Ck ′)k /2 ( n1 i hx ′i , Thus, applying Lemma 4.5, we obtain for every vector u ∈ ’d , certifiably 2C-subgaussian. Further, for z i  x ′i − µ i AY,ε AY,ε ∪ BC,k,ℓ ui 2)k/2 . w,X ′ ℓ hu, (Σ − Σ′ − (µ − µ′)(µ − µ′)T )ui k/2 6   0 . Further, (4.10) k/2 δ2 · hu, (Σ + Σ′ + (µ − µ′)(µ − µ′)T )ui k/2  6 (2δ2 )k/2 hu, (Σ + Σ′)ui k/2 + hu, (µ − µ′)(µ − µ′)T ui k/2 . 19 (4.11) (4.12) Here, δ 2 6 O(Ck · ε 1−2/k ). Further, X ′ ,w k  hu, (Σ − Σ′)ui k/2 6 2k/2 hu, (Σ − Σ′ − (µ − µ′)(µ − µ′)T )ui k/2 + 2k/2 hu, (µ − µ′)(µ − µ′)T ui k/2 . By Lemma 2.1 (mean estimation), hu, (µ − µ′)(µ − µ′)T ui k/2 6 δ 21 · hu, (Σ + Σ′ )ui k/2 , where δ1 6 O(Ck · ε 1−1/k ). Combining the above inequalities, we obtain: w,X ′ ℓ AY,ε ∪ BC,k,ℓ hu, (Σ − Σ′)ui k/2 6 O(δ 2 )k/2 hu, (Σ + Σ′)ui k/2 . (4.13) Further, w,X ′ k hu, (Σ + Σ′ )ui k/2  (hu, 2Σui − hu, (Σ − Σ′ )ui)k/2 6 2k/2 hu, 2Σui + 2k/2 hu, (Σ − Σ′)ui k/2 . Using this in conjunction with (4.13), we obtain: AY,ε ∪ BC,k,ℓ w,X ′ ℓ Now, using Lemma A.3 with f  hu, (Σ − Σ′)ui k/2 6 O(δ 2 )k/2 hu, Σui k/2. hu,(Σ−Σ′ )ui O(δ2 ) k/2 hu,Σui completes the proof.  The following is the analog of the Corollary 2.4. Corollary 4.7 (Stronger Mean Estimation). Let D be the uniform distribution over X ⊆ ’d with mean, covariance µ ∈ ’d and Σ ∈ ’d×d . Suppose D is (k, ℓ)-certifiably subgaussian with parameter C > 0 with k divisible by 4 and that Ck · ε 1−2/k 6 Ω(1). Let Y be an ε-corruption of X. Let µ′(X ′) be the linear Í polynomial n1 i x ′i in the vector-valued variables x1′ , x2′ , . . . , x ′n . ∀u ∈ ’d .AY,ε ∪ BC,k,ℓ w,X ′ ℓ n ± hu, µ − µ′i 6 δ · hu, Σui 1/2 w,X ′ where δ  O √ Ck · ε 1−1/k o . (4.14) In particular, AY,ε ∪ BC,k,ℓ ℓ kµ − µ′ k 6 δkΣk 1/2 and kΣ−1/2 (µ − µ′)k 6 δ. Furthermore, the bit complexity of the sum-of-squares proof is polynomial. √ Proof. Combining Lemma 4.4 and Corollary 4.6 we have for δ 1  O( Ckε 1−1/k ) and δ 2  O(Ck)ε 1−2/k , w,X ′ ℓ ′ k AY,ε ∪ BC,k,ℓ hu, µ − µ′(X )i 6 δ 1k hu, (Σ + Σ′)ui k/2 2 6 (2δ1 )  hu, Σui k/2 6 (6δ1 )2 hu, Σui k/2. ′ + hu, Σ ui k/2  Observe that hu, Σui k/2 is a constant, i.e., does not depend on any variable in the SoS proof. hu,µ−µ′ (X ′ )i  Using Lemma A.3 with f  (6δ )2 hu,Σui1/2 now completes the proof. 1 Finally, the following lemma is the sum-of-squares analog of Lemma 2.5. The proof is similar to the ones presented before, so we omit it here. 20 Lemma 4.8. Let D be the uniform distribution over X ⊆ ’d with second moments M2 ∈ ’d×d . Suppose D is (k, ℓ)-certifiably subgaussian with parameter C > 0 for even k divisible by r. Let Y be ε-corruption sample Í be a r-order tensor with entries that are r-degree polynomial in vector-valued of X. Let M r′  n1 ni1 x ′⊗r i ′ ′ variables x1 , . . . , x n . Then, the following inequality has a sum of squares proof of degree ℓ in variables of AY,ε ∪ BC,k,ℓ and u ∈ ’d , u,w,X ′ ℓ    −δhu, M2 + M2′ ui k/2 6 hM r − M r′ , u ⊗r i k/r 6 δhu, M2 + M2′ ui k/2 , (4.15) where δ  O (C r k r ) · ε 1−r/k . Consequently, AY,ε ∪ BC,k,ℓ  AY,ε ∪ BC,k,ℓ ∪ ku k22 6 1 u,w,X ′ ℓ  −δhu, M2 ui r/2 6 hM r − M r′ , u ⊗r i 6 δhu, M2 ui r/2 . (4.16) Furthermore, the bit complexity of the sum-of-squares proof is polynomial. Remark 4.9. We point out a subtle technical difference with respect to the two results above. In Lemma 4.4 and Corollary 4.6, we do not force u to be a variable in the sum of squares proof system. On the other hand, in Lemma 4.8, the resulting polynomial inequality has a sum of squares proof in u (in addition to the other variables). This is important because, the inequalities are degree 2 (in u) for Corollary 4.7 and Corollary 4.6 while higher degree in general. While unconstrained (in u) inequalities of degree 6 2 always have a SoS proof, this is not true for higher degrees. The lemma above relies on Proposition A.5 to obtain this stronger conclusion. 4.1 Putting things together We can now complete the proof of Theorem 1.2. Proof of Theorem 1.2. We just put together the implications of the above results to analyze Algorithm 4.3 here. Let S′ be the true sample used to produce the ε-corrupted sample S. Then, US′ - the uniform distribution on S′ is (k, ℓ)-certifiably 2C-subgaussian by an application of Lemma 5.5. Let ζ be the pseudo-distribution of degree at least ℓ over variables w, y satisfying AY,ε ∪ BC,k,ℓ Í Í output by Algorithm 4.3. Define µ′  n1 ni1 yi and Σ′  n1 ni1 (yi − µ̂)(yi − µ̂)⊤ be the polynomials defined in Corollary 4.7 and Corollary 4.6 respectively. We round the pseudo-distribution simply by setting µ̂  ˜ ζ µ′. From Corollary 4.7, we have that:  AY,ε ∪ BC,k,ℓ ⊢ℓ hu, µ − µ′i 2 6 δ 2 hu, Σui , Thus, for every u ∈ ’d , ˜ ζ hu, µ − µ′i 2 6 δ 2 hu, Σui By Cauchy-Schwarz inequality for pseudo-distributions, we have: 2 6 hu, µ − ˜ ζ µ′i 2 6 δ2 hu, Σui. 2 Using that ˜ µ′  µ̂ thus yields hu, µ − µ̂i 6 δ 2 hu, Σui. ˜ ζ hu, µ − µ′i 21 (4.17) Using u → Σ−1/2 u thus immediately yields:  −1/2 hu, Σ (µ − µ̂)i 2 6 δ2 ku k 2. Taking square roots thus implies hu, Σ−1/2 (µ − µ̂)i 6 δku k or equivalently, kΣ−1/2 (µ − µ̂)k 6 δ. On the other hand, taking square roots in (4.17) gives: |hu, µ − µ̂i|  kµ − µ̂k 6 δkΣk 1/2 as required. Next, we analyze the covariance estimation procedure from Algorithm 4.3. As before, our rounding is simple: Σ̂  ˜ ζ Σ′ . Applying Corollary 4.6 and using that degree ℓ pseudo-expectations must respect degree ℓ SoS proofs (Fact 3.4), we get that for every u ∈ ’d , whenever δ < 1, (1 − δ)Σ  Σ̂  (1 + δ)Σ as required.  The proof of Theorem 1.3 is entirely analogous. That the inequality in the conclusion has a sum-of-squares proof in u follows from Proposition A.5. 5 Certifiably subgaussian distributions In this section, we give sufficient conditions for distributions to be certifably subgaussian. As concrete consequences, we derive that (affine transformations of) products of scalar-valued subgaussian distributions are certifiably subgaussian and that certain mixture models are certifiably subgaussian. For the benefit of the reader, we restate here the definition of certifiable subgaussianity. Definition (Certifiable subgaussianity). A distribution D over ’d with mean µ is (k, ℓ)-certifiably subgaussian with parameter C > 0 if there for every positive integer k ′ 6 k/2, there exists a degree-ℓ sum-of-squares proof of the polynomial inequality ∀u ∈ “ d−1 . D(x) hx − µ, ui 2k ′  6 C·k ′ D(x) hx − µ, ui 2  k′ (5.1) We first note some basic invariance properties of certifiable subgaussianity. Lemma 5.1 (Invariance Under Linear Transformations). Suppose x ∈ ’d is a random variable with a (2k, ℓ)-certifiably C-subgaussian distribution. Then, for every matrix A ∈ ’d×d , the random variable Ax is (2k, ℓ)-certifiably C-subgaussian. Proof. Fix any v ∈ ’d . Invoking certifiable subgaussianity of x, we get that there’s a degree ℓ-SoS proof of the inequality: ′ [hAx, vi 2k ]  ′ ′ [hx, A⊤ vi 2k ] 6 (Ck ′ [hx, A⊤ vi 2 ])k  (Ck ′ [hAx, vi 2])k ′  22 Lemma 5.2 (Shifts of certifiable subgaussian distributions). Let k ∈ Ž be even and let D be a mean zero, (k, ℓ)-certifiably C-subgaussian distribution over ’d . Then, D satisfies for every number k ′ 6 k/2 and vector s ∈ ’d , ′ ′ (5.2) hx + s, ui 2)k ∀u ∈ “d−1 . hx + s, ui 2k 6 (Ck ′ D(x) D(x) Furthermore, the above inequality Eq. (5.2) has a degree-ℓ sum-of-squares proof. Next, we construct examples of various classes of distributions that are certifiably subgaussian. We start by noting that the standard gaussian distribution is k-certifiably O(1)-subgaussian for every k. Lemma 5.3 (Certifiable Subgaussianity of Standard Gaussian). Let x be the standard gaussian random variable with mean 0 and covariance Id on ’d . Then, x is (2k, 2k)certifiably 1-subgaussian for every k. Proof. By rotational symmetry of the standard gaussian and the fact that [x 2k ]  Γk , [hx, ui 2k ]  i Γk [hx, ui 2]k for some Γk < (2k)k .  Next, we prove a similar fact about uniform mixtures of arbitrary mean gaussians. Lemma 5.4. Let µ1 , µ2 , . . . µ q be arbitrary. Let D be the mixture of N(µ i , I) with mixture weights each component. Then, D is (k, k)-certifiably C-subgaussian with C  O(q). 1 q for Proof. We have using Lemma 5.2 and certifiable C-subgaussianity of the gaussian distribution for C  O(1), hx, ui 2k  1Õ hx, ui 2k q x∼N(µ i ,I)  k 1Õ 2 hx, ui 6 (2Ck) q x∼N(µ i ,I) i k i 1 1Õ  (2Ck) (hu, µ i i 2 + 1)k 6 q k−1 (Ck)k (hu, µ i i 2 + 1) q q k hx, ui 2  On the other hand, subgaussian. 1 q Í i i (hu, !k . µ i i 2 + 1). Thus, we obtain that D is (k, k)-certifiably (2Cq)  Lemma 5.5 (Sampling Preserves Certifiable Subgaussianity). Let D be a (k, ℓ)-certifiably subgaussian distribution on ’d . Let S be an i.i.d. sample from D of size n > n 0  Ω((d log (d/δ))k/2 /ε 2 ). Then, with probability at least 1 − δ over the draw of S according to D, the uniform distribution on S, US , is (k, ℓ) certifiably (C + ε)-subgaussian. We will use the following matrix concentration inequality in the proof. Fact 5.6 (Matrix Rosenthal Inequality). Fix p > 1.5. For a finite sequence Pk for k > 1 of independent, 2p random psd matrices that satisfy kPk k2p < ∞, k Õ k 2p Pk k2p ! 1/2p © Õ 6 ­k « k 1/2 p Pk k2p + 4p − 2 23 Õ k 2p kPk k2p ! 1/4p 2 ª ® . ¬ Proof. Without loss of generality, assume that M2  D xx ⊤ is full-rank. If not, we just work in the range-space of D xx ⊤ instead, as every sample x would live in this subspace. −1/2 Now, consider the affine transformation x → M2 x which makes second moment is now I. By Lemma 5.1, it is enough to show certifiable subgaussianity of the distribution after this transformation. Thus, without loss of generality, we assume that the D is isotropic, that is, M2  I. Now, by sum-of-squares version of the Cauchy-Schwarz inequality u 2k ′ ±(Ck ′) k ′ /2 ( k′ US hx, ui 2) − US hx, ui * 2k ′  ± (Ck ′) k ′ /2 ′ k ′ /2 ′ where k(Ck ′)k /2 US x ⊗2  ⊗k′ 6 k(Ck ) − US ′   x ⊗2 US US x ⊗2  ⊗k′  ⊗k′ − − US x ⊗2k ′ ,u ′ US ⊗2k ′ + x ⊗2k k∞ ku k22k ′ ′ (5.3) ′ x ⊗2k k∞ is the spectral norm of the canonical d k × d k flattening  ′ ⊗k ′ ′ − US x ⊗2k . Let T(D) be the corresponding of the 2k ′ -order tensor T(US )  (Ck ′)k /2 US x ⊗2 tensor for D. By standard matrix concentration, whenever n 0 ≫ (d log d/ε 2 ), US hx, ui 2 > 1− ε. By standard argument combining Fact 5.6 with Matrix-Chebyshev inequality, we can obtain that with probability at least 1 − δ over the draw of the sample S of size n > n 0  Ω((d log (d/δ))k/2 /ε 2 ), for any k ′ 6 k/2, kT(D) − T(US )k∞ 6 ε. (5.4) Thus, along with (5.3), we obtain that: u 2k ′ ′ k ′ /2 (Ck ) 2 k′ ( US hx, ui ) − Rearranging thus yields that: u 2k ′  ′ US  hx, ui (Ck ′)k /2 + ε) ( US 2k ′ 6 ′ εku k22k US hx, ui 2  k′ . ′ ′ hx, ui 2)k − < 2ε  US hx, ui 2k > 0 completing the proof.  Lemma 5.7 (Certifiable Subgaussianity of Mixtures). Let D1 and D2 be two (k, ℓ)-certifiably Csubgaussian distributions with covariances Σ1 , Σ2 respectively. Suppose further that ∆Σ1  Σ2  Σ1 . Let D be a mixture of D1 and D2 with mixture weights 1 − λ and λ respectively such that λ < ∆−k/2 . Then, D is (k, ℓ)-certifiably 2C-subgaussian. Proof. Assume that D1 and D2 are mean zero. We have for any u, ⊢2 D1 hx, ui 2 6 D hx, ui 2 6 (1 − λ + λ∆) D1 hx, ui 2. 2k ′ 6 (1 − λ) Ck On the other hand, for any k ′ 6 k/2, ⊢2k′ D hx, ui λ D2 Ck ′ D2 hx,  k′ ui 2 6 (1 − λ + ′ λ∆k ) Ck ′ Thus, D is (k, ℓ)-certifiably 2C-subgaussian. D1 hx, 24  k′ ui 2 . D1 hx, ui 2  k′ +  The proof of this lemma follows immediately from the following proposition. Proposition 5.8 (Invariance Under Additive Shifts). Let k ∈ Ž be even and let X be a mean zero, scalar random variable satisfying ( X k )1/k 6 (C X 2 )1/2 . Then, every s ∈ R satisfies ( (X + s)k )1/k 6 2(C (X + s)2 )1/2 .  k 1/k  2 1/2 Proof. By the triangle inequality for L k norms, we have (X + s)k  1/k 6 Xk  1/k + |s |. Therefore, 6 C X + |s |. On the other hand (X + s)2  X 2 + s 2 using the fact that (X + s) X  0. Thus, 2( (X + s)2 )1/2 > ( X 2 )1/2 + |s |. Combining the last two inequalties yields the lemma.  Lemma 5.9 (Certifiable Subgaussianity of Product Distributions). Let D  D1 ⊗ D2 ⊗ · · · ⊗ Dd be a product distribution on ’d such that each Di has ean zero, variance 1 and certifiably subgaussian with constant C i . Then, D is (2k, 2k)-certifiably subgaussian with constant C  max i C i . Proof. Let 1 6 C  maxi C i . Observe that in that case, Di x 2 j 6 C 2 j · N(0,1) x 2 j for every j 6 2k. Let D ′ be the distribution obtained by the following process: 1) sample x ∼ D and 2) output x and −x with uniform probability. By linearity, D hx, ui 2k  D ′ hx, ui 2k . Call a multi-set S of [d] even if every individual element in S appears even number of times. Let S be an even multi-set of size |S|  2k. Then, D ′ x S  D x S 6 C |S| · N(0,I) x S using the sub-gaussianity of each x i . Now, let u ∈ ’d . Then, D hx, ui 2  D ′ hx, ui 2  ku k22. We have: D hx, ui 2k  ′ Õ S: even multi-set [x S ] · u S . Since S is even, we can upper bound the term corresponding to S above by C |S| N(0,I) x S u S . Thus, Í 2k 6 C |S| S S 2k 2k D ′ hx, ui N(0,I) x u . Thus, the RHS is at most σ · x∼N(0,I) hx, ui . S: even multi-set By rotational symmetry, x∼N(0,I) hx, ui 2k  Γk · ( x∼N(0,I) hx, ui 2)k  Γk · ( D ′ hx, ui 2)k , where Γk 6 (2k)k . Thus, D ′ hx, ui 2k 6 (2k)k C 2k ( D ′ hx, ui 2)k . And thus, D hx, ui 2k 6 (2k)k C 2k ( D hx, ui 2)k . Showing that D is (2k, 2k) certifiably C-subgaussian.  6 Applications 6.1 Outlier-Robust Independent Component Analysis In this section, we apply our robust moment estimation algorithm to give an efficient algorithm for outlier-robust independent component analysis. We recall the definition of subgaussian random variables first. Definition 6.1 (Subgaussian Distributions). A distribution D on ’ is said to be K-subgaussian if √ for every p > 1, x∼D [x p ]1/p 6 K( z∼N(0,1) [z p ])1/p 6 K · p. A distribution D on ’n is said to be K-subgaussian if for every unit vector v ∈ ’n , hx, vi is K-subgaussian random variable on ’. 25 Next, we describe the ICA problem: Definition 6.2 (Fully Determined Independent Component Analysis). Let A be an unknown mixing matrix in ’d×d of full rank. Let x be an unknown product distribution with C-subgaussian marginals satisfying [x 2i ]  1 and [x 4i ] − 1 , 0 for every i. Given samples from y  {Ax} for i.i.d. samples x, estimate a matrix  such that there’s a permutation π and some signing σ ∈ {±1} d on columns of  satisfying kA i − σi ·  π(i) k2 6 δkA i k for the smallest possible δ. Remark 6.3. 1. The assumption that [x 2i ]  1 is without the loss of generality. Since we only see samples from {Ax} one can’t uniquely recover the scalings on the column A i and [x 2i ] simultaneously for any i. 2. The goal is to recover a matrix A whose columns are close to that of A up to permutation and signs. This is again, necessary, since the order of the columns cannot be uniquely recovered from samples of {Ax}. 3. Information theoretically, columns of A can be recovered (up to permutation and scaling) if at most one x i is has gaussian distribution. Our main result is a outlier-robust algorithm for ICA. Our algorithm is based on 4th order tensor decomposition. We will use noise-tolerant generalization of the FOOBI algorithm due to Cardoso [DLCC07] presented in [MSS16a]. Theorem 6.4 (Outlier-Robust ICA). For every C > 1 and even k ∈ Ž, there exists a (dn)O(k) that given a (corrupted) sample S ⊆ ’d of size n outputs component estimates â1 , . . . , â d ∈ ’d with the following guarantees: Suppose A ∈ ’d×d is a non-singular matrix with condition number κ and columns a1 , . . . , a d ∈ ’d . Suppose x is a centered random vector with d independent coordinates such that every √ 2 4 k 1/k coordinate i ∈ [d] satisfies [xi ]  1, [xi ]−3  γ , 0, and [xi ] 6 Ck. Then, if S is an ε-corrupted sample of size |S| > n 0 from the distribution {Ax}, where n 0 6 (C + κ + d)O(k) , the component estimates satisfy with high probability max min hA−1 â i , A−1 a π(i) i 2 > 1 − δ π∈S d i∈[d] for δ < (1 + 1 |γ| ) · O(C 2 k 2 ) · ε 1−4/k . (6.1) Our algorithm is simple. It first estimates the 2nd and 4th moments of the observed distribution from an ε-corrupted sample using Algorithm 4.3 and then applies a blackbox tensor decomposition algorithm to a “whitened” version of 4th moments. Tensor Decomposition We state the tensor decomposition algorithm (implicit in [MSS16a], analog of Theorem 1.1 for 4th order tensor decomposition) we use before proceeding. It is convenient to define the following relaxation of the injective tensor norm in describing the guarantees of this algorithm. Definition 6.5 (Sum of Squares Norms). Given a tensor T ∈ (’d )⊗2t for any 2t, the degree 2t sum of squares norm of T, kT ksos2t is defined by kT ksos2t  2t  sup µ(u): deg 2t pseudo-dist over sphere 26  ˜ µ(u) hT, u ⊗2t i .  ⊗4 Fact 6.6 ([MSS16a]). There exists a polynomial time algorithm that given a symmetric 4-tensor T ∈ ’d , outputs a set of vectors {c 1′ , c 2′ , . . . , c ′n } ⊆ ’d , such that for every orthonormal set {c 1 , c 2 , . . . , c n } ∈ ’d , there’s a permutation π : [n] → [n] satisfying Õ T − c i⊗4 ksos4 . kT ksos4 n max kc i − c ′π(i) k 2 6 O(1) · k i i1 2 j−1 Without loss of generality, we can assume that the distribution x satisfies [x i ]  0 for every j > 0. This can be accomplished by negating each sample independently with probability 1/2. The fraction of corruptions in the modified sample is still ε. Algorithm 6.7 (Algorithm for Outlier-Robust ICA via Sum of Squares). Given: ε-corrupted sample from {Ax}: Y  { y1 , . . . , yn } ⊆ ’d . Estimate: Mixing Matrix A ∈ ’d×d . Operation: 1. Apply Algorithm 4.3 to estimate the covariance Σ̂ and 4th moments M̂4 of sample Y. 2. Define the “whitened” estimated 4th moment tensor M̂4′ so that for every u ∈ ’d , h M̂4′ , u ⊗4 i  h M̂4 , (Σ−1/2 u)⊗4 i. Let σ1 , σ2 , . . . , σd be the columns of Σ̂−1/2 . Then, the above is equivalent to setting: Í M̂4′  j1 , j2 ,..., j4 σ j1 ⊗ σ j2 ⊗ σ j3 ⊗ σ j4 · (M̂4 )(j1 , j2 , j3 , j4 ). 3. Let T̂  M̂4′ − 3I ⊗ I. 4. Apply Tensor Decomposition from Theorem 6.6 to T̂ to recover components â1 , â2 , . . . , â d . 5. Output Σ̂1/2  where   (â1 , . . . , â d ). We now analyze the algorithm. The following lemma shows that the components of the tensor T that we decompose are close to the columns of A. Lemma 6.8. Let T  M̂4′ − 3I ⊗ I for the whitenened 4th moments M̂4′ . Let a1 , a2 , . . . , a d be the columns of A and let b i  (AA⊤ )−1/2 a i be the whitened versions. Then, k T̂ − γ Õ i 1/4 b i⊗4 ksos4 6 δ 4 (2 + 3δ 2 )1/4 . Proof. First, let’s understand why the tensor decomposition should produce the columns of A if all the moment estimates were exactly correct. We will then account for the estimation errors. For the distribution {Ax}, we can compute M2  Σ  [(Ax)(Ax)⊤ ]  AA⊤ using [xx ⊤ ]  I. Í Í Thus, for b i  Σ−1/2 a i , i b i b ⊤  i Σ−1/2 a i a ⊤ Σ−1/2  I. i i 27 Similarly, we have hM4 , u ⊗4 i  [hAx, ui 4]  Õ i [x 4i ]ha i , ui 4 + 3 Õ Õ 4  (γ − 3)ha i , ui + 3( i, j i [x 2i x 2j ]ha i , ui 2 ha j , ui 2 ha i , ui 2)2 . The whitened and shifted 4th moment T then, by definition, satisfies hT, u ⊗4i  hM4′ − 3I ⊗ I, u ⊗4 i  hM4 , (Σ−1/2 u)⊗4 i − 3ku k24 Õ  γhb i , ui 4 + 3( i Í hb i , ui 2)2 − 3ku k24  γhb i , ui 4 . Í As a result, ⊢u,4 hT − γ i b i⊗4 , u ⊗4 i  0 and thus, kT − γ i b i⊗4 ksos4  0. Now, let Σ̂, M̂4 , M̂4′ , T̂, b̂ i stand for their estimated (via Algorithm 4.3) counterparts. To complete the analysis, we now understand the error kT − T̂ ksos4 . The proof is then complete by using: k T̂ − γ Õ i b i⊗4 ksos4 6 k T̂ − T ksos4 + kT − γ Õ i b i⊗4 ksos4  k T̂ − T ksos4 . First, we check that our algorithm (Algorithm 4.3) can indeed be applied to this setting. Since x is a product distribution with ith marginal C i -subgaussian. By Lemma 5.9, x is (k, k)-certifiably C-subgaussian. By Lemma 5.1, {Ax} is (k, k)-certifiably subgaussian. 13 Finally, applying Lemma 5.5 shows that with probability at least 1 − 1/n 2 , the sample the facts above all hold for sample Y of size m > Ω̃(d k/2 /ε 2 ). We will use Σ, M4 , M4′ , T, b i  Σ−1/2 a i for the covariance, 4th moment, whitenened 4th moment, shifted, whitenened 4th moment and orthogonalized columns of A respectively. By Corollary 4.6, for every u, hu, Σ̂ui  (1 ± δ 2 )hu, Σui for some δ2 < O(Ck)ε 1−2/k . Thus: (1 − δ 2 )I  By Theorem 1.3, for δ 4  O(C 2 k 2 )ε 1−4/k , u,k 4 Thus, u,k 4  Õ i  b̂ i b̂ ⊤ i  (1 + δ 2 )I. (6.2)  ± hM4 − M̂4 , u ⊗4 i 6 δ 4 · hΣ, u ⊗2 i 2 + hΣ̂, u ⊗2 i 2 . −1/2 ± hM4 − M̂4 , (M̂2    −1/2 u)⊗2 i 2 + h M̂2 , (M̂2 6 δ4 (2 + 3δ2 )ku k24 Or, u,k 4 −1/2 u)⊗4 i 6 δ4 · hM2 , (M̂2 (6.3) u)⊗2 i 2  ± hM4′ − M̂4′ , u ⊗4 i 6 δ 4 (2 + 3δ 2 )ku k24 13Negating samples with probability 1/2 preserves 2nd and 4th moments and doesn’t affect certifiable subgaussianity. 28 Since M4′ − M̂4′  T − T̂, this gives: u,k 4  ± hT − T ′ , u ⊗4 i 6 δ 4 (2 + 3δ 2 )ku k24 (6.4)  The proof of Theorem 1.4 then follows easily. Proof of Theorem 1.4. We are now ready to apply Fact 6.6 to obtain that decomposing the tensor T recovers components c 1 , c 2 , . . . , c d such that there’s a permutation π : [d] → [d] such that 1 kc i − b π(i) k22 6 O(1) · |γ| · δ 4 (2 + 3δ 2 ). Applying the reverse whitening transform â i  Σ̂1/2 c i then recovers the set of vectors  a1 , a2 , . . . , a d satisfying the requirements of Theorem 1.4. 6.2 Outlier-Robust Mixtures of Gaussians In this section, we describe our outlier-robust algorithm for learning mixtures of spherical gaussians. As before, this algorithm can be seen as a direct analog of the non-robust algorithm that learns mixtures of spherical gaussians with linearly independent means by decomposing the 3rd order moment tensor of the observed sample. Theorem 6.9 (Outlier-Robust Mixtures of Gaussians). Let D be mixtures of N(µ i , I) for i 6 q with uniform14 mixture weights. Assume that µ i s are linearly independent and, further, let κ be the smallest Í non-zero eigenvalue of 1q i µ i µ⊤ . i Given an ε-corrupted sample of size n > n 0  Ω((κd log (d))k/2 ), for every k > 4, there’s a poly(n)d O(k) time algorithm that recovers µ̂1 , µ̂2 , . . . µ̂ q so that there’s a permutation π : [q] → [q] satisfying max k( i p p √ 1Õ µ i µ⊤ )−1/2 (µ̂ i − µ π(i) )k 6 O( qk)ε 1−1/k / κ + O(qk)ε 3/2−2/k /κ + O( qk)ε 1/3−1/k . i q i We will rely on 3rd order tensor decomposition algorithm (Theorem 1.1) from [MSS16b] here.  ⊗3 Fact 6.10 ([MSS16a]). There exists a polynomial time algorithm that given a symmetric 3-tensor T ∈ ’d , outputs a set of vectors {c 1′ , c 2′ , . . . , c ′n } ⊆ ’d , such that for every orthonormal set {c 1 , c 2 , . . . , c n } ∈ ’d , there’s a permutation π : [n] → [n] satisfying Õ T − c i⊗3 ksos4 . 6 O(1) · k kT ksos4 n max kc i − i c ′π(i) k 2 i1 14While our algorithm generalizes naturally to arbitrary mixture weights, we restrict to this situation for simplicity 29 Algorithm 6.11 (Algorithm for Outlier-Robust Mixture of Spherical Gaussians via Sum of Squares). Given: ε-corrupted sample X from Estimate: µ1 , µ2 , . . . , µ q . Í 1 i q N(µ i , I) for µ1 , µ2 , . . . , µ q . Operation: 1. Apply Algorithm 4.3 to estimate the first three moments, M̂1 , M̂2 , and M̂3 of sample X0. 2. Define the “whitened” version of the estimated 3rd moment tensor T̂ so that for every u ∈ ’d , hT̂, u ⊗3 i  h M̂3 − 3M̂1 ⊗ I, ((M̂2 − I)−1/2 u)⊗3 i. Concretely, set M̂1′  (M̂2 −I)−1/2 M̂1 . Let σ1 , σ2 , . . . , σd be the columns of (M̂2 −I)−1/2 . Í ′ Set T̂  j1 , j2 , j3 σ j1 ⊗ σ j2 ⊗ σ j3 · (M̂3 − 3M̂1 ⊗ I)(j1 , j2 , j3 ). 3. Apply Tensor Decomposition from Theorem 6.6 to T̂ to recover components ν̂1 , ν̂2 , . . . , ν̂q . 4. Output µ̂ i  (M̂2 − I)1/2 ν̂i for i 6 q as the estimated means. The following is the main technical lemma required in the proof of 6.9. Lemma 6.12. Let Y be an ε-corrupted sample of size n > n 0  Ω((κd log d)k /ε 2 ) from D0  Í 1 i q N(µ i , I). Let M̂ i be the estimated ith moment of D0 given by Algorithm 4.3 when run on Y. Let M̂ i′ be the whitening of the estimated ith moment of D0 as in Algorithm 6.11. Then, with high probability over the draw of the ε-corrupted sample Y, for T̂  M̂3′ − 3M̂1′ ⊗ I and νi  (M2 − I)−1/2 µ i for i 6 q and δ i  O(qk)i/2 ε 1−i/k for i  1, 2, 3, k T̂ − √ 1 Õ ⊗3 1/3 1/2 νi ksos4 6 δ1 / κ + δ1 δ 2 /κ + 2δ 3 . q i Proof. As before, we will first analyze the tensor T assuming all estimates of the moments are exactly correct and then account for the estimation errors. We will write M1 , M2 , M3 for the true first three moments of the input mixture of gaussian and Σ  M2 − I. We will let M̂1 , M̂2 , M̂3 , Σ̂ be the corresponding quantities estimated by our algorithm. We will write M1′ , M3′ for the true whitened moments and M̂1′ , M̂3′ , the estimated counterparts. Í By direct computtion, hM3 , u ⊗3 i  1q i (hµ i , ui 3 + 3hµ i , ui). Thus, hM3 − 3M1 ⊗ I, u ⊗3 i  1 q Í i hµ i , ui 3. Now, by definition, hM3′ − 3M1′ ⊗ I, u ⊗3i  hM3 − 3M1 ⊗ I, (Σ−1/2 u)⊗3 i. Í As a result, for T  M3′ − 3M1′ ⊗ I, we have: kT − 1q i (Σ−1/2 µ i )⊗3 ksos4  0. We now account for the estimation errors and bound kT − T̂ ksos4 . First, we use Lemma 5.4 and Lemma 5.5 to conclude that with high probability, the uniform distribution on the input sample is k-certifiably O(q)-subgaussian. 30 Now (via triangle inequality), kT − T̂ ksos4 6 kM3′ − M̂3′ ksos4 + 3kM1′ ⊗ I − M̂1′ ⊗ I ksos4 . (6.5) |hM1′ − M̂1′ , ui| 6 |hΣ−1/2 (M1 − M̂1 ), ui| + |h(Σ−1/2 − Σ̂−1/2 )M̂1 , ui| (6.6) Now, p From Theorem 1.2, for δ 1  O( qk)ε 1−1/k , δ 2  O(qk)ε 1−2/k , hM1 − M̂1 , Σ−1/2 ui 6 δ1 hu, M2Σ−1 ui 1/2 . Í Í Using that M2  I + 1q i µ i µ⊤ , M2 Σ−1  ( 1q i µ i µ⊤ )−1 + I. Thus, the first term in (6.6) can be i i √ upper bounded by δ 1 ku k2 + δ 1 hu, Σ−1 , ui 1/2 6 δ 1 (1 + 1/ κ)ku k2 . From Theorem 1.2, for any u, hu, (M2 − M̂2 )ui 6 δ 2 hu, M2 ui. Writing M2  Σ + I for Σ  1q µ i µ⊤ , i we obtain that hu, Σ̂ui  (1 ± O(δ2 /κ))hu, Σui. 1/2 √ Thus, h(Σ−1/2 − Σ̂−1/2 )M̂1 , ui 6 O(δ2 / κ) · hΣ−1/2 M̂1 , ui. Writing hΣ−1/2 M̂1 , ui  hΣ−1/2 (M̂1 − Í M1 )ui + hΣ−1/2 M1 ui and using that 1q i hΣ−1/2 µ i , ui 6 √1q ku k2, we finally obtain anupper bound   √ √ √ 1/2 √ 1/2 on (6.6) of: δ1 (1 + 1/ κ) + O(δ2 / κ)δ 1 1 + 1/ κ + √1q ku k2 6 O(δ 2 δ1 /κ + δ1 / κ)ku k2 . For the first term of (6.5), we apply Theorem 1.3 to get that for δ 3  O(qk)3/2 ε 1−3/k , u 6 u ′ h M̂3 − M3 , u ⊗3 i 2 6 δ 23 hM2 , u ⊗2 i 3 6 2δ3 hΣ, u ⊗2i 3 ku k23 . As a result, 6 h M̂3 − M3′ , u ⊗3 i 2 6 2δ23 ku k23 giving us an upper bound on the first term in the RHS of (6.5). This completes the proof.  We can now complete the proof of Theorem 6.9. Proof of Theorem 6.9. Applying Fact 6.10 to T̂ as in Lemma 6.12 yields that the resulting components √ 1/3 1/2 satisfy min π:[q]→[q] max i kνi − ν̂π(i) k2 6 O(1)δ 1 / κ+ δ1 δ 2 /κ+2δ 3 . The theorem now follows.  7 Information Theoretic Lower Bounds In this section, we give simple examples that show that our robust moment estimation results achieve sharp information theoretic limits for (certifiably) subgaussian distributions. In both results below, for any choice of k, we will show that there are two distributions that 1) both satisfy (k, k)-certifiable C-subgaussianity and 2) are at most ε apart in total variation distance but have lowdegree moments that are exactly as far apart as the estimation error in our algorithm. In fact, these distributions are really simple and in fact just one dimensional and thus the certifiability follows from generic resuls about sum of squares representations for univariate non-negative polynomials. We begin by showing a lower bound on the error incurred in robust estimation of mean. Lemma 7.1. There exist two distributions D1 and D2 on ’ such that: 1. D1 , D2 are both 2-subgaussian in all moments up to k. 31 2. dTV (D1 , D2 ) 6 ε. 3. k D1 x− D2 √ xk > kε 1−1/k . Proof. Let D1 be the standard gaussian distribution N(0, 1). Let D2 be the distribution that with √ −1/k . Observe probability 1 − ε outputs a sample from N(0, 1) and with probability ε, outputs kε that by construction, dTV (D1 , D2 ) 6 ε.√ The mean of D2 is easily seen to be kε 1−1/k . The variance of D2 can be computed to be between 1 and (1 + kε 1−2/k ). It is standard to check that D1 is C-subgaussian in all moments for C  1 < 2. For D2 , we √ k observe that D2 (x − D2 x)2k 6 D2 x 2k 6 (1− ε)k k/2 + ε k ε −1 < 2k k/2 6 2k k/2 ( D2 (x − D2 x)2 )k/2 . Thus, D2 is 2-subgaussian. This completes the proof.  Next, we show that our robust estimation error for the covariance and higher moments are also tight up to constant factors. As before, we will construct two zero-mean, 2-subgaussian distributions on ’ that are at most ε apart in total variation but the variances differ by at least k/2ε 1−2/k . The construction is very similar to the one above for mean estimation. Lemma 7.2. There exist two mean zero distributions D1 and D2 on ’ such that: 1. D1 , D2 are both 2-subgaussian in all moments up to k. 2. dTV (D1 , D2 ) 6 ε. 3. For any ε < 1/2, | D1 4. For any ε < 2−k/2r , | x2 − D1 D2 x 2r − x 2 | > 2k ε 1−2/k . D2 x 2r | > kr 2 ε 1−2r/k . Proof. Let D1 be the standard gaussian distribution N(0, 1). Let D2 be the distribution that with probability 1 − ε outputs a sample from N(0, 1) and with probability ε, outputs a uniform element √ −1/k . Observe that by construction, dTV (D1 , D2 ) 6 ε. of ± kε The mean of D1 , D2 is easily seen to be 0. The same computation as in the proof of Lemma 7.1 establishes that D1 , D2 are both 2-subgaussian. The variance of D1 is 1 and for D2 can be computed to be at least (1 − ε)(1 + kε 1−2/k − kε 2−2/k ) > 1 + 2k ε 1−2/k . The 2rth moment of D1 is ∼ k r and for D2 can be computed to as: k r (1 − ε + ε 1−2r/k ) > r k (1 + 0.5ε 1−2r/k ) for ε < 2−k/2r . This completes the proof.  32 References [AFH+ 12] Anima Anandkumar, Dean P. Foster, Daniel J. Hsu, Sham Kakade, and Yi-Kai Liu, A spectral algorithm for latent dirichlet allocation, NIPS, 2012, pp. 926–934. 1 [AGH+ 14] Animashree Anandkumar, Rong Ge, Daniel J. Hsu, Sham M. Kakade, and Matus Telgarsky, Tensor decompositions for learning latent variable models, Journal of Machine Learning Research 15 (2014), no. 1, 2773–2832. 1 [BCMV14] Aditya Bhaskara, Moses Charikar, Ankur Moitra, and Aravindan Vijayaraghavan, Smoothed analysis of tensor decompositions, STOC, ACM, 2014, pp. 594–603. 1, 9 [Ber06] Thorsten Bernholt, Robust estimators are hard to compute, Tech. report, Technical Report/Universität Dortmund, SFB 475 Komplexitätsreduktion in Multivariaten Datenstrukturen, 2006. 1 [BKS15] Boaz Barak, Jonathan A. Kelner, and David Steurer, Dictionary learning and tensor decomposition via the sum-of-squares method, STOC, ACM, 2015, pp. 143–151. 1, 2, 36 [BM16] Boaz Barak and Ankur Moitra, Noisy tensor completion via the sum-of-squares hierarchy, COLT, JMLR Workshop and Conference Proceedings, vol. 49, JMLR.org, 2016, pp. 417– 445. 2 [BS16] Boaz Barak and David Steurer, Proofs, beliefs, and algorithms through the lens of sum-ofsquares, 2016, Lecture notes in preparation, available on http://sumofsquares.org. 13 [CJN17] Yeshwanth Cherapanamjeri, Prateek Jain, and Praneeth Netrapalli, Thresholding based outlier robust PCA, COLT, Proceedings of Machine Learning Research, vol. 65, PMLR, 2017, pp. 593–628. 1 [CR09] Emmanuel J. Candès and Benjamin Recht, Exact matrix completion via convex optimization, Foundations of Computational Mathematics 9 (2009), no. 6, 717–772. 2 [CRTV05] Emmanuel J. Candès, Mark Rudelson, Terence Tao, and Roman Vershynin, Error correction via linear programming, FOCS, IEEE Computer Society, 2005, pp. 295–308. 2 [CSV17] Moses Charikar, Jacob Steinhardt, and Gregory Valiant, Learning from untrusted data, STOC, ACM, 2017, pp. 47–60. 1 [DKK+ 16] Ilias Diakonikolas, Gautam Kamath, Daniel M. Kane, Jerry Li, Ankur Moitra, and Alistair Stewart, Robust estimators in high dimensions without the computational intractability, FOCS, IEEE Computer Society, 2016, pp. 655–664. 1, 5, 7, 10 [DKK+ 17a] [DKK+ 17b] , Being robust (in high dimensions) can be practical, CoRR abs/1703.00893 (2017). 5 , Robustly learning a gaussian: abs/1704.03866 (2017). 6, 7 33 Getting optimal error, efficiently, CoRR [DLCC07] Lieven De Lathauwer, Joséphine Castaing, and Jean-François Cardoso, Fourth-order cumulant-based blind identification of underdetermined mixtures, IEEE Trans. Signal Process. 55 (2007), no. 6, part 2, 2965–2973. MR 2473563 8, 26 [GLS81] M. Grötschel, L. Lovász, and A. Schrijver, The ellipsoid method and its consequences in combinatorial optimization, Combinatorica 1 (1981), no. 2, 169–197. MR 625550 14, 15 [GM15] Rong Ge and Tengyu Ma, Decomposing overcomplete 3rd order tensors using sum-of-squares algorithms, APPROX-RANDOM, LIPIcs, vol. 40, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015, pp. 829–849. 1 [Gro11] David Gross, Recovering low-rank matrices from few coefficients in any basis, IEEE Trans. Information Theory 57 (2011), no. 3, 1548–1566. 2 [GVX13] Navin Goyal, Santosh Vempala, and Ying Xiao, Fourier PCA, CoRR abs/1306.5825 (2013). 8 [HK13a] Daniel Hsu and Sham M. Kakade, Learning mixtures of spherical Gaussians: moment methods and spectral decompositions, ITCS’13—Proceedings of the 2013 ACM Conference on Innovations in Theoretical Computer Science, ACM, New York, 2013, pp. 11–19. MR 3385380 9 [HK13b] Daniel J. Hsu and Sham M. Kakade, Learning mixtures of spherical gaussians: moment methods and spectral decompositions, ITCS, ACM, 2013, pp. 11–20. 1, 10 [HL17] Sam B. Hopkins and Jerry Li, Mixture models, robustness, and sum of squares proofs, 2017. 5 [HRRS11] Frank R Hampel, Elvezio M Ronchetti, Peter J Rousseeuw, and Werner A Stahel, Robust statistics: the approach based on influence functions, vol. 114, John Wiley & Sons, 2011. 1 [HSS15] Samuel B. Hopkins, Jonathan Shi, and David Steurer, Tensor principal component analysis via sum-of-square proofs, COLT, JMLR Workshop and Conference Proceedings, vol. 40, JMLR.org, 2015, pp. 956–1006. 2 [Hub11] Peter J Huber, Robust statistics, International Encyclopedia of Statistical Science, Springer, 2011, pp. 1248–1251. 1 [KMV10] Adam Tauman Kalai, Ankur Moitra, and Gregory Valiant, Efficiently learning mixtures of two gaussians, STOC, ACM, 2010, pp. 553–562. 1 [KS17] Pravesh K. Kothari and Jacob Steinhardt, Better agnostic clustering via relaxed tensor norms, 2017. 4, 5 [Las01] Jean B. Lasserre, New positive semidefinite relaxations for nonconvex quadratic programs, Advances in convex analysis and global optimization (Pythagorion, 2000), Nonconvex Optim. Appl., vol. 54, Kluwer Acad. Publ., Dordrecht, 2001, pp. 319–331. MR 1846160 14 34 [LRV16] Kevin A. Lai, Anup B. Rao, and Santosh Vempala, Agnostic estimation of mean and covariance, FOCS, IEEE Computer Society, 2016, pp. 665–674. 1, 6, 9 [MMY06] RARD Maronna, R Douglas Martin, and Victor Yohai, Robust statistics, John Wiley & Sons, Chichester. ISBN, 2006. 1 [MR05] Elchanan Mossel and Sébastien Roch, Learning nonsingular phylogenies and hidden markov models, STOC, ACM, 2005, pp. 366–375. 1 [MSS16a] Tengyu Ma, Jonathan Shi, and David Steurer, Polynomial-time tensor decompositions with sum-of-squares, FOCS, IEEE Computer Society, 2016, pp. 438–446. 1, 2, 8, 26, 27, 29 [MSS16b] , Polynomial-time tensor decompositions with sum-of-squares, CoRR abs/1610.01980 (2016). 13, 29 [MV10] Ankur Moitra and Gregory Valiant, Settling the polynomial learnability of mixtures of gaussians, FOCS, IEEE Computer Society, 2010, pp. 93–102. 1 [Nes00] Yurii Nesterov, Squared functional systems and optimization problems, High performance optimization, Appl. Optim., vol. 33, Kluwer Acad. Publ., Dordrecht, 2000, pp. 405–440. MR 1748764 14 [Par00] Pablo A Parrilo, Structured semidefinite programs and semialgebraic geometry methods in robustness and optimization, Ph.D. thesis, California Institute of Technology, 2000. 14 [Pea94] Karl Pearson, Contributions to the mathematical theory of evolution, Philosophical Transactions of the Royal Society of London. A 185 (1894), 71–110. 1 [PS17] Aaron Potechin and David Steurer, Exact tensor completion with sum-of-squares, COLT, Proceedings of Machine Learning Research, vol. 65, PMLR, 2017, pp. 1619–1673. 2 [Rec11] Benjamin Recht, A simpler approach to matrix completion, Journal of Machine Learning Research 12 (2011), 3413–3430. 2 [SCV17] Jacob Steinhardt, Moses Charikar, and Gregory Valiant, Resilience: A criterion for learning in the presence of arbitrary outliers, CoRR abs/1703.04940 (2017). 5 [Sho87] N. Z. Shor, Quadratic optimization problems, Izv. Akad. Nauk SSSR Tekhn. Kibernet. (1987), no. 1, 128–139, 222. MR 939596 14 [Tuk75] John W. Tukey, Mathematics and the picturing of data, 523–531. MR 0426989 1 [VX15] Santosh Vempala and Ying Xiao, Max vs min: Tensor decomposition and ICA with nearly linear sample complexity, COLT, JMLR Workshop and Conference Proceedings, vol. 40, JMLR.org, 2015, pp. 1710–1723. 8 35 A Sum-of-squares toolkit We will rely on the following sum-of-squares version of the inequality between the arithmetic and geometric mean. Lemma A.1 ([BKS15, Lemma A.1]). w k {w 1 > 0, . . . , w n > 0} Lemma A.2. For all even k, a,b k  ( Ö wi 6 i Í i w ik k ) . (a + b)k 6 2k−1 (a k + b k ) . (A.1) Proof. The lemma holds for k  2, because a,b 2  2a 2 + 2b 2 − (a + b)2  a 2 + b 2 − 2ab  (a − b)2 > 0 . (A.2) By iterating this inequality k/2 times, we obtain  a,b k (a + b)k  ((a + b)2 )k/2 6 2k/2 (a 2 + b 2 )k/2 . (A.3) By the binomial identity and Lemma A.1, a,b k 2 2 k/2 (a + b )  k/2   Õ k i0 6 a 2i b k−2i k/2    Õ k i0  i i k/2    Õ k i i0 k/2−1 2 2i k 1 2 (A.4) k k−2i k ·a + · ak + 1 2 · (a k + b k ) . · bk ·b  k    k . Here, we used for the second equality that ki  k−i Combining the previous two inequalities yields the desired inequality, a,b k (a + b)k 6 2k/2 · (a 2 + b 2 )k/2 6 2k−1 · (a k + b k ) . (A.5)  Lemma A.3. For any even k,  fk 6 1 f k 36  f 61 . (A.6) Proof. We apply Lemma A.1 for k/2 and substitute w 1  f 2 and w 2 , w 3 , . . . , w k/2  1 to obtain f k f2 6 f k +(k/2−1) . k/2 Thus,   f fk 6 1 k f2 6 1 . (A.7) At the same time, since 2 − 2 f  1 − f 2 + (1 − f )2 , we have   f f2 6 1 2 1− f > 0 . (A.8) Combining (A.7) and (A.8) yields the desired conclusion. Lemma A.4. For every even k ∈ Ž and all δ < 0.1,  where δ′  100δ. ( f − 1)k 6 δ k ( f + 1)k  f k  1 − δ′ 6 f 6 1 + δ′ , (A.9) Proof. By Lemma A.2, f Therefore, for δ′    4δ k 1−2δ ,  k ( f + 1)k  (( f − 1) + 2)k 6 2k ( f − 1)k + 4k . f ( f − 1)k 6 δ k ( f + 1)k By Lemma A.3, ( f − 1)k 6 (δ′)k equality gives the desired inequality.  k  ( f − 1)k 6 (δ′)k . 1 − δ′ 6 f 6 1 + δ , which together with the previous in-  Proposition A.5. Let A be a set of polynomial equality axioms in variable x such that: A 2t p(x, u) > 0, for a polynomial p with total degree at most 2t. Then, for any pseudo-distribution D of degree 2t on x satisfying A, u 2t ˜ D p(x, u) > 0. Proof. Let A  {q 1  0, q 2  0, . . . , q r  0}. Since A 2t p(x, u) > 0, there are degree 6 2t (in x) polynomials 11 (x, u), 12(x, u), . . . 1r ′ (x, u) and polynomials h1 (x, u), . . . , h r (x, u) such that Í Í p(x, u)  i 1i (x, u)2 + j h j (x, u)q j (x) where the degree of any h j (x, u) · q j (x) in x is at most 2t. Í Taking pseudo-expectations with respect to D, we have: ˜ D p(x, u)  i ˜ D 1i (x, u)2 + Í ˜ Í ˜ 2 j D h j (x, u)q j (x)  i D 1 i (x, u) , where we used that the second term must vanish as D satisfies A. Now, observe that by matching degrees on both sides as polynomial in u, no ˜ D 1i (x, u)2 has degree more than 2t. Further, each 1i (x, u)2 can be written as hG(x)G(x)⊤ , (1, u)⊗2t i for some matrix valued polynomial G(x). 37 Í ˜ ⊤ ⊗2t i. i h D [G(x)G(x) ], (1, u) Taking pseudo-expectations with respect to D yields that: ˜ D p(x, u)  Now, we claim that ˜ D G(x)G(x)⊤ is psd. To see this, let w be any vector and consider the quadratic form of this matrix in w. Then, the quadratic form equals ˜ D hG(x), wi 2 which is nonnegative by the positivity of pseudo-expectations. u  Thus, ˜ D p(x, u) is a sum of squares polynomial in u implying that 2t ˜ D p(x, u) > 0. 38
8cs.DS
INSPECTRE: Privately Estimating the Unseen arXiv:1803.00008v1 [cs.DS] 28 Feb 2018 Jayadev Acharya∗ ECE, Cornell University [email protected] Gautam Kamath† EECS & CSAIL, MIT [email protected] Ziteng Sun‡ ECE, Cornell University [email protected] Huanyu Zhang§ ECE, Cornell University [email protected] March 2, 2018 Abstract We develop differentially private methods for estimating various distributional properties. Given a sample from a discrete distribution p, some functional f , and accuracy and privacy parameters α and ε, the goal is to estimate f (p) up to accuracy α, while maintaining ε-differential privacy of the sample. We prove almost-tight bounds on the sample size required for this problem for several functionals of interest, including support size, support coverage, and entropy. We show that the cost of privacy is negligible in a variety of settings, both theoretically and experimentally. Our methods are based on a sensitivity analysis of several state-of-the-art methods for estimating these properties with sublinear sample complexities. 1 Introduction How can we infer a distribution given a sample from it? If data is in abundance, the solution may be simple – the empirical distribution will approximate the true distribution. However, challenges arise when data is scarce in comparison to the size of the domain, and especially when we wish to quantify “rare events.” This is frequently the case: for example, it has recently been observed that there are several very rare genetic mutations which occur in humans, and we wish to know how many such mutations exist [KC12, TBO+ 12, NWE+ 12]. Many of these mutations have only been seen once, and we can infer that there are many which have not been seen at all. Over the last decade, a large body of work has focused on developing theoretically sound and effective tools for such settings [OSW16] and references therein, including the problem of estimating the frequency distribution of rare genetic variations [ZVV+ 16]. However, in many settings where one wishes to perform statistical inference, data may contain sensitive information about individuals. For example, in medical studies, where the data may contain individuals’ health records and whether they carry some disease which bears a social stigma. Alternatively, one can consider a map application which suggests routes based on aggregate positions of individuals, which contains delicate information including users’ residence data. In these settings, it is critical that our methods protect sensitive information contained in the dataset. This does not preclude our overall goals of statistical analysis, as we are trying to infer properties of the population p, and not the samples which are drawn from said population. That said, without careful experimental design, published statistical findings may be prone to leaking sensitive information about the sample. As a notable example, it was recently shown that one can determine the identity of some individuals who participated in genome-wide association studies [HSR+ 08]. This ∗ Supported by NSF CCF-1657471 and a Cornell University startup grant. by ONR N00014-12-1-0999, NSF CCF-1617730, CCF-1650733, and CCF-1741137. Work partially done while author was an intern at Microsoft Research, New England. ‡ Supported by NSF CCF-1657471 and a Cornell University startup grant. § Supported by NSF CCF-1657471 and a Cornell University startup grant. † Supported 1 realization has motivated a surge of interest in developing data sharing techniques with an explicit focus on maintaining privacy of the data [JS13, USF13, YFSU14, SSB16]. Privacy-preserving computation has enjoyed significant study in a number of fields, including statistics and almost every branch of computer science, including cryptography, machine learning, algorithms, and database theory – see, e.g., [Dal77, AW89, AA01, DN03, Dwo08, DR14] and references therein. Perhaps the most celebrated notion of privacy, proposed by theoretical computer scientists, is differential privacy [DMNS06]. Informally, an algorithm is differentially private if its outputs on neighboring datasets (differing in a single element) are statistically close (for a more precise definition, see Section 2). Differential privacy has become the standard for theoretically-sound data privacy, leading to its adoption by several large technology companies, including Google and Apple [EPK14, Dif17]. Our focus in this paper is to develop tools for privately performing several distribution property estimation tasks. In particular, we study the tradeoff between statistical accuracy, privacy, and error rate in the sample size. Our model is that we are given sample access to some unknown discrete distribution p, over a domain of size k, which is possibly unknown in some tasks. We wish to estimate the following properties: • Support Coverage: If we take m samples from the distribution, what is the expected number of unique elements we expect to see? • Support Size: How many elements of the support have non-zero probability? • Entropy: What is the Shannon entropy of the distribution? For more formal statements of these problems, see Section 2.1. We require that our output is α-accurate, satisfies (ε, 0)-differential privacy, and is correct with probability 1 − β. The goal is to give an algorithm with minimal sample complexity n, while simultaneously being computationally efficient. Theoretical Results. Our main results show that privacy can be achieved for all these problems at a very low cost. For example, if one wishes to privately estimate entropy, this incurs an additional additive cost in the sample complexity which is very close to linear in 1/αε. We draw attention to two features of this bound. First, this is independent of k. All the problems we consider have complexity Θ(k/ log k), so in the primary regime of study where k  1/αε, this small additive cost is dwarfed by the inherent sample complexity of the non-private problem. Second, the bound is almost linear in 1/αε. We note that performing even the most basic statistical task privately, estimating the bias of a coin, incurs this linear dependence. Surprisingly, we show that much more sophisticated inference tasks can be privatized at almost no cost. In particular, these properties imply that the additive cost of privacy is o(1) in the most studied regime where the support size is large. In general, this is not true – for many other problems, including distribution estimation and hypothesis testing, the additional cost of privacy depends significantly on the support size or dimension [DHS15, CDK17, ASZ17, ADR17]. We also provide lower bounds, showing that our upper bounds are almost tight. A more formal statement of our results appears in Section 3. Experimental Results. We demonstrate the efficacy of our method with experimental evaluations. As a baseline, we compare with the non-private algorithms of [OSW16] and [WY18]. Overall, we find that our algorithms’ performance is nearly identical, showing that, in many cases, privacy comes (essentially) for free. We begin with an evaluation on synthetic data. Then, inspired by [VV13, OSW16], we analyze text corpus consisting of words from Hamlet, in order to estimate the number of unique words which occur. Finally, we investigate name frequencies in the US census data. This setting has been previously considered by [OSW16], but we emphasize that this is an application where private statistical analysis is critical. This is proven by efforts of the US Census Bureau to incorporate differential privacy into the 2020 US census [DLS+ 17]. Techniques. Our approach works by choosing statistics for these tasks which possess bounded sensitivity, which is well-known to imply privacy under the Laplace or Gaussian mechanism. We note that bounded sensitivity of statistics is not always something that can be taken for granted. Indeed, for many fundamental tasks, optimal algorithms for the non-private setting may be highly sensitive, thus necessitating crucial modifications to obtain differential privacy [ADK15, CDK17]. Thus, careful choice and design of statistics must be a priority when performing inference with privacy considerations. To this end, we leverage recent results of [ADOS17], which studies estimators for non-private versions of the problems we consider. The main technical work in their paper exploits bounded sensitivity to show sharp cutoff-style concentration bounds for certain estimators, which operate using the principle of best-polynomial approximation. They use these results to show that a single algorithm, the Profile Maximum Likelihood (PML), can estimate all these properties simultaneously. On the other hand, we consider the sensitivity of these estimators for purposes of privacy – the same property is utilized by both works for very different 2 purposes, a connection which may be of independent interest. We note that bounded sensitivity of a statistic may be exploited for purposes other than privacy. For instance, by McDiarmid’s inequality, any such statistic also enjoys very sharp concentration of measure, implying that one can boost the success probability of the test at an additive cost which is logarithmic in the inverse of the failure probability. One may naturally conjecture that, if a statistical task is based on a primitive which concentrates in this sense, then it may also be privatized at a low cost. However, this is not true – estimating a discrete distribution in `1 distance is such a task, but the cost of privatization depends significantly on the support size [DHS15]. One can observe that, algorithmically, our method is quite simple: compute the non-private statistic, and add a relatively small amount of Laplace noise. The non-private statistics have recently been demonstrated to be practical [OSW16, WY18], and the additional cost of the Laplace mechanism is minimal. This is in contrast to several differentially private algorithms which invoke significant overhead in the quest for privacy. Our algorithms attain almost-optimal rates (which are optimal up to constant factors for most parameter regimes of interest), while simultaneously operating effectively in practice, as demonstrated in our experimental results. Related Work. Over the last decade, there have been a flurry of works on the problems we study in this paper by the computer science and information theory communities, including Shannon and Rényi entropy estimation [Pan03, VV17, JVHW17, AOST17, OS17, WY18], support coverage and support size estimation [OSW16, WY18]. A recent paper studies the general problem of estimating functionals of discrete distribution from samples in terms of the smoothness of the functional [FS17]. These have culminated in a nearly-complete understanding of the sample complexity of these properties, with optimal sample complexities (up to constant factors) for most parameter regimes. Recently, there has been significant interest in performing statistical tasks under differential privacy constraints. Perhaps most relevant to this work are [CDK17, ASZ17, ADR17], which study the sample complexity of differentialy privately performing classical distribution testing problems, including identity and closeness testing. Other works investigating private hypothesis testing include [WLK15, GLRV16, KR17, KSF17, Rog17, GR17], which focus less on characterizing the finite-sample guarantees of such tests, and more on understanding their asymptotic properties and applications to computing p-values. There has also been study on private distribution learning [DHS15, DJW17, KV18], in which we wish to estimate parameters of the distribution, rather than just a particular property of interest. A number of other problems have been studied with privacy requirements, including clustering [WWS15, BDL+ 17], principal component analysis [CSS13, KT13, HP14], ordinary least squares [She17], and much more. 2 Preliminaries We will start with some definitions. Pk def Let ∆ = {(p(1), . . . , p(k)) : p(i) ≥ 0, i=1 p(i) = 1, 1 ≤ k ≤ ∞} be the set of discrete distributions over a countable support. Let ∆k be the set of distributions in ∆ with at most k non-zero probability values. A property f (p) is a mapping from ∆ → R. We now describe the classical distribution property estimation problem, and then state the problem under differential privacy. Property Estimation Problem. Given α, β, f , and independent samples X1n from an unknown distribution p, design an estimator fˆ : X1n → R such that with probability at least 1 − β, fˆ(X1n ) − f (p) < α.   def The sample complexity of fˆ, is Cfˆ(f, α, β) = min{n : Pr fˆ(X1n ) − f (p) > α < β} is the smallest number of samples to estimate f to accuracy α, and error β. We study the problem for β = 1/3, and by the median trick, we can boost the error probability to β with an additional multiplicative log(1/β) more samples: def Cfˆ(f, α) = Cfˆ(f, α, 1/3). The sample complexity of estimating a property f (p) is the minimum sample complexity over all estimators: C(f, α) = minfˆ Cfˆ(f, α). An estimator fˆ is ε-differentially private (DP) [DMNS06] if for any X1n and Y1n , with dham (X1n , Y1n ) ≤ 1, Pr (f (X1n )∈S) ≤ eε , for all i, and measurable S. Pr (f (Y1n )∈S ) 3 Private Property Estimation. Given α, ε, β, f , and independent samples X1n from an unknown distribution p, design an ε-differentially private estimator fˆ : X1n → R such that with probability at least 1 − β, fˆ(X1n ) − f (p) < α. Similar to the non-private setting, the sample complexity of ε-differentially private estimation problem is C(f, α, ε) = minfˆ:fˆis ε-DP Cfˆ(f, α, 1/3), the smallest number of samples n for which there exists such a ±α estimator with error probability at most 1/3. In their original paper [DMNS06] provides a scheme for differential privacy, known as the Laplace mechanism. This method adds Laplace noise to a non-private scheme in order to make it private. We first define the sensitivity of an estimator, and then state their result in our setting. def Definition 1. The sensitivity of an estimator fˆ : [k]n → R is ∆n,fˆ = maxdham (X1n ,Y1n )≤1 fˆ(X1n ) − fˆ(Y1n ) . Let Dfˆ(α, ε) = min{n : ∆n,fˆ ≤ αε}. Lemma 1.  n  α o ,ε . C(f, α, ε) = O min Cfˆ(f, α/2) + Dfˆ 4 fˆ Proof. [DMNS06] showed that for a function with sensitivity ∆n,fˆ, adding Laplace noise X ∼ Lap(∆n,fˆ/ε) makes the output ε-differentially private. By the definition of Dfˆ( α4 , ε), the Laplace noise we add has |x| 1 − b , hence we have e parameter at most α4 . Recall that the probability density function of Lap(b) is 2b 1 α Pr (|X| > α/2) < e2 . By the union bound, we get an additive error less than α = 2 + α2 with probability at most 1/3 + e12 < 0.5. Hence, with the median trick, we can boost the error probability to 1/3, at the cost of a constant factor in the number of samples. To prove sample complexity lower bounds for differentially private estimators, we observe that the estimator can be used to test between two distributions with distinct property values, hence is a harder problem. For lower bounds on differentially private testing, [ASZ17] gives the following argument based on coupling: Lemma 2. Suppose there is a coupling between distributions p and q over X n , such that E [dham (X1n , Y1n )] ≤ D. Then, any ε-differentially private algorithm that distinguishes between p and q with error probability at  most 1/3 must satisfy D = Ω 1ε . 2.1 Problems of Interest Support Size. The support size of a distribution p is S(p) = |{x : p(x) > 0}|, the number of symbols with non-zero probability values. However, notice that estimating S(p) from samples can be hard due to the presence of symbols with negligible, yet non-zero probabilities. To circumvent this issue, [RRSS09] proposed def to study the problem when the smallest probability is bounded. Let ∆≥ k1 = {p ∈ ∆ : p(x) ∈ {0} ∪ [1/k, 1]} be the set of all distributions where all non-zero probabilities have value at least 1/k. For p ∈ ∆≥ k1 , our goal is to estimate S(p) up to ±αk with the least number of samples from p. P Support Coverage. For a distribution p, and an integer m, let Sm (p) = x (1 − (1 − p(x))m ), be the expected number of symbols that appear when we obtain m independent samples from the distribution p. The objective is to find the least number of samples n in order to estimate Sm (p) to an additive ±αm. Support coverage arises in many ecological and biological studies [CCG+ 12] to quantify the number of new elements (gene mutations, species, words, etc) that can be expected to be seen in the future. Good and Toulmin [GT56] proposed an estimator that for any constant α, requires m/2 samples to estimate Sm (p). P 1 Entropy. The Shannon entropy of a distribution p is H(p) = x p(x) log p(x) , H(p) is a central object in information theory [CT06], and also arises in many fields such as machine learning [Now12], neuroscience [BWM97, NBdRvS04], and others. Estimating H(p) is hard with any finite number of samples due to the possibility of infinite support. To circumvent this, a natural approach is to consider distributions in ∆k . The goal is to estimate the entropy of a distribution in ∆k to an additive ±α, where ∆k is all discrete distributions over at most k symbols. 4 3 Statement of Results Our theoretical results for estimating support coverage, support size, and entropy are given below. Algorithms for these problems and proofs of these statements are provided in Section 4. Our experimental results are described and discussed in Section 5. Theorem 1. For any ε = Ω(1/m), the sample complexity of support coverage is   m log(1/α) m log(1/α) C(Sm , α, ε) = O + . log m log(εm) Furthermore,   m log(1/α) 1 C(Sm , α, ε) = Ω + . log m αε Theorem 2. For any ε = Ω(1/k), the sample complexity of support size estimation is  C(S, α, ε) = O  k log2 (1/α) k log2 (1/α) + . log k log(εk) Furthermore,   k log2 (1/α) 1 C(S, α, ε) = Ω + . log k αε Theorem 3. Let λ > 0 be any small fixed constant. For instance, λ can be chosen to be any constant between 0.01 and 1. We have the following upper bounds on the sample complexity of entropy estimation:    log2 (min{k, n}) 1 1 k + + log C(H, α, ε) = O α α2 αε αε and k log2 (min{k, n}) C(H, α, ε) = O + + λ2 α log k α2  1 αε 1+λ ! . Furthermore,  C(H, α, ε) = Ω  k log2 (min{k, n}) log k + + . α log k α2 αε We provide some discussion of our results. At a high level, we wish to emphasize the following two points: 1. Our upper bounds show that the cost of privacy in these settings is often negligible compared to the sample complexity of the non-private statistical task, especially when we are dealing with distributions over a large support. Furthermore, our upper bounds are almost tight in all parameters. 2. The algorithmic complexity introduced by the requirement of privacy is minimal, consisting only of a single step which noises the output of an estimator. In other words, our methods are realizable in practice, and we demonstrate the effectiveness on several synthetic and real-data examples. First, we examine our results on support size and support coverage estimation. We note that we focus on the regime where ε is not exceptionally small, as the privacy requirement becomes somewhat unusual. For instance, non-privately, if we have m samples for the problem of support coverage, then the empirical plug-in estimator is the best we can do. However, if ε = O(1/m), then group privacy [DR14] implies that the algorithm’s output distribution on any dataset of m samples must be very similar – however, these samples may have an arbitrary value of support coverage ∈ [m], which precludes hopes for a highly accurate estimator. To avoid degeneracies of this nature, we restrict our attention to ε = Ω(1/m). In this regime, if ε = Ω(mγ /m) for any constant γ > 0, then up to constant factors, our upper bound is within a constant factor of the optimal sample complexity without privacy constratints. In other words, for most meaningful values of ε, privacy comes for free. 5 Next, we turn our attention to entropy estimation. We note that the second upper bound in Theorem 3 has a parameter λ that indicates a tradeoff between the sample complexity incurred in the first and third term. This parameter determines the degree of a polynomial to be used for entropy estimation. As the degree becomes smaller (corresponding to a large λ), accuracy of the polynomial estimator decreases, however, at the same time, low-degree polynomials have a small sensitivity, allowing us to privatize the outcome. In terms of our theoretical results, one can think of λ = 0.01. With this parameter setting, it can be observed that our upper bounds are almost tight. For example, one can see that the upper and lower bounds match to either logarithmic factors (when looking at the first upper bound), or a very small polynomial factor in 1/αε (when looking at the second upper bound). For our experimental results, we experimentally determined an effective value for the parameter λ on a single synthetic instance. We then show that this choice of parameter generalizes, giving highly-accurate private estimation in other instances, on both synthetic on real-world data. 4 Algorithms and Analysis In this section, we prove our results for support coverage in Section 4.1, support size in Section 4.2, and entropy in Section 4.3. In each section, we first describe and analyze our algorithms for the relevant problem. We then go on to describe and analyze a lower bound construction, showing that our upper bounds are almost tight. All our algorithms fall into the following simple framework: 1. Compute a non-private estimate of the property; 2. Privatize this estimate by adding Laplace noise, where the parameter is determined through analysis of the estimator and potentially computation of the estimator’s sensitivity. 4.1 Support Coverage Estimation In this section, we prove Theorem 1, about support coverage estimation: Theorem 1. For any ε = Ω(1/m), the sample complexity of support coverage is   m log(1/α) m log(1/α) + . C(Sm , α, ε) = O log m log(εm) Furthermore,   1 m log(1/α) + . C(Sm , α, ε) = Ω log m αε Our upper bound is described and analyzed in Section 4.1.1, while our lower bound appears in Section 4.1.2. 4.1.1 Upper Bound for Support Coverage Estimation Let ϕi be the number of symbols that appear i times in X1n . We will use the following non-private support coverage estimator from [OSW16]: Ŝm (X1n ) = n X  ϕi 1 + (−t)i · Pr (Z ≥ i) , i=1 where Z is a Poisson random variable with mean r (which is a parameter to be instantiated later), and t = (m − n)/n. Our private estimator of support coverage is derived by adding Laplace noise to this non-private estimator with the appropriate noise parameter, and thus the performance of our private estimator, is analyzed by bounding the sensitivity and the bias of this non-private estimator according to Lemma 1. The sensitivity and bias of this estimator is bounded in the following lemmas. Lemma 3. Suppose m > 2n, then the maximum coefficient of ϕi in Ŝm (p) is at most 1 + er(t−1) . 6 Proof. By the definition of Z, we know Pr (Z ≥ i) = P∞ k=i k e−r rk! , hence we have: |1 + (−t)i · Pr (Z ≥ i)| ≤ 1 + ti ≤1+e ∞ X e−r k=i ∞ X −r k=i ≤ 1 + e−r =1+e rk k! (rt)k k! ∞ X (rt)k k=0 r(t−1) k! The bias of the estimator is bounded in Lemma 4 of [ADOS17]: Lemma 4. Suppose m > 2n, then h i E Ŝm (X1n ) − Sm (p) ≤ 2 + 2er(t−1) + min(m, S(p)) · e−r . Using these results, letting r = log(1/α), [OSW16] showed that there is a constant C, such that with n = C logmm log(1/α) samples, with probability at least 0.9, Ŝm (X1n ) Sm (p) − ≤ α. m m Ŝ (X n ) Our upper bound in Theorem 1 is derived by the following analysis of the sensitivity of m m 1 . If we change one sample in X1n , at most two of the ϕj ’s change. Hence by Lemma 3, the sensitivity of the estimator satisfies !  2  Ŝm (X1n ) ≤ · 1 + er(t−1) . (1) ∆ m m By Lemma 1, there is a private algorithm for support coverage estimation as long as ! Ŝm (X1n ) ≤ αε, ∆ m which by (1) holds if 2(1 + exp(r(t − 1))) ≤ αεm. Let r = log(3/α), note that t − 1 = m n − 2. Suppose αεm > 2, then, the condition above reduces to       1 3 m · − 2 ≤ log αεm − 1 . log α n 2 This is equivalent to m log(3/α) log( 21 αεm − 1) + 2 log(3/α) m log(3/α) = log( 23 εm − 3/α) + log(3/α) n≥ Suppose αεm > 2, then the condition above reduces to the requirement that   m log(1/α) n=O . log(εm) 7 (2) 4.1.2 Lower Bound for Support Coverage Estimation We now prove the lower bound described in Theorem 1. Note that the first term in the lower bound is the sample complexity of non-private support coverage estimation, shown in [OSW16]. Therefore, we turn our attention to prove the latter term in the sample complexity. Consider the following two distributions. u1 is uniform over [m(1 + α)]. u2 is distributed over m + 1 α 1 ∀i ∈ [m] and u2 [4] = 1+α . Moreover, 4 ∈ / [m(1 + α)]. Then, elements [m] ∪ {4} where u2 [i] = m(1+α)   m  1 Sm (u1 ) = m(1 + α) · 1 − 1 − , m(1 + α) and Sm (u2 ) =   m· 1− 1− 1 m(1 + α) m    + 1− 1− α 1+α m  hence, Sm (u2 ) − Sm (u1 )   = mα · 1 − 1 − 1 m(1 + α) m    − 1− 1− α 1+α m  = Ω(αm) α Hence we know there support coverage differs by Ω(αm). Moreover, their total variation distance is 1+α . The following lemma is folklore, based on the coupling interpretation of total variation distance, and the fact that total variation distance is subadditive for product measures. Lemma 5. For any two distributions p, and q, there is a coupling between n iid samples from the two distributions with an expected Hamming distance of dTV (p, q) · n. Using Lemma 5 and dTV (u1 , u2 ) = α 1+α , we have Lemma 6. Suppose u1 and u2 are as defined before, there is a coupling between un1 and un2 with expected α Hamming distance equal to 1+α n. Moreover, given n samples, we must be able to privately distinguish between u1 and u2 given an α accurate estimator of support coverage with privacy considerations. Thus, according to Lemma 2 and 6, we have:   α 1 1 n≥ ⇒n=Ω . 1+α ε εα 4.2 Support Size Estimation In this section, we prove our main theorem about support size estimation, Theorem 2: Theorem 2. For any ε = Ω(1/k), the sample complexity of support size estimation is  C(S, α, ε) = O  k log2 (1/α) k log2 (1/α) + . log k log(εk) Furthermore,   1 k log2 (1/α) + . C(S, α, ε) = Ω log k αε Our upper bound is described and analyzed in Section 4.2.1, while our lower bound appears in Section 4.2.2. 8 4.2.1 Upper Bound for Support Size Estimation In [OSW16], it is shown that the support coverage estimator can be used to obtain optimal results for estimating the support size of a distribution. In this fashion, taking m = k log(3/α), we we may use an estimate of the support coverage Sm (p) as an estimator of S(p). In particular, their result is based on the following observation. Lemma 7. Suppose m ≥ k log(3/α), then for any p ∈ ∆≥ k1 , |Sm (p) − S(p)| ≤ αk . 3 Proof. From the definition of Sm (p), we have Sm (p) ≤ S(p). For the other side, X X m S(p) − Sm (p) = (1 − p(x)) ≤ e−mp(x) x x ≤ k · e− log(3/α) = kα . 3 Therefore, estimating Sm (p) for m = k log(3/α), up to ±αk/3, also estimates S(p) up to ±αk. Therefore, the goal is to estimate the smallest value of n to solve the support coverage problem. Suppose r = log(3/α), and m = k log(3/α) = k · r in the support coverage problem. Then, we have t= k log(3/α) m −1= − 1. n n (3) Then, by Lemma 4 in the previous section, we have h i E Ŝm (X1n ) − S(p) h i ≤ E Ŝm (X1n ) − Sm (p) + |Sm (p) − S(p)| ≤ 2 + 2er(t−1) + min{m, k} · e−r + ≤ 2 + 2er(t−1) + k · e− log(3/α) + ≤ 2 + 2er(t−1) + 2 kα 3 kα 3 kα . 3 We will find conditions on n such that the middle  term above is at most kα. Toward this end, note that 2er(t−1) ≤ αk holds if and only if r(t − 1) ≤ log αk 2 . Plugging in (3), this holds when     k log(3/α) αk log(3/α) · − 2 ≤ log , n 2 which is equivalent to n≥ k log2 (3/α) log αk 2 + 2 log  3 α =O k log2 (1/α) log k  where we have assumed without loss of generality that α > k1 . The computations for sensitivity are very similar. From Lemma 1 1, we need to find the value of n such that 2 + 2er(t−1) ≤ αεk, where we assume that n ≤ 12 k log(3/α), else we just add noise to the true number of observed distinct elements By computations similar to the expectation case, this reduces to n≥ k log2 (3/α) . 3 log αεk 2 + log α 9 Therefore, this gives us a sample complexity of k log2 (1/α) n=O log (εk)   (4) for the sensitivity result to hold. We note that the bound above blows up when ε ≤ k1 . However, we note that our lower bound implies that we need at least Ω(1/ε) = Ω(k) samples in this case, which is not in the sub-linear regime that we are interested in. We therefore consider only the regime where the privacy parameter ε is at least 1/k. 4.2.2 Lower Bound for Support Size Estimation In this section, we prove a lower bound for support size estimation, as described in Theorem 2. The techniques are similar to those for support coverage in Section 4.1.2. The first term of the complexity is the lower bound for non-private setting. This follows by combining the lower bound of [OSW16] for support coverage, with the equivalence between estimation of support size and coverage as implied by Lemma 7. We focus on the second term in the sequel. Consider the following two distributions: u1 is a uniform distribution over [k] and u2 is a uniform distribution over [(1−α)k]. Then the support size of these two distribution differs by αk, and dTV (u1 , u2 ) = α. Hence by Lemma 5, we know the following: Lemma 8. Suppose u1 ∼ U [k] and u2 ∼ U [(1 − α)k], there is a coupling between un1 and un2 with expected Hamming distance equal to αn. Moreover, given n samples, we must be able to privately distinguish between u1 and u2 given an α accurate estimator of entropy with privacy considerations. Thus, according to Lemma 2 and Lemma 8, we have:   1 1 . αn ≥ ⇒ n = Ω ε εα 4.3 Entropy Estimation In this section, we prove our main theorem about entropy estimation, Theorem 3: Theorem 3. Let λ > 0 be any small fixed constant. For instance, λ can be chosen to be any constant between 0.01 and 1. We have the following upper bounds on the sample complexity of entropy estimation:    log2 (min{k, n}) 1 1 k + + log C(H, α, ε) = O α α2 αε αε and k log2 (min{k, n}) C(H, α, ε) = O + + λ2 α log k α2  1 αε 1+λ ! . Furthermore,  C(H, α, ε) = Ω  k log2 (min{k, n}) log k + + . α log k α2 αε We describe and analyze two upper bounds. The first is based on the empirical entropy estimator, and is described and analyzed in Section 4.3.1. The second is based on the method of best-polynomial approximation, and appears in Section 4.3.2. Finally, our lower bound is in Section 4.3.3. 4.3.1 Upper Bound for Entropy Estimation: The Empirical Estimator Our first private entropy estimator is derived by adding Laplace noise into the empirical estimator. The parameter of the Laplace distribution is ∆(H(ε p̂n )) , where ∆(H(p̂n )) denotes the sensitivity of the empirical estimator. By analyzing its sensitivity and bias, we prove an upper bound on the sample complexity for private entropy estimation and get the first upper bound in Theorem 3. 10 Let p̂n be the empirical distribution, and let H(p̂n ) be the entropy of the empirical distribution. The theorem is based on the following three facts:   log n ∆(H(p̂n )) = O . (5) n   k |H(p) − E [H(p̂n )]| = O , (6) n  2  log (min{k, n}) V ar (H(p̂n )) = O , (7) n With these three facts in hand, the sample complexity of the empiricalestimator can be bounded as follows. 1 1 log( αε ) . We also need |H(p) − E [H(p̂n )]| = By Lemma 1, we need ∆(H(p̂n )) ≤ αε, which gives  n = O αε  2 log (min{k,n}) k 2 O(α) and V ar (H(p̂n )) = O α , which gives n = O n + . α2 Proof of (5). The largest change in any Nx when we change one symbol is one. Moreover, at most two Nx change. Therefore, n j n j+1 log − log n j+1 n j j j 1 n = 2 · max log + log j=1...n−1 n j+1 n j+1   j 1 n j log , log ≤ 2 · max max j=1...n−1 n j+1 n j+1   1 log n ≤ 2 · max , , n n log n . =2· n ∆(H(p̂n )) ≤ 2 · Proof of (6). max j=1...n−1 (8) (9) (10) By the concavity of entropy function, we know that E [H(p̂n )] ≤ H(p). Therefore, E [|H(p) − H(p̂n )|] = H(p) − E [H(p̂n )] " # X =E (p̂n (x) log p̂n (x) − p(x) log p(x)) x " =E X x # " # X p̂n (x) p̂n (x) log +E (p̂n (x) − p(x)) log p(x) p(x) x = E [D(p̂n kp)]   ≤ E dχ2 (p̂n || p) " # X (p̂n (x) − p(x))2 =E p(x) x X (p(x)/n) ≤ p(x) x = (11) (12) (13) k . n (14) 11 2 Proof of (7). The variance bound of logn k is given precisely in Lemma 15 of [JVHW17]. To obtain the other half of the bound of, we apply the bounded differences inequality in the form stated in Corollary 3.2 of [BLM13]. Lemma 9. Let f : Ωn → R be a function. Suppose further that max 0 0 z1 ,...,zn ,zi f (z1 , . . . , zn ) − f (z1 , . . . , zi−1 , zi , . . . , zn ) ≤ ci . Then for independent variables Z1 , . . . , Zn , n Var(f (Z1 , . . . , Zn )) ≤ 1X 2 c . 4 i=1 i Therefore, using Lemma 9 and Equation (5)  V ar (H(p̂n )) ≤ n · 4.3.2 4 log2 n n2  = 4 log2 n . n Upper Bound for Entropy Estimation: Best-Polynomial Approximation We prove an upper bound on the sample complexity for private entropy estimation if one adds Laplace noise into best-polynomial estimator.This will give us the second upper bound in Theorem 3. In the non-private setting the optimal sample complexity of estimating H(p) over ∆k is given by Theorem 1 of [WY16]   k log2 (min{k, n}) Θ + . α log k α2 However, this estimator can have a large sensitivity. [ADOS17] designed an estimator that has the same sample complexity but a smaller sensitivity. We restate Lemma 6 of [ADOS17] here: Lemma 10. Let λ > 0 be a fixed small constant, which may be taken to be any value between 0.01 and 1. Then there is an entropy estimator with sample complexity   1 k log2 (min{k, n}) + Θ 2· , λ α log k α2 and has sensitivity nλ /n. We can now invoke Lemma 1 on the estimator in this lemma to obtain the upper bound on private entropy estimation. 4.3.3 Lower Bound for Entropy Estimation We now prove the lower bound for entropy estimation. Note that any lower bound on privately testing two distributions p, and q such that H(p) − H(q) = Θ(α) is a lower bound on estimating entropy. We analyze the following construction for Proposition 2 of [WY16]. The two distributions p, and q over [k] are defined as: 1 − p(1) 2 ,p(i) = , for i = 2, . . . , k, 3 k−1 2−η 1 − q(1) q(1) = ,q(i) = , for i = 2, . . . , k. 3 k−1 p(1) = Then, by the grouping property of entropy, H(p) = h(2/3) + 1 1+η · log(k − 1), and H(q) = h((2 − η)/3) + · log(k − 1), 3 3 12 (15) (16) which gives H(p) − H(q) = Ω(η log k). For η = α/ log k, the entropy difference becomes Θ(α). The total variation distance between p and q is η/3. By Lemma 5 in the paper, there is a coupling over X1n , and Y1n generated from p and q with expected Hamming distance at most dTV (p, q) · n. This along with Lemma 2 in the paper gives a lower bound of Ω(log k/αε) on the sample complexity. 5 Experiments We evaluated our methods for entropy estimation and support coverage on both synthetic and real data. Overall, we found that privacy is quite cheap: private estimators achieve accuracy which is comparable or near-indistinguishable to non-private estimators in many settings. Our results on entropy estimation and support coverage appear in Sections 5.1 and 5.2, respectively. Code of our implementation is available at https://github.com/HuanyuZhang/INSPECTRE. 5.1 Entropy We compare the performance of our entropy estimator with a number of alternatives, both private and non-private. Non-private algorithms considered include the plug-in estimator (plug-in), the Miller-Madow Estimator (MM) [Mil55], the sample optimal polynomial approximation estimator (poly) of [WY16]. We analyze the privatized versions of plug-in, and poly in Sections 4.3.1 and 4.3.2, respectively. The implementation of the latter is based on code from the authors of [WY16]1 . We compare performance on different distributions including uniform, a distribution with two steps, Zipf(1/2), a distribution with Dirichlet-1 prior, and a distribution with Dirichlet-1/2 prior, and over varying support sizes. While plug-in, and MM are parameter free, poly (and its private counterpart) have to choose the degree L of the polynomial to use, which manifests in the parameter λ in the statement of Theorem 3. [WY16] suggest the value of L = 1.6 log k in their experiments. However, since we add further noise, we choose a single L as follows: (i) Run privatized poly for different L values and distributions for k = 2000, ε = 1, (b) Choose the value of L that performs well across different distributions (See Figure 1). We choose L = 1.2·log k from this, and use it for all other experiments. To evaluate the sensitivity of poly, we computed the estimator’s value at all possible input values, computed the sensitivity, (namely, ∆ = maxdham (X1n ,Y1n )≤1 |poly(X1n )−poly(Y1n )|),  and added noise distributed as Lap 0, ∆ ε . 0.6 0.4 0.8 0.6 500 750 1000 1250 1500 Number of samples 1750 2000 0.0 1.0 0.8 500 750 1000 1250 1500 Number of samples 1750 2000 0.0 1.50 1.25 1.00 0.75 L=0.3log(k) L=0.6log(k) L=0.9log(k) L=1.2log(k) L=1.5log(k) L=1.8log(k) 2.0 1.5 1.0 0.5 0.25 0.2 250 L=0.3log(k) L=0.6log(k) L=0.9log(k) L=1.2log(k) L=1.5log(k) L=1.8log(k) 1.75 0.50 0.4 0.2 250 1.2 0.6 0.4 0.2 0.0 1.0 1.4 Dirichlet-1/2 prior Dirichlet-1 prior L=0.3log(k) L=0.6log(k) L=0.9log(k) L=1.2log(k) L=1.5log(k) L=1.8log(k) 1.6 RMSE 0.8 1.2 RMSE RMSE 1.0 Zipf 1/2 L=0.3log(k) L=0.6log(k) L=0.9log(k) L=1.2log(k) L=1.5log(k) L=1.8log(k) 1.4 RMSE Two steps L=0.3log(k) L=0.6log(k) L=0.9log(k) L=1.2log(k) L=1.5log(k) L=1.8log(k) RMSE Uniform 1.2 250 500 750 1000 1250 1500 Number of samples 1750 2000 0.00 250 500 750 1000 1250 1500 Number of samples 1750 2000 0.0 250 500 750 1000 1250 1500 Number of samples 1750 2000 Figure 1: RMSE comparison for private Polynomial Approximation Estimator with various values for degree L, k = 2000, ε = 1. The RMSE of various estimators for k = 1000, and ε = 1 for various distributions are illustrated in Figure 2. The RMSE is averaged over 100 iterations in the plots. We observe that the performance of our private-poly is near-indistinguishable from the non-private poly, particularly as the number of samples increases. It also performs significantly better than all other alternatives, including the non-private Miller-Madow and the plug-in estimator. The cost of privacy is minimal for several other settings of k and ε, for which results appear in Section A. 1 See https://github.com/Albuso0/entropy for their code for entropy estimation. 13 2.5 1.5 2.5 2.0 1.5 2.0 2.0 1.5 1.0 1.0 0.5 0.5 0.5 0.5 400 600 800 0.0 1000 Number of samples 200 400 600 800 0.0 1000 Number of samples 200 400 600 800 Number of samples 2.0 1.5 1.0 0.5 0.0 1000 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 1.5 1.0 200 Dirichlet-1/2 prior Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 1.0 0.0 Dirichlet-1 prior 3.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 RMSE 2.0 RMSE RMSE 2.5 Zipf 1/2 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 RMSE Two steps Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 RMSE Uniform 3.5 200 400 600 800 0.0 1000 Number of samples 200 400 600 800 1000 Number of samples Figure 2: Comparison of various estimators for the entropy, k = 1000, ε = 1. 5.2 Support Coverage We investigate the cost of privacy for the problem of support coverage. We provide a comparison between the Smoothed Good-Toulmin estimator (SGT) of [OSW16] and our algorithm, which is a privatized version of their statistic (see Section 4.1.1). Our implementation is based on code provided by the authors of [OSW16]. As shown in our theoretical results, the sensitivity of SGT is at most 2(1 + er (t − 1)), necessitating the addition of Laplace noise with parameter 2(1 + er(t−1) )/ε. Note that while the theory suggests we select the 2 1 loge n(t+1) parameter r = log(1/α), α is unknown. We instead set r = 2t t−1 , as previously done in [OSW16]. 5.2.1 Evaluation on Synthetic Data In our synthetic experiments, we consider different distributions over different support sizes k. We generate n = k/2 samples, and then estimate the support coverage at m = n · t. For large t, estimation is harder. Some results of our evaluation on synthetic are displayed in Figure 3. We compare the performance of SGT, and privatized versions of SGT with parameters ε = 1, 2, and 10. For this instance, we fixed the domain size k = 20000. We ran the methods described above with n = k/2 samples, and estimated the support coverage at m = nt, for t ranging from 1 to 10. The performance of the estimators is measured in terms of RMSE over 1000 iterations. 1500 Uniform 1000 Non-private Private eps=10 Private eps=2 Private eps=1 800 Two steps 1200 Non-private Private eps=10 Private eps=2 Private eps=1 1000 Zipf 1/2 1400 Non-private Private eps=10 Private eps=2 Private eps=1 1200 1000 800 400 RMSE RMSE RMSE RMSE 600 1000 600 400 500 0 1 200 2 3 4 5 t 6 7 8 9 10 0 1 3 4 5 t 6 7 8 9 10 0 1 1800 1600 1400 600 3 4 5 t 6 7 8 9 10 0 1 1000 800 600 400 200 2 Dirichlet-1/2 prior Non-private Private eps=10 Private eps=2 Private eps=1 1200 800 400 200 2 Dirichlet-1 prior Non-private Private eps=10 Private eps=2 Private eps=1 RMSE 2000 200 2 3 4 5 t 6 7 8 9 10 0 1 2 3 4 5 t 6 7 8 9 10 Figure 3: Comparison between the private estimator with the non-private SGT when k = 20000 We observe that, in this setting, the cost of privacy is relatively small for reasonable values of ε. This is as predicted by our theoretical results, where unless ε is extremely small (less than 1/k) the non-private sample complexity dominates the privacy requirement. However, we found that for smaller support sizes (as shown in Section A.2), the cost of privacy can be significant. We provide an intuitive explanation for why no private estimator can perform well on such instances. To minimize the number of parameters, we instead argue about the related problem of support-size estimation. Suppose we are trying to distinguish between distributions which are uniform over supports of size 100 and 200. We note that, if we draw n = 50 samples, the “profile” of the samples (i.e., the histogram of the histogram) will be very similar for the two distributions. In particular, if one modifies only a few samples (say, five or six), one could convert one profile into the other. In other words, these two profiles are almost-neighboring datasets, but simultaneously correspond to very different support sizes. This pits the two goals of privacy and accuracy at odds with each other, thus resulting in a degradation in accuracy. 5.2.2 Evaluation on Census Data and Hamlet We conclude with experiments for support coverage on two real-world datasets, the 2000 US Census data and the text of Shakespeare’s play Hamlet, inspired by investigations in [OSW16] and [VV17]. Our investigation 14 on US Census data is also inspired by the fact that this is a setting where privacy is of practical importance, evidenced by the proposed adoption of differential privacy in the 2020 US Census [DLS+ 17]. The Census dataset contains a list of last names that appear at least 100 times. Since the dataset is so oversampled, even a small fraction of the data is likely to contain almost all the names. As such, we make the task non-trivial by subsampling mtotal = 86080 individuals from the data, obtaining 20412 distinct last names. We then sample n of the mtotal individuals without replacement and attempt to estimate the total number of last names. Figure 4 displays the RMSE over 100 iterations of this process. We observe that even an exceptionally stringent privacy budget of ε = 0.5, the performance is almost indistinguishable from the non-private SGT estimator. 6000 Non-private Private eps=2 Private eps=1 Private eps=0.5 5000 RMSE 4000 3000 2000 1000 0 0.1 0.2 0.3 0.4 0.5 0.6 Fraction of seen names 0.7 0.8 0.9 1.0 Figure 4: Comparison between our private estimator with the SGT on Census Data The Hamlet dataset has mtotal = 31, 999 words, of which 4804 are distinct. Since the distribution is not as oversampled as the Census data, we do not need to subsample the data. Besides this difference, the experimental setup is identical to that of the Census dataset. Once again, as we can see in Figure 5, we get near-indistinguishable performance between the non-private and private estimators, even for very small values of ε. Our experimental results demonstrate that privacy is realizable in practice, with particularly accurate performance on real-world datasets. 1800 Non-private Private eps=2 Private eps=1 Private eps=0.5 1600 1400 RMSE 1200 1000 800 600 400 200 0 0.1 0.2 0.3 0.4 0.5 0.6 Fraction of seen words 0.7 0.8 0.9 1.0 Figure 5: Comparison between our private estimator with the SGT on Hamlet References [AA01] Dakshi Agrawal and Charu C. Aggarwal. On the design and quantification of privacy preserving data mining algorithms. In Proceedings of the 20th ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems, PODS ’01, pages 247–255, New York, NY, USA, 2001. ACM. 15 [ADK15] Jayadev Acharya, Constantinos Daskalakis, and Gautam Kamath. Optimal testing for properties of distributions. In Advances in Neural Information Processing Systems 28, NIPS ’15, pages 3577–3598. Curran Associates, Inc., 2015. [ADOS17] Jayadev Acharya, Hirakendu Das, Alon Orlitsky, and Ananda Theertha Suresh. A unified maximum likelihood approach for estimating symmetric properties of discrete distributions. In Proceedings of the 34th International Conference on Machine Learning, ICML ’17, pages 11–21. JMLR, Inc., 2017. [ADR17] Maryam Aliakbarpour, Ilias Diakonikolas, and Ronitt Rubinfeld. Differentially private identity and closeness testing of discrete distributions. arXiv preprint arXiv:1707.05497, 2017. [AOST17] Jayadev Acharya, Alon Orlitsky, Ananda Theertha Suresh, and Himanshu Tyagi. Estimating rényi entropy of discrete distributions. IEEE Transactions on Information Theory, 63(1):38–56, 2017. [ASZ17] Jayadev Acharya, Ziteng Sun, and Huanyu Zhang. Differentially private testing of identity and closeness of discrete distributions. arXiv preprint arXiv:1707.05128, 2017. [AW89] Nabil R. Adam and John C. Worthmann. Security-control methods for statistical databases: A comparative study. ACM Computing Surveys (CSUR), 21(4):515–556, 1989. [BDL+ 17] Maria-Florina Balcan, Travis Dick, Yingyu Liang, Wenlong Mou, and Hongyang Zhang. Differentially private clustering in high-dimensional euclidean spaces. In Proceedings of the 34th International Conference on Machine Learning, ICML ’17, pages 322–331. JMLR, Inc., 2017. [BLM13] Stephane Boucheron, Gabor Lugosi, and Pierre Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, 2013. [BWM97] Michael J. Berry, David K Warland, and Markus Meister. The structure and precision of retinal spike trains. Proceedings of the National Academy of Sciences, 94(10):5411–5416, 1997. [CCG+ 12] Robert K. Colwell, Anne Chao, Nicholas J. Gotelli, Shang-Yi Lin, Chang Xuan Mao, Robin L. Chazdon, and John T. Longino. Models and estimators linking individual-based and samplebased rarefaction, extrapolation and comparison of assemblages. Journal of Plant Ecology, 5(1):3–21, 2012. [CDK17] Bryan Cai, Constantinos Daskalakis, and Gautam Kamath. Priv’it: Private and sample efficient identity testing. In Proceedings of the 34th International Conference on Machine Learning, ICML ’17, pages 635–644. JMLR, Inc., 2017. [CSS13] Kamalika Chaudhuri, Anand D. Sarwate, and Kaushik Sinha. A near-optimal algorithm for differentially-private principal components. Journal of Machine Learning Research, 14(Sep):2905–2943, 2013. [CT06] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley, 2006. [Dal77] Tore Dalenius. Towards a methodology for statistical disclosure control. Statistisk Tidskrift, 15:429–444, 1977. [DHS15] Ilias Diakonikolas, Moritz Hardt, and Ludwig Schmidt. Differentially private learning of structured discrete distributions. In Advances in Neural Information Processing Systems 28, NIPS ’15, pages 2566–2574. Curran Associates, Inc., 2015. [Dif17] Differential Privacy Team, Apple. Learning with privacy at scale. https://machinelearning. apple.com/docs/learning-with-privacy-at-scale/appledifferentialprivacysystem. pdf, December 2017. [DJW17] John C. Duchi, Michael I. Jordan, and Martin J. Wainwright. Minimax optimal procedures for locally private estimation. Journal of the American Statistical Association, 2017. 16 [DLS+ 17] Aref N. Dajani, Amy D. Lauger, Phyllis E. Singer, Daniel Kifer, Jerome P. Reiter, Ashwin Machanavajjhala, Simson L. Garfinkel, Scot A. Dahl, Matthew Graham, Vishesh Karwa, Hang Kim, Philip Lelerc, Ian M. Schmutte, William N. Sexton, Lars Vilhuber, and John M. Abowd. The modernization of statistical disclosure limitation at the U.S. census bureau, 2017. Presented at the September 2017 meeting of the Census Scientific Advisory Committee. [DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Proceedings of the 3rd Conference on Theory of Cryptography, TCC ’06, pages 265–284, Berlin, Heidelberg, 2006. Springer. [DN03] Irit Dinur and Kobbi Nissim. Revealing information while preserving privacy. In Proceedings of the 22nd ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems, PODS ’03, pages 202–210, New York, NY, USA, 2003. ACM. [DR14] Cynthia Dwork and Aaron Roth. The Algorithmic Foundations of Differential Privacy. Now Publishers, Inc., 2014. [Dwo08] Cynthia Dwork. Differential privacy: A survey of results. In Proceedings of the 5th International Conference on Theory and Applications of Models of Computation, TAMC ’08, pages 1–19, Berlin, Heidelberg, 2008. Springer. [EPK14] Úlfar Erlingsson, Vasyl Pihur, and Aleksandra Korolova. RAPPOR: Randomized aggregatable privacy-preserving ordinal response. In Proceedings of the 2014 ACM Conference on Computer and Communications Security, CCS ’14, pages 1054–1067, New York, NY, USA, 2014. ACM. [FS17] Kazuto Fukuchi and Jun Sakuma. Minimax optimal estimators for additive scalar functionals of discrete distributions. In Proceedings of the 2017 IEEE International Symposium on Information Theory, ISIT ’17, pages 2103–2107, Washington, DC, USA, 2017. IEEE Computer Society. [GLRV16] Marco Gaboardi, Hyun-Woo Lim, Ryan M. Rogers, and Salil P. Vadhan. Differentially private chi-squared hypothesis testing: Goodness of fit and independence testing. In Proceedings of the 33rd International Conference on Machine Learning, ICML ’16, pages 1395–1403. JMLR, Inc., 2016. [GR17] Marco Gaboardi and Ryan Rogers. Local private hypothesis testing: Chi-square tests. arXiv preprint arXiv:1709.07155, 2017. [GT56] I.J. Good and G.H. Toulmin. The number of new species, and the increase in population coverage, when a sample is increased. Biometrika, 43(1-2):45–63, 1956. [HP14] Moritz Hardt and Eric Price. The noisy power method: A meta algorithm with applications. In Advances in Neural Information Processing Systems 27, NIPS ’14, pages 2861–2869. Curran Associates, Inc., 2014. [HSR+ 08] Nils Homer, Szabolcs Szelinger, Margot Redman, David Duggan, Waibhav Tembe, Jill Muehling, John V. Pearson, Dietrich A. Stephan, Stanley F. Nelson, and David W. Craig. Resolving individuals contributing trace amounts of dna to highly complex mixtures using high-density snp genotyping microarrays. PLoS Genetics, 4(8):1–9, 2008. [JS13] Aaron Johnson and Vitaly Shmatikov. Privacy-preserving data exploration in genome-wide association studies. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’13, pages 1079–1087, New York, NY, USA, 2013. ACM. [JVHW17] Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman. Minimax estimation of functionals of discrete distributions. IEEE Transactions on Information Theory, 61(5):2835– 2885, 2017. 17 [KC12] Alon Keinan and Andrew G. Clark. Recent explosive human population growth has resulted in an excess of rare genetic variants. Science, 336(6082):740–743, 2012. [KR17] Daniel Kifer and Ryan M. Rogers. A new class of private chi-square tests. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS ’17, pages 991–1000. JMLR, Inc., 2017. [KSF17] Kazuya Kakizaki, Jun Sakuma, and Kazuto Fukuchi. Differentially private chi-squared test by unit circle mechanism. In Proceedings of the 34th International Conference on Machine Learning, ICML ’17, pages 1761–1770. JMLR, Inc., 2017. [KT13] Michael Kapralov and Kunal Talwar. On differentially private low rank approximation. In Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’13, pages 1395–1414, Philadelphia, PA, USA, 2013. SIAM. [KV18] Vishesh Karwa and Salil Vadhan. Finite sample differentially private confidence intervals. In Proceedings of the 9th Conference on Innovations in Theoretical Computer Science, ITCS ’18, New York, NY, USA, 2018. ACM. [Mil55] George A. Miller. Note on the bias of information estimates. Information Theory in Psychology: Problems and Methods, 2:95–100, 1955. [NBdRvS04] Ilya Nemenman, William Bialek, and Rob de Ruyter van Steveninck. Entropy and information in neural spike trains: Progress on the sampling problem. Physical Review E, 69(5):056111:1– 056111:6, 2004. [Now12] Sebastian Nowozin. Improved information gain estimates for decision tree induction. In Proceedings of the 29th International Conference on Machine Learning, ICML ’12, pages 571–578. JMLR, Inc., 2012. [NWE+ 12] Matthew R. Nelson, Daniel Wegmann, Margaret G. Ehm, Darren Kessner, Pamela St. Jean, Claudio Verzilli, Judong Shen, Zhengzheng Tang, Silviu-Alin Bacanu, Dana Fraser, Liling Warren, Jennifer Aponte, Matthew Zawistowski, Xiao Liu, Hao Zhang, Yong Zhang, Jun Li, Yun Li, Li Li, Peter Woollard, Simon Topp, Matthew D. Hall, Keith Nangle, Jun Wang, Gonçalo Abecasis, Lon R. Cardon, Sebastian Zöllner, John C. Whittaker, Stephanie L. Chissoe, John Novembre, and Vincent Mooser. An abundance of rare functional variants in 202 drug target genes sequenced in 14,002 people. Science, 337(6090):100–104, 2012. [OS17] Maciej Obremski and Maciej Skorski. Renyi entropy estimation revisited. In Proceedings of the 20th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems, APPROX ’17, pages 20:1–20:15, Dagstuhl, Germany, 2017. Schloss Dagstuhl– Leibniz-Zentrum fuer Informatik. [OSW16] Alon Orlitsky, Ananda Theerta Suresh, and Yihong Wu. Optimal prediction of the number of unseen species. Proceedings of the National Academy of Sciences, 113(47):13283–13288, 2016. [Pan03] Liam Paninski. Estimation of entropy and mutual information. 15(6):1191–1253, 2003. [Rog17] Ryan Michael Rogers. Leveraging Privacy in Data Analysis. PhD thesis, University of Pennsylvania, May 2017. [RRSS09] Sofya Raskhodnikova, Dana Ron, Amir Shpilka, and Adam Smith. Strong lower bounds for approximating distribution support size and the distinct elements problem. SIAM Journal on Computing, 39(3):813–842, 2009. [She17] Or Sheffet. Differentially private ordinary least squares. In Proceedings of the 34th International Conference on Machine Learning, ICML ’17, pages 3105–3114. JMLR, Inc., 2017. 18 Neural Computation, [SSB16] Sean Simmons, Cenk Sahinalp, and Bonnie Berger. Enabling privacy-preserving GWASs in heterogeneous human populations. Cell Systems, 3(1):54–61, 2016. [TBO+ 12] Jacob A. Tennessen, Abigail W. Bigham, Timothy D. O’Connor, Wenqing Fu, Eimear E. Kenny, Simon Gravel, Sean McGee, Ron Do, Xiaoming Liu, Goo Jun, Hyun Min Kang, Daniel Jordan, Suzanne M. Leal, Stacey Gabriel, Mark J. Rieder, Goncalo Abecasis, David Altshuler, Deborah A. Nickerson, Eric Boerwinkle, Shamil Sunyaev, Carlos D. Bustamante, Michael J. Bamshad, Joshua M. Akey, Broad GO, Seattle GO, and on behalf of the NHLBI Exome Sequencing Project. Evolution and functional impact of rare coding variation from deep sequencing of human exomes. Science, 337(6090):64–69, 2012. [USF13] Caroline Uhler, Aleksandra Slavković, and Stephen E. Fienberg. Privacy-preserving data sharing for genome-wide association studies. The Journal of Privacy and Confidentiality, 5(1):137– 166, 2013. [VV13] Gregory Valiant and Paul Valiant. Estimating the unseen: Improved estimators for entropy and other properties. In Advances in Neural Information Processing Systems 26, NIPS ’13, pages 2157–2165. Curran Associates, Inc., 2013. [VV17] Gregory Valiant and Paul Valiant. Estimating the unseen: Improved estimators for entropy and other properties. Journal of the ACM, 64(6):37:1–37:41, 2017. [WLK15] Yue Wang, Jaewoo Lee, and Daniel Kifer. Revisiting differentially private hypothesis tests for categorical data. arXiv preprint arXiv:1511.03376, 2015. [WWS15] Yining Wang, Yu-Xiang Wang, and Aarti Singh. Differentially private subspace clustering. In Advances in Neural Information Processing Systems 28, NIPS ’15, pages 1000–1008. Curran Associates, Inc., 2015. [WY16] Yihong Wu and Pengkun Yang. Minimax rates of entropy estimation on large alphabets via best polynomial approximation. IEEE Transactions on Information Theory, 62(6):3702–3720, 2016. [WY18] Yihong Wu and Pengkun Yang. Chebyshev polynomials, moment matching, and optimal estimation of the unseen. The Annals of Statistics, 2018. [YFSU14] Fei Yu, Stephen E. Fienberg, Aleksandra B. Slavković, and Caroline Uhler. Scalable privacypreserving data sharing methodology for genome-wide association studies. Journal of Biomedical Informatics, 50:133–141, 2014. [ZVV+ 16] James Zou, Gregory Valiant, Paul Valiant, Konrad Karczewski, Siu On Chan, Kaitlin Samocha, Monkol Lek, Shamil Sunyaev, Mark Daly, and Daniel G. MacArthur. Quantifying unobserved protein-coding variants in human populations provides a roadmap for large-scale sequencing projects. Nature Communications, 7, 2016. A Additional Experimental Results This section contains additional plots of our synthetic experimental results. Section A.1 contains experiments on entropy estimation, while Section A.2 contains experiments on estimation of support coverage. A.1 Entropy Estimation We present four more plots of our synthetic experimental results for entropy estimation. Figures 6 and 7 are on a smaller support of k = 100, with ε = 1 and 2, respectively. Figures 8 and 9 are on a support of k = 1000, with ε = 0.5 and 2. 19 Uniform 3.0 2.0 2.5 1.5 2.0 1.5 2.5 1.5 1.0 0.5 0.5 0.5 40 60 Number of samples 80 100 20 Zipf 1 2.0 40 60 Number of samples 80 0.0 100 2.5 RMSE RMSE 1.5 0.5 0.5 0.5 80 100 20 40 60 Number of samples (a) 100 1.5 1.0 60 80 80 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.0 1.0 40 60 2.5 1.0 Number of samples 40 Number of samples Dirichlet-1/2 prior Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.0 1.5 20 20 Dirichlet-1 prior 3.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 RMSE 2.0 1.0 20 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 1.0 0.0 Zipf 1/2 3.5 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 RMSE RMSE 2.5 Two steps 3.5 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace RMSE 3.5 100 20 40 60 Number of samples (b) 80 100 (c) Figure 6: Comparison of various estimators for the entropy, k = 100, ε = 1. Uniform 3.0 2.0 1.5 Zipf 1/2 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 2.5 RMSE RMSE 2.5 Two steps 3.5 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.0 1.5 2.5 2.0 1.5 1.0 1.0 1.0 0.5 0.5 0.5 20 40 60 Number of samples 80 0.0 100 Zipf 1 2.5 40 60 Number of samples 80 100 20 Dirichlet-1 prior 3.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.0 20 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 RMSE 3.5 60 80 100 Dirichlet-1/2 prior Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 40 Number of samples Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 2.0 1.5 1.0 1.0 0.5 0.5 20 40 60 Number of samples (a) 80 100 RMSE 1.5 RMSE RMSE 2.0 1.5 1.0 0.5 20 40 60 Number of samples (b) 80 100 20 40 60 Number of samples (c) Figure 7: Comparison of various estimators for the entropy, k = 100, ε = 2. 20 80 100 Two steps Uniform Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 2.0 3.0 2.5 RMSE 1.5 2.5 1.5 2.0 1.5 1.0 1.0 0.5 0.5 0.5 0.0 0.0 1.0 200 400 600 800 Number of samples 1000 Zipf 1 1.75 1.50 1.00 400 600 800 Number of samples 0.0 1000 200 Dirichlet-1 prior 400 600 800 Number of samples 1000 Dirichlet-1/2 prior Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 2.0 2.0 RMSE 1.25 200 3.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.00 RMSE 2.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 RMSE RMSE 2.5 Zipf 1/2 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace RMSE 3.5 1.5 0.75 1.0 0.50 0.5 1.5 1.0 0.5 0.25 200 400 600 800 Number of samples 0.0 1000 200 400 600 800 Number of samples (a) 0.0 1000 200 400 600 800 Number of samples (b) 1000 (c) Figure 8: Comparison of various estimators for the entropy, k = 1000, ε = 0.5. Two steps Uniform Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 2.0 3.0 2.5 RMSE 1.5 2.5 1.5 2.0 1.5 1.0 1.0 0.5 0.5 0.5 0.0 0.0 1.0 200 400 600 800 Number of samples 1000 Zipf 1 2.00 1.50 1.00 0.75 400 600 800 Number of samples 200 400 600 800 Number of samples 1000 Dirichlet-1/2 prior Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.0 1.5 1.0 0.50 0.0 1000 Dirichlet-1 prior 2.5 RMSE 1.25 200 3.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 1.75 RMSE 2.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 3.0 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace 2.5 2.0 RMSE RMSE 2.5 Zipf 1/2 Plug-in Miller Poly Poly-Laplace Plug-in-Laplace RMSE 3.5 1.5 1.0 0.5 0.5 0.25 0.00 200 400 600 800 Number of samples (a) 1000 0.0 200 400 600 800 Number of samples (b) 1000 0.0 200 400 600 (c) Figure 9: Comparison of various estimators for the entropy, k = 1000, ε = 2. 21 800 Number of samples 1000 200 150 100 50 0 1 300 250 Two steps 200 Non-private Private eps=10 Private eps=2 Private eps=1 150 RMSE RMSE 150 Uniform Non-private Private eps=10 Private eps=2 Private eps=1 RMSE 200 100 50 2 3 4 5 6 t Zipf 1 7 8 9 0 1 10 200 Non-private Private eps=10 Private eps=2 Private eps=1 150 Zipf 1/2 Non-private Private eps=10 Private eps=2 Private eps=1 100 50 2 3 4 5 6 t Dirichlet-1 prior 7 8 9 0 1 10 200 Non-private Private eps=10 Private eps=2 Private eps=1 150 2 3 4 5 6 7 8 9 10 6 7 8 9 10 t Dirichlet-1/2 prior Non-private Private eps=10 Private eps=2 Private eps=1 150 RMSE RMSE RMSE 200 100 100 100 50 50 50 0 1 2 3 4 5 t 6 7 8 9 (a) 10 0 1 2 3 4 5 t (b) 6 7 8 9 10 0 1 2 3 4 5 t (c) Figure 10: Comparison between the private estimator with the non-private SGT when k = 1000. A.2 Support Coverage We present three additional plots of our synthetic experimental results for support coverage estimation. In particular, Figures 10, 11, and 12 show support coverage for k = 1000, 5000, 100000. 22 450 Non-private Private eps=10 Private eps=2 Private eps=1 500 400 350 RMSE RMSE 400 300 200 100 Two steps 450 Non-private Private eps=10 Private eps=2 Private eps=1 400 350 300 300 250 250 RMSE Uniform 600 200 0 1 1000 800 2 3 4 5 6 t Zipf 1 7 8 9 150 100 100 0 1 10 600 Non-private Private eps=10 Private eps=2 Private eps=1 500 50 2 3 4 5 6 t Dirichlet-1 prior 7 8 9 0 1 10 600 Non-private Private eps=10 Private eps=2 Private eps=1 500 RMSE RMSE 400 200 0 1 2 3 4 5 t 6 7 8 9 300 3 4 5 6 7 8 9 10 6 7 8 9 10 t Dirichlet-1/2 prior Non-private Private eps=10 Private eps=2 Private eps=1 300 200 200 100 100 0 1 10 2 400 RMSE 400 600 200 150 50 Zipf 1/2 Non-private Private eps=10 Private eps=2 Private eps=1 2 3 4 5 (a) t 6 7 8 9 0 1 10 2 3 4 5 (b) t (c) Figure 11: Comparison between the private estimator with the non-private SGT when k = 5000. 3000 2500 RMSE RMSE 4000 12000 2 3 4 5 6 t Zipf 1 7 8 9 1500 0 1 10 4500 Non-private Private eps=10 Private eps=2 Private eps=1 4000 3500 6000 2 3 4 5 6 t Dirichlet-1 prior 7 8 9 6000 Non-private Private eps=10 Private eps=2 Private eps=1 5000 2000 3 4 5 (a) t 6 7 8 9 10 3 4 5 6 7 8 9 10 6 7 8 9 10 t Dirichlet-1/2 prior Non-private Private eps=10 Private eps=2 Private eps=1 3000 2000 1000 500 2 2 4000 2500 0 1 1500 0 1 10 1000 2000 2000 500 1500 4000 Zipf 1/2 Non-private Private eps=10 Private eps=2 Private eps=1 1000 3000 8000 0 1 2500 500 RMSE RMSE 10000 3000 1000 2000 14000 3500 2000 6000 0 1 Two steps Non-private Private eps=10 Private eps=2 Private eps=1 RMSE 8000 Uniform Non-private Private eps=10 Private eps=2 Private eps=1 RMSE 10000 2 3 4 5 (b) t 6 7 8 9 10 0 1 2 3 4 5 t (c) Figure 12: Comparison between the private estimator with the non-private SGT when k = 100000. 23
7cs.IT
Chaos, Solitons & Fractals When Darwin meets Lorenz: Evolving new chaotic attractors through genetic programming Indranil Pan a,* and Saptarshi Das b P P P a) Department of Earth Science and Engineering, Imperial College London, Exhibition Road, London SW7 2AZ, United Kingdom (Email: [email protected] ) 31TU U 31T b) School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, United Kingdom (Email: [email protected] ) 31T 31T Abstract: In this paper, we propose a novel methodology for automatically finding new chaotic attractors through a computational intelligence technique known as multi-gene genetic programming (MGGP). We apply this technique to the case of the Lorenz attractor and evolve several new chaotic attractors based on the basic Lorenz template. The MGGP algorithm automatically finds new nonlinear expressions for the different state variables starting from the original Lorenz system. The Lyapunov exponents of each of the attractors are calculated numerically based on the time series of the state variables using time delay embedding techniques. The MGGP algorithm tries to search the functional space of the attractors by aiming to maximise the largest Lyapunov exponent (LLE) of the evolved attractors. To demonstrate the potential of the proposed methodology, we report over one hundred new chaotic attractor structures along with their parameters, which are evolved from just the Lorenz system alone. Keywords: chaos; genetic programming; Lorenz family; Lyapunov exponent 1. Introduction: An important research theme in non-linear dynamics is to identify sets of differential equations along with their parameters which give rise to chaos. Starting from the advent of Lorenz attractors in three dimensional nonlinear differential equations [1][2], its several other family of attractors have been developed like Rossler, Rucklidge, Chen, Lu, Liu, Sprott, Genesio-Tesi, Shimizu-Morioka etc. [3]. Extension of the basic attractors to four or even higher dimensional systems has resulted in a similar family of hyper-chaotic systems [4]. In addition, several other structures like multi-scroll [5] and multi-wing [6] versions of the Lorenz family of attractors have also been developed by increasing the number of equilibrium points. Development of new chaotic attractors has huge application especially in data encryption, secure communication [7] etc. and in understanding the dynamics of many real world systems whose governing equations match with the template of these chaotic systems [8]. There has been several research reports on the application of master-slave chaos synchronization in secure communication [9], where the fresh set of chaotic attractors can play a big role due to their rich phase space dynamics. Here we explore the potential of automatic generation of chaotic attractors which is developed from the basic three dimensional structure of the Lorenz system. We use the Lyapunov exponent – the most popular signature of chaos, to judge whether a computer generated arbitrary nonlinear structure of third order differential equation along with some chosen parameters exhibits chaotic motion in the phase space. Since there could be chaotic behaviour or complex limit cycles or even stable/unstable motions in the phase space, for unknown mathematical expressions of similar third order dynamical system, the computationally tractable way for the 1 Chaos, Solitons & Fractals investigation of the chaos seems to be the characterisation of the observed time series. The Taken’s theorem says that the original attractor could be reconstructed from the observation of just one state variable using the time delay embedding method [10]. In the time delay reconstruction method, the dynamics of the chaotic attractors is approximated in the phase space by plotting the observed time series and its delayed versions along orthogonal axes and finding the delay that has got maximum span in the phase space [10], [11]. We predominantly report similar chaotic attractors evolved over the basic Lorenz system of equations by changing two and three state equations together using the GP. We use standard nonlinear terms in the Lorenz family like the cross product and square terms in combination with sinusoidal terms, giving rise to multiple equilibrium points and hence very complex dynamical behaviour in the phase space e.g. with infinite number of equilibrium points [12][13]. Previous attempts have looked at chaotic dynamical systems modelling through the use of genetic programming [14], [15] using nonlinear autoregressive moving average with exogenous inputs or NARMAX models. The papers try to reproduce the dynamics of the Chua circuit through the nonlinear autoregressive models with different lags and orders. They also use multi-objective genetic algorithms to obtain a set of parsimonious models which can reproduce the original behaviour of the known chaotic attractor. The disadvantage of the paper is that the obtained expressions have multiple time delay terms in them, which are more complex than the ones obtained intuitively (i.e. which have three coupled first order nonlinear differential terms). Also, due to the method employed, the obtained chaotic dynamics are very similar to the original attractor and their phase portraits do not differ drastically, i.e. new types of chaotic attractors with totally different phase space dynamics are difficult to obtain using this method. There have been attempts of evolving new single state discrete time chaotic dynamical systems using the concept of GP and study of their bifurcation diagrams [16]. Similar work has been done in [17] using evolutionary algorithms and in [18] using analytical programming techniques. The technique has also been extended to higher number of states and used for evolutionary reconstruction of continuous time chaotic systems [19]. The present paper uses a MGGP paradigm to evolve multiple expressions for the state variables simultaneously. The multi-gene approach helps in obtaining new sets of dynamical equations which can show completely different phase space dynamics. This fact is exploited in the GP algorithm to find new sets of chaotic dynamical systems by maximising the LLE. This is the first paper of its kind that reports not only a single or a handful of attractors by varying the basic Lorenz system of equations like Rössler, Chen, Lu, Liu etc., but over hundreds of new interesting differential equation structures along with their parameters and Lyapunov exponents. The generic method proposed in this paper can be viewed as the first step towards explaining very complex chaotic motions hidden in various physical systems, which could be obtained by mixing the Lorenz equation with simple transcendental like sinusoids [20], which has got infinite number of equilibria. 2. Genetic programming to evolve new chaotic attractors 2.1. Basics of Genetic programming Genetic programming is an intelligent algorithm which is capable of automatically evolving computer programs to perform a given task [21], [22]. The GP algorithm has been applied to a variety of practical applications in human competitive engineering design [23]. GP has been used for symbolic regression to fit analytical nonlinear expressions to do prediction for any given experimental input2 Chaos, Solitons & Fractals output data set [24][25]. Essentially it can evolve the structure and parameters of a nonlinear expression by minimising the mean squared error (MSE) between the predicted and the observed values. This expression is analytic in nature and is therefore amenable to mathematical analysis. The present work exploits this paradigm to evolve analytical expressions for chaotic attractors, by trying to maximise the LLE of each of the nonlinear dynamical systems. The GP algorithm is based on the Darwinian principle of evolution and survival of the fittest. The recently introduced multi-gene GP or MGGP [24] is used in the present study. Each state equation which the GP searches is represented by one gene. If the GP simultaneously searches for two state equations, then they are represented by two separate genes and together they constitute one individual. These genes can be represented in the form of a tree structure as shown in Figure 1. In the beginning, the individuals are randomly initialized within the feasible space. Then they undergo reproduction, crossover and mutation to evolve fitter individuals in the subsequent generations. Crossover refers to the interchange of genetic material among the solutions. Mutation on the other hand refers to a random change within a gene itself. The crossover and mutation operations are stochastic ones and their probability of occurrence is pre-specified by the user. The tree structure representation as shown in Figure 1 is useful for doing the cross over and mutation operations algorithmically. Generally the maximum number of levels in a tree is confined to a specific small number to decrease the bloat in the solutions and also to reduce the run time of the algorithms. Here bloat refers to the case where the GP goes on evolving very complicated nonlinear expressions, without significant increase in the performance metric (i.e. the objective function) [23]. Figure 1: Schematic of the automatic evolution of new chaotic attractors using GP, based on the Lorenz system 3 Chaos, Solitons & Fractals The MGGP algorithm, as introduced in [24], uses a tree structure to represent each sub-gene and a weighted combination of these sub-genes are used to represent one individual gene. In this paper, we use the sub-genes to represent each of the state equations of the chaotic attractor, but we do not use the weights. Therefore the whole dynamical system is represented by one gene which has multiple sub-genes for each of the different state variables. So for example, if the objective is to evolve only two of the three state equations of the dynamical system, then two sub-genes would represent the overall system (since the other state equation is constant). 2.2. Objective function For any experimentally observed time series, the Lyapunov exponent of the reconstructed attractor having value greater than one is considered as a genuine signature of chaos [26]. In the present paper, we exploit this fact to identify whether a given set of equations are chaotic or not. There are many different methods of calculating the LLE or the Largest Lyapunov Exponent from the analytical set of differential equations. However, most of them require calculation of the Jacobian, which is difficult to do automatically through a computer program as they might involve complicated nonlinear expressions and unnecessarily increase the computation time within each run of the GP. Therefore to circumvent this problem we resorted to a numerical method of evaluating the LLE [27]. We simulated each of the GP evolved expression with an ordinary differential equation solver using a fixed time step of 0.01 seconds for a total time of 25 seconds. Then from the short time series obtained from the x-state variable [28], we used the time delay embedding method to reconstruct the attractor dynamics as done in Rosenstein et al. [27] and computed the LLE numerically. The GP algorithm aims to maximise the LLE in order to obtain systems which exhibit strong chaotic behaviour. It is to be noted that since the LLE is calculated from only one of the state variables by reconstructing the original attractor dynamics, it might not always give the true Lyapunov exponent of the system. However, we found that this is able to easily distinguish between chaotic and non-chaotic sequences and is therefore suitable as a performance metric to evaluate the chaotic dynamics of a given system. One of the problems is that the GP algorithm frequently evolves expressions whose solutions blow up. These expressions are not useful and a high value of fitness function is assigned to these solutions to discourage the algorithm from exploring these parts of the search space. 2.3. Different cases considered in this study Depending on which state equations are modified in the original Lorenz system (1), four different cases are identified as shown below in equations (2)-(5) by modifying two/three states together in the template of the celebrated Lorenz system [1]. x = 10 ( y − x ) , y = 28 x − xz − y, z = xy − ( 8 / 3) z i) Case A: Generalized type y-z Lorenz family x = 10 ( y − x ) , y = f ( x, y, z ) , z = h ( x, y, z ) ii) (2) Case B: Generalized type x-y Lorenz family = x g ( x, y, z ) , = y iii) (1) f ( x, y , z ) , = z xy − ( 8 / 3) z Case C: Generalized type x-z Lorenz family 4 (3) Chaos, Solitons & Fractals x = g ( x, y, z ) , y = 28 x − xz − y, z = h ( x, y, z ) iv) (4) Case D: Generalized x-y-z family (no structure) = x g= ( x, y, z ) , y f= ( x, y, z ) , z h ( x, y, z ) (5) In case A (Lorenz y-z family), the original expression of the Lorenz system is retained for the x state variable, while the expressions for the y and z states are evolved by the GP algorithm in each iteration. For cases B (Lorenz x-y family) and C (Lorenz x-z family) the algorithm evolves the equations for the states x-y and x-z respectively. Case D (Lorenz x-y-z family) is a generalized template which can evolve any chaotic attractor and the algorithm evolves all the three state equations involving the rate of x, y and z. Therefore this template has the potential to evolve even the other well-known chaotic attractors like Rössler, Chen, Lu, Rabinovich-Fabrikant systems etc. 3. Simulation results The GP algorithm is run for the different cases A to D for 100 times each and the unique structures with positive LLEs are identified and saved for post-processing. The parameter settings of the GP algorithm are shown in Table 1. One of the important parameters is the ratio of the mathematical operators to that of the numeric values in each gene which is taken as 0.5. A higher value would imply more mathematical operators which would make the expressions more complicated. A lower value would have more number of constants in the expression, which would make the expressions simpler but would reduce the capability of the expression to exhibit rich time domain and phase space dynamics. Table 1: Parameter settings for the MGGP algorithm GP Algorithm parameters Parameter settings Population size 25 Number of generation 50 Selection method Plain lexicographic tournament selection Tournament size 3 Termination criteria Maximum number of iterations Maximum depth of tree 5 Maximum tree depth for mutation 2 Ratio of mathematical operators to numeric values 0.5 Mathematical operators {+, −, ×, ÷,sin} In general the objective is to obtain all dynamical systems which have a positive Lyapunov exponent and not the one with the maximum Lyapunov exponent alone. Therefore we added a history variable in the GP algorithm which keeps track of all the nonlinear expressions which had a Lyapunov exponent greater than zero. It is essential to preserve the chaotic attractors in the intermediate generations as subsequent generations might destroy the attractor structure and evolve more fitter structures (with higher LLE) through the crossover and mutation operations. The simulations were run on a 64 bit Windows desktop with 16 GB memory and an Intel I7, 3.4 GHz processor. 100 runs of 5 Chaos, Solitons & Fractals each generalized family of attractors (i.e. each of the cases A, B, C, D) took approximately two days each. After the simulation runs, we screened the attractors which have a positive LLE as given by the time delay embedding technique. However the LLE obtained from this method is not reliable since it is computed by the phase space reconstruction using time delay embedding of just one of the state variables using a very short time series. Therefore we symbolically computed the Jacobian of each of the automatically evolved differential equations and calculated all the Lyapunov exponents and hence the LLE numerically, using the standard method as reported in [29]. We have also validated our algorithm with the Lyapunov exponents of standard attractors given in [29]. The algorithm to find the LLE of the new structures comprises of symbolic differentiation of the state equations to obtain the Jacobian, as an intermediate step. All our simulation results are reported with an initial condition of x0 {= 2,= y0 3,= z0 1} using a fixed step-size of 0.5 sec in the ode45 routine of Matlab (which uses Dormand-Price algorithm). For this numerical integration a total time span of 10000 second is simulated. Table 2: List of generalized Lorenz x-z family of attractors Name x expression z expression Coefficient values LLE LorXZ1 a y − b x + sin( z ) − c xy − dz a=10.0, b=10.0, c=5.9666, d=2.6667 0.8824 LorXZ2 a y − b x + sin( sin( y ) ) − c xy − dz a=10.0, b=10.0, c=2.6667, d=2.6667 0.9040 LorXZ3 a y − b x + sin( z ) − c xd − e z a=10.0, b=10.0, c=5.9666, d=2, e=2.6667 0.9809 LorXZ4 a y − b x − c z + sin( z ) − d xy − ez a=10.0, b=10.0, c=1.0, d=2.6667, e=2.6667 0.8952 LorXZ5 x − a y + b z + sin( sin( z ) − c y + d ) sin( x ) a=1.0, b=7.6825, c=1.0, d=3.9819 0.0606 LorXZ6 ay − bx x( y + c ) − d z a=10.0, b=10.0, c=0.24176, d=2.6667 0.8985 LorXZ7 ay − bx x( x + c ) − d z a=10.0, b=10.0, c=0.24176, d=2.6667 0.9626 LorXZ8 ay − bx x y − c z − d y + x ( sin( z ) − e ) a=10.0, b=10.0, c=2.6667, d=1.0, e=2.3465 0.7488 6 Chaos, Solitons & Fractals LorXZ9 a sin( z ) b sin( y ) a=10.0, b=100.0 0.0036 LorXZ10 a y − b x + sin( sin( x ) ) + sin( x − c y ) xy − dz a=10.0, b=10.0, c=1.0, d=2.6667 0.9031 LorXZ11 a y − b x + sin( sin( x ) ) xy − cz a=10.0, b=10.0, c=2.6667 0.9053 LorXZ12 a y − b x + sin( x − c y ) xy − dz a=10.0, b=10.0, c=1.0, d=2.6667 0.8985 LorXZ13 a y − b x + sin( x − c ) x − ez a=10.0, b=10.0, c=4.5177, d=2, e=2.6667 0.9863 LorXZ14 a y − bx − c sin( y − d z ) xy − ez a=11.0, b=10.0, c=1.0, d=1.0, e=2.6667 0.9674 LorXZ15 a x( z − b ) c sin( d y ) − e x a=10.0, b=1.0, c=10.0, d=6.6163, e=10.0 0.0722 LorXZ16 ay − bx x ( x + sin( z ) ) − c z a=10.0, b=10.0, c=2.6667 0.9325 d Here we explore the four different cases – Lorenz XZ, XY and YZ and XYZ family of attractors evolved using GP. These are reported in Table 2-Table 5 respectively. In the first three cases, one expression is kept common as in the Lorenz system and in the fourth case, all the state equations are evolved using GP. Symbolically computed LLE and the co-efficient values (exhibiting chaos) are also reported along with their nomenclature in the following tables which describes their dynamical behaviour in the 3D phase space. It is to be noted that we have reported only those chaotic systems which has at least one sine term in any of its state equations because it opens up the possibility of exhibiting rich phase space dynamics [20]. 7 Chaos, Solitons & Fractals Figure 2: Phase space dynamics of LorXZ3 (trunk) Figure 3: Phase space dynamics of LorXZ5 (toroidal spring) 8 Chaos, Solitons & Fractals Figure 4: Phase space dynamics of LorXZ9 (tangled string) Figure 5: Phase space dynamics of LorXZ15 Due to the large number of chaotic attractors obtained through GP, most of the phase portraits of the new systems are shown in the supplementary material. We here show the 2D and 3D phase space 9 Chaos, Solitons & Fractals diagrams of only those systems which has some interesting patterns in contrast with the standard Lorenz system. For example in Figure 2 for the LorXZ3 system, especially in the y-z plane, there is a prominently large ‘trunk’ between the two wings and its LLE is 0.98 which is close to that of the classical Lorenz system. Whereas the LorXZ5 system in Figure 3 shows very complex dynamics like a ‘toroidal spring’ and its LLE has deviated significantly from the original Lorenz system. LorXZ9 system in Figure 4 shows trapping of the phase space dynamics around several equilibrium points which takes a shape of a ‘tangled string’. System LorXZ15 shows high oscillations near low y values and takes multiple triangular excursions in the phase space combining both fast and slow dynamics together. Both the LorXZ5 and LorXZ15 have a small LLE of 0.06 and 0.07 respectively, whereas the LorXZ9 shows zero LLE up to two decimal place which refers to a weak chaos [30]. Phase portraits of the rest of the chaotic attractors in Table 2 are shown in the supplementary material and have the LLEs ranging between 0.74-0.98 which is pretty close to that of the original Lorenz system. In Table 3, similar results are reported for the Lorenz x-y family of attractors. In Figure 6, the phase space diagram of LorXY10 is shown which has a relatively smaller LLE of 0.5803. This system shows ‘uneven wings’ and the trajectories are more concentrated around one wing as compared to the other. A similar LLE of 0.58 and the resulting phase space dynamics can also be found in LorXY15. It is also interesting to note that in this family some of them have LLE greater than one e.g. LorXY4, LorXY5, LorXY21, LorXY22, LorXY23, LorXY25, LorXY28, LorXY29, LorXY31, LorXY33, LorXY34, indicating strong chaotic dynamics [30]. For example the LorXY21 system in Figure 7 shows a dense phase space diagram (especially in y-z plane) which is also confirmed by its high LLE of 1.1416. In Table 3, LorXY22 has the highest LLE of 1.2 among all the attractors in the LorXY family. Rest of the attractors’ LLEs in the LorXY family vary between 0.87-1.06. Table 3: List of generalized Lorenz x-y family of attractors Name x expression y expression LorXY1 ay − bx c x + sin( d z ) − e x z LorXY2 ay − bx c x + sin( d x ) − e x z LorXY3 ay − bx c x − d sin( sin( x + e ) ) − f x z LorXY4 ay − bx c x + y − d sin( x ) − e x z − f LorXY5 ay − bx c x + d y + esin( y ) − f x z LorXY6 ay − bx c x + sin( y − d )( sin( x ) + e ) − f x z LorXY7 ay − bx c x − d y − esin( sin( y ) ) − f x z LorXY8 ay − bx c x + sin( x ) − d x z LorXY9 ay − bx c x + sin( y + d ) − e x z LorXY10 ay − bx − c d x + sin( y + e ) − f x z LorXY11 ay − bx c x − d sin( x + y − e ) − f x z 10 Coefficient values a=10.0, b=10.0, c=28.0, d=5.0302, e=1.0 a=10.0, b=10.0, c=28.0, d=5.0302, e=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=6.2452, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=16.727 a=10.0, b=10.0, c=28.0, d=0.044006, e=0.044006, f=1.0 a=10.0, b=10.0, c=28.0, d=9.09, e=0.82679, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0 a=10.0, b=10.0, c=27.0, d=1.0 a=10.0, b=10.0, c=28.0, d=14.276, e=1.0 a=10.0, b=10.0, c=62.569, d=28.0, e=14.276, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=6.2569, f=1.0 LLE 0.9894 0.9783 0.9978 1.0598 1.0037 0.9924 0.8791 0.9765 0.9858 0.5803 0.9919 Chaos, Solitons & Fractals LorXY12 ay − bx c x + sin( d x − e y ) − f x z LorXY13 ay − bx c x − d sin( x − e y − f ) − g x z LorXY14 ay − bx c x − d sin( x − e y ) − f x z LorXY15 ay − bx − c d x + sin( e x − f y ) − g x z LorXY16 ay − bx c x − d sin( sin( x ) ) − e x z LorXY17 ay − bx c x − d sin( z ) − e x z + f LorXY18 ay − bx c x − d sin( e y ) − f x z LorXY19 ay − bx c x − d sin z e − f x z LorXY20 ay − bx c x − d sin( x y ) − e x z LorXY21 ay − bx c x − d sin( y z ) − e x z f LorXY22 ay − bx −c x z d + e x + sin( x − f y ) LorXY23 ay − bx c x + y + sin( x ) − d x z + e LorXY24 ay − bx c x − d sin sin x e LorXY25 ay − bx c x − d z − esin( z ) − f x z g + h LorXY26 ay − bx c x + sin( x − d y ) + sin( z ) − e x z − f LorXY27 ay − bx c x + sin( x − d y ) − esin( y ) + sin( z ) − f x z LorXY28 ay − bx c x − d sin( y ) + sin( z ) − e x z − f LorXY29 ay − bx c x + y − d sin( y ) + sin( z ) − e x z − f LorXY30 ay − bx c x + sin( x − d ) − esin( y ) + sin( z ) − f x z LorXY31 ay − bx c x + y + sin( x − d y ) − esin( y ) − f x z LorXY32 ay − bx c x + sin( x − d y ) − esin( x ) + sin( z ) − f x z LorXY33 ay − bx c x + y + sin( z ) − d x z − e LorXY34 ay − bx c x − d sin( x ) − e x z ( ) ( ( )) − f x z 11 a=10.0, b=10.0, c=28.0, d=10.0, e=10.0, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=6.386, g=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0 a=10.0, b=10.0, c=62.569, d=28.0, e=2.0, f=1.0, g=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=9.3771 a=10.0, b=10.0, c=28.0, d=1.0, e=4.8349, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=2, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=2 a=10.0, b=10.0, c=1.0, d=2, e=28.0, f=1.0 a=10.0, b=10.0, c=27.0, d=1.0, e=7.9648 a=10.0, b=10.0, c=28.0, d=1.0, e=2, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0, g=2, h=1.8865 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=13.173 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0 a=10.0, b=10.0, c=29.0, d=1.0, e=1.0, f=6.6666 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=6.6666 a=10.0, b=10.0, c=28.0, d=8.2736, e=1.0, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0 a=10.0, b=10.0, c=27.0, d=1.0, e=6.6666 a=10.0, b=10.0, c=30.0, d=1.0, e=1.0 0.9765 0.9965 0.9978 0.5800 0.9970 0.9618 0.9155 0.9684 0.9755 1.1416 1.2034 1.0629 0.9791 1.0457 0.9085 0.9853 1.0073 1.0635 0.9907 1.0438 0.9879 1.0592 1.0373 Chaos, Solitons & Fractals Figure 6: Phase space dynamics of LorXY10 (uneven wings) Figure 7: Phase space dynamics of LorXY21 12 Chaos, Solitons & Fractals Table 4: List of generalized Lorenz y-z family of attractors Name y expression LorYZ1 z expression Coefficient values LLE a x + sin sin( sin( y ) ) − b x z xy − cz a=28.0, b=1.0, c=2.6667 0.9942 LorYZ2 a x + sin( y ) − b x z xy − cz a=28.0, b=1.0, c=2.6667 0.9893 LorYZ3 a x + sin( sin( z ) ) − b x z xy − cz a=28.0, b=1.0, c=2.6667 0.9958 LorYZ4 a x + sin( sin( x + b ) ) − c x z xy − dz LorYZ5 a x + sin sin( sin( y ) ) − b x z xc − d z LorYZ6 a x − b sin( x ) − c x z xy − d LorYZ7 −a x ( z + sin( x ) − b ) xc − d z LorYZ8 −a x ( z − b ) x ( c x − d sin( x ) ) − e z LorYZ9 a x − b sin( c z ) − d x z xy − ez LorYZ10 a x − b sin( z ( y + c ) ) − d x z xy − ez LorYZ11 a x − b sin z c − d x z xy − ez LorYZ12 a x − b sin( c x ) − d x z xy − ez LorYZ13 a x − b sin( c x − d ) − e x z xy − fz LorYZ14 a x − b sin( c x z ) − d x z xy − ez LorYZ15 a x − b sin( z ) − c x z xy − d LorYZ16 a x − b sin( c x ) − d x z xy − e LorYZ17 a x + sin( b x ) − c x z xy − dz LorYZ18 −a x z + sin( sin( z ) ) − b LorYZ19 −a sin( y + z )( z − b sin( y ) + c ) −d sin( x ) LorYZ20 a x − b sin( x y ) − c x z xy − dz LorYZ21 a x − b sin( y − c z + d ) − e x z xy − fz LorYZ22 −a x z b + c x − d sin( x ) ex f − g z LorYZ23 −a x z b + c x − d sin( x ) exy − fz LorYZ24 −a x z b + c x − d sin( y ) ex f − g z LorYZ25 a x − b sin( x ) − c x z d xe − f z LorYZ26 sin( y ) − a y − b x z xy − c a=1.0, b=1.0, c=74.667 1.4151 LorYZ27 sin( y ) − a y − b x z yc − d 1.0990 LorYZ28 x sin( y ) − a x z − b y a=1.0, b=1.0, c=2, d=74.667 xy − c a=1.0, b=1.0, c=74.667 1.4381 ( ) ( ) ( ) ( ) c x d − ez 13 a=28.0, b=28.0, c=1.0, d=2.6667 a=28.0, b=1.0, c=2, d=2.6667 a=2.6667, b=1.0, c=1.0, d=74.746 a=1.0, b=28.0, c=2, d=2.6667 a=1.0, b=28.0, c=2.6667, d=1.0, e=2.6667 a=28.0, b=1.0, c=2.6667, d=1.0, e=2.6667 a=28.0, b=1.0, c=5.1757, d=1.0, e=2.6667 a=28.0, b=1.0, c=2, d=1.0, e=2.6667 a=28.0, b=1.0, c=2.6667, d=1.0, e=2.6667 a=28.0, b=1.0, c=2.6667, d=10.882, e=1.0, f=2.6667 a=28.0, b=1.0, c=2.6667, d=1.0, e=2.6667 a=28.0, b=1.0, c=1.0, d=74.667 a=28.0, b=1.0, c=28.0, d=1.0, e=74.667 a=28.0, b=4.7585, c=1.0, d=2.6667 a=1.0, b=28.0, c=28.0, d=2, e=2.6667 a=1.0, b=1.0, c=7.6203, d=34.054 a=28.0, b=1.0, c=1.0, d=2.6667 a=28.0, b=1.0, c=1.0, d=2.6667, e=1.0, f=2.6667 a=1.0, b=2, c=28.0, d=1.0, e=28.0, f=2, g=2.6667 a=1.0, b=2, c=28.0, d=1.0, e=28.0, f=2.6667 a=1.0, b=2, c=28.0, d=1.0, e=28.0, f=2, g=2.6667 a=28.0, b=1.0, c=2.6667, d=28.0, e=2, f=2.6667 0.9970 1.3102 1.6529 1.1940 1.3731 0.9988 0.8833 0.9605 0.9951 0.9987 0.8842 1.6428 1.5952 0.9861 1.1573 0.4993 0.9760 0.9983 2.4512 1.0945 2.3313 1.1666 Chaos, Solitons & Fractals ( ) LorYZ29 a x − b sin c x d − e x z xy − fz LorYZ30 a x − b sin( c y ) − d x z xy − ez LorYZ31 sin( z ) − a y − b x z cxy − d LorYZ32 a x − b sin( c z ) − d sin( x ) − e x z xy − fz LorYZ33 a x + sin( y + b ) − c sin( d x ) − e x z xy − fz LorYZ34 a x + sin( y + b ) − c sin( d z ) − e x z xf −gz LorYZ35 a x + sin( b z ) − c sin( d x ) − e x z xy − fz LorYZ36 a x + sin( y + b ) − c sin( y ) − d x z xy − ez LorYZ37 a x + sin( y + b ) − c x z + d xy − ez LorYZ38 a x + sin( x + y ) − b x z + c xy − dz LorYZ39 a x + sin( y + b ) + sin( c y ) − d x z xy − ez LorYZ40 a x + sin( y + b ) − c x z − d xy − ez LorYZ41 a x − b sin( y + z ) − c x z xy − dz LorYZ42 a x − b sin( y + z ) − c x z y( d x + y z ) − e z LorYZ43 a x − b sin( sin( x ) ) − c x z xy − dz LorYZ44 a x − b sin( x + x y ) − c x z xy − dz LorYZ45 a x − b sin( x + y ) − c x z y( d x + y z ) − e z LorYZ46 a x − b sin( x + c ) − d x z exy − fz LorYZ47 a x − b sin( x z sin( z ) ) − c x z d xy − ez LorYZ48 a x − b sin z c sin( z ) − d x z e xy − fz LorYZ49 a x − b sin( z ) − c x z d xy − ez LorYZ50 a x − b sin( z ) − c x z + d xy − ez LorYZ51 a x + sin( x y ) − b x z − c xy − dz LorYZ52 sin( x − a y ) − b y − c x z xy − d LorYZ53 a x − b sin( y ) − c x z − d xy − ez LorYZ54 a x + sin( x ) − b x z xy − cz LorYZ55 a x − b sin( y ) − c x z xy − dz LorYZ56 −a x ( z − b ) x ( x + sin( y ) ) − c z LorYZ57 −a x ( z − b ) x ( c x + sin( y ) ) − d z ( ) 14 a=28.0, b=1.0, c=2.6667, d=2, e=1.0, f=2.6667 a=28.0, b=1.0, c=2.6667, d=1.0, e=2.6667 a=2.0, b=1.0, c=2.6667, d=74.667 a=28.0, b=1.0, c=2.6667, d=4.6373, e=1.0, f=2.6667 a=28.0, b=8.0335, c=1.0, d=28.0, e=1.0, f=2.6667 a=28.0, b=8.0335, c=1.0, d=28.0, e=1.0, f=2, g=2.6667 a=28.0, b=2.6667, c=1.0, d=28.0, e=1.0, f=2.6667 a=28.0, b=8.0335, c=1.0, d=1.0, e=2.6667 a=28.0, b=8.0335, c=1.0, d=0.7661, e=2.6667 a=28.0, b=1.0, c=0.7661, d=2.6667 a=28.0, b=8.0335, c=2.9484, d=1.0, e=2.6667 a=28.0, b=8.0335, c=1.0, d=0.9016, e=2.6667 a=28.0, b=1.0, c=1.0, d=2.6667 a=28.0, b=1.0, c=1.0, d=28.0, e=2.6667 a=28.0, b=1.0, c=1.0, d=2.6667 a=28.0, b=1.0, c=1.0, d=2.6667 a=28.0, b=1.0, c=1.0, d=28.0, e=2.6667 a=28.0, b=1.0, c=28.0, d=1.0, e=56.0, f=2.6667 a=28.0, b=1.0, c=1.0, d=2, e=2.6667 a=28.0, b=1.0, c=2, d=1.0, e=2, f=2.6667 a=29.0, b=1.0, c=1.0, d=2, e=2.6667 a=28.0, b=1.0, c=1.0, d=2.174, e=2.6667 a=28.0, b=1.0, c=0.065552, d=2.6667 a=1.0, b=1.0, c=1.0, d=74.667 a=28.0, b=1.0, c=1.0, d=8.396, e=2.6667 a=27.0, b=1.0, c=2.6667 a=28.0, b=1.0, c=1.0, d=1.8597 0.9423 0.9579 1.1891 0.9667 0.9360 1.2302 0.9479 0.9827 0.9869 0.9972 0.9520 0.9875 0.9955 1.0601 0.9962 0.9671 0.9556 1.0033 1.1070 1.0440 1.1915 0.9979 0.9698 1.2824 0.9629 0.9779 0.8615 a=1.0, b=28.0, c=2.6667 1.2818 a=1.0, b=28.0, c=28.0, d=2.6667 1.3103 Chaos, Solitons & Fractals LorYZ58 −a x ( z − b ) x ( y + sin( y ) ) − c z Figure 8: Phase space dynamics of LorYZ19 (gaming console) Figure 9: Phase space dynamics of LorYZ22 (honey pot) 15 a=1.0, b=28.0, c=2.6667 0.9897 Chaos, Solitons & Fractals Figure 10: Phase space dynamics of LorYZ27 (Batman’s glasses) Figure 11: Phase space dynamics of LorYZ47 (snake’s hood) 16 Chaos, Solitons & Fractals In the LorYZ family, the obtained LLE could be as low as 0.49 (LorYZ19) and as high as 2.45 (LorYZ22). In Figure 8 for the LorYZ19 system, the phase space dynamics takes a shape similar to a ‘gaming console’. It has an LLE of 0.49 which is a significant deviation from the Lorenz system. In Figure 9 for the LorYZ22, the LLE reaches 2.45 which is the maximum amongst all the attractors found through GP. Also, the phase space dynamics of the strong chaotic system takes a shape similar to a ‘honey pot’ (in the y-z plane). Figure 10 for the LorYZ27 looks like ‘Batman’s glasses’ (in x-z plane) which has a LLE slightly higher than one. Also, a ‘snake’s hood’ like phase space diagram in the y-z plane has been observed in Figure 11, for the LorYZ47 system. Table 5: List of generalized Lorenz y-z family of attractors Name x expression y expression z expression LorXYZ1 ay − bx c x − d sin( y z ) − esin( y + z ) − f x z xy − gz LorXYZ2 ay − bx sin( y ) − c y − d x z xy − e LorXYZ3 ay − bx c x + sin( z ) − d x z + e xy − fz LorXYZ4 ay − bx c x + sin( z ) − d x z + e xf −gz LorXYZ5 ay − bx c x + sin( y ) − d x z − e xy − fz LorXYZ6 ay − bx c x + sin( y ) − d x z − e xf −gz LorXYZ7 ay − bx c x − d sin( y + e ) − f x z xy − gz LorXYZ8 ay − bx c x − d sin( e y + f ) − g x z xy − hz LorXYZ9 ay − bx c x − d sin( y ) − e x z xy − fz LorXYZ10 ay − bx c x − d sin( x ) − e x z fx − gz + xy LorXYZ11 ay − bx c x − d sin( z − e ) − f x z x( x + g ) − h z LorXYZ12 ay − bx c x − d sin( x − e ) − f x z x( x + g ) − h z LorXYZ13 ay − bx c x − d sin( x + y ) − e x z x( x + f ) − g z 17 Coefficient values a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=1.0, g=2.6667 a=10.0, b=10.0, c=1.0, d=1.0, e=74.667 a=10.0, b=10.0, c=29.0, d=1.0, e=10.434, f=2.6667 a=10.0, b=10.0, c=29.0, d=1.0, e=10.434, f=2, g=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=7.5891, f=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=7.5891, f=2, g=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=28.0, f=1.0, g=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=10.0, f=28.0, g=1.0, h=2.6667 a=10.0, b=10.0, c=29.0, d=1.0, e=1.0, f=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, f=2.6667, g=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=2.6674, f=1.0, g=1.2624, h=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=2.6674, f=1.0, g=1.2624, h=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=1.0, LLE 0.0177 1.4151 0.9821 1.3045 0.9625 0.1414 0.9749 0.8807 1.0127 0.9462 1.2253 1.1576 1.2002 Chaos, Solitons & Fractals f=1.2624, g=2.6667 ay − bx + c d x − esin( x − f ) − g x z h xi − j z z + sin y a z + sin( x y )( sin( y ) − b ) x ( sin( y ) − c ) LorXYZ16 ( )( z + sin( y ) ) z + sin( y )( z + sin( y ) ) a z + sin( y z )( sin( y ) − b ) x ( sin( y ) − c ) LorXYZ17 a y − b x − c sin( y ) −d x ( z − e ) fxy − gz LorXYZ18 ay − bx c x − d sin( z − e ) − f x z xy − gz LorXYZ19 a y sin( y + b ) c − dx ey − fx LorXYZ14 LorXYZ15 a=10.0, b=10.0, c=26.674, d=28.0, e=1.0, f=2.6674, g=1.0, h=2.0, i=2, j=2.6667 a=2, b=0.82942, c=7.178 a=2, b=0.82942, c=7.178 a=10.0, b=10.0, c=10.0, d=1.0, e=27.0, f=5.6746, g=2.6667 a=10.0, b=10.0, c=28.0, d=1.0, e=4.2886, f=1.0, g=2.6667 a=0.76197, b=9.9326, c=1.5855, d=8.466, e=0.99989, f=9.465 Figure 12: Phase space dynamics of LorXYZ15 (lasso). For the MGGP based evolution of all the three state variables, the chaotic dynamical systems with sine terms are reported in Table 5. It is interesting to note that all of them have a simple straight line state trajectory in the x-y plane. For the LorXYZ family also, high LLE values like 1.3 (LorXYZ4) is obtained. Here we report some interesting patterns whose LLEs are relatively less than the Lorenz system. Drawing comparison of the phase space dynamics with physical objects, we can identify several interesting attractors like the ‘lasso’ in LorXYZ15, ‘gramophone record’ in LorXYZ16, 18 1.2119 0.0005 0.0170 0.0482 0.9970 0.0026 Chaos, Solitons & Fractals ‘wrinkled butterfly’ in LorXYZ17, ‘sieve’ in LorXYZ19 etc. LorXYZ19 has a low LLE of 0.0026 and the evidence of chaos is even weaker in the LorXYZ15 which needs to be considered with caution and further investigation with regards to its dynamical behaviour needs to be done. Figure 13: Phase space dynamics of LorXYZ16 (gramophone record) 19 Chaos, Solitons & Fractals Figure 14: Phase space dynamics of LorXYZ17 (wrinkled butterfly) Figure 15: Phase space dynamics of LorXYZ19 (sieve) 4. Discussions 20 Chaos, Solitons & Fractals We ran the GP algorithm for the cases of individual state variables as well, i.e. where the expressions of two of the variables are fixed (e.g. x and y states are fixed) and the expression for the remaining state variable (e.g. z) is evolved through GP. However, we obtained very few new chaotic attractors for these cases. Most of the obtained expressions were slight variation from the nominal case of the Lorenz system and therefore we did not report them. One of the reasons for this can be that since the dynamical system is constrained by the other two expressions, the search space is relatively limited and there is not much potential for the GP algorithm to explore and evolve new chaotic structures. For the case of two and three variables, the search space is very large and the GP was not able to find good candidate solutions which exhibit chaos. Therefore we gave the Lorenz attractor as an initial guess value to the GP algorithm. This ensured that the GP found a good starting solution and in subsequent generations it was able to evolve good chaotic attractors. The two variable cases seems to be the most conducive for evolving chaotic attractors and we obtained many new chaotic structures for these cases. However for the two variable cases, there is a significant difference in the number of chaotic attractors produced. For example, the x-y family has much more attractors than y-z or x-z, even though all are run for 100 times each. This implies that some structures or expressions are more favourable in obtaining chaotic attractors than others. This is an interesting insight and can be leveraged to evolve more chaotic attractors for a particular family by putting them as an initial guess, in the first iteration of the GP algorithm. Many other interesting limit cycles and oscillators were evolved during the runs which are interesting in their own right, but are not reported here as they do not exhibit chaos. It would be interesting to explore chaotic attractors other than the Lorenz system to get new and richer dynamics. In some cases, the numerically calculated Lyapunov exponent is positive but on screening them manually, by looking at the phase portraits, we found several of them were complicated limit cycles. Therefore, after the machine generated a list of possible new chaotic attractor candidates, a manual screening was necessary to verify whether the expressions were actually chaotic or not. We found that in most cases if the LLE (with the time delay embedding method) was less than 0.1, the evolved dynamical systems were not chaotic. This can be mostly attributed to numerical issues, since the analysed time series was of considerably short duration (25 seconds) and also we are essentially trying to infer the chaotic properties of a three dimensional system from the observation of just the first state variable. The time delay reconstruction method for estimating LLE does not give the actual Lyapunov exponent as the delay embedding is on one of the states and not on all of them. Therefore we used LLE = 0.1 with the delay embedding method within the GP cost function, as a threshold for a first hand screening of the expressions before looking at the phase portraits. We also report here, only those expressions which has at least one sine term in it. We observed that the other expressions which only involved the terms +, -, ×, ÷, without the sine term, were mostly variants of the Lorenz system with almost similar time domain dynamics. After screening with the evolved attractors with LLE > 0.1 in the time delay embedding technique, their true LLE was again computed directly from the structure of the differential equations using symbolic differentiation and are reported in Table 1-4. There are a few research results on the increase in the number of equilibrium points in the phase space due to the sine terms which is commonly known as the ‘Labyrinth chaos’ [31], [32]. Similar complex dynamics have also been observed in the phase space due to increase in the number of equilibrium points beside the standard two wing Lorenz like attractor dynamics. Since the reported chaotic attractors have been obtained by keeping a history list of all those which were evolved during one GP run, some of them have nearly similar structure with slight modification in some of the terms. This is due to the fact that once the GP algorithm finds a good solution with a 21 Chaos, Solitons & Fractals high value of Lyapunov exponent, it gives more importance to this structure and tries to find solutions in the vicinity of the obtained one. Also, we used only a limited set of mathematical functions for evolving the chaotic attractors, i.e. only +, -, ×, ÷, sine. It is possible to use other mathematical functions like cosine, tangent, hyperbolic sine, hyperbolic tangent, signum function, exponential, logarithm etc. to obtain a diverse variety of chaotic attractors exhibiting rich dynamics. This might result in more complex expressions and can be pursued as scope of future work. Since our proposed methodology mostly relies on a numerical technique, one of its drawbacks is that it reports only a limited set of parameters for a chaotic structure. To make each structure truly a generalised one, each of the individual parameter sets needs to be varied and detailed bifurcation analysis needs to be done to obtain the properties of each dynamical system. This would give detailed results like the range of parameters that results in chaotic behaviour for each system etc. Though this kind of analysis is important and interesting in its own right, it is a separate stream of work and is therefore not included in the present paper. 5. Conclusions We proposed a GP based intelligent technique to evolve new chaotic dynamical systems based on the original Lorenz system of equations. We reported over one hundred new chaotic attractors to validate our proposed methodology. The GP algorithms employed a strategy of maximising the largest Lyapunov exponent to evolve these new attractors. The GP runs also simultaneously evolved new types of limit cycles and oscillators but they were not reported as they did not exhibit chaotic dynamics. The study also showed that some of the structures of the attractors (like the x-y family) were favourable in terms of producing a larger number of attractors. For most of the newly found chaotic attractors, we obtained phase space dynamics which are qualitatively similar in nature to that of the original Lorenz attractors. However some of them resulted in completely new kind of dynamics which could not be apprehended before. GP being a numerical algorithm reports only one set of parameters for which the system of equations exhibit chaos. If the numerical parameters for each family of attractors are varied, a different type of chaotic dynamics might be obtained or the chaos might disappear altogether. This needs further investigation. We believe that such symbiotic confluence of machine intelligence and classical dynamical systems theory, would play a big role in investigating complexity in nature, in the near future. One stream of extension of this work can look at mathematical analysis of such structures to find out properties of the dynamical system, carry out bifurcation analysis etc. Another extension of this work can investigate more new chaotic attractors which are evolved from other established chaotic systems like the unified chaotic systems, chaotic neural networks etc. References [1] E. N. Lorenz, “Deterministic nonperiodic flow,” Journal of the atmospheric sciences, vol. 20, no. 2, pp. 130– 141, 1963. [2] I. Stewart, “Mathematics: The Lorenz attractor exists,” Nature, vol. 406, no. 6799, pp. 948–949, 2000. [3] J. Lü, G. Chen, and D. Cheng, “A new chaotic system and beyond: the generalized Lorenz-like system,” International Journal of Bifurcation and Chaos, vol. 14, no. 05, pp. 1507–1537, 2004. [4] S. Yu, J. Lu, X. Yu, and G. Chen, “Design and implementation of grid multiwing hyperchaotic Lorenz system family via switching control and constructing super-heteroclinic loops,” Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. 59, no. 5, pp. 1015–1028, 2012. 22 Chaos, Solitons & Fractals [5] J. Lü and G. Chen, “Generating multiscroll chaotic attractors: theories, methods and applications,” International Journal of Bifurcation and Chaos, vol. 16, no. 04, pp. 775–858, 2006. [6] S. Yu, W. K. Tang, J. Lü, and G. Chen, “Generating 2n-wing attractors from Lorenz-like systems,” International Journal of Circuit Theory and Applications, vol. 38, no. 3, pp. 243–258, 2010. [7] A. Abel and W. Schwarz, “Chaos communications-principles, schemes, and system analysis,” Proceedings of the IEEE, vol. 90, no. 5, pp. 691–710, 2002. [8] C. Letellier, Chaos in nature, vol. 81. World Scientific, 2013. [9] B. Jovic, Synchronization Techniques for Chaotic Communication Systems. Springer, 2011. [10] M. Small, Applied nonlinear time series analysis: applications in physics, physiology and finance, vol. 52. World Scientific, 2005. [11] H. Kantz and T. Schreiber, Nonlinear time series analysis, vol. 7. Cambridge university press, 2004. [12] P. Zhou and F. Yang, “Hyperchaos, chaos, and horseshoe in a 4D nonlinear system with an infinite number of equilibrium points,” Nonlinear Dynamics, vol. 76, no. 1, pp. 473–480, 2014. [13] P. Zhou, K. Huang, and C. Yang, “A fractional-order chaotic system with an infinite number of equilibrium points,” Discrete Dynamics in Nature and Society, vol. 2013, 2013. [14] K. Rodriguez-Vázquez and P. J. Fleming, “Evolution of mathematical models of chaotic systems based on multiobjective genetic programming,” Knowledge and Information Systems, vol. 8, no. 2, pp. 235–256, 2005. [15] K. Rodriguez-Vazquez and P. J. Fleming, “Genetic programming for dynamic chaotic systems modelling,” in Evolutionary Computation, 1999. CEC 99. Proceedings of the 1999 Congress on, vol. 1, 1999. [16] I. Zelinka, S. Celikovsky, H. Richter, and G. Chen, Evolutionary algorithms and chaotic systems, vol. 267. Springer, 2010. [17] I. Zelinka, G. Chen, and S. Celikovsky, “Chaos synthesis by means of evolutionary algorithms,” International Journal of Bifurcation and Chaos, vol. 18, no. 04, pp. 911–942, 2008. [18] I. Zelinka, D. Davendra, R. Senkerik, R. Jasek, and Z. Oplatkova, “Analytical programming-a novel approach for evolutionary synthesis of symbolic structures,” Evolutionary Algorithms. InTech, 2011. [19] I. Zelinka, M. Chadli, D. Davendra, R. Senkerik, and R. Jasek, “An investigation on evolutionary reconstruction of continuous chaotic systems,” Mathematical and Computer Modelling, vol. 57, no. 1, pp. 2– 15, 2013. [20] W. K. Tang, G. Zhong, G. Chen, and K. Man, “Generation of n-scroll attractors via sine function,” Circuits and Systems I: Fundamental Theory and Applications, IEEE Transactions on, vol. 48, no. 11, pp. 1369–1372, 2001. [21] J. R. Koza, F. H. Bennett III, and O. Stiffelman, “Genetic programming as a Darwinian invention machine,” in Genetic Programming, Springer, 1999, pp. 93–108. [22] J. R. Koza, Genetic programming III: Darwinian invention and problem solving, vol. 3. Morgan Kaufmann, 1999. [23] J. R. Koza, Genetic programming: on the programming of computers by means of natural selection, vol. 1. MIT press, 1992. [24] D. P. Searson, D. E. Leahy, and M. J. Willis, “GPTIPS: an open source genetic programming toolbox for multigene symbolic regression,” International MultiConference of Engineers and Computer Scientists 2010, 2010. [25] I. Pan, D. S. Pandey, and S. Das, “Global solar irradiation prediction using a multi-gene genetic programming approach,” Journal of Renewable and Sustainable Energy, vol. 5, no. 6, p. 063129, 2013. 23 Chaos, Solitons & Fractals [26] A. Wolf, J. B. Swift, H. L. Swinney, and J. A. Vastano, “Determining Lyapunov exponents from a time series,” Physica D: Nonlinear Phenomena, vol. 16, no. 3, pp. 285–317, 1985. [27] M. T. Rosenstein, J. J. Collins, and C. J. De Luca, “A practical method for calculating largest Lyapunov exponents from small data sets,” Physica D: Nonlinear Phenomena, vol. 65, no. 1, pp. 117–134, 1993. [28] M. Banbrook, G. Ushaw, and S. McLaughlin, “How to extract Lyapunov exponents from short and noisy time series,” Signal Processing, IEEE Transactions on, vol. 45, no. 5, pp. 1378–1382, 1997. [29] K. Ramasubramanian and M. Sriram, “A comparative study of computation of Lyapunov spectra with different algorithms,” Physica D: Nonlinear Phenomena, vol. 139, no. 1, pp. 72–86, 2000. [30] S. Heiligenthal et al., “Strong and weak chaos in nonlinear networks with time-delayed couplings,” Physical review letters, vol. 107, no. 23, p. 234102, 2011. [31] J. C. Sprott, Elegant chaos: algebraically simple chaotic flows. World Scientific, 2010. [32] J. C. Sprott and K. E. Chlouverakis, “Labyrinth chaos,” International Journal of Bifurcation and Chaos, vol. 17, no. 06, pp. 2097–2108, 2007. 24 Supplementary Material a) Phase portraits of generalized Lorenz x-z family of attractors Figure 1: Phase space dynamics of LorXZ1 Figure 2: Phase space dynamics of LorXZ2 1 Figure 3: Phase space dynamics of LorXZ4 Figure 4: Phase space dynamics of LorXZ6 2 Figure 5: Phase space dynamics of LorXZ7 Figure 6: Phase space dynamics of LorXZ8 3 Figure 7: Phase space dynamics of LorXZ10 Figure 8: Phase space dynamics of LorXZ11 4 Figure 9: Phase space dynamics of LorXZ12 Figure 10: Phase space dynamics of LorXZ13 5 Figure 11: Phase space dynamics of LorXZ14 Figure 12: Phase space dynamics of LorXZ16 6 b) Phase portraits of generalized Lorenz x-y family of attractors Figure 13: Phase space dynamics of LorXY1 Figure 14: Phase space dynamics of LorXY2 7 Figure 15: Phase space dynamics of LorXY3 Figure 16: Phase space dynamics of LorXY4 8 Figure 17: Phase space dynamics of LorXY5 Figure 18: Phase space dynamics of LorXY6 9 Figure 19: Phase space dynamics of LorXY7 Figure 20: Phase space dynamics of LorXY8 10 Figure 21: Phase space dynamics of LorXY9 Figure 22: Phase space dynamics of LorXY11 11 Figure 23: Phase space dynamics of LorXY12 Figure 24: Phase space dynamics of LorXY13 12 Figure 25: Phase space dynamics of LorXY14 Figure 26: Phase space dynamics of LorXY15 13 Figure 27: Phase space dynamics of LorXY16 Figure 28: Phase space dynamics of LorXY17 14 Figure 29: Phase space dynamics of LorXY18 Figure 30: Phase space dynamics of LorXY19 15 Figure 31: Phase space dynamics of LorXY20 Figure 32: Phase space dynamics of LorXY22 16 Figure 33: Phase space dynamics of LorXY23 Figure 34: Phase space dynamics of LorXY24 17 Figure 35: Phase space dynamics of LorXY25 Figure 36: Phase space dynamics of LorXY26 18 Figure 37: Phase space dynamics of LorXY27 Figure 38: Phase space dynamics of LorXY28 19 Figure 39: Phase space dynamics of LorXY29 Figure 40: Phase space dynamics of LorXY30 20 Figure 41: Phase space dynamics of LorXY31 Figure 42: Phase space dynamics of LorXY32 21 Figure 43: Phase space dynamics of LorXY33 Figure 44: Phase space dynamics of LorXY34 22 c) Phase portraits of generalized Lorenz y-z family of attractors Figure 45: Phase space dynamics of LorYZ1 Figure 46: Phase space dynamics of LorYZ2 23 Figure 47: Phase space dynamics of LorYZ3 Figure 48: Phase space dynamics of LorYZ4 24 Figure 49: Phase space dynamics of LorYZ5 Figure 50: Phase space dynamics of LorYZ6 25 Figure 51: Phase space dynamics of LorYZ7 Figure 52: Phase space dynamics of LorYZ8 26 Figure 53: Phase space dynamics of LorYZ9 Figure 54: Phase space dynamics of LorYZ10 27 Figure 55: Phase space dynamics of LorYZ11 Figure 56: Phase space dynamics of LorYZ12 28 Figure 57: Phase space dynamics of LorYZ13 Figure 58: Phase space dynamics of LorYZ14 29 Figure 59: Phase space dynamics of LorYZ15 Figure 60: Phase space dynamics of LorYZ16 30 Figure 61: Phase space dynamics of LorYZ17 Figure 62: Phase space dynamics of LorYZ18 31 Figure 63: Phase space dynamics of LorYZ20 Figure 64: Phase space dynamics of LorYZ21 32 Figure 65: Phase space dynamics of LorYZ23 Figure 66: Phase space dynamics of LorYZ24 33 Figure 67: Phase space dynamics of LorYZ25 Figure 68: Phase space dynamics of LorYZ26 34 Figure 69: Phase space dynamics of LorYZ28 Figure 70: Phase space dynamics of LorYZ29 35 Figure 71: Phase space dynamics of LorYZ30 Figure 72: Phase space dynamics of LorYZ31 36 Figure 73: Phase space dynamics of LorYZ32 Figure 74: Phase space dynamics of LorYZ33 37 Figure 75: Phase space dynamics of LorYZ34 Figure 76: Phase space dynamics of LorYZ35 38 Figure 77: Phase space dynamics of LorYZ36 Figure 78: Phase space dynamics of LorYZ37 39 Figure 79: Phase space dynamics of LorYZ38 Figure 80: Phase space dynamics of LorYZ39 40 Figure 81: Phase space dynamics of LorYZ40 Figure 82: Phase space dynamics of LorYZ41 41 Figure 83: Phase space dynamics of LorYZ42 Figure 84: Phase space dynamics of LorYZ43 42 Figure 85: Phase space dynamics of LorYZ44 Figure 86: Phase space dynamics of LorYZ45 43 Figure 87: Phase space dynamics of LorYZ46 Figure 88: Phase space dynamics of LorYZ48 44 Figure 89: Phase space dynamics of LorYZ49 Figure 90: Phase space dynamics of LorYZ50 45 Figure 91: Phase space dynamics of LorYZ51 Figure 92: Phase space dynamics of LorYZ52 46 Figure 93: Phase space dynamics of LorYZ53 Figure 94: Phase space dynamics of LorYZ54 47 Figure 95: Phase space dynamics of LorYZ55 Figure 96: Phase space dynamics of LorYZ56 48 Figure 97: Phase space dynamics of LorYZ57 Figure 98: Phase space dynamics of LorYZ58 49 d) Phase portraits of generalized Lorenz x-y-z family of attractors Figure 99: Phase space dynamics of LorXYZ1 Figure 100: Phase space dynamics of LorXYZ2 50 Figure 101: Phase space dynamics of LorXYZ3 Figure 102: Phase space dynamics of LorXYZ4 51 Figure 103: Phase space dynamics of LorXYZ5 Figure 104: Phase space dynamics of LorXYZ6 52 Figure 105: Phase space dynamics of LorXYZ7 Figure 106: Phase space dynamics of LorXYZ8 53 Figure 107: Phase space dynamics of LorXYZ9 Figure 108: Phase space dynamics of LorXYZ10 54 Figure 109: Phase space dynamics of LorXYZ11 Figure 110: Phase space dynamics of LorXYZ12 55 Figure 111: Phase space dynamics of LorXYZ13 Figure 112: Phase space dynamics of LorXYZ14 56 Figure 113: Phase space dynamics of LorXYZ18 57
9cs.NE
On Colouring (2P2 , H)-Free and (P5 , H)-Free Graphs⋆ Konrad K. Dabrowski and Daniël Paulusma arXiv:1712.02447v1 [cs.CC] 6 Dec 2017 Department of Computer Science, Durham University, Science Laboratories, South Road, Durham DH1 3LE, United Kingdom {konrad.dabrowski,daniel.paulusma}@durham.ac.uk Abstract. The Colouring problem asks whether the vertices of a graph can be coloured with at most k colours for a given integer k in such a way that no two adjacent vertices receive the same colour. A graph is (H1 , H2 )-free if it has no induced subgraph isomorphic to H1 or H2 . A connected graph H1 is almost classified if Colouring on (H1 , H2 )-free graphs is known to be polynomial-time solvable or NP-complete for all but finitely many connected graphs H2 . We show that every connected graph H1 apart from the claw K1,3 and the 5-vertex path P5 is almost classified. We also prove a number of new hardness results for Colouring on (2P2 , H)-free graphs. This enables us to list all graphs H for which the complexity of Colouring is open on (2P2 , H)-free graphs and all graphs H for which the complexity of Colouring is open on (P5 , H)-free graphs. In fact we show that these two lists coincide. Moreover, we show that the complexities of Colouring for (2P2 , H)-free graphs and for (P5 , H)-free graphs are the same for all known cases. 1 Introduction Graph colouring is an extensively studied concept in both Computer Science and Mathematics due to its many application areas. A k-colouring of a graph G = (V, E) is a mapping c : V → {1, . . . , k} such that c(u) 6= c(v) whenever uv ∈ E. The Colouring problem that of deciding whether a given graph G has a k-colouring for a given integer k. If k is fixed, then we write k-Colouring instead. It is well known that even 3-Colouring is NP-complete [22]. Due to the computational hardness of Colouring, it is natural to restrict the input to special graph classes. A class is hereditary if it closed under vertex deletion. Hereditary graph classes form a large collection of well-known graph classes for which the Colouring problem has been extensively studied. A classical result in the area is due to Grötschel, Lovász, and Schrijver [14], who showed that Colouring is polynomial-time solvable for perfect graphs. Graphs with no induced subgraph isomorphic to a graph in a set H are said to be H-free. It is readily seen that a graph class G is hereditary if and only if it there exists a set H such that every graph in G is H-free. If the graphs of H are required to be minimal under taking induced subgraphs, then H is unique. For example, the set H of minimal forbidden induced subgraphs for the class of perfect graphs consists of all odd holes and odd antiholes [6]. Král’, Kratochvíl, Tuza, and Woeginger [21] classified the complexity of Colouring for the case where H consists of a single graph H. They proved that Colouring on H-free graphs is polynomial-time solvable if H is an induced subgraph of P4 or P1 + P3 and NP-complete otherwise.1 Král’ et al. [21] also initiated a complexity study of Colouring for graph classes defined by two forbidden induced subgraphs H1 and H2 . Such graph classes are said to be bigenic. For bigenic graph classes, no dichotomy is known or even conjectured, despite many results [1,2,4,5,7,8,10,15,16,18,20,21,23,26,27,28,31]. For instance, if we forbid two graphs H1 and H2 with |V (H1 )| ≤ 4 and |V (H2 )| ≤ 4, then there are three open cases left, namely when (H1 , H2 ) ∈ {(K1,3 , 4P1 ), (K1,3 , 2P1 + P2 ), (C4 , 4P1 )} (see [23] and Fig. 1). If H1 and H2 are connected with |V (H1 )| ≤ 5 and |V (H2 )| ≤ 5, then there are four open cases left, namely when H1 = P5 and H2 ∈ {C3 + 2P1 , C3 + P2 , P1 + 2P2 } (see [20] and Fig. 2) and when H1 = K1,3 and ⋆ 1 This paper was supported by EPSRC (EP/K025090/1) and the Leverhulme Trust (RPG-2016-258). We refer to Section 2 for notation used throughout Section 1. H2 = C4 + P1 (see [28] and Fig. 2). To give another example, Blanché et al. [1] determined the complexity of Colouring for (H, H)-free graphs for every graph H except when H = P3 + sP1 for s ≥ 3 or H = P4 + sP1 for s ≥ 2. K1,3 4P1 C4 2P1 + P2 Fig. 1. The graphs from the three pairs (H1 , H2 ) ∈ {(K1,3 , 4P1 ), (K1,3 , 2P1 + P2 ), (C4 , 4P1 )} of graphs on at most four vertices, for which the complexity of Colouring on (H1 , H2 )-free graphs is still open. K1,3 C4 + P1 C3 + 2P1 P5 C3 + P2 P1 + 2P2 Fig. 2. The graphs from the four pairs (H1 , H2 ) ∈ {(K1,3 , C4 + P1 ), (P5 , C3 + 2P1 ), (P5 , C3 + P2 ), (P5 , P1 + 2P2 )} of connected graphs on at most five vertices, for which the complexity of Colouring on (H1 , H2 )-free graphs is still open. The related problems Precolouring Extension and List Colouring have also been studied for bigenic graph classes. For the first problem, we are given a graph G, an integer k and a k-colouring c′ defined on an induced subgraph of G. The question is whether G has a k-colouring c extending c′ . For the second problem, each vertex u of the input graph G has a list L(u) of colours. Here the question is whether G has a colouring c that respects L, that is, with c(u) ∈ L(u) for all u ∈ V (G). For the Precolouring Extension problem no classification is known and we refer to the survey [12] for an overview on what is known. In contrast to the incomplete classifications for Colouring and Precolouring Extension, Golovach and Paulusma [13] showed a dichotomy for the complexity of List Colouring on bigenic graph classes. Our Approach. To get a handle on the computational complexity classification of Colouring for bigenic graph classes, we continue the line of research in [2,16,20,26,27,28] by considering pairs (H1 , H2 ), where H1 and H2 are both connected. We introduce the following notion. We say that a connected graph H1 is almost classified if Colouring on (H1 , H2 )-free graphs is known to be either polynomial-time solvable or NP-complete for all but finitely many connected graphs H2 . This leads to the following research question: Which connected graphs are almost classified? Our Results. In Section 3 we show, by combining known results from the literature, that every connected graph H1 apart from the claw K1,3 and the 5-vertex path P5 is almost classified. In fact we show that the number of pairs (H1 , H2 ) of connected graphs for which the complexity of Colouring is unknown is finite if neither H1 nor H2 is isomorphic to K1,3 or P5 . In Section 4 we prove a number of new hardness results for Colouring restricted to (2P2 , H2 )-free graphs (which form a subclass of (P5 , H2 )-free graphs). We do the latter by adapting the NP-hardness construction from [11] for List Colouring restricted to complete bipartite graphs. In Section 5, we first summarize our knowledge on the complexity of Colouring restricted to (2P2 , H)-free graphs and (P5 , H)-free graphs. Afterwards, we list all graphs H for which the complexity of 2 Colouring on (2P2 , H)-free graphs is still open, and all graphs H for which the complexity of Colouring on (P5 , H)-free graphs is still open. As it turns out, these two lists coincide. Moreover, the complexities of Colouring for (2P2 , H)-free graphs and for (P5 , H)-free graphs turn out to be the same for all cases that are known. 2 Preliminaries We consider only finite, undirected graphs without multiple edges or self-loops. Let G = (V, E) be a graph. The complement G of G is the graph with vertex set V (G) and an edge between two distinct vertices if and only if these two vertices are not adjacent in G. For a subset S ⊆ V , we let G[S] denote the subgraph of G induced by S, which has vertex set S and edge set {uv | u, v ∈ S, uv ∈ E}. Let {H1 , . . . , Hp } be a set of graphs. A graph G is (H1 , . . . , Hp )-free if G has no induced subgraph isomorphic to a graph in {H1 , . . . , Hp }. If p = 1, we may write H1 -free instead of (H1 )-free. The disjoint union G + H of two vertex-disjoint graphs G and H is the graph (V (G) ∪ V (H), E(G) ∪ E(H)). The disjoint union of r copies of a graph G is denoted by rG. A linear forest is the disjoint union of one or more paths. The graphs Cr , Kr and Pr denote the cycle, complete graph and path on r vertices, respectively. The graph K3 is also known as the triangle. The graph Kr,s denotes the complete bipartite graph with partition classes of size r and s, respectively. The graph K1,3 is also called the claw. The graph Sh,i,j , for 1 ≤ h ≤ i ≤ j, denotes the subdivided claw, that is, the tree that has only one vertex x of degree 3 and exactly three leaves, which are of distance h, i and j from x, respectively. Observe that S1,1,1 = K1,3 . The graph S1,1,2 is also known as the fork or the chair. The graph Th,i,j with 0 ≤ h ≤ i ≤ j denotes the graph with vertices a0 , . . . , ah , b0 , . . . , bi and c0 , . . . , cj and edges a0 b0 , b0 c0 , c0 a0 , ap ap+1 for p ∈ {0, . . . , h − 1}, bp bp+1 for p ∈ {0, . . . , i − 1} and cp cp+1 for p ∈ {0, . . . , j − 1}. Note that T0,0,0 = C3 . The graph T0,0,1 = P1 + P3 is known as the paw, the graph T0,1,1 as the bull, the graph T1,1,1 as the net, and the graph T0,0,2 is known as the hammer; see also Fig. 3. Also note that Th,i,j is the line graph of Sh+1,i+1,j+1 . T0,0,0 = C3 T0,0,1 = paw T0,1,1 = bull T1,1,1 = net T0,0,2 = hammer Fig. 3. Examples of Th,i,j graphs. Let T be the class of graphs every component of which is isomorphic to a graph Th,i,j for some 1 ≤ h ≤ i ≤ j or a path Pr for some r ≥ 1. The following result, which is due to Schindl and which we use in Section 5, shows that the Th,i,j graphs play an important role for our study. Theorem 1 ([31]). For p ≥ 1, let H1 , . . . , Hp be graphs whose complement is not in T . Then Colouring is NP-complete for (H1 , . . . , Hp )-free graphs. 3 Almost Classified Graphs In this section we prove the following result, from which it immediately follows that every connected graph apart from K1,3 and P5 is almost classified. In Section 5 we discuss why K1,3 and P5 are not almost classified. 3 Theorem 2. There are only finitely many pairs (H1 , H2 ) of connected graphs with {H1 , H2 }∩{K1,3 , P5 } = ∅, such that the complexity of Colouring on (H1 , H2 )-free graphs is unknown. Proof. We first make a useful observation. Let H be a tree that is not isomorphic to K1,3 or P5 and that is not an induced subgraph of P4 . If H contains a vertex of degree at least 4 then it contains an induced K1,4 . If H has maximum degree 3, then since H is connected and not isomorphic to K1,3 , it must contain an induced S1,1,2 . If H has maximum degree at most 2, then it is a path, and since it is not isomorphic to P5 and not an induced subgraph of P4 , it follows that H must be a path on at least six vertices. We conclude that if H is a tree that is not isomorphic to K1,3 or P5 and that is not an induced subgraph of P4 , then H contains K1,4 or S1,1,2 as an induced subgraph or H is a path on at least six vertices. Now let (H1 , H2 ) be a pair of connected graphs with {H1 , H2 } ∩ {K1,3 , P5 } = ∅. If H1 or H2 is an induced subgraph of P4 , then Colouring is polynomial-time solvable for (H1 , H2 )-free graphs, as Colouring is polynomial-time solvable for P4 -free graphs (see, for example, [21]). Hence we may assume that this is not the case. If H1 and H2 both contain at least one cycle [9] or both contain an induced K1,3 [17], then even 3-Colouring is NP-complete. Hence we may also assume that at least one of H1 , H2 is a tree and that at least one of H1 , H2 is a K1,3 -free graph. This leads, without loss of generality, to the following two cases. Case 1. H1 is a tree and K1,3 -free. Then H1 is a path. First suppose that H1 has at least 22 vertices. It is known that 4-Colouring is NPcomplete for (P22 , C3 )-free graphs [19] and that Colouring is NP-complete for (P9 , C4 )-free graphs [10] and for (2P2 , Cr )-free graphs for all r ≥ 5 [21]. Hence we may assume that H2 is a tree. By the observation at the start of the proof, this implies that H2 contains an induced K1,4 , S1,1,2 or P6 . Therefore H contains an induced 4P1 or 2P1 + P2 . Since H1 is a path on at least 22 vertices, H1 contains an induced 4P1 . As Colouring is NP-complete for (4P1 , 2P1 +P2 )-free graphs [21], Colouring is NP-complete for (H1 , H2 )-free graphs. Now suppose that H1 has at most 21 vertices. By the observation at the start of the proof, H1 is a path on at least six vertices. It is known that 5-Colouring is NP-complete for P6 -free graphs [18]. As K6 is not 5-colourable, this means that 5-Colouring is NP-complete for (P6 , K6 )-free graphs, as observed in [12]. Therefore we may assume that H2 is K6 -free. Recall that Colouring is NP-complete for (2P1 + P2 , 4P1 )-free graphs [21], which are contained in the class of (P6 , 4P1 )-free graphs. Therefore we may assume that H2 is 4P1 -free. Since H2 is (K6 , 4P1 )-free, Ramsey’s Theorem [29] implies that |V (H2 )| is bounded by a constant. We conclude that both H1 and H2 have size bounded by a constant. Case 2. H1 is a tree and not K1,3 -free, and H2 is K1,3 -free and not a tree. Then H1 contains a vertex of degree at least 3 and H2 contains an induced cycle Cr for some r ≥ 3. It is known that 3-Colouring is NP-complete for (K1,5 , C3 )-free graphs [25] and for (K1,3 , Cr )-free graphs whenever r ≥ 4 [21]. We may therefore assume that H1 is a tree of maximum degree at most 4 and that H2 contains at least one induced C3 but no induced cycles on more than three vertices. Recall that 4-Colouring is NP-complete for (P22 , C3 )-free graphs [19]. Hence we may assume that H1 is a P22 -free tree. As H1 has maximum degree at most 4, we find that H1 has a bounded number of vertices. By assumption, H1 contains a vertex of degree at least 3. As Colouring is NP-complete for (K1,3 , K4 )free graphs [21], we may assume that H2 is K4 -free. By the observation at the start of the proof, H1 must contain an induced K1,4 or S1,1,2 . Recall that Colouring is NP-complete for the class of (2P1 + P2 , 4P1 )-free graphs [21], which is contained in the class of (K1,4 , S1,1,2 , 4P1 )-free graphs. Hence we may assume that H2 is 4P1 -free. Since H2 is (K4 , 4P1 )-free, Ramsey’s Theorem [29] implies that |V (H2 )| is bounded by a constant. Again, we conclude that in this case both H1 and H2 have size bounded by a constant. ⊓ ⊔ Corollary 1. Every connected graph apart from K1,3 and P5 is almost classified. 4 Hardness Results In this section we prove that Colouring restricted to (2P2 , H)-free graphs is NP-complete for several graphs H. To prove our results we adapt a hardness construction from Golovach and Heggernes [11] for 4 proving that List Colouring is NP-complete for complete bipartite graphs. As observed in [13], a minor modification of this construction yields that List Colouring is NP-complete for complete split graphs, which are the graphs obtained from complete bipartite graphs by changing one of the bipartition classes into a clique. We first describe the construction of [11], which uses a reduction from the NP-complete [30] problem NotAll-Equal 3-Satisfiability with positive literals only. To define this problem, let X = {x1 , x2 , . . . , xn } be a set of logical variables, and let C = {C1 , C2 , . . . , Cm } be a set of 3-literal clauses over X in which all literals are positive and every literal appears at most once in each clause. The question is whether X has a truth assignment such that each clause in C contains at least one true literal and at least one false literal. If so, we say that such a truth assignment is satisfying. Let (X, C) be an instance of Not-All-Equal 3-Satisfiability with positive literals only. We construct an instance (G1 , L) of List Colouring as follows. For each xi we introduce a vertex, which we also denote by xi and which we say is of x-type. We define L(x1 ) = {1, 2}, L(x2 ) = {3, 4}, . . . , L(xn ) = {2n − 1, 2n}. In this way, each xi has one odd colour and one even colour in its list, and all lists L(xi ) are pairwise disjoint. For each Cj we introduce two vertices, which we denote by Cj and Cj′ and which we say are of C-type. If Cj = {xg , xh , xi } with L(xg ) = {a, a + 1}, L(xh ) = {b, b + 1} and L(xi ) = {c, c + 1}, then we set L(Cj ) = {a, b, c} and L(Cj′ ) = {a + 1, b + 1, c + 1}. Hence each Cj has only odd colours in its list and each Cj′ has only even colours in its list. To obtain the graph G1 we add an edge between every vertex of x-type and every vertex of C-type. Note that G1 is a complete bipartite graph with bipartition classes {x1 , . . . , xn } and ′ {C1 , . . . , Cm } ∪ {C1′ , . . . , Cm }. We also construct an instance (G2 , L) where G2 is obtained from G1 by adding edges between every pair of vertices of x-type. Note that G2 is a complete split graph. The following lemma is straightforward. We refer to [11] for a proof for the case involving G1 . The case involving G2 follows from this proof and the fact that the lists L(xi ) are pairwise disjoint, as observed in [13]. Lemma 1 ([11]). (C, X) has a satisfying truth assignment if and only if G1 has a colouring that respects L if and only if G2 has a colouring that respects L. We now extend G1 and G2 into graphs G′1 and G′2 , respectively, by adding a clique K consisting of 2n new vertices k1 , . . . , k2n and by adding an edge between a vertex kℓ and a vertex u of the original graph if and only if ℓ ∈ / L(u). We say that the vertices k1 , . . . , k2n are of k-type. Lemma 2. (C, X) has a satisfying truth assignment if and only if G′1 has a 2n-colouring if and only if G′2 has a 2n-colouring. Proof. Let i ∈ {1, 2}. By Lemma 1, we only need to show that Gi has a colouring that respects L if and only if G′i has a 2n-colouring. First suppose that Gi has a colouring c that respects L. We extend c to a colouring c′ of G′i by setting c′ (kℓ ) = ℓ for ℓ ∈ {1, . . . , 2n}. Now suppose that G′i has a 2n-colouring c′ . As the k-type vertices form a clique, we may assume without loss of generality that c′ (kℓ ) = ℓ for ℓ ∈ {1, . . . , 2n}. Hence the restriction of c′ to Gi yields a colouring c that respects L. ⊓ ⊔ In the next two lemmas we show forbidden induced subgraphs in G′1 and G′2 , respectively. The complements of these forbidden graphs are shown in Figs. 4 and 5, respectively. 2P2 = C4 3P2 T0,2,2 Fig. 4. Graphs that are not induced subgraphs of the complement of G′1 . 5 Lemma 3. The graph G′1 is (2P2 , 3P2 , T0,2,2 )-free. Proof. We will prove that G′1 is (C4 , 3P2 , T0,2,2 )-free. Observe that in G′1 , the set of x-type vertices is a clique, the set of C-type vertices is a clique and the set of k-type vertices is an independent set. Furthermore, in G′1 , no x-type vertex is adjacent to a C-type vertex. C4 -freeness. For contradiction, suppose that G′1 contains an induced subgraph H isomorphic to C4 ; say the vertices of H are u1 , u2 , u3 , u4 in that order. As the union of the set of x-type and C-type vertices induces a P3 -free graph in G′1 , there must be at least two vertices of the C4 that are neither x-type nor C-type. Since the k-type vertices form an independent set, we may assume without loss of generality that u1 and u3 are of k-type. It follows that u2 and u4 cannot be of k-type. As the set of x-type vertices and the set of C-type vertices each from a clique in G′1 , but u2 is non-adjacent to u4 , we may assume without loss of generality that u2 is of x-type and u4 is of C-type. Then u4 is adjacent to the two k-type neighbours of an x-type vertex, which correspond to an even and odd colour. This is not possible as u4 , being a C-type vertex, is adjacent in G′1 to (exactly three) k-type vertices, which correspond either to even colours only or to odd colours only. We conclude that G′1 is C4 -free. 3P2 -freeness. For contradiction, suppose that G′1 contains an induced subgraph H isomorphic to 3P2 . As the C-type vertices and x-type vertices each form a clique in G′1 , one edge of H must consist of two k-type vertices. This is not possible, as k-type vertices form an independent set in G′1 . We conclude that G′1 is 3P2 -free. T0,2,2 -freeness. For contradiction, suppose that G′1 contains an induced subgraph H isomorphic to T0,2,2 with vertices a0 , a1 , a2 , b0 , b1 , b2 , c0 and edges a0 b0 , b0 c0 , c0 a0 , a0 a1 , a1 a2 , b0 b1 , b1 b2 . As the k-type vertices form an independent set in G′1 , at least one of a1 , a2 and at least one of b1 , b2 is of x-type or C-type. As the x-type vertices and the C-type vertices form cliques in G′1 , we may assume without loss of generality that at least one of a1 , a2 is of C-type and at least one of b1 , b2 is of x-type. As the C-type vertices and the x-type vertices each form a clique in G′1 , this means that c0 must be of k-type, a0 cannot be of x-type and b0 cannot be of C-type. As k-type vertices form an independent set in G′1 , a0 and b0 cannot be of k-type. Therefore a0 is of C-type and b0 is of x-type. This is a contradiction, as C-type vertices are non-adjacent to ⊓ ⊔ x-type vertices. We conclude that G′1 is T0,2,2 -free. 2P2 = C4 2C3 = K3,3 C3 + P4 2P4 T0,0,4 Fig. 5. Graphs that are not induced subgraphs of the complement of G′2 . Lemma 4. The graph G′2 is (2P2 , 2C3 , C3 + P4 , 2P4 , T0,0,4 )-free. Proof. We will prove that G′2 is (C4 , 2C3 , C3 + P4 , 2P4 , T0,0,4 )-free. Observe that in G′2 , the set of C-type vertices is a clique, the set of x-type vertices is an independent set and the set of k-type vertices is an independent set. Furthermore, in G′2 , no x-type vertex is adjacent to a C-type vertex and every x-type vertex has degree 2. In fact, the union of the set of x-type vertices and the k-type vertices induces a disjoint union of P3 s in G′2 . C4 -freeness. For contradiction, suppose that G′2 contains an induced subgraph H isomorphic to C4 , say the vertices of H are u1 , u2 , u3 , u4 in that order. As the union of the set of x-type and C-type vertices induces a P3 -free graph in G′2 , there must be at least two vertices of the C4 that are neither x-type nor C-type. Since 6 the k-type vertices form an independent set, we may assume without loss of generality that u1 and u3 are of k-type. It follows that u2 and u4 cannot be of k-type. As the set of vertices of C-type form a clique in G′2 , at least one of u2 , u4 , say u2 , is of x-type. If u4 is also of x-type, then u2 and u4 are x-type vertices with the same two colours in their list, namely those corresponding to u1 and u3 . This is not possible. Thus u4 must be of C-type. Then u4 is adjacent to the two k-type neighbours of an x-type vertex, which correspond to an even and odd colour. This is not possible as u4 , being a C-type vertex, is adjacent in G′2 to (exactly three) k-type vertices, which correspond either to even colours only or to odd colours only. We conclude that G′2 is C4 -free. (2C3 , C3 + P4 , 2P4 )-freeness. For contradiction, suppose that G′2 contains an induced subgraph H isomorphic to 2C3 , C3 + P4 or 2P4 . As the k-type and x-type vertices induce a disjoint union of P3 s in G′2 , both components of H must contain a C-type vertex. This is not possible, as C-type vertices form a clique in G′2 . We conclude that G′2 is (2C3 , C3 + P4 , 2P4 )-free. T0,0,4 -freeness. For contradiction, suppose that G′2 contains an induced subgraph H isomorphic to T0,0,4 with vertices a0 , a1 , a2 , a3 , a4 , b0 , c0 and edges a0 b0 , b0 c0 , c0 a0 , a0 a1 , a1 a2 , a2 a3 , a3 a4 . First suppose, that neither b0 nor c0 is of C-type. Since the union of the set of x-type vertices and the k-type vertices induces a disjoint union of P3 s in G′2 it follows that a0 is of C-type. Since b0 and c0 are not of C-type and no vertex of C-type has a neighbour of x-type in G′2 , it follows that b0 and c0 must be of k-type. This is not possible, because the k-type vertices form an independent set in G′2 . Now suppose that at least one of b0 , c0 is of C-type. Since the vertices of C-type induce a clique in G′2 , it follows that no vertex in A := {a1 , a2 , a3 , a4 } is of C-type. Since A induces a P4 in G′2 , but the union of the set of x-type vertices and the k-type vertices induces a disjoint union of P3 s in G′2 , this is a contradiction. We conclude that G′2 is T0,0,4 -free. ⊓ ⊔ We are now ready to state the two main results of this section. It is readily seen that Colouring belongs to NP. Then the first theorem follows from Lemma 2 combined with Lemma 3, whereas the second one follows from Lemma 2 combined with Lemma 4. Note that 2C3 is isomorphic to K3,3 . Theorem 3. Colouring is NP-complete for (2P2 , 3P2 , T0,2,2 )-free graphs. Theorem 4. Colouring is NP-complete for (2P2 , 2C3 , C3 + P4 , 2P4 , T0,0,4 )-free graphs. 5 Conclusions We showed that every connected graph is almost classified except for the claw and the P5 . Our notion of almost classified graphs originated from recent work [16,20,27,28] on Colouring for (H1 , H2 )-free graphs for connected graphs H1 and H2 , in particular when H1 = P5 . We decreased the number of open cases for the latter graph by showing new NP-hardness results for (2P2 , H)-free graphs. In the following theorem we summarize all known results for Colouring restricted to (2P2 , H)-free graphs and (P5 , H)-free graphs. Theorem 5. Let H be a graph on n vertices. Then the following two statements hold: (i) If H contains a graph in {C3 + P4 , 3P2 , 2P4 } as an induced subgraph, or H is not an induced subgraph of T1,1,3 + P2n−1 , then Colouring is NP-complete for (2P2 , H)-free graphs. (ii) If H is an induced subgraph of a graph in {2P1 + P3 , P1 + P4 , P2 + P3 , P5 , T0,0,1 + P1 , T0,1,1 , T0,0,2 } or of sP1 + P2 for some integer s ≥ 0, then Colouring is polynomial-time solvable for (P5 , H)-free graphs. Proof. If H contains a graph in {2C3 , C3 + P4 , 3P2 , 2P4 , T0,2,2 , T0,0,4 } as an induced subgraph, then Colouring is NP-complete for (2P2 , H)-free graphs due to Theorems 3 and 4. We may therefore assume that H is (2C3 , C3 + P4 , 3P2 , 2P4 , T0,2,2 , T0,0,4 )-free. (Note that T1,1,3 + P2n−1 is (2C3 , T0,2,2 , T0,0,4 )-free.) Recall that T is the class of graphs every component of which is isomorphic to a graph Th,i,j for some / T . Therefore, if H ∈ / T , then Colouring 1 ≤ h ≤ i ≤ j or a path Pr for some r ≥ 1. Note that 2P2 = C4 ∈ is NP-complete for (2P2 , H)-free graphs by Theorem 1. We may therefore assume that H ∈ T . Since H is 7 2C3 -free, H can contain at most one component that is not a path. Since H is (T0,2,2 , T0,0,4 )-free, if H does have a component that is not a path, then this component must be an induced subgraph of T1,1,3 . The union of components of H that are isomorphic to paths form an induced subgraph of P2n−1 . Therefore H is an induced subgraph of T1,1,3 + P2n−1 . It is known that Colouring is polynomial-time solvable for (P5 , H)-free graphs if H is an induced subgraph of 2P1 + P3 [27], P1 + P4 (this follows from the fact that (P5 , P1 + P4 )-free graphs have cliquewidth at most 5 [3]; see also [2] for a linear-time algorithm), P2 + P3 [28], P5 [16], T0,0,1 + P1 [20], T0,1,1 [20], T0,0,2 [20] or sP1 + P2 for some integer s ≥ 0 [28]. ⊓ ⊔ Theorem 5 leads to the following open problem, which shows how the P5 is not almost classified. Recall that T0,0,0 = C3 . Open Problem 1 Determine the complexity of Colouring for (2P2 , H)-free graphs and for (P5 , H)-free graphs if – – – – – – H H H H H H = sP1 + Pt + Th,i,j for 0 ≤ h ≤ i ≤ j ≤ 1, s ≥ 0 and 2 ≤ t ≤ 3 = sP1 + Th,i,j for 0 ≤ h ≤ i ≤ 1 ≤ j ≤ 3 and s ≥ 0 such that h + i + j + s ≥ 3 = sP1 + T0,0,0 for s ≥ 2 = sP1 + Pt for s ≥ 0 and 3 ≤ t ≤ 7 such that s + t ≥ 6 = sP1 + Pt + Pu for s ≥ 0, 2 ≤ t ≤ 3 and 3 ≤ u ≤ 4 such that s + t + u ≥ 6 = sP1 + 2P2 for s ≥ 1. Open Problem 1 shows the following. – The open cases for Colouring restricted to (2P2 , H)-free graphs and (P5 , H)-free graphs coincide. – The graph H in each of the open cases is connected. – The number of minimal open cases is 10, namely when H ∈ {C3 + 2P1 , C3 + P2 , P1 + 2P2 } (see also Section 1) and when H ∈ {3P1 + P3 , 2P1 + P4 , 2P3 , P6 , T0,1,1 + P1 , T0,1,2 , T1,1,1 }. As every graph H listed in Open Problem 1 appears as an induced subgraph in both the graph G′1 and the graph G′2 defined in Section 4, we need new arguments to solve the open cases in Problem 1. The complexity of Colouring for (K1,3 , H)-free graphs is less clear. As mentioned in Section 1, the cases where H ∈ {4P1 , 2P1 + P2 , C4 + P1 } are still open. Moreover, K1,3 is not almost classified, as the case H = Pt is open for all t ≥ 6 (polynomial-time solvability for t = 5 was shown in [26]). Note that |E(H)| may be arbitrarily large, while Open Problem 1 shows that |E(H)| ≤ 8 in all open cases for the P5 . Since we have no new results for the case H1 = K1,3 , we refer to [24] for further details or to the summary of Colouring restricted to (H1 , H2 )-free graphs in [12]. References 1. A. Blanché, K. K. Dabrowski, M. Johnson, and D. Paulusma. Hereditary graph classes: When the complexities of Colouring and Clique Cover coincide. CoRR, abs/1607.06757, 2016. 2. H. L. Bodlaender, A. Brandstädt, D. Kratsch, M. Rao, and J. Spinrad. On algorithms for (P5 ,gem)-free graphs. Theoretical Computer Science, 349(1):2–21, 2005. 3. A. Brandstädt, H.-O. Le, and R. Mosca. Chordal co-gem-free and (P5 ,gem)-free graphs have bounded clique-width. Discrete Applied Mathematics, 145(2):232–241, 2005. 4. H. Broersma, P. A. Golovach, D. Paulusma, and J. Song. Determining the chromatic number of triangle-free 2P3 -free graphs in polynomial time. Theoretical Computer Science, 423:1–10, 2012. 5. K. Cameron and C. T. Hoàng. Solving the clique cover problem on (bull,C4 )-free graphs. CoRR, abs/1704.00316, 2017. 6. M. Chudnovsky, N. Robertson, P. Seymour, and R. Thomas. The strong perfect graph theorem. Annals of Mathematics, 164(1):51–229, 2006. 7. K. K. Dabrowski, F. Dross, and D. Paulusma. Colouring diamond-free graphs. Journal of Computer and System Sciences, 89:410–431, 2017. 8 8. K. K. Dabrowski, P. A. Golovach, and D. Paulusma. Colouring of graphs with Ramsey-type forbidden subgraphs. Theoretical Computer Science, 522:34–43, 2014. 9. T. Emden-Weinert, S. Hougardy, and B. Kreuter. Uniquely colourable graphs and the hardness of colouring graphs of large girth. Combinatorics, Probability and Computing, 7(04):375–386, 1998. 10. S. Gaspers, S. Huang, and D. Paulusma. Colouring square-free graphs without long induced paths. Manuscript, 2017. 11. P. A. Golovach and P. Heggernes. Choosability of P5 -free graphs. Proc. MFCS 2009, LNCS, 5734:382–391, 2009. 12. P. A. Golovach, M. Johnson, D. Paulusma, and J. Song. A survey on the computational complexity of colouring graphs with forbidden subgraphs. Journal of Graph Theory, 84(4):331–363, 2017. 13. P. A. Golovach and D. Paulusma. List coloring in the absence of two subgraphs. Discrete Applied Mathematics, 166:123–130, 2014. 14. M. Grötschel, L. Lovász, and A. Schrijver. Polynomial algorithms for perfect graphs. Annals of Discrete Mathematics, 21:325–356, 1984. 15. P. Hell and S. Huang. Complexity of coloring graphs without paths and cycles. Discrete Applied Mathematics, 216, Part 1:211–232, 2017. 16. C. T. Hoàng and D. A. Lazzarato. Polynomial-time algorithms for minimum weighted colorings of (P5 , P5 )-free graphs and similar graph classes. Discrete Applied Mathematics, 186:106–111, 2015. 17. I. Holyer. The NP-Completeness of edge-coloring. SIAM Journal on Computing, 10(4):718–720, 1981. 18. S. Huang. Improved complexity results on k-coloring Pt -free graphs. European Journal of Combinatorics, 51:336– 346, 2016. 19. S. Huang, M. Johnson, and D. Paulusma. Narrowing the complexity gap for colouring (Cs , Pt )-free graphs. The Computer Journal, 58(11):3074–3088, 2015. 20. T. Karthick, F. Maffray, and L. Pastor. Polynomial cases for the vertex coloring problem. CoRR, abs/1709.07712, 2017. 21. D. Král’, J. Kratochvíl, Zs. Tuza, and G. J. Woeginger. Complexity of coloring graphs without forbidden induced subgraphs. Proc. WG 2001, LNCS, 2204:254–262, 2001. 22. L. Lovász. Coverings and coloring of hypergraphs. Congressus Numerantium, VIII:3–12, 1973. 23. V. V. Lozin and D. S. Malyshev. Vertex coloring of graphs with few obstructions. Discrete Applied Mathematics, 216, Part 1:273–280, 2017. 24. V. V. Lozin and C. Purcell. Coloring vertices of claw-free graphs in three colors. Journal of Combinatorial Optimization, 28(2):462–479, 2014. 25. F. Maffray and M. Preissmann. On the NP-completeness of the k-colorability problem for triangle-free graphs. Discrete Mathematics, 162(1–3):313–317, 1996. 26. D. S. Malyshev. The coloring problem for classes with two small obstructions. Optimization Letters, 8(8):2261– 2270, 2014. 27. D. S. Malyshev. Two cases of polynomial-time solvability for the coloring problem. Journal of Combinatorial Optimization, 31(2):833–845, 2016. 28. D. S. Malyshev and O. O. Lobanova. Two complexity results for the vertex coloring problem. Discrete Applied Mathematics, 219:158–166, 2017. 29. F. P. Ramsey. On a problem of formal logic. Proceedings of the London Mathematical Society, s2-30(1):264–286, 1930. 30. T. J. Schaefer. The complexity of satisfiability problems. Proc. STOC 1978, pages 216–226, 1978. 31. D. Schindl. Some new hereditary classes where graph coloring remains NP-hard. Discrete Mathematics, 295(1– 3):197–202, 2005. 9
8cs.DS
On the Theory of Structural Subtyping Viktor Kuncak and Martin Rinard Laboratory for Computer Science Massachusetts Institute of Technology Cambridge, MA 02139, USA {vkuncak, rinard}@lcs.mit.edu arXiv:cs/0408015v1 [cs.LO] 5 Aug 2004 MIT-LCS-TR-879, January 2003 Abstract Contents We show that the first-order theory of structural subtyping of non-recursive types is decidable. Let Σ be a language consisting of function symbols (representing type constructors) and C a decidable structure in the relational language L containing a binary relation ≤. C represents primitive types; ≤ represents a subtype ordering. We introduce the notion of Σ-term-power of C, which generalizes the structure arising in structural subtyping. The domain of the Σ-term-power of C is the set of Σ-terms over the set of elements of C. We show that the decidability of the first-order theory of C implies the decidability of the first-order theory of the Σterm-power of C. This result implies the decidability of the first-order theory of structural subtyping of non-recursive types. Our decision procedure is based on quantifier elimination and makes use of quantifier elimination for term algebras and Feferman-Vaught construction for products of decidable structures. We also explore connections between the theory of structural subtyping of recursive types and monadic second-order theory of tree-like structures. In particular, we give an embedding of the monadic second-order theory of infinite binary tree into the first-order theory of structural subtyping of recursive types. 1 Introduction 2 2 Preliminaries 2.1 Term Algebra . . . . . . . . . . . . . . . . . . 2.2 Terms as Trees . . . . . . . . . . . . . . . . . 2.3 First Order Structures with Partial Functions 2 3 3 3 . . . . . . . . . . 25 26 27 27 28 29 Keywords: Structural Subtyping, Quantifier Elimination, Term Algebra, Decision Problem, Monadic SecondOrder Logic 6 Term-Powers of Decidable Theories 6.1 Product Theory of Terms of a Given Shape 6.2 A Logic for Term-Power Algebras . . . . . . 6.3 Some Properties of Term-Power Structure . 6.4 Quantifier Elimination . . . . . . . . . . . . 6.5 Handling Contravariant Constructors . . . . 6.6 A Note on Element Selection . . . . . . . . . . . . . . 31 31 32 32 37 43 44 . . . . . . . . . . . . . . . Sets 44 45 46 47 47 47 48 3 Some Quantifier Elimination Procedures 3.1 Quantifier Elimination . . . . . . . . . . . . 3.2 Quantifier Elimination for Boolean Algebras 3.3 Feferman-Vaught Theorem . . . . . . . . . . 3.4 Term Algebras . . . . . . . . . . . . . . . . 3.4.1 Term Algebra in Selector Language . 3.4.2 Quantifier Elimination . . . . . . . . 4 The 4.1 4.2 4.3 7 Some Connections with MSOL 7.1 Structural Subtyping Recursive Types . 7.2 A Decidable Substructure . . . . . . . . 7.3 Embedding Terms into Terms . . . . . . 7.4 Subtyping Trees of Known Shape . . . . 7.5 Recursive Feature Trees . . . . . . . . . 7.6 Reversed Binary Tree with Prefix-Closed 8 Conclusion 1 6 6 7 8 11 11 12 Pair Constructor and Two Constants 16 Boolean Algebras on Equivalent Terms . . . . 16 A Multisorted Logic . . . . . . . . . . . . . . 17 Quantifier Elimination for Two Constants . . 18 5 A Finite Number of Constants 5.1 Extended Term-Power Structure . . . . 5.2 Structural Base Formulas . . . . . . . . 5.3 Conversion to Base Formulas . . . . . . 5.4 Conversion to Quantifier-Free Formulas 5.5 One-Relation-Symbol Variance . . . . . ∗ This research was supported in part by DARPA Contract F3361500-C-1692, NSF Grant CCR00-86154, NSF Grant CCR00-63513, and the Singapore-MIT Alliance. † Draft of February 5, 2018, 11:28pm, see http://www.mit.edu/~vkuncak/papers for later versions. . . . . . . . . . . . 48 1 Introduction [31, Chapter 23], [39, 30], with membership constraints [10] and with queues [43], decidability of products [35, 14], [31, Chapter 12], and algebraically closed fields [50], The complexity of the decision problem for the first-order theory of structural subtyping has a non-elementary lower bound. This is a consequence of a general theorem about pairing functions [15, Theorem 1.2, Page 163] and applies to term algebras already, as observed in [39, 43]. Subtyping constraints are an important technique for checking and inferring program properties, used both in type systems and program analyses [34, 16, 13, 28, 23, 4, 3, 1, 2, 20, 41, 17, 54, 7, 8, 5, 42, 47, 19]. This paper presents a decision procedure for the firstorder theory of structural subtyping of non-recursive types. This result solves (for the case of non-recursive types) a problem left open in [48]. [48] provides the decidability result for structural subtyping of only unary type constructors, whereas we solve the problem for any number of constructors of any arity. Furthermore, we do not impose any constraints on the subtyping relation ≤, it need not even be a partial order. The generality of our construction makes it potentially of independent interest in logic and model theory. We approach the problem of structural subtyping using quantifier elimination and, to some extent, using monadic second-order logic of tree-like structures. This paper makes the contributions: 2 Preliminaries In this section we review some notions used in the this paper. If w is a word over some alphabet, we write |w| for the length of w. We write w1 · w2 to denote the concatenation of words w1 and w2 . A node v in a directed graph is a sink if v has no outgoing edges. A node v in a directed graph is a source if v has no incoming edges. We write E1 ≡ E2 to denote equality of syntactic entities E1 and E2 . We write x̄ to denote some sequence of variables x1 , . . . , xn . We assume that formulas are built from propositional connectives ∧, ∨, ¬, the remaining connectives are defined as shorthands. Connective ¬ binds the strongest, followed by ∧ and ∨. A literal L is an atomic formula A or a negation of an atomic formula ¬A. We define complementation of a literal by A = ¬A and ¬A = A. A formula ψ is in prenex form if it is of the form • we give a new presentation of Feferman-Vaught theorem for direct products using a multisorted logic (Section 3.3); for completeness we also include proof of quantifier-elimination for boolean algebras of sets (Section 3.2); • we give a new presentation of decidability of the firstorder theory of term algebras; the proof uses the language of both constructor and selector symbols (Section 3.4); • as an introduction to main result, we show decidability of structural subtyping with one covariant binary constructor and two constants (Section 4), this result does not rely on Feferman-Vaught technique; Q1 x1 . . . . Qn xn .φ where Qi ∈ {∀, ∃} for 1 ≤ i ≤ n and φ is a quantifier free formula. We call φ a matrix of ψ. If φ is a formula then FV(φ) denotes the set of free variables in φ. We write [x1 7→ a1 , . . . , xk 7→ ak ] for the substitution σ such that σ(xi ) = ai for 1 ≤ i ≤ k. If φ is a formula and t1 , . . . , tk terms, we write φ[x1 := t1 , . . . , xk := tk ] for the result of simultaneously substituting free occurrences of variables xi with term ti , for 1 ≤ i ≤ k. We write h(t) for the height of term t. h(a) = 0 if a is a constant, h(x) = 0 if x is a variable. If f (t1 , . . . , tk ) is a term then • we present a new construction, term-power algebra for creating tree-like theories based on existing theories (Section 5); • as a central result, we prove that if the base theory is decidable, so is the theory of term-power with arbitrary variance of constructors; we give an effective decision procedure for quantifier elimination in termpower structure; the procedure combines elements of quantifier elimination in Feferman-Vaught theorem and quantifier elimination in term algebras (Sections 5, 6). h(f (t1 , . . . , tk )) = 1 + max(h(t1 ), . . . , h(tk )) • we show the decidability of structural subtyping nonrecursive types as a direct consequence of the main result; We assume that all function symbols are of finite arity. If there are finitely many function symbols then for any nonnegative integer k there is only a finite number of terms t such that h(t) ≤ k. If φ(u) is a conjunction of literals, we say that φ′ results from ∃u.φ(u) by dropping quantified variable u iff φ′ is the result of eliminating from φ(u) all conjunctions containing u. More generally, if ψ is a formula of form • we give a simple embedding of monadic second-order theory of infinite binary tree into the theory of structural subtyping of recursive types with two primitive types (Section 7.1); • we show that structural subtyping of recursive types where terms range over constant shapes is decidable (Section 7.4); Q1 x1 . . . Qu . . . Qk xk . ψ0 then the result of dropping u from ψ is In addition to showing the decidability of structural subtyping, our hope is to promote the important technique of quantifier elimination, which forms the basis of our result. Quantifier elimination [22, Section 2.7] is a fruitful technique that was used to show decidability and classification of boolean algebras [46, 51] decidability of term algebras Q1 x1 . . . Qk xk . ψ0′ where ψ0′ is the result of dropping u from ∃u.φ0 . An equality is an atomic formula t1 = t2 where t1 and t2 are terms. A disequality is negation of an equality. 2 Let Cons(Σ) be the term algebra interpretation of signature Σ, defined as follows [22, Page 14]. For every f ∈ Σ with ar(f ) = k define Jf K ∈ Cons(Σ), with Jf K : FT(Σ)k → FT(Σ) by Jf K(t1 , . . . , tk ) = f (t1 , . . . , tk ) We use the usual Tarskian semantics of formulas. Unless otherwise stated φ |= ψ will denote that formula φ ⇒ ψ is true in a fixed relational structure that is under current consideration. Occasionally we find it convenient to work with multisorted logic, where domain is union of disjoint sets called sorts, and arity specifies the sorts of all operations. Constants are operations with zero arguments. Relations are operations that return the result in a distinguished sort bool interpreted over the boolean lattice {false, true} or over the distributive lattice of three-valued logic {false, true, undef} from Section 2.3). A structure C of a given language L is a pair of domain C and the interpretation function J KC . Hence, we name operations of the structure using symbols of the language and the interpretation function. If C is clear from the context we write simply J K for J KC . In Section 3.3 and Section 6 we use logic with several kinds of quantifiers. Our logic is first-order, but we give higher-order types to quantifiers. For example, a quantifier We will write f instead of Jf K when it causes no confusion. 2.2 We define trees representing terms as follows. We use sequences of nonegative integers to denote paths in the tree. Let Σ be a signature. A tree over Σ is a partial function t from the set Nat∗ of paths to the set Σ of function symbols such that: 1. if w ∈ Nat∗ , x ∈ Nat, and t(w · x) is defined, then t(w) is defined as well; 2. if t(w) = f with ar(f ) = k, then {i | t(w · i) is defined } = {1, . . . , k} Q :: (A → B) → B A finite tree is a tree with a finite domain. denotes a quantifier that binds variables of A sort enclosed within an expression of B sort and returns an expression of B sort. If X and Y are sets then X → Y denotes the set of all functions from A to B. When specifying the semantics of the quantifier Q we specify a function 2.3 The semantics of an expression M of sort B takes an environment σ which is a function from variable names to elements of A and produces an element of B, hence JM Kσ ∈ JBK. We define the semantics of an expression Qx. M by: JQx. M Kσ = JQKh where h : JAK → JBK is the function h(a) = JM K(σ[x := a]) σ[x := a](y) = ( σ(y), if y 6≡ x a, if y ≡ x Specifying types for quantifiers allows to express more Let σA be some arbitrary dummy global environment. If F is a formula without global variables we write JF KσA to denote the truth value of F ; clearly JF KσA does not depend on σA and we denote it simply JF K when no ambiguity arises. We use Hilbert’s epsilon as a notational convenience in metatheory. If P (x) is a unary predicate, then εx.P (x) denotes an arbitrary element d such that P (d) holds, if such element exists, or an arbitrary object otherwise. 2.1 First Order Structures with Partial Functions We make use of partial functions in our quantifier elimination procedures. In this section we briefly describe the approach to partial functions we chose to use; other approaches would work as well, see e.g. [24]. A language of partial functions Σ1 contains partial function symbols in addition to total function symbols and relation symbols. Consider a structure with the domain A interpreting a language with partial function symbols Σ1 . Given some environment σ, we have JtKσ ∈ A ∪ {⊥} where ⊥ ∈ /A is a special value denoting undefined results. We require the interpretations of total and partial function symbols to be strict in ⊥, i.e. f (a1 , . . . , ai , ⊥, ai+2 , . . . , ak ) = ⊥. We interpret atomic formulas and their negations over the three-valued domain {false, true, undef} using strong Kleene’s three-valued logic [26, 24, 44]. We require that JRK(a1 , . . . , ai , ⊥, ai+2 , . . . , ak ) = undef for every relational symbol R. Logical connectives in Kleene’s strong threevalued logic are the strongest “regular” extension of the corresponding connectives on the two-valued domain [26]. The regularity requirement means that the three-valued logic is a sound approximation of two-valued logic in the following sense. We may obtain the truth tables for three-valued logic by considering the truth values false, true, undef as shorthands for sets {false}, {true}, {false, true} and defining each logical operation ∗ by: JQK : (JAK → JBK) → JBK Here Terms as Trees s1 J∗K s2 = {b1 ◦ b2 | b1 ∈ s1 ∧ b2 ∈ s2 } where ◦ denotes the corresponding operation in the twovalued logic. As in a call-by-value semantics of lambda calculus, variables in the environments (σ) do not range over ⊥. We interpret quantifiers as ranging over the domain A or its subset if the logic is multisorted; the interpretation of quantifiers are similarly the best regular approximations of the corresponding two-valued interpretations. These properties of Kleene’s three-valued logic have the following important consequence. Suppose that we extend the definition of all partial functions to make them total functions on the domain A by assigning arbitrary values outside the original domain. Suppose that a formula φ evaluates Term Algebra We introduce the notion of term algebra [22, Page 14]. Let Nat be the set of natural numbers. Let the signature Σ be a finite set of function symbols and constants and let ar : Σ → Nat be a function specifying arity ar(f ) for every function symbol or constant f ∈ Σ. Let FT(Σ) denote the set of finite ground terms over signature Σ. We assume that Σ contains at least one constant c ∈ Σ, ar(c) = 0, and at least one function symbol f ∈ Σ, ar(f ) > 0. Therefore, FT(Σ) is countably infinite. 3 to an element of b ∈ {false, true} in Kleene’s logic. Then φ evaluates to the same truth-value b in the new logic of total functions. This property of three-valued logic implies that the algorithms that we use to transform formulas with partial functions will apply even for the logic that makes all functions total by completing them with arbitrary elements of A. We say that a formula ψ is well-defined iff its truth value is an element of {false, true}.  To enable the transformation of each well-defined formula into a disjunction of well-defined conjunctions of literals, we enrich the language of function and relation symbols as follows. With each partial function symbol f ∈ Σ1 of arity k = ar(f ) we associate a domain description Df = hhx1 , . . . , xk i, φi specifying the domain of f . Here x1 , . . . , xk are distinct variables and φ is an unnested conjunction of literals such that FV(φ) ⊆ {x1 , . . . , xk }. We require every interpretation of a first-order structure with partial function symbols to satisfy the following property: Example 1 Consider the domain of real numbers. The following formulas are not well-defined: 3 = 1/0 Jf K(a1 , . . . , ak ) 6= ⊥ ⇐⇒ JφK[x1 7→ a1 , . . . , xk 7→ ak ] ∀x. 1/x > 0 ∨ 1/x < 0 ∨ 1/x = 0 for all a1 , . . . , ak ∈ A. We henceforth assume that every structure with partial functions is equipped with a domain description Df for every partial function symbol f . The Proposition 8 below gives an algorithm for transforming a given well-defined formula into a disjunction of well-defined conjunctions. We first give some definitions and lemmas. The following formulas are well-defined: ∃x. 1/x = 3 ∀x. 1/x 6= 3 x = 0 ∨ 1/x > 0 Definition 3 If ψ is a formula with free variables, a domain formula for ψ is a formula φ not containing partial function symbols such that, for every valuation σ,  We say that a formula φ1 is equivalent to a formula φ2 and write φ1 ∼ = φ2 iff JψKσ 6= undef ⇐⇒ JφKσ = true From Definition 3 we obtain the following Lemma 4. Jφ1 Kσ = Jφ2 Kσ Lemma 4 Let ψ be a formula and φ a domain formula for ψ. Then ψ∼ = (ψ ∧ φ) ∨ (undef ∧ ¬φ) for all valuations σ (including those for which Jφ1 Kσ = undef). Sections below perform equivalence-preserving transformations of formulas. This means that starting from a welldefined formula we obtain an equivalent well-defined formula. When doing equivalence preserving transformations it is useful to observe that ∧, ∨ still form a distributive lattice. The partial order of this lattice is the chain false ≤ undef ≤ true. The element undef does not have a complement in the lattice; unary operation ¬ does not denote the lattice complement. However, the following laws still hold: Proof. Let σ be arbitrary valuation. Let v = JψKσ. If v ∈ {true, false} then JφKσ = true and J(ψ ∧ φ) ∨ (undef ∧ ¬φ)Kσ = (v ∧ true) ∨ (undef ∧ false) = v. If v = undef then JφK = false, so J(ψ ∧ φ) ∨ (undef ∧ ¬φ)Kσ = ¬(x ∧ y) ∼ = ¬x ∨ ¬y ∼ ¬x ∧ ¬y ¬(x ∨ y) = (undef ∧ false) ∨ (undef ∧ true) = undef. ¬¬x ∼ =x Observe that ψ ∧φ in Lemma 4 is a well-defined conjunction. We use this property to construct domain formulas using partial function domain descriptions. Let Df = hhx1 , . . . , xk i, B1f ∧ . . . ∧ Blff i The properties of ∧, ∨, ¬ are sufficient to transform any quantifier-free formula into disjunction of conjunctions of literals using the well-known straightforward technique. However, this straightforward technique in some cases yields conjunctions that are not well-defined, even though the formula as a whole is well-defined. for each partial function symbol f ∈ Σ1 of arity k, where B1f , . . . , Blff are unnested literals. If t1 , . . . , tk are terms, we write Bif (t1 , . . . , tk ) for Bif [x1 := t1 , . . . , xk := tk ]. Let subt(t) denote the set of all subterms of term t. For any literal B(t1 , . . . , tn ) where B(t1 , . . . , tn ) ≡ R(t1 , . . . , tn ) or B(t1 , . . . , tn ) ≡ ¬R(t1 , . . . , tn ), define Example 2 Transforming a negation of well-defined formula: ¬(x 6= 0 ∧ (y = 1/x ∨ z = x + 1)) may yield the following disjunction of conjunctions: DomForm(B(t1 , . . . , tn )) = V x = 0 ∨ (y 6= 1/x ∧ z 6= x + 1) f (s1 ,...,sk )∈∪1≤i≤n subt(ti ) where y 6= 1/x ∧ z 6= x + 1 is not a well-defined conjunction for x = 0. 1≤j≤lf 4 Bjf (s1 , . . . , sk ) (1) Lemma 5 Let B(t1 , . . . , tn ) be a literal containing partial function symbols. Then DomForm(B(t1 , . . . , tn )) is a domain formula for B(t1 , . . . , tn ). Proposition 8 Every well-defined quantifier-free formula ψ can be transformed into an equivalent disjunction ψ ′ of well-defined conjunctions of literals. Proof. Let σ be a valuation. By strictness of interpretations of function and predicate symbols, JB(t1 , . . . , tn )Kσ 6= undef iff Jf (s1 , . . . , sk )Kσ 6= ⊥ for every subterm f (s1 , . . . , sk ) of every term ti , iff JBjf (s1 , . . . , sk )Kσ = true for every 1 ≤ j ≤ lf and every subterm f (s1 , . . . , sk ). Proof. Using the standard procedure, convert ψ to disjunction of conjunctions C1 ∨ . . . ∨ Cn ∧ Ci′ Let Ci = B where B is a literal and let DomForm(B) = F1 ∧ . . . ∧ Fm . Replace B ∧ Ci′ by (B ∧ F1 ∧ . . . ∧ Fm ∧ Ci′ ) ∨ W ′ 1≤i≤m (undef ∧ ¬Fi ∧ DomForm(Fi ) ∧ Ci ) Lemma 6 Let B be a literal and let DomForm(B) = F1 ∧ . . . ∧ Fm . By Lemma 6 and distributivity, the result is an equivalent formula. Repeat this process for every literal in C1 ∨. . .∨Cn . The result can be written in the form Then B∼ = (B ∧ F1 ∧ . . . ∧ Fm ) ∨ W 1≤i≤m (undef ∧ ¬Fi ∧ DomForm(Fi )) (undef ∧ φ1 ) ∨ . . . ∨ (undef ∧ φp ) ∨ φp+1 ∨ . . . ∨ φp+q where each φi for 1 ≤ i ≤ p + q is a well-defined conjunction. Formula (2) is equivalent to Proof. If JBKσ 6= undef, then JFi Kσ = true for every 1 ≤ i ≤ m, and (undef ∧ (φ1 ∨ . . . ∨ φp )) ∨ φp+1 ∨ . . . ∨ φp+q Jundef ∧ ¬Fi ∧ DomForm(Fi )Kσ = false (3) and is equivalent to the well-defined formula ψ, so it is welldefined. Formulas φ1 ∨ . . . ∨ φp and φp+1 ∨ . . . ∨ φp+q are also well-defined. By Lemma 7, we conclude that formula (3) is equivalent to so the right-hand side evaluates to JBKσ as well. Now consider the case when JBKσ = undef. Then there exists a term f (s1 , . . . , sk ) such that Jf (s1 , . . . , sk )Kσ = undef. Because σ(x) 6= ⊥ for every variable x, there exists a term f (s1 , . . . , sk ) such that Jf (s1 , . . . , sk )Kσ = undef and Jsi Kσ 6= undef for 1 ≤ i ≤ k. Then there exists a formula Fp of form Bjf (s1 , . . . , sk ) such that JBjf (s1 , . . . , sk )Kσ = false, and Jundef ∧ ¬Fp ∧ DomForm(Fp )Kσ = undef. φp+1 ∨ . . . ∨ φp+q (4) Because (4) is a disjunction of well-defined formulas, (4) is the desired result ψ ′ . The following proposition presents transformation to unnested form for the structures with equality and partial function symbols, building on Proposition 8. For a similar unnested form in the first-order logic containing only total function symbols, see [22, Page 58]. Because JB ∧ F1 ∧ . . . ∧ Fm Kσ = false, and for every q, Proposition 9 Every well-defined quantifier-free formula ψ in a language with equality can be effectively transformed into an equivalent formula ψ ′ where ψ ′ is a disjunction of existentially quantified well-defined conjunctions of the following kinds of literals: Jundef ∧ ¬Fq ∧ DomForm(Fq )Kσ ∈ {undef, false}, the right-hand side evaluates to undef. Lemma 7 Let φ0 (ȳ) and φ1 (ȳ) be well-defined formulas whose free variables are among ȳ and let • R(x1 , . . . , xk ) where R is some relational symbol of arity k and x1 , . . . , xk are variables; ψ(ȳ) ≡ (undef ∧ φ0 (ȳ)) ∨ φ1 (ȳ) • ¬R(x1 , . . . , xk ) where R is some relational symbol of arity k and x1 , . . . , xk are variables; If ψ(ȳ) is well-defined for all values of variables ȳ, then • x1 = x2 where x1 , x2 are variables; ψ(ȳ) ∼ = φ1 (ȳ) • x = f (x1 , . . . , xk ) where f is some partial or total function symbol of arity k and x, x1 , . . . , xk are variables; Proof. Consider any valuation σ. Let • x1 6= x2 where x1 and x2 are variables. v = Jφ1 (ȳ)Kσ and (2) Proof. Transform the formula to disjunction of wellformed conjunctions of literals as in the proof of Proposition 8. Then repeatedly perform the following transformation on each well-defined conjunction φ. Let A(f (x1 , . . . , xk )) be an atomic formula containing term f (x1 , . . . , xk ). Replace φ ∧ A(f (x1 , . . . , xk )) with ′ v = Jψ(ȳ)Kσ We need to show v = v ′ . Because φ(ȳ) and ψ(ȳ) are welldefined, v, v ′ ∈ {false, true}. We consider two cases. Case 1. v = true. Then also v ′ = true. Case 2. v = false. Then v ′ = undef ∧ φ0 (ȳ). Because v ′ 6= undef, we conclude v ′ = false. ∃x0 . φ ∧ x0 = f (x1 , . . . , xk ) ∧ A(x0 ) 5 Replace x 6= f (x1 , . . . , xk ) with Proof. The following sequence of equivalences holds. ∀x. (undef ∧ φ0 (x, ȳ)) ∨ φ1 (x, ȳ) ∼ = x0 = f (x1 , . . . , xk ) ∧ x0 6= x ¬∃x.¬(undef ∧ φ0 (x, ȳ)) ∨ φ1 (x, ȳ) ∼ = Repeat this process until the resulting conjunction φ′ is in unnested form. φ′ is clearly equivalent to the original conjunction φ when all partial functions are well-defined. When some partial function is not well-defined, then both φ and φ′ evaluate to false, because by construction of φ in the proof of Proposition 8, each conjunction contains conjuncts that evaluate to false when some application of a function symbol is not well-defined. ¬∃x. (undef ∨ ¬φ0 (x, ȳ)) ∧ ¬φ1 (x, ȳ) ∼ = ¬∃x. (undef ∧ ¬φ1 (x, ȳ)) ∨ (¬φ0 (x, ȳ) ∧ ¬φ1 (x, ȳ)) ∼ = ¬ ((undef ∧ ∃x.¬φ1 (x, ȳ)) ∨ (∃x. ¬φ0 (x, ȳ) ∧ ¬φ1 (x, ȳ))) ∼ = ∼ (undef ∨ ∀x.φ1 (x, ȳ)) ∨ (∀x. φ0 (x, ȳ) ∨ φ1 (x, ȳ)) = (undef ∧ ∀x.φ0 (x, ȳ) ∨ φ1 (x, ȳ)) ∨ ∀x.φ1 (x, ȳ) Let a left-strict conjunction in Kleene logic be denoted by ∧′ and defined by p ∧′ q = (p ∧ q) ∨ (p ∧ ¬p) Proposition 13 Let φ0 (x̄, ȳ) and φ1 (x̄, ȳ) be well-defined formulas whose free variables are among ȳ and let The correctness of the transformation to unnested form in Proposition 9 relies on the presence of conjuncts that ensure that the entire conjunction evaluates to false whenever some term is undefined. The following Lemma 10 enables transformation to unnested form in an arbitrary context, allowing the transformation to unnested form to be performed independently from ensuring well-definedness of conjuncts. ψ(ȳ) ≡ Q1 x1 . . . Qn xn . (undef ∧ φ0 (x̄, ȳ)) ∨ φ1 (x̄, ȳ) where Q1 , . . . , Qn are quantifiers. If ψ(ȳ) is well-defined for all values of variables ȳ, then ψ(ȳ) ∼ = Q1 x1 . . . Qn xn . φ1 (x̄, ȳ) Proof. Applying successively Lemmas 11 and 12 to quantifiers Qn , . . . , Q1 , we conclude Lemma 10 Let φ(x) be a formula with free variable x and let t be a term possibly containing partial function symbols. Then 1. φ(t) ∼ = (∃x. x = t ∧ φ(x)) ∨ (undef ∧ ∀x.¬φ(x)) ; 2. φ(t) ∼ = ∃x. x = t ∧′ φ(x) ; 3. φ(t) ∼ = (∃x. x = t ∧ φ(x)) ∨ (t 6= t) . ψ(ȳ) ∼ = (undef ∧ φ2 (ȳ)) ∨ Q1 x1 . . . Qn xn . φ1 (x̄, ȳ) for some formula φ2 (ȳ). Then by Lemma 7, ψ(ȳ) ∼ = Q1 x1 . . . Qn xn . φ1 (x̄, ȳ). Proof. Straightforward. Proposition 13 below shows that a simplification similar to one in Lemma 7 can be applied even within the scope of quantifiers. To show Proposition 13 we first show two lemmas. 3 Some Quantifier Elimination Procedures ∃x. (undef ∧ φ0 (x, ȳ)) ∨ φ1 (x, ȳ) ∼ = As a preparation for the proof of the decidability of term algebras of decidable theories, we present quantifier elimination procedures for some theories that are known to admit quantifier elimination. We use the results and ideas from this section to show the new results in Sections 4, 5, 6. (undef ∧ ∃x.φ0 (x, ȳ)) ∨ ∃x.φ1 (x, ȳ) 3.1 Lemma 11 For all formulas φ0 (x, ȳ) and φ1 (x, ȳ), Proof. By distributivity of quantifiers and propositional connectives in Kleene logic we have: Quantifier Elimination Our technique for showing decidability of structural subtyping of recursive types is based on quantifier elimination. This section gives some general remarks on quantifier elimination. We follow [22] in describing quantifier elimination procedures. According to [22, Page 70, Lemma 2.7.4] it suffices to eliminate ∃y from formulas of the form ^ ψi (x̄, y) (5) ∃y. ∃x. (undef ∧ φ0 (x, ȳ)) ∨ φ1 (x, ȳ) ∼ = (∃x.undef ∧ φ0 (x, ȳ)) ∨ ∃x.φ1 (x, ȳ) ∼ = (undef ∧ ∃x.φ0 (x, ȳ)) ∨ ∃x.φ1 (x, ȳ) 0≤i<n Lemma 12 For all formulas φ0 (x, ȳ) and φ1 (x, ȳ), where x̄ is a tuple of variables and ψi (x̄, y) is a literal whose all variables are among x̄, y. The reason why eliminating formulas of the form (5) suffices is the following. Suppose that the formula in prenex form and consider the innermost quantifier of a formula. Let φ be the subformula containing the quantifier and the subformula that is the scope of the quantifier. If φ is of the form ∀x. φ0 we may replace φ with ¬∃x.¬φ0 . Hence, we may assume that φ is of the form ∀x. (undef ∧ φ0 (x, ȳ)) ∨ φ1 (x, ȳ) ∼ = (undef ∧ ∀x.φ0 (x, ȳ) ∨ φ1 (x, ȳ)) ∨ ∀x.φ1 (x, ȳ) 6 ∃x. φ1 . We then transform φ1 into disjunctive normal form and use the fact ∃x. (φ2 ∨ φ3 ) ⇐⇒ (∃x. φ2 ) ∨ (∃x. φ3 ) denoted by t has exactly k elements. These properties are first-order definable as follows. |t| ≥ 0 (6) We conclude that elimination of quantifiers from formulas of form (5) suffices to eliminate the innermost quantifier. By repeatedly eliminating innermost quantifiers we can eliminate all quantifiers from a formula. We may also assume that y occurs in every literal ψi , otherwise we would place the literal outside the existential quantifier using the fact for y not occurring in B. To eliminate variables we often use the following identity of a theory with equality: |t| ≥ k+1 ≡ ∃x. x ⊂ t ∧ |x| ≥ k |t| = k |t| ≥ k ∧ ¬|t| ≥ k+1 ≡ |(t3 ∩ tc4 ) ∪ (t4 ∩ tc3 )| = 0 It is therefore sufficient to consider the first-order formulas whose only atomic formulas are of the form |t| = 0. For the purpose of quantifier elimination we will additionally consider formulas that contain atomic formulas |t|=k for all k ≥ 1, as well as |t|≥k for k ≥ 0. Note that we can eliminate negative literals as follows: (7) Section 2.3 presents analogous identities for partial functions. Quantifier elimination procedures we give imply the decidability of the underlying theories. In this paper the interpretations of function and relation symbols on some domain A are effectively computable functions and relations on A. Therefore, the truth-value of every formula without variables is computable. The quantifier elimination procedures we present are all effective. To determine the truth value of a closed formula φ it therefore suffices to apply the quantifier elimination procedure to φ, yielding a quantifier free formula ψ, and then evaluate the truth value of ψ. 3.2 true We call a language which contains terms |t| ≥ k and |t| = k the language of boolean algebras with finite cardinality constraints. Because finite cardinality constraints are first-order definable, the language with finite cardinality constraints is equally expressive as the language of boolean algebras. Every inequality t1 ⊆ t2 is equivalent to the equality t1 ∩ t2 = t1 , and every equality t3 = t4 is equivalent to the cardinality constraint ∃y. (A ∧ B) ⇐⇒ (∃y.A) ∧ B ∃x.x = t ∧ φ(x) ⇐⇒ φ(t) ≡ ¬|t| = k ⇐⇒ |t| = 0 ∨ · · · ∨ |t| = k−1 ∨ |t| ≥ k+1 ¬|t| ≥ k ⇐⇒ |t| = 0 ∨ · · · ∨ |t| = k−1 (8) Every formula in the language of boolean algebras can therefore be written in prenex normal form where the matrix of the formulas is a disjunction of conjunctions of atomic formulas of the form |t| = k and |t| ≥ k, with no negative literals. Note that if a term t contains at least one operation of arity one or more, we may assume that the constants 0 and 1 do not appear in t, because 0 and 1 can be simplified away. Furthermore, the expression |0| denotes the integer zero, so all terms of form |0| = k or |0| ≥ k evaluate to true or false. We can therefore simplify every nontrivial term t so that it either t contains no occurrences of constants 0 and 1, or t ≡ 1. We next describe a quantifier elimination procedure for finite boolean algebras. We first transform the formula into prenex normal form and then repeatedly eliminate the innermost quantifier. As argued in Section 3.1, it suffices to show that we can eliminate an existential quantifier from any existentially quantified conjunction of literals. Consider therefore an arbitrary existentially quantified conjunction of literals ^ ψi (x̄, y) ∃y. Quantifier Elimination for Boolean Algebras This section presents a quantifier elimination procedure for finite boolean algebras. This result dates back at least to [46], see also [51, 27, 32, 6, 49], [22, Section 2.7 Exercise 3]. Note that the operations union, intersection and complement are definable in the first-order language of the subset relation. Therefore, quantifier elimination for the first-order theory of the boolean algebra of sets is no harder than the quantifier elimination for the first-order theory of the subset relation. However, the operations of boolean algebra are useful in the process of quantifier elimination, so we give the quantifier elimination procedure for the language containing boolean algebra operations. Instead of the first-order theory of the subtype relation we could consider monadic second-order theory with no relation or function symbols. These two languages are equivalent because the first-order quantifiers can be eliminated from monadic second-order theory using the subset relation (see Section 7.1). Finite boolean algebras are isomorphic to boolean algebras whose elements are all subsets of some finite set. We therefore use the symbols for the set operations as the language of boolean algebras. t1 ∩t2 , t1 ∪t2 , tc1 , 0, 1, correspond to set intersection, set union, set complement, empty set, and full set, respectively. We write t1 ⊆ t2 for t1 ∩ t2 = t1 , we write t1 ⊂ t2 for the conjunction t1 ⊆ t2 ∧ t1 6= t2 . For every nonnegative integer k we introduce formulas |t| ≥ k expressing that the set denoted by t has at least k elements, and formulas |t| = k expressing that the set 1≤i≤n where ψi is of the form |t| = k or of the form |t| ≥ k. We assume that y occurs in every formula ψi . It follows that no ψi contains |0| or |1|. Let x1 , . . . , xm , y be the set of variables occurring in formulas ψi for 1 ≤ i ≤ n. First consider the more general case m ≥ 1. Let for i1 , . . . , im ∈ {0, 1}, ti1 ...im = xi11 ∩ · · · ∩ ximm where t0 = t and t1 = tc . The terms in the set P = {ti1 ...im | i1 , . . . , im ∈ {0, 1}} 7 original formula ∃y. |s ∩ y| ≥ k ∧ |s ∩ y c | ≥ l eliminated form |s| ≥ k + l ∃y. |s ∩ y| = k ∧ |s ∩ y c | ≥ l |s| ≥ k + l c ∃y. |s ∩ y| ≥ k ∧ |s ∩ y | = l |s| ≥ k + l ∃y. |s ∩ y| = k ∧ |s ∩ y c | = l |s| = k + l This completes the description of quantifier elimination from an existentially quantified conjunction. By repeating this process for all quantifiers we arrive at a quantifier-free formula ψ. Hence we have the following theorem. Theorem 14 For every first-order formula φ in the language of boolean algebras with finite cardinality constraints there exists a quantifier-free formula ψ such that ψ is a disjunction of conjunctions of literals of form |t| ≥ k and |t| = k where t are terms of boolean algebra, the free variables of ψ are a subset of the free variables of φ, and ψ is equivalent to φ on all algebras of finite sets. Figure 1: Rules for Eliminating Quantifiers Remark 15 Now consider the case when formula φ has no free variables. By Theorem 14, φ is equivalent to ψ where ψ contains only terms without variables. A term without variables in boolean algebra can always be simplified to 0 or 1. Because |0| = 0, the literals with |0| reduce to true or false, so we may simplify them away. The expression |1| evaluates to the number of elements in the boolean algebra. We call literals |1| = k and |1| ≥ k domain cardinality constraints. A quantifier-free formula ψ can therefore be written as a propositional combination of domain cardinality constraints. We can simplify ψ into a disjunction of conjunctions of domain cardinality constraints and transform each conjunction so that it contains at most one literal. The result ψ ′ is a single disjunction of domain cardinality constraints. We may further assume that the disjunct of form |1| ≥ k occurs at most once. Therefore, the truth value of each closed boolean algebra formula is characterized by a set C of possible cardinalities of the domain. If ψ ′ does not contain any |1| ≥ k literals, the set C is finite. Otherwise, C = C0 ∪ {k, k + 1, . . .} for some k where C0 is a finite subset of {1, . . . , k − 1}. form a partition; moreover every boolean algebra expression whose variables are among xi can be written as a disjoint union of some elements of the partition P . Any boolean algebra expression containing y can be written, for some p, q ≥ 0 as (s1 ∩ y) ∪ · · · ∪ (sp ∩ y)∪ (t1 ∩ y c ) ∪ · · · ∪ (tq ∩ y c ) where s1 , . . . , sp ∈ P are pairwise distinct elements from the partition and t1 , . . . , tq ∈ P are pairwise distinct elements from the partition. Because |(s1 ∩ y) ∪ · · · ∪ (sp ∩ y) ∪ (t1 ∩ y c ) ∪ · · · ∪ (tq ∩ y c )| = |s1 ∩ y| + · · · + |sp ∩ y| + |t1 ∩ y c | + · · · + |tq ∩ y c | the constraint of form |t| = k can be written as _ |s1 ∩ y| = k1 ∧ · · · ∧ |sp ∩ y| = kp ∧ k1 ,...,kp ,l1 ,...,lq |t1 ∩ y c | = l1 ∧ · · · ∧ |tq ∩ y c | = lp where the disjunction ranges over nonnegative integers k1 , . . . , kp , l1 , . . . , lq ≥ 0 that satisfy 3.3 Feferman-Vaught Theorem The Feferman-Vaught technique is a way of discovering the first-order theories of complex structures by analyzing their components. This description is a little vague, and in fact the Feferman-Vaught technique itself has something of a floating identity. It works for direct products, as we shall see. Clever people can make it work in other situations too. — [22], page 458 k1 + · · · + kp + l 1 + · · · + l q = k From (8) it follows that we can perform a similar transformation for constraints of form |t| ≥ k. After performing this transformation, we bring the formula into disjunctive normal form and continue eliminating the existential quantifier separately for each disjunct, as argued in Section 3.1. We may therefore assume that all conjuncts ψi are of one of the forms: |s ∩ y| = k, |s ∩ y c | = k, |s ∩ y| ≥ k, and |s ∩ y c | ≥ k where s ∈ P . If there are two conjuncts both of which contain |s∩y| for the same s, then either they are contradictory or one implies the other. We therefore assume that for any s ∈ P , there is at most one conjunct ψi containing |s ∩ y|. For analogous reasons we assume that for every s ∈ P there is at most one conjunct ψi containing |s ∩ y c |. The result of eliminating the variable y is then given in Figure 1. The case when a literal containing |s ∩ y| does not occur is covered by the case |s ∩ y| ≥ k for k = 0, similarly for a literal containing |s ∩ y c |. It remains to consider the case m = 0. Then y is the only variable occurring in conjuncts ψi . Every cardinality expression t containing only y reduces to one of |y| or |y c |. If there are multiple literals containing |y|, they are either contradictory or one implies the others. We may therefore assume there is at most one literal containing |y| and at most one literal containing |y c |. We eliminate quantifier by applying rules in Figure 1 putting formally s = 1 where 1 is the universal set. We next review Feferman-Vaught theorem for direct products [14] which implies that the products of structures with decidable first-order theories have decidable first-order theories. The result was first obtained for strong and weak powers of theories in [35]; [35] also suggests the generalization to products. Our sketch here mostly follows [14] and [35], see also [31, Chapter 12] as well as [22, Section 9.6]. Somewhat specific to our presentation is the fact that we use a multisorted logic and build into the language the correspondence between formulas interpreted over C and the cylindric algebra of sets of positions. Let LC be a relational language. Let further I be some nonempty finite or countably infinite index set. For each i ∈ I let Ci = hCi , J KCi i be a decidable structure interpreting the language LC . We define direct product of the family of structures Ci , i ∈ I, as the structure P = Πi∈I Ci 8 where P = hP, J KP i. P is the set of all functions t such that t(i) ∈ Ci for i ∈ I, and J KP is defined by JrKP (t1 , . . . , tk ) = ∀i. JrKCi (t1 (i), . . . , tk (i)) for each relation symbol r ∈ LC . inner formula relations for r ∈ LC r :: tuplek → indset inner logical connectives ∧I , ∨I :: indset × indset → indset I ¬ :: indset → indset trueI , falseI :: indset inner formula relations for r ∈ LC JrKP2 (t1 , . . . , tk ) inner formula quantifiers I ∃ ,∀ I :: inner logical connectives (tuple → indset) → indset = :: I P2 (A1 , A2 ) = A1 ∧ A2 I P2 (A1 , A2 ) = A1 ∪ A2 J∧ K index set equality I J∨ K indset × indset → bool I P2 J¬ K logical connectives ∧, ∨ :: bool × bool → bool ¬ :: bool → bool true, false :: bool :: = I JfalseI KP2 = ∅ inner formula quantifiers S J∃I KP2 f = t∈P f (t) T I P2 J∀ K f = t∈P f (t) (indset → bool) → bool index set equality tuple quantifiers ∃, ∀ :: (A) = I \ A JtrueI KP2 index set quantifiers ∃L , ∀L = { i | JrKCi (t1 (i), . . . , tk (i)) } I P2 J= K (A1 , A2 ) (tuple → bool) → bool = (A1 = A2 ) logical connectives (interpreted as usual) Figure 2: Operations in product structure index set quantifiers S J∃ K f = A∈2I f (A) T L P2 J∀ K f = A∈2I f (A) L P2 For the purpose of quantifier elimination we consider a richer language of statements about product structure P. Figure 2 shows this richer language. The corresponding structure P2 = hP2 , J KP2 i contains, in addition to the function space P , a copy of the boolean algebra 2I of subsets of the index set I. We interpret a relation r ∈ LC by tuple quantifiers JrKP2 (t1 , . . . , tk ) = { i | JrKCi (t1 (i), . . . , tk (i)) } J∃K P2 f = ∃t ∈ P. f (t) J∀K P2 f = ∀t ∈ P. f (t) We let JtrueI KP2 = I and write Figure 3: Semantics of operations in product structure P2 r(t1 , . . . , tk ) =I trueI to express JrKP (t1 , . . . , tk ). Hence P2 is at least as expressive as P. Note that Figure 2 does not contain an equality relation between tuples. If we need to express the equality between tuples, we assume that some binary relation r0 ∈ LC in the base structure is interpreted as equality, and express the equality between tuples t1 and t2 using the formula: r0 (t1 , t2 ) =I trueI . Figure 3 shows the semantics of the language in Figure 2. (The logic has no partial functions, so we interpret the sort bool over the set {true, false}.) 9 We let A1 ⊆I A2 stand for A1 ∧I A2 =I A2 . Note that the interpretations of ∧I , ∨I , ¬I , trueI , falseI , =I , ∃L , ∀L form a first-order structure of boolean algebras of subsets of the set I. We call formulas in this boolean algebra sublanguage index-set algebra formulas. On the other hand, relations r for r ∈ LC , together with ∧I , ∨I , ¬I , ∃I , ∀I form the signature of first-order logic with relation symbols. We call formulas built only from these operations inner formulas. Let φ be a an inner formula with free tuple variables t1 , . . . , tm and no free indset variables. Then φ specifies a relation ρ ⊆ Dm . Consider the corresponding first-order formula φ′ interpreted in the base structure C; formula φ′ specifies a relation ρ′ ⊆ C m . The following property follows from the semantics in Figure 3: ρ(t1 , . . . , tm ) = { i ∈ I | ρ′ (t1 (i), . . . , tm (i)) } where partition(B1 , . . . , Bn ) denotes a boolean algebra expression expressing that sets B1 , . . . , Bn form the partition of trueI . It is easy to see that 11 and 12 are equivalent. By repeating this construction we eliminate all term quantifiers from a formula. We then eliminate all set quantifiers as in Section 3.2. For that purpose we extend the language with cardinality constraints. As the result we obtain cardinality constraints on inner formulas. Closed inner formulas evaluate to trueI or falseI depending on their truth value in base structure C. Hence, if C is decidable, so is P2 . Theorem 16 (Feferman-Vaught) Let C be a decidable structure. Then every formula in the language of Figure 2 is equivalent on the structure P2 to a propositional combination of cardinality constraints of the index-set boolean algebra i.e. formulas of form |φ| ≥ k and |φ| = k where φ is an inner formula. (9) Sort constraints imply that quantifiers ∃I , ∀I are only applied to inner formulas. Let φ be a formula of sort bool. By labelling subformulas of sort indset with variables A1 , . . . , An , we can write φ in form φ1 : Example 17 Let r ∈ LC be a binary relation on structure C. Let us eliminate quantifier ∃t from the formula φ(t1 , t2 ): ∃t.∃L A1 , A2 , A3 . ∃L A1 , . . . , An . A 1 =I φ1 ∧ . . . ∧ A n =I φn ∧ A1 =I r(t, t1 ) ∧ A2 =I r(t1 , t) ∧ A3 =I r(t2 , t) ∧ ψ(A1 , . . . , An ) |¬I A1 | = 0 ∧ |¬I A2 | = 0 ∧ |¬I A3 | ≥ 1 where We first introduce sets B0 , . . . , B7 that form partition of trueI . The formula is then equivalent to φ1 : φ = ψ(φ1 , . . . , φn ) Furthermore, by defining B1 , . . . , Bm to be the partition of trueI consisting of terms of form ∃t.∃L B0 , B1 , B2 , B3 , B4 , B5 , B6 , B7 . B0 =I r(t, t1 ) ∧I r(t1 , t) ∧I r(t2 , t) ∧ Ap11 ∧I . . . ∧I Apnn B1 =I ¬I r(t, t1 ) ∧I r(t1 , t) ∧I r(t2 , t) ∧ ′ for p1 , . . . , pn ∈ {0, 1}, we can find a formula ψ and formulas φ′1 , . . . , φ′m such that φ1 is equivalent to φ2 : B2 =I r(t, t1 ) ∧I ¬I r(t1 , t) ∧I r(t2 , t) ∧ B3 =I ¬I r(t, t1 ) ∧I ¬I r(t1 , t) ∧I r(t2 , t) ∧ ∃L B1 , . . . , Bm . B1 =I φ′1 ∧ . . . ∧ Bm =I φ′m ∧ B4 =I r(t, t1 ) ∧I r(t1 , t) ∧I ¬I r(t2 , t) ∧ (10) B5 =I ¬I r(t, t1 ) ∧I r(t1 , t) ∧I ¬I r(t2 , t) ∧ ψ ′ (B1 , . . . , Bn ) B6 =I r(t, t1 ) ∧I ¬I r(t1 , t) ∧I ¬I r(t2 , t) ∧ φ′1 , . . . , φ′m evaluate to sets that form partition of and where trueI for all values of free variables. (By partition of trueI we here mean a family of pairwise disjoint sets whose union is trueI , but we do not require the sets to be non-empty.) Now consider a formula of form ∃t.φ where φ is without ∃, ∀ quantifiers (but possibly contains ∃I , ∀I and ∃L , ∀L quantifiers). We transform φ into φ2 as described, and then replace ∃t. ∃L B1 , . . . , Bm . B1 = I φ′1 ∧ . . . ∧ Bm = I φ′m ∧ B7 =I ¬I r(t, t1 ) ∧I ¬I r(t1 , t) ∧I ¬I r(t2 , t) ∧ φ0 where φ0 ≡ |B1 | = 0 ∧ |B2 | = 0 ∧ |B3 | = 0 ∧ |B5 | = 0 ∧ (11) |B6 | = 0 ∧ |B7 | = 0 ∧ ′ ψ (B1 , . . . , Bn ) |B4 | ≥ 1 with We now eliminate the quantifier ∃t from the formula φ1 , L L ∃ D1 , . . . , Dm . ∃ B1 , . . . , Bm . D1 =I (∃I t.φ′1 ) ∧ . . . ∧ Dm =I (∃I t.φ′m ) ∧ B1 ⊆I D1 ∧ . . . ∧ Bm ⊆I Dm ∧ (12) partition(B1 , . . . , Bn ) ∧ ψ ′ (B1 , . . . , Bn ) 10 B0 ⊆I D0 ∧ . . . ∧ B7 ⊆I D7 ∧ problems are special cases of decision problem for first-order theory of term algebras, for a survey see e.g. [45, 9]. We believe that our proof provides some insight into different variations of quantifier elimination procedures for term algebras. Like [22] we use selector language symbols, but retain the usual constructor symbols as well. The advantage of the selector language is that ∃y. z = f (x, y) is equivalent to a quantifier-free formula x = f1 (z) ∧ Isf (z). On the other hand, constructor symbols also increase the set of relations on terms definable via quantifier-free formulas, which can slightly simplify quantifier-elimination procedure, as will be seen by comparing Proposition 34 and Proposition 38. Compared to [22, Page 70], we find that the termination of our procedure is more evident and the extension to the term-power algebra in Section 6 easier. Our base formulas somewhat resemble formulas arising in other quantifier elimination procedures [31, 11, 30]. Our terminology also borrows from congruence closure graphs like those of [39, 38], although we are not primarily concerned with efficiency of the algorithm described. Term algebra is an example of a theory of pairing functions, and [15] shows that non-empty family of theories of pairing functions as nonelementary lower bound on time complexity. φ0 3.4.1 obtaining formula φ2 : ∃L D0 , D1 , D2 , D3 , D4 , D5 , D6 , D7 . D0 =I ∃I t. r(t, t1 ) ∧I r(t1 , t) ∧I r(t2 , t) ∧ D1 =I ∃I t. ¬I r(t, t1 ) ∧I r(t1 , t) ∧I r(t2 , t) ∧ D2 =I ∃I t. r(t, t1 ) ∧I ¬I r(t1 , t) ∧I r(t2 , t) ∧ D3 =I ∃I t. ¬I r(t, t1 ) ∧I ¬I r(t1 , t) ∧I r(t2 , t) ∧ D4 =I ∃I t. r(t, t1 ) ∧I r(t1 , t) ∧I ¬I r(t2 , t) ∧ D5 =I ∃I t. ¬I r(t, t1 ) ∧I r(t1 , t) ∧I ¬I r(t2 , t) ∧ D6 =I ∃I t. r(t, t1 ) ∧I ¬I r(t1 , t) ∧I ¬I r(t2 , t) ∧ D7 =I ∃I t. ¬I r(t, t1 ) ∧I ¬I r(t1 , t) ∧I ¬I r(t2 , t) ∧ φ3 where φ3 ≡ ∃L B0 , B1 , B2 , B3 , B4 , B5 , B6 , B7 . To facilitate quantifier elimination we use a selector language Sel(Σ) for term algebra [22, Page 61]. We define term algebra in selector language as a first-order structure with partial functions. The set Sel(Σ) contains, for every function symbol f ∈ Σ of arity ar(f ) = k, a unary predicate Isf ⊆ FT(Σ) and functions f1 , . . . , fk : FT(Σ) → FT(Σ) such that We next apply quantifier elimination for boolean algebras to formula φ3 and obtain formula φ′3 : φ′3 ≡ |D4 | ≥ 1 ∧ |¬I D0 ∧I ¬I D4 | = 0 Hence φ(t1 , t2 ) is equivalent to ∃L D0 , D4 . Isf (t) fi (f (t1 , . . . , tk )) fi (t) D0 =I ∃I t. r(t, t1 ) ∧I r(t1 , t) ∧I r(t2 , t) ∧ D4 =I ∃I t. r(t, t1 ) ∧I r(t1 , t) ∧I ¬I r(t2 , t) ∧ |D4 | ≥ 1 ∧ |¬I D0 ∧I ¬I D4 | = 0 ⇐⇒ = = ∃t1 , . . . , tk . t = f (t1 , . . . , tk ()13) ti , 1≤i≤k (14) ⊥, ¬Isf (t) (15) For every f ∈ Σ and 1 ≤ i ≤ ar(f ), expression fi (t) defined iff Isf (t) holds, so we let Df = hx, Isf (x)i. As a special case, if d is a constant, then ar(d) = 0 and Isd (t) ⇐⇒ t = d. After substituting the definitions of D0 and D4 , formula φ(t1 , t2 ) can be written without quantifiers ∃, ∀, ∃L , ∀L . Proposition 18 For every formula φ1 in the language Cons(Σ) there exists an equivalent formula φ2 in the selector language.  3.4 Term Algebra in Selector Language Term Algebras Proof Sketch. Because of the presence of equality symbol, every formula in language Cons(Σ) can be written in unnested form such that every atomic formula is of two forms: x1 = x2 , or f (x1 , . . . , xk ) = y, where y and xi are variables. We keep every formula x1 = x2 unchanged and transform each formula In this section we present a quantifier elimination procedure for term algebras (see Section 2.1). A quantifier elimination procedure for term algebras implies that the first-order theory of term algebras is decidable. In the sections below we build on the procedure in this section to define quantifier elimination procedures for structural subtyping. The decidability of the first-order theory of term algebras follows from Mal’cev’s work on locally free algebras [31, Chapter 23]. [39] also gives an argument for decidability of term algebra and presents a unification algorithm based on congruence closure [38]. Infinite trees are studied in [12]. [30] presents a complete axiomatization for algebra of finite, infinite and rational trees. A proof in the style of [22] for an extension of free algebra with queues is presented in [43]. Decidability of an extension of term algebras with membership tests is presented in [10] in the form of a terminating term rewriting system. Unification and disunification f (x1 , . . . , xk ) = y into the well-defined conjunction x1 = f1 (y) ∧ · · · ∧ xk = fk (y) ∧ Isf (y) Note that predicates Isf form a partition of the set of all terms i.e. the following formulas are valid: W Isf (x) ∀x. f ∈Σ (16) ∀x. ¬(Isf (x) ∧ Isg (x)), for f 6≡ g 11 ¬, ∧, ∨ for some function j : {1, . . . , m} → {1, . . . , n}. We require each semi-base formula to satisfy the following congruence closure property: there are no two distinct variables ui and ui′ such that both ui = f (ul1 , . . . , ulk ) and ui′ = f (ul1 , . . . , ulk ) occur as conjuncts φj in formula structure. We denote by U the set of internal variables of a given semi-base formula, U = {u1 , . . . , un }. ∃ Proposition 28 ✲ quantifier-free formula ✲ ✛ disjunction of base formulas ✛ Proposition 34 Definition 20 A semi-base formula in selector language is obtained from the base formula in constructor language by replacing every conjunct of form Figure 4: Quantifier Elimination for Term Algebra ui = f (ul1 , . . . , ulk ) A constructor-selector language contains both constructor symbols f ∈ Cons(Σ) and selector symbols fi ∈ Sel(Σ). with the well-defined conjunction Isf (ui ) ∧ ul1 = f1 (ui ) ∧ · · · ∧ ulk = fk (ui ) 3.4.2 Quantifier Elimination A semi-base formula in selector language is clearly a wellformed conjunction of literals. All atomic formulas in a semibase formula are unnested, in both constructor and selector language. We can represent a base formula as a labelled directed graph with the set of nodes U ; we call this graph graph associated with a semi-base formula. Nodes of the graph are in a bijection with internal variables of the semi-base formula. We call nodes corresponding to parameter variables up+1 , . . . , up+q parameter nodes; nodes u1 , . . . , up are nonparameter nodes. Each non-parameter node is labelled by a function symbol f ∈ Σ and has exactly ar(f ) successors, with edge from uk to ul labelled by the positive integer i iff fi (uk ) = ul occurs in the semi-base formula written in selector language. A constant node is a node labelled by some constant symbol c ∈ Σ, ar(c) = 0. A constant node is a sink in the graph; every sink is either a constant or a parameter node. In addition to the labelling by function symbols, each node u ∈ U of the graph is labelled by zero or more free variables x such that equation x = u occurs in the semi-base formula. We proceed to quantifier elimination for term algebra. A schematic view of our proof is in Figure 4. The basic insight is that any quantifier-free formula can be written in a particular unnested form, as a disjunction of base formulas. Base formulas trivially permit elimination of an existential quantifier, yet every base formula can be converted back to a quantifier-free formula. A semi-base formula is almost the base formula, except that it may be cyclic. We introduce cyclicity after explaining the graph representation of a semi-base formula. Definition 19 (Semi-Base Formula) A semi-base formula β with • free variables x1 , . . . , xm , • internal non-parameter variables u1 , . . . , up , and • internal parameter variables up+1 , . . . , up+q is a formula of form ∃u1 , . . . , un Definition 21 (Base Formula) A semi-base formula φ is a base formula iff the graph associated with φ is acyclic. distinct(u1 , . . . , un ) ∧ structure(u1 , . . . , un ) ∧ A semi-base formula whose associated graph is cyclic is unsatisfiable in the term algebra of finite terms. Checking the cyclicity of a base formula corresponds to occur-check in unification algorithms (see e.g. [29, 11]). labels(u1 , . . . , un ; x1 , . . . , xm ) distinct(u1 , . . . , un ) enforces that variables are distinct ^ ui 6= uj . distinct(u1 , . . . , un ) ≡ Definition 22 By height H(u) of a node u in the acyclic graph we mean the length of the longest path starting from u. 1≤i<j≤n structure(u1 , . . . , un ) specifies relationships between terms denoted by variables: A node u is sink iff H(u) = 0. structure(u1 , . . . , un ) ≡ p V Definition 23 We say that an internal variable ul is a source variable of a base formula β iff ul is represented by a node that is source in the directed acyclic graph corresponding to β. Equivalently, if β is written in the selector language, then ul is a source variable iff β contains no equations of form ul = fi (uk ). ui = ti (u1 , . . . , un ) i=1 where each ti (u1 , . . . , un ) is a term of form f (ul1 , . . . , ulk ) for f ∈ Σ, k = ar(f ). labels(u1 , . . . , un ; x1 , . . . , xm ) identifies some free variables with some parameter and non-parameter variables: ^ xi = uji labels(u1 , . . . , un ; x1 , . . . , xm ) ≡ Definition 24 If ui and uj are internal variables, we write ui ֌∗ uj if there is a path in the underlying graph from node ui to node uj . Equivalently, ui ֌∗ uj iff there exists a term t(ui ) in the selector language such that |= β ⇒ uj = t(ui ). 1≤i≤m 12 Relation ֌∗ is a partial order on internal variables of β. The following Lemma 25 is similar to the Independence of Disequations Lemma in e.g. [10, Page 178]. Proof. Consider a formula ∃x.β where β is a base formula. The only place where x occurs in β is x = us1 in the subformula labels. By dropping the conjunct x = us1 from β we obtain a base formula β1 where β1 is equivalent to ∃x.β. Lemma 25 Let β be a base formula of the form ∃u1 , . . . , up , up+1 , . . . , up+q . β0 Proposition 28 (Quantifier-Free to Base) Every welldefined quantifier-free formula in constructor-selector language can be written as true, false, or a disjunction of base formulas. where up+1 , . . . , up+q are parameter variables of β, and β0 is quantifier-free. Let Sp+1 , . . . , Sp+q be infinite sets of terms. Then there exists a valuation σ such that Jβ0 Kσ = true and Jui Kσ ∈ Si for p + 1 ≤ i ≤ p + q. Proof Sketch. Let φ be a well-defined quantifier-free formula in constructor-selector language. By Proposition 8 we can transform φ into an equivalent formula in disjunctive normal form ψ1 ∨ · · · ∨ ψp Proof. To construct σ assign first the values to parameter variables, as follows. Let hG be the length of the longest path in the graph associated with β. Pick σ(up+1 ) ∈ Sp+1 so that h(σ(up+1 )) > hG , and for each i where p + 2 ≤ i ≤ p + q pick σ(ui ) ∈ Si so that h(σ(ui )) > h(σ(ui−1 )) + hG . The set of heights of an infinite set of terms is infinite, so it is always possible to choose such σ(ui ). Next consider internal nodes u1 , . . . , up+q in some topological order. For each non-parameter node ui such that ui = f (ul1 , . . . , ulk ) occurs in β0 , let σ(ui ) = f (σ(uli ), . . . , σ(ulk )). Finally assign the values to free variables by σ(x) = σ(u) where x = u occurs in β0 . By construction, JstructureKσ = true and JlabelsKσ = true. It remains to show JdistinctKσ = true i.e. σ(ui ) 6= σ(uj ) for 1 ≤ i, j ≤ p + q, i 6= j. We show this property of σ by induction on m = min(H(ui ), H(uj )). Without loss of generality we assume H(ui ) ≤ H(uj ). Consider first the case m = 0. Then ui is a parameter or a constant node. If ui is a constant and uj is a non-parameter variable then ui and uj are labelled by different function symbols so σ(ui ) 6= σ(uj ). If ui is a constant and uj is a parameter variable then h(σ(ui )) = 0 whereas h(σ(uj )) > hG ≥ 0. Consider the case where ui is a parameter variable and uj is a non-parameter variable. Let where each ψi is a well-defined conjunction of literals. Consider an arbitrary ψi . There exists an unnested quantifierfree formula ψi′ with additional fresh free variables x1 , . . . , xq such that ψi is equivalent to ∃x1 , . . . , xq . ψi′ By distributivity and (6) it suffices to transform each conjunction of unnested formulas into disjunction of base formulas. In the sequel we will assume transformations based on distributivity and (6) are applied whenever we transform conjunction of literals into a formula containing disjunction. We also assume that every equation f (x1 , . . . , xn ) = y is replaced by the equivalent one y = f (x1 , . . . , xn ) and every equation fi (x) = y is replace by y = fi (x). Because of our assumption that Σ is finite, we can eliminate every literal of form ¬Isf (x) using the equivalence _ Isg (x) (17) ¬Isf (x) ⇐⇒ g∈Σ\{f } which follows from (16). We then transform formula back into disjunctive normal form and propagate the existential quantifiers to the conjunctions of literals. We may therefore assume that there are no literals of form ¬Isf (x) in the conjunction. Furthermore, Isf (x) ∧ Isg (x) ⇐⇒ false for f 6≡ g, so we may assume that for variable x there is at most one literal Isf (x) for some f . If fi (x) occurs in the conjunction, because the conjunction is well-defined, we may always add the conjunct Isf (x). This way we ensure that exactly one literal of form Isf (x) occurs in the conjunction. We next ensure that every variable has either none or all of its components named by variables. If the conjunction contains literal Isf (x) but does not contain x = f (x1 , . . . , xn ) and does not contain an equation of form y = fi (x) for every i, 1 ≤ i ≤ ar(f ), we introduce a fresh existentially quantified variable for each i such that a term of form y = fi (x) does not appear in the conjunction. At this point we may transform the entire conjunction into constructor language by replacing J = {j1 | uj1 is a parameter variable s.t. uj ֌∗ uj1 } If J = ∅, then β0 uniquely specifies σ(uj ), and h(σ(uj )) = H(uj ) ≤ hG < h(σ(ui )) Let J 6= ∅ and j0 = max J. If i ≤ j0 , then h(σ(ui )) ≤ h(σ(uj0 )) < h(σ(uj )) If j0 < i then h(σ(uj )) ≤ h(σ(uj0 )) + hG < h(σ(uj0 +1 )) ≤ h(σ(ui )) Now consider the case m > 0. ui and uj are nonparameter nodes, so let ui = f (ui1 , . . . , uik ) and uj = g(uj1 , . . . , ujl ). If f 6= g then clearly σ(ui ) 6= σ(uj ). Otherwise, by congruence closure property of base formulas, there exists d such that uid 6= ujd . Then by induction hypothesis σ(uid ) 6= σ(ujd ), so σ(ui ) 6= σ(uj ). Isf (ui ) ∧ vl1 = f1 (ui ) ∧ · · · ∧ vlk = fk (ui ) with ui = f (vl1 , . . . , vlk ) for k = ar(f ). We next ensure that for every two variables x1 and x2 occurring in the conjunction exactly one of the conjunct x1 = x2 or x1 6= x2 is present. Namely if both conjuncts x1 = x2 and x1 6= x2 are present, the conjunction is false. If none of the conjuncts is present, we insert the disjunction Corollary 26 Every base formula is satisfiable. Proposition 27 (Quantification of Base Formula) If β is a base formula and x a free variable in β, then there exists a base formula β1 equivalent to ∃x.β. 13 x1 = x2 ∨ x1 6= x2 as one of the conjuncts and transform the result into disjunction of existentially quantified conjunctions. We next perform congruence closure for finite terms [38] on the resulting conjunction, using the fact that equality is reflexive, symmetric, transitive and congruent with respect to free operations f ∈ Cons(Σ) and that t(x) 6= x for every term t 6≡ x. Syntactically, the result of congruence closure can be viewed as adding new equations to the conjunction. If the congruence closure procedure establishes that the formula is unsatisfiable, the result is false. Otherwise, all variables are grouped into equivalence classes. If a u1 = u2 occurs in the conjunction where both u1 and u2 are internal variables, we replace u1 with u2 in the formula and eliminate the existential quantifier. If for some free variable x there is no internal variable u such that conjunction x = u occurs, we introduce a new existentially quantified variable and a conjunct x = u. These transformations ensure that for every equivalence class there exists exactly one internal variable in the formula. It is now easy to pick representative conjuncts from the conjunction to obtain conjunction of the syntactic form in Definition 19 of semi-base formula. The resulting formula is a base formula because congruence closure algorithm ensures that the associated graph is acyclic. every variable u we obtain a quantifier-free formula equivalent to β. Definition 31 Let β be a base formula. The set determined of determined variables of β is the smallest set S that contains the set covered and satisfies the following condition: if u is a non-parameter node and all successors u1 , . . . , uk (k ≥ 0) of u in the associated graph are in S, then u is also in S. In particular, every constant node is determined. A parameter node w is determined iff w is covered. Lemma 32 If a node u is not determined, then there exists an uncovered parameter node v such that u ֌∗ v. Proof. The proof is by induction on H(()u). If H(()u) = 0 then u has no successors, and u cannot be a constant node because it is not determined. Therefore, u is a parameter node, so we may let v ≡ u. Assume that the statement holds for for every node u′ such that H(()u′ ) = k and let H(()u) = k + 1. Because u is not determined, there exists a successor u′ of u such that u′ is not determined, so by induction hypothesis there exists an uncovered parameter node v such that u′ ֌∗ v. Hence u ֌∗ u′ ֌∗ v. We next turn to the problem of transforming a base formula into a quantifier-free formula. We will present two constructions. The first construction yields a quantifier-free formula in constructor-selector language and is sufficient for the purpose of quantifier elimination. The second construction yields a quantifier-free formula in selector language and is slightly more involved; we present it to provide additional insight into the quantifier elimination approach to term algebras. We first introduce notions of covered and determined variables of a base formula β. The basic idea behind these notions is that β implies a functional dependence from the free variables of β to each of the determined variables. In both constructions we use the notion of a a covered variable, which denotes a component of a term denoted by some free variable. In the first construction we also use the notion of determined variable, which includes covered variables as well as variables constructed from covered variables using constructor operations f ∈ Cons(Σ). Lemma 33 Every base formula β is equivalent to a base formula β ′ obtained from β by eliminating all nodes that are not determined. Proof. Construct β ′ from β by eliminating all terms containing a variable u ∈ U \ determined and eliminating the corresponding existential quantifiers. Then all variables in β ′ are determined. β ′ has fewer conjuncts than β, so |= β ⇒ β ′ . To show |= β ′ ⇒ β, let σ be any assignment of terms to determined variables of β such that β evaluate to true under σ. As in the proof of Lemma 25, define the extension σ ′ of σ as follows. Choose sufficiently large values σ ′ (v) for every uncovered sink variable v, so that σ ′′ defined as the unique extension of σ ′ to the remaining undetermined variables assigns different terms to different variables. This is possible because the term model is infinite. The resulting assignment σ ′′ satisfies the matrix of the base formula β. Therefore, |= β ′ ⇒ β, so β and β ′ are equivalent base formulas. Definition 29 Consider an arbitrary base formula β. We say that an internal variable u is covered by a free variable x iff x = u′ occurs in β for some u′ such that u ֌∗ u′ . An internal variable u is covered iff u is covered by x for some free variable x (in particular, if x = u occurs in β then u is covered). Let covered denote the set of covered internal variables of base formula, and let uncovered = U \ covered where U is the set of all internal variables of β. First Construction Proposition 34 (Base to Constructor-Selector) Every base formula β is equivalent to a quantifier-free formula φ in constructor-selector language. Proof. By Lemma 33 we may assume that all variables in β are determined. To every variable u we assign a term τ (u). Term τ (u) is in constructor-selector language and the variables of τ (u) are among the free variables of β. If u ∈ covered, we assign τ (u) as in the proof of Lemma 30. If u1 , . . . , uk are the successors of a determined node u, we put τ (u) = f (τ (u1 ), . . . , τ (uk )) Lemma 30 (Covered Base to Selector) Every base formula without uncovered variables is equivalent to a quantifier free formula in selector language. Proof. Consider a base formula β where every variable is covered. Consider an arbitrary quantified variable u. Because u is covered, there exists variable x free in β such that u = t(x) for some term t in the selector language. Replace every occurrence of u in the matrix of β by t(x) and eliminate the quantification over u. Repeating this process for where f is the label of node u. This definition uniquely determines τ (u) for all u ∈ determined. We obtain the quantifier-free formula φ by replacing every variable u with τ (u) and eliminating all quantifiers. 14 For every u we have |= β ⇒ u = τ (u), so |= β ⇒ φ. Conversely, if φ is satisfied then τ defines an assignment for u variables which makes the matrix of β true. Therefore β and φ are equivalent. and determined, it is not a parameter node. We show how to eliminate u0 without introducing new uncovered variables. Our goal is to eliminate u0 from the associated graph. We need to preserve information that u0 is distinct from variables u ∈ U \ {u0 } in the graph. We consider two cases. If u is not a parameter node, then by congruence closure either u0 and u are labelled by different function symbols, or they are labelled by the same function symbol f ∈ Σ with ar(f ) = k and there exists i, 1 ≤ i ≤ k and variables ui = fi (u0 ) and u′i = fi (u′ ) such that ui 6≡ u′i . Hence the constraint u0 6= u is deducible from the inequalities of other variables in β and we can eliminate u0 without changing the truth value of β. Next consider the case when u is a parameter node. By assumption u is determined, and because it is parameter, it is covered. We then perform parameter node expansion as described above. The result of elimination of u0 in β is a disjunction of base formulas β ′ , in each β ′ every parameter node is expanded. If u is a parameter node in β then the constraint u0 6= u is preserved in each β ′ because u is not a parameter node in β ′ so the previous argument applies. Because the parameter nodes being expanded are covered, so are their successor nodes introduced by parameter expansion. Therefore, by repeatedly applying elimination of uncovered variables for every uncovered variable u0 , we obtain a disjunction ∆ of formulas β ′ where each β ′ has no uncovered variables, and ∆ is equivalent to β. Second Construction The reason for using constructor symbols f ∈ Cons(Σ) in the first construction is to preserve the constraints of form u 6= v when eliminating node u with successors u1 , . . . , uk . Using constructor symbols we would obtain the constraint f (u1 , . . . , uk ) 6= v. Our second construction avoids introducing constructor operations by decomposing f (u1 , . . . , uk ) 6= v into disjunction of inequalities of form ui 6= fi (v). When v is a parameter node, the presence of term fi (v) potentially requires introducing a new node in the associated graph, we call this process parameter expansion. Parameter expansion may increase the total number of nodes in the graph, but it decreases the number of uncovered nodes, so the process of converting a base formula to a quantifier-free formula in the selector language terminates. Lemma 35 Let β be an arbitrary base formula. 1. If u is covered and u ֌∗ u′ then u′ is covered as well. 2. If u′ is uncovered and u′ is not a source, then there exists u 6≡ u′ such that u ֌∗ u′ and u is also uncovered. 3. If β contains an uncovered variable then β contains an uncovered variable that is a source. Proposition 38 (Base to Selector) For every base formula β there exists an equivalent quantifier-free formula ψ in selector language. Proof. By definition. Proof. By Lemma 37, β is equivalent to a disjunction β1 ∨ · · · ∨ βn where each βi has no uncovered variables. By Lemma 30, each βi is equivalent to some quantifier free formula ψi , so β is equivalent to the quantifier-free formula ψ1 ∨ · · · ∨ ψn . Parameter Expansion We define the operation of expanding a parameter node in a base formula as follows. Let β be an arbitrary base formula and w a parameter variable in β. The result of expansion of w is a disjunction of base formulas β ′ generated by applying (13) to w. In each of the resulting formulas β ′ variable w is not a parameter any more. Each β ′ contains Isf (w) for some f ∈ Σ and node w has successors u1 , . . . , uk for k = ar(f ). Each successor ui is either an existing internal variable or a fresh variable. For a given β, sink expansion generates disjunction of formulas β ′ for every choice of f ∈ Σ and every choice of successors ui , subject to congruence closure so that β ′ is a base formula: we discard the choices of successors of w that yield formulas β ′ violating congruence of equality. (This process is similar to converting quantifier-free formulas into disjunction of base formulas in the proof of Proposition 28.) The following lemma shows the correctness of parameter expansion. The final theorem in this section summarizes quantifier elimination for term algebra. Theorem 39 (Term Algebra Quantifier Elimination) There exist algorithms A, B, C such that for a given formula φ in constructor-selector language of term algebras: a) A produces a quantifier-free formula φ′ in constructorselector language b) B produces a quantifier-free formula φ′ in selector language c) C produces a disjunction φ′ of base formulas Lemma 36 (Parameter expansion soundness) Let ∆ = β1′ ∨ · · · βk′ be the disjunction generated by parameter expansion of a base formula β. Then ∆ is equivalent to β. Proof. a): Transform formula φ into prenex form Lemma 36 justifies the use of parameter expansion in the following Lemma 37. where φ∗ is quantifier free, as in Section 3.1. We eliminate the innermost quantifier Qn as follows. Suppose first that Qn is ∃. Transform the matrix φ∗ into disjunctive normal form C1 ∨ · · · ∨ Cn . By Proposition 28, transform C1 ∨ · · · ∨ Cn into disjunction β1 ∨ · · · ∨ βm of base formulas. Then propagate ∃ into individual disjuncts, using Q1 x1 . . . Qn−1 xn−1 Qn xn .φ∗ Lemma 37 Every base formula β can be written as a disjunction of base formulas without uncovered variables. Proof Sketch. By Lemma 33 we may assume that all variables of β are determined. Suppose β contains an uncovered variable. Then by Lemma 35, β contains an uncovered variable u0 such that u0 is a source. Because u0 is uncovered ∃xn . β1 ∨ · · · ∨ βm ⇐⇒ (∃xn .β1 ) ∨ · · · ∨ (∃xn .βm ) 15 Consider a signature Σ = {a, b, g} where a and b are constant symbols and g is a function symbol of arity 2. We define a partial order ≤ on the set FT(Σ) of ground terms over Σ as the least reflexive partial order relation ρ satisfying By Proposition 27, an existentially quantified base formula is again a base formula, so ∃xn .βi ⇐⇒ βi′ for some βi′ . We thus obtain the ′ Q1 x1 . . . Qn−1 xn−1 . β1′ ∨ · · · ∨ βm (18) 1. a ρ b; By Proposition 34, every base formula is equivalent to a quantifier-free formula in selector language, so 18 is equivalent to Q1 x1 . . . Qn−1 xn−1 .ψ 2. (s1 ρ t1 ) ∧ (s2 ρ t2 ) ⇒ g(s1 , s2 ) ρ g(t1 , t2 ). The structure with equality in the language {a, b, g, ≤}, where ≤ is interpreted as above and a, b, g are interpreted as free operations on term algebra corresponds to the structural subtyping with two base types a and b and one binary type constructor g, with g covariant in both arguments. We denote this structure by BS. We proceed to show that BS admits quantifier elimination and is therefore decidable. where ψ is a quantifier free formula. Hence, we have eliminated the innermost existential quantifier. Next consider the case when Qn is ∀. Then φ is equivalent to Q1 x1 . . . Qn−1 xn−1 ¬∃xn .¬φ∗ Apply the procedure for eliminating xn to ¬φ∗ . The result is formula of form Q1 x1 . . . Qn−1 xn−1 . ¬ψ 4.1 Boolean Algebras on Equivalent Terms In preparation for the quantifier elimination procedure we define certain operations and relations on terms. We also establish some fundamental properties of the structure BS. Define a new signature Σ0 = {cs , g s } as an abstraction of signature Σ = {a, b, g}. Define function shapified : Σ → Σ0 by shapified(a) = cs (19) where ψ is quantifier free. But ¬ψ is also quantifier free, so we have eliminated the innermost universal quantifier. By repeating this process we eliminate all quantifiers, yielding the desired formula φ′ . The direct construction for showing b) is analogous to a), but uses Proposition 38 in place of Proposition 34. To show c), apply e.g. construction a) to obtain a quantifierfree formula ψ and then transform ψ into disjunction of base formulas using Proposition 28. shapified(b) = cs shapified(g) = g s Let ar(shapified(f )) = ar(f ) for each f ∈ Σ; in this case cs is a constant and g s is a binary function symbol. Let FT(Σ0 ) be the set of ground terms over the signature Σ0 . Define shape of a term t, as the function sh : FT(Σ) → FT(Σ0 ), by letting sh(f (t1 , . . . , tk )) = This completes our description of quantifier elimination for term algebras. We remark that there are alternative ways to define base formula. In particular the requirement on disequality of all variables is not necessary. This requirement may lead to unnecessary case analysis when converting a quantifier-free formula to disjunction of base formulas, but we believe that it simplifies the correctness argument. shapified(f )(sh(t1 ), . . . , sh(tk )) for k = ar(f ). In this case we have sh(a) = cs 4 The Pair Constructor and Two Constants sh(b) = cs In this section we give a quantifier elimination procedure for structural subtyping of non-recursive types with two constant symbols and one covariant binary constructor. Two constants corresponds to two primitive types; one binary covariant constructor corresponds to the pair constructor for building products of types. The construction in this section is an introduction to the more general construction in Section 5, where we give a quantifier elimination procedure for any number of constant symbols and relations between them. The construction in this section demonstrates the interaction between the term and boolean algebra components of the structural subtyping. We therefore believe the construction captures the essence of the general result of Section 5. The basic observation behind the quantifier elimination procedure for two constant symbols is that the structure of terms in this language is isomorphic to a disjoint union of boolean algebras with some additional term structure connecting elements from different boolean algebras. As we argue below, the structural subtyping structure contains one copy of boolean algebra for every equivalence class of terms that have the same “shape” i.e. are same up to the constants in the leaves. sh(g(t1 , t2 )) = g s (sh(t1 ), sh(t2 )) Define t1 ∼ t2 iff sh(t1 ) = sh(t2 ). Then ∼ is the smallest equivalence relation ρ such that 1. a ρ b; 2. (s1 ρ t1 ) ∧ (s2 ρ t2 ) ⇒ g(s1 , s2 ) ρ g(t1 , t2 ). For every term t define the word tCont(t) ∈ {0, 1}∗ by letting tCont(a) = 0 tCont(b) = 1 tCont(f (t1 , t2 )) = tCont(t1 ) · tCont(t2 ) The set of all words w ∈ {0, 1}n is isomorphic the boolean algebra of Bn of all subsets of some finite sets of cardinality n, so we write w1 ∩w2 , w1 ∪w2 , wc for operations corresponding to intersection, union, and set complement in the set of words w ∈ {0, 1}n . We write w1 ⊆ w2 for w1 ∩ w2 = w1 . Define function δ by δ(t) = hsh(t), tCont(t)i 16 For term t in any language containing constant symbols, let tLen(t) denote the number of occurrences of constant symbols in t. If w is a sequence of elements of some set, let sLen(w) denote the length of the sequence. Observe that sLen(tCont(t)) = tLen(t) and tLen(sh(t)) = tLen(t). Moreover, t1 ∼ t2 implies sLen(tCont(t1 )) = sLen(tCont(t2 )). Define the set B by a, b :: term g :: term × term → term Isg :: term → bool g1 , g2 :: term → term B = {hs, wi | s ∈ FT(Σ0 ), w ∈ {0, 1}∗ , tLen(s) = sLen(w)} Function δ is a bijection from the set FT(Σ) to the set B. For b1 , b2 ∈ B define b1 ≤ b2 iff δ −1 (b1 ) ≤ δ −1 (b2 ). From the definitions it follows = :: term × term → bool ≤ :: term × term → bool s :: shape s :: shape × shape → shape hs1 , w1 i ≤ hs2 , w2 i ⇐⇒ s1 = s2 ∧ w1 ⊆ w2 Isgs :: shape → bool If g is defined on B via isomorphism δ we also have g1s , g2s :: shape → shape c g sh :: s g(hs1 , w1 i, hs2 , w2 i) = hg (s1 , s2 ), w1 · w2 i = For any fixed s ∈ FT(Σ0 ), the set B(s0 ) = {hs, wi ∈ B | s = s0 } ∩ ,∪ (20) c is isomorphic to the boolean algebra Bn , where n = tLen(s). Accordingly, we introduce on each B(s) the set operations t1 ∩s t2 , t1 ∪s t2 , tc1s . Expressions t1 ∩s t2 and t1 ∪s t2 are defined iff sh(t1 ) = s and sh(t2 ) = s, whereas expression tc1s is defined iff sh(t1 ) = s. We also introduce cardinality expressions as in Section 3.2. If t denotes a term, then the expression |t|s denotes the number of elements of the set corresponding to t. Here we require s = sh(t). We use expressions |t|s = k and |t|s ≥ k as atomic formulas for constant integer k ≥ 0. Note that t1 ≤ t2 ⇐⇒ sh(t1 ) = sh(t2 ) ∧ |t1 ∩ tc2 |sh(t1 ) = 0 t1 = t2 ⇐⇒ sh(t1 ) = sh(t2 ) ∧ |(t1 ∩ tc2 ) ∪ (tc1 ∩ t2 )|sh(t1 ) = 0 1 ,0 | | ≥ k, | | = k 4.2 = g(t1 ∩ t′1 , t2 ∩ t′2 ) g(t1 , t2 ) ∪ g(t′1 , t′2 ) = g(t1 ∪ t′1 , t2 ∪ t′2 ) g(t1 , t2 )c (21) (22) (23) = g(tc1 , tc2 ) If E(x1 , . . . , xn ) denotes an expression consisting only of operations of boolean algebra, then from (4.1) by induction follows that E(g(t11 , t21 ), . . . , g(t1n , t2n )) = :: shape × shape → bool :: shape × term × term → term :: shape × term → term :: shape → term :: shape × term → bool A Multisorted Logic To show the decidability of structure BS, we give a quantifier elimination procedure for an extended structure, denoted FT2 . We use a first-order two-sorted logic with sorts term and shape interpreted over FT2 . The domain of structure FT2 is FT(Σ) ∪ FT(Σ0 ) with elements FT(Σ) having sort term and elements FT(Σ0 ) having sort shape. Variables in Var have term sort, variables in Vars have shape sort. In general, if t denotes an element of FT2 , we write tS to indicate that the element has sort shape. Figure 5 shows operations and relations in FT2 with their sort declarations. The signature is infinite because operations |t|s ≥ k and |t|s = k are parameterized by a nonnegative integer k. We require all terms to be well-sorted. Functions g1 and g2 are interpreted as partial selector functions in the term constructor-selector language, so Dg1 = Dg2 = hhxi, Isg (x)i. Similarly, g1s and g2s are partial selector functions in the shape constructor-selector language, so Dg1s = Dg2s = hhxi, Isgs (x)i. The expressions t1 ∩s t2 and t1 ∪s t2 are defined iff sh(t1 ) = sh(t2 ) = s, and tcs is defined iff sh(t) = s. We therefore let Equation 23 allows decomposing formulas of form |g(t1 , t2 )|s ≥ k into propositional combinations of formulas of form |t1 |s1 ≥ k and |t2 |s2 ≥ k. Note further that the following equations hold: g(t1 , t2 ) ∩ g(t′1 , t′2 ) term → shape Figure 5: Operations and relations in structure FT2 Let sh(t1 ) = s1 , sh(t2 ) = s2 , and s = g s (s1 , s2 ). Then |g(t1 , t2 )|s = |t1 |s1 + |t2 |s2 s D∩s = D∪s = hhy s , x1 , x2 i, sh(x1 ) = y s ∧ sh(x2 ) = y s i g(E(t11 , . . . , t1n ), E(t21 , . . . , t2n )) and (24) D c = hhy s , xi, sh(x) = y s i For atomic formulas |t|s ≥ k and |t|s = k we require atomic formula sh(t) = s to ensure well-definedness: Equations (24) and (23) imply |E(g(t11 , t21 ), . . . , g(t1n , t2n ))| = |E(t11 , . . . , t1n )| + |E(t21 , . . . , t2n )| (25) Boolean algebra B(g s (s1 , s2 )) is isomorphic to the product of boolean algebras B(s1 ) and B(s2 ); the constructor g acts as union of disjoint sets. D| | =k = D| | ≥k = hhy s , xi, sh(x) = y s i Note that the language of Figure 5 subsumes the language {a, b, g, ≤} for the structural subtyping structure. The 17 where each ti is a shape term of form f (usi1 , . . . , usik ) for some f ∈ Σ0 , k = ar(f ), and j : {1, . . . , ms } → {1, . . . , ns } is a function mapping indices of free shape variables to indices of internal shape variables. quantifier-elimination procedure we present in Section 4.3 is therefore sufficient for quantifier elimination in the firstorder logic interpreted over the structural subtyping structure FT2 . 4.3 termBase(u1 , . . . , un , x1 , . . . , xm ) = Quantifier Elimination for Two Constants p V We are now ready to present a quantifier elimination procedure for the structure FT2 . The quantifier elimination procedure is based on the quantifier elimination for term algebras of Section 3.4 as well as the quantifier elimination for boolean algebras of Section 3.2. We first define an auxiliary notion of a us -term as a term formed starting from shape us term variables and shape us constants, using operations ∩us , ∪us , and cus . i=1 cardin(up+1 , . . . , up+q , usps +1 , . . . , usps +qs ) = ψ1 ∧ · · · ∧ ψr t ∩us t′ ∈ Term(us ), t ∪us t′ ∈ Term(us ), and where each ψi is of form tcus ∈ Term(us ) |t(up+1 , . . . , up+q )|us = k Similarly to base formulas of Section 3.4, we define structural base formulas for FT2 structure. A structural base formula contains a copy of a base formula for the shape sort (shapeBase), a base formula for the term sort without term disequalities (termBase), a formula expressing mapping of term variables to shape variables (hom), and cardinality constraints on term parameter nodes of the term base formula (cardin). or |t(up+1 , . . . , up+q )|us ≥ k s for some u -term t(up+1 , . . . , up+q ) that contains no variables other than some of the variables up+1 , . . . , up+q , and the following condition holds: If a variable up+j occurs in term t(up+1 , . . . , up+q ), then sh(up+j ) = us occurs in formula hom. Definition 41 (Structural Base Formula) A structural base formula with: • internal non-parameter term variables u1 , . . . , up ; P0) the graph associated with shape base formula • internal parameter term variables up+1 , . . . , up+q ; ∃us1 , . . . , usns . shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) • free shape variables xs1 , . . . , xsms ; is acyclic (compare to Definition 21); • internal non-parameter shape variables us1 , . . . , usps ; P1) congruence closure property for shapeBase subformula: there are no two distinct variables usi and usj such that both usi = f (usl1 , . . . , uslk ) and usj = f (usl1 , . . . , uslk ) occur as conjuncts in formula shapeBase; • internal parameter shape variables usps , . . . , usps +qs is a formula of form: ∃u1 , . . . , un , us1 , . . . , usns . P2) congruence closure property for termBase subformula: there are no two distinct variables ui and uj such that both ui = f (ul1 , . . . , ulk ) and uj = f (ul1 , . . . , ulk ) occur as conjuncts in formula termBase; shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) ∧ termBase(u1 , . . . , un , x1 , . . . , xm ) ∧ hom(u1 , . . . , un , us1 , . . . , usns ) ∧ P3) homomorphism property of sh: for every non-parameter term variable u such that u = f (ui1 , . . . , uik ) occurs in termBase, if the conjunct sh(u) = us occurs in hom, then for some shape variables usj1 , . . . , usjk the term us = f s (usj1 , . . . , usjk ) occurs in shapeBase where f s = shapified(f ) and for every r where 1 ≤ r ≤ k, conjunct sh(uir ) = usjr occurs in hom. cardin(up+1 , . . . , un , usps +1 , . . . , usns ) where n = p + q, ns = ps + q s , and formulas shapeBase, termBase, hom, and cardin are defined as follows. shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) = m Vs i=1 (26) We require each structural base formula to satisfy the following conditions: • free term variables x1 , . . . , xm ; ∧ distinct(us1 , . . . , usn ) sh(ui ) = usji where j : {1, . . . , n} → {1, . . . , ns } is some function such that {j1 , . . . , jp } ⊆ {1, . . . , ps } and {jp+1 , . . . , jp+q } ⊆ {ps + 1, . . . , ps + q s } (a term variable is a parameter variable iff its shape is a parameter shape variable). 3. if t, t′ ∈ Term(us ), then also i=1 n V i=1 2. 0us , 1us ∈ Term(us ) usi = ti (us1 , . . . , usn ) ∧ xi = uji i=1 hom(u1 , . . . , un , us1 , . . . , usns ) = 1. Var ⊆ Term(us ) s m V where each ti is a term of form f (ui1 , . . . , uik ) for some f ∈ Σ, k = ar(f ), and j : {1, . . . , m} → {1, . . . , n} is a function mapping indices of free term variables to indices of internal term variables. Definition 40 (us -terms) Let us ∈ Vars be a shape variable. The set of us -terms Term(us ) is the least set such that: p V ui = ti (u1 , . . . , un ) ∧ xsi = usji 18 u svz u sw v u vz uv We next use (21) to eliminate atomic formulas t1 ≤ t2 and replace them with cardinality constraints, resulting in formula φ2 equivalent to φ1 : u zv z w uz φ2 ≡ φ2,1 ∧ φ2,2 uw where s uw1 u v1 u z1 uw1 s uw2 u v2 u z2 φ2,1 ≡ |g(v, z) ∩ g(z, v)c |sh(g(v,z)) = 0 ∧ uw2 sh(g(v, z)) = sh(g(z, v)) ∧ (30) Isg (v) ∧ Isg (w) Figure 6: One of the Base Formulas Resulting from (28) and φ2,2 ≡ According to Definition 41 a structural base formula contains no selector function symbols. Formulation using selector symbols is also possible, as in Definition 20. The only partial function symbols occurring in a structural base formula of Definition 41 are in cardin subformula. Condition (26) therefore ensures that functions in cardin and thus the entire base formula are well-defined. Note that acyclicity of shape base formula shapeBase (condition P0) implies acyclicity of term base formula as well. Namely, condition P3 ensures that any cycle in termBase implies a cycle in shapeBase. As in Section 3.4 we proceed to show that each quantifierfree formula can be written as a disjunction of base formulas and each base formula can be written as a quantifier-free formula. We strongly encourage the reader to study the following example because it illustrates the idea behind our quantifierelimination decision procedure.  ¬ |g1 (w) ∩ g1 (v)c |sh(g1 (w)) = 0 ∧ sh(g1 (w)) = sh(g1 (v)) (31) Here we have written e.g. |g(v, z) ∩ g(z, v)c |sh(g(v,z)) = 0 as a shorthand for |g(v, z) ∩sh(g(v,z)) g(z, v)csh(g(v,z)) |sh(g(v,z)) = 0 (In general, we omit term shape arguments for boolean algebra operations if the arguments are identical to the enclosing term shape argument of the cardinality constraint.) We next transform φ2 into disjunction of well-defined conjunctions. Following the ideas in Proposition 8, we transform φ2,2 into φ3,1 ∨ φ3,2 where φ3,1 ≡ Example 42 The following sentence is true in structure FT2 . |g1 (w) ∩ g1 (v)c |sh(g1 (w)) ≥ 1 ∧ sh(g1 (w)) = sh(g1 (v)) (32) and φ3,2 ≡ sh(g1 (w)) 6= sh(g1 (v)) ∀x, y. x ≤ y ⇒ ∃z. z ≤ x ∧ z ≤ y ∧ and then transform φ2,1 ∧ φ2,2 into ∀w. w ≤ x ∧ w ≤ y ⇒ (φ2,1 ∧ φ3,1 ) ∨ (φ2,1 ∧ φ3,2 ) ∀v. g(v, z) ≤ g(z, v) ∧ Isg (v) ∧ Isg (w) ⇒ g1 (w) ≤ g1 (v) (27) An informal proof of sentence (27) is as follows. Suppose that x ≤ y. Then sh(x) = sh(y) = xs . Let z = x ∩xs y. Now consider some w such that w ≤ x and w ≤ y. Then sh(w) = xs , so w ≤ z. Suppose that v is such that g(v, z) ≤ g(z, v). Then by covariance of g we have z ≤ v, so w ≤ v. If we assume Isg (w) and Isg (v), then g1 (w) and g1 (v) are well defined and by covariance of g we conclude g1 (w) ≤ g1 (v), as desired. We now give an alternative argument that shows that sentence (27) is true. This alternative argument illustrates the idea behind our quantifier-elimination decision procedure. For the sake of brevity we perform some additional simplifications along the way that are not part of the procedure we present (although they could be incorporated to improve efficiency), and we skip consideration of some uninteresting cases during the case analyses. Let us first eliminate the quantifier from formula For the sake of brevity we ignore the case φ2,1 ∧ φ3,2 ; it is possible to show that φ2,1 ∧ φ3,2 is equivalent to false in the context of the entire formula. We transform φ2,1 ∧ φ3,1 into unnested form, introducing fresh existentially quantified variables uvz ,uzv ,uw1 ,uv1 , usvz , usw1 that denote terms occurring in φ2,1 ∧ φ3,1 . The result is formula φ4 where φ4 ≡ ∃uvz , uzv , uw1 , uv1 , usvz , usw1 . uvz = g(v, z) ∧ uzv = g(z, v) ∧ uw1 = g1 (w) ∧ uv1 = g1 (v) ∧ usvz = sh(uvz ) ∧ usw1 = sh(uw1 ) ∧ sh(uzv ) = usvz ∧ sh(uv1 ) = usw1 ∧ Isg (v) ∧ Isg (w) ∧ |uvz ∩ uczv |usvz = 0 ∧ |uw1 ∩ ucv1 |usw1 ≥ 1 ∀v. g(v, z) ≤ g(z, v) ∧ Isg (v) ∧ Isg (w) ⇒ g1 (w) ≤ g1 (v) (28) (33) To transform φ4 into disjunction of structural base formulas we keep introducing new existentially quantified variables and adding derived conjuncts to satisfy the invariants of Definition 41. Formula (28) is equivalent to ¬∃v.φ1 where φ1 ≡ g(v, z) ≤ g(z, v) ∧ Isg (v) ∧ Isg (w) ∧ ¬(g1 (w) ≤ g1 (v)) (29) 19 Because Isg (v) and Isg (w) appear in the conjunct, we give names to the remaining successors of v, w, by introducing uw2 = g2 (w), uv2 = g2 (v). We may now write the constraints in constructor language, using e.g. conjunct v = g(uv1 , uv2 ) instead of v, z, w, respectively. The resulting structural base formula is ∃uvz , uzv , uv , uz , uw , uv1 , uv2 , uz1 , uz2 , uw1 , uw2 , usvz , usw , usw1 , usw2 . Isg (v) ∧ uv1 = g1 (v) ∧ uv2 = g2 (v) shapeBase1 ∧ termBase1 ∧ hom1 ∧ cardin1 To ensure that every term variable has an associated shape variable, we introduce fresh variables usv , usw , usz , usw2 , usv2 with conjuncts usv = sh(v), usw = sh(w), usz = sh(z), usw2 = sh(uw2 ), usv2 = sh(uv2 ). Note that base formula contains distinct(us1 , . . . , usn ) subformula. In the case when the current conjunction is not strong enough to entail the disequality between shape variables usi and usj , we perform case analysis, considering the case usi = usj (then usi can be replaced by usj ), and the case usi 6= usj . This case analysis will lead to a disjunction of structural base formulas (unless some of the formulas is shown contradictory in the transformation process). In contrast to shape variables, we do not not perform case analysis for disequality of term variables, because termBase in Definition 41 does not contain a distinct subformula. In this example we perform case analysis on whether usw = usz and usw = usv should hold. For the sake of the example let us consider the case when usw = usz = usv , usv2 = usw2 and usvz , usw , usw1 , usw2 are all distinct. In that case shape variables usw , usz , usv denote the same shape, so let us replace e.g. usz and usv with usw . Similarly, we replace usv2 with usw2 . We obtain conjuncts sh(v) = usw , sh(z) = usw , sh(uv2 ) = usw2 . We next ensure homomorphism property P3 in Definition 41. From conjuncts uvz = g(v, z), sh(uvz ) = usvz , and sh(v) = usw , we conclude where shapeBase1 = usvz = g s (usw , usw ) ∧ usw = g s (usw1 , usw2 ) ∧ distinct(usvz , usw , usw1 , usw2 ) termBase1 = uvz = g(uv , uz ) ∧ uzv = g(uz , uv ) ∧ uv = g(uv1 , uv2 ) ∧ uz = g(uz1 , uz2 ) ∧ uw = g(uw1 , uw2 ) ∧ v = uv ∧ z = uz ∧ w = uw hom1 = sh(uvz ) = usvz ∧ sh(uzv ) = usvz ∧ sh(uv ) = usw ∧ sh(uz ) = usw ∧ sh(uw ) = usw ∧ sh(uv1 ) = usw1 ∧ sh(uz1 ) = usw1 ∧ sh(uw1 ) = usw1 ∧ sh(uv2 ) = usw2 ∧ sh(uz2 ) = usw2 ∧ sh(uw2 ) = usw2 cardin1 = |uv1 ∩ ucz1 |usw1 = 0 ∧ |uz1 ∩ ucv1 |usw1 = 0 ∧ |uv2 ∩ ucz2 |usw2 = 0 ∧ |uz2 ∩ ucv2 |usw2 = 0 ∧ |uw1 ∩ ucv1 |usw1 ≥ 1 usvz = sh(uvz ) = Figure 6 shows a graph representation of the subformulas shapeBase1 , termBase1 , and hom1 of the resulting structural base formula. Recall that we are eliminating the quantification over v from ¬∃v.φ1 . We can now existentially quantify over v. As in Proposition 27, we simply remove the conjunct v = uv from termBase and the quantifier ∃v. As in Figure 4 of Section 3.4 the structural base formula form allows us to eliminate an existential quantifier, whereas the quantifier-free form allows us to eliminate a negation. We transform the structural base formula (35) into a quantifier-free formula as follows. We first use rule (7) to eliminate variable uvz , replacing it with g(v, z). In the resulting formula g(v, z) occurs only in hom1 in the form sh(g(v, z)) = g s (sh(v), sh(z)) = g s (usw , usw ) so we add the conjunct usvz = g s (usw , usw ) to the formula. Similarly, from w = g(uw1 , uw2 ), sh(w) = usw , sh(uw1 ) = usw1 , sh(uw2 ) = usw2 we conclude usw = g(uw1 , uw2 ) and add this conjunct to the formula. Adding these two conjuncts makes property P3 hold. (Note that, had we decided to consider the case where sh(v) 6= sh(z) we would have arrived at a contradiction due to sh(usvz ) = sh(uszv ).) We next apply rule (25) to reduce all cardinality constraints into cardinality constraints on parameter nodes (nodes u for which there there is no conjunct of form u = f (ui1 , . . . , uik )). We replace |uvz ∩ uczv |usvz = 0 with |uv ∩ ucz |usw = 0 ∧ |uz ∩ ucv |usw =0 (35) sh(g(uv , uz )) = usvz (34) (36) But (36) is a consequence of conjuncts usvz = g s (usw , usw ), sh(uv ) = usw and sh(uw ) = usw , so we omit (36) from the formula. In analogous way we eliminate variable uzv and the conjuncts that contain it. We also eliminate uv , analogously to uvz and uzv . In the resulting formula usvz occurs only in distinct subformula of shapeBase. Conjuncts usvz 6= usw , usvz 6= usw1 , and usvz 6= usw2 follow from the remaining conjuncts in shapeBase by acyclicity. Hence we may replace distinct(usvz , usw , usw1 , usw2 ) by distinct(usw , usw1 , usw2 ). Now usvz does not occur in the matrix of the formula, so we may eliminate ∃usvz altogether. Variable v is a parameter variable, but z is not, which prevents application of (25). We therefore introduce uz1 and uz2 such that z = g(uz1 , uz2 ). Because sh(z) = usw , we have sh(uz1 ) = usw1 and sh(uz2 ) = usw2 by homomorphism property. We can now continue applying rule (25) to (34). The result is: |uv1 ∩ ucz1 |usw1 = 0 ∧ |uz1 ∩ ucv1 |usw1 = 0 ∧ |uv2 ∩ ucz2 |usw2 = 0 ∧ |uz2 ∩ ucv2 |usw2 = 0 To make the formula conform to Definition 41 we introduce internal variables uv , uz , uw corresponding to free variables 20 variables are “covered”): The resulting formula is: φ5 ≡ ∃uz , uw , uv1 , uv2 , uz1 , uz2 , uw1 , uw2 , usw , usw1 , usw2 . usw =g s (usw1 , usw2 ) ∧ distinct(usw , usw1 , usw2 ) uz = z ∧ uz = g(uz1 , uz2 ) ∧ uw = g(uw1 , uw2 ) ∧ sh(uz ) = usw ∧ sh(uw ) = usw sh(uv2 ) = |uv1 ∩ ucz1 |usw1 ∧ ∧ sh(uz2 ) = usw2 = 0 ∧ |uz1 ∩ ∧ sh(uw2 ) = ucv1 |usw1 usw2 ∧ uw2 = g2 (w) φ7,2 ≡ sh(uv1 ) = usw1 ∧ sh(uz1 ) = usw1 ∧ sh(uw1 ) = usw1 ∧ |g1 (w) ∩ g1 (z)c |g1s (sh(w)) ≥ 1 |uw1 ∩ ucv1 |usw1 ≥ 1 (44) Our quantifier elimination would also consider the case sh(g2 (w)) 6= sh(g2 (z)). The procedure finds the case contradictory in a larger context, when eliminating ∃z, because sh(z) = sh(x) = sh(w) follows from z ≤ x and w ≤ x. Ignoring this case, we observe that φ7,1 ∨ φ7,2 is equivalent to the quantifier-free formula φ8 , where (38) Note that all variables from φ5,1 belong to B(s) where s is the value of shape variable usw1 (see (20)). This means that we can apply quantifier elimination for boolean algebra (Section 3.2) to eliminate uv1 . The result is (39) Isgs (sh(w)) ∧ Isg (w) ∧ Isg (z) ∧ φ8 ≡ ≥1 sh(z) = sh(w) ∧ |g1 (w) ∩ g1 (z)c |g1s (sh(w)) ≥ 1 (45) Let us therefore assume that the result of quantifier elimination in (28) is ¬φ8 . We proceed to eliminate the next quantifier, ∀w, from Similarly, to eliminate uv2 we consider formula φ5,3 : φ5,3 ≡ ∃uv2 . sh(uv2 ) = usw2 ∧ sh(uz2 ) = usw2 ∧ sh(uw2 ) = usw2 ∧ ∀w. w ≤ x ∧ w ≤ y ⇒ ¬φ8 |uv2 ∩ ucz2 |usw2 = 0 ∧ |uz2 ∩ ucv2 |usw2 = 0 (40) The result of boolean algebra quantifier elimination on φ5,3 is true (indeed, one may let uv2 = uz2 ). The resulting base formula with uv1 and uv2 eliminated is φ6 : (46) (46) is equivalent to ¬∃w. w ≤ x ∧ w ≤ y ∧ φ8 After eliminating ≤ we obtain φ6 ≡ ∃uz , uw , uz1 , uz2 , uw1 , uw2 , usw , usw1 , usw2 . ¬∃w. |w ∩ xc |sh(w) = 0 ∧ sh(x) = sh(w) ∧ usw = g s (usw1 , usw2 ) ∧ distinct(usw , usw1 , usw2 ) ∧ |w ∩ y c |sh(w) = 0 ∧ sh(y) = sh(w) ∧ uz = g(uz1 , uz2 ) ∧ uw = g(uw1 , uw2 ) ∧ sh(uz ) = usw ∧ sh(uw ) = usw ∧ Isgs (sh(w)) ∧ Isg (w) ∧ Isg (z) ∧ sh(z) = sh(w) ∧ g1s (sh(w)) = g2s (sh(w)) ∧ |uv1 ∩ ucz1 |usw1 = 0 ∧ |uz1 ∩ ucv1 |usw1 = 0 ∧ z = uz ∧ w = uw ∧ = g1s (sh(w)) usw2 = g2s (sh(w)) sh(z) = sh(w) ∧ |g1 (w) ∩ g1 (z)c |g1s (sh(w)) ≥ 1 (43) When transforming formula φ4 we chose the case usw1 6= usw1 . If we choose the case usw1 = usw2 , we obtain quantifier-free formula φ7,2 : φ5,1 ≡ ∃uv1 . ucz1 |usw1 = sh(w) distinct(g1s (sh(w)), g2s (sh(w))) (37) We next eliminate uv1 . It suffices to eliminate it from conjuncts where it occurs, so we consider formula φ5,1 : sh(uz1 ) = usw1 ∧ sh(uw1 ) = usw1 ∧ (42) Isgs (sh(w)) ∧ Isg (w) ∧ Isg (z) ∧ φ7,1 ≡ =0∧ |uw1 ∩ ucv1 |usw1 ≥ 1 |uw1 ∩ uw1 = g1 (w) Structural base formula φ6 is therefore equivalent to the quantifier-free formula φ7,1 : |uv2 ∩ ucz2 |usw2 = 0 ∧ |uz2 ∩ ucv2 |usw2 = 0 ∧ φ5,2 ≡ uz2 = g2 (z) usw1 sh(uv1 ) = usw1 ∧ sh(uz1 ) = usw1 ∧ sh(uw1 ) = usw1 ∧ usw2 uz1 = g1 (z) usw z = uz ∧ w = uw ∧ uw = w Isgs (sh(w)) ∧ Isg (w) ∧ Isg (z) ∧ sh(z) = sh(w) ∧ |g1 (w) ∩ g1 (z)c |g1s (sh(w)) ≥ 1 (41) (47) We now proceed similarly as in eliminating variable v. The result is ¬φ9 where sh(uz1 ) = usw1 ∧ sh(uw1 ) = usw1 ∧ sh(uz2 ) = usw2 ∧ sh(uw2 ) = usw2 ∧ φ9 ≡ |uw1 ∩ ucz1 |usw1 ≥ 1 sh(x) = sh(z) ∧ sh(y) = sh(z) ∧ Isg (x) ∧ Isg (y) ∧ Isg (z) ∧ Isgs (sh(z)) ∧ Observe that the equalities in φ6 are sufficient to express all variables bound in φ6 in terms of free variables (all internal (48) c |g1 (x) ∩ g1 (y) ∩ g1 (z) |g1s (sh(z)) ≥ 1 The remaining quantifiers that bind z, y, and x are eliminated similarly. 21 To eliminate the quantifier ∃z, we need to transform ¬φ9 into disjunction of base formulas. This transformation requires negation of φ9 and creates several disjuncts. We consider only the two cases, φ10 and φ11 , that are not contradictory in the enclosing context of conjuncts z ≤ x and z ≤ y: unnested form x = x1 ∩s x2 cardinality constraint |x + (x1 ∩ x2 )|s = 0 x = x1 ∪s x2 |x + (x1 ∪ x2 )|s = 0 x= φ10 ≡ φ11 sh(x) = sh(z) ∧ sh(y) = sh(z) ∧ ¬Isgs (sh(z)) (49) ≡ sh(x) = sh(z) ∧ sh(y) = sh(z) ∧ Isg (x) ∧ Isg (y) ∧ Isg (z) ∧ Isgs (sh(z)) ∧ |x + xc1 |s = 0 xc1s Figure 7: Elimination of Boolean Algebra Unnested Formulas . Expression x + y is a shorthand for (x ∩ y c ) ∪ (y ∩ xc ). (50) c |g1 (x) ∩ g1 (y) ∩ g1 (z) |g1s (sh(z)) = 0 In this case, formula φ6 ∧ψ1 is contradictory. Formula φ6 ∧ψ2 is equivalent to φ′′6 where φ10 is equivalent to sh(x) = cs ∧ sh(y) = cs ∧ sh(z) = cs φ′′6 ≡ ∃uz , uw , uz1 , uz2 , uw1 , uw2 , usw , usw1 , usw2 . (51) usw = g s (usw1 , usw2 ) ∧ distinct(usw , usw1 , usw2 ) ∧ The result of eliminating ∃z from uz = g(uz1 , uz2 ) ∧ uw = g(uw1 , uw2 ) ∧ ∃z. |z ∩ xc |sh(z) = 0 ∧ |z ∩ y c |sh(z) = 0 ∧ φ10 z = uz ∧ w = uw ∧ is therefore sh(uz ) = usw ∧ sh(uw ) = usw ∧ φ10,2 ≡ sh(x) = sh(y) ∧ ¬Isgs (sh(x)) (52) The result of eliminating ∃z from c c ∃z. |z ∩ x |sh(z) = 0 ∧ |z ∩ y |sh(z) = 0 ∧ φ11 is φ11,2 ≡ sh(x) = sh(y) ∧ Isgs (sh(x)) |x ∩ y |sh(x) = 0 ∧ sh(x) = sh(y) ⇒ sh(x) = sh(y) ∪ ∧ sh(uw2 ) = ∧ ≥1∧ (ucz ∩ uw )|usw ≥ 1  We proceed to sketch the general case of quantifier elimination. The following Proposition 44 is analogous to Proposition 27; the proof is again straightforward.  Formulas in the Example 42 do not contain disequalities between terms variables, only disequalities between shape variables. If a conjunction contains disequalities between term variables, we eliminate the disequalities using rule (22) in the process of converting formula to disjunction of structural base formulas. The following Example 43 illustrates this process. Proposition 44 (Quantification of Structural Base) If β is a structural base formula and x a free term variable in β, then there exists a base structural formula β1 equivalent to ∃x.β. The following Proposition 45 corresponds to Proposition 28. Proposition 45 (Quantifier-Free to Structural Base) Every well-defined quantifier-free formula φ in the language of Figure 5 can be written as true, false, or a disjunction of structural base formulas. Example 43 Consider the formula ≡ φ6 ∧ uz 6= uw Where φ6 is given by (41). By (22), literal uz 6= uw is equivalent to ψ1 ∨ ψ2 where Proof Sketch. Let φ be a well-defined quantifier-free formula in the language of Figure 5. We first use rule (21) to eliminate occurrences of ≤ in the formula replacing them with cardinality constraints. We then convert formula into disjunction φ1 ∨ · · · ∨ φn of well-formed conjunctions of literals. We next describe how to transform each conjunction φi into a disjunction of base formulas. Let φi be a conjunction of literals. Using the technique of Proposition 9, we convert the formula to unnested form, adding existential quantifiers. We then eliminate unnested conjuncts that contain boolean algebra operations, according to Figure 7. The only atomic formulas in the resulting (53) and |(uz ∩ ucw ) ∪ (ucz ∩ uw )|sh(uz ) ≥ 1 |(uz ∩ ucw ) ∧ usw2 and transform φ′′6 into a disjunction of base formulas. to structural base formula yields true. We conclude that (27) is a true sentence in the structure FT2 , which completes our quantifier elimination procedure example. ψ2 ≡ sh(uz ) = sh(uw ) ∧ |uw1 ∩ ucz1 |usw1 ∧ sh(uw1 ) = |(uz ∩ ucw ) ∪ (ucz ∩ uw )|usw ≥ 1 c ψ1 ≡ sh(uz ) 6= sh(uw ) sh(uz2 ) = usw2 (55) usw1 As in Example 42, we now apply rule (25) to φ10,2 ∨ φ11,2 is equivalent to sh(x) = sh(y). Converting φ′6 sh(uz1 ) = usw1 (54) 22 3. if hu, ti ∈ S and u = f (u1 , . . . , uk ) occurs in termBase for some f ∈ Σ then {hu1 , f1 (t)i, . . . , huk , fk (t)i} ⊆ S; existentially quantified conjunction are of form x = a, x = b, x = g(x1 , x2 ), Isg (x), x1 = g1 (x), x2 = g2 (x), x1 = x2 , xs = cs , xs = g s (xs1 , xs2 ), Isgs (xs ), xs1 = g1s (xs ), xs2 = g2s (xs ), xs1 = xs2 , xs = sh(x), as well as |t1 |xs ≥ k and |t2 |xs = k for some xs -terms t1 and t2 . The only negated atomic formulas are of form x1 6= x2 , xs1 6= xs2 , ¬Isg (x) and ¬Isgs (xs ). As in the proof of Proposition 28, we use (17) to eliminate ¬Isg (x) and ¬Isgs (xs ). This process leaves formulas of form x1 6= x2 and xs1 6= xs2 as the only negated atomic formulas. In the sequel, whenever we perform case analysis and generate a disjunction of conjunctions, existential quantifiers propagate to the conjunctions, so we keep working with a existentially quantified conjunction. The existentially quantified variables will become internal variables of a structural base formula. We next convert conjuncts that contain only term variables to a base formula, and convert shape part to base formula, as in the proof of Proposition 28. We simultaneously make sure every term variable has an associated shape variable, introducing new shape variables if needed. (This process is interleaved with conversion to base formula, to ensure that there is always a conjunct stating that newly introduced shape variables are distinct.) We also ensure homomorphism requirement by replacing internal variables when we entail their equality. Another condition we ensure is that parameter term variables map to parameter shape variables, and non-parameter term variables to non-parameter shape variables; we do this by performing expansion of term and shape variables. We perform expansion of shape variables as in Section 3.2. Expansion of term variables is even simpler because there is no need to do case analysis on equality of term variable with other variables. The resulting existentially quantified conjunction might contain disequalities u 6= u′ between term variables. We eliminate these disequalities as explained in Example 43, by converting each disequality into a cardinality constraint using (22). In general, we need to consider the case when sh(u) 6= sh(u′ ) and generate another disjunct. Elimination of disequalities might violate previously established homomorphism invariants, so we may need to reestablish these invariants by repeating the previously described steps. The overall process terminates because we never introduce new inequalities between term variables. As a final step, we convert all cardinality constraints into constraints on parameter term variables, using (25). In the case when the shape of cardinality constraint is cs , we cannot apply (25). However, in that case the homomorphism condition ensures that each of the participating variables is equal to a or equal to b. This means that we can simply evaluate the cardinality constraint in the boolean algebra {a, b}. If the result is true we simply drop the constraint, otherwise the entire base formula becomes false. This completes our sketch of transforming a quantifierfree formula into disjunction of structural base formulas. 4. if hus , ts i ∈ S and us = f s (us1 , . . . , usk ) occurs in shapeBase then {hus1 , f1s (ts )i, . . . , husk , fks (ts )i} ⊆ S; 5. if hu, ti ∈ S and sh(u) = us occurs in hom then hus , sh(t)i ∈ S. Definition 47 An internal term variable u is covered iff there exists a term t such that hu, ti ∈ S. An internal shape variable us is covered iff there exists a term ts such that hus , ts i ∈ S. Lemma 48 Let β be a structural base formula with matrix β0 and let covering be the covering of β. 1. If hu, ti ∈ S then |= β0 ⇒ u = t. 2. If hus , ts i ∈ S then |= β0 ⇒ us = ts . Proof. By induction, using Definition 46. Corollary 49 Let β be a structural base formula such that every internal variable is covered. Then β is equivalent to a well-defined quantifier-free formula. Proof. By Lemma 48 using (7). Lemma 50 Let u be an uncovered non-parameter term variable in a structural base formula β such that u is a source i.e. no conjunct of form u′ = f (u1 , . . . , u, . . . , uk ) occurs in termBase. Let β ′ be the result of dropping u from β. Then β is equivalent to β ′ . Proof. Let u occur in termBase in form u = f (u1 , . . . , uk ) The only other occurrence of u in β is in hom and has the form sh(u) = us . Because non-parameter term variables are mapped to non-parameter shape variables, shapeBase contains formula us = shapified(f )(us1 , . . . , usk ) (56) us1 , . . . , usk are such that, by = usi occurs in hom. This s where homomorphism property, sh(ui ) means that the conjunct sh(u) = u is a consequence of the remaining conjuncts, so it may be omitted. After that, applying (7) yields a structural base formula β ′ not containing u, where β ′ is equivalent to β. Corollary 51 Every base formula is equivalent to a base formula without uncovered non-parameter term variables. Proof. If a structural base formula has an uncovered non-parameter term variable, then it has an uncovered nonparameter term variable that is a source. By repeated application of Lemma (50) we eliminate all uncovered nonparameter term variables. We introduce the notion of covered variables in structural base formula by generalizing Definition 29. Definition 46 The set covering of variable coverings of a structural base formula β is the least set S of pairs hu, ti where u is an internal (shape or term) variable and t is a term over the free variables of β, such such that: The next example illustrates how we deal with cardinality constraints |1s |s ≥ k and |1s | = k, which contain no term variables. These constraints restrict the size of shape s. Luckily, we can be translate them into shape base formula constraints. 1. if x = u occurs in termBase then hu, xi ∈ S; 2. if xs = us occurs in shapeBase then hus , xs i ∈ S; 23 termBase, it occurs only in hom and cardin. Let sh(u) = us occur in hom. Let ψ1 , . . . , ψp be all conjuncts of cardin that contain u. Each ψi is of form |ti |us ≥ ki or |ti |us = ki for some us -term ti . Let uj1 , . . . , ujq be all term variables appearing in ti terms other than u. Conjunct sh(ujr ) = us occurs in hom for each r where 1 ≤ r ≤ q. The base formula can therefore be written in form Example 52 (Shape Term Size Constraints) Let x < y denote conjunction x ≤ y∧x 6= y. Let us eliminate quantifiers from formula ∃x.φ(x) where φ(x) ≡ ¬(∃y.∃z. x < y ∧ y < z) ∧ (57) ¬(∃u. u < x) Eliminating variables y, z from the first conjunct and variable u from the second conjunct yields β1 ≡ ∃x1 , . . . , xe , xs1 , . . . , xsf . φ ∧ φ1 where ¬|xc |sh(x) ≥ 2 ∧ ¬|x|sh(x) ≥ 1 φ1 ≡ ∃u. sh(u) = us ∧ which is equivalent to c sh(uj1 ) = us ∧ . . . ∧ sh(ujq ) = us ∧ c (|x |sh(x) = 0 ∨ |x |sh(x) = 1) ∧ |x|sh(x) = 0 ψ1 ∧ . . . ∧ ψp and further to disjunction c (58) All term variables in ψ1 , . . . , ψk range over terms of shape us . Therefore, φ1 defines a relation in the boolean algebra B(Jus K). This allows us to apply construction in Section 3.2. We eliminate u from ψ1 ∧. . .∧ ψp and obtain a propositional combination ψ0 of cardinality constraints with us -terms. φ0 does not contain variable u. We may assume that ψ0 is in disjunctive normal form c (|x |sh(x) = 0 ∧ |x|sh(x) = 0) ∨ (|x |sh(x) = 1 ∧ |x|sh(x) = 0) The first disjunct can be shown contradictory. Let us transform the second disjunct into a structural base formula. After introducing u = x and us = sh(u), we obtain ∃u, us . x = u ∧ sh(u) = us ∧ |u|us = 0 ∧ |uc |us = 1 ψ0 ≡ α1 ∨ . . . ∨ αw Then ∃x.φ(x) is equivalent to Let ∃u, us . sh(u) = us ∧ |u|us = 0 ∧ |uc |us = 1 φ1,i ≡ sh(uj1 ) = us ∧ . . . ∧ sh(ujq ) = us ∧ αi Eliminating parameter term variable u yields for 1 ≤ i ≤ w. Base formula β1 is equivalent to disjunction of base formulas β1,i where ∃us . |1|us = 1 β1,i ≡ ∃x1 , . . . , xe , xs1 , . . . , xsf . φ ∧ φ1,i Constraint |1|us = 1 means that the largest set in the boolean algebra B(s) where s is the value of us has size one. There exists exactly one boolean algebra of size one in the structure FT2 , namely {a, b}. Therefore, |1|us = 1 is equivalent to us = cs . We may now eliminate us by letting us = cs . We conclude that the sentence ∃x.φ(x) is true. Notice that we have also established that formula φ(x) is equivalent to sh(x) = cs , as a consequence of We have thus eliminated an uncovered parameter term variable u from β1 . By repeating this process we eliminate all uncovered parameter term variables from a base formula. The resulting formula contains no uncovered term variables. It remains to eliminate uncovered shape variables. This process is similar to term algebra quantifier elimination in Section 3.4. An essential part of construction in Section 3.4 is Lemma 25, which relies on the fact that uncovered parameter variables may take on infinitely many values. We therefore ensure that uncovered parameter shape variables are not constrained by term variables through conjuncts outside shapeBase. Suppose that us is an uncovered parameter shape variable in a base formula β. us does not occur in termBase. us does not occur in hom either, because all term variables are covered, and a conjunct sh(u) = us would imply that us is covered. The only possible occurrence of us is in cardinality constraint ψ of subformula cardin, where ψ is of form |t|us = k or of form |t|us ≥ k. Suppose there is some term variable u occurring in t. Then sh(u) = us so us is covered, which is a contradiction. Therefore, t has no variables. t can thus be simplified to either 0us or 1us . In general, a constraint of form |1|us = k or |1|us ≥ k is a domain cardinality constraint for boolean algebra B(Jus K) (see Remark 15 as well as (20)). A constraint containing |0us | is equivalent to true or false. A constraint |1|us = 0 is equivalent to false. A constraint |1|us = k for k ≥ 1 is equivalent to |1sh(x) |sh(x) = 1  The following Proposition 53 corresponds to Proposition 38. Proposition 53 (Struct. Base to Quantifier-Free) Every structural base formula β is equivalent to a quantifierfree formula φ in the language of Figure 5. Proof Sketch. By Corollary 51 we may assume that β has no uncovered non-parameter term variables. By Corollary 49 we are done if there are no uncovered variables, so it suffices to eliminate uncovered parameter term variables and uncovered shape variables. Let u be an uncovered parameter term variable. Then u does not occur in termBase. Indeed, suppose for the sake of contradiction that u occurs in termBase in some formula u′ = f (u1 , . . . , u, . . . , uk ) us = ts1 ∨ · · · ∨ us = tsp Then u′ is an uncovered non-parameter variable in β, which is a contradiction because we have assumed β has no uncovered non-parameter variables. Therefore, u does no occur in where ts1 , . . . , tsp is the list of all ground terms in signature Σ0 that have exactly k occurrences of constant cs . We therefore 24 generate a disjunction of base formulas β1 , . . . , βp where βi results from β by replacing |1|us = k with us = tsi . We convert each βi to a disjunction of base formulas by labelling subterms of ti by internal shape variables and doing case analysis on the equality between new internal shape variables to ensure the invariants of a base formula. The result is a disjunction of base formulas where variable us occurs only in shapeBase subformula. Similarly, |1|us ≥ k + 1 is equivalent to ¬(|1|us = k) and thus to us 6= ts1 ∧ · · · ∧ us 6= tsq (59) Theorem 54 (Two Constants Quant. Elimination) There exist algorithms A, B such that for a given formula φ in the language of Figure 5: a) A produces a quantifier-free formula φ′ in selector language b) B produces a disjunction φ′ of structural base formulas Proof. Analogous to proof of Theorem 39, using Proposition 45 in place of Proposition 28 and Proposition 53 in place of Proposition 38. where ts1 , . . . , tsq is the list of all ground terms in signature Σ0 that have at most k occurrences of constant cs . We replace |1|us ≥ k + 1 by (59) and again convert the result to a disjunction of base formulas where us occurs only in shapeBase subformula. Each of the resulting base formulas β 1 are such that every uncovered variable in β 1 is a shape variable that occurs only in shapeBase. Let β1 ≡ Corollary 55 The first-order theory of the structure FT2 is decidable. This completes description of our quantifier elimination for the first-order theory of structure FT2 , which models structural subtyping with two base types and one binary constructor. It is straightforward to extend the construction of this section to any number of covariant constructors if the base formula has only two constants. In Section 5 we extend the result to any number of constants as well. Finally, in Section 6 we extend the result to allow arbitrary decidable structures for primitive types, even if the number of primitive types is infinite. ∃u1 , . . . , un , us1 , . . . , usps , usps +1 , . . . , usps +qs . shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) ∧ termBase(u1 , . . . , un , x1 , . . . , xm ) ∧ hom(u1 , . . . , un , us1 , . . . , usn ) ∧ cardin(up+1 , . . . , up+q , usps +1 , . . . , usps +qs ) 5 1 us1 , . . . , usps 2 where are uncovered shape variables. Then β is equivalent to β : β2 ≡ A Finite Number of Constants In this section we prove the decidability of structural subtyping of any finite number of constant symbols (primitive types) and any number of function symbols (constructors). We first show the result when all constructors are covariant, we then show the result when some of the constructors are contravariant. We introduce the notion of Σ-term-power of some structure C as a generalization of the structure of structural subtyping. We represent primitive types in structural subtyping as a structure C with a finite carrier C. We call C the base structure. Without loss of generality, we assume that C has only relations; functions and constants are definable using relations. Let LC be a set of relation symbols and let ≤ ∈ LC be a distinguished binary relation symbol. ≤ represents the subtype ordering between types. C is finite, so C is decidable (see Section 6 for the case when C is infinite but decidable). We represent type constructors as free operations in the term algebra with signature Σ. To represent the variance of constructors we define for each constructor f ∈ Σ of arity ar(f ) = k and each argument 1 ≤ i ≤ k the value variance(f, i) ∈ {−1, 1}. The constructor f is covariant in argument i iff variance(f, i) = 1. For convenience we assume ar(f ) ≥ 1 for each f ∈ Σ. The Σ-term-power of C is a structure P defined as follows. Let Σ′ = Σ ∪ C. The domain of P is the set P of finite ground Σ′ -terms. Elements of C are viewed as constants of arity 1. The structure P has signature Σ∪LC . The constructors f ∈ Σ are interpreted in P as in a free term algebra: Jf KP (t1 , . . . , tk ) = f (t1 , . . . , tk ) ∃u1 , . . . , un , usps +1 , . . . , usps +qs . φ2 (usps +1 , . . . , usps +qs , xs1 , . . . , xsms ) termBase(u1 , . . . , un , x1 , . . . , xm ) ∧ hom(u1 , . . . , un , us1 , . . . , usn ) ∧ cardin(up+1 , . . . , up+q , usps +1 , . . . , usps +qs ) Here φ2 is a base formula (Definitions 19 and 21) whose free variables are variables free in β 2 as well as all covered shape variables: φ2 (usps +1 , . . . , usps +qs , xs1 , . . . , xsms ) ≡ ∃us1 , . . . , usps . shapeBase(us1 , . . . , usps , usps +1 , usps +qs , xs1 , . . . , xsms ) Applying Lemma 37 we conclude that φ2 is equivalent to some disjunction k _ φ3,i i=1 of base formulas without uncovered variables. Let β 3,i be the result of replacing φ2 with φ3,i in β 2 . Then β 2 is equivalent to k _ β 3,i i=1 and each β 3,i has no uncovered variables either, because every free variable of φ3,i is either free or covered in β 3,i . By Corollary 49 each β 3,i can be written as a quantifier free formula. A relation r ∈ LC \{≤} is interpreted pointwise on the terms of same “shape” as follows. JrKP is the least relation ρ such that: The following Theorem 54 corresponds to 39 of Section 3.4. 1. if JrKC (c1 , . . . , cn ) then ρ(c1 , . . . , cn ) 25 2. if ρ(ti1 , . . . , tin ) for all i where 1 ≤ i ≤ k, then 8. cardinality constraint relations J|φ(x1 , . . . , xk )|ts = kKPE (t1 , . . . , tk ) = ρ(f (t11 , . . . , t1k ), . . . , f (tn1 , . . . , tnk )) |Jφ(x1 , . . . , xk )KPE (t1 , . . . , tk )| = k The relation ≤ ∈ LC is interpreted similarly, but taking into account the variance. J≤KP is the least relation ρ such that and J|φ(x1 , . . . , xk )|ts ≥ kKPE (t1 , . . . , tk ) = 1. if J≤KC (c1 , c2 ) then ρ(c1 , c2 ) 2. if ρvariance(f,i) (ti1 , . . . , tin ) |Jφ(x1 , . . . , xk )KPE (t1 , . . . , tk )| ≥ k ρ(f (t11 , . . . , t1k ), . . . , f (tn1 , . . . , tnk )) It remains to complete the semantics of cardinality constraint relations, by defining the set Jφ(x1 , . . . , xk )KPE (t1 , . . . , tk ). If s is a shape, we call the set of positions of constant cs in s leaves of s, and denote it by leaves(s). We represent a leaf as a sequence of pairs hf, ii where f is a constructor of arity k and 1 ≤ i ≤ k. If l ∈ leaves(s) and sh(t) = s, then t[l] denotes the element c ∈ C at position l in term t i.e. if l = hf 1 , i1 i . . . hf n , in i then t[l] = finn (. . . fi22 (fi11 (t)) . . .) (64) We define: Here we use the notation ρv for v ∈ {−1, 1} with the meaning: ρ1 = ρ and ρ−1 = {hy, xi | hx, yi ∈ ρ}. We next sketch the decidability of structural subtyping for any finite number of primitive types C. For now we assume that all constructors f ∈ Σ are covariant, the relation ≤ thus does not play a special role. Extended Term-Power Structure For the purpose of quantifier elimination we define the structure PE by extending the domain and the set of operations of the term-power structure P. The domain of PE is PE = P ∪ PS where PS is the set of shapes defined as follows. Let Σs = {cs } ∪ {f s | f ∈ Σ} be a set of function symbols such that cs is a fresh constant symbol with ar(cs ) = 0 and f s are fresh distinct constant symbols with ar(f s ) = ar(f ) for each f ∈ Σ. The set of shapes PS is the set of ground Σs -terms. When referring to elements of PE by term we mean an element of P ; by shape we mean an element of PS . We write X s to denote an entity pertaining to shapes as opposed to terms, so xs , us denote variables ranging over shapes, and ts to denotes terms that evaluate to shapes. The extended structure PE contains term algebra operations on terms and shapes (including selector operations and tests, [22, Page 61]), the homomorphism sh, and cardinality constraint relations |φ|ts = k and |φ|ts ≥ k: Jφ(x1 , . . . , xk )KPE (t1 , . . . , tk ) = {l | Jφ(x1 , . . . , xk )KC (t1 [l], . . . , tk [l])} shapified(f )(JshKPE (t1 ), . . . , JshKPE (tn )) |Jφ(x1 , . . . , xk )KPE (c1 , . . . , ck )| = ( 1, Jφ(x1 , . . . , xk )KC (c1 , . . . , ck ) 0, shapified(f ) = cs , if x ∈ C s if f ∈ Σ f , (66) C ¬Jφ(x1 , . . . , xk )K (c1 , . . . , ck ) |Jφ(x1 , . . . , xk )KPE (f (t11 , . . . , t1l ), . . . , f (tk1 , . . . , tkl ))| = |Jφ(x1 , . . . , xk )KPE (t11 , . . . , tk1 )| + . . . + |Jφ(x1 , . . . , xk )KPE (t1l , . . . , tkl )| (67) Definition (65) generalizes [14, Definition 2.1, Page 63]. We write |φ(t1 , . . . , tk )|ts = k as a shorthand for the atomic formula (|φ(x1 , . . . , xk )|ts = k)(t1 , . . . , tk ), similarly for |φ(t1 , . . . , tk )|ts ≥ k. This is more than a notational convenience, see Section 6 for an approach which introduces sets of leaves as elements of the domain of PE and defines a cylindric algebra interpreted over sets of leaves. The approach in this section follows [35] in merging the quantifier elimination for products and quantifier elimination for boolean algebras. Some of the operations in PE are partial. We use the definitions and results of Section 2.3 to deal with partial functions. fi (t) is defined iff Isf (t) holds, fis (ts ) is defined iff Isf s (ts ) holds. Cardinality constraints |φ(t1 , . . . , tk )|ts = k and |φ(t1 , . . . , tk )|ts ≥ k are defined iff sh(t1 ) = . . . = sh(tk ) = ts holds. The structure PE is at least as expressive as P because the only operations or relations present in P but not in PE are JrKP for r ∈ LC , and we can express JrKP (t1 , . . . , tk ) as |¬ r(t1 , . . . , tk )|sh(t1 ) = 0. Our goal is to give a quantifier elimination for first-order formulas of structure PE . By a quantifier-free formula we mean a formula without quantifiers outside cardinality constraints, e.g. the formula |∀x.x ≤ t|xs = k is quantifier-free. (60) where shapified(x) = (65) The following equations follow from (65) and can be used as an equivalent alternative definition for cardinality relations: 1. constructors in the term algebra of terms, f ∈ Σ′ Jf KPE (t1 , . . . , tk ) = f (t1 , . . . , tk ); 2. selectors in term the algebra of terms, Jfi KPE (f (t1 , . . . , tk )) = ti ; 3. constructor tests in the term algebra of terms, JIsf KPE (t) = ∃t1 , . . . , tk . t = f (t1 , . . . , tk ); 4. constructors in the term algebra of shapes, f s ∈ Σs Jf s KPE (ts1 , . . . , tsk ) = f s (ts1 , . . . , tsk ); 5. selectors in the term algebra of shapes, Jfis KPE (f s (ts1 , . . . , tsk )) = tsi ; 6. constructor tests in the term algebra of shapes, JIsf s KPE (ts ) = ∃ts1 , . . . , tsk . ts = f s (ts1 , . . . , tsk ); 7. the homomorphism mapping terms to shapes such that: JshKPE (f (t1 , . . . , tn )) = (63) where φ(x1 , . . . , xk ) is is a first-order formula over the base-structure language LC with free variables x1 , . . . , xk , term ts denotes a shape, and k is a nonnegative integer constant. for all i where 1 ≤ i ≤ k, then 5.1 (62) (61) 26 5.2 Structural Base Formulas a parameter variable iff its shape is a parameter shape variable). In this section we define the notion of structural base formulas for any base structure C with a finite carrier. Definition 56 of structural base formula for quantifier elimination in PE differs from Definition 41 in the conjuncts of cardin subformula. Instead of cardinality constraints on boolean algebra terms, Definition 56 contains cardinality constraints on first-order formulas. The notion of base formula and Lemma 25 apply to terms P as well as shapes PS in the structure PE because shapes are also terms over the alphabet Σs . For brevity we write u∗ for an internal shape or term variable, and similarly x∗ for a free shape or term variable, t∗ for terms, f ∗ for term or shape term algebra constructor and fi∗ for a term or shape term algebra selector. cardin(up+1 , . . . , un , usps +1 , . . . , usns ) = ψ1 ∧ · · · ∧ ψd where each ψi is a cardinality constraint of the form |φ(uj1 , . . . , ujl )|us = k or |φ(uj1 , . . . , ujl )|us ≥ k where {j1 , . . . , jl } ⊆ {p + 1, . . . , n} and the conjunct sh(ujd ) = us occurs in termHom for 1 ≤ d ≤ l. We require each structural base formula to satisfy the following conditions: P0) the graph associated with shape base formula Definition 56 (Structural Base Formula) A structural base formula with: • • • • • • ∃us1 , . . . , usns . shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) free term variables x1 , . . . , xm ; internal non-parameter term variables u1 , . . . , up ; internal parameter term variables up+1 , . . . , up+q ; free shape variables xs1 , . . . , xsms ; internal non-parameter shape variables us1 , . . . , usps ; internal parameter shape variables usps , . . . , usps +qs is acyclic; P1) congruence closure property for shapeBase subformula: there are no two distinct variables usi and usj such that both usi = f (usl1 , . . . , uslk ) and usj = f (usl1 , . . . , uslk ) occur as conjuncts in formula shapeBase; is a formula of the form: P2) congruence closure property for termBase subformula: there are no two distinct variables ui and uj such that both ui = f (ul1 , . . . , ulk ) and uj = f (ul1 , . . . , ulk ) occur as conjuncts in formula termBase; ∃u1 , . . . , un , us1 , . . . , usns . shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) ∧ termBase(u1 , . . . , un , x1 , . . . , xm ) ∧ termHom(u1 , . . . , un , us1 , . . . , usns ) P3) homomorphism property of sh: for every non-parameter term variable u such that u = f (ui1 , . . . , uik ) occurs in termBase, if conjunct sh(u) = us occurs in termHom, then for some shape variables usj1 , . . . , usjk term us = f s (usj1 , . . . , usjk ) occurs in shapeBase where f s = shapified(f ) and for every r where 1 ≤ r ≤ k, conjunct sh(uir ) = usjr occurs in termHom. ∧ cardin(up+1 , . . . , un , usps +1 , . . . , usns ) where n = p + q, ns = ps + q s , and formulas shapeBase, termBase, termHom, cardin are defined as follows. shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) = s p V usi = ti (us1 , . . . , usns ) ∧ i=1 ∧ distinct(us1 , . . . , usn ) m Vs Note that the validity of the occur check for term variables follows from P0) and P3). Another immediate consequence of Definition 56 is the following Proposition 57. xsi = usji i=1 Proposition 57 (Quantification of Str. Base Form.) If β is a structural base formula and x a free shape or term variable in β, then there exists a base structural formula β1 equivalent to ∃x.β. where each ti is a shape term of the form f s (usi1 , . . . , usik ) for some f ∈ Σ0 , k = ar(f ), and j : {1, . . . , ms } → {1, . . . , ns } is a function mapping indices of free shape variables to indices of internal shape variables. We proceed to show that a quantifier-free formula can be written as a disjunction of base formulas, and a base formula can be written as a quantifier-free formula. termBase(u1 , . . . , un , x1 , . . . , xm ) = p V ui = ti (u1 , . . . , un ) ∧ i=1 m V xi = uji i=1 5.3 where each ti is a term of the form f (ui1 , . . . , uik ) for some f ∈ Σ, k = ar(f ), and j : {1, . . . , m} → {1, . . . , n} is a function mapping indices of free term variables to indices of internal term variables. termHom(u1 , . . . , un , us1 , . . . , usns ) = n V i=1 Conversion to Base Formulas Conversion from a quantifier-free formula to the structural base formula is given by Proposition 57. The proof of Proposition 58 is analogous to the proof of Proposition 45 but uses of (67) instead of (25). Proposition 58 (Quantifier-Free to Structural Base) Every well-defined quantifier-free formula φ is equivalent on PE to true, false, or some disjunction of structural base formulas. sh(ui ) = usji where j : {1, . . . , n} → {1, . . . , ns } is some function such that {j1 , . . . , jp } ⊆ {1, . . . , ps } and {jp+1 , . . . , jp+q } ⊆ {ps + 1, . . . , ps + q s } (a term variable is 27 5.4 Conversion to Quantifier-Free Formulas Lemma 65 Let α ≡ ∃u. sh(u) = us ∧ The conversion from structural base formulas to quantifierfree formulas is similar to the case of two constant symbols in Section 4.3, but requires the use of Feferman-Vaught technique. sh(u1 ) = us ∧ . . . ∧ sh(uk ) = us ∧ (69) ψ1 ∧ . . . ∧ ψp Definition 59 The set dets of variable determinations of a structural base formula β is the least set S of pairs hu∗ , t∗ i where u∗ is an internal term or shape variable and t∗ is a term over the free variables of β, such such that: where each ψi is a cardinality constraint of the form |φ|us = k or |φ|us ≥ k where all free variables of φ are among u, u1 , . . . , uk . Then there exists formula ψ such that ψ is a disjunction of conjunctions of cardinality constraints |φ′ | = k and |φ′ | ≥ k where the free variables in each φ′ are among u1 , . . . , uk and formula α is equivalent on PE to α′ where 1. if x∗ = u∗ occurs in termBase or shapeBase, then hu∗ , x∗ i ∈ S; 2. if hu∗ , t∗ i ∈ S and u∗ = f ∗ (u∗1 , . . . , u∗k ) occurs in shapeBase or termBase then {hu∗1 , f1∗ (t∗ )i, . . . , hu∗k , fk∗ (t∗ )i} ⊆ S; α′ ≡ sh(u1 ) = us ∧ . . . ∧ sh(uk ) = us ∧ ψ (70) 3. if {hu∗1 , f1∗ (t∗ )i, . . . , hu∗k , fk∗ (t∗ )i} ⊆ S and u∗ = f ∗ (u∗1 , . . . , u∗k ) occurs in shapeBase or termBase then hu∗ , t∗ i ∈ S; Proposition 66 (Struct. Base to Quantifier-Free) Every structural base formula β is equivalent on PE to some well-defined quantifier-free formula φ. 4. if hu, ti ∈ S and sh(u) = us occurs in termHom then hus , sh(t)i ∈ S. Proof Sketch. By Corollary 64 we may assume that β has no undetermined non-parameter term variables. By Corollary 62 we are done if there are no undetermined variables, so it suffices to eliminate undetermined parameter term variables and undetermined shape variables. Let u be an undetermined parameter term variable. u does not occur in termBase because it cannot have a successor or a predecessors in the graph associated with term base formula. Therefore, u′ occurs only in termHom and cardin. Let us be the shape variable such that us = sh(u) occurs in termHom. Let ψ1 , . . . , ψp be all conjuncts of cardin that contain u. Each ψi is of the form |φ|us ≥ ki or |φ|us = ki and for each variable u′ free in φ the conjunct sh(u) = us occurs in termHom. The base formula can therefore be written in form β1 ≡ ∃x1 , . . . , xe , xs1 , . . . , xsf . φ ∧ α Definition 60 An internal variable u∗ is determined if hu∗ , t∗ i ∈ dets for some term ts . An internal variable is undetermined if it is not determined. Lemma 61 Let β be a structural base formula with matrix β0 and let dets be the determinations of β. If hu∗ , t∗ i ∈ S then |= β0 ⇒ u∗ = t∗ . Corollary 62 Let β be a structural base formula such that every internal variable is determined. Then β is equivalent to a well-defined quantifier-free formula. Proof. By Lemma 61 using ∃x.x = t ∧ φ(x) ⇐⇒ φ(t) (68) where α has the form as in Lemma W 65. Applying Lemma 65 we eliminate u and obtain ψ = w i=1 αi where and each αi is a conjunction of cardinality constraints. Base formula β1 is W β thus equivalent to the disjunction w i=1 1,i where each β1,i is a base formula Lemma 63 Let u be an undetermined non-parameter term variable in a structural base formula β such that u is a source i.e. no conjunct of the form u′ = f (u1 , . . . , u, . . . , uk ) β1,i ≡ ∃x1 , . . . , xe , xs1 , . . . , xsf . φ ∧ φ1,i ′ occurs in termBase. Let β be the result of removing u and conjuncts containing u from β. Then β is equivalent to β ′ . By repeating this process we eliminate all undetermined parameter term variables from a base formula. Each of the resulting base formulas contains no undetermined term variables. It remains to eliminate undetermined shape variables. This process is similar to term algebra quantifier elimination; the key ingredient is Lemma 25, which relies on the fact that undetermined parameter variables may take on infinitely many values. We therefore ensure that undetermined parameter shape variables are not constrained by term and parameter variables through conjuncts outside shapeBase. Consider an undetermined parameter shape variable us . us does not occur in termHom, because all term variables are determined and a conjunct us = sh(u) would imply that us is determined as well. us can thus occur only in cardin within some cardinality constraint |φ|us = k or |φ|us ≥ k. Moreover, formula φ in each such cardinality constraint is closed: otherwise φ would contain some free variable u, by Proof. The conjunct containing u in termHom is a consequence of the remaining conjuncts, so we drop it. We then apply (68). Corollary 64 Every base formula is equivalent to a base formula without undetermined non-parameter term variables. Proof. If a structural base formula has an undetermined non-parameter term variable, then it has an undetermined non-parameter term variable that is a source. Repeatedly apply Lemma 63 to eliminate all undetermined non-parameter term variables. The following Lemma 65 is a consequence of the fact that terms of a fixed shape s form a substructure of P isomorphic to the finite power C m where m = |leaves(s)| and follows from Feferman-Vaught theorem in Section 3.3. 28 definition of base formula u would have to be a parameter variable, all parameter term variables are determined, so us would be determined as well. Let us denote some shape s. Because φ is a closed formula, |φ| is equal to 0 if JφKC = false and to the shape size m = |leaves(s)| if JφKC = true. (The fact that closed formulas reduce to the constraints on domain size appears in [35, Theorem 3.36, Page 13].) After eliminating constraints equivalent to 0 = k and 0 ≥ k, we obtain a conjunction of simple linear constraints of the form m = k and m ≥ k. These constraints specify a finite or infinite set S ⊆ {0, 1, . . .} of possible sizes m. Let A = {s | |leaves(s)| ∈ S}. If the set S is infinite then it contains an infinite interval of form {m0 , m0 + 1, . . .} so the set A is infinite. If Σ contains a unary constructor and S is nonempty, then A is infinite. If Σ contains no unary constructors and S is finite then A is finite and the cardinality constraints containing us are equivalent to W p s s s s i=1 u = ti where A = {t1 , . . . , tp }. We therefore generate a disjunction of base formulas β1 , . . . , βp where βi results from β by replacing cardinality constraints containing us with with us = tsi . We convert each βi to a disjunction of base formulas by labelling subterms of ti with internal shape variables and doing case analysis on the equality between new internal shape variables to ensure the invariants of a base formula, as in the proof of 58. By repeating this process for all shape variables us where the set S is finite, we obtain base formulas where the set A is infinite for every undetermined parameter shape variable us . We may then eliminate all undetermined parameter and non-parameter shape variables along with the conjuncts that contain them. The result is an equivalent formula by Lemma 25. All variables in each of the resulting base formulas are determined. By Corollary 62 each formula can be written as a quantifier-free formula, and the resulting disjunction is a quantifier-free formula. 5.5 We clearly have for every formula φ and every valuation σ: Jφ(−1) KC = JφKC n Y Jφ(x1 , . . . , xk )KPE (t1 , . . . , tk ) = ′ {l | Jφ(x1 , . . . , xk )KC (t1 [l], . . . , tk [l])} = t2 ≤ t1 (φ1 ∧ φ2 ) = φ1 (−1) ∧ φ2 (−1) (φ1 ∨ φ2 )(−1) = φ1 (−1) ∧ φ2 (−1) (¬φ)(−1) = ¬φ(−1) (∃t.φ)(−1) = ∃t.φ(−1) (∀t.φ)(−1) = ∀t.φ(−1) = (J≤KC )−1 |{l ∈ leaves(s) | variance(l) = 1}| N (s) |{l ∈ leaves(s) | variance(l) = −1}| = (76) |Jφ(x1 , . . . , xk )KPE (f (t11 , . . . , t1l ), . . . , f (tk1 , . . . , tkl ))| P = li=1 |Jφ(variance(f,l)) (x1 , . . . , xk )KPE (t1i , . . . , tki )| (77) The only change in the proof of Proposition 58 is the use of (77) instead of (67). Most of the proof of Proposition 66 remains unchanged as well; the only additional difficulty is eliminating constraints of the form |φ|us = k and |φ|us ≥ k where us is a parameter shape variable and φ is a closed formula. Lemma 68 below addresses this problem. We say that an algorithm g finitely computes some function f : A → 2B where B is an infinite set iff g is a function from A to the set Fin(B) ∪ {∞} where Fin(B) is the set of finite subsets of set B, ∞ is a fresh symbol, and ( f (a), if f (a) ∈ Fin(B) (78) g(a) = ∞, if f (a) ∈ / Fin(B) Lemma 68 There exists an algorithm V that, given a shape variable us and a conjunction ψ ≡ n i=1 ψi of cardinality constraints where each ψi is of form |φi |us = ki or |φi |us ≥ ki for some closed formula φi , finitely computes the set A = {s | JψKP [us 7→ s]} (79) of shapes which satisfy ψ in P. Proof Sketch. Let φ be a closed formula in language LC . Compute JφKC and Jφ(−1) KC and then replace |φ|s with one of the expressions P (s) + N (s), P (s), N (s), 0 according to the following table. Define C −1 to have the same domain and same interpretation of operations and relations r ∈ LC \ {≤} but where −1 P (s) = Because of this isomorphism, Lemma 65 still holds and we may still use Feferman-Vaught theorem from Section 3.3. Equation (67) generalizes to: (71) J≤KC (75) where C ′ denotes C for variancel = 1 and C −1 for variancel − 1. Hence, isomorhism between terms of some fixed shape s with |leaves(s)| = m and C m breaks, but there is still an isomorphism with C P (s) × (C −1 )N(s) where = r(t1 , . . . , tk ), if r ∈ LC \ {≤} (−1) (74) We generalize (65) to Definition 67 If φ is a first-order formula in the language LC the contravariant version of φ, denoted φ(−1) , is defined by induction on the structure of formula by: (t1 ≤ t2 ) variance(f j , ij ) j=1 So far we have assumed that all constructors are covariant. In this section we describe the changes needed to extend the result to the case when the constructors have arbitrary variance with respect to some distinguished binary relation denoted ≤. (−1) (73) If l ∈ leaves(s) is a leaf l = hf 1 , i1 i . . . hf n , in i, define variance(l) as the product of integers One-Relation-Symbol Variance (r(t1 , . . . , tk ))(−1) −1 (72) 29 JφKC true Jφ(−1) KC true |φ|s = P (s) + N (s) true false P (s) false true N (s) false false 0 (80) Let The constraints of the form N (s) + P (s) = k and N (s) + P (s) = k can be expressed as propositional combinations of constraints of the form N (s) = k, P (s) = k, P (s) ≥ k and N (s) ≥ k. Therefore, ψ can be written as a propositional combination of these four kinds of constraints and each conjunction C(s) can further be assumed to have one of the forms: F2) CkP ,k+ (s) ≡ P (s) = kP ∧ N (s) ≥ kN ; N F3) Ck+ ,kN (s) ≡ P (s) ≥ kP ∧ N (s) = kN ; P Ti = {hp, ni ∈ U | n = i} p1 + · · · + pl + nl+1 + . . . + nk = i F4) Ck+ ,k+ (s) ≡ P (s) ≥ kP ∧ N (s) ≥ kN . N (85) (86) for nonnegative integers p1 , . . . , pl , nl+1 , . . . , nk . First consider solution solutions where no variable is equal to i. If for one of the solutions, one of the sets Sp1 , . . . , Spl is infinite, then Si is infinite, otherwise add to Si all elements hi, ni where n = n1 + · · · + nl + pl+1 + . . . + pk (87) If n ≤ kP then also add the same elements hi, ni to Tn . Next, proceed analogously with Ti , considering solutions of Let A = {s ∈ PS | C(s)}. To compute A when Σ contains unary constructors, we first restrict Σ to the language Σ′ with no unary constructors, and compute the set A′ ⊆ A using language Σ′ . If A′ is empty, so is A, otherwise A is infinite. Assume that Σ contains no unary constructors. Assume further Σ contains at least one binary constructor and at lest one constructor is contravariant in some argument. Let S = {hP (s), N (s)i | s ∈ A} n1 + · · · + nl + pl+1 + . . . + pk = i Because P (s) + N (s) = |leaves(s)| and there are only finitely many shapes of any given size (every constructor is of arity at least two), it suffices to finitely compute S. S can be given an alternative characterization as follows. If f ∈ Σ, ar(f ) = k, f is covariant in l arguments and contravariant in k−l arguments define Jf KS (hp1 , n1 i, . . . , hpk , nk i) = P P P P h li=1 pi + ki=l+1 ni , li=1 ni + ki=l+1 pi i = {hp, ni ∈ U | p = i} To finitely compute S, finitely compute the sets Si and Ti for 0 ≤ i ≤ kP . The algorithm starts with all sets Si and Ti empty and keeps adding elements according to operations Jf KS . Assume that S0 , T0 , . . . , Si−1 , Ti−1 are finitely computed. The computation of Si and Ti proceeds as follows. Let f ∈ Σ be a constructor of arity k with l covariant arguments. For Si we consider all solutions of the equation F1) CkP ,kN (s) ≡ P (s) = kP ∧ N (s) = kN ; P Si (88) If at this point Si is not infinite and not empty, then also consider the solutions of (87) where pj = i for some j. If such solution exists, then mark Si as infinite. Proceed analogously with Ti . Finally, if both Si and Ti are still finite but there exists a solution for Si where nl+j = i for some j and exists a solution for Tj where pl+d = i for some d, then mark both Si and Ti as infinite. This completes the sketch of one step of the computation. (This step also applies to S0 and T0 ; we initially assume that h1, 0i ∈ T0 .) (81) Let U be the subset of {hp, ni | p, n ≥ 0} generated from element h1, 0i using operations Jf KS for f ∈ Σ. Then Example 69 Let us apply this algorithm to the special case where Σ = {f, g} and S = {hp, ni ∈ U | c(p, n)} variance(g) = h1, 1i (82) variance(f ) = h−1, 1i Let us see what the set S looks like. If hx, yi ∈ S define khx, yi = hkx, kyi as in a vector space. First, h1, 0i ∈ S because of cs . Next h1, 1i ∈ S because of f s and h2, 0i ∈ S because of g s . More generally, we have the following composition rule: If hp1 , n1 i, hp2 , n2 i then where c(p, n) is the linear constraint corresponding to the constraint C(s). Let C(s) = CkP ,kN (s). Then S ⊆ {hp, ni | p + n = kP + kN }. S is therefore a subset of a finite set and is easily computable, which solves case F1). Let C(s) = Ck+ ,k+ (s). Because Σ contains a binary conN P structor, S contains pairs hp, ni with arbitrarily large p+n, so either the p components or n component of elements of S grows unboundedly. Because Σ contains a constructor f contravariant in some argument, we can define using f an operation o acting as a constructor covariant in at least one argument and contravariant in at least one argument. Using operation on tuples whose one component grows unboundedly yields tuples whose both components grow unboundedly. Therefore, S is infinite, which solves case F4). Finally, consider the case C(s) = CkP ,k+ (s) (this will N solve the case C(s) = Ck+ ,kN (s) as well). Observe that hp1 + p2 , n1 + n2 i ∈ S s because of g , and hn1 + p2 , p1 + n2 i ∈ S s because of f . Using g s we obtain all pairs hp, 0i for p ≥ 1. Using f s once on those we obtain h1, ni for n ≥ 0. Adding these we additionally obtain hp, ni for p ≥ 2 and n ≥ 0. Hence we have all pairs hp, ni for p ≥ 1 and n ≥ 0 and those are the only ones that can be obtained. Thus, P kP −1 CkP ,k+ (s) = CkP ,0+ (s) ∧ N ^ ¬CkP ,i (s) S = {hp, ni | p ≥ 1 ∧ n ≥ 0} (83) i=0 As expected, the case F1) yields a finite and the case F4) an infinite set. The case F2) for kP = 0 is an empty set, otherwise it is an infinite set. The case F3) always yields an infinite set. This solves the problem for two constructors f, g. Because the set S for each CkP ,i (s) is finite, it suffices to finitely compute S for CkP ,0+ (s). In that case S = {hp, ni ∈ U | p = kP } (84) 30 we obtain the theory in Section 4. When R = {r} where r is a partial order on types, we obtain the theory of structural subtyping of non-recursive covariant types. For arbitrary relational structure C, if f ∈ Σ for ar(f ) = k we obtain a structure that properly contains the k-th strong power of structure C, in the terminology of [35]. The structure of this section follows Sections 4. We also associate a boolean algebra of sets with each term t. However, in this case, the elements of the associated boolean algebra are sets of occurrences of the constants that satisfy the given first-order formula interpreted over C. The occurrences of constants within the terms of a given shape correspond to the indices of the product structure in Section 3.3. We call these occurrences leaves, because they can be represented as leaves of the tree corresponding to a term. lifted relations r ′ for r ∈ LC r′ :: termk → bool term algebra on terms constructors, f ∈ Σ: f :: termk → term constructor test, f ∈ Σ: Isf :: term → bool selectors, f ∈ Σ: fi :: term → term 6.1 Figure 8: Basic Operations of Σ-term-power Structure In this section we define the notions shape and leafset, and state some properties that we use in the sequel. Let Σ0 = {cs } ∪ {f s | f ∈ Σ}  Lemma 68 allows to carry our the proof of Proposition 66 so we obtain our main result for finite C. be a set of function symbols such that cs is a fresh constant symbol with ar(cs ) = 0 and f s are fresh distinct constant symbols with ar(f s ) = ar(f ) for each f ∈ Σ. Let shapified : Σ′ → Σ0 be defined by Theorem 70 (Term Power Quant. Elimination) There exists an algorithm that for a given well-defined formula φ produces a quantifier-free formula φ′ that is equivalent to φ on PE . shapified(x) = cs , if x ∈ C s if f ∈ Σ shapified(f ) Corollary 71 (Decidability of Structural Subtyping) Let C be a structure with a finite carrier and P a Σ-termpower of C. Then the first-order theory of P is decidable. 6 Product Theory of Terms of a Given Shape = f , Let FT(Σ0 ) be the set of ground terms with signature Σ0 and FT(Σ′ ) the set of ground terms of signature Σ′ . Define function sh :: FT(Σ′ ) → FT(Σ0 ) mapping each term to its shape by Term-Powers of Decidable Theories In this section we extend the result of Section 5 on decidability of term-powers of a base structure C to allow C to be an arbitrary decidable theory, even if the carrier C is infinite. To keep a finite language in the case when C is infinite, we introduce a predicate IsPRI that allows testing whether t ∈ C for a term t ∈ P . In structural base formulas, we now distinguish between 1) composed variables, denoting elements t ∈ P for which Isf (t) holds for some constructor f ∈ Σ, and 2) primitive variables, denoting elements t ∈ P for which IsPRI (t) holds. Another generalization compared to Section 5 is the use of a syntactically richer language for term power algebras; to some extent this richer language can be viewed as syntactic sugar and can be simplified away. The generalization to infinitely many primitive types and the generalization to a richer language are orthogonal. For most of the section we focus on covariant constructors, Section 6.5 discusses a generalized notion of variance. As in Section 3.3 let C = hC, Ri be a decidable structure where C is a non-empty set and R is a set of relations interpreting some relational language LC , such that each r ∈ R is a relation of arity ar(r) on set C, i.e. r ⊆ C ar(r) . We assume that R contains a binary relation symbol r = ∈ R, interpreted as equality on the set C. Operations and relations of the Σ-term-power structure are summarized in Figure 8. We will show the decidability of the first-order theory of the structure with these operations. In the special case when C = {a, b} and sh(f (t1 , . . . , tn )) = shapified(f )(sh(t1 ), . . . , sh(tn )) for each f ∈ Σ′ . Define t1 ∼ t2 iff sh(t1 ) = sh(t2 ). Let t be a term or shape and t′ the tree representing t as in Section 2.2. If p is a path such that t′ (p) is defined and denotes a constant, we write t[p] to denote t′ (p) and call p a leaf. Note that t[p] is defined iff sh(t)[p] is defined. On the set of equivalent terms leaves act as indices of Section 3.3. If s is a shape, let leaves(s) denote the set of all leaves defined on shape s. Generalizing tCont of Section 4.1, define function tCont : FT(Σ′ ) → C ∗ by: tCont(c) = cs , if c ∈ C tCont(f (t1 , . . . , tk )) = tCont(t1 ) · . . . · tCont(tk ) Define δ(t) = hsh(t), tCont(t)i and B = {hs, wi | s ∈ FT(Σ0 ), w ∈ C ∗ , tLen(s) = sLen(w)} If all constructors f ∈ Σ are covariant then δ is a bijection between FT(Σ′ ) and B. Let B(s0 ) = {hs, wi ∈ B | s = s0 } For a fixed s0 , the set B(s0 ) is isomorphic to the power structure C n where n = tLen(s). For each shape s we introduce operations from Section 3.3. To distinguish the sets of positions belonging to different shapes, we tag each set of positions L with a shape r = {ha, ai, ha, bi, hb, bi} 31 s. We call the pair hs, Li a leafset. The interpretation of each relation r ∈ LC is the leafset: 6.2 To show the decidability of the first-order theory of the structure FT∗ with operations in Figure 8, we show decidability for a richer structure. Figure 9 shows the operations and relations of this richer structure. The structure has four sorts: bool representing truth values, term representing terms, shape representing shapes, and leafset representing sets of leaves within a given shape. The structure can be seen as as a combination of the operations of Figure 5 and Figure 2. For each relation symbol r ∈ R we define a relation symbol r ∗ of sort shape × termk → bool acting on terms of the same shape. While in Section 4.2 we associate a boolean algebra with the terms of same shape, in this section we associate a cylindric algebra [21] with terms of the same shape. This is a particularly simple cylindric algebra resulting from lifting first-order logic on the base structure C so that elements are replaced by terms of a given shape (which are isomorphic to functions from leaves to elements), and boolean values are replaced by sets of leaves (isomorphic to functions from leaves to booleans). In both cases, operations on the set X are lifted to operations on the set leaves(s) → X. Syntactically, we introduce a copy of all propositional connectives and quantifiers: ∧I , ∨I , ¬I , trueI , falseI . Like boolean algebra operations in Figure 5, these syntactic constructs in Figure 9 take an additional shape argument, because term-power algebra contains one copy of a strong power C n of base structure for each shape. We call formulas built using the operations of the cylindric algebra inner formulas. For each operation in Figure 2 there is an operation in Figure 9, potentially taking a shape as an additional argument (for operations used to build inner formulas). The logic further contains term algebra operations on terms, leafsets, and shapes. We use undecorated identifiers (e.g. u) to denote variables of term sort, variables with superscript S to denote shape variables (e.g. us ) and variables with superscript L to denote leafset variables (e.g. uL ). Figures 10 and 11 show the semantics of logic in Figure 9. The first row specifies semantics of operations in the case when all arguments are defined and are in the domain of the operation. The domain of each operation is in the second column, it is omitted if it is equal to the entire domain resulting from interpreting the sort of the operation. All operations except for plain logical operations and quantifiers over the bool domain are strict. Logical operations and quantifiers over the bool domain are defined as in the three-valued logic of Section 2.3. We remark that values of leafset act as terms with two constants in Figure 5. In fact, if the base structure C has only two constants then the formula x = a and its propositional combinations are sufficient to express all facts about C, so in that case there is no need to distinguish between terms and leafsets. Jrs K(t1 , . . . , tk ) = hs, {p | JrKC (t1 [p], . . . , tk [p])}i We let ∧Is , ∨Is , ¬Is , trueIs , falseIs stand for intersection, union, complement, full set and empty set in the algebra of subsets of the set leaves(s). We also introduce ∃Is as the union of a family of subsets indexed by a term of shape s and ∀Is as the intersection of a family of subsets indexed by a term. We use constructor-selector language for the term algebra on terms. We introduce constructor-selector language on shapes by generalizing operations in Section 4.1 in a natural way. In addition, we introduce a constructor-selector language on leafsets. For each f ∈ Σ we introduce a constructor symbol f L on leafsets and define leafified(f ) = f L Constructors f L act on leafsets as follows. If Li ⊆ leaves(si ) for 1 ≤ i ≤ k define f L (hs1 , L1 i, . . . , hsk , Lk i) = hs, Li where s = f s (s1 , . . . , sk ), and L ⊆ leaves(s) is given by L = ({1} · L1 ) ∪ · · · ∪ ({k} · Lk ) (Here we define A · B = {a · b | a ∈ A ∧ b ∈ B}.) We define selector functions on leafsets as follows. If s = f s (s1 , . . . , sk ) and L ⊆ leaves(s), then fiL (hs, Li) = hsi , Li i where Li ⊆ leaves(si ) is defined by Li = {w | w · i ∈ A} Equivalently, we require that fiL (f L (hs1 , L1 i, . . . , hsn , Ln i)) = hsi , Li i We can now express relations r ′ in Figure 8 using the fact: r ′ (t1 , . . . , tk ) ⇐⇒ sh(t2 ) = sh(t1 ) ∧ . . . ∧ sh(tk ) = sh(t1 ) ∧ rsh(t1 ) (t1 , . . . , tk ) = (89) trueIsh(t1 ) To handle an infinite number of elements of the base structure C, we do not introduce into the language constants for every element of C as in Section 5. Instead, we introduce the predicate IsPRI :: term → bool called primitive-term test that checks whether a term is a constant: IsPRI (x) = (x ∈ C) and the predicate IsPRIL :: leafset → bool called primitiveleafset test: IsPRIL (hs, Li) = (s = cs ) Instead of the rule (16), we have for f, g ∈ Σ ∪ {PRI}: W ∀x. Isf (x) f ∈Σ∪{PRI} ∀x. ¬(Isf (x) ∧ Isg (x)), (90) for f 6≡ g Analogous rules hold for term algebra of leafsets: W ∀x. Isf L (x) f ∈Σ∪{PRI} ∀x. ¬(Isf L (x) ∧ IsgL (x)), A Logic for Term-Power Algebras 6.3 Some Properties of Term-Power Structure In this section we establish some further properties of the term-power structure, including the homomorphism properties between the term algebra of terms and the term algebra of leafsets. We also argue that it suffices to consider a restricted class of formulas called simple formulas. (91) for f 6≡ g Term algebra of shapes satisfies the original rules (16) of term algebra. 32 term algebra on terms constructors, f ∈ Σ: per-shape product structure inner formula relations for r ∈ LC : r :: :: shape × leafset × leafset → leafset ¬I :: leafset → leafset trueI , falseI :: leafset :: term → bool IsPRI :: term → bool selectors, f ∈ Σ: fi inner formula quantifiers: ∃ ,∀ Isf primitive-term test: ∧I , ∨I I :: termk → term constructor test, f ∈ Σ: shape × termk → leafset inner logical connectives: I f :: term → term term shape: :: shape × (term → leafset) → leafset :: leafset × leafset → bool sh :: term → shape leafset equality: =L constructors, f ∈ Σ: leafset cardinality constraints, k ≥ 0: | | ≥ k, | | = k :: shape × leafset → bool fL :: leafsetk → leafset constructor test, f ∈ Σ: leafset quantifiers: ∃L , ∀L term algebra on leafsets :: (leafset → bool) → bool Isf L :: leafset → bool primitive-leafset test: term equality: = :: IsPRIL term × term → bool selectors, f ∈ Σ: term quantifiers: ∃, ∀ :: (term → bool) → bool fiL :: leafset → leafset leafset shape: shape equality: =s :: leafset → bool :: shape × shape → bool :: (shape → bool) → bool lssh :: leafset → shape shape quantifiers: ∃s , ∀s term algebra on shapes constructors, f ∈ Σ0 : logical connectives: ∧, ∨ :: bool × bool → bool ¬ :: bool → bool true, false, undef :: bool fs :: shapek → shape constructor test, f ∈ Σ0 : Isf s :: shape → bool selectors, f ∈ Σ: fis :: shape → shape Figure 9: Operations and relations in structure P 33 interpretation of sorts JtermK = FT(Σ′ ) JshapeK = FT(Σ0 ) JleafsetK JboolK = {hs, Li | L ⊆ leaves(s)} = {true, false, undef} semantics inner formula relations for r ∈ LC : JrK(s, t1 , . . . , tk ) well-definedness = hs, {l | JrKC (t1 [l], . . . , tk [l])}i sh(t1 ) = s ∧ . . . ∧ sh(tk ) = s = hs, L1 ∩ L2 i s1 = s ∧ s2 = s = hs, L1 ∪ L2 i s1 = s ∧ s2 = s = hs, leaves(s) \ L1 i s1 = s inner logical connectives: J∧I K(s, hs1 , L1 i, hs2 , L2 i) I J∨ K(s, hs1 , L1 i, hs2 , L2 i) I J¬ K(s, hs1 , L1 i) I = hs, leaves(s)i I = hs, ∅i Jtrue K(s) Jfalse K(s) inner formula quantifiers, for h : JtermK → JleafsetK: S J∃I K(s, h) = hs, {L | ∃t ∈ JtermK. sh(t) = s ∧ h(t) = hs, Li}i T J∀I K(s, h) = h {L | ∃t ∈ JtermK. sh(t) = s ∧ h(t) = hs, Li}, i ∀t ∈ JtermK. lssh(h(t)) = s ∀t ∈ JtermK. lssh(h(t)) = s leafset equality: J=L K(hs1 , L1 i, hs2 , L2 i) = s1 = s2 ∧ L1 = L2 leafset cardinality constraints: J|hs1 , L1 i|s ≥ kK = (|L1 | ≥ k) s1 = s J|hs1 , L1 i|s = kK = (|L1 | = k) s1 = s leafset quantifiers, for h : JleafsetK → JboolK: J∃L Kh = ∃hs, ti ∈ JleafsetK. h(hs, ti) L = ∀hs, ti ∈ JleafsetK. h(hs, ti) J∀ Kh term equality: J=K(t1 , t2 ) = (t1 = t2 ) term quantifiers, for h : JtermK → JboolK: J∃Kh = ∃t ∈ JtermK. h(t) J∀Kh = ∀t ∈ JtermK. h(t) shape equality: J=s K(ts1 , ts2 ) = (ts1 = ts2 ) shape quantifiers, for h : JshapeK → JboolK: J∃s Kh = ∃t ∈ JshapeK. h(t) s = ∀t ∈ JshapeK. h(t) J∀ Kh Figure 10: Semantics for Logic of Term-Power Algebra (Part I) 34 semantics term algebra on terms well-definedness constructors, f ∈ Σ: Jf K(t1 , . . . , tk ) = f (t1 , . . . , tk ) constructor test, f ∈ Σ: JIsf K(t) = ∃t1 , . . . , tk . t = f (t1 , . . . , tk ) primitive-term test: JIsPRI K(t) = (t ∈ C) selectors, f ∈ Σ: Jfi K(t) = ǫti . t = f (t1 , . . . , ti , . . . , tk ) JIsf K(t) term shape: Jsh(f (t1 , . . . , tn ))K = shapified(f )(sh(t1 ), . . . , sh(tn )) term algebra on leafsets constructors, f ∈ Σ: Jf L K(hs1 , L1 i, . . . , hsk , Lk i) = hf (s1 , . . . , sk ), ({1} · L1 ) ∪ · · · ∪ ({k} · Lk )i constructor test, f ∈ Σ: JIsf L K(hs, Li) = ∃s1 , L1 , . . . , sk , Lk . hs, Li = Jf L K(hs1 , L1 i, . . . , hsk , Lk i) primive-leafset test: JIsPRIL K(hs, Li) = (s = cs ) selectors, f ∈ Σ: JfiL K(hs, Li) = ǫhsi , Li i. hs, Li = Jf L K(hs1 , L1 i, . . . , hsi , Li i, . . . , hsk , Lk i) JIsf L K(hs, Li) leafset shape: JlsshK(hs, Li) = s term algebra on shapes constructors, f ∈ Σ: Jf s K(s1 , . . . , sk ) = f s (s1 , . . . , sk ) constructor test, f ∈ Σ0 : JIsf s K(s) = ∃s1 , . . . , sk . s = f s (s1 , . . . , sk ) Jfis K(s) = ǫsi . s = f s (s1 , . . . , si , . . . , sk ) selectors, f ∈ Σ: Figure 11: Semantics for Logic of Term-Power Algebra (Part II) 35 JIsf s K(s) Recall that r = ∈ R is the equality relation on C. Given r = , we can express the equality between terms by: t1 = t2 ⇐⇒ According to the semantics in Figure 10, sh is a homomorphism from the term algebra of terms to the term algebra of shapes. In addition, lssh is a homomorphism from the term algebra of leafsets to the term algebra of shapes. We also have the following important property. Let r ∈ LC be a relation symbol of arity n, let f ∈ Σ be a function symbol of arity k, and let r =′ (t1 , t2 ) sh(t2 ) = sh(t1 ) ∧ r = (t1 , t2 ) = trueIsh(t1 ) (92) We define the notion of a us -term as in Definition 40 except that we use different symbols for boolean algebra operations. ⇐⇒ sh(t1j ) = . . . = sh(tnj ) = sj Definition 72 (us -terms) Let us ∈ Vars be a shape variable. The set of us -terms Term(us ) is the least set such that: for 1 ≤ j ≤ k. If f s = shapified(f ), f L = leafified(f ), and s = f s (s1 , . . . , sk ) then 1. uL ∈ Term(us ) for every leafset variable uL ; 2. falseIus , trueIus rs (f (t11 , . . . , t1k ), . . . , f (tn1 , . . . , tnk )) = s ∈ Term(u ); f L (rs1 (t11 , . . . , tn1 ), . . . , rsk (t1k , . . . , tnk )) 3. if tL1 , tL2 ∈ Term(us ), then also (93) Furthermore, if lssh(lj ) = lssh(lj′ ) = sj for 1 ≤ j ≤ k and s = f s (s1 , . . . , sk ) then tL1 ∧Ius tL2 ∈ Term(us ), tL1 ∨Ius tL2 ∈ Term(us ), and f L (l1 , . . . , lk ) ∧Is f L (l1′ , . . . , lk′ ) =L ¬Ius tL1 ∈ Term(us ) f L (l1 ∧Is1 l1′ , . . . , lk ∧Isk lk′ ) f L (l1 , . . . , lk ) ∨Is f L (l1′ , . . . , lk′ ) =L If ts is a term of shape sort, the notion of ts -inner formula is defined as follows. f L (l1 ∧Is1 l1′ , . . . , lk ∧Isk lk′ ) Definition 73 (us -inner formula) Let us ∈ Vars be a shape variable. The set of us -inner formulas Inner(us ) is the least set such that: ¬Is f L (l1 , . . . , lk ) =L f L (¬Is1 l1 , . . . , ¬Isk lk ) 1. if u1 , . . . , uk are term variables and r ∈ LC such that ar(r) = k, then (94) ∃Is t.f L (h1 (t), . . . , hk (t)) =L f L (∃Is1 t.h1 (t), . . . , ∃Isk t.hk (t)) rus (u1 , . . . , uk ) ∈ Inner(us ) ∀Is t.f L (h1 (t), . . . , hk (t)) =L 2. falseIus , trueIus ∈ Inner(us ) f L (∀Is1 t.h1 (t), . . . , ∀Isk t.hk (t)) 3. if φ1 , φ2 ∈ Inner(us ) then also From these properties by induction we conclude that if φ(us , u1 , . . . , un ) is an inner formula, then φ1 ∧Ius φ2 ∈ Inner(us ) φ1 ∨Ius φ2 ∈ Inner(us ) φ(s, f (t11 , . . . , t1k ), . . . , f (tn1 , . . . , tnk )) = ¬Ius φ1 ∈ Inner(us ) f L (φ(s1 , t11 , . . . , tn1 ), . . . , φ(sk , t1k , . . . , tnk )) 4. if φ ∈ Inner(us ) and u is a term variable that does not occur in us , then also (95) Let φ(us , u1 , . . . , un ) be an inner formula and let φ′ (u1 , . . . , un ) be a first-order formula that results from replacing operations ∧Is , ∨Is , ¬Is , ∀Is , ∃Is by ∧, ∨, ¬, ∀, ∃. Interpreting φ′ (u1 , . . . , un ) over the structure C yields a relation ρ′ ⊆ C n . If sh(t1 ) = . . . = sh(tk ) = s ∃Ius u.φ ∈ Inner(us ) ∀Ius u.φ ∈ Inner(us ) If φ ∈ Inner(us ) and u1 , . . . , un is the set of free term variables of φ, we write φ(us , u1 , . . . , un ) for φ. Furthermore, if ts is a term of shape sort and t1 , . . . , tn terms of term sort, we write φ(ts , t1 , . . . , tn ) for then JφK(s, t1 , . . . , tk ) = hs, {l | ρ′ (t1 [l], . . . , tk [l])}i φ[us := ts , u1 := t1 , . . . , un := tn ] The following Definition 75 introduces a more restricted set of formulas than the set of formulas permitted by sort declarations in Figure 9. We call this restricted set of formulas simple formulas. One of the main properties of simple formulas compared to arbitrary formulas is that simple formulas allow the use of operations ∃I , ∀I , and relations r , r ∈ LC only within instances of us -inner formulas. where we assume that variables bound by ∃I and ∀I are renamed to avoid the capture of variables that are free in ts , t1 , . . . , tn . We call φ(ts , t1 , . . . , tn ) an instance of the us -inner formula φ(us , u1 , . . . , un ). If φ(us , u1 , . . . , un ) is an inner formula, we abbreviate it by writing [φ′ (u1 , . . . , un )]us where φ′ results from φ(us , u1 , . . . , un ) by omitting the shape argument us from the operations occurring in φ(us , u1 , . . . , un ). Similarly, we write [φ′ (t1 , . . . , tn )]ts for φ(ts , t1 , . . . , tn ). Definition 74 A simple operation is any operation or relation in Figure 9 except for operations ∃I , ∀I , and relations r for r ∈ LC . 36 Because the top-level sort of a formula is bool, every term tL0 of sort leafset occurs within some formula tL1 =L tL2 or Isf L (tL ), |tL |ts = k, |tL |ts ≥ k or as part of some term lssh(tL ). We can replace Isf L (tL ) with Definition 75 (Simple Formulas) The set of simple formulas is the least set that satisfies the following. 1. if φ(us , u1 , . . . , un ) is a an inner formula, ts a term of shape sort, t1 , . . . , tn terms of term sort and uL is a leafset variable, then L L ∃uL . uL =L tL ∧′ Isf L (uL ) s u = φ(t , t1 , . . . , tn ) according to Lemma 10, so we need not consider that case. We can similarly eliminate non-variable leafset terms from cardinality constraints. If a leafset term tL occurs in an expression lssh(tL ), we consider the smallest atomic formula ψ(lssh(tL )) enclosing lssh(tL ), and replace ψ(tL ) with is a simple formula. 2. applying simple operations to simple formulas yields simple formulas. ∃uL . uL =L tL ∧′ ψ(uL ) Example 76 A formula (96) This transformation is valid by Lemma 10 because ψ and lssh are strict. We further assume that in every atomic formula tL1 =L tL2 , the term tL1 is a leafset variable. Suppose that a term tL in a formula uL =L tL is not an instance of an inner formula. Then there are two possibilities. is a simple formula equivalent to formula (96). We abbreviate ∃Ius1 u. rus1 (u, u) as [∃I u. r(u, u)]us1 . 1. There are some occurrences of leafset term algebra operations f L , fiL or leafset variables uL1 in tL . Here by “occurrence” in tL we mean occurrence that is reachable without going through a shape argument or a relation, but only through operations ∀I , ∃I , ∧I , ∨I , ¬I . For example, we ignore the occurrences of f L , fiL within terms ts that occur in ∧Its . uL =L ∃Ius1 u. rus2 (u, u) is not a simple formula for us1 6≡ us2 . Formula (us1 = us2 ∧ uL =L ∃Ius1 u. rus1 (u, u)) ∨ (us1 6= us2 ∧ undef)  Lemma 77 shows that for every formula in the logic of Figure 9 there exists an equivalent simple formula. Note that even simple formulas are sufficient to express the relations of structural subtyping. A reader not interested in the decidability of the more general logic of Figure 9 may therefore ignore Lemma 77. 2. not all shape arguments in ∀I , ∃I , ∧I , ∨I , ¬I , trueI , falseI , r occurring in tL are syntactically identical. We eliminate the first possibility by propagating leafset term algebra operations f L , fiL inwards until they reach expressions of form Lts (t1 , . . . , tn ), applying the equations (94) from left to right. We then convert f L , fiL operations of term algebra of leafsets into operations of the term algebra of terms applying (93) from right to left. To eliminate the second possibility, let ts1 , . . . , tsn be the occurrences (reachable through trueI , falseI , ∧I , ∨I , ¬I , ∀I , ∃I ) in term tL of the shape arguments of operations trueI , falseI , ∧I , ∨I , ¬I , ∀I , ∃I . Then replace Lemma 77 (Formula Simplification) For every welldefined formula in the logic of Figure 9 there exists an equivalent well-defined simple formula. Proof Sketch. According to the definition of simple formula, we need to ensure that every occurrence of quantifiers ∀I , ∃I and relations r is an occurrence in some inner-formula instance φ(ts , t1 , . . . , tn ). Each occurrence rts (t1 , . . . , tn ) is an inner formula instance by itself, so the main difficulty is fitting the quantifiers ∀I and ∃I into inner formulas. Let us examine the syntactic structure of formulas of logic in Figure 9. This syntactic structure is determined by sort declarations. Each expression of leafset is formed starting from uL = tL (ts1 , . . . tsn ) with s s s ′ ′ CL s s s ′ (∃s us . ∀CL 1 (u = t1 ) ∧ . . . ∧ ∀n (u = tn )∧ 1. relations r ∈ LC ; uL =WL tL (us , . . . , us )) ∨ (undef ∧ 1≤i<j≤n tsi 6= tsj ) 2. leafset variables; 3. trueI , falseI I I I I I L Here ∀CL denotes universal quantification ∀ui,1 , . . . , ui,ni i where ui,1 , . . . , ui,ni is a list of those term variables occurring in tsi that are bound by some quantifier ∃I , ∀I within tL . fiL . using operations ∧ , ∨ , ¬ , ∀ , ∃ , as well as f and The leafset expressions can be used in a formula in the following ways (in addition to constructing new leafset expressions): 1. to compare for equality using =L ; 6.4 2. to test for the top-level constructor using Isf L ; Quantifier Elimination In this section we give a quantifier elimination procedure for the term-power structure. The procedure of this section is applicable whenever C is a structure with a decidable firstorder theory. 3. to form leafset cardinality constraints; 4. to form a shape using lssh. 37 where n = p + q, nL = pL + q L , ns = ps + q s , and formulas shapeBase, leafsetBase, termBase, leafsetHom, termHom, cardin, innerMap are defined as follows. Definition 78 below generalizes the notion of structural base formula of Definition 41, Section 4.3. There are two main differences between Definition 41 and the present Definition 78. The first difference is the presence of three (instead of two) base formulas: shape base, leafset base, and term base. This difference is a consequence of the distinction between leafsets and terms and is needed whenever base structure C has more than two elements. There is a homomorphism formula relating leafset base formula to shape base formula and a homomorphism formula relating term base formula to shape base formula. Furthermore, some of the leafset variables are determined by term variables using inner formula maps, which establishes the relationship between term base formula and leafset base formula. Cardinality constraints now apply to leafset variables. The second difference is the distinction between composed and primitive non-parameter leafset and term variables. A composed non-parameter variable denotes a leafset or a term whose shape s has property Isf s (s) for some f ∈ Σ. A primitive non-parameter variable denotes a leafset or a term whose shape is cs and has property IsPRI or IsPRIL . The purpose of this distinction is to allow cardinality constraints and inner formula maps not only on parameter variables, but also on primitive non-parameter variables, which is useful when the base structure C is decidable but infinite. shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) = s p V usi = ti (us1 , . . . , usns ) ∧ i=1 ∧ distinct(us1 , . . . , usn ) • internal primitive ur+1 , . . . , up ; non-parameter term term rL V i=1 pL V uLi = ti (uL1 , . . . , uLnL ) ∧ IsPRIL (uLi ) ∧ i=r L +1 m VL i=1 xLi = uLji where each ti is a term of form f (ui1 , . . . , uik ) for some f ∈ Σ, k = ar(f ), and j : {1, . . . , mL } → {1, . . . , nL } is a function mapping indices of free leafset variables to indices of internal leafset variables. leafsetHom(uL1 , . . . , uLnL , us1 , . . . , usns ) = variables n VL i=1 lssh(uLi ) = usji where j : {1, . . . , nL } → {1, . . . , ns } is some function such that {j1 , . . . , jp } ⊆ {1, . . . , ps } and {jpL +1 , . . . , jpL +qL } ⊆ {ps +1, . . . , ps +q s } (a leafset variable is a parameter variable iff its shape is a parameter shape variable). variables • internal parameter term variables up+1 , . . . , up+q ; • free leafset variables xL1 , . . . , xLmL ; termBase(u1 , . . . , un , x1 , . . . , xm ) = • internal composed non-parameter leafset variables uL1 , . . . , uLrL ; r V i=1 p V • internal primitive non-parameter leafset variables uLrL +1 , . . . , uLpL ; i=r+1 m V • internal parameter leafset variables upL +1 , . . . , upL +qL ; • free shape variables xsi = usji leafsetBase(uL1 , . . . , uLnL , xL1 , . . . , xLmL ) = • free term variables x1 , . . . , xm ; non-parameter i=1 where each ti is a shape term of form f s (usi1 , . . . , usik ) for some f ∈ Σ0 , k = ar(f ), and j : {1, . . . , ms } → {1, . . . , ns } is a function mapping indices of free shape variables to indices of internal shape variables. Definition 78 (Structural Base Formula) A structural base formula with: • internal composed u1 , . . . , ur ; m Vs ui = ti (u1 , . . . , un ) ∧ IsPRI (ui ) ∧ xi = uji i=1 xs1 , . . . , xsms ; where each ti is a term of form f (ui1 , . . . , uik ) for some f ∈ Σ, k = ar(f ), and j : {1, . . . , m} → {1, . . . , n} is a function mapping indices of free term variables to indices of internal term variables. • internal non-parameter shape variables us1 , . . . , usps ; • internal parameter shape variables usps , . . . , usps +qs is a formula of form: termHom(u1 , . . . , un , us1 , . . . , usns ) = ∃u1 , . . . , un , uL1 , . . . , uLnL , us1 , . . . , usns . n V i=1 shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) ∧ leafsetBase(uL1 , . . . , uLnL , xL1 , . . . , xLmL ) ∧ leafsetHom(uL1 , . . . , uLnL , us1 , . . . , usns ) ∧ sh(ui ) = usji where j : {1, . . . , n} → {1, . . . , ns } is some function such that {j1 , . . . , jp } ⊆ {1, . . . , ps } and {jp+1 , . . . , jp+q } ⊆ {ps + 1, . . . , ps + q s } (a term variable is a parameter variable iff its shape is a parameter shape variable). termBase(u1 , . . . , un , x1 , . . . , xm ) ∧ cardin(uLrL +1 , . . . , uLnL , usps +1 , . . . , usns ) = ψ1 ∧ · · · ∧ ψd termHom(u1 , . . . , un , us1 , . . . , usns ) ∧ where each ψi is of form cardin(uLrL +1 , . . . , uLnL , usps +1 , . . . , usns ) ∧ |tL (uLrL +1 , . . . , uLnL )|us = k innerMap(ur+1 , . . . , un , uLrL +1 , . . . , uLnL , usps +1 , . . . , usns ) 38 or As in Section 3.4 and Section 4.3 we proceed to show that each quantifier-free formula can be written as a disjunction of base formulas and each base formula can be written as a quantifier-free formula. We first give a small example to illustrate how the techniques of Section 4.3 extend to the more general case of Σ-term-power. |tL (uLrL +1 , . . . , uLnL )|us ≥ k for some us -term tL (uLrL +1 , . . . , uLnL ) that contains no variables other than some of the variables uLrL +1 , . . . , uLnL , and the following condition holds: If a variable uLj for r L + 1 ≤ j ≤ nL occurs in the term tL (uLrL +1 , . . . , uLnL ), then lssh(uLj ) = us occurs in formula leafsetHom. Example 79 We solve one subproblem from Example 42 using the language of term-power algebras. Consider the formula (97) ∃v. g(v, z) ≤ g(z, v) ∧ Isg (v) ∧ Isg (w) ∧ innerMap(ur+1 , . . . , un , uLrL +1 , . . . , uLnL , usps +1 , . . . , usns ) = ¬(g1 (w) ≤ g1 (v)) η1 ∧ · · · ∧ ηe Formula (99) is in the language of Figure 8, with ≤ a binary lifted relation. After converting (99) into the language of Figure 9 we obtain as one of the possible cases formula: where each ηi is of form uLj I s = φ (u , ui1 , . . . , uik ) for some inner formula φI (us , ui1 , . . . , uik ) ∈ Inner(us ) where L + 1 ≤ j ≤ nL i.e. uLj is a primitive non-parameter leafset variable or parameter leafset variable, {ui1 , . . . , uik } ⊆ {ur+1 , . . . , un } are primitive nonparameter term variables and parameter variables, the conjunct lssh(uL ) = us occurs in leafsetHom, and the following condition holds: sh(uij ) = us occurs in formula termHom for every j where 1 ≤ j ≤ k. (99) ∃v. [g(v, z)  g(z, v)]sh(g(z,v)) =L trueIsh(g(z,v)) ∧ sh(g(z, v)) =s sh(g(z, v)) ∧ Isg (v) ∧ Isg (w) ∧ [g1 (w)  g1 (v)]sh(g1 (w)) 6= (100) L trueIsh(g1 (w)) ∧ s sh(g1 (v)) = sh(g1 (w)) (98) where  is the subtyping relation on the base structure C so that ≤ = ′ . We next transform the formula into unnested form, obtaining: We require each structural base formula to satisfy the following conditions: P0) the graph associated with shape base formula ∃v, uvz , uzv , uw1 , uv1 . ∃L uLvz , uLw1 . ∃s usvz , usw1 . ∃us1 , . . . , usns . shapeBase(us1 , . . . , usns , xs1 , . . . , xsms ) uvz = g(v, z) ∧ uzv = g(z, v) ∧ is acyclic (compare to Definition 21); uw1 = g1 (w) ∧ uv1 = g1 (v) ∧ P1) congruence closure property for shapeBase subformula: there are no two distinct variables usi and usj such that both usi = f (usl1 , . . . , uslk ) and usj = f (usl1 , . . . , uslk ) occur as conjuncts in formula shapeBase; usvz =s sh(uvz ) ∧ usw1 =s sh(uw1 ) ∧ P2) congruence closure property for leafsetBase subformula: there are no two distinct variables uLi and uLj such that both uLi = f L (uLl1 , . . . , uLlk ) and uLj = f L (uLl1 , . . . , uLlk ) occur as conjuncts in formula leafsetBase; uLvz =L [uvz  uzv ]usvz sh(uzv ) =s usvz ∧ sh(uv1 ) =s usw1 ∧ Isg (v) ∧ Isg (w) ∧ (101) |¬uLvz |usvz = 0 ∧ uLw1 =L [uw1  uv1 ]usw1 ∧ P3) congruence closure property for termBase subformula: there are no two distinct variables ui and uj such that both ui = f (ul1 , . . . , ulk ) and uj = f (ul1 , . . . , ulk ) occur as conjuncts in formula termBase; |¬uLw1 | ≥ 1 We next transform (101) into disjunction of base formulas. A typical base formula is: P4) homomorphism property of lssh: for every nonparameter leafset variable uL such that uL = f L (uLi1 , . . . , uLik ) occurs in leafsetBase, if conjunct lssh(uL ) = us occurs in leafsetHom, then for some shape variables usj1 , . . . , usjk term us = f s (usj1 , . . . , usjk ) occurs in shapeBase where f s = shapified(f ) and for every r where 1 ≤ r ≤ k, conjunct lssh(uir ) = usjr occurs in leafsetHom. ∃uvz , uzv , uv , uz , uw , uv1 , uv2 , uz1 , uz2 , uw1 , uw2 . ∃L uLvz , uLv , uLz , uLv1 , uLv2 , uLz1 , uLz2 , uLw1 . ∃s usvz , usw , usw1 , usw2 . shapeBase1 ∧ (102) leafsetBase1 ∧ leafsetHom1 ∧ termBase1 ∧ termHom1 ∧ P5) homomorphism property of sh: for every non-parameter term variable u such that u = f (ui1 , . . . , uik ) occurs in termBase, if conjunct sh(u) = us occurs in termHom, then for some shape variables usj1 , . . . , usjk term us = f s (usj1 , . . . , usjk ) occurs in shapeBase where f s = shapified(f ) and for every r where 1 ≤ r ≤ k, conjunct sh(uir ) = usjr occurs in termHom. cardin1 ∧ innerMap1 shapeBase1 = usvz = g s (usw , usw ) ∧ usw = g s (usw1 , usw2 ) ∧ distinct(usvz , usw , usw1 , usw2 ) 39 leafsetBase1 = uLvz = g L (uLv , uLz ) ∧ uLv =g L (uLv1 , uLv2 ) ∧ and uLz =g L (uLz1 , uLz2 ) ψ2 ≡ ∃uv2 .∃L uLv2 , uLz2 . leafsetHom1 = lssh(uLvz ) = usvz ∧ lssh(uLv ) = usw ∧ lssh(uLz ) = usw ∧ sh(uv2 ) =s usw2 ∧ sh(uz2 ) =s usw2 ∧ lssh(uLv1 ) = usw1 ∧ lssh(uLv2 ) = usw2 ∧ sh(uLv2 ) =s usw2 ∧ sh(uLz2 ) =s usw2 ∧ lssh(uLz1 ) = usw1 ∧ lssh(uLz2 ) = usw2 ∧ uLv2 =L [uv2  uz2 ]usw2 ∧ uLz2 =L [uz2  uv2 ]usw2 ∧ lssh(uLw1 ) = usw1 |¬uLv2 |usw2 = 0 ∧ |¬uLz2 |usw2 = 0 Formula ψ1 expresses a fact in a structure isomorphic to the power C n where n is the number of leaves in the shape denoted by usw1 . Similarly, ψ2 expresses a fact in a product structure C m where m is the number of leaves in the shape denoted by usw2 . We can therefore use the technique of Feferman-Vaught technique (Section 3.3) to eliminate the quantifiers from formulas ψ1 and ψ2 . According to Example 17, ψ1 is equivalent to: termBase1 = uvz = g(uv , uz ) ∧ uzv = g(uz , uv ) ∧ uv = g(uv1 , uv2 ) ∧ uz = g(uz1 , uz2 ) ∧ uw = g(uw1 , uw2 ) ∧ z = uz ∧ w = uw termHom1 = sh(uvz ) = usvz ∧ sh(uzv ) = usvz ∧ ∃L uL0 , uL4 . sh(uv ) = usw ∧ sh(uz ) = usw ∧ sh(uw ) = usw ∧ sh(uv1 ) = usw1 sh(uv2 ) = usw2 ∧ sh(uz1 ) = usw1 ∧ sh(uz2 ) = usw2 ∧ sh(uw1 ) = usw1 ∧ sh(uw2 ) = usw2 uL0 =L [∃I t. t  uz1 ∧I uz1  t ∧I uw1  t]usw1 ∧ ∧ uL4 =L [∃I t. t  uz1 ∧I uz1  t ∧I ¬I uw1  t]usw1 ∧ |uL4 |usw1 ≥ 1 ∧ |¬I uL0 ∧I ¬I uL4 |usw1 = 0 innerMap1 = We similarly apply Feferman-Vaught construction to ψ2 and obtain the result true. We may now substitute the results of quantifier elimination in ψ1 and ψ2 . The resulting formula is: uLv1 =L [uv1  uz1 ]usw1 ∧ uLz1 =L [uz1  uv1 ]usw1 ∧ uLv2 =L [uv2  uz2 ]usw2 ∧ uLz2 =L [uz2  uv2 ]usw2 ∧ uLw1 =L [uw1  uv1 ]usw1 ∃uvz , uzv , uv , uz , uw , uv1 , uv2 , uz1 , uz2 , uw1 , uw2 . cardin1 = |¬uLv1 |usw1 = 0 ∧ |¬uLz1 |usw1 = 0 ∧ ∃L uLvz , uLv , uLz , uLv1 , uLv2 , uLz1 , uLz2 , uLw1 . ∃s usvz , usw , usw1 , usw2 . |¬uLv2 |usw2 = 0 ∧ |¬uLz2 |usw2 = 0 ∧ |¬uLw1 |usw1 | ≥ 1 shapeBase1 ∧ leafsetHom2 ∧ We next show how to transform the base formula (102) into quantifier-free form. We substitute away non-parameter term variables uvz , uzv , uv and non-parameter leafset variables uLvz , uLv , uLz , because the homomorphism constraints they participate in may be derived from the remaining conjuncts. We next eliminate parameter term variables uv1 , uv2 and parameter leafset variables uLv1 , uLv2 , uLz1 , uLz2 , uLw1 . Grouping the conjuncts in cardin1 and innerMap1 by their shape, we may extract the subformulas ψ1 and ψ2 of (102). termBase2 ∧ termHom2 ∧ cardin1 ∧ innerMap1 where leafsetHom2 = lssh(uL0 ) = usw1 ∧ lssh(uL4 ) = usw1 termBase2 = uz = g(uz1 , uz2 ) ∧ uw = g(uw1 , uw2 ) ∧ z = uz ∧ w = uw ψ1 ≡ innerMap2 = ∃uv1 .∃L uLv1 , uLz1 , uLw1 . sh(uv1 ) =s usw1 ∧ sh(uz1 ) =s usw1 ∧ sh(uw1 ) =s usw1 ∧ uL0 =L [∃I t. t  uz1 ∧I uz1  t ∧I uw1  t]usw1 ∧ lssh(uLv1 ) =s usw1 ∧ lssh(uLz1 ) =s usw1 ∧ uL4 =L [∃I t. t  uz1 ∧I uz1  t ∧I ¬I uw1  t]usw1 ∧ lssh(uLw1 ) =s usw1 ∧ cardin2 = |uL4 |usw1 ≥ 1 ∧ |¬I uL0 ∧I ¬I uL4 |usw1 = 0 uLw1 =L [uw1  uv1 ]usw1 ∧ In the resulting formula all variables are expressible in terms of free variables, so we can write the formula without quantifiers ∃, ∀, ∃L , ∀L . |¬uLv1 |usw1 = 0 ∧ |¬uLz1 |usw1 = 0 ∧  uLv1 =L [uv1  uz1 ]usw1 ∧ uLz1 =L [uz1  uv1 ]usw1 ∧ |¬uLw1 |usw1 | ≥1 The following Proposition 80 is analogous to Proposition 44; the proof is straightforward. 40 5. formulas of term algebra of shapes: Proposition 80 (Quantification of Struct. Base) If β is a structural base formula and x a free shape, leafset, or term variable in β, then there exists a base structural formula β1 equivalent to ∃x.β. us1 =s us2 , us1 6=s us2 us0 =s f s (us1 , . . . , usn ) The following Proposition 81 corresponds to Proposition 45. us =s fis (us0 ) Proposition 81 (Quantifier-Free to Structural Base) Let φ be a well-defined simple formula without quantifiers ∃L , ∀L , ∃, ∀, ∃s , ∀s . Then φ can be written as true, false, or a disjunction of structural base formulas. Isf s (us0 ), ¬Isf s (us0 ) We next describe transformation of each existentially quantified conjunction. In the sequel, whenever we perform case analysis and generate a disjunction of conjunctions, existential quantifiers propagate to the conjunctions, so we keep working with a existentially quantified conjunction. The existentially quantified variables will become internal variables of a structural base formula. Analogously to the proof of Proposition 28, we use (90), (91), (16) to eliminate literals ¬Isf (u0 ), ¬Isf L f L (uL0 ), ¬Isgs (us0 ). As in the proof of Proposition 45, we replace formulas of leafset boolean algebra by cardinality constraints, similarly to Figure 7. We next convert formulas of term algebra of terms into a base formula, formulas of term algebra of leafsets into a base formula, and formulas of term algebra of shapes into a base formula. We simultaneously make sure that every term or leafset variable has an associated associated shape variable, introducing new shape variables if needed. We also ensure homomorphism requirements by replacing internal variables when we entail their equality. Another condition we ensure is that parameter term variables map to parameter shape variables, and non-parameter term variables to non-parameter shape variables; we do this by performing expansion of term and shape variables. We perform expansion of shape variables as in Section 3.2. Expansion of term and variables is even simpler because there is no need to do case analysis on equality of term variable with other variables. We eliminate disequality between term variables using (92). We eliminate disequalities between leafset variables as in Example 43, by converting each disequality into a cardinality constraint. Elimination of disequalities might violate previously established homomorphism invariants, so we may need to reestablish these invariants by repeating the previously described steps. The overall process terminates because we never introduce new inequalities between term or leafset variables. As a final step, we convert all cardinality constraints into constraints on parameter term variables, using (95). In the case when the shape of cardinality constraint is cs , we cannot apply (95). However, in this case, unlike Proposition 45, we do not do case analysis on all possible constant leafsets (this is not even possible in general). This is because Definition 78, unlike Definition 41 implies no need to further decompose cardinality constraints in that case, because we allow primitive non-parameter leafset variables. This completes our sketch of transforming a quantifierfree formula into disjunction of structural base formulas. Proof Sketch. The overall idea of the transformation to base formula is similar to the transformation in the proof of Proposition 45. Additional complexity is due to inner formulas. However, note that an inner formula φ(us , u1 , . . . , un ) is well-defined iff δ(us , u1 , . . . , un ) holds where δ(us , u1 , . . . , un ) ≡ sh(u1 ) = us ∧ . . . ∧ sh(un ) = us Hence, each formula φ(us , u1 , . . . , un ) can be treated as a partial operation p of sort shape × termn → leafset and the domain given by Dp = hhus , u1 , . . . , un i, δ(us , u1 , . . . , un )i This means that we may apply Proposition 9 and convert formula to disjunction existentially quantified well-defined conjunctions of literals in one of the following forms: 1. equality with inner formulas: uL0 =L φ(us , u1 , . . . , un ) where φ(us , u1 , . . . , un ) is a us -inner formula; 2. formulas of leafset boolean algebra: uL0 =L uL1 ∧Ius uL2 uL0 =L uL1 ∨Ius uL2 uL0 =L ¬Ius uL1 uL0 =L trueIus uL0 =L falseIus 3. formulas of term algebra of terms: u1 = u2 , u1 6= u2 u0 = f (u1 , . . . , un ) u = fi (u0 ) Isf (u0 ), ¬Isf (u0 ) sh(u) = us 4. formulas of term algebra of leafsets: uL1 =L uL2 , uL1 6=L uL2 uL0 =L f L (uL1 , . . . , uLn ) We introduce the notion of determined variables in structural base formula generalizing Definition 29 and Definition 46. For brevity, we write u∗ for internal shape, term, or leafset variables, similarly x∗ for a free variable, t∗ for a term uL =L fiL (uL0 ) Isf L (uL0 ), ¬Isf L (uL0 ) lssh(uL ) =L us 41 and f ∗ for a shape, term, or leafset term algebra constructor and fi∗ for a shape, term, or leafset term algebra selector. Lemma 87 Let uL be an undetermined composed nonparameter leafset variable in a structural base formula β such that uL is a source i.e. no conjunct of form Definition 82 The set dets of variable determinations of a structural base formula β is the least set S of pairs hu∗ , t∗ i where u∗ is an internal term, leafset, or shape variable and t∗ is a term over the free variables of β, such such that: ∗ ′ uL = f L (uL1 , . . . , uL , . . . , uLk ) occurs in leafsetBase. Let β ′ be the result of dropping uL from β. Then β is equivalent to β ′ . ∗ 1. if x = u occurs in termBase, leafsetBase, or shapeBase, then hu∗ , x∗ i ∈ S; ∗ ∗ ∗ ∗ Proof. Because uL is a composed non-parameter term variable, it does not occur in innerMap or cardin, so it only occurs in leafsetBase and leafsetHom. The conjunct containing uL in leafsetHom is a consequence of the remaining conjuncts, so it may be dropped. After that, applying (7) yields a structural base formula β ′ not containing uL , where β ′ is equivalent to β. (u∗1 , . . . , u∗k ) 2. if hu , t i ∈ S and u = f occurs in shapeBase, termBase, or leafsetBase then {hu∗1 , f1∗ (t∗ )i, . . . , hu∗k , fk∗ (t∗ )i} ⊆ S; 3. if {hu∗1 , f1∗ (t∗ )i, . . . , hu∗k , fk∗ (t∗ )i} ⊆ S and u∗ = f ∗ (u∗1 , . . . , u∗k ) occurs in shapeBase, termBase, or leafsetBase then hu∗ , t∗ i ∈ S; 4. if hu, ti ∈ S and sh(u) = us occurs in termHom then hus , sh(t)i ∈ S; Corollary 88 Every base formula is equivalent to a base formula without undetermined composed non-parameter term variables and without undetermined composed nonparameter leafset variables. 5. if huL , tL i ∈ S and lssh(uL ) = us occurs in leafsetHom then hus , lssh(tL )i ∈ S; Proof. If a structural base formula has an undetermined composed non-parameter term variable, then it has an undetermined composed non-parameter term variable that is a source, similarly for leafset variables. By repeated application of Lemma 86 and Lemma 87 we eliminate all undetermined non-parameter term and leafset variables. 6. if uL = φ(us , u1 , . . . , un ) occurs in innerMap where φ(us , u1 , . . . , un ) is an inner formula and {hus , ts i, hu1 , t1 i, . . . , hun , tn i} ⊆ S, then huL , φ(ts , t1 , . . . , tn )i ∈ S. (In the special case when φ contains no free term variables, if hus , ts i ∈ S then huL , φ(us )i ∈ S. The following Proposition 89 corresponds to Proposition 53 and Proposition 66. ∗ Definition 83 An internal variable u is determined if hu∗ , t∗ i ∈ dets for some term ts . An internal variable is undetermined if it is not determined. Proposition 89 (Struct. Base to Quantifier-Free) Every structural base formula β is equivalent to a welldefined simple formula φ without quantifiers ∃L , ∀L , ∃, ∀, ∃s , ∀s . Lemma 84 Let β be a structural base formula with matrix β0 and let dets be the determinations of β. If hu∗ , t∗ i ∈ S then |= β0 ⇒ u∗ = t∗ . Proof Sketch. By Corollary 88 we may assume that β has no undetermined composed non-parameter term and leafset variables. By Corollary 85 we are done if there are no undetermined variables, so it suffices to eliminate: Proof. By induction, using Definition 82. Corollary 85 Let β be a structural base formula such that every internal variable is determined. Then β is equivalent to a well-defined formula without quantifiers ∃L , ∀L , ∃, ∀, ∃s , ∀s . 1. undetermined parameter term variables, 2. undetermined primitive non-parameter term variables, 3. undetermined parameter leafset variables, Proof. By Lemma 84 using (7). 4. undetermined primitive non-parameter leafset variables, and Lemma 86 Let u be an undetermined composed nonparameter term variable in a structural base formula β such that u is a source i.e. no conjunct of form 5. undetermined shape variables. If u is an undetermined parameter term variable or a primitive non-parameter term variable, then u does not occur in termBase, so it occurs only in termHom and innerMap. If uL is an undetermined parameter leafset variable or a primitive non-parameter leafset variable then uL does not occur in leafsetBase, so it occurs only in leafsetHom, innerMap, and cardin. For a undetermined term or leafset variable of shape us such that there is an uncovered parameter or primitive nonparameter term or leafset variable with shape us , consider all conjuncts γi in innerMap of form ′ u = f (u1 , . . . , u, . . . , uk ) occurs in termBase. Let β ′ be the result of dropping u from β. Then β is equivalent to β ′ . Proof. Because u is a composed non-parameter term variable, it does not occur in innerMap, so it only occurs in termBase and termHom. The conjunct containing u in termHom is a consequence of the remaining conjuncts, so it may be dropped. After that, applying (7) yields a structural base formula β ′ not containing u, where β ′ is equivalent to β. uLj = φ(us , ui1 , . . . , uik ) 42 and all conjuncts δi from cardin of form: 6.5 In this section we discuss the decidability of the Σ-termpower structure for a decidable theory C when some of the function symbols f ∈ Σ are contravariant. We then suggest a generalization of the notion of variance to multiple relations and to relations with arity greater than two. The modifications needed to accommodate contravariance with respect to some distinguished relation symbol ≤∈ R for the case of infinite C are analogous to the modifications in Section 5.5. We this obtain a quantifier elimination procedure for any decidable theory C in the presence of contravariant constructors. |tL (uLrL +1 , . . . , uLnL )|us = k or Handling Contravariant Constructors |tL (uLrL +1 , . . . , uLnL )|us ≥ k Together with formulas from termHom and leafsetHom that contain term and leafset variables free in formulas γi and δi , these conjuncts form a formula η which expresses a relation in the substructure of term-power algebra which (because constructors are covariant) is isomorphic to a term-power of C. We therefore use Feferman-Vaught theorem from Section 3.3 to eliminate all term and parameter variables from η. By repeating this process we eliminate all undetermined parameter and leafset variables. It remains to eliminate undetermined shape variables. This process is similar to term algebra quantifier elimination in Section 3.4. An essential part of construction in Section 3.4 is Lemma 25, which relies on the fact that undetermined parameter variables may take on infinitely many values. We therefore ensure that undetermined parameter shape variables are not constrained by term and parameter variables through conjuncts outside shapeBase. An undetermined parameter shape variable us does not occur in termHom or leafsetHom because there are no parameter term and leafset variables, so us can occur only in innerMap and cardin. However, because undetermined parameter and leafset variables are eliminated from the formula, if us is a parameter shape variable then exactly one of these two cases holds: Theorem 92 (Decidability of Structural Subtyping) Let C be a decidable structure and P a Σ-term-power of C. Then the first-order theory of P is decidable. In the rest of this section we consider a generalization that allows defining variance for every relation symbol r ∈ R of any arity, and not just the relation symbol ≤∈ R. For a given relation symbol r ∈ R, function symbol f ∈ Σ, with k = ar(f ), and integer i where 1 ≤ i ≤ k, let Pr (f, i) denote a permutation of the set {1, . . . , k} that specifies the variance of the i-th argument of f with respect to the relation r. For example, if r is a binary relation then Pr (f, i) is the identity permutation {h1, 1ih2, 2i} if i-th argument of f is covariant, or a the transpose permutation {h1, 2i, h2, 1i} if i-th argument of f is contravariant. If l ∈ leaves(s) is a leaf l = hf 1 , i1 i . . . hf n , in i, define the permutation variance(l) as the composition of permutations: 1. there are some conjuncts in innerMap and cardin that contain us and contain some determined term and leafset variables, in this case us is determined, or variance(l) = Pr (f n , in ) ◦ · · · ◦ Pr (f 1 , i1 ) Then define JrK by 2. there are no conjuncts in innerMap containing us and cardin contains only domain cardinality constraints of form |1|us = k and |1|us ≥ k. JrK(s, t1 , . . . , tk ) = hs, {l | JrKC (tp1 [l], . . . , tpk [l]) ∧ Hence, if us is a shape variable it remains to eliminate the constraints of form |1|us = k and |1|us ≥ k. We eliminate these constraints as in the proof of Proposition 66. In the resulting formula all variables are determined. By Corollary 85 the formula can be written as a formula without quantifiers ∃L , ∀L , ∃, ∀, ∃s , ∀s . hp1 , . . . , pk i = variance(l) }i We generalize (76) by defining Nπ (s) = |{l ∈ leaves(s) | variance(l) = π}| The following is the main result of this paper. As in Section 5.5, we can transform the constraints |1|us = k and |1|us ≥ k on each parameter shape variable into a conjunction of constraints of form: Theorem 90 (Term Power Quant. Elimination) There exist algorithms A, B such that for a given formula φ in the language of Figure 9: Nπ (us ) = k a) A produces a quantifier-free formula φ′ in selector language or b) B produces a disjunction φ′ of structural base formulas A problem on nonnegative integers. To solve the problem of variance with any number of relation symbols of any arity, it suffices to solve the following problem on sets of tuples of non-negative integers. Let Nat = {0, 1, 2, . . .}. Consider the structure St = Natd for some d ≥ 2 and let D = {1, 2, . . . , d}. If p is a permutation on D, let Mp denote an operation St → St defined by We also explicitly state the following corollary. Corollary 91 Let C be a structure with decidable first-order theory. Then the set of true sentences in the logic of Figure 9 interpreted in the structure P according to Figures 10 and 11 is decidable. Nπ (us ) ≥ k Mp (x1 , . . . , xd ) = (xp1 , . . . , xpd ) 43 If hx1 , . . . , xd i, hy1 , . . . , yd i ∈ St define The resulting language is at least as expressive as the language in Figure 5. This language is interesting because it allows reasoning about updates to leaves of a tree of fixed shape, thus generalizing the theory of updatable arrays [33] to the theory of trees with update operations, which would be useful for program verification. We did not choose this more expressive language in this report for the following reason. If the base structure C has a finite domain C, then for certain reasonable choice of the relations interpreting LC it is possible to express statements of this extended language in the logic of Figure 9. The idea is to assume a partial order on the elements of C with a minimal element, and use terms t with exactly one leaf non-minimal to model the leaves. On the other hand, in the more interesting case when C is infinite, we can easily obtain undecidable theories in the presence of selection operation. Namely, the selection operation allows terms to be used as finite sets of elements of C. The term-power therefore increases the expressiveness from the first-order theory to the weak monadic second-order theory, which allows quantification over finite sets of objects. Weak monadic theory allows in particular inductive definitions. If theory of structure C is decidable, weak monadic theory might therefore still be undecidable, as an example we might take the term algebra itself, whose weak monadic theory would allow defining subterm relation, yielding an undecidable theory [55, Page 508]. hx1 , . . . , xd i + hy1 , . . . , yd i = hx1 + y1 , . . . , xd + yd i Consider a finite set of operations f : Stk → St where each operation f is determined by k permutations pf1 , . . . , pfk in the following way: f (t1 , . . . , tk ) = Mpf (t1 ) + . . . + Mpf (tk ) 1 k Hence, each operation f of arity k is given by a permutation which specifies how to exchange the order of arguments in the tuple. After permuting the arguments the tuples are summed up. Given a finite set F of operations f , let S be the set generated by operations in F starting from the element (1, 0, . . . , 0) ∈ St. Let C(n1 , . . . , nd ) be a conjunction of simple linear constraints of the forms ni = ai and ni ≥ ai Consider the set AC = {(n1 , . . . , nd ) ∈ S|C(n1 , . . . , nd )} The problem is: For given set of operations F , is there an algorithm that given C(n1 , . . . , nd ) finitely computes the set AC . End of a problem on nonnegative integers. We conjecture that the technique of Lemma 68 can be generalized to yield a solution to the problem on nonnegative integers and thus establish the decidability for the notion of variance with respect to any number of relations with any number of arguments. 6.6 7 Some Connections with MSOL This section explores some relationships between the theory of structural subtyping and monadic second-order logic (MSOL) interpreted over tree-like structures. We present it as a series of remarks that are potentially useful for understanding the first-order theory of structural subtyping of recursive types, see [36, 37] for similar results in the context of the theory of feature trees. In Section 7.1 we exhibit an embedding of MSOL of infinite binary tree into the first-order theory of structural subtyping of recursive types with two constant symbols a,b and one covariant binary function symbol f . MSOL of infinite binary tree is decidable. Although the embedding does not give an answer to the decidability of the structural subtyping of recursive types, it does show that the problem is at least as difficult as decidability of MSOL over infinite trees. We therefore expect that, if the theory of structural subtyping of recursive types is decidable, the decidability proof will likely either use decidability of MSOL over infinite trees, or use directly techniques similar to those of [18, 56]. In Section 7.2 we use the embedding in Section 7.1 to argue the decidability of formulas of the first-order theory of structural subtyping of recursive types where variables range over terms of certain fixed infinite shape se . In Section 7.3 we present an encoding of all terms using terms of shape se . We argue that the main obstacle in using this encoding to show the decidability of the first-order theory of structural subtyping recursive types is inability to define the set of all prefix-closed terms of the shape se . In Section 7.4 we generalize the decidability result of Section 7.2 by allowing different variables to range over different constant shapes. In Section 7.5 we illustrate some of the difficulties in reducing first-order theory of structural subtyping to MSOL over tree-like structures. We show that if we use a certain A Note on Element Selection We make a brief note related to the choice of the language for making statements in term-power algebras. In Section 5 we avoided the use of leafset variables by substituting them into cardinality constraints. In this section we use a cylindric algebra of leafsets. An apparently even more flexible alternative is to allow the element selection operation select :: term × leaf → elem where elem is a new sort, interpreted over the set C, and leaf is a sort interpreted over the set of pairs of a shape and a leaf. Instead of the formula rus (t1 , . . . , tn ) =L trueIus we would then write ∀l. rus (select(t1 , l), . . . , select(tn , l)) =L trueIus Using select operation we can define update relation: update(t1 , l0 , e, t2 ) ≡ ∀l. ((l = l0 ∧ select(t2 , l) = e) ∨ (l 6= l0 ∧ select(t2 , l) = select(t1 , l))) 44 form of infinite feature trees instead of infinite terms, the decidability follows. In Section 7.6 we point out that monadic second-order logic with prefix-closed sets is undecidable, which follows from [48]. This fact indicates that if we hope to show the decidability of structural subtyping of recursive types, it is essential to maintain the incomparability of types of different shape. 7.1 and Succ1 (Y1 , Y2 ) as ∀x.Y2 (x) ⇐⇒ ∃y.y = succ1 (x). The statement follows by induction on the structure of formulas. ∗ Let R ⊆ (2{0,1} )k × ({0, 1}∗ )n be relation of arity k + n. ∗ ∗ Define R∗ ⊆ (2{0,1} )k × (2{0,1} )n by Structural Subtyping Recursive Types R∗ (Y1 , . . . , Yk , X1 , . . . , Xn ) ≡ In this section we define the problem of structural subtyping of recursive types. We then give an embedding of MSOL of the infinite binary tree into the first-order theory of structural subtyping of infinite terms over the signature Σ = {a, b, g} with the partial order ≤. We define MSOL over infinite binary tree [6, Page 317] as the structure MSOL(2) = h{0, 1}∗ , succ0 , succ1 i. The domain of the structure is the set {0, 1}∗ of all finite strings over the alphabet {0, 1}. We denote first-order variables by lowercase letters such as x, y, z. First-order variables range over finite words w ∈ {0, 1}∗ . We denote second-order variables by uppercase letters such as X, Y, Z. Second-order variables range over finite and infinite subsets S ⊆ {0, 1}∗ . The only relational symbol is equality, with the standard interpretation. There are two function symbols, denoting the appending of the symbol 0 and the appending of the symbol 1 to a word: succ0 w = w · 0 succ1 w = ∃x1 , . . . , xn . X1 = {x1 } ∧ · · · ∧ Xn = {xn } ∧ R(Y1 , . . . , Yk , x1 , . . . , xn ) Lemma 94 (MSOL(1) expresses MSOL(2) ) If R is definable in MSOL(2) , then R∗ is definable in MSOL(1) . Proof Sketch. Property of being an empty set is definable in MSOL(1) by the formula φ0 (Y1 ) ≡ ∀Y2 .Y1 j Y2 The relation ⊂ of being a proper subset is definable in MSOL(1) by formula φ1 (Y1 , Y2 ) ≡ Y1 j Y2 ∧ Y1 6= Y2 and the relation ⊂1 of having one element more is definable by formula w·1 φ2 (Y1 , Y2 ) ≡ Y1 ⊂ Y2 ∧ ¬∃Z. Y1 ⊂ Z ∧ Z ⊂ Y2 For the purpose of embedding into the first-order theory of structural subtyping, we consider a structure MSOL(1) = h{0, 1}∗ , j, Succ0 , Succ1 i equivalent to MSOL(2) . We use the language of MSOL without first-order variables to make statements within MSOL(1) . j is a binary relation on sets denoting the subset relation: The property of being a singleton set can then be expressed by formula φ3 (Y1 ) ≡ ∃Y0 . φ0 (Y0 ) ∧ Y0 ⊂1 Y1 We define the relation on singletons corresponding to succ0 by Y1 j Y2 ⇐⇒ ∀x. x ∈ Y1 ⇒ x ∈ Y2 φ4 (Y1 , Y2 ) ≡ φ3 (Y1 ) ∧ φ3 (Y2 ) ∧ Succ0 (Y1 , Y2 ) Succ0 and Succ1 are binary relations on sets, Succ0 , Succ1 ⊆ ∗ ∗ 2{0,1} × 2{0,1} , defined as follows: Succ0 (Y1 , Y2 ) ⇐⇒ Y2 = {w · 0 | w ∈ Y1 } Succ1 (Y1 , Y2 ) ⇐⇒ Y2 = {w · 1 | w ∈ Y1 } Similarly, the relation corresponding to succ1 is defined by φ5 (Y1 , Y2 ) ≡ φ3 (Y1 ) ∧ φ3 (Y2 ) ∧ Succ1 (Y1 , Y2 ) If R is expressible by some formula ψ in MSOL(2) , then R is expressible by a formula in prenex normal form, so suppose ψ is of form Q1 V1 . . . Qn Vn .ψ0 The structure MSOL(1) is similar to one in [18]; the difference is that relations Succ0 and Succ1 are true even for non-singleton sets. Lemmas 93 and 94 show the expected equivalence of MSOL(2) and MSOL(1) . where ψ0 is quantifier free. We construct a formula ψ ′ expressing R∗ in MSOL(1) . We obtain the matrix ψ0′ of ψ ′ by translating ψ0 as follows. If x is a first-order variable in ψ0 , we represent it with a second-order variable X denoting a singleton set. We replace membership relation Y (x) with subset relation X ⊂ Y . We replace succ0 with φ4 and succ1 with φ5 . We construct ψ ′ by adding quantifiers to ψ0′ as follows. Second-order quantifiers remain the same. First-order quantifiers are relativized to range over singleton sets: ∀x.ψi becomes ∀X.φ3 (X) ⇒ ψi′ and ∃x.ψi becomes ∃X. φ3 (X) ∧ ψi′ (X). Lemma 93 (MSOL(2) expresses MSOL(1) ) Every relation on sets definable in MSOL(1) is definable in MSOL(2) . Proof. We express relations ⊆, Succ0 , Succ1 as formulas in MSOL(2) , as follows. We express Y1 ⊆ Y2 as ∀x. Y1 (x) ⇒ Y2 (x), Succ0 (Y1 , Y2 ) as We can view MSOL(1) as a first-order structure with the ∗ domain 2{0,1} . We show how to embed MSOL(1) into the first-order theory of structural subtyping. ∀x.Y2 (x) ⇐⇒ ∃y.y = succ0 (x), 45 If w = a1 . . . an is a word, then wR denotes the reverse of the word, wR = an . . . a1 . We define the embedding ι to map a set Y ⊆ {0, 1}∗ into the unique tree t such that t ∼ te and for every w ∈ {0, 1}∗ , We define the first-order structure of structural subtyping of recursive types similarly to the corresponding structure for non-recursive types in Section 4; the only difference is that the domain contains both finite and infinite terms. Infinite terms correspond to infinite trees [12, 30]. We define infinite trees as follows. We use alphabet {l, r} to denote paths in the tree. A tree domain D is a finite or infinite subset of the set {l, r}∗ such that: w ∈ Y ⇐⇒ T (h(wR ) · r) = b Observe that ι(∅) = te . Define formulas TSucc0 (t1 , t2 ) and TSucc1 (t1 , t2 ) as follows: ∗ 1. D is prefix-closed: if w ∈ {l, r} , x ∈ {l, r} then w · x ∈ D implies w ∈ D; 2. if w ∈ D then exactly one of the following two properties hold: (a) w is an interior node: {w · l, w · r} ⊆ D (b) w is a leaf: {w · l, w · r} ∩ D = ∅. 7.2 T (r · w) = T2 (w), if w ∈ T2 the term g(x1 , x2 ) denotes a term of the shape g s (se , se ). First we show that all atomic formulas are of one of the following forms: 1. x0 = g(g(x1, x2 ), a); 2. x0 = g(g(x1, x2 ), b); 3. x1 = x2 ; 4. x1 ≤ x2 . Consider an atomic formula t1 = t2 . The key idea is that if sh(t1 ) 6= sh(t2 ) then the formula t1 = t2 is false. If none of the term t1 and t2 is a variable then one of them is a constant or a constructor application. If t1 ≡ g(t11 , t12 ) then either t1 = t2 is false or t2 ≡ g(t21 , t22 ) for some t21 , t22 . We may therefore decompose t1 = t2 into t11 = t21 and t12 = t22 . By repeating this decomposition we arrive at terms of form t1 = t2 where both t1 and t2 are constants or at the equality of form x0 = t(x1 , . . . , xn ). The equalities between the constants can be trivially evaluated. This leaves only terms of form x0 = t(x1 , . . . , xn ). Let ts (xs1 , . . . , xsn ) be a shape term that results from replacing a and b with cs and replacing g with g s in t. Because all variables range over Te , we conclude that x0 = t(x1 , . . . , xn ) can be true only if te = g(g(te, te ), a) Trees in the ∼-equivalence class of te have the tree domain D = TDom(te ) given by the regular context-free grammar D → ǫ | r | l | lrD | llD whereas the leaves L of D are given by the context-free grammar L → ǫ | r | lrL | llL or the regular expression (lr|ll)∗ r. Let h be the homomorphism of words from {0, 1}∗ to {l, r}∗ such that h(1) = lr A Decidable Substructure g(x1 , x2 ) ≤ g(x3 , g(g(x4 , x5 ), b)) Observe that t1 ∼ t2 iff TDom(t1 ) = TDom(t2 ). We next present an embedding ι of MSOL(1) into SIT. The image of the embedding ι are the infinite trees that are in the same ∼-equivalence-class with the tree te . We define te as the unique solution of the equation: ll ≡ t2 = g(g(te, t1 ), te ) We show that the set of closed formulas from BF that are true in SIT is decidable. Although the quantifiers are bounded, terms in this logic can still denote elements of shape other than se . For example, the in the atomic formula t1 ∼ t2 ≡ ∃t0 . t0 ≤ t1 ∧ t0 ≤ t2 = TSucc1 (t1 , t2 ) Te = {t | sh(t) = se } Let IT denote the set of all infinite trees. The structural subtyping structure is the structure SIT = hIT, g, a, b, ≤i. SIT is an infinite-term counterpart to the structure BS from Section 4. Similarly to the case of finite terms, define the relation ∼ of “being of the same shape” in SIT by h(0) ≡ t2 = g(g(t1, te ), te ) Section 7.1 shows that terms of shape se form a substructure within SIT that is isomorphic to MSOL(1) . In this section we consider the following converse problem. Consider the formulas BF that, instead of quantifiers ∃, ∀, contain bounded quantifiers ∃e , ∀e that range over the elements of the set TDom(T ) = {l · w | w ∈ T1 } ∪ {r · w | w ∈ T2 } if w ∈ T1 TSucc0 (t1 , t2 ) It is straightforward to show that ι is an injection and that ι maps relation ⊂ into ≤, relation Succ0 into TSucc0 , and relation Succ1 into TSucc1 . Moreover, the range of ι is the set of all terms t such that sh(t) = se where se = sh(te ). A tree with a tree domain D is a total function T from the set of leaves of D to the set {a, b}. Note that the tree domain D of a tree T can be reconstructed from T as the prefix closure of the domain of the graph of function T ; we write TDom(T ) for the tree domain of tree T . Two trees are equal if they are equal as functions. Hence, equal trees have equal function domains and equal tree domains. We say that T1 ≤ T2 iff TDom(T1 ) = TDom(T2 ) and T1 (w) ≤0 T2 (w) for every word w ∈ TDom(T1 ). Here ≤0 is the relation {ha, ai, ha, bi, hb, bi}. If T1 and T2 are trees, then g(T1 , T2 ) denotes the tree T such that T (l · w) = T1 (w), (103) se = ts (se , . . . , se ) 46 then we could express all statements of SIT within the BF subtheory, and therefore SIT would be decidable. The fundamental problem with specifying P (x) is not the use of two bits to encode the three possible elements {a, b, g}, but the constraint that if a term contains a subterm of the form t4 (t1 , t2 , a, a) or t4 (t1 , t2 , a, b) at some even depth, then t1 ≡ t2 ≡ te . Compared to the relationships given by constructor g, this constraint requires taking about successor relation at the opposite side of the paths within a tree, see Section 7.6. If t(x1 , . . . , xn ) ∈ {a, b} is then (7.2) is false. If t(x1 , . . . , xn ) ≡ x1 , we obtain formula of the desired form. So assume t(x1 , . . . , xn ) ≡ g(t21 , t22 ). Then sh(t21 ) = g s (se , se ) and sh(t22 ) = cs . Therefore, t21 ≡ g(t211 , t212 ) where either sh(t211 ) = sh(t212 ) = se or t1 = t2 is false. Similarly, either t22 ∈ {a, b} or t1 = t2 is false. Therefore, t(x1 , . . . , xn ) ≡ g(g(t211 , t212 ), a), t(x1 , . . . , xn ) ≡ g(g(t211 , t212 ), b), or t1 = t2 is false. If t(x1 , . . . , xn ) ≡ g(g(t211 , t212 ), a) then we may replace the t1 = t2 with the formula ∃e y1 , y2 . x0 = g(g(y1 , y2 ), a) ∧ y1 = t211 ∧ y2 = t212 7.4 and similarly in the other case. By continuing this process by the induction on the structure of the term t(x1 , . . . , xn ) we either conclude that t1 = t2 is false, or we conclude that t1 = t2 is equivalent to a conjunction of formulas of the desired form. Conversion of atomic formula of form t1 ≤ t2 is analogous to the conversion of formulas t1 = t2 . To see the decidability it now suffices to convert the formulas of the form x0 = g(g(x1 , x2 ), a) and x0 = g(g(x1, x2 ), b) into formulas TSucc0 (t1 , t2 ) and TSucc1 (t1 , t2 ). Expressibility of x0 = g(g(x1, x2 ), a) follows from the fact that the following relationship between X0 , X1 , X2 is expressible in MSOL: We next argue that if we allow the logic to have a copy of bounded quantifiers ∃s , ∀s for every constant shape s, we obtain a decidable theory. To denote constant shapes in a finite number of symbols we consider in addition to term algebra symbols g s , cs the expressions that yield solutions of mutually recursive equations on shapes; the details of the representation of types are not crucial for our argument, see e.g. [12] Consider a closed formula in such language. Because every variable has an associated constant shape, we can compute the set of all shapes occurring in the formula. This means that all variables of the formula range over a finite known set of shapes. This allows us to define the predicate P given by (104) as a disjunction of cases, one case for every shape. Define hmin , hmax functions that take a shape and produce a lower and upper bound for terms of that shape: X0 = {w · 0 | w ∈ X1 } ∪ {w · 1 | w ∈ X2 } Similarly, the expressibility of x0 = g(g(x1, x2 ), b) follows from the fact that hmin (g X0 = {w · 0 | w ∈ X1 } ∪ {w · 1 | w ∈ X2 } ∪ {ǫ} is expressible in MSOL. We conclude that the set of closed BF formulas that are true in SIT is decidable. 7.3 Subtyping Trees of Known Shape hmax (g s hmin (cs ) = ta (ts1 , ts2 )) tg (hmin (ts1 ), hmin (ts2 )) = hmax (cs ) = tb s tg (hmax (ts1 ), hmax (ts2 )) (ts1 , ts2 )) = If s1 , . . . , sn is the list of shapes occurring in a formula, we then define a predicate P specific to that formula by Embedding Terms into Terms We next give an embedding of the set of all terms into Te . As in Section 7.1 te be the unique solution of the equation te = g(g(te, te ), a) and let P (t) = n _ (hmin (si ) ≤ t ∧ t ≤ hmax (t)) i=1 t4 (x1 , x2 , x3 , x4 ) ≡ g(g(g(g(x1, x2 ), x3 ), te ), x4 ) Define We can therefore define P (t) and use it to translate the formula into a BF formula of the same truth value. Therefore, structural subtyping with quantification bounded to constant shapes is decidable. For decidability of the structural subtyping recursive types it would be interesting to examine the decision procedure for MSOL and determine whether there is some uniformity in it that would allow us to handle even quantification over shapes that are determined by variables. ta ≡ t4 (te , te , a, a) tb ≡ t4 (te , te , a, b) tg (x1 , x2 ) ≡ t4 (x1 , x2 , b, b) Then define the homomorphism hT from the set of all terms to the set Te by hT (a) = ta 7.5 hT (b) = tb We next remark that certain notion of subtyping of recursive feature trees is decidable. By a feature tree we mean an infinite tree built using a constructor which takes other feature trees and an optional node label as an argument. In this section we consider the simple case of one binary constructor f and assume only one label denoted by 1. Hence, an empty feature tree is a feature tree, and if t1 and t2 are feature trees then so are f ǫ (t1 , t2 ) and f 1 (t1 , t2 ). We represent an empty feature tree e by an infinite tree that has all features ǫ. We compare feature trees as follows. Let ≤ be defined on the hT (g(t1 , t2 )) = tg (hT (t1 ), hT (t2 )) Then hT is embedding of the set of all terms into the subset subset Te of all terms. The term algebra operations a, b, g map to ta , tb , tg and ≤ maps to ≤. Note that, if it were possible to define a predicate P (t) such that P (x) ⇐⇒ ∃y.hT (y) = x (104) 47 Recursive Feature Trees features {ǫ, 1} as the relation {hǫ, ǫi, hǫ, 1i, h1, 1i}. Define ≤ on trees as the least relation such that: Every subset of C(w) in PCl is a of the form C(w1 ) for some word w1 . Define PSucc0 and PSucc1 on PCl by: 1. e ≤ t for all terms t; 2. t1 ≤ t′1 and t2 ≤ t′2 implies r1 f (t1 , t2 ) ≤ f r2 (t′1 , t′2 ) The decidability of feature trees follows from Section 7.1 because of the isomorphism hF between the set of terms Te and the set of feature trees. Here hF is defined by: hF (e) = te g(hF (t1 ), hF (t2 ), a) 1 g(hF (t1 ), hF (t2 ), b) hF (f (t1 , t2 )) = hF (f (t1 , t2 )) = The feature trees as we defined them have a limited feature and node label alphabet. This is not a fundamental problem. Muchnik’s theorem [56] gives the decidability of MSOL of trees over arbitrary decidable structures. It is reasonable to expect that the decidability of MSOL over decidable structures yields a generalization of the result of Section 7.1 and therefore the decidability of feature trees with a richer vocabulary of features. The crucial property of our definition of feature trees is that features can appear in any node of the tree. Hence, there are no prefix closure requirements on trees as in Section 7.3, which is responsible for relatively simple reduction to MSOL. 7.6 ∃w. X1 = C(w) ∧ X2 = C(0 · w) PSucc1 (X1 , X2 ) = ∃w. X1 = C(w) ∧ X2 = C(1 · w) Consider a monadic theory PrefT with relations PSucc0 and PSucc1 where second-order variables range over the subsets of PCl. It is easy to see that PrefT corresponds to the firstorder theory of non-structural subtyping of recursive types, with subset relation ⊆ corresponding to subtype relation ≤, empty set corresponding to the least type ⊥, PSucc0 (X1 , X2 ) corresponding to X2 = f (X1 , ⊥), and PSucc1 (X1 , X2 ) corresponding to X2 = f (⊥, X2 ). The first-order theory of non-structural subtyping was shown undecidable in [48], so PrefT is undecidable. An interesting open problem is the decidability of fragments of the first-order theory of structural subtyping. This problem translates directly to the decidability of the fragments of PrefT, a monadic theory with prefix-closed sets, or, under the word-reversal isomorphism, the decidability of fragments of the monadic theory of two successor symbols with suffix-closed sets. for all r1 , r2 ∈ {ǫ, 1} such that r1 ≤ r2 . ǫ PSucc0 (X1 , X2 ) = 8 Conclusion In this paper we presented a quantifier elimination procedure for the first-order theory of structural subtyping of non-recursive types. Our proof uses quantifier elimination. Our decidability proof for the first-order theory of structural subtyping clarifies the structure of the theory of structural subtyping by introducing explicitly the notion of shape of a term. We presented the proof in several stages with the hope of making the paper more accessible and self-contained. Our result on the decidability of Σ-term-power is more general than the decidability of structural subtyping non-recursive types, because we allow even infinite decidable base structures for primitive types. We view this decidability result as an interesting generalization of the decidability for term algebras and decidability of products of decidable theories. This generalization is potentially useful in theorem proving and program verification. Of potential interest might be the study of axiomatizability properties; the quantifier elimination approach is appropriate for this purpose [31, 30], we did not pay much attention to this because we view the language and the mechanism for specifying the axioms of secondary importance. Our goal in describing quantifier elimination procedure was to argue the decidability of the theory of structural subtyping. While it should be relatively easy to extract an algorithm from our proofs, we did not give a formal description of the decision procedure. One possible formulation of the decision procedure would be a term-rewriting system such as [11]; this formulation is also appropriate for implementation within a theorem prover. Our approach eliminates quantifiers as opposed to quantifier alternations. For that purpose we extended the language with partial functions. The use of Kleene logic for partial functions seems to preserve most of the properties of two valued logic and appears to agree with the way partial functions are used in informal mathematical practice. An alternative direction for proving decidability of structural subtyping would be to use Ehrenfeucht-Fraisse games [53, Page 405]; [15] uses techniques based on games to study both the decidability and the computational complexity of theories. Reversed Binary Tree with Prefix-Closed Sets It is instructive to compare the difficulties our approach faces in showing the decidability of structural subtyping of recursive types with the difficulties reported in [48]. In [48, Section 5.3] the authors remark that the difficulty with applying tree automata is that the set x = f (y, z) is not regular. By reversing the set of paths in a tree representing a term we have shown in Section 7.1 that the relationship x = f (y, z) becomes expressible. However, the difficulty now becomes specifying a set of words that represents a valid term, because there is no immediate way of stating that a set of words is prefix-closed. If we add an operation that allows expressing relationship at both “ends” of the words, we obtain a structure whose MSOL is undecidable due to the following result [52, Page 183]. Theorem 95 MSOL theory of the structure with two successor operations w · 0 and w · 1 and one inverse successor operation 0 · w is undecidable. The case that is of interest of us is the dual to Theorem 95 under the word-reversing isomorphism: a structure with operations 0 · w, 1 · w, w · 0 has undecidable MSOL closed formulas. Instead of expressing prefix-closure using operations w·0, w · 1, let us consider MSOL over the structure that contains only operations 0 · w and 1 · w, but where all second-order variables range over prefix-closed sets. This logic also turns out to be undecidable. Let PCl be the set of prefix-closed sets. For each word w, there exists the smallest PCl set containing w, namely the set C(w) given by: C(w) = {w′ | w′ ≺ w} 48 The complexity of our the decidability for structural subtyping non-recursive types is non-elementary and is a consequence of the non-elementary complexity of the term algebra, whose elements and operations are present in the theory of structural subtyping. Tools like MONA [25] show that non-elementary complexity does not necessarily make the implementation of a decision procedure uninteresting. An interesting property of quantifier elimination is that it can be applied partially to elimination an innermost quantifier from some formula. This property makes our decision procedure applicable as part of an interactive theorem prover or a subroutine of a more general decision procedure. In this paper we have left open the decidability of structural subtyping of recursive types, giving only a few remarks in Section 7. In particular we have observed in Section 7.1 that every formula in the monadic second-order theory of the infinite binary tree [6, Page 317] has a corresponding formula in the first-order theory of structural subtyping of recursive types. In that sense, the decision problem for structural subtyping recursive types is at least as hard as the decision problem for the monadic second-order logic interpreted over the infinite binary tree. This observation is relevant for two reasons. First, it is unlikely that a minor modification of the quantifier elimination technique we used to show the decidability of structural subtyping non-recursive types can be used to show the decidability of recursive types. Because of the embedding in Section 7.1 such a quantifier-elimination proof would have to subsume the determinization of tree automata over infinite trees. Second, the embedding suggests even greater difficulties in implementing a decision procedure for the first-order theory of structural subtyping (provided that it exists). While we know at least one interesting example of weak monadic second-order logic decision procedure, namely [25] we are not aware of any implementation of the full monadic secondorder logic decision procedure for the infinite tree. The relationship between the non-structural as well as structural subtyping and monadic second-order logic of the infinite binary tree and tree like structures [57] requires further study. In that respect the work on feature trees [36, 37] appears particularly relevant. [5] L. O. Andersen. Program Analysis and Specialization of the C Programming Language. PhD thesis, DIKU, University of Copenhagen, 1994. [6] Egon Börger, Erich Gräedel, and Yuri Gurevich. The Classical Decision Problem. Springer-Verlag, 1997. [7] Witold Charatonik and Leszek Pacholski. Set constraints with projections are in NEXPTIME. In Proc. 35th Annual Symposium on Foundations of Computer Science (FOCS), pages 642–653, 1994. [8] Witold Charatonik and Andreas Podelski. Set constraints with intersection. In Proc. 12th IEEE LICS, pages 362–372, 1997. [9] Hubert Comon. Disunification: A survey. In JeanLouis Lassez and Gordon Plotnik, editors, Computational Logic: Essays in Honor of Alan Robinson. The MIT Press, Cambridge, Mass., 1991. [10] Hubert Comon and Catherine Delor. Equational formulae with membership constraints. Information and Computation, 112(2):167–216, 1994. [11] Hubert Comon and Pierre Lescanne. Equational problems and disunification. Journal of Symbolic Computation, 7(3):371, 1989. [12] Bruno Courcelle. Fundamental properties of infinite trees. Theoretical Computer Science, 25(2):95–169, March 1983. [13] Rowan Davies and Frank Pfenning. Intersection types and computational effects. In Proc. ICFP, pages 198– 208, 2000. [14] S. Feferman and R. L. Vaught. The first order properties of products of algebraic systems. Fundamenta Mathematicae, 47:57–103, 1959. [15] Jeanne Ferrante and Charles W. Rackoff. The Computational Complexity of Logical Theories, volume 718 of Lecture Notes in Mathematics. Springer-Verlag, 1979. [16] Tim Freeman and Frank Pfenning. Refinement types for ML. In Proc. ACM PLDI, 1991. Acknowledgements The first author would like to thank Albert Meyer for pointing out to the work [15], and Jens Palsberg and Jakob Rehof for useful discussions about the subject of this paper. [17] Alexandre Frey. Satisfying subtype inequalities in polynomial space. Theoretical Computer Science, 277:105– 117, 2002. References [18] Yuri Gurevich and Leo Harrington. Trees, automata, and games. In Proceedings of the fourteenth annual ACM symposium on Theory of computing, pages 60– 65, 1982. [1] Alexander Aiken. Introduction to set constraint-based program analysis. Science of Computer Programming, 35:79–111, 1999. [19] Nevin Heintze and Olivier Tardieu. Ultra-fast aliasing analysis using CLA: A million lines of C code in a second. In Proc. ACM PLDI, 2001. [2] Alexander Aiken, Dexter Kozen, and Ed Wimmers. Decidability of systems of set constraints with negative constraints. Information and Computation, 122, 1995. [20] Fritz Henglein and Jakob Rehof. The complexity of subtype entailment for simple types. In Proc. 12th IEEE LICS, pages 352–361, 1997. [3] Alexander Aiken, Edward L. Wimmers, and T. K. Lakshman. Soft typing with conditional types. In Proc. 21st ACM POPL, pages 163–173, New York, NY, 1994. [21] L. Henkin, J. D. Monk, and A. Tarski. Cylindric Algebras, Part I. North Holland, 1971. [4] Roberto M. Amadio and Luca Cardelli. Subtyping recursive types. Transactions on Programming Languages and Systems, 15(4):575–631, 1993. [22] Wilfrid Hodges. Model Theory, volume 42 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, 1993. 49 [39] Derek C. Oppen. Reasoning about recursively defined data structures. Journal of the ACM, 27(3), 1980. [23] Trevor Jim and Jens Palsberg. Type inference in systems of recursive types with subtyping. http://www. cs.purdue.edu/homes/palsberg/, 1999. [40] Frank Pfenning. Unification and anti-unification in the calculus of constructions. In Proc. 6th IEEE LICS, pages 74–85, 1991. [24] Manfred Kerber and Michael Kohlhase. A mechanization of strong Kleene logic for partial functions. In Alan Bundy, editor, Proc. 12th CADE, pages 371–385, Nancy, France, 1994. Springer Verlag, Berlin, Germany. LNAI 814. [41] Francois Pottier. Simplifying subtyping constraints: A theory. Information and Computation, 170(2):153–183, November 2001. [25] Nils Klarlund, Anders Møller, and Michael I. Schwartzbach. MONA implementation secrets. In Proc. 5th International Conference on Implementation and Application of Automata. Lecture Notes in Computer Science, 2000. [42] Jakob Rehof. The Complexity of Simple Subtyping Systems. PhD thesis, Computer Science Department, Univ. of Copenhagen (DIKU), April 1998. [43] Tatiana Rybina and Andrei Voronkov. A decision procedure for term algebras with queues. ACM Transactions on Computational Logic (TOCL), 2(2):155–181, 2001. [26] Stephen Cole Kleene. Introduction to Metamathematics. D. Van Nostrand Company, Inc., Princeton, New Jersey, 1952. fifth reprint, 1967. [27] Dexter Kozen. Complexity of boolean algebras. Theoretical Computer Science, 10:221–247, 1980. [44] Mooly Sagiv, Thomas Reps, and Reinhard Wilhelm. Parametric shape analysis via 3-valued logic. ACM TOPLAS, 24(3):217–298, 2002. [28] Dexter Kozen, Jens Palsberg, and Michael I. Schwartzbach. Efficient recursive subtyping. Mathematical Structures in Computer Science, 5(1):113–125, 1995. [45] Jörg H. Siekmann. Unification theory. Journal of Symbolic Computation, 7(3):207–274, 1989. [46] Thoralf Skolem. Untersuchungen über die Axiome des Klassenkalküls and über “Produktations- und Summationsprobleme”, welche gewisse Klassen von Aussagen betreffen. Skrifter utgit av Vidnskapsselskapet i Kristiania, I. klasse, no. 3, Oslo, 1919. [29] John W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 2nd edition, 1987. [30] Michael J. Maher. Complete axiomatizations of the algebras of the finite, rational, and infinite trees. Proc. 3rd IEEE LICS, 1988. [47] Bjarne Steensgaard. Points-to analysis in almost linear time. In Proc. 23rd ACM POPL, St. Petersburg Beach, FL, January 1996. [31] Anatolii Ivanovic Mal’cev. The Metamathematics of Algebraic Systems, volume 66 of Studies in Logic and The Foundations of Mathematics. North Holland, 1971. [48] Zhendong Su, Alexander Aiken, Joachim Niehren, Tim Priesnitz, and Ralf Treinen. First-order theory of subtyping constraints. In Proc. 29th ACM POPL, 2002. [32] Ursula Martin and Tobias Nipkow. Boolean unification: The story so far. Journal of Symbolic Computation, 7(3):275–293, 1989. [49] Madhu Sudan. Quantifier elimination for boolean algebras is trivial. Personal Communication, MIT LCS Elevator, 9 October 2002. [33] John McCarthy and James Painter. Correctness of a compiler for arithmetic expressions. In Proceedings of Symposia in Applied Mathematics. American Mathematical Society, 1967. [50] Alfred Tarski. Arithmetical classes and types of algebraically closed and real-closed fields. Bull. Amer. Math. Soc., 55, 64, 1192, 1949. [34] John C. Mitchell. Type inference with simple types. Journal of Functional Programming, 1(3):245–285, 1991. [51] Alfred Tarski. Arithmetical classes and types of boolean algebras. Bull. Amer. Math. Soc., 55, 64, 1192, 1949. [35] Andrzej Mostowski. On direct products of theories. Journal of Symbolic Logic, 17(1):1–31, March 1952. [52] Wolfgang Thomas. Automata on infinite objects. In Handbook of Theoretical Computer Science, Volume B, pages 133–191. Elsevier and The MIT Press, 1990. [36] Martin Mueller and Joachim Niehren. Ordering constraints over feature trees expressed in secondorder monadic logic. Information and Computation, 159(1/2):22–58, 2000. [53] Wolfgang Thomas. Languages, automata, and logic. In Handbook of Formal Languages Vol.3: Beyond Words. Springer-Verlag, 1997. [54] Jerzy Tiuryn. Subtype inequalities. In Proc. 7th IEEE LICS, 1992. [37] Martin Mueller, Joachim Niehren, and Ralf Treinen. The first-order theory of ordering constraints over feature trees. Discrete Mathematics and Theoretical Computer Science, 4(2):193–234, September 2001. [55] K. N. Venkataraman. Decidability of the purely existential fragment of the theory of term algebras. Journal of the ACM (JACM), 34(2):492–510, 1987. [38] Greg Nelson and Derek C. Oppen. Fast decision procedures based on congruence closure. Journal of the ACM (JACM), 27(2):356–364, 1980. [56] Igor Walukiewicz. Monadic second-order logic on treelike structures. In STACS’96, volume 1046 of Lecture Notes in Computer Science, 1996. 50 [57] Igor Walukiewicz. Monadic second-order logic on treelike structures. Theoretical Computer Science, 275(1– 2):311–346, March 2002. 51
6cs.PL
Effectively integrating information content and structural relationship to improve the GO-based similarity measure between proteins Bo Li1, James Z. Wang1, F. Alex Feltus2, Jizhong Zhou3, Feng Luo1, 2, § 1 School of Computing, Clemson University, Clemson, SC, USA 2 Department of Biochemistry and Genetics, Clemson University, Clemson, SC, USA 3 Institute for Environmental Genomics, and Department of Botany and Microbiology, University of Oklahoma, Norman, OK, 73019, USA, § Corresponding author Email addresses: BL: [email protected] JZW: [email protected] FAF: [email protected] JZZ: [email protected] FL: [email protected] -1- Abstract Background The Gene Ontology (GO) provides a knowledge base to effectively describe proteins. However, measuring similarity between proteins based on GO remains a challenge. Results In this paper, we propose a new similarity measure, information coefficient similarity measure (SimIC), to effectively integrate both the information content (IC) of GO terms and the structural information of GO hierarchy to determine the similarity between proteins. Testing on yeast proteins, our results show that SimIC efficiently addresses the shallow annotation issue in GO, thus improves the correlations between GO similarities of yeast proteins and their expression similarities as well as between GO similarities of yeast proteins and their sequence similarities. Furthermore, we demonstrate that the proposed SimIC is superior in predicting yeast protein interactions. We predict 20484 yeast protein-protein interactions (PPIs) between 2462 proteins based on the high SimIC values of biological process (BP) and cellular component (CC). Examining the 214 MIPS complexes in our predicted PPIs shows that all members of 159 MIPS complexes can be found in our PPI predictions, which is more than those (120/214) found in PPIs predicted by relative specificity similarity (RSS). Conclusions Integrating IC and structural information of GO hierarchy can improve the effectiveness of the semantic similarity measure of GO terms. The new SimIC can effectively correct the effect of shallow annotation, and then provide an effective way to measure similarity between proteins based on Gene Ontology. -2- Background The Gene Ontology [1] is a database of controlled vocabulary terms that represent human knowledge about genes and gene products. GO terms are divided into three categories: biological process (BP), molecular function (MF), and cellular component (CC). The cellular component terms characterize the location of gene products in the cell. The molecular function terms represent the molecular level activities of proteins. The biological process terms describe a series of events accomplished by one or more proteins [1]. The internal GO structure is a hierarchical directed acyclic graph (DAG), in which the terms are vertices and the relationships form the edges. Each of the three categories forms a branch of the GO DAG. There are five types of semantic relationships in the GO DAG, and we focus our study on two of them, namely, “is_a”, which implies that the child is a subclass of the parent, and “part_of”, which indicates that the child is component of the parent. Currently, the GO is an important knowledge resource to describe proteins, and using GO annotations to measure the similarity between gene products has attracted a lot of attention [2-20]. In these approaches, the similarity between two proteins is calculated based on semantic similarities among GO terms that annotate them. There are many challenges in the measurement of semantic similarity between GO terms and between proteins. First, the depth of empirical knowledge and the number of associated GO terms on each aspect of a given protein is highly variable. This creates a situation where the hierarchical level of annotated GO terms may not accurately represent the underlying concept that GO terms at the same level in the DAG may not be equally specific [7]. Second, annotating gene products with GO terms is often limited by bioinformatics techniques including the sensitivity of homology detection for functional mapping. Proteins may be annotated with high level GO terms which are close to the root of the DAG or with GO terms associated with a large number of genes thereby diluting functional information. -3- This phenomenon is called “shallow” annotation. Furthermore, a large number of gene products have not been annotated with any GO terms at all. The similarity measures between GO terms are usually based on two factors: information content (IC) of GO terms and structural information of GO hierarchy. A GO term is less informative if it is frequently used to annotate gene products. That is, the information content of a GO term is inversely proportional to its probability of correctly annotating genes. The probability of a given term t is defined as the number of genes annotated with it over the total number of annotations. As a gene that is annotated with a GO term is also annotated with all ancestors of the GO term in the DAG, the total number of annotations is the frequency of DAG root. Namely, the probability of a term t is given by: anno(t )  prob(t )   anno( d ) d descendant ( t ) (1)  anno( c ) cdescendant ( root ) where anno(t) is the number of genes directly annotated with the term t. The hierarchical structure of GO makes the probability of GO terms non-decreasing from the descendants to the root of the DAG and the probability of the DAG root has a value of 1. The IC of a GO term is usually defined as negative logarithm of the term's probability [21]: IC (t )   log( prob(t )) (2) Thus, the information content of DAG root is 0. As GO terms ascend the hierarchical tree of DAG from the leaf, the information content would not increase. Commonly used structural information of GO hierarchy is the topological path between two GO terms, which is the number of steps taken to reach one term from another in the DAG. However, the same distance between two GO terms close to the root of the DAG (more general) and between two GO terms close to the leaf of the DAG (more specific) doesn’t mean the similarities of these two GO term pairs are conceptually equal. Furthermore, the GO terms in the same level may also not be of the same specificity. Thus, -4- the path between two GO terms is not the ideal similarity measure. Meanwhile, two structural concepts that are frequently used in GO similarity measures are the “lowest common ancestor” (LCA) and “most informative common ancestor” (MICA) [21]. Given two GO terms, the MICA is their common ancestor with the lowest probability in the DAG structure. In the example of gene ontology DAG shown in Figure 1, there are three common ancestors of terms H and I which are A, D and F. As term A definitely has higher probability than terms D and F have. The MICA of term H and I is one of terms D and F, which should have lower probability than term A. The LCA of two GO terms is their common ancestor term at the lowest level in the DAG structure. As shown in Figure 1, the LCA of terms H and I is F. Notice that there also exists a case where two terms have multiple ancestor terms at the same level. In this case, the term with lowest probability will be the final LCA. The relationship between two terms and their LCA or MICA has been widely used to designing the similarity measures. Many similarity measures have been developed based on structural relationship, information content [13, 21-26] or both. Most of these similarity measures are originally designed for other ontologies, such as WordNet [27]. Unfortunately, most existing similarity measures have not effectively integrated information content and structural information. Some similarity measures (e.g. Resnik [21]) employ information content of the MICA of two concepts only. Although these measures can reflect the semantic difference among the DAG levels, they lose the structural information in the DAG and cannot reveal the semantic difference among term pairs with same MICA. As shown in Figure 1, both the (L, D) pair and the (L, K) pair have the same MICA, B. Thus, they will have the same Resnik similarity. However, conceptually, L should be more similar to D. Other similarity measures such as Lin [23], Jiang [22], and graph information content (GIC) of Pesquita et al [26] (see Methods section for detailed definitions) utilize both information content and structural information, -5- but they actually can not distinguish semantic differences among different levels in DAG when both proteins are annotated by the same GO term(s), in which case the GO term similarity is just the protein similarity. For the example of Figure 1, if two proteins are uniquely annotated by term J and two proteins are uniquely annotated by term C, these two pairs of proteins will have same similarities in Lin, Jiang and GIC, which is misleading. On the other hand, the Resnik measure can give different values in this case. Similarity measures, like the Relative Specificity Similarity (RSS) of Wu et al. [28] and Wang et al. [13], consider the structure information of the GO DAG only. These similarities are also affected by shallow annotations because no information content is incorporated into them. For the example of Figure 1, if a pair of proteins (pair A) are only annotated by GO term J and another pair of proteins (pair B) are only annotated by GO term C, the Wang measure will give both pairs of proteins with similarity 1.0 and can not distinguish the semantic difference of GO terms J and C. The RSS is also insufficient to measure the similarity between protein pairs (see Results sections for more discussion). Recently, Schlicker et al. [24] proposed a relevance similarity that counts both the structural relationship between two terms and their MICA and the probability of the MICA. The relevance similarity is defined as the product of Lin similarity and a coefficient, (1prob(MICA)). The adjustment of Lin similarity with (1-prob(MICA)) can reduce the effect of shallow annotations. Unfortunately, the probability adjustment becomes insensitive when the probability value is close to 0 and 1. For example, if there are about ten thousand biological process annotations, a GO term with 100 annotations will have a probability of 0.01. On the other hand, a GO term with ten annotations will have a probability of 0.001. Then, the adjustments in relevance similarity are 0.99 and 0.999 respectively. These adjustments are not sensitive enough to detect their information difference, which is a ten-fold difference in annotation number. -6- In this paper, we propose a new method to calculate similarity between two GO terms by effectively integrating structural information with IC. Our SimIC similarity measure is sensitive to IC differences among GO terms and efficiently addresses the shallow annotation problem. We have compared SimIC with Lin, Jiang and GIC, relevance, Resnik and adjusted Resnik similarities by calculating the correlations between semantic similarities of yeast protein pairs and their gene expression similarities and correlations between semantics similarities of yeast protein pairs and sequence similarities. All comparisons are based on the GO annotations of yeast from Saccharomyces Genome Database (SGD) [29]. Furthermore, we show that SimIC is superior in predicting yeast protein interactions. Results New semantic similarity measures by integrating IC and structure of GO DAG The relevance similarity proposed by Shlicker et al. [24] is not sensitive to the semantic difference among GO term close to the leaf of DAG (probability close to 0) or close to the root of the DAG (probability close to 1). In order to enforce more impact of information content on the similarity detection, we introduce a new information coefficient similarity (SimIC) measure to effectively integrate information content and structural relationship. The SimIC is defined as: (3) 1 2  log( prob( MICA)) ) SimIC (t1, t2 )   (1  1  IC ( MICA) log( prob(t1 ))  log( prob(t2 )) Using the example in background section, the SimIC adjustments for the GO term with 100 annotations and the GO term with ten annotations are 0.67 and 0.75 respectively. The adjustment difference is 0.08, which is much higher than the difference between probability adjustments used in relevance similarity, which is 0.009. The first part of SimIC is simply the Lin similarity. We also can obtain other flavours of SimIC by replacing the first part of SimIC with Jiang or other measures. Unless explicitly stated otherwise, we will refer SimIC as the one defined in Equation 3. -7- Correction of the shallow annotation by SimIC Many proteins are annotated by general GO terms due to limited knowledge about these proteins and their homologs. As discussed in Background section, these shallow annotations will affect the accuracy of similarity measures, such as Lin, Jiang, Wang and GIC measures. Both the relevance similarity and SimIC are able to correct the shallow annotation effect using information content. Comparing to the relevance similarity, the SimIC will be provide more efficient correction for the GO terms close to root. Figure 2 shows the distributions of Lin, relevance and SimIC of all pairs of annotated yeast proteins. As each similarity measure has different range of values, we take a normalization step by evenly partitioning values of each similarity measure into 21 bins. As shown in Figure 2, the SimIC has a higher percentage of low values relative to Lin and relevance similarities. Moreover, the relevance similarity has similar percentage of the highest values (last bin) as the Lin similarity has, which means that the probability adjustment is limited to the highest values. On the other hand, the SimIC is able to efficiently adjust the high similarity values caused by shallow annotations and lead to a very low percentage of highest values as shown in inner figure of Figure 2. For example, as shown in Figure 3A, yeast proteins YDR387C and YHL035C both annotated with term “transport”, a level-three GO term with 1726 annotations in total associated with the term and its descendants. And as shown in Figure 3B, yeast proteins YNL108C and YJR051W both are annotated with “metabolic process”, a level-two GO term with 7027 annotations in total associated with the term and its descendants. Both pairs of yeast proteins will have similarity values of 1.0 based on Lin, Wang, Jiang and GIC measures. The relevance similarity has a value of 0.8512 for YDR387C and YHL035C and a value of 0.5689 for YNL108C and YJR051W. Meanwhile, the SimIC has a value of 0.6829 for YDR387C and YHL035C and a value of 0.4284 for YNL108C and YJR051W. Figure 4 compares SimIC similarities with relevance similarities for yeast protein pairs with Lin’s similarities greater than 0.9. While -8- relevance similarities have values between 0.85 and 1.0, the corresponding SimIC similarities have values between 0.7 and 0.9. Figure 3 C shows an example of different levels of correction by relevance and SimIC similarities for the MICA with low probability. Yeast proteins YDR130C and YOR058C both annotated with term “microtubule binding”, a levelsix GO term with 10 annotations in total associated with it and its descendants. Yeast proteins YDR130C and YOR058C will have similarity values of 1.0 based on Lin, Wang, Jiang and GIC measures. The relevance similarity has a value of 0.9984 and the SimIC has a value of 0.8658. Improvement of correlation between GO semantic similarities and sequence similarities of yeast proteins by SimIC In order to further demonstrate the effectiveness of SimIC, we first calculated the correlation between GO semantic similarities and sequence similarities of yeast proteins. Ten GO semantic similarities including the Lin, Jiang, GIC, RSS, Resnik and adjusted Resnik as well as two types of relevance and SimIC similarities were used for analysis. Previous studies [4, 5, 24] already showed that high sequence similarities are correlated with high semantic similarities although high semantic similarities do not always mean high sequence similarities. Therefore, we only employed protein pairs with both significant sequence similarity and semantic similarity in the comparison (see detail in Methods). As a protein can be annotated by multiple GO terms, three approaches have been used to calculate protein semantic similarity from the GO term similarity. Namely, the semantic similarity between proteins is the maximum, average and best match average of GO term similarities. Table 1, 2 and 3 show the correlations between protein semantic similarities based on twelve GO similarity measures of three GO branches and the sequence similarity for each of these three protein similarity approaches. For the average protein similarity approach, the Resnik similarity gave higher average correlation than those original Lin, Jiang gave, which is -9- consistent with Lord’s results [5]. For all three approaches of protein semantic similarities, the SimIC always give higher average correlation values and lower standard deviation than those of Lin, Jiang and GIC similarities. However, the relevance similarities do not always lead to higher average correlations. This indicated that the SimIC will be more effective than relevance similarity. Furthermore, it is noted that the SimIC has dramatically improved the correlations between sequence similarities and GO semantic similarities of cellular components (CC), which usually have higher shallow annotation rate. Moreover, the adjusted Resnik similarity gave the highest average correlation for maximum protein similarity approach, and the SimIC resulted in the highest average correlations for both average and best match average protein similarity approaches. The best match average approach based on SimIC give overall best average correlation value of 0.893. The scatter plots of best match average based on BP (Figure 5), CC (Figure 6) and MF (Figure 7) SimIC measured against corresponding sequence similarities showed a strong linear relationship. These results suggest that SimIC measure can improve the average correlation between semantic similarity and sequence similarity of proteins. Evaluation of semantic similarities of protein pairs with various levels of sequences similarities To further test the correlation between semantic and sequences similarities, we calculated the distribution of the protein BP and MF semantic similarity of proteins pairs in different levels of evolutionary relationship [24]. The protein pairs between human and yeast were divided into four categories based on the level of sequence similarity: orthology (from Inparanoid database [30]), high similarity (HS), low similarity (LS) and no similarity (NS). The distributions of BP and MF SimIC similarities in four categories are shown in Figure 8 and 9. The NS dataset has the highest percentage of low semantic similarity and lowest percentage of high semantic similarity. On the other hand, the orthology data has the lowest percentage - 10 - of low semantic similarity and highest percentage of high semantic similarity. Figure 9 shows a histogram of the relationship between MF and BP semantic similarity for the protein pairs in the orthology dataset. Similar to results of Schlicker et al., the highest peak occurs at M0.9 and B0.9. Next, we compared our distributions with distributions of GIC (Figures 11, 12) and relevance similarity (Figures 13, 14) in four datasets. The distributions of BP GIC similarities show a higher percentage in the low semantic similarity range and a lower percentage in the high semantic similarity range A Chi-square test showed that the distributions of BP SimIC similarities are significant different from those of GIC similarity with p-value less than 10-7 in all four datasets. The distribution of MF SimIC similarities are also significant different from those of GIC similarity with p-value less than 0.025, except for NS dataset. Meanwhile, the distributions of BP and MF relevance similarities are very similar to those of SimIC similarities. A Chi-square test showed no significant differences with pvalues greater than 0.75. Improvement of correlation between GO semantic similarities and expression similarities of yeast proteins by SimIC Another widely used method to evaluate GO semantic similarities is to calculate the correlation between semantic similarities and gene expression similarities [7]. Here, we calculated the correlations between each of these three approaches of protein semantic similarities based on the Lin, Jiang, GIC, RSS, Resnik and adjusted Resnik as well as two types of relevance and SimIC similarities of three GO branches and each of the expression similarities from four yeast microarray studies [31-34]. The results are shown in Tables 4, 5 and 6. For the maximum protein similarity approach, the Resnik similarity yielded higher average correlations than Lin and Jiang similarities gave, which is consistent with Sevilla’s results [7]. For all these three approaches of protein semantic similarities, the SimIC always gave higher average correlation values than those Lin, Jiang and GIC similarities give, - 11 - although the standard deviation of correlations are not always lower. Meanwhile, the relevance similarities only gave lower average correlations than those Lin, Jiang and GIC similarities gave in some of cases. These results further confirm that the SimIC will be more effective than relevance similarity in fine tuning up the semantic similarities. The Resnik similarity resulted in the highest average correlation for maximum protein similarity approach and the SimIC yielded the highest average correlations for both average and best match average protein similarity approaches. Overall, the best match average approach based on SimIC gave the highest average correlation value of 0.785. These results suggest that SimIC will improve the average correlations between semantics similarities and gene expression similarities of proteins. Improvement of protein interaction prediction using SimIC Recently, Wu et al. [28] predicted protein interactions based on RSS between protein pairs. They predicted “gold standard” protein interactions using 0.8 as threshold for BP and CC RSS similarity values. Their predictions are proven to be consistent with known experimental interactions and complexes. RSS is a similarity measure based on the structural information only. As discussed in the introduction above, it may be insufficient to measure the similarity between protein pairs. So we hypothesized that predicting protein interactions based on RSS may not be effective. In an example provided in Wu’s paper, the authors proposed that LRP1 is a new member of complex DNA ligase IV (MIPS complex 510.180.30.20) because LRP1 has high RSS similarities with other two member of this complex, DNL4 and LIF1. For the maximum protein similarity approach, the RSS between DNL4 and LRP1and between LRP1 and LIF1 are 0.818 for CC and 1.000 for BP. Meanwhile, for the maximum protein similarity approach, the SimIC similarities between DNL4 and LRP1and between LRP1 and LIF1 are both 0.194 for CC and 0.864 for BP. By looking into cellular component GO terms annotated LRP1, DNL4 and LIF1, we find that the MICA of CC GO terms annotated LRP1 - 12 - and LIF1 is “nuclear part” with id GO:0044428 (see Figure 15), also inferred by the MICA of CC GO terms annotated LRP1 and DNL4 (see Figure 16). Although the “nuclear part” is at the fifth level of the GO cellular component tree, there are 1662 annotations in total associated with this GO term and its descendants, which means that this GO term is a general GO term and is weakly informative. Our SimIC considers this fact, hence providing lower similarity values. We also observed that some protein pairs with low RSS actually score quite high with SimIC similarities. An example is yeast protein pair HSP12 and MUC1. The BP RSS of this pair is 0.270 while the SimIC is 0.847. After further investigating the BP GO terms annotated HSP12 and MUC1, we found that the MICA (also the most recent common ancestor (MRCA)) of these BP GO terms is “cell adhesion” with id GO:0007155 (see Figure 17). Although the “cell adhesion” is at the third level of GO biological process tree, there are only 9 annotations associated with this GO term and its descendants, which means this term is very informative. Our SimIC similarities take into account both structure and information content, and then give out high similarity value between HSP12 and MUC1. On the other hand, RSS gives low similarity value purely based on the structure information of the GO terms. By integrating the information content and structural information of the GO, our SimIC similarities should be more effective in predicting protein interactions. To further investigate the effect of information content on RSS measure, we compared the RSS values of all possible yeast protein pairs with those of SimIC. As Wu et al. used the RSS values 0.8 as threshold to determine their gold standard positive, we wanted to examine how many protein pairs with high RSS values (>0.8) will have a low SimIC (<=0.2). A statistical study (see Methods) showed that protein pairs were unlikely to interact if their SimIC is less than or equal to 0.2. As shown in Table 7, we obtained 234976 protein pairs with high BP RSS (>0.8) (see Supplementary Table 1). Among these 234976 protein pairs, - 13 - 681 pairs have low SimIC similarities (<=0.2) with a percentage of 0.29%. We also obtained 172862 pairs of proteins with high CC RSS (>0.8) (see Supplementary Table 2). Among these, 81515 pairs had low SimIC values (<=0.2) with a percentage of 5.96%. Furthermore, for 12956 pairs of proteins with both BP and CC RSS values greater than 0.8 (see supplementary table 3), 1460 pairs with either BP or CC SimIC values are less than or equal 0.2 with a percentage of 11.27%. We believe the reason that we detected fewer pairs of interacting proteins than those obtained by Wu et al. may be because of the different version of GO annotation we used or because of only “is_a” and “part_of” relationship in DAG tested in this study. Our results show that “shallow annotation” affects the calculation of semantics similarity. RSS, which utilize only the structural information of GO DAG, may lead to a ~10% increase in false positive protein interaction predictions. Then, we applied our SimIC to predict yeast protein interactions. In order to compare with Wu’s study, we used the maximum protein similarity approach. First, the statistical significances of SimIC were obtained by comparing the SimIC similarities of interactions between proteins from MIPS complex [35, 36] with those of randomly sampled protein interactions (see Methods section for details). As shown in Figure 18, in two of 100 bins: BP (0.7, 0.8], CC (0.8, 0.9] and BP (0.8, 0.9], CC (0.8, 0.9], the MIPS complex protein interactions were 1580 and 774 standard deviation greater than the average number of random protein pair in those bins, respectively. Following Wu et al, we use 0.7 as threshold for the BP SimIC and 0.8 as threshold for CC SimIC to predict yeast protein interactions. If a protein pair’s BP SimIC was greater than 0.7 and its CC SimIC was greater than 0.8, they were predicted to interact. In total, we predicted 20484 protein-protein interactions (PPIs) between 2462 proteins (Table 8 and Supplemental Table 4). Among 20484 PPIs, there were 6076 PPIs are in the MIPS data, which includes 8250 protein pairs with both BP and CC annotations. Thus, our predictions revealed 73.65% of MIPS PPIs - 14 - Figure 19 shows the predicted protein interaction network. The network consists of 128 connected components. The largest component has 1871 proteins and 18269 interactions. As an example, we examined the predicted interactions of two proteins mentioned above in complex DNA ligase IV: DNL4 and LIF1. Figure 20 shows that, in predicted PPIs, the DNL4 and LIF1 interact and both are connected to two proteins in the Ku complex: YKU80 and YKU70. The link between DNL4 and LIF1 and the link between YKU70 and YKU80 are verified in both MIPS database [35] and DIP database [37]. The links between DNL4, LIF1 and YKU70, YKU80 are not presented in both databases. However, there is experimental evidence showed that Ku complex can form a temporary complex with DNA ligase IV complex [38]. Then, we examined how MIPS complexes [35] exist in our predicted PPIs. Only 214 MIPS complexes with at least two distinct members were included in our study. Members of 202 complexes were found in our predicted PPIs (See Supplemental Table 5). There were 159 complexes of which all members were detected. Meanwhile, only 120 of 214 complexes of which all members were found in Wu’s predicted protein network [28]. This result implies our predicted networks are more consistent in detecting experimentally verified protein complexes. Assessment of prediction of protein-protein interactions using different similarity measures To further examine the effectiveness of SimIC in PPI prediction, we used the area under receiver operating characteristics (ROC) curve (AUC) [39] method to evaluate the performance of classifiers based on different semantic similarity measures. The 8250 protein pairs from MIPS [35, 36] complex data with both BP and CC annotation were used as positive controls and an identical number of protein pairs were selected randomly as a negative control dataset. Three types of protein similarity approach based on Lin, Jiang, Resnik, GIC, relevance, SimIC, and RSS were tested and the combination of BP and CC - 15 - similarities were used to predict PPIs. Tests were repeated ten times with different sampling negative control dataset have been performed. As shown in Table 9, although the performances of relevance, SimIC and GIC semantic measures were close, the best match average approach based on SimIC give overall best performance as indicated by the AUC value. This comparison demonstrates that the proposed SimIC is superior in predicting yeast protein interactions. Conclusions The Gene Ontology has been increasingly used to annotate proteins, and in some cases is the only clue to biological function in a newly sequenced gene. The effectiveness of semantic similarity measures is an important consideration for the comparison of proteins based solely on their GO annotations. Most current existing semantic similarity measures do not effectively combine the structural information of GO DAG and information content of GO terms. In this paper, we proposed a new measure, SimIC, which combines both information content of GO terms and structural information of Gene Ontology hierarchy. The SimIC can efficiently address the shallow annotation problem. We compared our proposed SimIC measure with Lin, Jiang, GIC, relevance, RSS, Resnik and adjusted Resnik similarities through the study of the correlation between semantic similarities of yeast protein pairs and their gene expression similarities and the correlation between semantic similarities of yeast protein pairs and their sequence similarities. The results show that the similarities obtained by our SimIC measure have the highest average correlation with gene expression similarities and sequence similarities. Furthermore, we demonstrate that our SimIC measure is superior to RSS in predicting protein interactions. We have successfully predicted 20484 protein-protein interactions (PPIs) between 2462 proteins, and the predicted protein interaction network covers more MIPS protein complex that the protein interaction network of Wu et al. does. Moreover, a quick examination of our PPI network detected the predicted protein interactions - 16 - between Ku complex and DNA ligase IV complex, which are supported by biological evidence. Although the SimIC measure is promising, there are reasons to further improve the semantic similarity measure. In our current approach, we have not distinguished the “is_a” and the “part_of” relationships among GO terms. In the future, we will investigate the difference of the information content between the “is_a” and the “part_of” relationships. In addition, two terms are semantically close if they have “regulates”, “positively_regulates” and “negatively_regulates” relation with the same term, even though they are not directly related. Utilization of the “regulates”, “positively_regulates” and “negatively_regulates” relationships in the semantic similarity measure will require further investigation. Furthermore, it is worthwhile to explore the role of evidence codes of GO annotation in the semantic similarity measure. Methods GO database and Yeast GO annotation The Gene Ontology dated April 16th 2008 was downloaded from gene ontology consortium website (www.geneontology.org). The yeast gene annotation database was from Saccharomyces Genome Database (www.yeastgenome.org) and the April 12th 2008 release was used. Proteins only annotated with “biological process”, “molecular function” and “cellular component” were filtered out from the analyses. And the annotations with IEA evidence code are also ignored in this study. After this filtering, there were 4620 proteins with biological process annotations, 3876 proteins with molecular function annotation and 5077 proteins with cellular component annotation. Furthermore, only the “is_a” and “part_of” relationship are used to build the GO DAG. The “regulates”, “positively_regulates” and “negatively_regulates” relationships are not included in our present study. - 17 - New semantic similarity measures by integrating IC and structure of GO DAG Information coefficient similarity (SimIC). The SimIC between two GO term t1 and t2 is defined as: (3) 1 2  log( prob( MICA)) ) SimIC (t1, t2 )   (1  1  IC ( MICA) log( prob(t1 ))  log( prob(t2 )) Theoretically, the values of SimIC are also between zero and one. Practically, the yeast has only about 13683 BP annotations, and then the highest IC of BP GO term will be about 9.52 on natural log base. Therefore, the BP SimIC will have maximum values around 0.9. Adjusted Resnik similarity. In order to add structural information into Resnik similarity, we simply adjust it with Lin similarity: Sim Adjusted  Re snik  SimRe snik  SimLin (4) Similar to original Resnik similarity, this adjusted Resnik similarity has no maximum value and has a minimum zero. Existing semantic similarity measures Resnik similarity. Resnik introduced a semantic similarity for “is_a” ontologies based the highest IC values among IC values of all common ancestors of two terms [21]: Sim Re snik ( t1 , t2 )  max( IC ( t )) (5) tS ( t1 ,t2 ) where S(t1, t2) is the set of common ancestors of two term t1 and t2. The Resnik similarity has no maximum values and has a minimum zero. Lin similarity. Lin [23] developed a information-theoretic similarity applicable to any domain that can be described by a probabilistic model. Lin measure is based on the relative probability between two terms and their MICA. The Lin measure between two GO term t1 and t2 is defined as: Sim Lin (t1 , t 2 )  2  log( prob( MICA)) log( prob(t1 ))  log( prob(t 2 )) (6) The value of Lin similarity ranges from zero to one. - 18 - Jiang similarity. The Jiang and Conrath [22] integrated the edge-based method with the node-based approach of the information content calculation to develop a new distance measure. For its simple case in which factors related to local density, node depth and link type are ignored, the Jiang measure between two GO term t1 and t2 is defined as: DisJiang(t1, t2 )  IC(t1 )  IC(t2 )  2  IC(MICA(t1, t2 )) (7) Jiang distance measure can easily be transformed into a similarity measure by adding one and inverting it [10]. Sim Jiang ( t1 , t 2 )  1 1  Dis J  C ( t1 , t 2 ) (8) If terms t1 and t2 are the same, DistJiang(t1,t2) should be 0. Adding one is to avoid the division of 0. The value of Jiang similarity ranges from zero to one Graph information content (GIC) similarity. Let DAGT1 and DAGT2 be two ancestor DAGs induced by two GO terms t1 and t2 [21]. Then, the graph information content (GIC) [26] measure is defined as the ratio between sum of information content of GO terms in the intersection of DAGT1 and DAGT2 and sum of information content of GO terms in the union of DAGT1 and DAGT2: SimGIC (t1 , t 2 )    tDAGT 1  DAGT 2 tDAGT 1  DAGT 2 IC (t ) (9) IC (t ) The values of GIC similarity ranges from zero to one. Relative Specificity Similarity (RSS). The RSS of two GO terms is obtained from three components α, β and γ. Component α is length of path from the root to most recent common ancestor (MRCA). Component β is the maximum length of paths from each of GO terms to its leaf descendants. Component γ is the sum of distance between MRCA and two GO terms. Thus, the RSS between two GO terms can be calculated by[28]: RSS (t1 , t2 )  max Depth GO max Depth GO       (10) - 19 - where the maxDepthGO is the maximum distance from the GO root to the leaf term. The values of RSS are between 0 and 1. Schlicker’s relevance similarity. Schlicker et al. pointed out that both the specific of MICA and the relation between two GO terms with MICA are needed to be considered in a similarity measure. Then, they developed a new relevance similarity measure by combining the Lin similarity with the probability of MICA [24]. (11) Sim Re l (t1 , t 2 )  max ( Sim Lin  (1  prob( MICA)) tS ( t1 ,t2 ) The values of relevance similarity are also between zero and one. We also evaluate the varieties of relevance similarity by using probability of MICA to adjust Jiang similarity. Calculation of similarities between proteins A protein can be annotated with a set of GO terms. Therefore, similarity between two proteins can be calculated based on the similarities between two set of GO terms. Here, we investigated three approaches for computing protein similarity based on GO term similarities [4, 13, 24, 26]. Best match average similarity. The similarity between two proteins is based on average of best match GO term pairs [13]. Given two proteins p1 and p2, let go1 and go2 represent their corresponding sets of GO terms. First, the similarity between one GO term t and a set of GO terms, go = {t1, t2, … tk) is defined as the maximum similarity between the t and any member in set go. Sim(t, go)  max( Sim(t , ti )) (12) 1i k Therefore, the similarity between two proteins can be defined as the weighted average of the term similarity scores:  Sim (t1i , go 2 )   Sim (t 2 j , go1 ) Sim ( p1 , p 2 )  1 i  m 1 j  n (13) mn - 20 - where m is the number of terms in go1 and n is the number of terms in go2. In the best match average protein similarity measures, proteins with more GO terms annotated will have more influence on the overall similarity score. Average similarity. The similarity between two proteins, p1 and p2, is the average of similarities among all pairs of two GO term sets, go1 and go2:  Sim ( p1 , p 2 )  1 i  m ,1 j  n Sim (t i , t j ) (14) mn where m is the number of terms in go1 and n is the number of terms in go2. Maximum similarity. The similarity between two proteins, p1 and p2, is the maximum similarity among all pairs of two GO term sets, go1 and go2: Sim( p1 , p2 )  max 1i m,1 j n (15) ( Sim(ti , t j )) where m is the number of terms in go1 and n is the number of terms in go2. Calculation of correlation between semantic similarity and sequence similarity The correlation between semantic similarity and sequence similarity is measured by the correlation between semantic similarity values and the BLAST log bit scores [4, 5] between sequences. An all-against-all BLAST search for yeast proteins are performed on a local copy of NCBI BLAST [40]. The expect value threshold 100 was used. As the BLAST result is not symmetric, for each protein pair, the final sequence similarity is the average of two BLAST results between them [26]. Sim seq ( P1 , P2 )  log10 ( BScore ( P1 , P2 )  BScore ( P2 , P1 ) ) 2 (16) where BScore is BLAST bit score. Then, the correlation is calculated based on protein pairs with both BLAST bit scores and semantic similarity values. First, the semantic similarities were split into 50 intervals. For each interval, the average of semantic similarities that fall in this interval and the average of corresponding BLAST log bit scores are calculated. Therefore, the Pearson correlation is calculated based on these 50 pairs of average values. - 21 - Calculation of correlation between semantic similarity and gene expression similarity Four microarray profiles have been used for the calculation [31-34]. All four data sets have at least 50 experiment data points, which should ensure the robust of the calculation of correlation between gene expressions. First, the Pearson correlations between expressions of genes were calculated. Only genes with the number of missing values less than ten percentage of total experiments were included. Missing values were estimated using nearest neighbour based method [41]. Only the protein pairs with both expression correlations value and semantic similarity values were included for analysis. The semantic similarities were split into 50 intervals. Then, the correlation between semantic similarity and expression similarity was calculated based on the average values of intervals [7]. For each interval, the average of semantic similarities that fall in this interval and the average of expression correlations between corresponding protein pairs were calculated. Then, the Pearson correlation is calculated based on the average values of semantic similarities and expression correlation values of 50 intervals. Determination of statistical significant SimIC of yeast protein interactions First, the BP and CC SimIC similarities of 8250 interactions between proteins from MIPS complex [35, 36] were calculated. Then, the BP or CC SimIC similarities values were partitioned into 10 bins, respectively. The ith bin includes values with i*bin_size <values≤ (i+1)*bin_size, where bin_size=0.1. Combining the BP and CC similarity bins, there were 100 bins in total. The protein pairs were distributed into 100 bins based on the combination of BP and CC similarity values. Thereafter, we randomly sample 8250 protein interactions and calculate their BP and CC SimIC similarities. The number of protein pairs in each of 100 bins was counted based on the BP and CC similarity values. We repeated this process one thousand times. Finally, the number of MIPS complex protein interactions in each bin was - 22 - compared to the average and standard deviation of number of random protein interactions in each bin to obtain a Z score: Z Score  (# parisMIPS  Average# pairsrandom ) / SDrandom (17) The Figure 18 shows the Z score in each of 100 bins. The MIPS complex protein interactions have shown extremely high Z scores in two of 100 bins: BP (0.7, 0.8], CC (0.8, 0.9] and BP (0.8, 0.9], CC (0.8, 0.9]. ROC curve analysis The receiver operating characteristics (ROC) [39] evaluates the performance of classifiers based on the tradeoff between specificity and sensitivity. The area under the ROC curve (AUC) can be used to compare the prediction performance. While an area of 1 means perfect prediction, an area of 0.5 indicates random prediction. We employed the ROCR [42] library to draw the ROC curves and calculate the AUC values. Authors' contributions FL developed the methods. BL and FL implemented the methods. BL, JZW, AF JZZ and FL wrote the paper. Acknowledgements FL and BL are supported by NSF EPSCoR grants EPS-0447660. Feng Luo would like to acknowledge financial support from the Institute for Modeling and Simulation Applications at Clemson University. The authors would like to thank Dr. Andread Schlicker for providing the IO, HSS, LSS and NSS datasets. JZW’ work is partially supported by the National Institutes of Health contract 1R15CA131808-01. References 1. 2. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT et al: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nature genetics 2000, 25(1):25-29. del Pozo A, Pazos F, Valencia A: Defining functional distances over gene ontology. BMC bioinformatics 2008, 9(1):50-50. - 23 - 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. Couto FM, Silva MJ, Coutinho PM: Implementation of a functional semantic similarity measure between gene-products. 2003. Lord PW, Stevens RD, Brass A, Goble CA: Investigating semantic similarity measures across the Gene Ontology: the relationship between sequence and annotation. Bioinformatics (Oxford, England) 2003, 19(10):1275-1283. Lord PW, Stevens RD, Brass A, Goble CA: Semantic similarity measures as tools for exploring the gene ontology. Pacific Symposium on Biocomputing 2003:601612. Wang H, Azuaje F, Bodenreider O, Dopazo J: Gene expression correlation and gene ontology-based similarity: an assessment of quantitative relationships. Proceedings of the 2004 IEEE Symposium on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB2004) 2004:25-31. Sevilla JL, Segura V, Podhorski A, Guruceaga E, Mato JM, Martinez-Cruz LA, Corrales FJ, Rubio A: Correlation between gene expression and GO semantic similarity. IEEE/ACM transactions on computational biology and bioinformatics / IEEE, ACM 2005, 2(4):330-338. Guo X, Liu R, Shriver CD, Hu H, Liebman MN: Assessing semantic similarity measures for the characterization of human regulatory pathways. Bioinformatics (Oxford, England) 2006, 22(8):967-973. Zhang P, Zhang J, Sheng H, Russo JJ, Osborne B, Buetow K: Gene functional similarity search tool (GFSST). BMC bioinformatics 2006, 7:135. Couto FM, Silva MJ, Coutinho PM: Measuring semantic similarity between Gene Ontology terms. Data & Knowledge Engineering 2007, 61:137-152. Schlicker A, Albrecht M: FunSimMat: a comprehensive functional similarity database. Nucleic acids research 2007. Tao Y, Sam L, Li J, Friedman C, Lussier YA: Information theory applied to the sparse gene ontology annotation network to predict novel gene function. Bioinformatics (Oxford, England) 2007, 23(13):i529-538. Wang JZ, Du Z, Payattakool R, Yu PS, Chen CF: A new method to measure the semantic similarity of GO terms. Bioinformatics (Oxford, England) 2007, 23(10):1274-1281. Yu H, Jansen R, Stolovitzky G, Gerstein M: Total ancestry measure: quantifying the similarity in tree-like classification, with genomic applications. Bioinformatics (Oxford, England) 2007, 23(16):2163-2173. Couto FM, Silva MJ, Coutinho PM: Semantic similarity over the gene ontology: Family correlation and selecting disjunctive ancestors. In: Proceedings of the 14th ACM international conference on Information and knowledge management: 2005: ACM New York, NY, USA; 2005: 343-344. Lerman G, Shakhnovich BE: Defining functional distance using manifold embeddings of gene ontology annotations. Proceedings of the National Academy of Sciences 2007, 104(27):11334-11334. Speer N, Frohlich H, Spieth C, Zell A: Functional grouping of genes using spectral clustering and gene ontology. In: Proc Int Joint Conf Neural Networks: 2005; 2005: 298–303-298–303. Frohlich H, Speer N, Spieth C, Zell A: Kernel based functional gene grouping. In: Neural Networks, 2006 IJCNN'06 International Joint Conference on: 3580-3585. Mistry M, Pavlidis P: Gene ontology term overlap as a measure of gene functional similarity. BMC bioinformatics 2008, 9(1):327-327. - 24 - 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 36. 37. Pesquita C, Faria D, Bastos H, Ferreira AE, Falcao AO, Couto FM: Metrics for GO based protein semantic similarity: a systematic evaluation. BMC bioinformatics 2008, 9 Suppl 5:S4. Resnik P: Using Information Content to Evaluate Semantic Similarity in a Taxonomy. Proceedings of the 14th International Joint Conference on Artifical Intelligence (IJCAI-95) 1995. Jiang JJ, Conrath DW: Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy In the Proceedings of ROCLING X 1997. Lin D: An Information-Theoretic Definition of Similarity. Proceedings of International Conference on Machine Learning 1998. Schlicker A, Domingues FS, Rahnenfuhrer J, Lengauer T: A new measure for functional similarity of gene products based on Gene Ontology. BMC bioinformatics 2006, 7:302. Wu Z, Palmer M: Verb semantics and lexical selection. Proceeding of 32th annual meeting of the association for computational linguistics 1994. Pesquita C, Faria D, Bastos H, Falcão AO, Couto FM: Evaluating GO-based Semantic Similarity Measures. In: Proceedings of the 10th Annual Bio-Ontologies Meeting (Bio-Ontologies 2007) 2007; Vienna, Austria; 2007: 37-40. WordNet, An Electronic Lexical Database: The MIT Press; 1998. Wu X, Zhu L, Guo J, Zhang DY, Lin K: Prediction of yeast protein-protein interaction network: insights from the Gene Ontology and annotations. Nucleic acids research 2006, 34(7):2137-2150. Dwight SS, Balakrishnan R, Christie KR, Costanzo MC, Dolinski K, Engel SR, Feierbach B, Fisk DG, Hirschman J, Hong EL et al: Saccharomyces genome database: underlying principles and organisation. Brief Bioinform 2004, 5(1):9-22. Berglund AC, Sjolund E, Ostlund G, Sonnhammer ELL: InParanoid 6: eukaryotic ortholog clusters with inparalogs. Nucleic acids research 2007. Gasch AP, Huang M, Metzner S, Botstein D, Elledge SJ, Brown PO: Genomic expression responses to DNA-damaging agents and the regulatory role of the yeast ATR homolog Mec1p. Molecular biology of the cell 2001, 12(10):2987-3003. Gasch AP, Spellman PT, Kao CM, Carmel-Harel O, Eisen MB, Storz G, Botstein D, Brown PO: Genomic expression programs in the response of yeast cells to environmental changes. Molecular biology of the cell 2000, 11(12):4241-4257. O'Rourke SM, Herskowitz I: Unique and redundant roles for HOG MAPK pathway components as revealed by whole-genome expression analysis. Molecular biology of the cell 2004, 15(2):532-542. Spellman PT, Sherlock G, Zhang MQ, Iyer VR, Anders K, Eisen MB, Brown PO, Botstein D, Futcher B: Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Molecular biology of the cell 1998, 9(12):3273-3297. Mewes HW, Amid C, Arnold R, Frishman D, Guldener U, Mannhaupt G, Munsterkotter M, Pagel P, Strack N, Stumpflen V et al: MIPS: analysis and annotation of proteins from whole genomes. Nucleic acids research 2004, 32(Database issue):D41-44. Lu LJ, Xia Y, Paccanaro A, Yu H, Gerstein M: Assessing the limits of genomic data integration for predicting protein networks. Genome research 2005, 15(7):945953. Salwinski L, Miller CS, Smith AJ, Pettit FK, Bowie JU, Eisenberg D: The Database of Interacting Proteins: 2004 update. Nucleic acids research 2004, 32(Database issue):D449-451. - 25 - 38. 39. 40. 41. 42. 43. Nick McElhinny SA, Snowden CM, McCarville J, Ramsden DA: Ku recruits the XRCC4-ligase IV complex to DNA ends. Mol Cell Biol 2000, 20(9):2996-3003. Zweig MH, Campbell G: Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin Chem 1993, 39(4):561-577. http://blast.ncbi.nlm.nih.gov/Blast.cgi. Troyanskaya O, Cantor M, Sherlock G, Brown P, Hastie T, Tibshirani R, Botstein D, Altman RB: Missing value estimation methods for DNA microarrays. Bioinformatics (Oxford, England) 2001, 17(6):520-525. Sing T, Sander O, Beerenwinkel N, Lengauer T: ROCR: visualizing classifier performance in R. Bioinformatics (Oxford, England) 2005, 21(20):3940-3941. Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, Amin N, Schwikowski B, Ideker T: Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome research 2003, 13(11):2498-2504. - 26 - Figures Figure 1 – Gene ontology DAG illustration Figure 2 – Distribution of Lin, relevance and SimIC of all pairs of annotated yeast proteins The values of all similarities are partitioned into 21 bins. The first bin includes number of protein pairs with zero values. The following ith bin includes values number of protein pairs with i*bin_size <values≤ (i+1)*bin_size, where bin_size=(max-min)/20. Inner figure shows the distribution of last four bins. Figure 3 – Examples of different levels of information coefficient corrections Figure 4 – Scatter plot of relevance similarities against SimIC similarites of yeast protein pairs with Lin’s similarity greater than 0.9 Figure 5 – Scatter plot of yeast protein pair BP semantic similarity against their sequence similarity The protein similarities are calculated using best math average method and the SimIC is used to calculate semantic similarities among GO terms. The semantic similarities are split into 50 intervals. For each interval, the average of BP semantic similarities that fall in this interval and the average of corresponding BLAST log bit scores are calculated. The trend line is a linear regression of the data. Figure 6 – Scatter plot of yeast protein pair CC semantic similarity against their sequence similarity The protein similarities are calculated using best math average method and the SimIC is used to calculate semantic similarities among GO terms. The semantic similarities are split into 50 intervals. For each interval, the average of CC semantic similarities that fall in this interval and the average of corresponding BLAST log bit scores are calculated. The trend line is a linear regression of the data. - 27 - Figure 7 – Scatter plot of yeast protein pair MF semantic similarity against their sequence similarity The protein similarities are calculated using best math average method and the SimIC is used to calculate semantic similarities among GO terms. The semantic similarities are split into 50 intervals. For each interval, the average of MF semantic similarities that fall in this interval and the average of corresponding BLAST log bit scores are calculated. The trend line is a linear regression of the data. Figure 8 – Distribution of the SimIC BP similarities of human and yeast protein pairs in four different sets The bins correspond to the following intervals of protein similarity values: S0.0: [min, min+0.2*range]; S0.2: [min+0.2*range, min+ 0.4*range]; S0.4: [min+0.4*range, min+0.6*range]; S0.6: [min+0.6*range, min+0.8*range]; S0.8: [min+0.8*range, max]. Here, min and max is the minimum and maximum of similarities and range = max-min. Figure 9 – Distribution of the SimIC MF similarities of human and yeast protein pairs in four different sets The bins correspond to the same intervals as in legend of Figure 7. Figure 10 – Distribution of the SimIC BP and MF human and yeast protein similarities for orthology dataset The bins correspond to the same intervals as in legend of Figure 7. Figure 11 – Distribution of the GIC BP similarities of human and yeast protein pairs in four different sets The bins correspond to the same intervals as in legend of Figure 7. Figure 12 – Distribution of the GIC MF similarities of human and yeast protein pairs in four different sets The bins correspond to the same intervals as in legend of Figure 7. Figure 13 – Distribution of the relevance BP similarities of human and yeast protein pairs in four different sets The bins correspond to the same intervals as in legend of Figure 7. - 28 - Figure 14 – Distribution of the relevance MF similarities of human and yeast protein pairs in four different sets The bins correspond to the same intervals as in legend of Figure 7. Figure 15 – Visualization of GO annotations for yeast genes LRP1 and LIF1 The orange nodes are GO terms annotating LRP1 and cyan nodes are GO terms annotating gene LIF1. Figure 16 – Visualization of GO annotations for yeast genes LRP1 and DNL4 The orange nodes are GO terms annotating gene LRP1 and the cyan nodes are GO terms annotating gene DNL4. Figure 17 – Visualization of GO annotations for yeast genes HSP12 and MUC1 The orange nodes are GO terms annotating gene HSP12 and the cyan nodes are GO terms annotating gene MUC1. Figure 18 – Statistical significance of SimICsimilarity (SimIC) The values of BP SimICand CC SimIC are partitioned into 10 bins, respectively. The ith bin includes values number of protein pairs with i*bin_size <values≤ (i+1)*bin_size, where bin_size=0.1. The Z-score for each combination bin of BP and CC are calculated by (#pairsMIPS –Average #pairsrandom)/SDrandom. Figure 19 – The predicted yeast protein interaction networks using SimIC The graph is draw using cytoscape [43]. Figure 20 – Predicted protein interactions between Ku complex (YKU70, YKU80) and DNA ligase IV complex (DNL4, LIF1) The existing protein interactions are represented by black links. The predicted protein interactions are represented by red links. - 29 - Tables Table 1 - Correlations between semantic similarities (maximum) and sequence similarities of yeast proteins The varieties of relevance similarity and SimIC also are used in comparison. The highest average correlation is indicated by bold. BP MF CC Average Std RSS 0.691 0.694 0.535 0.640 0.091 Resnik 0.899 0.905 0.861 0.888 0.024 No adjustment Resnik*lin Lin 0.911 0.714 0.894 0.839 0.870 0.732 0.762 0.892 0.021 0.068 jiang 0.505 0.311 0.465 0.427 0.102 GIC 0.777 0.705 0.702 0.728 0.042 Relevance similarities lin Jiang 0.701 0.593 0.879 0.544 0.657 0.558 0.746 0.565 0.118 0.025 SimIC lin 0.738 0.855 0.767 0.787 0.061 jiang 0.709 0.765 0.738 0.737 0.028 Table 2 - Correlations between semantic similarities (average) and sequence similarities of yeast proteins The varieties of relevance similarity and SimIC also are used in comparison. The highest average correlation is indicated by bold. BP MF CC Average Std RSS 0.788 0.803 0.631 0.741 0.095 Resnik 0.895 0.898 0.616 0.803 0.162 No adjustment Resnik*lin lin 0.894 0.833 0.856 0.906 0.460 0.570 0.737 0.770 0.240 0.177 jiang 0.649 0.687 0.361 0.566 0.178 GIC 0.777 0.705 0.702 0.728 0.042 Relevance similarities lin Jiang 0.829 0.841 0.868 0.623 0.664 0.681 0.787 0.715 0.108 0.113 SimIC lin 0.863 0.872 0.735 0.823 0.077 jiang 0.798 0.656 0.774 0.743 0.076 Table 3 - Correlations between semantic similarities (best match average) and sequence similarities of yeast proteins The varieties of relevance similarity and SimIC also are used in comparison. The highest average correlation is indicated by bold. BP MF CC Average Std RSS 0.825 0.832 0.714 0.790 0.066 Resnik 0.931 0.942 0.647 0.840 0.167 No adjustment Resnik*lin lin 0.922 0.910 0.910 0.848 0.545 0.770 0.792 0.843 0.214 0.070 jiang 0.889 0.754 0.759 0.801 0.077 - 30 - GIC 0.777 0.705 0.702 0.728 0.042 Relevance similarities lin Jiang 0.895 0.908 0.938 0.710 0.840 0.789 0.891 0.802 0.049 0.100 SimIC lin 0.915 0.892 0.871 0.893 0.022 jiang 0.935 0.838 0.826 0.866 0.060 Table 4 - Correlations between semantic similarities (maximum) and expression similarities of yeast proteins The varieties of relevance similarity and SimIC also are used in comparison. The highest average correlation is indicated by bold. Four microarray profiles [31-34] are denoted as “MAPK” [33], “Damage”[31], “cell cycle”[34] and “ENV” [32], respectively. Noted that the GIC semantic similarity measure itself is a protein similarity measure and has the same values in Table 4, 5, 6. BP-MAPK MF-MAPK CC-MAPK BP-Damage MF-Damage CC-Damage BP-cell cycle MF-cell cycle CC-cell cycle BP-ENV MF-ENV CC-ENV Average Std RSS 0.511 0.419 0.583 0.497 0.464 0.672 0.422 0.537 0.668 0.427 0.36 0.707 0.522 0.114 Resnik 0.631 0.564 0.550 0.849 0.752 0.640 0.894 0.844 0.806 0.884 0.786 0.718 0.743 0.122 No adjustment Resnik*lin lin 0.575 0.575 0.638 0.505 0.62 0.323 0.715 0.727 0.800 0.682 0.671 0.472 0.733 0.760 0.838 0.615 0.824 0.699 0.797 0.731 0.791 0.541 0.733 0.398 0.728 0.586 0.086 0.141 jiang 0.200 0.007 0.209 0.398 0.275 0.043 0.501 0.022 0.616 0.466 0.290 0.040 0.252 0.211 GIC 0.774 0.539 0.530 0.716 0.611 0.666 0.793 0.636 0.721 0.757 0.604 0.658 0.667 0.09 Relevance similarities lin Jiang 0.499 0.354 0.556 0.147 0.299 0.392 0.568 0.505 0.533 0.479 0.477 0.205 0.583 0.528 0.581 0.455 0.687 0.393 0.612 0.527 0.543 0.348 0.484 0.295 0.535 0.386 0.094 0.124 lin 0.568 0.640 0.519 0.716 0.707 0.630 0.756 0.774 0.798 0.731 0.737 0.686 0.689 0.084 SimIC jiang 0.378 0.532 0.372 0.531 0.669 0.469 0.580 0.746 0.697 0.547 0.681 0.520 0.560 0.121 Table 5 - Correlations between semantic similarities (average) and expression similarities of yeast proteins The varieties of relevance similarity and SimIC also are used in comparison. The highest average correlation is indicated by bold. Four microarray profiles [31-34] are denoted as “MAPK” [33], “Damage”[31], “cell cycle”[34] and “ENV” [32], respectively. Noted that the GIC semantic similarity measure itself is a protein similarity measure and has the same values in Table 4, 5, 6. BP-MAPK MF-MAPK CC-MAPK BP-Damage MF-Damage CC-Damage BP-cell cycle MF-cell cycle CC-cell cycle BP-ENV MF-ENV CC-ENV Average Std RSS 0.658 0.392 0.570 0.641 0.577 0.534 0.682 0.604 0.728 0.695 0.649 0.632 0.613 0.089 Resnik 0.580 0.614 0.321 0.803 0.782 0.516 0.822 0.882 0.529 0.794 0.822 0.544 0.667 0.173 No adjustment Resnik*lin lin 0.435 0.762 0.633 0.531 0.254 0.426 0.640 0.794 0.732 0.704 0.420 0.610 0.707 0.820 0.856 0.768 0.387 0.735 0.670 0.789 0.768 0.660 0.456 0.608 0.580 0.684 0.184 0.120 jiang 0.497 0.372 0.185 0.657 0.358 0.517 0.578 0.433 0.629 0.597 0.122 0.526 0.456 0.170 - 31 - GIC 0.774 0.539 0.530 0.716 0.611 0.666 0.793 0.636 0.721 0.757 0.604 0.658 0.667 0.09 Relevance similarities lin Jiang 0.788 0.450 0.507 0.300 0.384 0.423 0.832 0.593 0.578 0.581 0.530 0.565 0.881 0.661 0.696 0.553 0.708 0.606 0.836 0.631 0.595 0.561 0.568 0.569 0.659 0.541 0.155 0.101 lin 0.734 0.631 0.462 0.832 0.746 0.602 0.889 0.812 0.661 0.870 0.751 0.650 0.720 0.125 SimIC jiang 0.444 0.552 0.355 0.558 0.705 0.529 0.647 0.770 0.564 0.562 0.715 0.553 0.580 0.116 Table 6 - Correlations between semantic similarities (best match average) and expression similarities of yeast proteins The varieties of relevance similarity and SimIC also are used in comparison. The highest average correlation is indicated by bold. Four microarray profiles [31-34] are denoted as “MAPK” [33], “Damage”[31], “cell cycle”[34] and “ENV” [32], respectively. Noted that the GIC semantic similarity measure itself is a protein similarity measure and has the same values in Table 4, 5, 6. BP-MAPK MF-MAPK CC-MAPK BP-Damage MF-Damage CC-Damage BP-cell cycle MF-cell cycle CC-cell cycle BP-ENV MF-ENV CC-ENV Average Std RSS 0.835 0.46 0.688 0.82 0.677 0.703 0.855 0.683 0.708 0.847 0.638 0.835 0.46 0.688 Resnik 0.686 0.690 0.389 0.870 0.820 0.587 0.792 0.666 0.637 0.863 0.841 0.686 0.690 0.389 No adjustment Resnik*lin lin 0.582 0.930 0.656 0.622 0.351 0.540 0.808 0.904 0.816 0.750 0.537 0.711 0.721 0.900 0.618 0.752 0.554 0.732 0.798 0.908 0.813 0.686 0.582 0.930 0.656 0.622 0.351 0.540 jiang 0.883 0.661 0.426 0.878 0.649 0.576 0.884 0.651 0.600 0.887 0.693 0.883 0.661 0.426 GIC 0.774 0.539 0.530 0.716 0.611 0.666 0.793 0.636 0.721 0.757 0.604 0.658 0.667 0.09 Relevance similarities lin Jiang 0.938 0.888 0.551 0.534 0.599 0.588 0.926 0.893 0.613 0.640 0.685 0.706 0.922 0.913 0.743 0.666 0.819 0.806 0.921 0.896 0.647 0.685 0.938 0.888 0.551 0.534 0.599 0.588 Table 7 - Protein pairs with high RSS values and low SimIC values Number of protein pairs Number of protein pairs with with RSS>0.8 RSS>0.8 and SimIC<= 0.2 BP 234976 681 CC 172862 81515 BP-CC 12956 1460 - 32 - SimIC lin 0.847 0.665 0.562 0.897 0.767 0.696 0.942 0.831 0.771 0.919 0.787 0.738 0.785 0.111 jiang 0.688 0.632 0.590 0.774 0.754 0.727 0.836 0.840 0.765 0.789 0.802 0.761 0.747 0.077 Percentage 0.29% 5.96% 11.27% Table 8 Proteins pairs with SimIC greater than threshold Number of protein pairs with BP SimIC >0.7 are listed in first row and number of protein pairs with CC SimIC >0.8 are listed in second row. Protein pairs with BP SimIC >0.7 and CC SimIC>0.8 are predicted to be physically interacted (third row). The MIPS dataset include 8250 protein pairs with both BP and CC annotations are used to evaluate recall of prediction. BP CC BP-CC Number of protein 4775 2698 2462 Number of protein pairs with SimIC greater than threshold 520506 (4.57%) 58478 (0.47%) 20484 (0.08%) Number (percentage) of predicted protein pairs in MIPS dataset 6629 (80.35%) 6502 (78.81%) 6076 (73.65%) Table 9 – Comparison of AUC values using different BP and CC similarities to predict protein- protein interactions The highest average correlation is indicated by bold. The data showed is the average (Ave) and standard deviation (Std) of ten times of tests with different sampling negative control dataset. Protein Similarity measures Semantic Similarity measures BMA Ave±Std MAX Ave±Std Average Ave±Std Resnik 0.9586 ±0.007 0.9448 ±0.0014 0.9480 ±0.0008 Lin 0.9526 ±0.008 0.9252 ±0.0012 0.9391 ±0.0011 Jiang 0.9372 ±0.0011 0.9161 ±0.0011 0.9224 ±0.0013 - 33 - GIC 0.9566 ±0.0006 0.9566 ±0.0006 0.9566 ±0.0006 Relevance 0.9655 ±0.0004 0.9637 ±0.0006 0.9541 ±0.0008 SimIC 0.9664 ±0.0004 0.9637 ±0.0006 0.9553 ±0.0007 RSS 0.9316 ±0.0006 0.9479 ±0.0006 0.8726 ±0.0008 Additional files Additional file 1 – Supplemental Table S1 Protein pairs with high BP RSS values (>0.8) and low BP SimIC values (<=0.2). Additional file 2 – Supplemental Table S2 Protein pairs with high CC RSS values (>0.8) and low CC SimIC values (<=0.2). Additional file 3 – Supplemental Table S3 Protein pairs with both high BP RSS values (>0.8) and high CC RSS values (>0.8) and either low BP or low CC SimIC values (<=0.2). Additional file 4 – Supplemental Table S4 Protein pairs with both high BP SimIC values (>0.7) and CC SimIC values (>0.8), which is predicted as true protein interactions. Additional file 5 – Supplemental Table S5 Members of MIPS protein complexes exist in predicted protein networks. - 34 - Figure 1 - 35 - Figure 2 - 36 - Figure 3 - 37 - Figure 4 - 38 - Figure 5 - 39 - Figure 6 - 40 - Figure 7 - 41 - Figure 8 - 42 - Figure 9 - 43 - Figure 10 - 44 - Figure 11 - 45 - Figure 12 - 46 - Figure 13 - 47 - Figure 14 - 48 - Figure 15 - 49 - Figure 16 - 50 - Figure 17 - 51 - Figure 18 - 52 - Figure 19 - 53 - Figure 20 - 54 -
5cs.CE
arXiv:1509.07808v2 [cs.DS] 16 Feb 2018 A (1 + ε)-Approximation for Makespan Scheduling with Precedence Constraints using LP Hierarchies∗ Elaine Levey† Thomas Rothvoss‡ Abstract In a classical problem in scheduling, one has n unit size jobs with a precedence order and the goal is to find a schedule of those jobs on m identical machines as to minimize the makespan. It is one of the remaining four open problems from the book of Garey & Johnson whether or not this problem is NP-hard for m = 3. We prove that for any fixed ε and m, an LP-hierarchy lift of the time-indexed LP with a slightly super poly-logarithmic number of r = (log(n))Θ(log log n) rounds provides a (1 + ε)-approximation. For example Sherali-Adams suffices as hierarchy. This implies an algorithm that yields a (1 + ε)-approximation in time nO(r) . The 7 previously best approximation algorithms guarantee a 2 − 3m+1 -approximation in 4 polynomial time for m ≥ 4 and 3 for m = 3. Our algorithm is based on a recursive scheduling approach where in each step we reduce the correlation in form of long chains. Our method adds to the rather short list of examples where hierarchies are actually useful to obtain better approximation algorithms. 1 Introduction One of the landmarks in the theory of scheduling is the paper of Graham [Gra66] from 1966, dealing with the following problem: suppose we have a set J of n jobs, each one with a running time pj along with m identical parallel machines that we can use to process the jobs. Moreover, the input contains a precedence order on the jobs; we write j ≺ j ′ if job j has to be completed before job j ′ can be started. The goal is to schedule the jobs in a non-preemptive fashion so that the makespan is minimized. Here, the makespan gives the time that the last job is finished. In the 3-field notation 1 , this problem is abbreviated as P | prec | Cmax . Graham showed that the following ∗ The conference version of this work appeared in the 48th ACM Symposium on Theory of Computing (STOC 2016). † University of Washington, Seattle, USA. Email: [email protected] ‡ University of Washington, Seattle, USA. Email: [email protected]. Supported by NSF grant 1420180 with title “Limitations of convex relaxations in combinatorial optimization” and an Alfred P. Sloan Research Fellowship. 1 In the 3-field notation, the first field specifies the available processors, the 2nd field the jobs and the last field the objective function. In our case, P m means that we have m identical machines; pj = 1, prec indicates that the jobs have unit length and precedence constraints and the last field Cmax specifies that the objective function is to minimize the maximum completion time. 1 1 )-approximation on the makespan: compute an arbitrary list schedule gives a (2 − m topological ordering of the jobs and whenever a machine becomes idle, select the first available job from the list. It had been known since the late 70’s that it is NP-hard to approximate the problem better than within a factor of 4/3 due to Lenstra and Rinnooy Kan [LK78] and Schuurman and Woeginger [SW99] prominently placed the quest for any improvement on their well known list of 10 open problems in scheduling. Finally in 2010, Svensson [Sve10] showed that assuming a variant of the unique games conjecture [BK09], there is no (2 − ε)-approximation algorithm for P | prec, pj = 1 | Cmax . However, for unit size jobs, Lam and Sethi [LS77] analyzed an algorithm of Coffman and Graham and 2 for P | prec, pj = 1 | Cmax . showed that it provides a slighly better guarantee of 2 − m 7 Later, Gangal and Ranade [GR08] gave an algorithm with a 2 − 3m+1 guarantee for m ≥ 4. In a typical scheduling application, the number of jobs might be huge compared to the number of machines, which does justify to ask for the complexity status of such problems if the number m of machines is a constant. Even under the additional restriction of unit size jobs, no better approximation result is known. In fact, it is one of the remaining four open problems from the book of Garey and Johnson [GJ79] whether P 3 | prec, pj = 1 | Cmax is even NP-hard. Also Schuurman and Woeginger [SW99] list under “Open Problem 1” the question whether there is a PTAS for this problem (recall that for m = 2, the result of [LS77] gives an optimum schedule). To understand where the lack of progress is coming from, one has to go back to the list scheduling algorithm of Graham. If we schedule the jobs in a greedy manner, then one can argue that there is always a chain of jobs j1 ≺ j2 ≺ . . . ≺ jk so that at any point in time either all m machines are P fully busy or a job from that chain was 1 processed. Since both quantities, the load m j∈J pj and the length of any chain are lower bounds on any schedule, we can conclude that the schedule has length at most 1 even for general running times, by observing 2 · OP T . One can shave off a factor of m that the processing times of the jobs in the longest chain do not need to be again counted in the load bound. Also the papers [LS77] and [GR08] effectively rely on those two lower bounds. In fact, [Cha95] showed that a large class of algorithms including the ones of [Gra66, GR08] cannot beat a bound of 2 − √2m ; moreover Graham’s algorithm 2 )-approximation for unit size jobs, see [GR08]. is indeed not better than a (2 − m Of course, one always has the option to study the strength of linear programs for an optimization problem. The most natural one for P m | prec, pj = 1 | Cmax is certainly the following time-indexed LP: For a parameter T that denotes the length of the time 2 horizon, we define a set K(T ) as the set of fractional solutions to: T X xj,t = 1 ∀j ∈ J X xj,t ≤ m ∀t ∈ [T ] X ≥ t=1 j∈J t′ <t xi,t′ X xj,t′ t′ ≤t 0 ≤ xj,t ≤ 1 (1) ∀i ≺ j ∀t ∈ [T ] ∀j ∈ J ∀t ∈ [T ] Here xj,t is a decision variable that is supposed to tell whether job j ∈ J is scheduled in time slot t ∈ [T ], where [T ] := {1, . . . , T }. The constraints guarantee that in an integral solution each job is assigned to one time slot; no time slot receives more than m jobs and for a pair of jobs i ≺ j, job i has to be scheduled before j. Unsurprisingly, this LP has a constant integrality gap as one can see from the following construction: take k blocks J1 , . . . , Jk of |Ji | = m + 1 jobs each and define the precedence order so that all the jobs in Ji have to be finished before any job in Ji+1 can be started. Any integral schedule needs two time units per block, hence OP T = 2k. On the other hand, the LP solution can schedule the m + 1 jobs of each block “in m and finish the schedule after k · m+1 parallel”, each at a rate of m+1 m time units in which each machine has always been fully busy. This results in an integrality gap of at least 2 . 2 − m+1 It has been long known, that in principle one can take the linear program for any optimization problem and strengthen it automatically by applying an LP or SDP hierarchy lift. We will provide formal definitions later, but basically these operators ensure that for any set of at most r variables, the LP solution indeed lies in the convex hull of integral combinations. Here, r is the number of levels or rounds and one typically needs time nO(r) to solve an r-level hierarchy. Some known approximation results have been reinterpreted in hindsight in this framework, for example a constant number of Lasserre rounds applied to a basic LP suffices for the Goemans-Williamson algorithm for MaxCut [GW95] and also√a constant number of Lasserre rounds implies the triangle inequalities in the O( log n)approximation algorithm by Arora, Rao and Vazirani [ARV09]. Moreover, the subspace enumeration component in the subexponential time algorithm of Arora, Barak and Steurer [ABS10] for Unique Games could be replaced with a Lasserre SDP. However, there are relatively few results where hierarchies have been genuinely useful (at least fewer than researchers have hoped for). For example Chlamtáč [Chl07] used SDP hierarchies to find better colorings in 3-colorable graphs and Raghavendra and Tan [RT12] apply them to obtain approximation algorithms for CSPs with cardinality constraints. An application to color hypergraphs can be found in [CS08]. Hierarchies also turned out to be the right approach for Sparsest Cut in bounded tree width graphs, see the paper by Chlamtáč, Krauthgamer and Raghavendra [CKR10] and the 2-approximation by Gupta, Talwar and Witmer [GTW13]. For an application of the Lasserre hierarchy 3 in the context of scheduling, see the recent work of Bansal, Srinivasan and Svensson [BSS16]. Throughout this paper, logarithms will be with respect to base 2, that means log(T ) := log2 (T ). 1.1 Our Contribution Our main result is that an LP lift with 2 /ε2 )·log log n) (log(n))O((m rounds closes the integrality gap of LP (1) to at most 1 + ε. This implies: Theorem 1. For the problem P m | prec, pj = 1 | Cmax one can compute a (1 + ε)2 2 approximate solution in time nO(r) where r := (log(n))O((m /ε )·log log n) . This gives a partial answer to one of the questions under “Open Problem 1” in [SW99] which asked whether there is a PTAS for this problem. In a Dagstuhl workshop, Mathieu [Dag10] asked the more specific question whether the Sherali-Adams hierarchy gives a (1 + ε)-approximation after c(ε, m) rounds. We also make progress on the question from the book of Garey and Johnson [GJ79] by improving the 34 -polynomial time approximation for m = 3 [LS77] to a 1 + ε in slightly more than quasi-polynomial time. In particular, this implies that P m | prec, pj = 1 | Cmax is not APX-hard, O(log log n) ). assuming that NP 6⊆ DTIME(nlog(n) 2 An Explicit LP Hierarchy for Makespan Scheduling In principle, our result can be obtained by applying the well-known Sherali-Adams hierarchy to the linear program in (1) — of course the same still holds true for even more powerful hierarchies such as the Lasserre SDP hierarchy. While this may be the preferable option for experts, we will work with an explicit strengthening of the above linear program that hopefully will be more accessible to non-experts in LP hierarchies. For a set K ⊆ Rm we denote cone(K) := k nX i=1 o λi xi | k ∈ N; xi ∈ K ∀i ∈ [k]; λi ≥ 0 ∀i ∈ [k] as the convex cone that is spanned by K. Let us fix a parameter r. Let σ : J → [T ] ∪ {∗} be a partial assignment that assigns slots only for a subset of jobs. All the jobs with σ(j) = ∗ are unassigned. Let supp(σ) := {j ∈ J | j is assigned in σ} be the support of that partial assignment. We denote ∅ as the partial assignment that assigns no job at all. Moreover for a partial assignment σ and j ∈ / supp(σ) and t ∈ [T ], let σ ∪ (j, t) be the partial assignment augmented by σ(j) = t. 4 We say that a solution to SA(K(T ), r) is a set of vectors x := {xσ }|supp(σ)|≤r , where we define x := x∅ satisfying the following program: P σ∪(j,t) xσ = ∀σ : |supp(σ)| < r and j ∈ / supp(σ) (I) t∈[T ] x  xσ ∈ cone K(T ) ∩ {x | xj,σ(j) = 1 ∀j ∈ supp(σ)} ∀σ : |supp(σ)| ≤ r (II) x ∈ K(T ) (III) In other words, x is a collection of nO(r) many vectors xσ that each has dimension |J| · T . Note that if xσ is a non-zero vector, then it can be scaled to be a fractional solution in K(T ) that has all assignments of the partial assignment σ integral. Notice that we have a variable for each σ with |supp(σ)| ≤ r, so one can find a feasible solution of the program in nO(r) time. One can think of this system as basically being the Sherali-Adams system, just that we do include more redundant variables that will make it easy to prove the needed properties. First, we claim that if there exists a valid schedule σ ∗ , then SA(K(T ), r) 6= ∅. Here we can build a valid solution by simply choosing xσ as the characteristic vector of σ ∗ if σ and σ ∗ agree. We set xσ = 0 if there is a job j ∈ supp(σ) so that σ(j) 6= σ ∗ (j). We give the following useful properties: P Lemma 2. Fix some r. Let x ∈ SA(K(T ), r). Let2 λσ := Tt=1 xσj,t . Then the following holds xσ λσ ∈ K(T ) ∩ {x | xj,σ(j) = 1 ∀j ∈ supp(σ)}.  b) If r = n, then x ∈ conv K(T ) ∩ {0, 1}J×[T ] . a) If λσ > 0, then ∗ ∗ c) Let j ∗ ∈ J and t∗ ∈ [T ] so that ρ := xj ∗ ,t∗ > 0. Then taking y σ := 1ρ · xσ∪(j ,t ) for each σ, one has y = {y σ }|supp(σ)|≤r−1 ∈ SA(K(T ), r − 1) and yj ∗,t∗ = 1. Moreover, xj,t = 0 ⇒ yj,t = 0 for all j ∈ J and t ∈ [T ]. Proof. We prove the following: a) Follows from (II) and the definition of λσ . b) We can iteratively apply (I) to obtain x= X xσ = σ:J→[T ] By a), xσ λσ X σ:J→[T ]:λσ >0 λσ · xσ . λσ are 0/1 vectors. c) From the definition we can see that (I), (II) are just inherited. (III) and yj ∗ ,t∗ = 1 follow from the scaling. The implication xj,t = 0 ⇒ yj,t = 0 follows from (j ∗ ,t∗ ) ≤ ρ1 xj,t . yj,t = 1ρ · xj,t 2 Here j ∈ J is any fixed job. But note that this definition does not depend on the choice of j. 5 If we have solution x ∈ SA(K(T ), r) and variables j ∗ , t∗ with xj ∗ ,t∗ > 0, then conditioning on xj ∗ ,t∗ = 1 means to replace the solution x with the solution y = {y σ }|supp(σ)|≤r−1 ∈ SA(K(T ), r − 1) described in Lemma 2.c. 3 An Overview In this section, we will give an overview over the different steps in our algorithm; the detailed implementation of some of the steps will be given in Section 4, Section 5 and Section 6. For a given time horizon T , a feasible schedule is an assignment σ : J → {1, . . . , T } with |σ −1 (t)| ≤ m for all t ∈ [T ] and for all j, j ′ ∈ J one has j ≺ j ′ ⇒ σ(j) < σ(j ′ ). Formally, our main technical theorem is as follows: 2 Theorem 3. For any solution x ∈ SA(K(T ), r) with r := (log n)O((m /ε can find a feasible schedule σ : J → N of the jobs in time nO(r) so that 2 )·log log n) , one max σ(j) ≤ (1 + ε) · T. j∈J To obtain a (1 + ε)-approximation, we can find the minimum value of T so that SA(K(T ), r) 6= ∅ with binary search and then compute a solution x ∈ SA(K(T ), r). In particular, by virtue of being a relaxation, that value of T will satisfy T ≤ OP T , where OP T is the makespan of the optimum schedule. For the sake of a simpler notation, we will assume that T is a power of 2 — if 2z−1 < T ≤ 2z for some integer z, then one can add m · (2z − T ) many dummy jobs that all depend on each original job so that the algorithm will schedule the dummy jobs at the very end. Moreover we will assume that 1ε , m ≤ log(n) as otherwise the bound is meaningless. The main routine of our algorithm will schedule jobs only within the time horizon T of the LP-hierarchy solution, but we will allow it to discard jobs. Formally this means, we will find an assignment σ : J \ Jdiscarded → [T ] that will not have assigned slots to jobs in Jdiscarded . Such an assignment will still be called “feasible” if apart from the load bound, the condition j ≺ j ′ ⇒ σ(j) < σ(j ′ ) is satisfied for all j, j ′ ∈ J \ Jdiscarded . In particular dependencies with discarded jobs play no role in this definition. The reason for this definition is that one can easily insert the discarded jobs at the very end of the algorithm: Lemma 4. Any feasible schedule σ : J \ Jdiscarded → {1, . . . , T } can be modified in polynomial time to a feasible schedule σ ∗ : J → {1, . . . , T + |Jdiscarded |} which also includes the previously discarded jobs. 6 Proof. Select any job j ∗ ∈ Jdiscarded . Since σ is a valid schedule which respects all precedence constraints in J \ Jdiscarded , there must be a time t∗ so that all jobs j ≺ j ∗ have σ(j) ≤ t∗ and all jobs j with j ∗ ≺ j have σ(j) > t∗ . Then we insert an extra time unit after time t∗ ; in this extra time slot, we only process j ∗ . We continue the procedure with inserting the next job from Jdiscarded \ {j ∗ }. Now, let us introduce some notation: We can imagine the precedence order “≺” as a directed transitive graph G = (J, E) with the nodes as jobs and edges (j, j ′ ) ∈ E ⇔ j ≺ j ′ . In that view, let δ+ (j) := {j ′ ∈ J | j ≺ j ′ } be the jobs depending on j and let δ− (j) := {j ′ ∈ J | j ′ ≺ j} be the jobs on which j depends. Note that δ+ (j) and δ− (j) are always distinct. We abbreviate δ(j) := δ+ (j) ∪ δ− (j) as the jobs that have any dependency with j. Finally, for a subset of jobs J ′ ⊆ J, let ∆(J ′ ) := max{|δ(j) ∩ J ′ | + 1 | j ∈ J ′ } be the maximum degree of a node in the subgraph induced by J ′ , counting also the node itself. We partition the time horizon [T ] into a balanced binary family I of intervals of lengths T, T2 , 2T2 , . . . , 2, 1. Let I := I0 ∪˙ . . . ∪˙ Ilog(T ) be the binary laminar family of intervals that we obtain by repeatedly partitioning intervals into two equally-sized subintervals. Recall that each level Iℓ contains 2ℓ many interval I ∈ Iℓ ; each one T consisting of |I| P= 2ℓ many time units. For each job j ∈ J and each interval I, we now define xj,I := t∈I xj,t , which denotes how much of job j will be scheduled somewhere within that interval I. Our algorithm will schedule the jobs in a recursive manner. The main claim is that for any interval I ∗ , LP-hierarchy solution x∗ and a set of jobs J ∗ with x∗j,I ∗ = 1 we can schedule almost all jobs from J ∗ within I ∗ while respecting all precedence constraints. We use parameters k := c1εm log log(T ) where c1 > 0 is a large enough constant . To get some intuition for the that we will choose in Section 6, and δ := 2 2kε2 8k m2 log(T ) parameters, considering ε and m as fixed constants, one would have k = Θ(log log n) and δ = 1/ log(n)Θ(log log n) . Formally, the main technical lemma is the following: Lemma 5. Fix ε > 0. Let I ∗ ∈ I be an interval from the balanced family of length T ∗ := |I ∗ |. Let x∗ ∈ SA(K(T ), r ∗ ) be an LP-hierarchy solution with 2 r ∗ ≥ log(T ∗ ) · 2mk2 · 2k /δ. Let J ∗ ⊆ {j ∈ J : x∗j,I ∗ = 1}. Then one can find a feasible assignment σ : J ∗ \ ∗ Jdiscarded → I ∗ that discards only ∗ |Jdiscarded |≤ ε log(T ∗ ) ε · · T∗ + · |J ∗ | 2 log(T ) 2m many jobs. Before we move on to explain the procedure behind Lemma 5, we want to argue that it implies our main result, Theorem 3: 7 Proof. We set I ∗ := {1, . . . , T } and x∗ := x, then J ∗ := J is a valid choice as trivially xj,{1,...,T } = 1 for any job. To satisfy the requirement of Lemma 5 we need 2 log(T ) · m 2 2mk2 · 2k ≤ (log(n))O(( ε ) log log(n)) δ k many levels of the hierarchy. Here we use that k = Θ( m ε log log T ), hence 2 = m 2 2 (log(T ))Θ(m/ε) and 2k = (2k )k = (log(T ))Θ(( ε ) log log T ) (we want to point out that many of the lower order terms are absorbed into the O-notation of the exponent and we assume that 1ε , m ≤ log(n)). Then Lemma 5 returns a valid assignment σ : J \ Jdiscarded → [T ] that discards only |Jdiscarded | ≤ ε ε log(T ) · ·T + · |J| ≤ ε · T 2 log(T ) 2m many jobs. Inserting those discarded jobs via Lemma 4 then results in a feasible schedule of makespan at most (1 + ε) · T . The rest of the manuscript will be devoted to proving Lemma 5. We fix a constant ε > 0 as the target value for our approximation ratio and denote T ∗ := |I ∗ | as the length of our interval. Let us first argue how to handle the base case, which for us is if log(T ∗ ) ≤ k2 . 2 In that case, we have at most mT ∗ ≤ m2k jobs. Hence, the LP-hierarchy lift has r ∗ ≥ mT ∗ many levels and one can repeatedly condition on events xj,t = 1 for j ∈ J ∗ and t ∈ I ∗ until one arrives at an LP hierarchy solution x∗∗ with x∗∗ j,t ∈ {0, 1} for all ∗ ∗ j ∈ J . This then represents a valid schedule of jobs J in the interval I ∗ without the need to discard any jobs. ∗ We now come to a high-level description of the algorithm. Let I0∗ , . . . , Ilog(T ∗ ) be ∗ the family of subintervals of I ∗ , where Iℓ∗ contains 2ℓ intervals of lengthPT2ℓ each, see ∗ Figure 1. For a job j ∈ J ∗ , we define ℓ(j, x∗ ) := max{ℓ : ∃I ∈ Iℓ∗ with t∈I xj,t = 1} as the level that owns the job in the current LP-hierarchy solution. We also abbreviate J(ℓ, x∗ ) := {j ∈ J ∗ | ℓ(j, x∗ ) = ℓ} as all jobs owned by level ℓ. The algorithm is as follows: • Step 1: Starting with the LP-hierarchy solution x∗ , we can iteratively condition on events until we arrive at a solution x∗∗ that has the property that for any interval I ∈ I0∗ ∪ . . . ∪ Ik∗2 −1 , the jobs owned by that interval have small dependence degree, that ∆(J(I, x∗∗ )) ≤ δ|I|, where J(I, x∗∗ ) := P means ∗ ∗∗ {j ∈ J | I minimal with t∈I xj,t = 1}. If we then consider the set of jobs ∗∗ ∗ ∗∗ J := {j ∈ J | 0 ≤ ℓ(j, x ) < k2 } owned by the first k2 levels, the longest chain in J ∗∗ will contain at most k2 δT ∗ jobs. We will show in Section 4 that the 2 number of required conditionings can be upperbounded by 2mk2 · 2k /δ, which 2 implies that x∗∗ ∈ SA(K(T ), r ∗ − 2mk2 · 2k /δ). • Step 2: From now on, we work with the modified LP-hierarchy solution x∗∗ . We select a level index ℓ∗ ∈ {k, . . . , k2 } and partition the jobs in J ∗ in three different groups: 8 – The jobs on the top levels: Jtop := J(0, x∗∗ ) ∪ . . . ∪ J(ℓ∗ − k − 1, x∗∗ ) – The jobs on the k middle levels: Jmiddle := J(ℓ∗ − k, x∗∗ ) ∪ . . . ∪ J(ℓ∗ − 1, x∗∗ ) – The jobs on the bottom levels: Jbottom := J(ℓ∗ , x∗∗ ) ∪ . . . ∪ J(log(T ∗ ), x∗∗ ) Then we discard all jobs in Jmiddle . In Section 6 we will describe how the index ℓ∗ is chosen and in particular we will provide an upper bound on the number of discarded middle jobs. • Step 3: In this step, we will find a schedule for the bottom jobs. For this purpose, we call Lemma 5 recursively for each interval I ∈ Iℓ∗ with a copy of the solution x∗∗ and jobs JI := {j ∈ Jbottom | x∗∗ j,I = 1}. Here it is crucial that the intervals are disjoint but also the sets JI are disjoint for different intervals I ∈ Iℓ∗ . Then Lemma 5 returns a valid schedule of theSform σI : JI \ JI,discarded → I for each interval I ∈ Iℓ∗ . Let Jbottom-discarded := I∈Iℓ∗ JI,discarded ⊆ Jbottom be the union of jobs that were discarded in those calls. The partial schedules σI satisfy / I. We combine those schedules to |σI−1 (t)| ≤ m for t ∈ I and |σI−1 (t)| = 0 for t ∈ a schedule σ : Jbottom /Jbottom-discarded → I ∗ . From the disjointness of the intervals, it is clear that again |σ −1 (t)| ≤ m for all t ∈ I ∗ . Moreover, if j ≺ j ′ and j, j ′ ∈ JI for some interval I ∈ Iℓ∗ , then by the inductive hypothesis σ(j) < σ(j ′ ). On the other hand, if j ∈ JI and j ′ ∈ JI ′ then ∗∗ we know by Lemma 2.c that I had to come before I ′ since x∗∗ j,I = 1 = xj ′ ,I ′ . • Step 4: We continue working with the previously constructed schedule σ that schedules the non-discarded bottom jobs. In this step, we will extend the schedule σ and insert the jobs of Jtop in the remaining free slots. We will prove in Section 5 that this can be done without changing the position of any scheduled bottom job and without violating any precedence constraints. Again, we allow that the procedure discards a small number of additional jobs from Jtop that we will account for later. Eventually, the schedule σ satisfies the claim for Lemma 5. The intuition behind the algorithm is as follows: When we call the procedure recursively for intervals I ∈ Iℓ∗∗ we cannot control where the jobs JI will be scheduled within that interval I. In particular the decisions made in different intervals I, I ′ ∈ Iℓ∗∗ will in general not be consistent. But the discarding of the middle jobs creates a gap between the top jobs and the bottom jobs in the sense that the intervals of the top jobs are at least a factor 2k longer than intervals of the bottom jobs. For a top job j ∈ Jtop we will be pessimistic and assume that all the bottom jobs that j depends on will be scheduled just at the very end of their interval. Still, as those intervals are very short, we will be able to argue that the loss in the flexibility is limited and most of the top jobs can be processed. As a second crucial ingredient, the conditioning had the implication that the top jobs do not contain any long chains any more. This will imply that a greedy schedule of the top jobs will leave little idle time, resulting in only few discarded top jobs. 9 Jtop Jmiddle Jbottom I0∗ .. . Iℓ∗∗ −k−1 Iℓ∗∗ −k .. . Iℓ∗∗ −1 .. Iℓ∗∗ .. . Ik∗2 −1 Ik∗2 J ∗∗ .. 1 2 ... T∗ time Figure 1: Binary dissection of the interval I ∗ used in the algorithm behind Lemma 5. A high-level pseudo-code description of the whole scheduling algorithm can be found in Figure 2: 4 Step (1) — Reducing Dependence In this section we will implement “Step (1)” and show that we can reduce the maximum dependence degrees of the jobs owned by the first k2 levels in order to bound the length 2 of chains. We will do this by conditioning on up to 2mk2 · 2k /δ many variables. We are considering an interval I ∗ and a subset of jobs J ∗ ⊆ J that the vector x∗ from the current LP-hierarchy solution x∗ fully schedules within I ∗ . Recall that for one of the subintervals I ∈ Iℓ∗ below I ∗ , we write J(I, x∗ ) = {j ∈ J(ℓ, x∗ ) | x∗j,I = 1} as the jobs owned by that particular interval. Lemma 6. Let x∗ ∈ SA(K(T ), r ∗ ). Then one can find an induced solution x∗∗ ∈ 2 SA(K(T ), r ∗∗ ) with r ∗∗ := r ∗ − 2mk 2 · 2k /δ so that ∆(J(I, x∗∗ )) ≤ δ · |I| for all intervals I ∈ I0∗ ∪ . . . ∪ Ik∗2 −1 . ˙ 2 ∈ I0 ∪ . . . ∪ Ik2 −1 Proof. We set initially x∗∗ := x∗ . If there is any interval I = I1 ∪I ∗∗ with ∆(J(I, x )) > δ · |I|, then we must have a job j ∈ J(I, x∗∗ ) that has either + − + δ δ |δJ(I,x ∗∗ ) (j) ∪ {j}| ≥ 2 · |I| or |δJ(I,x∗∗ ) (j) ∪ {j}| ≥ 2 · |I|. We assume that |δJ(I,x∗∗ ) (j) ∪ {j}| ≥ 2δ · |I| holds and omit the other case, which is symmetric. Then we pick a time ∗∗ by the LP-hierarchy solution conditioned on the t ∈ I2 with x∗∗ j,t > 0 and replace x + ∗∗ event “xj,t = 1”. Note that this means that all jobs in δJ(I,x ∗∗ ) (j) ∪ {j} will be removed ∗∗ ′ from J(I, x ). In fact, each such job will be moved to J(I , x∗∗ ) where I ′ ⊆ I2 is some 10 QPTAS for Makespan Scheduling Input: Scheduling instance (J, ≺), parameters m ∈ N and ε > 0 Output: (1 + ε)-approximate schedule σ (1) Compute a solution x = (xσ )|supp(σ)|≤r 2 2 (log(n))O(m /ε )·log log(n) and T minimal (2) Call RecursiveScheduling(J, x, [T ]) → σ (3) Insert discarded jobs into schedule σ ∈ SA(K(T ), r) with r := RecursiveScheduling P Input: Jobs J ∗ , LP lift x∗ , interval I ∗ with t∈I ∗ x∗j,t = 1 for j ∈ J ∗ Output: Schedule σ with some jobs discarded Build binary family of intervals I ∗ Call Breaking Long Chains → x∗∗ Select partition into top, middle, bottom jobs. Pick ℓ∗ . Discard middle jobs. For each I ∈ Iℓ∗ set JI := {j ∈ J ∗ | x∗∗ j,I = 1} and ∗∗ call RecursiveSchedule(JI , x , I) → σI (6) Combine σI ’s to one schedule σ (7) Call two-phased algorithm based on matching and EDF to insert top jobs into σ (1) (2) (3) (4) (5) Figure 2: High-level description of main algorithm. 11 subinterval. The conditioning can also change the owning interval of other jobs, but for each job j, the set of times t such that x∗∗ j,t > 0 can only shrink if we condition on any event, see Lemma 2.c. Hence jobs only move from intervals to subintervals. ∗ Since in each iteration, at least 2δ · |I| ≥ 2δ · Tk2 many jobs “move” and each job moves 2 at most k2 many times out of an interval in I0∗ ∪ . . . ∪ Ik∗2 −1 , we need to condition at most k2 2mT ∗ · k2 2 2 = 2mk · ∗ δ δ Tk2 2 many times. The implication of Lemma 6 is that the set of jobs owned by intervals I ∈ I0∗ ∪ . . . ∪ will not contain long chains, simply because we have only few intervals and none of jobs owned by a single interval contain long chains anymore. Ik∗2 −1 Lemma 7. After applying Lemma 6, the longest chain within jobs owned by intervals I ∈ I0∗ ∪ . . . ∪ Ik∗2 −1 has length at most k2 δT ∗ . Proof. First, let us argue how many jobs a chain can have that are all assigned to ∗ intervals of the same level ℓ. From each interval I, the chain can only include δ|I| = δ· T2ℓ many jobs. Since |Iℓ | = 2ℓ , the total number of jobs from level ℓ is bounded by δT ∗ . The claim follows from the pigeonhole principle and the fact that we have k 2 many levels in I0∗ ∪ . . . ∪ Ik∗2 −1 . We can summarize the algorithm from Lemma 6 as follows: Breaking long chains Input: Scheduling instance with jobs J ∗ , a precedence order, an LP-hierarchy solution x∗ ∈ SA(K(T ), r ∗ ), an interval I ∗ Output: An LP-hierarchy solution x∗∗ with maximum chain length k2 δT ∗ in Sk2 −1 ∗∗ ℓ=0 J(ℓ, x ) (1) Make a copy x∗∗ := x∗ Sk2 −1 ∗ ˙ 2 ) ∈ ℓ=0 (2) WHILE ∃I = (I1 ∪I Iℓ WITH ∆(J(I, x∗∗ )) > δ|I| DO (3) Choose a job j ∈ J(I, x∗∗ ) with |δJ(I,x∗∗ ) (j) ∪ {j}| ≥ δ|I|. + δ ∗∗ on x∗∗ = 1 for some (4) If |δJ(I,x ∗∗ ) (j) ∪ {j}| ≥ 2 · |I| THEN condition in x j,t t ∈ I2 ELSE condition on x∗∗ j,t = 1 for some t ∈ I1 . Note that after each conditioning in step (6), the solution x∗∗ will change and the set J(I, x∗∗ ) will be updated. 12 5 Step (4) — Scheduling Top Jobs Consider the algorithm from Section 3 and the state at the end of Step 3. At this point, we have a schedule σ that schedules most of the bottom jobs. The main argument that remains to be shown is how to add in the top jobs which are owned by intervals in I0∗ ∪ . . . ∪ Iℓ∗∗ −k−1 . This is done in two steps. First, we use a matching-based argument to show that most top jobs can be inserted in the existing schedule so that the precedence constraints with the bottom jobs are respected. In this step, we will be discarding up to 4m·2−k ·T ∗ many jobs. More crucially, the schedule will not have satisfied precedence constraints within Jtop . In a 2nd step, we temporarily remove the top jobs from the schedule and reinsert them with a variant of the Earliest Deadline First (EDF) scheduling. As we ε ∗ will see later in Theorem 9, this results in at most 8 log T · T additionally discarded jobs. 5.1 A Preliminary Assignment of Top Jobs Let us recall what we did so far. In Step 3, we applied Lemma 5 recursively on each interval I ∈ Iℓ∗ to schedule the bottom jobs. We already argued that the resulting schedules could be combined to a schedule σ : Jbottom /Jbottom-discarded → I ∗ that respects all precedence constraints. Let the intervals in Iℓ∗∗ be called I1 , . . . , Ip , so that the time horizon T ∗ is partitioned ∗ into p equally sized subintervals with p = 2ℓ . After reindexing the time horizon, let us assume for the sake of a simpler notation that I ∗ = {1, . . . , T ∗ }. If we abbreviate ∗ ti := i· Tp for i ∈ {0, . . . , p}, then the ith interval contains the time periods Ii := {ti−1 + 1, . . . , ti }. Each time t has an available capacity of cap(t) = m − |σ −1 (t)| ∈ {0, . . . , m} many machines, which Pis the number of machines not used by jobs in Jbottom . We abbreviate cap(Ii ) := t∈Ii cap(t) as the capacity of interval Ii . The available positions of jobs in Jtop are constrained by the scheduled times of jobs in Jbottom . As we had no further control over the exact position of the bottom jobs within their intervals Ii , we want to define for each job j ∈ Jtop a release time rj and a deadline dj determined by the most pessimistic outcome of how σ could have scheduled the bottom jobs. For all j ∈ Jtop , we define  rj := min ti + 1 | σ(j ′ ) ≤ ti ∀j ′ ∈ Jbottom : j ′ ≺ j (2)  ′ ′ ′ dj := max ti | σ(j ) ≥ ti + 1 ∀j ∈ Jbottom : j ≺ j In particular, the release time will be the first time unit of an interval Ii and the deadline will be the last time unit of an interval Ii . Let ir (j) and id (j) be the corresponding indices, so that the release time is of the form rj = tir (j)−1 + 1 and the deadline is dj = tid (j) . Then our goal is to show that most top jobs j can be scheduled somewhere in the time frame Iir (j) ∪ . . . ∪ Iid (j) . This would imply that at least all precedence constraints between bottom and top jobs are going to be satisfied. Notice here that the existing fractional assignment that x∗∗ provides for j ∈ Jtop might also be using the slots in the two intervals coming right before and right after the 13 rj dj Iir (j) I1 Iid (j) tir (j) Ip tid (j) tp = T∗ ˙ p and possible release times and Figure 3: Visualization of interval I ∗ = I1 ∪˙ . . . ∪I deadlines for a job j ∈ Jtop . Note that x∗∗ might schedule j over the whole hatched area, while our choice of rj and dj will force us to process j inside the black-hatched area (or to discard the job). t0 1 2 t1 range [rj , dj ] (see Figure 3). This is due to our rounding of release times and deadlines ′ to interval beginnings and ends. Let Jbottom := Jbottom \ Jbottom-discarded be the bottom jobs scheduled by the recursive calls of the algorithm. ′ → I ∗ of bottom jobs can be extended to a Lemma 8. A valid schedule σ : Jbottom ′ ′ ′ ⊆ Jtop that includes most of the top jobs. → I ∗ with Jtop ∪ Jtop schedule σ : Jbottom −1 ′ and The schedule satisfies (i) |σ (t)| ≤ m for t ∈ I ∗ ; (ii) rj ≤ σj ≤ dj for all j ∈ Jtop ′ −k ∗ (iii) one discards at most |Jtop \ Jtop | ≤ 4m · 2 · T many top jobs. Proof. We want to use a matching-based argument. For this sake, we consider the bipartite graph with jobs on one side and subintervals on the other. Formally, we define a graph G = (V, U, E + ) with V = Jtop , U = {1, . . . , p} where the nodes i ∈ U have capacity cap(Ii ), and edges E + = {(j, i) ∈ V × U | max{ir (j) − 1, 1} ≤ i ≤ min{id (j) + 1, p}}. We say that a matching M is V -perfect if it covers every node in V . Then the neighborhood of each top job includes every interval in which it is fractionally scheduled in x∗∗ . ′ has been assigned by σ to precisely that Moreover, each of the bottom jobs j ∈ Jbottom ∗∗ ∗∗ interval Ii with xj,Ii = 1. Hence x gives a V -perfect fractional matching that respects the given capacities cap(Ii ). In bipartite graphs, the existence of a fractional V -perfect matching implies the existence of an integral V -perfect matching, see e.g. [Sch03]. However, in order to assign the top jobs to slots within release times and deadlines we are only allowed to use the smaller set of edges E = {(j, i) ∈ V × U | ir(j) ≤ i ≤ idj }. For any J ∗ ⊆ V , we let N + (J ∗ ) be the neighborhood of J ∗ along edges in E + and let N (J ∗ ) be its neighborhood along edges in E. Let the magnitude of a neighborhood |N (J ∗ )| be defined as the sum of capacities of the nodes it contains. By Hall’s Theorem [Sch03], the minimum number of exposed V -nodes in a maximum matching in E 14 is precisely max{|J| − |N (J)|}. J⊆V J∗ Now, fix the set ⊆ V attaining the maximum; then |J ∗ | − |N (J ∗ )| is the number of jobs that we have to discard. Since E + allows for a V -perfect matching, the reverse direction of Hall’s Theorem gives that |J ∗ | ≤ |N + (J ∗ )|. Thus |J ∗ | − |N (J ∗ )| ≤ |N + (J ∗ )| − |N (J ∗ )|. Note that N (J ∗ ) is in general not a consecutive interval of ∗ {1, . . . , p}. We can upper bound the difference |N + (J ∗ )| − |N (J ∗ )| by 2m · Tp times the number of connected components of N (J ∗ ). This is the point where we take advantage of the “gap” between the levels of the top and bottom jobs: for each job j ∈ V there is an interval I ∈ I0∗ ∪˙ . . . ∪˙ Iℓ∗∗ −k−1 so that N + (j) contains the midpoint of that interval. Due to the gap, there are at most p · 2 · 2−k such intervals3 . Hence the number of discarded jobs can be bounded by |J ∗ | − |N (J ∗ )| ≤ 2m · T∗ · 2p · 2−k = 4m · 2−k · T ∗ . p Finally note that a corresponding maximum matching can be computed in polynomial time. 5.2 Reassigning the Top Jobs via EDF We have seen so far that we can schedule most of the bottom and top jobs so that all precedence constraints within the bottom jobs are satisfied and the top jobs are correctly scheduled between the bottom jobs that they depend on. However, the schedule as it is now ignores the precedence constraints within the top jobs. In this section, we will remove the top jobs from the schedule and then reinsert them using a variant of the Earliest Deadline First (EDF) scheduling policy. For the remainder of Section 5.2, we will show a stand-alone theorem that we will use as a black box. Imagine a general setting where we have m identical machines and n jobs J, each one with integer release times rj and deadlines dj and a unit processing time. The EDF scheduling rule picks at any time the available job with minimal dj for processing. It is a classical result in scheduling theory by Dertouzos [Der74] that for m = 1 and unit size jobs, EDF is an optimal policy. Here “optimal” means that if there is any schedule that finishes all jobs before their deadline, then EDF does so, too. The result extends to the case of arbitrary running times pj if one allows preemption. Now, our setting is a little bit different. For each time t, we have a certain number cap(t) ∈ {0, . . . , m} of slots. Additionally, we have a precedence order that we need to respect. But we can use to our advantage that the precedence order has only short chains; moreover, the number of different release times and deadlines is limited. Formally we will prove the following: It is possible that N (j) = ∅. Still N + (j) will contain a midpoint of a level 0, . . . , ℓ∗ − k − 1 interval, hence we have accounted for those jobs as well. Note that such jobs would automatically get discarded. 3 15 Theorem 9. Let J be a set of jobs with release times rj , deadlines dj and consistent precedence constraints4 with maximum chain length C. Suppose that {1, . . . , T } is the time horizon, partitioned into equally sized intervals I1 , . . . , Ip and all release times/deadlines correspond to beginnings and ends of those intervals. Let cap : [T ] → {0, . . . , m} be a capacity function and assume that there exists a schedule σ̃ : J → [T ] assigning each job to slots between its release time and deadline that respects capacities but does not necessarily respect precedence constraints within J. Then in polynomial time, one can find a schedule σ : J \ Jdiscarded → [T ] that respects capacities, release times, deadlines and precedence constraints and discards |Jdiscarded | ≤ p2 mC many jobs. We use the following algorithm, which is a variant of Earliest Deadline First, where we discard those jobs that we cannot process in time: Earliest Deadline First Input: Jobs J with deadlines, release times, precedence constraints; capacity function cap : [T ] → {0, . . . , m} Output: Schedule σ : J → [T ] ∪ {DISCARDED} (1) Set σ(j) := UNASSIGNED for all j ∈ J and Jdiscarded := ∅ (2) Sort the jobs J = {1, . . . , n} so that d1 ≤ d2 ≤ . . . ≤ dn (3) FOR t = 1 TO T DO (4) FOR cap(t) MANY TIMES DO (5) Select the lowest index j of a job with rj ≤ t ≤ dj that has not been scheduled or discarded and that has all jobs in δ− (j) already processed (or discarded). (6) Set σ(j) := t (if there was any such job) (7) FOR each j ∈ J with dj = t and σ(j) = UNASSIGNED, add j to Jdiscarded and set σ(j) := DISCARDED At the end all jobs j will be either scheduled between rj and dj (that means rj ≤ σ(j) ≤ dj ) or they are in Jdiscarded . We will say that a job j was discarded in the interval [t, t′ ] if j ∈ Jdiscarded and t ≤ dj ≤ t′ . We call a time t busy if |σ −1 (t)| = cap(t) and non-busy otherwise. Let us make a useful observation: Lemma 10. Let I = {t′ , . . . , t′′ } ⊆ Ii be part of one of the subintervals. Suppose that there is a non-busy time t∗ ∈ I and a job j with I ⊆ {rj , . . . , dj } and σ(j) ∈ {DISCARDED} ∪ {t′′ + 1, . . . , T }. Then there is a job j ∗ ∈ σ −1 (t∗ ) with j ∗ ≺ j. Proof. Consider any inclusion-wise maximal chain of jobs j1 ≺ j2 ≺ . . . ≺ jq that ends in j = jq and otherwise has only jobs j1 , . . . , jq−1 ∈ σ −1 ({t∗ , . . . , t′′ }). First suppose that q > 1 and hence j1 6= j. It is impossible that σ(j1 ) > t∗ because by assumption rj1 ≤ t∗ 4 Here “consistent” means that for a pair of dependent jobs j ≺ j ′ one has rj ≤ rj ′ and dj ≤ dj ′ . 16 and hence EDF would have processed j1 already earlier at time t∗ (by maximality of the chain, there is no job scheduled at times t∗ , . . . , σ(j1 ) − 1 on which j1 might depend). Hence σ(j1 ) = t∗ and by transitivity j1 ≺ j, which gives the claim. In the 2nd case, we have j1 = j, hence there is no job that j depends on scheduled between t∗ and t′′ . But we know that rj ≤ t∗ . Thus EDF would have processed j at time t∗ or earlier. With this observation we can easily limit the number of non-busy times per interval: Lemma 11. Let I = {t′ , . . . , t′′ } ⊆ Ii be part of one of the subintervals. Suppose that there is at least one job j with I ⊆ {rj , . . . , dj } and σ(j) ∈ {DISCARDED}∪{t′′ +1, . . . , T }. Then the number of non-busy times in I is bounded by C. Proof. By Lemma 10, for the latest time t∗ ∈ I with |σ −1 (t∗ )| < cap(t∗ ), there is at least one job j ∗ ∈ σ −1 (t∗ ) with j ∗ ≺ j. Then we can continue by induction, replacing t′′ with t∗ − 1 and replacing j by j ∗ to build a chain of jobs ending with j that includes a job scheduled at each non-busy time. Since no chains can be longer than C, this gives the claim. Now we come to the main argument where we give an upper bound of the number of discarded jobs: Lemma 12. One has |Jdiscarded | ≤ p2 mC. Proof. Suppose that |Jdiscarded | ≥ p · K; we will then derive a bound on K. By the pigeonhole principle, we can find an interval Ib so that at least K many jobs get discarded in Ib . Let us denote the lowest priority (i.e. the latest deadline) job that gets discarded in Ib by js . Now delete all those lower priority jobs js+1 , . . . , jn . Note that this does not affect how EDF schedules j1 , . . . , js and still we would have at least K jobs discarded in Ib , including js . By Lemma 11, the number of non-busy periods in Ib is bounded by C. Now, choose a minimal index a ∈ {1, . . . , b − 1} so that in each of the intervals Ia , . . . , Ib one has at most C many non-busy periods. We abbreviate I ′ := Ia ∪ . . . ∪ Ib . Note that by definition Ia−1 has more than C many non-busy periods5 . Define n o J ′ := j ∈ {j1 , . . . , js } : (σ(j) ∈ I ′ ) or (j discarded and dj ∈ I ′ ) . By Lemma 10, any job in J ′ has its release time in Ia or later, since otherwise we could not have C + 1 non-busy times in Ia−1 . Now, let us double count the number of jobs in J ′ . On the one hand, we have ′ |J | ≥ b X i=a |σ −1 b X (cap(Ii ) − mC) + |J ′ ∩ Jdiscarded | (Ii )| + |J ∩ Jdiscarded | ≥ {z } | ′ i=a ′ ≥ cap(I ) − pmC + K. ≥K 5 Admittedly it is possible that a = 1 in which case one might imagine I0 as an interval in which all times are non-busy and which does not contain any release times. 17 On the other hand, we know that there is an assignment σ̃ of jobs in J ′ to slots in Ia , . . . , Ib . That tells us that |J ′ | ≤ cap(I ′ ). Comparing both bounds gives that K ≤ pmC. 6 Step (2) — Accounting for Discarded Jobs In this section, we need to argue that the level ℓ∗ can be chosen so that the total number of jobs that are discarded in Steps (1)-(4) are bounded by ∗ |Jdiscarded |≤ ε log(|I ∗ |) ε · · |I ∗ | + · |J ∗ | 2 log(T ) 2m as claimed. Let us summarize the three occasions in the algorithm where a job might get discarded: ∗ (A) In Step (3), in order to schedule the bottom jobs, we have 2ℓ many recursive calls of Lemma 5 for intervals I ∈ Iℓ∗∗ . The cumulative number of discarded jobs from all those calls is bounded by ∗ T ε log( 2ℓ∗ ) T ∗ ε ε log(T ∗ ) − ℓ∗ ε 2 · · · ℓ∗ + · |Jbottom | = · · T∗ + · |Jbottom | 2 log(T ) 2 2m 2 log(T ) 2m ℓ∗ (B) As we have seen in Section 5, the number of top jobs that need to be discarded in Step (4) can be bounded by 2 4m · 2−k · T ∗ + p2 mC ≤ 4m · 2−k · T ∗ + k2 m22k T ∗ δ ≤ ε 1 · · T ∗. 4 log(T ) Here we use that the length of the maximum chain within the top jobs is C ≤ 2 ∗ k2 δT ∗ . Moreover, we have substituted the parameters p = 2ℓ ≤ 2k as well as and k = c1 m δ = 2 2kε2 ε log log(T ) with a large enough constant c1 > 0. 8k m2 log(T ) (C) In Step (2), we discard all the middle jobs. In the remainder of this section we prove that there is an index ℓ∗ so that |Jmiddle | ≤ ε 1 ε · · T∗ + · (|Jmiddle | + |Jtop |) 4 log(T ) 2m Let us first assume that we can indeed find a proper index ℓ∗ so that the bound in 18 (C) is justified. Then the total number of jobs that the algorithm discards is (A) (B) }| { z z }| { ε ε log(T ∗ ) − ℓ∗ ε 1 ∗ · ·T + · |Jbottom | + · · T∗ 2 log(T ) 2m 4 log(T ) (C) }| { z ε 1 ε ∗ + · ·T + · (|Jmiddle | + |Jtop |) 4 log(T ) 2m ε ε log(T ∗ ) · · T∗ + · (|Jtop | + |Jmiddle | + |Jbottom |) ≤ 2 log(T ) 2m | {z } =|J ∗ | which is the bound that we claimed in Lemma 5. It remains to justify the claim in (C). We abbreviate αi := |J(i · k, x∗∗ ) ∪ . . . ∪ J((i + 1) · k − 1, x∗∗ )| for i ∈ {0, . . . , k − 1}. In words, each number αi represents the number of jobs owned by k consecutive levels. We observe that if there is an index i ∈ {0, . . . , k − 1} so that ε · T∗ (I) αi ≤ 4 log(T ) or i ε X αj (II) αi ≤ · 2m j=1 then we can choose ℓ∗ := (i + 1) · k and (C) will be satisfied. Here we use that for this particular choice of ℓ∗ , we will have |Jmiddle | = αi and |Jtop | = α0 + . . . + αi−1 . So, we assume for the sake of contradiction that no index i satisfies either (I) or (II) (or both). Then one can easily show that the αi ’s have to grow exponentially. We show this in a small lemma: Lemma 13. Let q ∈ N and suppose we P have a sequence of numbers α0 , α1 , . . . , αN 1 satisfying αi ≥ αmin > 0 and αi ≥ q · ij=1 αj for all i = 0, . . . , N . Then αi ≥ 2⌊i/(2q)⌋ · αmin . Proof. Group the indices into consecutive blocks of 2q numbers, where α0 , . . . , α2q−1 is block 0, α2q , . . . , α4q−1 is block 1 and so on. We want to prove by induction that each αi in the jth block is at least 2j · αmin . For j = 0, the claim follows from the assumption. For j > 0, we use that αi is at least the sum of 2q numbers that by inductive hypothesis j−1 are all at least 2 q · αmin . The claim follows. 2m ε ∗ Applying Lemma 13 with αmin := 4 log(T ) · T and q := ε we obtain in particular that ε ε · T∗ αk−1 ≥ 2⌊(k−1) 4m ⌋ · 4 log(T ) If we set k = c1εm log(log(T )) for some adequately large c1 , then αk−1 > mT ∗ , which is a contradiction since we only have |J ∗ | ≤ m · |I ∗ | = mT ∗ many jobs with xj,I ∗ = 1. 19 7 Conclusion For the proof of Lemma 5 we already argued that the number of discarded jobs satisfies the claimed bound and that all precedence constraints will be satisfied. Regarding the number of rounds in the hierarchy, recall that we started with a solution x∗ ∈ 2 SA(K(T ), r ∗ ) with r ∗ ≥ log(T ∗ ) · 2mk2 · 2k /δ. Then we apply a round of conditionings 2 in Lemma 6 to obtain x∗∗ ∈ SA(K(T ), r ∗∗ ) with r ∗∗ := r ∗ − 2mk2 · 2k /δ. We use copies of the solution x∗∗ in our recursive application of Lemma 5 to intervals of size ∗ T ∗ /2ℓ . Since ℓ∗ ≥ k ≥ 1, the remaining number of LP-hierarchy rounds satisfies 2 ∗ r ∗∗ ≥ log(T ∗ /2ℓ ) · 2mk2 · 2k /δ. Thus we still have enough LP-hierarchy rounds for the recursion. Another remark concerns why we may assume that the precedence constraints are consistent in Theorem 9. Suppose we have jobs j, j ′ ∈ Jtop with j ≺ j ′ and consider the definition of release times and deadlines in Eq. 2. By transitivity, any job j ′′ ∈ Jbottom with j ′ ≺ j ′′ which limits the deadline of j ′ will also limit the deadline of j, hence dj ≤ dj ′ . Similarly one can argue that rj ≤ rj ′ . This concludes the proof of Lemma 5 and our main result follows. 8 Follow-up work and open problems A natural question that arises is whether the number of O(log n)O(log log n) rounds for constant ε, m can be improved. In fact, after the conference version of this paper appeared, Garg [Gar17] was able to reduce the number of rounds down to O(log n)O(1) , hence providing an actual QPTAS. It remains open whether c(m, ε) many rounds suffice as well. Another tantalizing question is whether a similar approach could give a (1 + ε)-approximation for P m | prec | Cmax , where the processing times pj ∈ N are arbitrary. Note that the difficulty in this setting comes from the issue that jobs have to be scheduled non-preemptively. References [ABS10] S. Arora, B. Barak, and D. Steurer. Subexponential algorithms for unique games and related problems. In 51th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2010, October 23-26, 2010, Las Vegas, Nevada, USA, pages 563–572, 2010. [ARV09] S. Arora, S. Rao, and U. V. Vazirani. Expander flows, geometric embeddings and graph partitioning. J. ACM, 56(2), 2009. [BK09] N. Bansal and S. Khot. Optimal long code test with one free bit. In Foundations of Computer Science, 2009. FOCS ’09. 50th Annual IEEE Symposium on, pages 453–462, Oct 2009. 20 [BSS16] N. Bansal, S. Srinivasan, and O. Svensson. Lift-and-round to improve weighted completion time on unrelated machines. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages 156–167, 2016. [Cha95] M. Charikar. Approximation algorithms for problems in combinatorial optimization. Technical report, B. Tech. Project Report, Department of Computer Sc. and Engg., IIT Bombay, 1995. [Chl07] E. Chlamtac. Approximation algorithms using hierarchies of semidefinite programming relaxations. In 48th Annual IEEE Symposium on Foundations of Computer Science (FOCS 2007), October 20-23, 2007, Providence, RI, USA, Proceedings, pages 691–701, 2007. [CKR10] E. Chlamtac, R. Krauthgamer, and P. Raghavendra. Approximating sparsest cut in graphs of bounded treewidth. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, 13th International Workshop, APPROX 2010, and 14th International Workshop, RANDOM 2010, Barcelona, Spain, September 1-3, 2010. Proceedings, pages 124–137, 2010. [CS08] E. Chlamtac and G. Singh. Improved approximation guarantees through higher levels of SDP hierarchies. In Approximation, Randomization and Combinatorial Optimization. Algorithms and Techniques, 11th International Workshop, APPROX 2008, and 12th International Workshop, RANDOM 2008, Boston, MA, USA, August 25-27, 2008. Proceedings, pages 49–62, 2008. [Dag10] Open problems – scheduling. In Susanne Albers, Sanjoy K. Baruah, Rolf H. Möhring, and Kirk Pruhs, editors, Scheduling, number 10071 in Dagstuhl Seminar Proceedings, Dagstuhl, Germany, 2010. Schloss Dagstuhl - LeibnizZentrum fuer Informatik, Germany. [Der74] M. L. Dertouzos. Control robotics: The procedural control of physical processes. In IFIP Congress, pages 807–813, 1974. [Gar17] S. Garg. Quasi-ptas for scheduling with precedences using LP hierarchies. CoRR, abs/1708.04369, 2017. [GJ79] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman and Company, New York, New York, 1979. [GR08] D. Gangal and A. Ranade. Precedence constrained scheduling in optimal. Journal of Computer and System Sciences, 74(7):1139 – 1146, 2008. [Gra66] R. L. Graham. Bounds for certain multiprocessing anomalies. Bell System Technical Journal, 45(9):1563–1581, 1966. 21 [GTW13] A. Gupta, K. Talwar, and D. Witmer. Sparsest cut on bounded treewidth graphs: algorithms and hardness results. In Symposium on Theory of Computing Conference, STOC’13, Palo Alto, CA, USA, June 1-4, 2013, pages 281–290, 2013. [GW95] M. X. Goemans and D. P. Williamson. Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming. J. ACM, 42(6):1115–1145, 1995. [LK78] J. K. Lenstra and A. H. G. Rinnooy Kan. Complexity of scheduling under precedence constraints. Operations Research, 26(1):22–35, 1978. [LS77] S. Lam and R. Sethi. Worst case analysis of two scheduling algorithms. SIAM J. Comput., 6(3):518–536, 1977. [RT12] P. Raghavendra and N. Tan. Approximating csps with global cardinality constraints using SDP hierarchies. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2012, Kyoto, Japan, January 17-19, 2012, pages 373–387, 2012. [Sch03] A. Schrijver. Combinatorial Optimization - Polyhedra and Efficiency. Springer, 2003. [Sve10] O. Svensson. Conditional hardness of precedence constrained scheduling on identical machines. In Proceedings of the Forty-second ACM Symposium on Theory of Computing, STOC ’10, pages 745–754, New York, NY, USA, 2010. ACM. [SW99] P. Schuurman and G. J. Woeginger. Polynomial time approximation algorithms for machine scheduling: ten open problems. Journal of Scheduling, 2(5):203–213, 1999. 22
8cs.DS
Noname manuscript No. (will be inserted by the editor) Adapting general-purpose speech recognition engine output for domain-specific natural language question answering C. Anantaram · Sunil Kumar Kopparapu arXiv:1710.06923v1 [cs.CL] 12 Oct 2017 the date of receipt and acceptance should be inserted later Abstract Speech-based natural language question-answering interfaces to enter- prise systems are gaining a lot of attention. General-purpose speech engines can be integrated with NLP systems to provide such interfaces. Usually, general-purpose speech engines are trained on large ‘general’ corpus. However, when such engines are used for specific domains, they may not recognize domain-specific words well, and may produce erroneous output. Further, the accent and the environmental conditions in which the speaker speaks a sentence may induce the speech engine to inaccurately recognize certain words. The subsequent natural language question-answering does not produce the requisite results as the question does not accurately represent what the speaker intended. Thus, the speech engine’s output may need to be adapted for a domain before further natural language processing is carried out. We present two mechanisms for such an adaptation, one based on evolutionary development and the other based on machine learning, and show how we can repair the speech-output to make the subsequent natural language question-answering better. 1 Introduction Speech-enabled natural-language question-answering interfaces to enterprise application systems, such as Incident-logging systems, Customer-support systems, Marketing-opportunities systems, Sales data systems etc., are designed to allow end-users to speak-out the problems/questions that they encounter and get automatic responses. The process of converting human spoken speech into text is performed by an Automatic Speech Recognition (ASR) engine. While functional examples of ASR with enterprise systems can be seen in day-to-day use, most of C. Anantaram TCS Innovation Labs - Delhi, ASF Insignia, Gwal Pahari, Gurgaon, India E-mail: [email protected] Sunil Kumar Kopparapu TCS Innovation Labs [email protected] Mumbai, Yantra Park, Thane (West) E-mail: sunilku- 2 Anantaram, Sunil these work under constraints of a limited domain, and/or use of additional domainspecific cues to enhance the speech-to-text conversion process. Prior speech-andnatural language interfaces for such purposes have been rather restricted to either Interactive Voice Recognition (IVR) technology, or have focused on building a very specialized speech engine with domain specific terminology that recognizes key-words in that domain through an extensively customized language model, and trigger specific tasks in the enterprise application system. This makes the interface extremely specialized, rather cumbersome and non-adaptable for other domains. Further, every time a new enterprise application requires a speech and natural language interface, one has to redevelop the entire interface again. An alternative to domain-specific speech recognition engines has been to repurpose general-purpose speech recognition engines, such as Google Speech API, IBM Watson Speech to text API which can be used across domains with natural language question answering systems. Such general-purpose automatic speech engines (gp-ASR) are deep trained on very large general corpus using deep neural network (DNN) techniques. The deep learnt acoustic and language models enhance the performance of a ASR. However, this comes with its own limitations. For freely spoken natural language sentences, the typical recognition accuracy achievable even for state-of-the-art speech recognition systems have been observed to be about 60% to 90% in real-world environments (Lee et al., 2010). The recognition is worse if we consider factors such as domain-specific words, environmental noise, variations in accent, poor ability to express on the part of the user, or inadequate speech and language resources from the domain to train such speech recognition systems. The subsequent natural language processing, such as that in a question answering system, of such erroneously and partially recognized text becomes rather problematic, as the domain terms may be inaccurately recognized or linguistic errors may creep into the sentence. It is, hence, important to improve the accuracy of the ASR output text. In this paper, we focus on the issues of using a readily available gp-ASR and adapting its output for domain-specific natural language question answering (Anantaram et al., 2015a). We present two mechanisms for adaptation, namely (a) an evolutionary development based artificial development mechanism of adaptation (Evo-Devo), where we consider the output of ASR as a biological entity that needs to adapt itself to the environment (in this case the enterprise domain) through a mechanism of repair and development of its genes and (b) a machine learning based mechanism where we examine the closest set of matches with trained examples and the number of adaptive transformations that the ASR output needs to undergo in order to be categorized as an acceptable natural language input for question-answering. We present the results of these two adaptation and gauge the usefulness of each mechanism. The rest of the paper is organized as follows, in Section 2 we briefly describe the work done in this area which motivates our contribution. The main contribution of our work is captured in Section 3 and we show the performance of our approach through experiments in Section 4. We conclude in Section 5. Adapting general-purpose speech recognition engine for QA 3 2 Related Work Most work on ASR error detection and correction has focused on using confidence measures, generally called the log-likelihood score, provided by the speech recognition engine; the text with lower confidence is assumed to be incorrect and subjected to correction. Such confidence based methods are useful only when we have access to the internals of a speech recognition engine built for a specific domain. As mentioned earlier, use of domain-specific engine requires one to rebuild the interface every time the domain is updated, or a new domain is introduced. As mentioned earlier, our focus is to avoid rebuilding the interface each time the domain changes by using an existing ASR. As such our method is specifically a post-ASR system. A post-ASR system provides greater flexibility in terms of absorbing domain variations and adapting the output of ASR in ways that are not possible during training a domain-specific ASR system (Ringger and Allen, 1996). Note that an erroneous ASR output text will lead to an equally (or more) erroneous interpretation by the natural language question-answering system, resulting in a poor performance of the overall QA system Machine learning classifiers have been used in the past for the purpose of combining features to calculate a confidence score for error detection. Non-linguistic and syntactic knowledge for detection of errors in ASR output, using a support vector machine to combine non-linguistic features was proposed in (Shi, 2008) and Naive Bayes classifier to combine confidence scores at a word and utterance level, and differential scores of the alternative hypotheses was used in (Zhou et al., 2005) Both (Shi, 2008) and (Zhou et al., 2005) rely on the availability of confidence scores output by the ASR engine. A syllable-based noisy channel model combined with higher level semantic knowledge for post recognition error correction, independent of the internal confidence measures of the ASR engine is described in (Jeong et al., 2004). In (López-Cózar and Callejas, 2008) the authors propose a method to correct errors in spoken dialogue systems. They consider several contexts to correct the speech recognition output including learning a threshold during training to decide when the correction must be carried out in the context of a dialogue system. They however use the confidence scores associated with the output text to do the correction or not. The correction is carried using syntactic-semantic and lexical models to decide whether a recognition result is correct. In (Bassil and Semaan, 2012) the authors proposes a method to detect and correct ASR output based on Microsoft N-Gram dataset. They use a contextsensitive error correction algorithm for selecting the best candidate for correction using the Microsoft N-Gram dataset which contains real-world data and word sequences extracted from the web which can mimic a comprehensive dictionary of words having a large and all-inclusive vocabulary. In (Jun and Lei, 2011) the authors assume the availability of pronunciation primitive characters as the output of the ASR engine and then use domain-specific named entities to establish the context, leading to the correction of the speech recognition output. The patent (Amento et al., 2007) proposes a manual correction of the ASR output transcripts by providing visual display suggesting the correctness of the text output by ASR. Similarly, (Harwath et al., 2014) propose a re-ranking and classification strategy based on logistic regression model to esti- 4 Anantaram, Sunil mate the probability for choosing word alternates to display to the user in their framework of a tap-to-correct interface. Our proposed machine learning based system is along the lines of (Jeong et al., 2004) but with differences: (a) while they use a single feature (syllable count) for training, we propose the use of multiple features for training the Naive Bayes classifier, and (b) we do not perform any manual alignment between the ASR and reference text – this is done using an edit distance based technique for sentence alignment. Except for (Jeong et al., 2004) all reported work in this area make use of features from the internals of the ASR engine for ASR text output error detection. We assume the use of a gp-ASR in the rest of the paper. Though we use examples of natural language sentences in the form of queries or questions, it should be noted that the description is applicable to any conversational natural language sentence. 3 Domain adaptation of ASR output 3.1 Errors in ASR output In this paper we focus on question answering interfaces to enterprise systems, though our discussion is valid for any kind of natural language processing sentences that are not necessarily a query. For example, suppose we have a retail-sales management system domain, then end-users would be able to query the system through spoken natural language questions (S ) such as ( S= /What is the total sales of miscellaneous store retailers from year two thousand ten to year two thousand fifteen?/ A perfect ASR would take S as the input and produce (T ), namely, ( T = what is the total sales of miscellaneous store retailers from year two thousand ten to year two thousand fifteen We consider the situation where a ASR takes such a sentence (S ) spoken by a person as input, and outputs an inaccurately recognized text (T 0 ) sentence. In our experiments, when the above question was spoken by a person and processed by a popular ASR engine such as Google Speech API, the output text sentence was (T 0 )  what is the total sales of miscellaneous storyteller T0 = from the year two thousand ten to two thousand fifteen Namely S −→ ASR −→ T 0 It should be noted that an inaccurate output by the ASR engine maybe the result of various factors such as background noise, accent of the person speaking the sentence, the speed at which he or she is speaking the sentence, domainspecific words that are not part of popular vocabulary etc. The subsequent natural Adapting general-purpose speech recognition engine for QA 5 language question answering system cannot answer the above output sentence from its retail sales data. Thus the question we tackle here is – how do we adapt or repair the sentence (T 0 ) back to the original sentence (T ) as intended by the speaker. Namely T 0 −→ adaptation, repair −→ T We present two mechanisms for adaptation or repair of the ASR output, namely T 0 −→ T , in this paper: (a) an evolutionary development based artificial develop- ment mechanism, and (b) a machine-learning mechanism. 3.2 Evo-Devo based Artificial Development mechanism of adaption Our mechanism is motivated by Evolutionary Development (Evo-Devo) processes in biology (Harding and Banzhaf, 2008; Anantaram et al., 2015b; Tufte, 2008) to help adapt/repair the overall content accuracy of an ASR output (T 0 ) for a domain. We give a very brief overview of Evo-Devo process in biological organisms and discuss how this motivates our mechanism. In a biological organism, evodevo processes are activated when a new biological cell needs to be formed or an injured cell needs to be repaired/replaced. During such cell formation or repair, the basic genetic structure consisting of the genes of the organism are replicated into the cell – the resultant set of ’genes in the cell’ is called the genotype of the cell. Once this is done, the genotype of the cell is then specialized through various developmental processes to form the appropriate cell for the specific purpose that the cell is intended for, in order to factor-in the traits of the organism – called the phenotype of the cell. For example, if a person has blue-eyes then a blue-eye cell is produced, or if a person has brown skin then a brown-skin cell is produced. During this process, environmental influence may also play a role in the cell’s development and such influences are factored into the genotype-phenotype development process. The field of Evo-Devo has influenced the field of Artificial Intelligence (AI) and a new sub-field called Artificial Development (Art-Dev) has been created that tries to apply Evo-Devo principles to find elegant solutions to adaptation and repair problems in AI. We take inspiration from the Evo-Devo biological process and suitably tailor it to our research problem of repairing the ASR output (T 0 ). In our approach we consider the erroneous ASR output text as the input for our method and treat it as an ’injured biological cell’. We repair that ’injured cell’ through the development of the partial gene present in the input sentence with respect to the genes present in the domain. We assume that we have been provided with the domain ontology describing the terms and relationships of the domain. In our framework, we consider the domain ontology as the true ’genetic knowledge’ of that ’biological organism’. In such a scenario, the ’genetic repair’ becomes a sequence of match-and-replace of words in the sentence with appropriate domain ontology terms and relationships. Once this is done, the ’genotype-to-phenotype repair’ is the repair of linguistic errors in the sentence after the ’genetic repair’. The following sub-section describes our method in detail. 6 Anantaram, Sunil 3.2.1 Repair method We assume that all the instances of the objects in the domain are stored in a database associated with the enterprise system, and can be expressed in relational form (such as [a R c]), for example ['INDUSTRY', 'has', 'PEAK SALES']. A relational database will store it as a set of tables and we treat the data in the database as static facts of the domain. The ontology of the domain can then be generated from this database. We assume that the data schema and the actual data in the enterprise application forms a part of the domain terms and their relationships in the ontology. This identifies the main concepts of the domain with a <subjectpredicate-object> structure for each of the concepts. The ontology thus generated describes the relations between domain terms, for example ['SALES', 'has code', 'NAICS CODE'] or ['OPTICAL GOODS', 'sales 2009', '8767 million'] and thus can be expressed using OWL schema as <s-p-o > structure. Each <s-p-o> entry forms the genes of the domain. We start by finding matches between domain ontology terms and words that appear in the input sentence. Some words of the input sentence will match domain ontology terms exactly. The corresponding domain ontology entry consisting of subject-predicate-object triple is put into a candidate set. Next, other words in the input sentence that are not exact matches of domain ontology terms but have a ’closeness’ match with terms in the ontology are considered. This ’closeness’ match is performed through a mix of phonetic match combined with Levenshtein distance match. The terms that match help identify the corresponding domain ontology entry (with its subject-predicate-object triple) is added to the candidate set. This set of candidate genes is a shortlist of the ’genes’ of the domain that is probably referred to in the input sentence. Next, our mechanism evaluates the ‘fittest’ domain ontology entry from the candidate set to replace the partial gene in the sentence. A fitness function is defined and evaluated for all the candidate genes short-listed. This is done for all words / phrases that appear in the input sentence except the noise words. The fittest genes replace the injured genes of the input sentence. The set of all genes in the sentence forms the genotypes of the sentence. This is the first-stage of adaptation. Once the genotypes are identified, we grow them into phenotypes to remove the grammatical and linguistic errors in the sentence. To do this, we find parts of the sentence that is output by the first-stage of adaptation (the gene-level repair) and that violate well-known grammatical/ linguistic rules. The parts that violate are repaired through linguistic rules. This is the second stage of adaptation/ repair. This process of artificial rejuvenation improves the accuracy of the sentence, which can then be processed by a natural language question answering system (Bhat et al., 2007). Thus, this bio-inspired novel procedure helps adapt/repair the erroneously recognized text output by a speech recognition engine, in order to make the output text suitable for deeper natural language processing. The detailed steps are described below. Step 1: Genes Identification: We match the sub-parts (or sub-strings) of the ASR-output sentence with the genes of the domain. The match may be partial due to the error present in the sentence. The genes in the domain that match the closest, evaluated by a phonetic and/or syntactic match between the ontology Adapting general-purpose speech recognition engine for QA 7 The fitness function F takes as input the asr word, the candidate gene, the Levenshtein distance weight (L), the Phonetic algorithm weight (wP ) and Threshold (T ). Fitness function F then tries to find the closeness of the match between asr word and the candidate gene. To do that, the function calculates two scores: 1. algoScore: is an aggregated score of the similarity of the gene with the asr word by various phonetic algorithms; and 2. editScore: is the Levenshtein distance between the asr word and the gene. The fitness function then calculates the final fitness of the gene using the formula: f inalScore = wP ∗ algoScore + L ∗ (1 − editScore). (1) If the f inalScore is greater than a given threshold T the asr word is replaced by the candidate gene, otherwise the asr word is kept as it is, namely, if (f inalScore > T ) asr word ← gene Fig. 1 Fitness Function. entity and the selected sub-part, are picked up and form the candidates set for the input sentence. For example, let the actual sentence that is spoken by an end-user be "which industry has the peak sales in nineteen ninety seven?". In one of our experiments, when Google Speech API was used as the ASR engine for the above sentence spoken by a user, then the speech engine’s output sentence was "which industry has the pixel in nineteen ninety seven?". This ASR output is erroneous (probably due to background noise or the accent of the speaker) and needs repair/ adaptation for the domain. As a first step, the ASR output sentence is parsed and the Nouns and Verbs are identified from part-of-speech (POS) tags. Syntactic parsing also helps get <subject-verb-object> relations to help identify a potential set of <s-p-o> genes from the ontology. For each of the Nouns and Verbs and other syntactic relations, the partially matching genes with respect to the domain ontology are identified; for this particular sentence the partially matching genes are, "industry" and "pixel". This leads us to identify the probable set of genes in the domain ontology that are most likely a possible match: 'INDUSTRY', 'has', 'PEAK SALES'. The set of all such probable genes need to be evaluated and developed further. Step 2: Developing the genes to identify the genotypes: Once the basic candidate genes are identified, we evaluate the genes to find the best fit for the situation on hand with evolution and development of the genes, and then test a fitness function (see Fig. 3.2 and select the most probable gene that survives. This gives us the set of genotypes that will form the correct ASR sentence. For example, the basic genes "INDUSTRY" and "PIXEL" are used to match the substring "industry has the pixel" with the gene "INDUSTRY', 'has field', 'PEAK SALES’. This is done through a matching and fitness function that would identify the most appropriate gene of the domain. We use a phonetic match function like Soundex, Metaphone, Double-metaphone (Naumann, 2015) to match "pixel" with "PEAK SALES" or an edit-distance match function like Levenshtein distance (Naumann, 2015) to find the closeness of the match. In a large domain there may be many such probable candidates. In such a case, a fitness function is used to decide which of the matches are most suitable. The genes identified are now collated together to repair the input 8 Anantaram, Sunil sentence. This is done by replacing parts of the input sentence by the genes identified in the previous step. In the above example the ASR sentence, "Which industry has the pixel in nineteen ninety seven?" would be adapted/repaired to "Which industry has the peak sales in nineteen ninety seven?". Step 3: Developing Genotypes to Phenotype of sentence: The repaired sen- tence may need further linguistic adaptation/ repair to remove the remaining errors in the sentence. To achieve this, the repaired ASR sentence is re-parsed and the POS tags are evaluated to find any linguistic inconsistencies, and the inconsistencies are then removed. For example, we may notice that there is a WP tag in a sentence that refers to a Wh-Pronoun, but a WDT tag is missing in the sentence that should provide the Determiner for the Wh-pronoun. Using such clues we can look for phonetically matching words in the sentence that could possibly match with a Determiner and repair the sentence. Linguistic repairs such as these form the genotype to phenotype repair/ adaptation of the sentence. The repaired sentence can then be processed for question-answering. We use open source tools like LanguageTool to correct grammatical errors. In addition we have added some domain specific grammar rules. As we understand, the LanguageTool has 54 grammar rules, 5 style rules and 4 built-in Python rules for grammar check and correction. Further we have added some 10 domain specific rules to our linguistic repair function. Our grammar rules can be extended or modified for any domain. 3.2.2 Algorithm of the Evo-Devo process The algorithm has two main functions: ONTOLOGY BASED REPAIR (that encode Steps 1 and 2 described above) and LINGUISTIC REPAIR (encoding Step 3 above). The input sentence is POS tagged and the nouns and verbs are considered. A sliding window allows the algorithm to consider single words or multiple words in a domain term. Let S = w1 , w2 , w3 , · · · , wn be the set of words in the ASR-output(asr out). Let D = dt1 , dt2 , dt3 , · · · , dtm be the domain-ontology-terms. These terms may be considered as candidate genes that can possibly replace the ASR output (asr out) words that may be erroneously recognized. A sliding window of length l consisting of words wi , · · · , wi+l−1 is considered for matching with domain-ontology-terms. The length l may vary from 1 to p, where p may be decided based on the environmental information. For example, if the domain under consideration has financial terms then p may be five words, while for a domain pertaining to car parts, p may be two words. The part match functionality described below evaluates a cost function, say C ({wi , · · · , wi+l−1 }, dtk ) such that minimizing C ({wi , · · · , wi+l−1 }, dtk ) would result in dt∗ which may be a possible candidate to replace {wi , · · · , wi+l−1 }, namely, dt∗ = min C ({wi , · · · , wi+l−1 }, dtk ) G The cost function C ({wi , · · · , wi+l−1 }, dtk ) = b1 ∗ soundex(φ{wi , · · · , wi+l−1 }, φ{dtk }) + b2 ∗ metaphone(φ{wi , · · · , wi+l−1 }, φ{dtk }) + Adapting general-purpose speech recognition engine for QA 9 b3 ∗ (edit distance(φ{wi , · · · , wi+l−1 }, φ{dtk }) + b4 ∗ (number of syllables(φ{wi , · · · , wi+l−1 }) − number of syllables(φ{dtk }) + b5 ∗ (word2vec(φ{wi , · · · , wi+l−1 }) − word2vec(φ{dtk })2 where weights, b1 + b2 + b3 + b4 + b5 = 1 and φ{k} represents each-element-in the set k. If the value of the cost function C ({wi , · · · , wi+l−1 }, dtk ) is greater than the pre-determined threshold T then {wi , · · · , wi+l−1 } may be replaced with the dt∗ , otherwise the {wi , · · · , wi+l−1 } is maintained as it is. The broad algorithm of Evolutionary Development mechanism is shown in Algorithm 1. Algorithm 1 Evo-Devo Mechanism INPUT: ASR output sentence, sentence; domain ontology OUTPUT: Repaired sentence, repaired sentence start 1: // parse the input sentence 2: parsed sentence ← POS tag(sentence) 3: // repair process starts - do genetic repair and find the genotypes 4: part repaired sentence ← ontology based repair(parsed sentence) 5: // grow the genotypes into phenotypes 6: repaired sentence ← linguistic repair(parsed sentence,part repaired sentence) end 7: function ontology based repair(parsed sentence) 8: nouns verbs ← find(parsed sentence, noun verb POStoken) 9: // for each noun verb entry in nouns verbs do next 4 steps 10: // find partially matching genes: match nouns and verbs with entries in domain ontology with phonetic algorithms and Levenshtein distance match 11: concepts referred ← part match(noun verb entry, domain ontology) 12: // find genes: get the subject-predicate-object for concepts 13: candidate genes ← add(spo entry, concepts referred) 14: // simulate the development process of the genes - find the fittest gene from candidate genes 15: fit gene ← fittest(candidate genes, POS token) 16: // add fittest gene into set of genotypes 17: genotypes ← add(fit gene) 18: // replace partially identified genes in input with genotypes identified 19: repaired sentence ← substitute(parsed sentence, nouns verbs, genotypes) 20: return repaired sentence 21: end function 22: 23: 24: 25: 26: 27: 28: 29: 30: 31: 32: 33: 34: 35: function linguistic repair(part repaired sentence) other POStags ← find(part repaired sentence, remaining POStokens) // find POS tags without linguistic completion ling err ← linguistic check( other POStags, part repaired sentence) // find candidate words for linguistic error candidate words ← add(part repaired sentence, ling err) // find the closest semantic match for error words fit word ← fittest word(candidate words, ling err) // add fittest word into repaired sentence fit words ← add(candidate word, fit word) // create the repaired sentence repaired sentence ← replace(part repaired sentence, fit words, other POStags) return repaired sentence end function 10 Anantaram, Sunil 3.2.3 Detailed example of our method Let us assume that we have the domain of retail sales data described in an ontology of <subject-predicate-object> structure as shown in Table 1. Subject INDUSTRY INDUSTRY CAR DEALERS CAR DEALERS CAR DEALERS OPTICAL GOODS OPTICAL GOODS OPTICAL GOODS Predicate BUSINESS BUSINESS SALES 2013 SALES 2011 SALES 2009 SALES 2013 SALES 2011 SALES 2009 Object CAR DEALERS OPTICAL GOODS 737640 million 610747 million 486896 million 10364 million 10056 million 8767 million Table 1 Ontology Structure. Now, let us consider that a user speaks the following sentence to Google Now speech engine: "Which business has more sales in 2013: Car dealers or optical goods?". In our experiment the Google Now speech engine produced the ASR output sentence as "which business has more sales in 2013 car dealers for optical quotes". The recognized ASR sentence has errors. In order to make this ASR sentence more accurate, we input this sentence into the Evo-Devo mechanism and run the process: – Genes Identification (Step 1): We parse the ASR sentence and identify the parts-of-speech in it as: which/WDT, business/NN, has/VBZ, more/JJR, sales/NNS, in/IN, 2013/CD, car/NN, dealers/NNS, for/IN, optical/JJ, quotes/NNS. Considering the words that have POS tags of Nouns (NN/NNS etc.) in the example sentence we get the words "business", "sales", "car", "dealers", "quotes". Based on these words we extract all the partially matching subject- predicate-object instances of the domain ontology. For example, we obtain instances such as [OPTICAL GOODS SALES 2013 10364 million], [INDUSTRY BUSINESS OPTICAL GOODS] and [INDUSTRY BUSINESS CAR DEALERS], etc. from the domain ontology that are partially matching with the words "business" and "sales" respectively. POS tag 2013/CD also leads to reinforcing the above <s-p-o> instance. – Developing the genes to identify the genotypes (Step 2): We replace the erroneous words in the sentence by using a fitness function. The fitness function is defined using string similarity metric (Levenshtein distance) and an aggregated score of phonetic algorithms such as Soundex, Metaphone and Double Metaphone as described in Fitness function in the section above. Thus we get the following adaptation: which business has more sales in 2013 car dealers for optical goods? – Developing Genotypes to Phenotype (Step 3): We now find the parts-ofspeech of the repaired sentence after the step 2 as: which/WDT, business/NN, has/VBZ, more/JJR, sales/NNS, in/IN, 2013/CD, car/NN, dealers/NNS, for/IN, optical/JJ, goods/NNS. In the linguistic repair step, we find that since there is no direct ontological relationship between "car dealers" and "optical goods", we cannot have the Adapting general-purpose speech recognition engine for QA 11 preposition for between these domain terms. Thus we have to find a linguistic relation that is permissible between these domain terms. One of the options is to consider linguistic relations like ‘conjunction’, ‘disjunction’ between domain terms. Thus, when we evaluate linguistic relations AND or OR between these domain terms, we find that OR matches closely with for through a phonetic match rather than AND. Thus we replace for with or in the sentence. Hence the final output of the Evo-Devo mechanism is "which business has more sales in 2013 car dealers or optical goods?". This sentence can now be processed by a question-answering (QA) system. In the above example, a QA system (Bhat et al., 2007) would parse the sentence, identify the known ontological terms {business, sales, 2013, car dealers, optical goods}, find the unknown predicates {which business, more sales}, form the appropriate query over the ontology, and return the answer "CAR DEALERS". 3.2.4 Limitations of the method We assume that there is a well-structured domain ontology for the domain and it is available in the form of <s-p-o> triples. We also assume that the speaker speaks mostly grammatically correct sentences using terms in the domain. While the method would work for grammatically incorrect sentences, the linguistic repair step would suffer. We assume that the speech is processed by a gp-ASR and the ASR-output forms the input sentence that needs repair. However, it is important to note that the input sentence (i.e. the ASR output) need not necessarily contain <s-p-o> triples for our method to work. The <s-p-o> triples that are short-listed from domain ontology aid in forming a candidate set of ’possible genes’ to consider and the fittest amongst them is considered (Step 2) in the context of the other words in the sentence. For example, if the input sentence was ’Who had pick sales’ would get repaired to ’Who had peak sales’ since the domain term of ’peak sales’ would match with ’pick sales’ in our method. Further, input sentences need not necessarily be queries; these can be just statements about a domain. For example, if the above ASR-output sentence was "We hit pick sales this season", the method would repair it as "We hit peak sales this season" using the same set of steps for repair. However, as of now, our method does not repair paraphrases of sentences like "which industry had biggest sales" to "which industry had peak sales". Such repairs need extension to our matching process. The method does not impose any restriction on the sentence or its formation; it can be a fully meaningful sentence in a domain or may contain partial information. The method finds the fittest repair for the inaccuracies occurring in an sentence, post-ASR recognition. It should also be noted that the method does not know the original sentence spoken by the speaker, but tries to get back the original sentence for a particular domain. 3.3 Machine Learning mechanism of adaptation In the machine learning based mechanism of adaptation, we assume the availability of example pairs of (T 0 , T ) namely (ASR output, the actual transcription of the spoken sentence) for training. We further assume that such a machine-learnt model 12 Anantaram, Sunil can help repair an unseen ASR output to its intended correct sentence. We address the following hypothesis Using the information from past recorded errors and the corresponding correction, can we learn how to repair (and thus adapt to a new domain) the text after ASR? Note that this is equivalent to, albiet loosely, learning the error model of a specific ASR. Since we have a small training set, we have used the Naive Bayes classifier that is known to perform well for small datasets with high bias and low variance. We have used the NLTK (Bird et al., 2009) Naive Bayes classifier in all our experiments. Let T 0 be the erroneous text (which is the ASR output), T the corresponding reference text (which is the textual representation of the spoken sentence) and F a feature extractor, such that fβ = F (T 0 β ) (2) fβ = (fβ 1 , fβ 2 , · · · fβn ) (3) where 0 is a set of n features extracted from T β . Suppose there are several pairs say (T 0 i , Ti ) for i = 1, 2, · · ·N . Then we can derive fi for each T 0 i using (2). The probability that T 0 k belongs to the class Tk can be derived through the feature set fk as follows. P (Tk ) ∗ P (fk |Tk ) P (Tk |fk ) = P (f k ) where P (Tk ) is the apriori probability of the class Tk and P (fk |Tk ) is the probability of occurrence of the features fk in the class Tk , and P (fk ) is the overall probability of the occurrence of the feature set fk . Making naive assumption of independence in the features fk1 , fk2 , · · ·fkn we get P (Tk |fk ) = P (Tk ) ∗ (P (fk1 |Tk ) ∗ P (fk2 |Tk ) ∗ · · ·P (fkn |Tk )) P (f k ) (4) In our experiments, the domain specific reference text Ti was spoken by several people and the spoken speech was passed through a general purpose speech recognition engine (ASR) that produced a (possibly) erroneous hypothesis T 0 i . Each pair of reference and the ASR output (i.e. hypothesis) was then word aligned using edit distance, and the mismatching pairs of words were extracted as (T 0 i , Ti ) pairs. For example, if we have the following spoken sentence:  /In which year beer wine and liquor stores has S1 : successful year/ and the corresponding true transcription  In which year beer wine and liquor stores has T1 : successful year One of the corresponding ASR output T 0 1 was  in which year dear wine and liquor stores have T 01 : successful year Adapting general-purpose speech recognition engine for QA 13 In this case the (T 0 , T ) pairs are (dear, beer) and (have, has). As another example consider that T2 was spoken but T 0 2 was recognized by the ASR.  T2 : T 02 : Whether the sales of jewelry business crosses fifty thousand in a year  whether the sales of than twenty business crosses fifty thousand in a year Clearly, in this case the (T 0 , T ) pair is (than twenty, jewelry). Let us assume two features, namely, fβ in (2) is of dimension n = 2. Let the two features be (number of words, number of syllables). Then, for the (T 0 , T ) pair (than twenty, jewelry) we have F ((than twenty)) = (2, 3) since the number of words in than twenty is 2 and than twenty contains 3 syllables. P (fk1 |Tk ) in this case would be the probability that the number of words in the input are two (fk1 = 2) when the correction is jewelry. A third example is: T3 : n In two thousand thirteen which industry had the peak sales T 0 3 : {in two thousand thirteen which industry have the pixels Note that in this case the (T 0 , T ) pair is (peak sales, pixel). Calculating thus the values of P (Tk ) for all reference corrections, P (fkj |Tk ) for all feature values, fkj for all the j features in fk , we are in a position to calculate the RHS of (4). When this trained classifier is given an erroneous text, features are extracted from this text and the repair works by replacing the erroneous word by a correction that maximizes (4), Tk∗ = max P (Tk |fk ) Tk Namely, the Tk∗ for which P (Tk |fk ) is maximum. 4 Experiments and results We present the results of our experiments with both the Evo-Devo and the Machine Learning mechanisms described earlier using the U.S. Census Bureau conducted Annual Retail Trade Survey of U.S. Retail and Food Services Firms for the period of 1992 to 2013 (USCensus, 2015). 14 Anantaram, Sunil Fig. 2 T 0 accuracy (y-axis) for the 250 utterance (x-axis) for Ga, Ki, Ku and Ps. 4.1 Data Preparation We downloaded this survey data and hand crafted a total of 293 textual questions (AwazYP, 2015) which could answer the survey data. A set of 6 people (L2 English) generated 50 queries each with the only constraint that these queries should be able to answer the survey data. In all a set of 300 queries were crafted of which duplicate queries were removed to leave 293 queries in all. Of these, we chose 250 queries randomly and distributed among 5 Indian speakers, who were asked to read aloud the queries into a custom-built audio data collecting application. So, in all we had access to 250 audio queries spoken by 5 different Indian speakers; each speaking 50 queries. Each of these 250 audio utterances were passed through 4 different ASR engines, namely, Google ASR (Ga), Kaldi with US acoustic models (Ku), Kaldi with Indian Acoustic models (Ki) and PocketSphinx ASR (Ps). In particular, that audio utterances were in wave format (.wav) with a sampling rate of 8 kHz and 16 bit. In case of Google ASR (Ga), each utterance was first converted into .flac format using the utility sound exchange (sox) commonly available on Unix machines. The .flac audio files were sent to the cloud based Google ASR (Ga) one by one in a batch mode and the text string returned by Ga was stored. In all 7 utterances did not get any text output, presumably Ga was unable to recognize the utterance. For all the other 243 utterances a text output was received. In case of the other ASR engines, namely, Kaldi with US acoustic models (Ku), Kaldi with Indian Acoustic models (Ki) and PocketSphinx ASR (Ps) we first took the queries corresponding to the 250 utterances and built a statistical language model (SLM) and a lexicon using the scripts that are available with PocketSphinx Adapting general-purpose speech recognition engine for QA 15 S : {/Which stores has total sales more than two hundred thousand/ T 0 : {which state has total sales more than twenty thousand Ga; 70.00% 0 T : {which stores has total sales more than two in two thousand Ki; 80.00% T 0 : {which stores has total sales more than point of sales Ku; 70.00% T 0 : {list the total sales more than Ps; 40.00% Fig. 3 Sample output (T 0 ) of four different ASR for the same spoken utterance (S). Also shown are the accuracy of the ASR output. ASR engine Result No result Google ASR (Ga) Kaldi US (Ku) Kaldi IN (Ki) PocketSphinx (Ps) 243 250 250 250 7 0 0 0 Total 993 7 Correct 55 103 127 44 Error (A+B) 188 147 123 206 >=70% (A) 143 123 111 109 <70% (B) 45 24 12 97 329 664 486 178 Table 2 ASR engines and their output %accuracy (CMU, 2017) and Kaldi (Kaldi, 2017). This language model and lexicon was used with the acoustic model that were readily available with Kaldi and Ps. In case of Ku we used the American English acoustic models, while in case of Ki we used the Indian English acoustic model. In case of Ps we used the Voxforge acoustic models (VoxForge, 2017). Each utterance was passed through Kaldi ASR for two different acoustic models to get T 0 corresponding to Ku and Ki. Similarly all the 250 audio utterance were passed through the Ps ASR to get the corresponding T 0 for Ps. A sample utterance and the output of the four engines is shown in Figure 3. Figure 2 and Table 2 capture the performance of the different speech recognition engines. The performance of the ASR engines varied, with Ki performing the best with 127 of the 250 utterances being correctly recognized while Ps returned only 44 correctly recognized utterances (see Table 2, Column 4 named ”Correct”) of 250 utterances. The accuracy of the ASR varied widely. For instance, in case of Ps there were as many as 97 instances of the 206 erroneously recognized utterances which had an accuracy of less than 70%. Note that the accuracy is computed as the number of deletions, insertions, substitutions that are required to convert the ASR output to the textual reference (namely, T 0 → T ) and is a common metric used in speech literature (Hunt, 1990). For all our analysis, we used only those utterances that had an accuracy 70% but less that 100%, namely, 486 instances (see Table 2, Figure 4). An example showing the same utterance being recognized by four different ASR engines is shown in Figure 3. Note that we used T 0 corresponding to Ga, Ki and Ku in our analysis (accuracy ≥ 70%) and not T 0 corresponding to Ps which has an accuracy of 40% only. This is based on our observation that any ASR output that is lower that 70% accurate is so erroneous that it is not possible to adapt and steer it towards the expected output. The ASR output (T 0 ) are then given as input in the Evo-Devo and Machine Learning mechanism of adaptation. 16 Anantaram, Sunil Fig. 4 All utterances that have and T 0 accuracy (y-axis) ≥ 70 and < 100 used in all our experiments. 4.2 Evo-Devo based experiments We ran our Evo-Devo mechanism with the 486 ASR sentences (see Table 2) and measured the accuracy after each repair. On an average we have achieved about 5 to 10% improvements in the accuracy of the sentences. Fine-tuning the repair and fitness functions, namely Equation (1), would probably yield much better performance accuracies. However, experimental results confirm that the proposed Evo-Devo mechanism is an approach that is able to adapt T 0 to get closer to T . We present a snapshot of the experiments with Google ASR (Ga) and calculate accuracy with respect to the user spoken question as shown in Table 3. Table 3 clearly demonstrates the promise of the evo-devo mechanism for adaptation/repair. In our experiments we observed that the adaptation/repair of subparts in ASR-output (T 0 ) that most probably referred to domain terms occurred well and were easily repaired, thus contributing to increase in accuracy. For nondomain-specific linguistic terms the method requires one to build very good linguistic repair rules, without which the method could lead to a decrease in accuracy. One may need to fine-tune the repair, match and fitness functions for linguistic terms. However, we find the abstraction of evo-devo mechanism is very apt to use. 4.3 Machine Learning experiments In the machine learning technique of adaptation, we considers (T 0 , T ) pairs as the predominant entity and tests the accuracy of classification of errors. Adapting general-purpose speech recognition engine for QA User’s Question (T ), Google ASR out (T 0 Ga ), After Evo-devo (T 0 ED ) T : {In two thousand fourteen which industry had the peak sales T 0 Ga : {in two thousand fourteen which industry had the pixels T 0 ED : {in two thousand fourteen which industry had the peak sales nin which year did direct selling establishments make the T : maximum and in which year did they do the minimum sales nwhichsales year did direct selling establishments make the maximum T 0 Ga : and in which year did they do the many muscles ncells which year did direct selling establishments make the maximum T 0 ED : sales and in which year did they do the many musical (Which one among the electronics and appliance store and food T : and beverage stores has sales in more than hundred thousand in(at least three years in a row which one among the electronics and appliance store and food T 0 Ga : and beverages stores have sales in more than one lakh in at least three years in a row (which one among the electronics and appliance store and food 0 T ED : and beverage stores have sales in more than one lakh in at least three years in a row 17 Acc Ga: 80% ED: 100% Ga:80.9% ED: 85.7% Ga: 85.7% ED:89.3% Table 3 Evo-Devo experiments with Google ASR (Ga). In our experiment, we used a total of 570 misrecognition errors (for example, (dear, beer) and (have, has) derived from (T 0 1 , T1 ) or (than twenty, jewelry) derived from (T 0 2 , T2 )) in the 486 sentences. We performed 10-fold cross validation, each fold containing 513 (T 0 , T ) pairs for training and 57 pairs for testing, Note that we assume the erroneous words in the ASR output being marked by a human oracle, in the training as well as the testing set. Suppose the following example (T4 ) occurs in the training set: T4 : n T 04 : n Which business has posted cumulative sales of more than one million dollars from the 2007 to 2012 which business has posted latest stills of more than one million dollars from 2007 to 2012 The classifier is given the pair {F (latest stills), cumulative sales} to the classifier. And if the following example occurs in the testing set (T5 ), T5 : n T 05 : n Did sales remain the same in retail between two thousand thirteen and two thousand fourteen sales wine same in retail between two thousand thirteen and two thousand fourteen the trained model or the classifier is provided F (wine) and successful repair would mean it correctly labels (adapts) it to remain the. The features used for classification were (n = 6 in Equation (3)) fβ 1 fβ 2 fβ 3 fβ 4 fβ 5 fβ 6 → → → → → → Left context (word to the left of T 0 ), Number of errors in the entire ASR sentence, Number of words in T 0 , Right context (word to the right of T 0 ), Bag of vowels of T 0 and Bag of consonants of T 0 . 18 Anantaram, Sunil The combination of features fβ 6 , fβ 5 , fβ 1 , fβ 3 , fβ 4 namely, (bag of consonants, bag of vowels, left context, number of words, right context) gave the best results with 32.28% improvement in accuracy in classification over 10-fold validation. The experimental results for both evo-devo and machine learning based approaches demonstrate that these techniques can be used to correct the erroneous output of ASR. This is what we set out to establish in this paper. 5 Conclusions General-purpose ASR engines when used for enterprise domains may output erroneous text, especially when encountering domain-specific terms. One may have to adapt/repair the ASR output in order to do further natural language processing such as question-answering. We have presented two mechanisms for adaptation/repair of ASR-output with respect to a domain. The Evo-Devo mechanism provides a bio-inspired abstraction to help structure the adaptation and repair process. This is one of the main contribution of this paper. The machine learning mechanism provides a means of adaptation and repair by examining the featurespace of the ASR output. The results of the experiments show that both these mechanisms are promising and may need further development. 6 Acknowledgments Nikhil, Chirag, Aditya have contributed in conducting some of the experiments. We acknowledge their contribution. References Cheongjae Lee, Sangkeun Jung, Kyungduk Kim, Donghyeon Lee, and Gary Geunbae Lee. Recent approaches to dialog management for spoken dialog systems. Journal of Computing Science and Engineering, 4(1):1–22, 2010. C. Anantaram, Rishabh Gupta, Nikhil Kini, and Sunil Kumar Kopparapu. Adapting general-purpose speech recognition engine output for domain-specific natural language question answering. In Workshop on Replicability and Reproducibility in Natural Language Processing: adaptive methods, resources and software at IJCAI 2015, Buenos Aires, 2015a. E.K. Ringger and J.F. Allen. Error correction via a post-processor for continuous speech recognition. In Acoustics, Speech, and Signal Processing, 1996. ICASSP-96. Conference Proceedings., 1996 IEEE International Conference on, volume 1, pages 427–430 vol. 1, May 1996. doi: 10.1109/ICASSP.1996.541124. Yongmei Shi. An Investigation of Linguistic Information for Speech Recognition Error Detection. PhD thesis, University of Maryland, Baltimore County, October 2008. Lina Zhou, Jinjuan Feng, A. Sears, and Yongmei Shi. Applying the naive bayes classifier to assist users in detecting speech recognition errors. In System Sciences, 2005. HICSS ’05. Proceedings of the 38th Annual Hawaii International Conference on, pages 183b–183b, Jan 2005. doi: 10.1109/HICSS.2005.99. Adapting general-purpose speech recognition engine for QA 19 Minwoo Jeong, Byeongchang Kim, and G Lee. Using higher-level linguistic knowledge for speech recognition error correction in a spoken q/a dialog. In HLTNAACL special workshop on Higher-Level Linguistic Information for Speech Processing, pages 48–55, 2004. Ramón López-Cózar and Zoraida Callejas. Asr post-correction for spoken dialogue systems based on semantic, syntactic, lexical and contextual information. Speech Commun., 50(8-9):745–766, August 2008. ISSN 0167-6393. doi: 10.1016/j. specom.2008.03.008. URL http://dx.doi.org/10.1016/j.specom.2008.03.008. Youssef Bassil and Paul Semaan. ASR context-sensitive error correction based on microsoft n-gram dataset. CoRR, abs/1203.5262, 2012. URL http://arxiv.org/ abs/1203.5262. J. Jun and L. Lei. Asr post-processing correction based on ner and pronunciation primitive. In 2011 7th International Conference on Natural Language Processing and Knowledge Engineering, pages 126–131, Nov 2011. doi: 10.1109/NLPKE. 2011.6138180. B. Amento, P. Isenhour, and L. Stead. Error correction in automatic speech recognition transcripts, September 6 2007. URL https://www.google.com/patents/ US20070208567. US Patent App. 11/276,476. David Harwath, Alexander Gruenstein, and Ian McGraw. ”choosing useful word alternates for automatic speech recognition correction interfaces. In INTERSPEECH-2014, pages 949–953, 2014. Simon Harding and Wolfgang Banzhaf. Artificial development. In Organic Computing, Understanding Complex Systems, pages 201–219. Springer Berlin Heidelberg, 2008. ISBN 978-3-540-77656-7. doi: 10.1007/978-3-540-77657-4 9. URL http://dx.doi.org/10.1007/978-3-540-77657-4_9. C Anantaram, Nikhil Kini, Chirag Patel, and Sunil Kopparapu. Improving asr recognized speech output for effective nlp. In The Ninth International Conference on Digital Society, pages 17–21, Lisbon, Portugal, Feb 2015b. Gunnar Tufte. From Evo to EvoDevo: Mapping and Adaptation in Artificial Development. Development, 2008. Shefali Bhat, C. Anantaram, and Hemant Jain. Framework for text-based conversational user-interface for business applications. In Proceedings of the 2Nd International Conference on Knowledge Science, Engineering and Management, KSEM’07, pages 301–312, Berlin, Heidelberg, 2007. Springer-Verlag. ISBN 3-540-76718-5, 978-3-540-76718-3. URL http://dl.acm.org/citation.cfm?id= 1775431.1775465. Felix Naumann. http://hpi.de/fileadmin/user_upload/fachgebiete/naumann/ folien/SS13/DPDC/DPDC_12_Similarity.pdf, 2015. Steven Bird, Ewan Klein, and Edward Loper. Natural Language Processing with Python. O’Reilly Media, Inc., 1st edition, 2009. ISBN 0596516495, 9780596516499. USCensus. http://www.census.gov/retail/index.html, 2015. Viewed Sep 2015. AwazYP. https://sites.google.com/site/awazyp/data/ijcai, 2015. Viewed Aug 2017. CMU. Building language model for pocketsphinx, 2017. URL http://cmusphinx. sourceforge.net/wiki/tutoriallm. Kaldi. Overview of graph creation in kaldi, 2017. URL http://kaldi-asr.org/ doc/graph.html. 20 Anantaram, Sunil VoxForge. Updated 8khz sphinx acoustic model, 2017. URL http://www.voxforge. org/home/news/news/updated-8khz-sphinx-acoustic-model. Melvyn J. Hunt. Figures of merit for assessing connected-word recognisers. Speech Communication, 9(4):329 – 336, 1990. ISSN 0167-6393. doi: http://dx.doi.org/10. 1016/0167-6393(90)90008-W. URL http://www.sciencedirect.com/science/ article/pii/016763939090008W.
2cs.AI
Algorithms to Approximate Column-Sparse Packing Problems Brian Brubach∗, Karthik A. Sankararaman†, Aravind Srinivasan‡, and Pan Xu§ arXiv:1711.02724v4 [cs.DS] 11 Dec 2017 Department of Computer Science, University of Maryland, College Park, MD 20742, USA Abstract Column-sparse packing problems arise in several contexts in both deterministic and stochastic discrete optimization. We present two unifying ideas, (non-uniform) attenuation and multiple-chance algorithms, to obtain improved approximation algorithms for some well-known families of such problems. As three main examples, we attain the integrality gap, up to lower-order terms, for known LP relaxations for k-column sparse packing integer programs (Bansal et al., Theory of Computing, 2012) and stochastic k-set packing (Bansal et al., Algorithmica, 2012), and go “half the remaining distance” to optimal for a major integrality-gap conjecture of Füredi, Kahn and Seymour on hypergraph matching (Combinatorica, 1993). This is a full version of the paper that appeared in SODA-2018. This version improves the presentation and fixes some typos. Acknowledgements. We would like to thank David Harris for pointing out the equivalence between Conjecture 2 and Conjecture 1. We also thank him for pointing out useful prior results on graph coloring in parallel and distributed settings. We would like to thank the anonymous reviewers for valuable comments on this paper. ∗ Email: [email protected] Supported in part by NSF Awards CNS 1010789 and CCF 1422569. † Email: [email protected] Supported in part by NSF Awards CNS 1010789 and CCF 1422569. ‡ Email: [email protected] Supported in part by NSF Awards CNS-1010789, CCF-1422569 and CCF-1749864, and by research awards from Adobe, Inc. § Email: [email protected] Supported in part by NSF Awards CNS 1010789 and CCF 1422569. 1 Introduction Column-sparse packing problems arise in numerous contexts (e.g., [2, 15, 12, 13, 47, 16, 25, 24, 52, 39, 46]). We present two unifying ideas (attenuation and multiple-chances) to obtain improved approximation algorithms and/or (constructive) existence results for some well-known families of such problems. These two unifying ideas help better handle the contention resolution [32] that is implicit in such problems. As three main examples, we attain the integrality gap (up to lower-order terms) for known LP relaxations for k-column sparse packing integer programs (k -CS-PIP: Bansal et al. [13]) and stochastic k-set packing (SKSP: Bansal et al. [12]), and go “half the remaining distance” to optimal for a major integrality-gap conjecture of Füredi, Kahn and Seymour on hypergraph matching [39]. Letting R+ denote the set of non-negative reals, a general Packing Integer Program (PIP) takes the form: max {f (x) | A · x ≤ b, x ∈ {0, 1}n } , m×n ; (1) where b ∈ Rm + , and A ∈ R+ here A · x ≤ b means, as usual, that A · x ≤ b coordinate-wise. Furthermore, n is the number of variables/columns, m is the number of constraints/rows, A is the matrix of sizes with the j th column representing the size vector SIj ∈ Rm + of item j, b is the capacity vector, and f is some non-decreasing function (often of the form w · x, where w is a nonnegative vector of weights). The items’ “size vectors” SIj can be deterministic or random. PIPs generalize a large class of problems in combinatorial optimization. These range from optimally solvable problems such as classical matching to much harder problems like independent set which is NP-Hard to approximate to within a factor of n1−ǫ [70]. A k-column sparse packing program (k -CS-PP) refers to a special case of packing programs wherein each size vector SIj (a column of A) takes positive values only on a subset C(j) ⊆ [m] of coordinates with |C(j)| ≤ k. The k -CS-PP family captures a broad class of packing programs that are well studied such as k-column sparse packing integer programs (k -CS-PIP), k-uniform hypergraph matching, stochastic matching, and stochastic k-set packing (SKSP). While we primarily focus on programs with linear objectives, some of these approaches can be extended to monotone submodular objectives as well from prior work (e.g., [13], [32]). We show randomized-rounding techniques (including non-uniform attenuation, multiple chances) that, along with the “nibble method” [4, 63] in some cases, yield improved results for some important families of Packing Integer Programs (PIPs). In the case of k -CS-PIP and SKSP, we show asymptotically optimal bounds matching the LP integrality gap (as a function of the column-sparsity k, which is our asymptotic parameter). For hypergraph matching, we make progress “half the remaining way” towards meeting a classic conjecture of Füredi et al. [39]. Additionally, we show a simple application of simulation-based attenuation to obtain improved ratios for the Unsplittable Flow Problem on trees (UFP-TREES: Chekuri et al. [30]) with unit demands and submodular objectives, a problem which admits a natural packing-LP relaxation. 1.1 Preliminaries and Main Results The natural LP relaxation is as follows (although additional valid constraints are necessary for k -CS-PIP [13]): max{w · x : A · x ≤ b, x ∈ [0, 1]n } (2) Typically, a rounding algorithm takes as input an optimal solution x ∈ [0, 1]n to LP (2) – or one of its relatives – and outputs an integral X ∈ {0, 1}n which is feasible for PIP (1) such that the resultant 1 approximation ratio, E[w·X] w·x , is maximized. Note that E[w · X] is the expected weight of the solution over all randomness in the algorithm and/or the problem itself. For a general weight vector w, we often seek to E[X ] maximize minj:xj 6=0 xj j , as the usual “local” strategy of maximizing the approximation ratio. As notation, we will denote the support of X as the set of rounded items. We say item j participates in constraint i if and only if Aij 6= 0. We say that a variable is safe to be rounded to 1 if doing so would not violate any constraint conditional on the variables already rounded; we call it unsafe otherwise. k-Column Sparse Packing Integer Programs (k -CS-PIP). Suppose we have n items and m constraints. Each item j ∈ [n] has a weight wj and a column aj ∈ [0, 1]m . Suppose we have a capacity vector b = 1 (this is w.l.o.g., see e.g., Bansal et al. [13]) and our goal is to select a subset of items such that the total weight is maximized while no constraint is violated. In addition, we assume each column aj has at most k non-zero entries. (The important special case where for all j, aj lies in {0, 1}m – and has at most k non-zero entries – is the classic k-set packing problem; it is NP-hard to approximate within o(k/ log k) [44]. This special case is generalized in two ways below: by allowing stochasticity in stochastic k-set packing, and by allowing the column-sparsity k to vary across columns as in hypergraph matching.) Observe that this problem (i.e., k -CS-PIP) can be cast as a special case of PIP shown in (1) with the j th column of A being A[j] = aj . The resultant LP relaxation is as follows (just as in Bansal et al. [13], we will ultimately use a stronger form of this LP relaxation which incorporates additional valid constraints; see (7) in Section 2.1). max{w · x : A · x ≤ 1, x ∈ [0, 1]n } where A[j] = aj (3) For general PIPs, the best-known approximation bounds are shown in Srinivasan [67]. The problem of k -CS-PIP, in its full generality, was first considered by Pritchard [61] and followed by several subsequent works such as Pritchard and Chakrabarty [62] and Bansal et al. [13]. Chekuri et al. [32] defined a contention resolution framework for submodular objectives and showed how the previous algorithms for k -CS-PIP fit into such a framework (and hence, extending the k -CS-PIP algorithms to non-negative submodular objectives by losing a constant factor in approximation)1 . Our main result for this problem is described in Theorem 1. Bansal et al. [13] showed that the stronger LP (which adds additional valid constraints to the natural LP relaxation) has an integrality gap of at least 2k − 1. We consider the same LP, and hence our result shown in Theorem 1 is asymptotically optimal w.r.t. this LP. The previous best known results for this problem were a factor of ek + o(k) due to Bansal et al. [13], a factor of O(k2 ) independently due to Chekuri et al. [26] 2 and Pritchard and Chakrabarty [62], and a factor of O(2k · k2 ) due to Pritchard [61]. Theorem 1. There exists a randomized rounding algorithm for k -CS-PIP with approximation ratio at most 2k + o(k) for linear objectives. Corollary 1. There exists a randomized rounding algorithm for k -CS-PIP with approximation ratio at most (2k + o(k))/ηf for non-negative submodular objectives, where ηf is the approximation ratio for max{F (x) : x ∈ PI ∩ {0, 1}n } (here, F (x) is the multi-linear extension of the sub-modular function f and PI is the k -CS-PIP polytope); ηf = 1 − 1/e and ηf = 0.385 in the cases of non-negative monotone and non-monotone submodular functions respectively3 . 1 In [13], the authors also show extensions to non-negative monotone submodular objectives. Note that this work is cited in [13]. 3 To keep consistent with prior literature, we state all approximation ratios for sub-modular maximization (i.e., ηf ) as a value less than 1. This is in contrast to the approximation ratios defined in this paper where the values are always greater than 1 2 2 Stochastic k-Set Packing (SKSP). The Stochastic k-Set Packing problem was first introduced in Bansal et al. [12] as a way to generalize several stochastic-optimization problems such as Stochastic Matching4 . The problem can be defined formally as follows. Suppose we have n items and that each item j has a random non-negative weight Wj and a random m-dimensional size vector SIj ∈ {0, 1}m . The random variables {Rj := (Wj , SIj ) : j ∈ [n]} are mutually independent5 . Each random vector Rj ∈ R+ × {0, 1}m is drawn from some probability distribution: our algorithm only needs to know the values of ui,j := E[SIi,j ] for all i, j – where SIi,j denotes the ith component of SIj – and wj := E[Wj ]. Moreover, for each item j, there is a known subset C(j) ⊆ [m] of at most k coordinates such that SIi,j can be nonzero only if i ∈ C(j): all coordinates in [m] \ C(j) will have value zero with probability 1. We are given a capacity vector b ∈ Zm +. The algorithm proceeds in multiple steps. At each step, we consider any one item j that has not been considered before, and which is safe with respect to the current remaining capacity, i.e., adding item j to the current set of already-added items will not cause any capacity constraint to be violated regardless of what random SIj materializes.6 Upon choosing to probe j, the algorithm observes its size realization and weight, and has to irrevocably include j. The task is to sequentially probe some subset of the items such that the expected total weight of items added is maximized. Let w denote (w1 , . . . , wn ) and xj denote the probability that j is added in the OPT solution. Bansal et al. [12] introduced the following natural LP to upper bound the optimal performance. max{w · x : A · x ≤ b, x ∈ [0, 1]n } where A[i, j] = ui,j (4) The previous best known bound for SKSP was 2k+o(k) due to Bansal et al. [12]. Our main contribution (Theorem 2) is to improve this bound to k + o(k), a result that is again asymptotically optimal w.r.t. the natural LP (4) considered (Theorem 1.3 from [39]). Theorem 2. There exists a randomized rounding algorithm achieving an approximation ratio of k + o(k) for the stochastic k-set packing problem, where the “o(k)” is a vanishing term when k → ∞. Hypergraph Matching. Suppose we have a hypergraph H = (V, E) with |V| = m and |E| = n. (This is the opposite of the usual graph notation, but is convenient for us since the LP here has |V| constraints and |E| variables.) Each edge e ∈ E has a weight we . We need to find a subset of edges with maximum total weight such that every pairwise intersection is empty (i.e., we obtain a hypergraph matching). Observe that the problem of finding a maximum weighted hypergraph matching can be cast as a special case of PIP. Let w = (we ) and e ∈ {0, 1}m be the canonical (characteristic-vector) representation of e. Then the natural LP relaxation is as follows: max{w · x : A · x ≤ 1, x ∈ [0, 1]n } where A[j] = ej (5) Note that in these natural IP and LP formulations, the number of vertices in an edge e, ke = |e|, can be viewed as the column-sparsity of the column associated with e. Thus, this again broadly falls into the class of column-sparse packing programs. For general hypergraphs, Füredi et al. [39] presented the following well-known conjecture. Conjecture 1 (Füredi et al. [39]). For any hypergraph H = (V, E) and a weight vector w = (we ) over all edges, there exists a matching M such that   P 1 (6) k − 1 + e e∈M ke we ≥ OPT(H, w) 4 Here, we use the definition from the journal version [12]; the conference version of [12] defines the problem slightly differently. Note that Wj can be correlated with SIj . 6 This is called the safe-policy assumption. This allows us to handle the correlations between Wj and coordinates of SIj . A detailed discussion of this model can be found in [12]. 5 3 where ke denotes the number of vertices in hyperedge e and OPT(H, w) denotes an optimal solution to the LP relaxation (5) of hypergraph matching. The function “ke − 1 + k1e ” is best-possible in the sense that certain hypergraph families achieve it [39]. We generalize Conjecture 1 slightly: Conjecture 2 (Generalization of Conjecture 1). For any given hypergraph H = (V, E) with notation as in Conjecture 1, let x = (xe : e ∈ E) denote a given optimal solution to the LP relaxation (5). Then: (i) there is a distribution D on the matchings of H such that for each edge e, the probability that it is present in a xe , and (ii) D is efficiently samplable. sample from D is at least ke −1+1/k e Part (i) of Conjecture 2 immediately implies Conjecture 1 via the linearity of expectation. In fact, Part (i) of Conjecture 2 and Conjecture 1 are equivalent which can be shown using a LP-duality argument for the standard LP for computing the fractional chromatic number. Füredi et al. [39] gave (non-constructive) proofs for Conjecture 1 for the three special cases where the hypergraph is either uniform, intersecting, or uniformly weighted. Chan and Lau [24] gave an algorithmic proof of Conjecture 1 for k-uniform hypergraphs, by combining the iterative rounding method and the fractional local ratio method. Using similar techniques, Parekh [58] and Parekh and Pritchard [59] generalized this to k- uniform b-hypergraph matching. We go “half the remaining distance” in resolving Conjecture 2 for all hypergraphs, and also do so algorithmically: the work of Bansal et al. [12] gives “ke + 1” instead of the target ke − 1 + 1/ke in Conjecture 2, and we improve this to ke + O(ke · exp(−ke )). Theorem 3. There exists an efficient algorithm to generate a random matching M for a hypergraph such xe that each edge e is added in M with probability at least ke +o(1) , where {xe } is an optimal solution to the standard LP (5) and where the o(1) term is O(ke exp(−ke )), a vanishing term when ke → ∞. UFP-TREES with unit demands. In this problem, we are given a tree T = (V, E) with each edge e having an integral capacity ue . We are given k distinct pairs of vertices (s1 , t1 ), (s2 , t2 ), . . . , (sk , tk ) each having unit demand. Routing a demand pair (si , ti ) exhausts one unit of capacity on all the edges in the path. With each demand pair i, there is an associated weight wi ≥ 0. The goal of the problem is to choose a subset of demand pairs to route such that no edge capacity is violated, while the total weight of the chosen subset is maximized. In the non-negative submodular version of this problem, we are given a non-negative submodular function f over all subsets of demand pairs, and aim to choose a feasible subset that maximizes f . This problem was introduced by Chekuri et al. [30] and the extension to submodular version was given by Chekuri et al. [32]. We show that by incorporating simple attenuation ideas, we can improve the analysis of the previous best algorithm for the Unsplittable Flow Problem in Trees (UFP-TREES) with unit demands and non-negative submodular objectives. Chekuri et al. [32] showed that they can obtain an approximation of 27/ηf , where ηf is the approximation ratio for maximizing a non-negative submodular function, via their contention-resolution scheme (henceforth abbreviated as CR schemes)7 . We improve their 1/27-balanced CR scheme to a 1/8.15-balanced CR scheme via attenuation and hence achieve an approximation of 8.15/ηf for non-negative sub-modular objectives. Theorem 4. There exists a 8.15/ηf -approximation algorithm to the UFP-TREES with unit demands and non-negative submodular objectives. 7 Please see Section 2.3 for formal definitions of CR schemes. 4 Extension to submodular objectives. Chekuri et al. [32] showed that given a rounding scheme for a PIP with linear objectives, we can extend it to non-negative submodular objectives by losing only a constant factor, if the rounding scheme has a certain structure (see Theorem 7, due to [32]). Our improved algorithm for k -CS-PIP and UFP-TREES admits this structure and hence can be extended to non-negative submodular functions. See Section B in the Appendix for the required background on submodular functions. A simple but useful device that we will use often is as follows. Simulation-based attenuation. We use the term simulation throughout this paper to refer to Monte Carlo simulation and the term simulation-based attenuation to refer to the simulation and attenuation techniques as shown in [2] and [19]. At a high level, suppose we have a randomized algorithm such that for some event E (e.g., the event that item j is safe to be selected into the final set in SKSP) we have Pr[E] ≥ c, then we modify the algorithm as follows: (i) We first use simulation to estimate a value Ê that lies in the range [Pr[E], (1 + ǫ) Pr[E]] with probability at least 1 − δ. (ii) By “ignoring” E (i.e., attenuation, in a problem-specific manner) with probability ∼ c/Ê, we can ensure that the final effective value of Pr[E] is arbitrarily close to c, i.e., in the range [c/(1 + ǫ), c] with probability at least 1 − δ. This simple idea of attenuating the probability of an event to come down approximately to a certain value c is what we term simulation-based attenuation. The number of samples needed to obtain the estimate Ê is Θ( cǫ12 · log( 1δ )) via a standard Chernoff-bound argument. In our applications, we will take ǫ = 1/poly(N ) where N is the problem-size, and the error ǫ will only impact lower-order terms in our approximations. 1.2 Our Techniques In this section, we describe our main technical contributions of the paper and the ingredients leading up to them. Achieving the integrality gap of the LP of [13] for k -CS-PIP. Our first main contribution in this paper is to achieve the integrality gap of the strenghthened LP of [13] for k -CS-PIP, up to lower-order terms: we improve the ek + o(k) of [13] to 2k + o(k). We achieve this by following the same overall structure as in [13] and improve the alteration steps using randomization. We view the alteration step as a question on an appropriately constructed directed graph. In particular, a key ingredient in the alteration step answers the following question. “Suppose we are given a directed graph G such that the maximum out-degree is bounded by an asymptotic parameter d. Find a random independent set I in the undirected version of this graph such that every vertex is added into I with probability at least 1/(2d)−o(1/d)”. It turns out that this question can be answered by looking at the more-general question of finding a good coloring of the undirected version of this graph. The key idea here is to “slow down” the contention-resolution approach of [13], leading to Theorem 1. However, motivated by works that obtain strong “negative correlation” properties – e.g., the papers [31, 60] obtain negative cylindrical correlation8 and the even-stronger negative association for rounding in matroid polytopes – we ask next if one can achieve this for k -CS-PIP. (It is well-known that even negative cylindrical correlation yields Chernoff-type bounds for sums of random variables [57]; we use this in Section 5.) We make progress toward this in Theorem 6. Achieving the integrality gap of the natural LP for SKSP via a “multiple chances” technique. Our second contribution in is to develop an algorithm that achieves the integrality gap of k + o(k) for SKSP, improving on the 2k of [12]. To achieve this, we introduce the “multiple-chances” technique. We will now informally describe this technique, which is motivated by the powerful “nibble” idea from probabilistic combinatorics (see, e.g., Ajtai, Komlós, and Szemerédi [4] and Rödl [63]). 8 This is sometimes simply called “negative correlation”. 5 The current-best ratios for many special cases of k -CS-PP are Θ(k); e.g., ek+o(k) for k -CS-PIP [13], the optimal approximation ratio (w.r.t. the integrality gap) of k − 1 + 1/k for k-uniform hypergraph matching [24], or the 2k-approximation for SKSP [12]. Thus, many natural approaches involve sampling items with a probability proportional to 1/k. Consider a k -CS-PP instance with budget b. Suppose we have a randomized algorithm ALG which outputs a solution SOL wherein each item j is added SOL with probability exactly equal to xj /(ck) for some constant c > 0.9 After running ALG, the expected usage of each budget i is bi /(ck); this follows directly from the budget constraint in the LP. This implies that after running ALG, we have only used a tiny fraction of the whole budget, in expectation. Thus, we may run ALG again on the remaining items to further improve the value/weight of SOL. Hence, an item that was previously not chosen, receives a “second chance” to be rounded up and included in the solution. The observation that only a tiny fraction of the budget is used can be made after running ALG for a second time as well. Hence, in principle, we can run ALG multiple times and we call the overarching approach a multiple chance algorithm. The analysis becomes rather delicate as we run for a large number of iterations in this manner. “Half the remaining distance” toward resolving the FKS conjecture and non-uniform attenuation approach. Our third contribution is in making significant progress on the well-known Conjecture 1 due to Füredi, Kahn and Seymour. To achieve this, we introduce a technique of non-uniform attenuation. A common framework for tackling k -CS-PP and related problems is random permutation followed by sampling via uniform attenuation: follow a random order π on the items and add each item j with probability αxj whenever it is safe, where xj is an optimal solution to an appropriate LP and α is the attenuation factor. Typically α is a parameter fixed in the analysis to get the best ratio (e.g., see the SKSP algorithm in Bansal et al. [12]). This method is called uniform attenuation, since all items share the same attenuation factor α. An alternative strategy used previously is that of weighted random permutations (see, e.g., Adamczyk et al. [2] and Baveja et al. [15]): instead of using a uniformly-random permutation, the algorithm “weights” the items and permutes them non-uniformly based on their xj values. We introduce a notion of non-uniform attenuation, which approaches the worst-case scenario in a different manner. We still stay within the regime of uniform permutations but will attenuate items non-uniformly, based on their xj values; a careful choice of attenuation function is very helpful here, as suggested by the optimization problem (12). This is a key ingredient in our improvement. 1.3 Other Related Work In this subsection, we list the related work not mentioned in previous sections and yet closely related to the problems we study. Note that packing programs, submodular maximization, and their applications to approximation algorithms have a vast literature. Our goal here is to list some papers in closely relevant areas and this is by no means an exhaustive list of references in each of these closely-aligned areas. For k -CS-PIP, related problems have been studied in discrepancy theory. In such problems, we have a k-column sparse LP and we want to round the fractional solution such that the violation (both above and below) of any constraint is minimized. This study started with the famous work of Beck and Fiala [17] and some of the previous work on k -CS-PIP (e.g., [61]) used techniques similar to Beck and Fiala. There has been a long line of work following Beck and Fiala, including [66, 7, 14, 55, 64, 43, 11, 53, 9]. One special case of k -CS-PIP is the k-set packing problem. Many works including [45, 5, 25, 18] studied this problem with [18] giving the best approximation of (k + 1)/2 + ǫ for this problem. Closely related to k -CS-PIP is the notion of column-restricted packing introduced by Kolliopoulos and Stein [47]. Many works have 9 Note that given an algorithm which adds each item with probability at least xj /(ck), we can use Monte Carlo simulation and attenuation techniques similar to [32, 2, 19] to get an algorithm which adds each item with probability essentially equal to xj /(ck). 6 studied this version of packing programs, including [30, 27, 16]. Similar to Bansal et al. [13], our algorithms also extend to submodular objective functions. In particular, we use tools and techniques from Calinescu et al. [22] and Chekuri et al. [32] for both k -CS-PIP and the UFP problem on trees. Monotone sub-modular function maximization subject to k-sparse constraints has been studied in the context of k-partition matroids, k-knapsacks, and the intersection of k partition matroids in many works including [38, 51, 69, 50]. Beyond the monotone case, there are several algorithms for the non-negative sub-modular maximization problem including [37, 28, 21, 36, 20]. Stochastic variants of PIPs have also been previously studied. Baveja et al. [15] considered the following stochastic setting of k-uniform hypergraph matching: the algorithm has to probe edge e to check its existence; each edge e is associated with a probability 0 < pe ≤ 1 with which it will be present (independently of other edges) on being probed; the task is to sequentially probe edges such that the expected total weight of matching obtained is maximized. The stochastic version of hypergraph matching can be viewed as a natural generalization of stochastic matching (e.g., Bansal et al. [12]) to hypergraphs. The work of [15] gave an (k + ǫ + o(1))–approximation algorithm for any given ǫ > 0 asymptotically for large k. Other work on stochastic variants of PIPs includes [33, 34, 54, 3, 42, 1, 41]. Later in this paper, we show yet another application of attenuation: UFP-TREES with unit demands. This problem is a more specific version of column-restricted packing problems mentioned previously. The Unsplittable Flow Problem in general graphs and its various specializations on different kinds of graphs has been extensively studied. Some of these works include [46, 65, 40, 48, 29, 8, 6, 49, 23, 10, 35]. 1.4 Outline In Section 2, we present a randomized rounding algorithm for k -CS-PIP using randomized alteration techniques. We analyze this algorithm to prove Theorem 1 and show an extension to submodular objectives. In Section 3, we apply second-chance techniques to SKSP. After analyzing this algorithm, we show how it can be extended to multiple chances, yielding the improved result of Theorem 2. In Section 4, we present an algorithm for hypergraph matching and analyze it to prove Theorem 3, making progress toward Conjecture 2 (and by extension Conjecture 1). In Section 5, we show how attenuation can lead to an improved contention resolution scheme for UFP-TREES, proving Theorem 4. We end with a brief conclusion and discussion of open problems in Section 6. All proofs from main section can be found in Section 7. Appendix A contains a few useful technical lemmas used in this paper while Appendix B gives a self-contained background on submodular functions. 2 k-Column Sparse Packing We describe a rounding algorithm for k -CS-PIP, which achieves the asymptotically optimal approximation ratio of (2k + o(k)) with respect to the strengthened LP shown in Bansal et al. [13] (see (7) in Section 2.1). Theorem 6 then develops a near-negative-correlation generalization of this result. Recall that we have a k-sparse matrix A ∈ [0, 1]m×n and a fractional solution x ∈ [0, 1]n such that A·x ≤ 1. Our goal is to obtain an integral solution X ∈ {0, 1}n (possibly random) such that A·X ≤ 1 and such that the expected value of the objective function w · X is “large”. (We will later extend this to the case where the objective function f (X) is monotone submodular.) At a very high level, our algorithm performs steps similar to the contention-resolution scheme defined by Chekuri et al. [32]; the main contribution is in the details10 . We first perform an independent-sampling step to obtain a random set R of variables; we 10 We would like to point out that the work of [13] also performs similar steps and fits into the framework of [32]. 7 then conduct randomized alterations to the set R to obtain a set of rounded variables that are feasible for the original program with probability 1. Note that the work of [13] uses deterministic alterations. Moving from deterministic alteration to careful randomized alteration, as well as using a much-less aggressive uniform attenuation in the initial independent sampling, yield us the optimal bound. 2.1 Algorithm Before describing the algorithm, we review some useful notations and concepts, some of which were introduced in [13]. For a row i of A and ℓ = k1/3 , let big(i) := {j : aij > 1/2}, med(i) := {j : 1/ℓ ≤ aij ≤ 1/2}, and tiny(i) := {j : 0 < aij < 1/ℓ}, which denote the set of big, medium, and tiny items with respect to constraint i. For a given randomly sampled set R and an item j ∈ R, we have three kinds of blocking events for j. Blocking events occur when a set of items cannot all be rounded up without violating some constraint. In other words, theses events, with probability 1, prevent j from being rounded up. We partition the blocking events into the following three types: • BB(j): There exists some constraint i with aij > 0 and an item j ′ 6= j such that j ′ ∈ big(i) ∩ R. • MB(j): There exists some constraint i with med(i) ∋ j such that | med(i) ∩ R| ≥ 3. X • TB(j): There exists some constraint i with tiny(i) ∋ j such that aij ′ > 1 − aij . j ′ 6=j: j ′ ∈(med(i)∪tiny(i))∩R Informally, we refer to the above three blocking events as big, medium and tiny blocking events for j with respect to R. The main algorithm of Bansal et al. [13]. As briefly mentioned in Section 1.1, Bansal et al. add certain valid constraints on big items to the natural LP relaxation in (3) as follows: X max{w · x s.t. A · x ≤ 1and ∀i ∈ [m] xj ≤ 1, x ∈ [0, 1]n }where A[j] = aj (7) j∈big(i) Algorithm 1, BKNS, gives a formal description of the algorithm of Bansal et al. [13], in which they set α = 1. Algorithm 1: BKNS(α) 1 2 3 4 Sampling: Sample each item j independently with probability (αxj )/k and let R0 be the set of sampled items. Discarding low-probability events: Remove an item j from R0 if either a medium or tiny blocking event occurs for j with respect to R0 . Let R1 ⊆ R0 be the set of items not removed. Deterministic alteration: Remove an item j from R1 if a big blocking event occurs for j with respect to R1 . Let RF ⊆ R1 be the set of items not removed; return RF . Theorem 5 (Bansal et al. [13]). By choosing α = 1, Algorithm 1 yields a randomized ek+o(k)-approximation for k -CS-PIP. 8 Our algorithm for k -CS-PIP via randomized alterations. Our pre-processing is similar to BKNS with the crucial difference that α ≫ 1 (but not too large), i.e., we do not attenuate too aggressively; furthermore, our alteration step is quite different. Let [n] = {1, 2, . . . , n} denote the set of items. We first sample each item independently using an appropriate product distribution over the items (as mentioned above, we crucially use a different value for α than BKNS). Let R0 denote the set of sampled items. We remove items j from R0 for which either a medium or tiny blocking event occurs to obtain a set R1 . We next perform a randomized alteration, as opposed to a deterministic alteration such as in line 3 of BKNS. We then randomly and appropriately shrink R1 to obtain the final set RF . We now informally describe our randomized alteration step. We construct a directed graph G = (R1 , E) from the constraints as follows. For every item j ∈ R1 , we create a vertex. We create a directed edge from item j to item j ′ 6= j in G iff j ′ causes a big blocking event for j (i.e., there exists a constraint i where j has a non-zero coefficient and j ′ is in big(i)). We claim that the expected degree of every vertex in this graph constructed with R1 is at most α = ω(1). If any vertex j has a degree greater than d := α + α2/3 , we will remove j from R1 . Hence we now have a directed graph with every vertex having a degree of at most d. We claim that we can color the undirected version of this directed graph with at most 2d + 1 colors. We choose one of the colors c in [2d + 1] uniformly at random and add all vertices of color c into RF . Algorithm 2 gives a formal description of our approach. Algorithm 2: The Algorithm for k -CS-PIP 1 2 3 Sampling: Sample each item j independently with probability αxj /k (where, say, α = log k) and let R0 be the set of sampled items. Discarding low-probability events: Remove an item j from R0 if either a medium or tiny blocking event occurs for j with respect to R0 . Let R1 be the set of items not removed. Randomized alteration: (a) Create a directed graph: For every item in R1 , create a vertex in graph G. Add a directed edge from item j to item j ′ if there exists a constraint i such that aij > 0 and aij ′ ≥ 1/2. (b) Removing anomalous vertices: For every vertex v in G, if the out-degree of v is greater than d := α + α2/3 , call v anomalous. Remove all anomalous vertices from G to obtain G′ and let the items corresponding to the remaining vertices in G′ be R2 . (c) Coloring G′ : Assign a coloring χ to the vertices of G′ using 2d + 1 colors as described in the text such that for any edge e (ignoring the direction), both end points of e receive different colors. 4 (d) Choosing an independent set: Choose a number c ∈ [2d + 1] uniformly at random. Add all vertices v from G′ into RF such that χ(v) = c. Return RF . Example. Before moving to the analysis, we will show an example of how the randomized alteration (i.e., lines 3(a-d) of Algorithm 2) works. We will illustrate this on the integrality gap example considered in [13]. In this example, we have n = 2k − 1 items and m = 2k − 1 constraints. The weights of all items are 1. For some 0 < ǫ ≪ 1/(nk), the matrix A is defined as follows. ∀i, j ∈ [2k − 1] we have,   if i = j 1 aij := ǫ if j ∈ {i + 1, i + 2, . . . , i + k − 1(modn)}   0 otherwise 9 As noted in [13], setting xj = (1 − kǫ) for all j ∈ [n] is a feasible LP solution, while the optimal integral solution has value 1. After running line 1 of the algorithm, each item j is selected with probability (1 − o(1))α/k independently. For simplicity, we will assume that there are no medium or tiny blocking events for every j (these only contribute to the lower-order terms). Note that in expectation the total number of chosen items will be approximately 2α; with high probability, the total number of vertices in the graph will be n1 := 2α+o(α). Let b1 , b2 , . . . , bn1 denote the set of items in this graph. The directed graph contains the edge (bi , bj ) for all distinct i, j; for simplicity, assume that the graph has no anomalous vertices. Since the undirected counterpart of this graph is a complete graph, every vertex will be assigned a unique color; thus the solution output will have exactly one vertex with probability 1 − o(1). 2.2 Analysis We prove the following main theorem using Algorithm 2 with α = log(k). Theorem 1. There exists a randomized rounding algorithm for k -CS-PIP with approximation ratio at most 2k + o(k) for linear objectives. We will divide the analysis into three parts. At a high-level the three parts prove the following. • Part 1 (Proved in Lemma 1). For directed graphs with maximum out-degree at most d, there exists a coloring χ and a corresponding algorithm such that the number of colors used, |χ|, is at most 2d + 1. • Part 2 (Proved in Lemma 2). For any item j ∈ R1 , the event that the corresponding vertex in G has an out-degree larger than d occurs with probability at most o(1). This implies that conditional on j ∈ R1 , the probability that j is present in G′ is 1 − o(1). • Part 3 (Proved in Lemma 3). For each item j ∈ R0 , either a medium or a tiny blocking event occurs with probability at most o(1) (again, for our choice α = log(k)). This implies that for each j ∈ R0 , it will be added to R1 with probability 1 − o(1). We assume the following lemmas which are proven in Section 7.1. Lemma 1. Given a directed graph G = (V, E) with maximum out-degree at most d, there is a polynomialtime algorithm that finds a coloring χ of G’s undirected version such that |χ|, the number of colors used by χ, is at most 2d + 1. Lemma 2. For any item j, the probability – conditional on the event “j ∈ R1 ” – that j is selected into R2 (i.e., j ∈ R2 ) is 1 − o(1), for α = log(k). Lemma 3. For each item j, either a medium or a tiny blocking event occurs with probability – conditional on the event “j ∈ R0 ” – of at most Θ(α4/3 /k1/3 ) = o(1), for α = log k. We can now prove the main theorem, Theorem 1. Proof. First we show that RF is feasible for our original IP. We have the following observations about Algorithm 2: (i) from line 2, “Discarding low-probability events”, we have that no item in RF can be blocked by either medium or tiny blocking events; (ii) from the “Randomized alteration” steps in line 3, we have that no item in RF has any neighbor in G′ that is also included in RF . This implies that no item in RF can be blocked by any big blocking events. Putting together the two observations implies that RF is a feasible solution to our IP. 10 We now show that the probability of j being in RF can be calculated as follows. Pr[j ∈ RF ] = Pr[j ∈ R0 ] · Pr[j ∈ R1 | j ∈ R0 ] · Pr[j ∈ R2 | j ∈ R1 ] · Pr[j ∈ RF | j ∈ R2 ] 1 αxj · (1 − o(1)) · (1 − o(1)) · ≥ k 2α + 2α2/3 + 1 xj = 2k(1 + o(1)) The first inequality is due to the following. From the sampling step we have that Pr[j ∈ R0 ] = αxj /k. From Lemma 3 we have that Pr[j ∈ R1 | j ∈ R0 ] = 1 − o(1). Lemma 2 implies that Pr[j ∈ R2 | j ∈ R1 ] = 1 − o(1). Finally from Lemma 1 we have that the total number of colors needed for items in R2 is at most 2d + 1, and hence the probability of picking j’s color class is 1/(2d + 1). Thus, Pr[j ∈ RF | j ∈ R2 ] = 1/(2α + 2α2/3 + 1) (recall that d := α + α2/3 ). Nearly-negative correlation. The natural approach to proving Lemma 1 can introduce substantial positive correlation among the items included in RF . However, by slightly modifying the algorithm, we can obtain the following Theorem, which induces nearly-negative correlation in the upper direction among the items. Theorem 6. Given any constant ǫ ∈ (0, 1), there is an efficient randomized algorithm for rounding a fractional solution within the k -CS-PIP polytope, such that 1. For all items j ∈ [n], Pr[j ∈ RF ] ≥ xj 2k(1+o(1)) . 2. For any t ∈ [n] and any t-sized subset {v1 , v2 , . . . , vt } of items in [n], we have (with d = α + α2/3 ∼ log k as above) Pr[v1 ∈ RF ∧ v2 ∈ RF ∧ . . . ∧ vt ∈ RF ] ≤ (2dǫ )t−1 · t Y xvj . 2k j=1 2.3 Extension to Submodular Objectives As described in the preliminaries, we can extend certain contention-resolution schemes to submodular objectives using prior work. We will now show that the above rounding scheme can be extended to submodular objectives; in particular, we will use the following definition and theorem from Chekuri et al. [32] 11 . Definition 1. (bc-BALANCED MONOTONE CR SCHEMES [32]) Let b, c ∈ [0, 1] and let N := [n] be a set of items. A bc-balanced CR scheme π for PI (where PI denotes the convex relaxation of the set of feasible integral solutions I ⊆ 2N ) is a procedure which for every y ∈ bPI and A ⊆ N , returns a set πy (A) ⊆ A ∩ support(y) with the following properties: (a) For all A ⊆ N, y ∈ bPI , we have that πy ∈ I with probability 1. (b) For all i ∈ support(y), y ∈ bPI , we have that Pr[i ∈ πy (R(y)) | i ∈ R(y)] ≥ c, where R(y) is the random set obtained by including every item i ∈ N independently with probability yi . (c) Whenever i ∈ A1 ⊆ A2 , we have that Pr[i ∈ πy (A1 )] ≥ Pr[i ∈ πy (A2 )]. 11 Specifically, Definition 1.2 and Theorem 1.5 from [32]. 11 Theorem 7 (Chekuri et al. [32]). Suppose PI admits a bc-balanced monotone CR scheme and ηf is the approximation ratio for max{F (x) : x ∈ PI ∩ {0, 1}n } (here, F (x) is the multi-linear extension of the sub-modular function f ). Then there exists a randomized algorithm which gives an expected 1/(bcηf )approximation to maxS∈I f (S), when f is a non-negative submodular function. For the case of monotone sub-modular functions, we have the optimal result ηf = 1−1/e (Vondrák [68]). For non-monotone sub-modular functions, the best-known algorithms have ηf ≥ 0.372 due to Ene and Nguyen [36] and more recently ηf ≥ 0.385 due to Buchbinder and Feldman [20] (it is not known if these are tight: the best-known upper bound is ηf ≤ 0.478 due to Oveis Gharan and Vondrák [56]). We will show that Algorithm 2 is a 1/(2k + o(k))-balanced monotone CR scheme, for some b, c such that bc = 1/(2k + o(k)). Hence, from Theorem 7 we have a (2k + o(k))/ηf -approximation algorithm for k -CS-PIP with sub-modular objectives. This yields Corollary 1. For ease of reading, we will first re-state notations used in Definition 1 in the form stated in the previous sub-section. The polytope PI represents the k -CS-PIP polytope defined by Eq. 7. The vector y is defined as yi := αxi /k which is used in the Sampling step of the algorithm (i.e., Line 1). The scheme πy is the procedure defined by lines 1, 2, 3 of the algorithm. In other words, this procedure takes a subset A of items and returns a feasible solution with probability 1 (and hence satisfying property (a) in the definition). Our goal then is to show that it further satisfies properties (b) and (c). The set R(y) corresponds to the set R0 , where every item i is included into R0 with probability yi , independently. From the sampling step of the algorithm, we have that b = α/k, since each item i is included in the set R(y) with probability yi := xi α/k and x ∈ PI and hence y ∈ (α/k)PI . From the alteration steps we have that c = (1 − o(1))/(2α + o(α)), since for any item i, we have Pr[i ∈ RF | i ∈ R0 ] ≥ (1 − o(1))/(2α + o(α)). Thus, πy satisfies property (b). Now we will show that the rounding scheme πy satisfies property (c) in Definition 1. Let A1 and A2 be two arbitrary subsets such that A1 ⊆ A2 . Consider a j ∈ A1 . We will now prove the following. Pr[j ∈ RF | R0 = A1 ] ≥ Pr[j ∈ RF | R0 = A2 ] Note that for i ∈ {1, 2} we have, Pr[j ∈ RF | R0 = Ai ] = Pr[j ∈ RF | j ∈ R2 , R0 = Ai ] Pr[j ∈ R2 | R0 = Ai ] (8) For both i = 1 and i = 2, the first term in the RHS of Eq. (8) (i.e., Pr[j ∈ RF | j ∈ R2 , R0 = Ai ]) is same and is equal to 1/(2d + 1). Note that the second term in the RHS of Eq. (8) (i.e., Pr[j ∈ R2 | R0 = Ai ]) can be rewritten as Pr[j ∈ R2 | j ∈ R0 , Ro = Ai ] since j ∈ A1 and R0 = Ai for i ∈ {1, 2}. From lines 2 and 3(b) of the algorithm we have that the event j ∈ R2 conditioned on j ∈ R0 occurs if and only if: (i) no medium or tiny blocking events occurred for j. (ii) vertex j did not correspond to an anomalous vertex in G (one with out-degree greater than d := α + α2/3 ). Both (i) and (ii) are monotonically decreasing in the set R0 (i.e., if it holds for R0 = A2 then it holds for R0 = A1 ). Hence we have Pr[j ∈ R2 | R0 = A1 ] ≥ Pr[j ∈ R2 | R0 = A2 ]. 3 The Stochastic k-Set Packing Problem Consider the stochastic k-set packing problem defined in the introduction. We start with a second-chancebased algorithm yielding an improved ratio of 8k/5 + o(k). We then improve this to k + o(k) via multiple 12 chances. Recall that if we probe an item j, we have to add it irrevocably, as is standard in such stochasticoptimization problems; thus, we do not get multiple opportunities to examine j. Let x be an optimal solution to the benchmark LP (4) and C(j) be the set of constraints that j participates in. Bansal et al. [12] presented Algorithm 3, SKSP(α). They show that SKSP(α) will add each item j with probability at least β (xj /k) where β ≥ α(1 − α/2). By choosing α = 1, SKSP(α) yields a ratio of 2k.12 Algorithm 3: SKSP(α) [12] 1 2 3 4 Let R denote the set of chosen items which starts out as an empty set. For each j ∈ [n], generate an independent Bernoulli random variable Yj with mean αxj /k. Choose a uniformly random permutation π over [n] and follow π to check each item j one-by-one: add j to R if and only if Yj = 1 and j is safe (i.e., each resource i ∈ C(j) has at least one unit of budget available); otherwise skip j. Return R as the set of chosen items. At a high level, our second-chance-based algorithm proceeds as follows with parameters {α1 , β1 , α2 } to be chosen later. During the first chance, we set α = α1 and run SKSP(α1 ). Let E1,j denote the event that j is added to R in this first chance. From the analysis in [12], we have that Pr[E1,j ] ≥ (xj /k) α1 (1 − α1 /2). By applying simulation-based attenuation techniques, we can ensure that each item j is added to R in the first chance with probability exactly equal to β1 xj /k for a certain β1 ≤ α1 (1 − α1 /2) of our choice.13 In other words, suppose we obtain an estimate Ê1,j := Pr[E1,j ]. When running the original randomized algorithm, whenever j can be added to R in the first chance, instead of adding it with probability 1, we add it with probability ((xj /k) β1 )/Ê1,j . In the second chance, we set α = α2 and modify SKSP(α2 ) as follows. We generate an independent Bernoulli random variable Y2,j with mean α2 xj /k for each j; let Y1,j denote the Bernoulli random variable from the first chance. Proceeding in a uniformly random order π2 , we add j to R if and only if j is safe, Y1,j = 0 and Y2,j = 1. Algorithm 4, SKSP(α1 , β1 , α2 ), gives a formal description. Algorithm 4: SKSP(α1 , β1 , α2 ) 1 2 3 4 Initialize R as the empty set. The first chance: Run SKSP(α1 ) with simulation-based attenuation such that Pr[E1,j ] = β1 xj /k for each j ∈ [n], with β1 ≤ α1 (1 − α1 /2). R now denotes the set of variables chosen in this chance. The second chance: Generate an independent Bernoulli random variable Y2,j with mean α2 xj /k for each j. Follow a uniformly random order π2 over [n] to check each item j one-by-one: add j to R if and only if j is safe, Y1,j = 0, and Y2,j = 1; otherwise, skip it. Return R as the set of chosen items. Lemma 4 lower bounds the probability that an item gets added in the second chance. For each j, let E2,j be the event that j is added to R in the second chance. 12 The terminology used in [12] is actually 1/α; however, we find the inverted notation α more natural for our calculations. See footnote in Section 1.2. Sampling introduces some small sampling error, but this can be made into a lower-order term with high probability. We thus assume for simplicity that such simulation-based approaches give us exact results. 13 13 Lemma 4. After running SKSP(α1 , β1 , α2 ) on an optimal solution x to the benchmark LP (4), we have   α x x Pr[E2,j ] ≥ kj α2 1 − 1k j − β1 − α22 We can use Lemma 4 to show that we get an approximation ratio of 8k/5. Observe that the events E1,j and E2,j are mutually exclusive. Let Ej be the event that j has been added to R after the two chances. Then, 5x x2 by choosing α1 = 1, β1 = 1/2, and α2 = 1/2, we have Pr[Ej ] = Pr[E1,j ]+Pr[E2,j ] = 8kj − 2kj2 . From the Linearity of Expectation, we get that the total expected weight of the solution is at least (5/8k − o(k))w.x. Theorem 8. By choosing α1 = 1, β1 = 1/2, and α2 = 1/2, SKSP(α1 , β1 , α2 ) achieves a ratio of for SKSP. 8k 5 +o(k) 3.1 Extension to T Chances Intuitively, we can further improve the ratio by performing a third-chance probing and beyond. We present a natural generalization of (α1 , β1 , α2 ) to ({αt , βt | t ∈ [T ]}) with T chances, where {αt , βt | t ∈ [T ]} are parameters to be fixed later. Note that (α1 , β1 , α2 ) is the special case wherein T = 2. During each chance t ≤ T , we generate an independent Bernoulli random variable Yt,j with mean αt xj /k for each j. Then we follow a uniform random order πt over [n] to check each item j one-by-one: we add j to R if and only if j is safe, Yt′ ,j = 0 for all t′ < t and Yt,j = 1; otherwise we skip it. Suppose for a chance t, we have that each j is added to R with probability at least βt xj /k. As before, we can apply simulation-based attenuation to ensure that each j is added to R in chance t with probability exactly equal to βt xj /k. Notice that to achieve this goal we need to simulate our algorithm over all previous chances up to the current one t. Algorithm 5, SKSP({αt , βt |t ∈ [T ]}), gives a formal description of the algorithm. Notice that during the last chance T , we do not need to perform simulation-based attenuation. For the sake of clarity in presentation, we still describe it in the algorithm description. Algorithm 5: SKSP({αt , βt |t ∈ [T ]}) 1 2 3 4 5 Initialize R as the empty set. for t=1, 2, . . . , T do Generate an independent Bernoulli random variable Yt,j with mean αt xj /k for each j. Follow a uniform random order πt over [n] to check each item j one-by-one: add j to R if and only if j is safe, Yt′ ,j = 0 for all t′ < t, and Yt,j = 1; otherwise, skip it. Apply simulation-based attenuation such that each j is added to R in the tth chance with probability equal to βt xj /k. Return R as the set of chosen items. ′ be the event that j is added to R in the tth chance before line 3 of the algorithm For each item j, let Et,j for t (i.e., before the start of the tth iteration of the loop). Lemma 5 lower bounds the probabilities of these events. Lemma 5. After running ({αt , βt | t ∈ [T ]}) on an optimal solution x to the benchmark LP (4), we have P     P αt t′ <t αt′ αt ′ ] ≥ xj α 1 − ′ Pr[Et,j − β − ′ t t <t t k 2 k 14 Combining Lemma 5 and simulation-based attenuation, we have that for any given {αt′ | t′ ≤ t} and in chance t with probability equal to βt xj /k for any {βt′ | t′ < t}, each  to RP P item j is added βt ≤ αt 1 − t′ <t βt′ − αt /2 − αt t′ <t αt′ /k. For each j, let Ej,t be the event that j is added to R 5, we have that the in chance t and Ej the event that j is added to R after T chances. From Algorithm P P events {Ej,t | t ≤ T } are mutually exclusive. Thus, Pr[Ej ] = t≤T Pr[Ej,t ] = t≤T βt xj /k. Therefore, to maximize the final ratio, we will solve the following optimization problem. max X t∈[T ] βt s.t. βt ≤ αt 1− X t′ <t αt βt′ − 2 ! − αt P t′ <t αt′ k ∀t ∈ [T ], αt ≥ 0 ∀t ∈ [T ] (9) Consider a simplified version of maximization program (9) by ignoring the O(1/k) term as follows. ! X X αt ∀t ∈ [T ], αt ≥ 0 ∀t ∈ [T ] (10) max βt , s.t. βt ≤ αt 1 − βt′ − 2 ′ t <t t∈[T ] Lemma 6. An optimal solution to the program (10) is βt∗ = X X 1 (1 − βt∗′ )2 , ∀t ≥ 1, α∗t = 1 − βt∗′ , ∀t ≥ 1 2 ′ ′ t <t t <t where β0∗ = 0, 0 ≤ α∗t ≤ 1 for all t ≥ 1 and limT →∞ PT ∗ t=1 βt = 1. Theorem 9. Let T P be some slowly-growing function of k, e.g., T = log k. For each t ∈ [T ], set ᾱt = α∗t ( t′ <t α∗t′ ) ∗ ∗ . Then we have (1) {ᾱt , β̄t | t ∈ [T ]} is feasible to the program (9) and (2) α k Pt , β̄t = βt − t∈[T ] β̄t = 1 + o(1) where o(1) goes to 0 when k goes to infinity. Thus, SKSP({ᾱt , β̄t |t ∈ [T ]}) achieves a ratio of k + o(k) for SKSP. 4 Hypergraph Matching In this section, we give a non-uniform attenuation approach to the hypergraph matching problem which leads to improved competitive ratios. Additionally as stated in the introduction, this takes us “half the remaining distance” towards resolving the stronger Conjecture 2. Consider a hypergraph H = (V, E). Assume each e ∈ E has cardinality |e| = ke . Let x = {xe } be an optimal solution to the LP (5). We start with a warm-up algorithm due to Bansal et al. [12]14 . Algorithm 6, HM(α), summarizes their approach. Lemma 7. Each edge e is added to R with probability at least xe ke +1 in HM(α) with α = 1. It can be shown that in HM(α = 1), the worst case occurs for the edges e with xe ≤ ǫ ≈ 0 (henceforth referred to as “tiny” edges). In contrast, for the edges with xe ≫ ǫ (henceforth referred to as “large” edges), the ratio is much higher than the worst case bound. This motivates us to balance the ratios among tiny and large edges. Hence, we modify Algorithm 6 as follows: we generate an independent Bernoulli random variable Ye with mean g(xe ) for each e, where g : [0, 1] → [0, 1] and g(x)/x is decreasing over [0, 1]. Algorithm 7 gives a formal description of this modified algorithm. 14 Similar to our algorithm for SKSP, we use the notation α while [12] use 1/α. 15 Algorithm 6: HM(α) 1 2 3 4 Initialize R to be the empty set. We will add edges to this set during the algorithm and return it at the end as the matching. For each e ∈ E, generate an independent Bernoulli random variable Ye with mean αxe . Choose a random permutation π over E and follow π to check each edge one-by-one: add e to R if and only if Ye = 1 and e is safe (i.e., none of the vertices in e are matched); otherwise skip it. Return R as the matching. Algorithm 7: HM(g) 1 2 3 4 Initialize R to be the empty set. For each e ∈ E, generate an independent Bernoulli random variable Ye with mean g(xe ). Choose a random permutation π over E and follow π to consider each edge one by one: add e to R if Ye = 1 and e is safe; otherwise skip it. Return R as the matching Observe that HM(α) is the special case wherein g(xe ) = αxe . We now consider the task of finding the optimal g such that the resultant ratio achieved by HM(g) is maximized. Consider a given e with xe = x. ′ 6= e, we say e′ is a neighbor of e (denoted by e′ ∼ e) if e′ ∋ v for some v ∈ e. From the LP (5), For any eP we have e′ ∼e xe′ ≤ ke (1 − x). Let Ee be the event that e is added to R. By applying an analysis similar to the proof of Lemma 7, we get the probability of Ee is at least R1Q (11) Pr[Ee ] ≥ g(x) 0 e′ ∼e (1 − tg(xe′ ))dt. Therefore our task of finding an optimal g to maximize the r.h.s. of (11) is equivalent to finding maxg F(g), where F(g) is defined in equation (12). i h . × r(x) (12) F(g) = minx∈[0,1] g(x) x In equation (12), r(x) is defined as  R1Q P . r(x) = min 0 e′ ∼e 1 − tg(xe′ ) dt, where e′ ∼e xe′ ≤ ke (1 − x), xe′ ∈ [0, 1], ∀e′ Lemma 8. By choosing g(x) = x(1 − x2 ), we have that the minimum value of F(g) in Eq. (12) is F(g) = 1 ke (1 − exp(−ke )). We now prove the main result, Theorem 3. Proof. Consider HM(g) as shown in Algorithm 7 with g(x) = x(1 − x/2). Let R be the random matching returned. From Lemma 8, we have that each e will be added to R with probability at least xe F(g) = xe ke (1 − exp(−ke )). 16 5 More Applications In this section, we briefly describe how a simple simulation-based attenuation can lead to improved contention resolution schemes for UFP-TREES with unit demands. This version of the problem was studied by Chekuri et al. [30] where they gave a 4-approximation for the linear objective case. They also described a simple randomized algorithm that obtains a 27-approximation15 . Later, Chekuri et al. [32] developed the machinery of contention resolution schemes, through which they extended it to a 27/ηf -approximation algorithm for non-negative submodular objective functions (where ηf denotes the approximation ratio for maximizing non-negative submodular functions16 ). We show that using simple attenuation ideas can further improve this 27/ηf -approximation to an 8.15/ηf -approximation. We achieve this by improving the 1/27-balanced CR scheme17 to a 1/8.15-balanced CR scheme and hence, from Theorem 1.5 of [32], the approximation ratio follows. Consider the natural packing P LP relaxation. Associate a variable xi with every demand pair. Our constraint set is: for every edge e, i:e∈Pi xi ≤ ue , where ue is the capacity of e. Our algorithm (formally described in Algorithm 8) proceeds similar to the one described in [30], except at line 3, where we use our attenuation ideas. Algorithm 8: Improved contention resolution Scheme UFP-TREES with unit demands 1 2 3 4 Root the tree T arbitrarily. Let LCA(si , ti ) denote the least common ancestor of si and ti . Construct a random set R of demand pairs, by including each pair in R independently with probability αxi . Consider the demand pairs in increasing distance of their LCA from the root. Let Rfinal denote the set of pairs included in the rounded integral solution. For every demand pair i, simulate the run of the algorithm from beginning (i.e., produce many samples of R and separately run the algorithm on these samples) to obtain the estimate ηi of the probability of i to be safe (i.e., none of the edges in the path has exhausted capacities). Suppose i is safe, add it to Rfinal with probability β/ηi . Return Rfinal as the set of demands chosen from routing. Analysis. For the most part, the analysis is similar to the exposition in Chekuri et al. [32]. We will highlight the part where attenuation immediately leads to improved bounds. Consider a fixed pair (si∗ , ti∗ ) and let ℓ := LCA(si∗ , ti∗ ) in T . Let P and P ′ denote the unique path in the tree from ℓ to si∗ and ℓ to ti∗ respectively. As in [32], we will upper bound the probability of i∗ being unsafe due to path P and a symmetric argument holds for P ′ . Let e1 , e2 , . . . , eλ be the edges in P from ℓ to si∗ . Let Ej denote the event that i∗ is not safe to be added in line 3 of Algorithm 8, because of overflow at edge ej . Note that for j > h and uej ≥ ueh , event Ej implies Eh and hence Pr[Ej ] ≤ Pr[Eh ]. Note, this argument does not change due to attenuation since the demands are processed in increasing order of the depth and any chosen demand pair using edge ej also has to use eh up until the time i∗ is considered. Thus, we can make a simplifying assumption similar to [32] and consider a strictly decreasing sequence of capacities ue1 > ue2 > . . . > ueλ ≥ 1. Let Sj denote the set of demand pairs that use edge ej . The following steps is the part where our analysis differs from [32] due to the introduction of attenuation. Define, β := 1 − 2αe/(1 − αe) and γ := αβ. Note that without attenuation, we have ηi ≥ β for all i from the analysis in [32]. 15 This can be obtained by maximizing over all 0 < b < 1/3e in Lemma 4.19 of [32], which yields approximately 1/27. See the section on k -CS-PIP for a discussion on various values of ηf known. 17 See the section on extension to sub-modular objectives in k -CS-PIP for defintion of a balanced CR scheme. 16 17 Let Ej′ denote the event that at least uej demand pairs Pout of Sj are included in the final solution. Note ′ that Pr[Ej ] ≤ Pr[Ej ]. From the LP constraints we have i∈Sj xi ≤ uej . LetPXi denote the indicator random variable for the following event: {i ∈ R ∧ i ∈ Rfinal }. We define X := i∈Sj Xi . Note that, event Ej′ happens if and only if X ≥ uej and hence we have, Pr[Ej′ ] = Pr[X ≥ uej ]. Additionally, we have that the Xi ’s are “cylindrically negatively correlated” and hence we can apply the Chernoff-Hoeffding bounds due to Panconesi and Srinivasan [57]. Observe that E[X] ≤ γuej (since each i is included in R independently with αxi and then included in Rfinal with probability exactly β) and for 1 + δ = 1/γ, we have Pr[X ≥ uej ] ≤ (e/(1 + δ))(1+δ)µ ≤ (γe)uej . Hence, taking a union bound over all the edges in the path, we have the probability of i∗ being unsafe due to an edge in P to be at P∞ most q=1 (γe)ℓ = (γe)/(1 − γe) (We used the fact that ue1 > ue2 > . . . > ueλ ≥ 1). Combining the symmetric analysis for the other path P ′ , we have the probability of i∗ being unsafe to be at most 2γe/(1 − γe). Note that we used the fact that γe < 1 in the geometric series. Additionally, since γ ≤ 1, we have that 2γe/(1 − γe) ≤ 2αe/(1 − αe). Hence, using ηi ≥ β is justified. Now to get the claimed approximation ratio, we solve the following maximization problem: max {α · (1 − 2γe/(1 − γe)) : β = 1 − 2αe/(1 − αe), γ = α ∗ β, 0 ≤ γe < 1/3} 0≤α≤1 which yields a value of 1/8.15. 6 Conclusion and Open Problems In this work, we described two unifying ideas, namely non-uniform attenuation and multiple-chance probing to get bounds matching integrality gap (up to lower-order terms) for the k -CS-PIP and its stochastic counterpart SKSP. We generalized the conjecture due to Füredi et al. [39] (FKS conjecture) and went “halfway” toward resolving this generalized form using our ideas. Finally, we showed that we can improve the contention resolution schemes for UFP-TREES with unit demands. Our algorithms for k -CS-PIP can be extended to non-negative submodular objectives via the machinery developed in Chekuri et al. [32] and the improved contention resolution scheme for UFP-TREES with unit demands leads to improved approximation ratio for submodular objectives via the same machinery. This work leaves a few open directions. The first concrete problem is to completely resolve the FKS conjecture and its generalization. We believe non-uniform attenuation and multiple-chances combined with the primal-dual techniques from [39] could give the machinery needed to achieve this. Other open directions are less well-formed. One is to obtain stronger LP relaxations for the k -CS-PIP and its stochastic counterpart SKSP such that the integrality gap is reduced. The other is to consider improvements to related packing programs, such as column-restricted packing programs or general packing programs. 7 Proofs 7.1 Proofs for Section 2 (k -CS-PIP) We will now provide the missing proofs for k -CS-PIP. Proof of Lemma 1. We will prove this Lemma by giving a coloring algorithm that uses at most 2d + 1 colors and prove its correctness. Recall that we have a directed graph such that the maximum out-degree ∆ ≤ d. The algorithm is a simple greedy algorithm, which first picks the vertex with minimum total 18 degree (i.e., sum of in-degree plus out-degree). It then removes this vertex from the graph and recursively colors the sub-problem. Finally, it assigns a color to this picked vertex not assigned to any of its neighbors. Algorithm 9 describes the algorithm formally. Algorithm 9: Greedy algorithm to color bounded degree directed graph Color-Directed-Graph(G, V, d, χ) 1 if V = φ then 2 return χ 3 4 5 6 7 else Let vmin denote the vertex with minimum total degree. χ = Color-Directed-Graph(G, V \ {vmin }, d, χ). Pick the smallest color c ∈ [2d + 1] that is not used to color any of the neighbors of vmin . Let χ(vmin ) = c. return χ We will now prove the correctness of the above algorithm. In particular, we need to show that in every recursive call of the function, there is always a color c ∈ [2d + 1] such that the assignment in line 6 of the algorithm is feasible. We prove this via induction on number of vertices in the graph G. Base Case: The base case is the first iteration is when the number of vertices is 1. In this case, the statement is trivially true since vmin has no neighbors. Inductive Case: We have that ∆ ≤ d for every recursive call. Hence, the sum of total degree of all vertices in the graph is 2nd (Each edge contributes 2 towards the total degree and there are nd edges). Hence, the average total degree is 2d. This implies that the minimum total degree in the graph is at most 2d. Hence, the vertex vmin has a total degree of at most 2d. From inductive hypothesis we have that V \ {vmin } can be colored with at most 2d + 1 colors. Hence, there exists a color c ∈ [2d + 1], such that χ(vmin ) = c is a valid coloring (since vmin has at most 2d neighbors). Proof of Lemma 2. Consider an item j ∈ R1 . We want to show that the Pr[δj > α + α2/3 ] ≤ o(1), where δj represents the out-degree of j in the directed graph G. Recall that from the construction of graph G, we have a directed edge from item j to item j ′ if and only there is a constraint i where aij ′ > 1/2 and aij > 0. For sake of simplicity, let {1, 2, . . . , Nj } denote the set of out-neighbors of j in graph G and let {X1 , X2 , . . . , XNj } denote the corresponding indicator random variable for them being included in R1 . Hence, for evey i ∈ [Nj ] we have E[Xi ] = α(1 − o(1))xi /k. From the strengthened constraints in LP (7) we have that E[δj ] = E[X1 + X2 + . . . + XNj ] ≤ α(1 − o(1)). Hence we have 1/3 ) Pr[δj > α + α2/3 ] ≤ Pr[δj > E[δj ] + α2/3 ] ≤ e−Ω(α = o(1) The last inequality is from the Chernoff bounds, while the last equality is true for α = log(k). Proof of Lemma 3. Consider the medium blocking event MB(j). Let i be a constraint that causes MB(j) and let j1 , j2 6= j be the two other variables such that j1 , j2 ∈ med(i). Denote, Xj , Xj1 and Xj2 to be the indicators that j ∈ R0 , j1 ∈ R0 , j2 ∈ R0 respectively. We know that aij xj + aij1 xj1 + aij2 xj2 ≤ 1 and since j, j1 , j2 ∈ med(i), we have xj + xj1 + xj2 ≤ ℓ for some constant value of ℓ. The probability that scenario MB is “bad” is if Xj + Xj1 + Xj2 ≥ 2. Note that E[Xj + Xj1 + Xj2 ] ≤ αℓ k . 19 Hence, using the the Chernoff bounds in the form denoted in Theorem 10 of Appendix, we have  2 2 α ℓ Pr[Xj1 + Xj2 ≥ 2 Xj1 = 1] = Pr[Xj1 + Xj2 ≥ 2] ≤ O k2 Note that the first equality is due to the fact that these variables are independent. Using a union bound 2 2 over the k constraints j appears in, the total probability of the “bad” event is at most O( α kℓ ). And since α = O(log k) and ℓ = Θ((k/α)1/3 ), this value is o(1). For scenario TB(j) we will do the following. If j is tiny, ourP“bad” event for constraint i is that the total size of remaining items to be at least 1 − 1/ℓ. We know that E[ h∈R\{j} Aih Xih ] ≤ α/k. We want, P ℓ Pr[ h∈R\{j} Aih Xih > 1 − 1/ℓ] ≤ αk . ℓ−1 Note that the inequality follows from Markov’s inequality. Since ℓ ≥ 2, we have that α α ℓ . ≤ .(1 + 2/ℓ) k ℓ−1 k Note that taking a union bound over the k constraints, setting ℓ = Θ((k/α)1/3 ) and α = log(k), we have that αk .(1 + 2/ℓ) = o(1). Proof of Theorem 6. To prove this theorem, we will now describe a modified version of Algorithm 9. The other steps in the proof remain the same and will directly imply the theorem. To this end, we will now describe the modified algorithm and its analysis. Algorithm 10: Greedy algorithm to color bounded out-degree directed graphs using 2d + d1−ǫ colors, and with near-negative correlation Color-Dir-Graph-Neg-Corr(G, V, d, ǫ, χ) 1 if V = φ then 2 return χ 3 4 5 6 7 else Let vmin denote the vertex with minimum total degree. χ = Color-Dir-Graph-Neg-Corr(G, V \ {vmin }, d, ǫ, χ). Among the smallest d1−ǫ colors in [2d + d1−ǫ ] that have not been used thus far to color any of the neighbors of vmin , choose a color cr uniformly at random. Let χ(vmin ) = cr . return χ As before, we will choose one of the colors c in [2d + d1−ǫ ] uniformly at random and include all vertices which received a color c in the set RF . Since α1−ǫ ≤ o(α), a similar analysis as before follows to give part (1) of Theorem 6. We will now prove part (2). Fix an arbitrary t ∈ [n] and any t-sized subset U := {v1 , v2 , . . . , vt } of items in [n]. A necessary condition for these items to be present in G′ is that they were all chosen into R0 , which happens with probability Qt xvj α (13) j=1 k ; suppose that this indeed happens (all the remaining probability calculations are conditional on this). Note that Algorithm 10 first removes the vertex vmin and then recurses; i.e., it removes the vertices one-by-one, 20 starting with vmin . Let σ be the reverse of this order, and suppose that the order of vertices in U according to σ is, without loss of generality, {v1 , v2 , . . . , vt }. Recall again that our probability calculations now are conditional on all items in U being present in G′ ; denote such conditional probabilities by Pr′ . Note that 1 . Next, a moment’s reflection shows that for any j with 2 ≤ j ≤ t, Pr′ [v1 ∈ RF ] = 2d+d1 1−ǫ ≤ 2α Pr ′ [vj ∈ RF | v1 ∈ RF , v2 ∈ RF , . . . , vj−1 ∈ RF ] ≤ 1 d1−ǫ ≤ 2dǫ . 2α Chaining these together and combining with (13) completes the proof. 7.2 Proofs for Section 3 (SKSP) We will now provide the missing proofs for SKSP. Proof of Lemma 4. Let us fix j. Note that “Y1,j = 0 and Y2,j = 1” occurs with probability (1 − α1 xj /k) (α2 xj /k). Consider a given i ∈ C(j) and let U2,i be the budget usage of resource i when the algorithm reaches j in the random permutation of the second chance. Pr[E2,j ] = Pr[Y1,j = 0 ∧ Y2,j = 1] Pr[E2,j | Y1,j = 0 ∧ Y2,j = 1] # "  ^ α1 xj  α2 xj = 1− (U2,i ≤ bi − 1) | Y1,j = 0 ∧ Y2,j = 1 Pr k k i !  X α1 xj  α2 xj 1− Pr [U2,i ≥ bi | Y1,j = 0] ≥ 1− k k i ! X α2 xj α1 xj ≥ 1− − Pr [U2,i ≥ bi ] k k (14) (15) (16) (17) i Let X1,ℓ be the indicator random variable showing if ℓ is added to R in the first chance and 1{2,ℓ} indicate if item ℓ falls before j in the random order π2 . Thus we have U2,i ≤ which implies E[U2,i ] ≤ X ℓ6=j ui,ℓ P  ℓ6=j SIi,ℓ X1,ℓ + (1 − Y1,ℓ )Y2,ℓ 1{2,ℓ} β1 xℓ 1  α1 xℓ  α2 xℓ 1− + k 2 k k  ≤   β1 1 α2 + k 2 k  bi Plugging the above inequality into (17) and applying Markov’s inequality, we complete the proof of Lemma 4. Proof of Theorem 8. Consider a given item j. We have, Pr[Ej ] = Pr[E1,j ] + Pr[E2,j ] ≥ = xj k xj k β1 + α2 α1 xj k  − β1 − α22   1 − β1 − α22 − O(1/k) β1 + α2 1 − To obtain the worst case, we solve the following optimization problem for an (hypothetical) adversary: 21  α2  max β1 + α2 1 − β1 − , s.t. 0 ≤ β1 ≤ α1 (1 − α1 /2), α1 ≥ 0, α2 ≥ 0 (18) 2 Solving the above program, the optimal solution is α1 = 1, β1 = 1/2 and α2 = 1/2 with a ratio of ( 58 + o(1))k. Bernoulli random variable Zt,j = 1 iff Yt′ ,j = 0 for all Proof of Lemma 5. Consider an item j and define α x Q α x  t′ < t and Yt,j = 1. Observe that E[Zt,j ] = tk j t′ <t 1 − tk′ j . Consider a given i ∈ C(j) and let Ut,j be the budget usage of resource i when the algorithm reaches j in the random permutation during chance t. Thus we have, ′ ] = Pr[Z ′ Pr[Et,j t,j = 1] Pr[Ej,t | Zt,j = 1] α x  α x Q = tk j t′ <t 1 − tk′ j V Pr [ i (Ut,i ≤ bi − 1) | Zt,j = 1] α x  α x Q ≥ tk j t′ <t 1 − tk′ j   P V 1 − i Pr Ut,i ≥ bi | t′ <t Yt′ ,j = 0  P α x Q αt′ xj  ≥ tk j − i Pr [Ut,i ≥ bi ] t′ <t 1 − k Notice that E[Ut,j ] ≤ X ℓ6=j P  1 αt 1 Y αt′ xℓ  αt xℓ xℓ X t′ <t βt′ )≤ + βt′ + 1− bi ui,ℓ ( k ′ 2 ′ k k k 2 k t <t t <t By applying Markov’s inequality, we get Q xj ′ ] ≥ Pr[Et,j t′ <t 1 − k αt ≥ xj k  αt (1− P t′ <t β t′ − αt′ xj  k αt 2 )− αt − P P t′ <t βt′  + t′ <t αt′ k αt 2   P Proof of Lemma 6. For any given P {βt | 1 ≤ t < T }, we have βTP≤ αT 1 − t′ <T βt′ − αT /2 . Thus, the optimal solution is βT∗ = 12 (1 − 1≤t′ <T βt∗′ )2 and αT = 1 − 1≤t′ <T βt∗′ . Thus,  2  P P P ∗ = 1 1+ β + β β = β t t t t<T T t∈[T ] t<T 2 P P Therefore the maximization of t≤T βt is reduced to that P of ∗t≤T −1 βt . By repeating the above analysis ∗ ∗ ∗ for T times, we get our claim for βt and αt . Let γT = t∈[T ] βt . From the above analysis, we see   ∗ )2 , ∀t ≥ 2 γ1∗ = 21 , γt∗ = 21 1 + (γt−1 ∗ )2 ≥ 0, ∗ = 12 (1−γt−1 Since γ1∗ ≤ 1, we can prove that γt∗ ≤ 1 for all t by induction. Notice that γt∗ −γt−1 which implies that {γt } is a non-decreasing. Since {γt } is non-decreasing and bounded, it has a limit ℓ. The only solution to the equation ℓ = (1 + ℓ2 )/2 is ℓ = 1, and hence limT →∞ γT∗ = 1. Proof of Theorem 9. Observe that for each t ∈ [T ], β̄t ≤ βt∗ . Also, 2 α∗ (P ′ α∗ ) P β̄t = 12 1 − t′ <t βt∗′ − t tk<t t′   α∗ P α∗ P ( ′ ) α∗t ∗ ∗ = αt 1 − t′ <t βt′ − 2 − t tk<t t′ ≤ ᾱt 1 − P t′ <t β̄t′ − 22 ᾱt 2  − ᾱt ( P t′ <t k ᾱt′ ) Thus we claim that {ᾱt , β̄t | t ∈ [T ]} is feasible for the program (9). Notice that  P ∗ X X α∗t X T2 log2 k t′ <t αt′ ∗ β̄t = βt − ≥ γT∗ − = 1 − (1 − γT∗ + ) k k k t∈[T ] t∈[T ] t∈[T ] From Lemma 6, we have that (1 − γT∗ ) = o(1) thus proving the theorem. 7.3 Proofs for Section 4 (Hypergraph Matching) We will now provide the missing proofs for Hypergraph Matching. Proof for Lemma 7. The proof is very similar to that in [15, 12]. For each e ∈ E, let Ee be the event that e is added to R in HM(α). Consider an edge e and for each v ∈ e, let Lv be the event that v is available when checking e (in the order given by π). Let π(e) = t ∈ (0, 1). R 1 V  Pr[Ee ] = αxe 0 Pr v∈e Lv | π(e) = t dt R1Q Q ≥ αxe 0 v∈e f :f ∋v (1 − tαxf )dt   R1 e 1 − (1 − α)ke +1 ≥ αxe 0 (1 − tα)ke dt = kex+1 Thus choosing α = 1 completes the proof of the lemma. Proof of Lemma 8. Consider a given xe = x with |e| = ke . Notice that for each given t ∈ (0, 1), P   Q ′ ′ ) = exp 1 − tg(x )) ln(1 − tg(x ′ ′ e e e ∼e e ∼e Note that g(x) = x(1 − x/2) satisfies the condition of Lemma 9 in the Appendix and hence, P for each given t  ∈ (0, 1), the function ln(1 − tg(x)) is convex over x ∈ [0, 1]. Thus to minimize e′ ∼e ln(1 − P tg(xe′ )) subject to 0 ≤ xe′ ≤ 1 and e′ xe′ ≤ κ with κ = ke (1 − x), an adversary will choose the following worst case scenario: create κ/ǫ neighbors for e with each xe′ = ǫ and let ǫ → 0. Thus, min Y e′ ∼e X   1 − tg(xe′ ) = min exp ln(1 − tg(xe′ )) = lim (1 − tg(ǫ))κ/ǫ = exp(−tκ) ǫ→0 e′ ∼e Therefore for each fixed xe = x, the inner minimization program in (12) has an analytic form of optimal value as follows: Z 1Y Z 1  1 − exp(−κ) exp(−tκ)dt = min 1 − tg(xe′ ) dt = (19) κ 0 ′ 0 e ∼e Plugging this back into (12), we obtain F(g) = minx∈[0,1] G(x), where     . 1 G(x) = 1 − x2 ke (1−x) 1 − exp(−ke (1 − x)) Note that G′ (x) ≥ 0 over x ∈ [0, 1] and thus, the minimum value of G(x) over x ∈ [0, 1] is G(0) = 1 ke (1 − exp(−ke )). 23 References [1] A DAMCZYK , M. Non-negative submodular stochastic probing via stochastic contention resolution schemes. CoRR abs/1508.07771 (2015). [2] A DAMCZYK , M., G RANDONI , F., AND M UKHERJEE , J. Improved approximation algorithms for stochastic matching. In Algorithms-ESA 2015. Springer, 2015, pp. 1–12. [3] A DAMCZYK , M., S VIRIDENKO , M., AND WARD , J. Submodular stochastic probing on matroids. In 31st International Symposium on Theoretical Aspects of Computer Science, vol. 25 of LIPIcs. Leibniz Int. Proc. Inform. 2014, pp. 29–40. [4] A JTAI , M., KOML ÓS , J., 29, 3 (1980), 354–360. AND S ZEMER ÉDI , E. A note on Ramsey numbers. J. Comb. Theory, Ser. A [5] A RKIN , E. M., AND H ASSIN , R. On local search for weighted k-set packing. Math. Oper. Res. 23, 3 (1998), 640–648. [6] A ZAR , Y., AND R EGEV, O. Combinatorial algorithms for the unsplittable flow problem. Algorithmica 44, 1 (2006), 49–66. [7] BANSAL , N. Constructive algorithms for discrepancy minimization. In 2010 IEEE 51st Annual Symposium on Foundations of Computer Science—FOCS 2010. IEEE Computer Soc., Los Alamitos, CA, 2010, pp. 3–10. [8] BANSAL , N., C HAKRABARTI , A., E PSTEIN , A., AND S CHIEBER , B. A quasi-PTAS for unsplittable flow on line graphs. In STOC’06: Proceedings of the 38th Annual ACM Symposium on Theory of Computing. ACM, New York, 2006, pp. 721–729. [9] BANSAL , N., DADUSH , D., AND G ARG , S. An algorithm for komlós conjecture matching banaszczyk’s bound. In IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS) (2016), IEEE, pp. 788–799. [10] BANSAL , N., F RIGGSTAD , Z., K HANDEKAR , R., AND S ALAVATIPOUR , M. R. A logarithmic approximation for unsplittable flow on line graphs. ACM Trans. Algorithms 10, 1 (2014), Art. 1, 15. [11] BANSAL , N., AND G ARG , S. Improved algorithmic bounds for discrepancy of sparse set systems. arXiv preprint arXiv:1601.03311 (2016). [12] BANSAL , N., G UPTA , A., L I , J., M ESTRE , J., NAGARAJAN , V., AND RUDRA , A. When LP is the cure for your matching woes: improved bounds for stochastic matchings. Algorithmica 63, 4 (2012), 733–762. [13] BANSAL , N., KORULA , N., NAGARAJAN , V., AND S RINIVASAN , A. Solving packing integer programs via randomized rounding with alterations. Theory Comput. 8 (2012), 533–565. [14] BANSAL , N., 451–471. AND S PENCER , J. Deterministic discrepancy minimization. Algorithmica 67, 4 (2013), 24 [15] BAVEJA , A., C HAVAN , A., N IKIFOROV, A., S RINIVASAN , A., AND X U , P. Improved bounds in stochastic matching and optimization. In Approximation, randomization, and combinatorial optimization. Algorithms and techniques, vol. 40 of LIPIcs. Leibniz Int. Proc. Inform. 2015, pp. 124–134. [16] BAVEJA , A., AND S RINIVASAN , A. Approximation algorithms for disjoint paths and related routing and packing problems. Math. Oper. Res. 25, 2 (2000), 255–280. [17] B ECK , J., AND F IALA , T. “Integer-making” theorems. Discrete Appl. Math. 3, 1 (1981), 1–8. [18] B ERMAN , P. A d/2 approximation for maximum weight independent set in d-claw free graphs. Nordic J. Comput. 7, 3 (2000), 178–184. [19] B RUBACH , B., S ANKARARAMAN , K. A., S RINIVASAN , A., AND X U , P. Attenuate locally, win globally: An attenuation-based framework for online stochastic matching with timeouts. In Sixteenth International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2017). 2017. [20] B UCHBINDER , N., AND F ELDMAN , M. Constrained submodular maximization via a non-symmetric technique. arXiv preprint arXiv:1611.03253 (2016). [21] B UCHBINDER , N., F ELDMAN , M., NAOR , J., AND S CHWARTZ , R. Submodular maximization with cardinality constraints. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms (2014), ACM, New York, pp. 1433–1452. [22] C ALINESCU , G., C HEKURI , C., P ÁL , M., AND VONDR ÁK , J. Maximizing a monotone submodular function subject to a matroid constraint. SIAM J. Comput. 40, 6 (2011), 1740–1766. [23] C HAKRABARTI , A., C HEKURI , C., G UPTA , A., AND K UMAR , A. Approximation algorithms for the unsplittable flow problem. Algorithmica 47, 1 (2007), 53–78. [24] C HAN , Y. H., AND L AU , L. C. On linear and semidefinite programming relaxations for hypergraph matching. Math. Program. 135, 1-2, Ser. A (2012), 123–148. [25] C HANDRA , B., AND H ALLD ÓRSSON , M. M. Greedy local improvement and weighted set packing approximation. J. Algorithms 39, 2 (2001), 223–240. [26] C HEKURI , C., E NE , A., AND KORULA , N. Personal communication. 537–538. [27] C HEKURI , C., E NE , A., AND KORULA , N. Unsplittable flow in paths and trees and columnrestricted packing integer programs. In Approximation, randomization, and combinatorial optimization, vol. 5687 of Lecture Notes in Comput. Sci. Springer, Berlin, 2009, pp. 42–55. [28] C HEKURI , C., JAYRAM , T. S., AND VONDR ÁK , J. On multiplicative weight updates for concave and submodular function maximization. In ITCS’15—Proceedings of the 6th Innovations in Theoretical Computer Science. ACM, New York, 2015, pp. 201–210. √ [29] C HEKURI , C., K HANNA , S., AND S HEPHERD , F. B. An O( n) approximation and integrality gap for disjoint paths and unsplittable flow. Theory Comput. 2 (2006), 137–146. [30] C HEKURI , C., M YDLARZ , M., AND S HEPHERD , F. B. Multicommodity demand flow in a tree and packing integer programs. ACM Trans. Algorithms 3, 3 (2007), Art. 27, 23. 25 [31] C HEKURI , C., VONDR ÁK , J., AND Z ENKLUSEN , R. Dependent randomized rounding via exchange properties of combinatorial structures. In 2010 IEEE 51st Annual Symposium on Foundations of Computer Science—FOCS 2010. IEEE Computer Soc., Los Alamitos, CA, 2010, pp. 575–584. [32] C HEKURI , C., VONDR ÁK , J., AND Z ENKLUSEN , R. Submodular function maximization via the multilinear relaxation and contention resolution schemes. SIAM J. Comput. 43, 6 (2014), 1831–1879. [33] D EAN , B. C., G OEMANS , M. X., AND VONDR ÁK , J. Adaptivity and approximation for stochastic packing problems. In Proceedings of the Sixteenth Annual ACM-SIAM Symposium on Discrete Algorithms (2005), ACM, New York, pp. 395–404. [34] D EAN , B. C., G OEMANS , M. X., AND VONDR ÁK , J. Approximating the stochastic knapsack problem: the benefit of adaptivity. Math. Oper. Res. 33, 4 (2008), 945–964. [35] E LBASSIONI , K., G ARG , N., G UPTA , D., K UMAR , A., NARULA , V., AND PAL , A. Approximation algorithms for the unsplittable flow problem on paths and trees. In 32nd International Conference on Foundations of Software Technology and Theoretical Computer Science, vol. 18 of LIPIcs. Leibniz Int. Proc. Inform. 2012, pp. 267–275. [36] E NE , A., AND N GUYEN , H. L. Constrained submodular maximization: Beyond 1/e. In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on (2016), IEEE, pp. 248–257. [37] F ELDMAN , M., NAOR , J., AND S CHWARTZ , R. A unified continuous greedy algorithm for submodular maximization. In 2011 IEEE 52nd Annual Symposium on Foundations of Computer Science— FOCS 2011. IEEE Computer Soc., Los Alamitos, CA, 2011, pp. 570–579. [38] F ISHER , M. L., N EMHAUSER , G. L., AND W OLSEY, L. A. An analysis of approximations for maximizing submodular set functions. II. Math. Programming Stud., 8 (1978), 73–87. Polyhedral combinatorics. [39] F ÜREDI , Z., K AHN , J., AND S EYMOUR , P. D. On the fractional matching polytope of a hypergraph. Combinatorica 13, 2 (1993), 167–180. [40] G ARG , N., VAZIRANI , V. V., AND YANNAKAKIS , M. Primal-dual approximation algorithms for integral flow and multicut in trees. Algorithmica 18, 1 (1997), 3–20. [41] G UPTA , A., NAGARAJAN , V., AND S INGLA , S. Adaptivity gaps for stochastic probing: Submodular and XOS functions. CoRR abs/1608.00673 (2016). [42] G UPTA , A., NAGARAJAN , V., AND S INGLA , S. Algorithms and adaptivity gaps for stochastic probing. In Proceedings of the Twenty-seventh Annual ACM-SIAM Symposium on Discrete Algorithms (Philadelphia, PA, USA, 2016), SODA ’16, Society for Industrial and Applied Mathematics, pp. 1731– 1747. [43] H ARVEY, N. J. A., S CHWARTZ , R., AND S INGH , M. Discrepancy without partial colorings. In Approximation, randomization, and combinatorial optimization, vol. 28 of LIPIcs. Leibniz Int. Proc. Inform. 2014, pp. 258–273. [44] H AZAN , E., S AFRA , S., AND S CHWARTZ , O. On the complexity of approximating k-set packing. Computational Complexity 15, 1 (2006), 20–39. 26 [45] H URKENS , C. A. J., AND S CHRIJVER , A. On the size of systems of sets every t of which have an SDR, with an application to the worst-case ratio of heuristics for packing problems. SIAM J. Discrete Math. 2, 1 (1989), 68–72. [46] K LEINBERG , J. M. Approximation algorithms for disjoint paths problems. PhD thesis, Citeseer, 1996. [47] KOLLIOPOULOS , S. G., AND S TEIN , C. Approximating disjoint-path problems using packing integer programs. Math. Program. 99, 1, Ser. A (2004), 63–87. [48] KOLMAN , P. A note on the greedy algorithm for the unsplittable flow problem. Inform. Process. Lett. 88, 3 (2003), 101–105. [49] KOLMAN , P., AND S CHEIDELER , C. Improved bounds for the unsplittable flow problem. J. Algorithms 61, 1 (2006), 20–44. [50] K ULIK , A., S HACHNAI , H., AND TAMIR , T. Maximizing submodular set functions subject to multiple linear constraints. In Proceedings of the Twentieth Annual ACM-SIAM Symposium on Discrete Algorithms (2009), SIAM, Philadelphia, PA, pp. 545–554. [51] L EE , J., M IRROKNI , V. S., NAGARAJAN , V., AND S VIRIDENKO , M. Maximizing nonmonotone submodular functions under matroid or knapsack constraints. SIAM J. Discrete Math. 23, 4 (2009/10), 2053–2078. [52] L EIGHTON , T., L U , C.-J., R AO , S., AND S RINIVASAN , A. New algorithmic aspects of the local lemma with applications to routing and partitioning. SIAM J. Comput. 31, 2 (2001), 626–641. [53] L EVY, A., R AMADAS , H., AND ROTHVOSS , T. Deterministic discrepancy minimization via the multiplicative weight update method. arXiv preprint arXiv:1611.08752 (2016). [54] L I , J., AND Y UAN , W. Stochastic combinatorial optimization via Poisson approximation. In Proceedings of the 2013 ACM Symposium on Theory of Computing—STOC’13. ACM, New York, 2013, pp. 971–980. [55] L OVETT, S., AND M EKA , R. Constructive discrepancy minimization by walking on the edges. SIAM J. Comput. 44, 5 (2015), 1573–1582. [56] OVEIS G HARAN , S., AND VONDR ÁK , J. Submodular maximization by simulated annealing. In Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on Discrete Algorithms (2011), SIAM, Philadelphia, PA, pp. 1098–1116. [57] PANCONESI , A., AND S RINIVASAN , A. Randomized distributed edge coloring via an extension of the Chernoff-Hoeffding bounds. SIAM J. Comput. 26, 2 (1997), 350–368. [58] PAREKH , O. Iterative packing for demand and hypergraph matching. In Integer programming and combinatorial optimization, vol. 6655 of Lecture Notes in Comput. Sci. Springer, Heidelberg, 2011, pp. 349–361. [59] PAREKH , O., AND P RITCHARD , D. Generalized hypergraph matching via iterated packing and local ratio. In Approximation and online algorithms, vol. 8952 of Lecture Notes in Comput. Sci. Springer, Cham, 2015, pp. 207–223. 27 [60] P ERES , Y., S INGH , M., AND V ISHNOI , N. Random walks in polytopes and negative dependence. In Proceedings of the Eighth Annual Innovations in Theoretical Computer Science (Berkeley, CA, 2017) (2017). [61] P RITCHARD , D. Approximability of sparse integer programs. In Algorithms—ESA 2009, vol. 5757 of Lecture Notes in Comput. Sci. Springer, Berlin, 2009, pp. 83–94. [62] P RITCHARD , D., AND C HAKRABARTY, D. Approximability of sparse integer programs. Algorithmica 61, 1 (2011), 75–93. [63] R ÖDL , V. On a packing and covering problem. Eur. J. Comb. 6, 1 (1985), 69–78. [64] ROTHVOSS , T. Constructive Discrepancy Minimization for Convex Sets. SIAM J. Comput. 46, 1 (2017), 224–234. [65] S RINIVASAN , A. Improved approximations for edge-disjoint paths, unsplittable flow, and related routing problems. In Foundations of Computer Science, 1997. Proceedings., 38th Annual Symposium on (1997), IEEE, pp. 416–425. [66] S RINIVASAN , A. Improving the discrepancy bound for sparse matrices: better approximations for sparse lattice approximation problems. In Proceedings of the Eighth Annual ACM-SIAM Symposium on Discrete Algorithms (1997), ACM, New York, pp. 692–701. [67] S RINIVASAN , A. Improved approximation guarantees for packing and covering integer programs. SIAM J. Comput. 29, 2 (1999), 648–670. [68] VONDR ÁK , J. Optimal approximation for the submodular welfare problem in the value oracle model. In STOC’08. ACM, New York, 2008, pp. 67–74. [69] WARD , J. A (k + 3)/2-approximation algorithm for monotone submodular k-set packing and general k-exchange systems. In 29th International Symposium on Theoretical Aspects of Computer Science. 2012, pp. 42–53. [70] Z UCKERMAN , D. Linear degree extractors and the inapproximability of max clique and chromatic number. Theory Comput. 3 (2007), 103–128. A Technical Lemmas In our main section, our algorithm for k -CS-PIP uses the following version of the Chernoff-Hoeffding bounds. This can be easily derived from the standard form and we give a proof here for completeness. Theorem 10 (Chernoff-Hoeffding bounds). Suppose c1 , c2 , and k are P positive values with c2 ≥ X1 , X2 , . . . , Xn ∈ [0, 1] be independent random variables satisfying E[ i Xi ] ≤ ck1 . Then,  c2 P c1 e Pr[ i Xi ≥ c2 ] ≤ kc 2 c1 k. Let  µ P δ Proof. The standard form of the Chernoff-Hoeffding bounds yields Pr[ i Xi ≥ (1+δ)µ] ≤ (1+δ)e (1+δ) , for δ ≥ 0. Note that we want (1 + δ)(c1 /k) = c2 , hence giving 1 + δ = form of the Chernoff-Hoeffding bounds gives us the desired bound. 28 kc2 c1 . Plugging this into the standard Lemma 9 (Convexity). Assume f : [0, 1] → [0, 1] and it has second derivatives over [0, 1]. Then we have that ln(1 − tf (x)) is a convex function of x ∈ [0, 1] for any given t ∈ (0, 1) iff (1 − f )(−f ′′ ) ≥ f ′2 for all x ∈ [0, 1]. Proof. Consider a given t ∈ (0, 1) and let F (x) = ln(1 − tf (x)). F (x) is convex over [0, 1] iff F ′′ ≥ 0 for all x ∈ [0, 1]. We can verify that it is equivalent to the condition that (1 − f )(−f ′′ ) ≥ f ′2 for all x ∈ [0, 1]. B Submodular Functions In this section, we give the required background needed for submodular functions. Definition 2 (Submodular functions). A function f : 2[n] → R+ + on a ground-set of elements [n] := {1, 2, . . . , n} is called submodular if for every A, B ⊆ [n], we have that f (A ∪ B) + f (A ∩ B) ≤ f (A) + f (B). Additionally, f is said to be monotone if for every A ⊆ B ⊆ [n], we have that f (A) ≤ f (B). For our algorithms, we assume a value-oracle access to a submodular function. This means that, there is an oracle which on querying a subset T ⊆ [n], returns the value f (T ). Definition 3 (Multi-linear extension). The multi-linear extension of a submodular function f is the continuous function F : [0, 1]n → R+ + defined as P Q Q F (x) := T ⊆[n] ( j∈T xj j6∈T (1 − xj ))f (T ) Note that the multi-linear function F (x) = f (x) for every x ∈ {0, 1}n . The multi-linear extension is a useful tool in maximization of submodular objectives. In particular, the above has the following probabilistic interpretation. Let S ⊆ [n] be a random subset of items where each item i ∈ [n] is added into S with probability xi . We then have F (x) = ES∼x [f (S)]. It can be shown that the two definitions of F (x) are equivalent. Hence, a lower bound on the value of F (x) directly leads to a lower bound on the expected value of f (S). 29
8cs.DS
Under consideration for publication in Theory and Practice of Logic Programming 1 Products of Weighted Logic Programs∗ arXiv:1006.3035v1 [cs.AI] 15 Jun 2010 SHAY B. COHEN, ROBERT J. SIMMONS, and NOAH A. SMITH School of Computer Science Carnegie Mellon University E-mail: {scohen,rjsimmon,nasmith}@cs.cmu.edu submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003 Abstract Weighted logic programming, a generalization of bottom-up logic programming, is a wellsuited framework for specifying dynamic programming algorithms. In this setting, proofs correspond to the algorithm’s output space, such as a path through a graph or a grammatical derivation, and are given a real-valued score (often interpreted as a probability) that depends on the real weights of the base axioms used in the proof. The desired output is a function over all possible proofs, such as a sum of scores or an optimal score. We describe the PRODUCT transformation, which can merge two weighted logic programs into a new one. The resulting program optimizes a product of proof scores from the original programs, constituting a scoring function known in machine learning as a “product of experts.” Through the addition of intuitive constraining side conditions, we show that several important dynamic programming algorithms can be derived by applying PRODUCT to weighted logic programs corresponding to simpler weighted logic programs. In addition, we show how the computation of Kullback-Leibler divergence, an information-theoretic measure, can be interpreted using PRODUCT. KEYWORDS: weighted logic programming, program transformations, natural language processing 1 Introduction Weighted logic programming is a technique that can be used to declaratively specify dynamic programming algorithms in a number of fields such as natural language processing (Manning and Schütze 1999) and computational biology (Durbin et al. 1998). Weighted logic programming is a generalization of bottom-up logic programming where each proof is assigned a score (or weight) that is a function of the scores of the axioms used in the proof. When these scores are interpreted as probabilities, then the solution to a whole weighted logic program can be interpreted in terms of probabilistic reasoning about unknowns, implying that the weighted logic program implements probabilistic inference.1 ∗ To appear in Theory and Practice of Logic Programming (TPLP). 1 The word inference has a distinct meaning in logic programming (e.g. “inference rule,” “valid inference”), and so we will attempt to avoid confusion by using the probabilistic modifier whenever we are talking about probabilistic reasoning about unknowns. 2 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith Even though weighted logic programming is not limited to probabilistic inference, it is worth detailing their relationship. Let I, A, and P be random variables, where the values of I and A are known and the value of P is not known. Often there is a correspondence where • I corresponds to a conditional “input,” encoded as axioms, known to be true; • A corresponds to a set of axioms known to be true; and • P corresponds to a deductive proof of the goal theorem using the axioms. In the setting of weighted logic programming, there may be many different proofs of the goal given the set of axioms. We must therefore distinguish the weighted logic program from the “world” we are reasoning about in which these many different proofs of the goal correspond to different, mutually exclusive events, each of which has some probability of occurring. Weighted logic programming implements probabilistic inference over the value of the proof random variable P given the values of A and I: the weighted logic program implies a probability distribution p(P | A, I), and it can be used to compute different useful quantities related to the distribution. Previous work on weighted logic programming has shown that certain families of probabilistic models lend themselves extremely well to weighted logic programming as an inference mechanism. In general, weighted logic programming deals with probability distributions over objects with combinatorial structure—paths through graphs, grammatical derivations, and sequence alignments—that are quite useful in computer science applications. In principle, one can think about combining such distributions with each other, creating distributions over even more complex structures that are related. This paper is about a natural extension to weighted logic programming as probabilistic inference over structures: combining weighted logic programs to perform inference over two or more structures. We describe a program transformation, PRODUCT, that implements joint probabilistic inference via weighted logic programming over two structured variables P1 and P2 , when (a) each of the two separate structures can be independently reasoned about using weighted logic programming, and (b) the joint model factors into a product of two distributions p(P1 | A1 , I1 ) and p(P2 | A2 , I2 ).2 As a program transformation on traditional logic programs, PRODUCT is not novel; it has existed as a compiler transformation for over a decade (Pettorossi and Proietti 1994; Pettorossi 1999). As a way of describing joint probabilistic inference in weighted logic programming, the transformation has been intuitively exploited in designing algorithms for specific applications, but has not, to our knowledge, been generalized. The contribution of this paper is a general, intuitive, formal setting for dynamic programming algorithms that process two or more conceptually distinct objects. Indeed, we show that many important dynamic programming algorithms can be derived using simpler “factor” programs and the PRODUCT transformation together with side conditions that capture the relationship between the structures. The paper is organized as follows. In §2 we give an overview of weighted logic 2 In the language of probability, this means that P1 and P2 are conditionally independent given A1 , A2 , I1 , and I2 . Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP).3 reachable(Q) :- initial(Q). (1) reachable(Q) :- reachable(P), edge(P, Q). (2) Fig. 1. A simple bottom-up logic program for graph reachability. a b c d initial(a) edge(a, c) edge(a, d) edge(b, b) edge(c, a) = = = = = t t t t t edge(c, d) edge(d, b) edge(d, c) edge(d, d) = = = = t t t t Fig. 2. A directed graph and the corresponding initial database. programming. In §3 we describe products of experts, a concept from machine learning that elucidates the kinds of probabilistic models amenable to our framework. In §4 we describe the PRODUCT transformation. In §5 we give show how several wellknown algorithms can be derived using the PRODUCT transformation applied to simpler algorithms. §6 presents some variations on the PRODUCT transformation. In §7 we show how to use the PRODUCT transformation and a specially designed semiring to calculate important information theoretic quantities related to probability distributions over proofs. 2 Weighted Logic Programming To motivate weighted logic programming, we begin with a logic program for singlesource connectivity on a directed graph, shown in Figure 1. In the usual bottomup interpretation of this program, an initial database (i.e., set of axioms) would describe the edge relation and one (or more) starting vertices as axioms of the form initial(a) for some a. Repeated forward inference can then be applied on the rules in Figure 1 to find the least database closed under those rules. However, in traditional logic programming this program can only be understood as a program calculating connectivity over a graph. Weighted logic programming generalizes traditional logic programming. In traditional logic programming, a proof is a tree of valid (deductive) inferences from axioms, and a valid atomic proposition is one that has at least one proof. In weighted logic programming we generalize this notion: axioms, proofs, and atomic propositions are said to “have values” rather than just “be valid.” Traditional logic programs can be understood as weighted logic programs with Boolean values: axioms all have the value “true,” as do all valid propositions. The single-source connectivity program would describe the graph in Figure 2 by assigning t as the value of all the existing edges and the proposition initial(a). 4 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith 2.1 Non-Boolean Programs With weighted logic programming, the axioms and propositions can be understood as having non-Boolean values. In Figure 3, each axiom of the form edge(X, Y) is given a value corresponding to the cost associated with that edge in the graph, and the axiom initial(a) is given the value 0. If we take the value or “score” of a proof to be the the sum of the values the axioms at its leaves and take the value of a proposition to be the minimum score over all possible proofs, then the program from Figure 1 gives a declarative specification of the single-source shortest path problem. Multiple uses of an axiom in a proof are meaningful: if a proof includes the edge(d, d) axiom once, it corresponds to a single traversal of the loop from d to d and adds a cost of 2, and if a proof includes the axiom twice, it corresponds to two distinct traversals and adds a cost of 4. We replace the connectives :- (disjunction) and , (conjunction) with min = and +, respectively, and interpret the WLP over the non-negative numbers. With a specific execution model, the result is Dijkstra’s single-source shortest-path algorithm. 8 a b 9 20 6 4 16 c 15 d 2 initial(a) edge(a, c) edge(a, d) edge(b, b) edge(c, a) = = = = = 0 4 20 8 9 edge(c, d) edge(d, b) edge(d, c) edge(d, d) = = = = 15 6 16 2 Fig. 3. A cost graph and the corresponding initial database. In addition to the cost-minimization interpretation in Figure 3, we can interpret weights on edges as probabilities and restate the problem in terms of probability maximization. In Figure 4, the outgoing edges from each vertex sum to at most 1. If we assign the missing 0.1 probability from vertex b to a “stopping” event—either implicitly or explicitly by modifying the axioms—then each vertex’s outgoing edges sum to exactly one and the graph can be seen as a Markov model or probabilistic finite-state network over which random walks are well-defined. If we replace the connectives :- (disjunction) and , (conjunction) with max = and ×, then the value of reachable(X) for any X is the probability of the most likely path from a to X. For instance, reachable(a) ends up with the value 1, and reachable(b) ends up with value 0.16, corresponding to the path a → d → b, whose weight is (value of initial(a) × value of edge(a, d) × value of edge(d, b)). If we keep the initial database from Figure 4 but change our operators from max = and × to += and ×, the result is a program for summing over the probabilities of all distinct paths that start in a and lead to X, for each vertex X. This quantity is known as the “path sum” (Tarjan 1981). The path sum for reachable(b), for Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP).5 0.9 a 0.2 0.6 c 0.8 0.3 0.4 b 0.2 d 0.5 initial(a) edge(a, c) edge(a, d) edge(b, b) edge(c, a) = = = = = 1 0.2 0.8 0.9 0.6 edge(c, d) edge(d, b) edge(d, c) edge(d, d) = = = = 0.4 0.2 0.3 0.5 Fig. 4. A probabilistic graph and the corresponding initial database. With stopping probabilities made explicit, this would encode a Markov model. reachable(Q) ⊕= initial(Q). (3) reachable(Q) ⊕= reachable(P) ⊗ edge(P, Q). (4) Fig. 5. The logic program from Figure 1, rewritten to emphasize that it is generalized to an arbitrary semiring. instance, is 10—this is not a probability, but rather an infinite sum of probabilities of many paths, some of which are prefixes of each other.3 These three related weighted logic programs are useful generalizations of the reachability logic program in Figure 1. Figure 5 gives a generic representation of all four algorithms in the Dyna language (Eisner et al. 2005). The key difference among them is the semiring in which we interpret the weights. An algebraic semiring consists of five elements hK, ⊕, ⊗, 0, 1i, where K is a domain closed under ⊕ and ⊗, ⊕ is a binary, associative, commutative operator, ⊗ is a binary, associative operator that distributes over ⊕, 0 ∈ K is the ⊕-identity, and 1 ∈ K is the ⊗-identity. We require, following Goodman (1999), that the semirings we use be complete. Complete semirings are semirings with the additional property that they are closed under finite products and infinite sums—in our running example, this corresponds to the idea that there may be infinitely many paths through a graph, all with finite length. Complete semirings also have the property that infinite sums behave like finite ones—they are associative and commutative, and the multiplicative operator distributes over them. In our running example, reachability uses the Boolean semiring h{t, f}, ∨, ∧, f, ti, single-source shortest path uses hR≥0 ∪ {∞}, min, +, ∞, 0i, the most-probable-path 3 Clearly “10” is not a meaningful probability, but that is a result of the loop from b to b with probability 0.9—in fact, one informal way of looking at the result is simply to observe that 10 = 1 + 0.9 + (0.9)2 + (0.9)3 + . . ., corresponding to proofs of reachable(b) that include edge(b, b) zero, one, two, three, . . . times. If we added an axiom edge(b, final) with weight 0.1 representing the 10% probability of stopping at any step in state b, then the path sum for reachable(final) would be 10 × 0.1 = 1, which is a reasonable probability that corresponds to the fact that a graph traversal can be arbitrarily long but has a 100% chance of eventually reaching b and then stopping. 6 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith variant uses h[0, 1], max, ×, 0, 1i, and the probabilistic path-sum variant uses the so-called “real” semiring hR≥0 ∪ {∞}, +, ×, 0, 1i. Weighted logic programming developed primarily within the computational linguistics community. Building upon the observations of Shieber, Schabes, and Pereira (1995) and Sikkel (1997) that many parsing algorithms for nondeterministic grammars could be represented as deductive logic programs, Goodman (1999) showed that the structure of the parsing algorithms was amenable to interpretation on a number of semirings. McAllester (1999) additionally showed that this representation facilitates reasoning about asymptotic complexity. Other developments include a connection between weighted logic programs and hypergraphs (Klein and Manning 2004), optimal A∗ search for maximizing programs (Felzenszwalb and McAllester 2007), semiring-general agenda-based implementations (Eisner et al. 2005), improved k-best algorithms (Huang and Chiang 2005), and program transformations to improve efficiency (Eisner and Blatz 2007). 2.2 Formal Definition Eisner and Blatz (2007) describe the semantics of weighted logic programs in detail; we summarize their discussion in this section and point the reader to that paper for further detail. A weighted logic program is a set of Horn equations describing a set of declarative, usually recursive equations over an abstract semiring. Horn equations, which we will refer to by the shorter and more traditional term rules, take the form consequent(U) ⊕= antecedent1 (W1 ) ⊗ · · · ⊗ antecedentn (Wn ). Here U and the Wi are sequences of terms which include free variables. If the variables in U are a subset of the variables in W1 . . . Wn for every rule, then the program is range restricted or fully grounded. We can also give rules side conditions. Side conditions are additional constraints that are added to a rule to remove certain proofs from consideration. For example, side conditions could allow us to modify rule 4 in Figure 5 to disallow self-loops and only allow traversal of an edge when there was another edge in the opposite direction: reachable(Q) ⊕= reachable(P) ⊗ edge(P, Q) if edge(Q, P) ∧ Q 6= P. (5) Side conditions do not change the value of any individual proof, they only filter out any proof that does not satisfy the side conditions. In this paper, we use mostly side conditions that enforce equality between variables. For a more thorough treatment of side conditions see Goodman (1999) or Eisner and Blatz (2007). A weighted logic program is specified on an arbitrary semiring, and can be interpreted in any semiring hK, ⊕, ⊗, 0, 1i as previously described. The meaning of a weighted logic program is determined by the rules together with a set of fully grounded axioms (or facts in the Prolog setting). Each axiom is assigned a value from the set K that is interpreted as a weight or score. A common idiom in weighted logic programming is to specify the query as a Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP).7 distinguished predicate goal that takes no arguments. A computationally uninteresting (because are no intermediate computation steps) but otherwise legitimate way to present a weighted logic program is as a single rule of the form goal ⊕= axiom1 (W1 ) ⊗ · · · ⊗ axiomn (Wn ). In this degenerate case, each distinct way of satisfying the premises using axioms in the database would correspond to a distinct proof of goal. The score of each proof would be given by the semiring-product of the scores of the axioms, and the value of goal would be determined by the semiring-sum of the scores of all the proofs. In the general case, the value of the proposition/theorem goal is a semiring-sum over all of its proofs, starting from the axioms, where the value of any single proof is the semiring-product of the axioms involved. This is effectively encoded using the inference rules as a sum of products of sums of products of ... sums of products, exploiting distributivity and shared substructure for efficiency. This inherent notion of shared substructure means that weighted logic programming can give straightforward declarative specifications for problems that are typically solved by dynamic programming. The Dyna programming language implements a particular dynamic programming strategy for implementing these declarative specifications (Eisner et al. 2005), though the agenda algorithm that it implements may potentially have significantly different behavior, in terms of time and space complexity, than other dynamic programming algorithms that meet the same specification. In many practical applications, as in our reachability example in §2.1, values are interpreted as probabilities to be maximized or summed or costs to be minimized. 3 Weighted Logic Programs and Probabilistic Reasoning In this section, we will return focus to the probabilistic interpretation of weighted logic programs that we first described in the introduction. In §3.1 we will describe in more detail how the results of weighted logic programs are interpreted as probabilities—readers with a background in statistics and machine learning can probably skip or skim this section. In §3.2, we will introduce the notion of a product of experts that motivates the PRODUCT transformation. Our running example for this section is a probabilistic finite-state automaton 0.5, 0 a 0.3, 1 0.2, 0 0.2, 0 b d 0.8, 1 1.0, 1 c initial(a) final(c) arc(a, b, 0) arc(a, b, 1) arc(a, d, 0) arc(b, c, 0) arc(b, c, 1) arc(d, c, 1) = = = = = = = = Fig. 6. A probabilistic FSA and the corresponding initial database. 1 1 0.5 0.3 0.2 0.2 0.8 1.0 8 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith goal ⊕= path(Q) ⊗ final(Q). (6) path(Q) ⊕= initial(Q). (7) path(Q) ⊕= path(P) ⊗ arc(P, Q, A). (8) Fig. 7. The weighted logic program for weighted FSA reachability. goal ⊕= path(Q, I) ⊗ final(Q) ⊗ length(I). path(Q, 0) ⊕= initial(Q). (10) path(Q, I) ⊕= path(P, I − 1) ⊗ arc(P, Q, A) ⊗ string(I, A). (11) (9) Fig. 8. The weighted logic program for weighted FSA recognition. over the alphabet {0, 1}, shown in Figure 6. The most probable path through the graph is the one that recognizes the string “01” by going through states a, b, and c, and that the probability of this path is 0.4. Other than the labels on the edges, this is the same setup used in the graph-reachability example from Figure 4. The edge predicate from the previous section is now called arc and has been augmented to carry a third argument representing an output character. 3.1 The Probabilistic Interpretation of Weighted Logic Programming Recall from the introduction that, in the context of weighted logic programming, we have random variables I, A, and P , where • I corresponds to a set of conditional “input” axioms known to be true; • A corresponds to a set of axioms known to be true; and • P corresponds to a deductive proof of the goal theorem using the axioms. In this case, I corresponds to one of the various possible sentences recognized by the FSA (i.e., 00, 01, 10, and 11). A corresponds to a particular directed graph with weighted edges, encoded by a set of axioms. P corresponds to an individual proof/path through the graph. In Figure 7, which is the straightforward adaptation of the reachability program in Figure 5 to labeled edges, the value of goal in the most-probable-path semiring is maxproof p(P = proof , I = sentence | A = graph)— the value of the most probable path emitting any possible sentence I. In order to talk about the input sentences I, we first add a set of axioms that describe I. If we are interested in the sentence “01” we would add axioms string(1, 0), string(2, 1), and length(2), whereas if we were interested in the sentence “hey” we would add axioms string(1, h), string(2, e), string(3, y), and length(3). These axioms are all given the value 1 (the multiplicative unit of the semiring), and so they could equivalently be treated as side conditions. With these new axioms, we modify Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP).9 Figure 7 to obtain Figure 8, a weighted logic program that limits the proofs/paths to the ones which represent recognition of the input string I.4 Now, Figure 8 interpreted over the most-probable-path semiring does allow us to find the proof that, given the edge weights and a specific sentence, maximizes p(P = proof | I = sentence, A = graph). It does not, however, give us p(P = proof | I = sentence, A = graph), but rather p(P = proof , I = sentence | A = graph), the joint probability of a path and a sentence given the weights on the edges. Concretely, in our running example there are five possible proofs of goal in Figure 7 whose probabilities sum to 1, but there are only two parses that also recognize the string “01,” which are a0 b1 c with weight 0.4 and a0 d1 c with weight 0.2—the route through b is twice as likely. The value of goal in Figure 8 interpreted in the most-probable-path semiring would be 0.4 (the joint probability of obtaining the proof a0 b1 c and of recognizing the string “01”) not 0.6 (the probability of the proof a0 b1 c given the sentence “01”). In other words, we have: p(P = a0 b1 c, I = 01 | A = Fig. 6) = 0.4, p(P = a0 d1 c, I = 01 | A = Fig. 6) = 0.2, p(P = a0 b1 c | I = 01, A = Fig. 6) = 0.6. The solution for correctly discovering the conditional probability lies in the fact that the joint and conditional probabilities are related in the following way: p(P | A, I) = p(P, I | A) p(I | A) This, combined with the knowledge that the marginal probability p(I | A) is the result of evaluating Figure 8 over the path-sum semiring (i.e., hR≥0 ∪{∞}, +, ×, 0, 1i), allows us to correctly calculate not only the most probable proof P of a given sentence but also the probability of that proof given the sentence. The marginal probability in our running example is 0.6, and 0.4/0.6 = 0.6, which is the desired result. To restate this in a way that is more notationally consistent with other work in machine learning, we first take the weighted axioms A as implicit. Then, instead of proofs P we talk about values y for a random variable Y drawn out of a domain Y (the space of possible structures, which in our setting corresponds to the space of possible proofs), and instead of inputs I we talk about values x for a random variable X drawn out of a domain X (the space of all possible inputs). Then, to predict the most likely observed value for y, denoted ŷ, we have the following formula: ŷ = argmax p(Y = y | X = x) = argmax y∈Y y∈Y p(Y = y, X = x) p(X = x) (12) Because p(X = x) does not depend on y, if we only want to know ŷ it suffices to find the y that maximizes p(Y = y, X = x) (which was written as p(P = proof , I = 4 Rule 11 in this figure uses “I − 1” in a premise: we assume that our formalism includes natural numbers that support increment/decrement operations, and our simple uses can be understood as syntactic shorthand either for structured terms (z, s(z), etc.) or for the use of primitive side conditions such as inc(I, I0 ). 10 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith sentence | A = axioms) above). One way to do this is to execute a weighted logic program in the most-probable-path semiring. 3.2 Products of Experts Of recent interest are probability models p that take a factored form, for example: p(Y = y | X = x) ∝ p1 (Y = y | X = x) × · · · × pn (Y = y | X = x) (13) where ∝ signifies “proportional to” and suppresses the means by which the probability distribution is renormalized to sum to one. This kind of model is called a product of experts (Hinton 2002). Intuitively, the probability of an event under p can only be relatively large if “all the experts concur,” i.e., if the probability is large under each of the pi . Any single expert can make an event arbitrarily unlikely (even impossible) by giving it very low probability, and the solution to Equation 12 for a product of experts model will be the y ∈ Y (here, a proof) least objectionable to all experts. The attraction of such probability distributions is that they modularize complex systems (Klein and Manning 2003; Liang et al. 2008). They can also offer computational advantages when solving Equation 12 (Chiang 2007). Further, the expert factors can often be trained (i.e., estimated from data) separately, speeding up expensive but powerful machine learning methods (Smith and Smith 2004; Sutton and McCallum 2005; Smith et al. 2005; Cohen and Smith 2007). To the best of our knowledge, there has been no attempt to formalize the following intuitive idea about products of experts: algorithms for reasoning about mutually constrained product proof values should resemble the individual algorithms for each of the two separate “factor” proofs’ values. Our formalization is intended to aid in algorithm development as new kinds of complex random variables are coupled, with a key practical advantage: the expert factors are known because they fundamentally underlie the main algorithm. Indeed, we call our algorithms “products” because they are derived from “factors,” analogous to the product of expert probability models that are derived from factor expert probability models. To relate this observation to the running example from this section, imagine we created two copies of Figure 8 which operated over the same sentence (as described by string and length predicates) but which had different predicates and axioms goal1 , path1 , final1 , initial1 , and arc1 (and likewise goal2 , path2 , etc.). Consider a combined goal predicate goal1•2 defined by the rule goal1•2 ⊕= goal1 ⊗ goal2 . (14) Now we have two experts (goal1 and goal2 ), and we literally take the (semiring) product of them, but this is still not quite the “product of experts,” because the proofs of the goals are allowed to be independent. In other words, what we have is the following: p(Y1 = y1 , Y2 = y2 | X = x) ∝ p1 (Y1 = y1 | X = x) × p2 (Y2 = y2 | X = x). The PRODUCT transformation is a meaning-preserving transformation on weighted Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 11 logic programs that exposes the joint structure in such a way that—depending on our domain-specific understanding of what it means for the two proofs y1 and y2 to match—allows us to add constraints that result in a weighted logic program that forces the structures to match, as required by the specification in Equation 13. 4 Products of Weighted Logic Programs In this section, we will motivate products of weighted logic programs in the context of the running example of generalized graph reachability. We will then define the PRODUCT transformation precisely and describe the process of specifying new algorithms as constrained versions of product programs. The PRODUCT transformation can be seen as an instance of the tupling program transformation combined with an unfold/fold transformation (Pettorossi and Proietti 1994; Pettorossi 1999) that preserves the meaning of programs. However, we are interested in this transformation not for reasons of efficiency, but because it has the effect of exposing the shared structure of the two individual programs in such a way that, by the manual addition of constraints, we can force the two original programs to optimize over the same structures, thereby implementing optimization over the product of experts as described in the previous section. The addition of constraints requires an understanding of the problem at hand, as we show in §5 by presenting a number of examples. 4.1 The Product of Graph Reachability Experts Figure 9 defines two experts, copies of the graph-reachability program from Figure 5. We are interested in a new predicate reachable1•2 (Q1 , Q2 ), which for any particular Q1 and Q2 should be equal to the product of reachable1 (Q1 ) and reachable2 (Q2 ). Just as we did in our thought experiment with goal1•2 in the previous section, we could define the predicate by adding the following rule to the program in Figure 9: reachable1•2 (Q1 , Q2 ) ⊕= reachable1 (Q1 ) ⊗ reachable2 (Q2 ). This program is a bit simplistic, however; it merely describes calculating the experts independently and then combining them at the end. The predicate reachable1•2 can alternatively be calculated by adding the following four rules to Figure 9: reachable1•2 (Q1 , Q2 ) reachable1•2 (Q1 , Q2 ) reachable1•2 (Q1 , Q2 ) reachable1•2 (Q1 , Q2 ) ⊕= ⊕= ⊕= ⊕= initial1 (Q1 ) ⊗ initial2 (Q2 ). initial1 (Q1 ) ⊗ reachable2 (P2 ) ⊗ edge2 (P2 , Q2 ). reachable1 (P1 ) ⊗ edge1 (P1 , Q1 ) ⊗ initial2 (Q2 ). reachable1 (P1 ) ⊗ edge1 (P1 , Q1 ) ⊗ reachable2 (P2 ) ⊗ edge2 (P2 , Q2 ). This step is described as an unfold by Pettorossi (1999). This unfold can then be followed by a fold: because reachable1•2 (Q1 , Q2 ) was defined above to be the product of reachable1 (Q1 ) and reachable2 (Q2 ), we can replace each instance of the two premises reachable1 (Q1 ) and reachable2 (Q2 ) with the single premise reachable1•2 (Q1 , Q2 ). 12 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith reachable1 (Q1 ) ⊕= initial1 (Q1 ). (15) reachable1 (Q1 ) ⊕= reachable1 (P1 ) ⊗ edge1 (P1 , Q1 ). (16) reachable2 (Q2 ) ⊕= initial2 (Q2 ). (17) reachable2 (Q2 ) ⊕= reachable2 (P2 ) ⊗ edge2 (P2 , Q2 ). (18) Fig. 9. Two identical experts for generalized graph reachability, duplicates of the program in Figure 5. reachable1•2 (Q1 , Q2 ) ⊕= initial1 (Q1 ) ⊗ initial2 (Q2 ). (19) reachable1•2 (Q1 , Q2 ) ⊕= reachable2 (P2 ) ⊗ edge2 (P2 , Q2 ) ⊗ initial1 (Q1 ). (20) (21) reachable1•2 (Q1 , Q2 ) ⊕= reachable1 (P1 ) ⊗ edge1 (P1 , Q1 ) ⊗ initial2 (Q2 ). reachable1•2 (Q1 , Q2 ) ⊕= reachable1•2 (P1 , P2 ) ⊗ edge1 (P1 , Q1 ) ⊗ edge2 (P2 , Q2 ). (22) Fig. 10. Four rules that, in addition to the rules in Figure 9, give the product of the two experts defined by the reachable1 and reachable2 predicates. The new rules that result from this replacement can be seen in Figure 10. 4.2 The PRODUCT Transformation The PRODUCT program transformation is shown in Figure 11. For each desired product of experts, where one expert, the predicate p, is defined by n rules and the other expert q by m rules, the transformation defines the product of experts for p•q with n × m new rules, the cross product of inference rules from the first and second experts. The value of a coupled proposition p•q in P 0 will be equal to the semiring product of p’s value and q’s value in P (or, equivalently, in P 0 ). Note that lines 6–8 are nondeterministic under certain circumstances, because if the antecedent of the combined program is a(X) ⊗ a(Y) ⊗ b(Z) and the algorithm is computing the product of a and b, then the resulting antecedent could be either a•b(X, Z) ⊗ a(Y) or a•b(Y, Z) ⊗ a(X). This nondeterminism usually does not arise, and when it does, as in §5.2, there is usually an obvious preference. The PRODUCT transformation is essentially meaning preserving: if the program 0 P is the result of the PRODUCT transformation on P, then the following is true: • Any ground instance p(X) that is given a value in P is given the same value in P 0 . This is immediately apparent because the program P 0 is stratified: none of the new rules are ever used to compute values of the form p(X), so their value is identical to their value in P. • Any ground instance p•q(X, Y) in P 0 has the same value as p(X) ⊗ q(Y). This is the result of the following theorem: Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 13 Input: A logic program P and a set S of pairs of predicates (p, q). Output: A program P 0 that extends P, additionally computing the product predicate p•q for every pair (p, q) ∈ S in the input. 1: P 0 ← P 2: for all pairs (p, q) in S do 3: for all rules in P, of the form p(W) ⊕= A1 ⊗ · · · ⊗ An do 4: for all rules in P, of the form q(X) ⊕= B1 ⊗ · · · ⊗ Bm do 5: let r ← [p•q(W, X) ⊕= A1 ⊗ · · · ⊗ An ⊗ B1 ⊗ · · · ⊗ Bm ] 6: for all pairs (s(Y), t(Z)) of antecedents in r such that (s, t) ∈ S do 7: remove the antecedents s(Y) and t(Z) from r 8: insert the antecedent s•t(Y, Z) into r 9: end for 10: add r to P 0 11: end for 12: end for 13: end for 14: return P 0 Fig. 11. Algorithmic specification of the PRODUCT transformation. Theorem 1 Let P be a weighted logic program over a set of predicates R, and let S be a set of pairs of predicates from P. Then after applying PRODUCT on (P, S), resulting in a new program P 0 , for every (p, q) ∈ S, the value p•q(X, Y) in P 0 is p(X) ⊗ q(Y). Proof: By distributivity of the semiring, we know that p(X) ⊗ q(Y) is the sum: M v(t) ⊗ v(r) where t and r range over proofs of p(X) and q(Y) respectively, with t,r their values being v(t) and v(r). This implies that we need to show that there is a bijection between the set A of proofs for p•q(X, Y) in P 0 and the set B of pairs of proofs for p(X) and q(Y) such that for every s ∈ A and (t, r) ∈ B we have v(s) = v(t) ⊗ v(r). Using structural induction over the proofs, we first show that every pair of proofs (t, r) ∈ B has a corresponding proof s ∈ A with the needed value. In the base case, where the proofs t and r include a single step, the correspondence follows trivially. Let (t, r) ∈ B. Without loss of generality, we will assume that both t and r contain more than a single step in their proofs. In the last step of its proof, t used a rule of the form p(X) ⊕= a1 (X1 ) ⊗ · · · ⊗ an (Xn ) (23) and r used a rule in its last step of the form q(Y) ⊕= b1 (Y1 ) ⊗ · · · ⊗ bm (Ym ) (24) Let ti be the subproofs of ai (Xi ) and let rj be subproofs of bj (Yj ). It follows that PRODUCT creates from those two rules a single inference rule of the form: p•q(X, Y) ⊕= c1 (W1 ) ⊗ · · · ⊗ cp (Wp ) (25) 14 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith reachable1•2 (Q1 , Q2 ) ⊕= initial1 (Q1 ) ⊗ initial2 (Q2 ). reachable1•2 (Q1 , Q2 ) ⊕= reachable1•2 (P1 , P2 ) ⊗ edge1 (P1 , Q1 ) ⊗ edge2 (P2 , Q2 ). (27) (26) Fig. 12. By removing all but these two rules from the product of experts in Figure 10, we constrain both paths to have the same number of steps. reachable1•2 (Q1 , Q2 ) ⊕= initial1 (Q1 ) ⊗ initial2 (Q2 ) if Q1 = Q2 . (28) reachable1•2 (Q1 , Q2 ) ⊕= reachable1•2 (P1 , P2 ) ⊗ edge1 (P1 , Q1 ) ⊗ edge2 (P2 , Q2 ) if Q1 = Q2 . Fig. 13. By further constraining the program in Figure 12 to demand that Q1 = Q2 at all points, we constrain both paths to be identical. reachable1•2 (Q) ⊕= initial1 (Q) ⊗ initial2 (Q). (29) reachable1•2 (Q) ⊕= reachable1•2 (P) ⊗ edge1 (P, Q) ⊗ edge2 (P, Q). (30) Fig. 14. We can simplify Figure 13 by internalizing the side condition and giving reachable1•2 only one argument. where ci (Wi ) is either ak (Yk ) for some k, or bl (Yl ) for some k, or ak • b` (Xk , Y` ) for some k, `. We resolve each case as following: 1. If ci (Wi ) = ak (Yk ) then we set si = tk . 2. If ci (Wi ) = bk (Yk ) then we set si = rk . 3. If ci (Wi ) = ak • b` (Xk , Y` ) then according to the induction hypothesis, we have a proof for ak • b` (Xk , Y` ) such that its value is v(tk ) ⊗ v(r` ). We set si to be that proof. Since we have shown there is a proof for each antecedent of p • q(X, Y), we have shown that there is a proof for p • q(X, Y). That its value is indeed p(X) ⊗ q(Y) is concluded trivially from the induction steps. The reverse direction for constructing the bijection is similar, again using structural induction over proofs.  4.3 From PRODUCT to a Product of Experts The output of the PRODUCT transformation is a starting point for describing dynamic programming algorithms that perform two actions—traversing a graph, scanning a string, parsing a sentence—at the same time and in a coordinated fashion. Exactly what “coordinated fashion” means depends on the problem, and answering that question determines how the problem is constrained. If we return to the running example of generalized graph reachability, the program as written has eight rules, four from Figure 9 and four from Figure 10. Two Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 15 examples of constrained product programs are given in Figures 12–14. In the first example in Figure 12, the only change is that all but two rules have been removed from the program in Figures 9 and 10. Whereas in the original product program reachable1•2 (Q1 , Q2 ) corresponded to the product of the weight of the best path from the initial state of graph one to Q1 and the weight of the best path from the initial state of graph two to Q2 , the new program computes the best paths from the two origins to the two destinations with the additional requirement that the paths be the same length—the rules that were deleted allowed for the possibility of a prefix on one path or the other. If our intent is for the two paths to not only have the same length but to visit exactly the same sequence of vertices, then we can further constrain the program to only define reachable1•2 (Q1 , Q2 ) where Q1 = Q2 , as shown in Figure 13. After adding this side condition, it is no longer necessary for reachable1•2 to have two arguments that are always the same, so we can simply further as shown in Figure 14. For simplicity’s sake, we will usually collapse arguments that have been forced by equality constraints to agree. The choice of paired predicates S is important for the final weighted logic program that PRODUCT returns and it also limits the way we can add constraints to derive a new weighted logic program. Future research might consider a machine learning setting for automatically deriving S from data, to minimize some cost (e.g., observed runtime). When PRODUCT is applied on two copies of the same weighted logic program (concatenated together to a single program), a natural schema for selecting paired predicates arises, in which we pair a predicate from one program with the same predicate from the other program. This “natural” pairing leads to the derivation of several useful, known algorithms, to which we turn in §5. 5 Examples In this section, we give several examples of constructing weighted logic programs as constrained products of simpler weighted logic programs. 5.1 Finite-State Algorithms We have already encountered weighted finite-state automata (WSFAs) in §3.1. Like WFSAs, weighted finite-state transducers (WFSTs) are a generalization of the graph-reachability problem: in WFSAs the edges are augmented with a symbol and represented as arc(P, Q, A), whereas in WFSTs edges are augmented with a pair of input-output symbols and represented as arc(P, Q, A, B). Weighted finite-state machines are widely used in speech and language processing (Mohri 1997; Pereira and Riley 1997). They are used to compactly encode many competing string hypotheses, for example in speech recognition, translation, and morphological (word-structure) disambiguation. Many sequence labeling and segmentation methods can also be seen as weighted finite-state models. 16 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith goal1•2 ⊕= path1•2 (Q1 , Q1 ) ⊗ final1 (Q2 ) ⊗ final2 (Q2 ). (31) path1•2 (Q1 , Q2 ) ⊕= initial1 (Q1 ) ⊗ initial2 (Q2 ). (32) path1•2 (Q1 , Q2 ) ⊕= path1•2 (P1 , P2 ) ⊗ arc1 (P1 , Q1 , A1 ) ⊗ arc2 (P2 , Q2 , A2 ) if A1 = A2 . (33) Fig. 15. The constrained product of two of the WFSA experts described in Figure 7. goal ⊕= path(Q) ⊗ final(Q). (34) path(Q) ⊕= initial(Q). (35) path(Q) ⊕= path(P) ⊗ arc(P, Q, A, B). (36) Fig. 16. The weighted logic program describing weighted finite-state transducers. goal1•2 ⊕= path1•2 (Q1 , Q2 ) ⊗ final1 (Q1 ) ⊗ final2 (Q2 ). path1•2 (Q1 , Q2 ) ⊕= initial1 (Q1 ) ⊗ initial2 (Q2 ). (37) (38) path1•2 (Q1 , Q2 ) ⊕= path1•2 (P1 , P2 ) ⊗ arc1 (P1 , Q1 , A1 , B1 ) ⊗ arc2 (P2 , Q2 , A2 , B2 ) if B1 = A2 . Fig. 17. The composition of two weighted finite-state transducers can be derived by constraining the product of two weighted finite-state transducers. Weighted finite-state automata. Our starting point for weighted finite-state automata will be the weighted logic program for WFSAs described in Figure 7, which is usually interpreted as a probabilistic automaton in the most-probable-path semiring (i.e., h[0, 1], max, ×, 0, 1i). If the PRODUCT of that algorithm with itself is taken, we can follow a series of steps similar to the ones described in §4.3. First, we remove rules that would allow the two WFSAs to consider different prefixes, and then we add a constraint to rule 33 that requires the two paths’ symbols to be identical. The result is a WFSA describing the (weighted) intersection of the two WFSAs. The intersection of two WFSAs is itself a WFSA, though it is a WFSA where states are described by two terms—Q1 and Q2 in path1•2 (Q1 , Q2 )—instead of a single term. Weighted intersection generalizes intersection and has a number of uses. For instance, consider an FSA that is “probabilistic” but that only accepts the single string “01” because the transitions are all deterministic and have probability 1: 0 1.0, 0 1 1.0, 1 2 If we consider the program in Figure 15 with axioms describing the above FSA and the probabilistic FSA given in Figure 6, then the resulting program is functionally equivalent to the weighted logic program in Figure 8 describing a WFSA specialized to a particular string. Alternatively, if we consider the program in Figure 15 with axioms describing the probabilistic FSA in Figure 6 and the following single-state Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 17 probabilistic FSA, the result will be a probabilistic FSA biased towards edges with the “1” symbol and against edges with the “0” symbol. 0.1, 0 s 0.9, 1 Both of the above examples can be understood as instances of the product of experts pattern discussed in §3.2. In the first case, the additional expert eliminates certain possibilities by assigning zero probability to them, and in the second case the additional expert merely modifies probabilities by preferring the symbol “1” to the symbol “0.” Weighted finite-state transducers. Suppose we take the PRODUCT transformation of the WFST recognition algorithm in Figure 16 with itself and constrain the result by removing all but the three interesting rules (as before) and requiring that B1 (the “output” along the first edge) always be equal to A2 (the “input” along the second edge). The result is shown in Figure 17; this is the recognition algorithm for the WFST resulting from composition of two WFSTs. Composition permits small, understandable components to be cascaded and optionally compiled, forming complex but efficient models of string transduction (Pereira and Riley 1997). 5.2 Context-Free Parsing Parsing natural languages is a difficult, central problem in computational linguistics (Manning and Schütze 1999). Consider the sentence “Alice saw Bob with binoculars.” One analysis (the most likely in the real world) is that Alice had the binoculars and saw Bob through them. Another is that Bob had the binoculars, and Alice saw the binocular-endowed Bob. Figure 18 shows syntactic parses into noun phrases (NP), verb phrases (VP), etc., corresponding to these two meanings. It also shows some of the axioms that could be used to describe a context-free grammar describing English sentences in Chomsky normal form (Hopcroft and Ullman 1979).5 A proof corresponds to a derivation of the given sentence in a context-free grammar, i.e., a parse tree. Shieber et al. (1995) show that parsing with CFGs can be formalized as a logic program, and in Goodman (1999) this framework is extended to the weighted case. If weights are interpreted as probabilities, then the h[0, 1], max, ×, 0, 1i semiring interpretation finds the probability of the parse with maximum probability and the hR≥0 ∪ {∞}, +, ×, 0, 1i semiring interpretation finds the total weight of all parse trees (a measure of the “total grammatically” of a sentence). In Figure 19, we give the specification of the weighted CKY algorithm (Cocke and Schwartz 1970; Kasami 5 Chomsky normal form (CNF) means that the rules in the grammar are either binary with two nonterminals or unary with a terminal. We do not allow  rules, which in general are allowed in CNF grammars. 18 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith S NP VP NP → Alice NP → Bob P → with ... S → NP VP VP → V NP PP → P NP NP → NP PP ... PP VP P V NP NP Alice saw Bob with binoculars V NP VP NP P NP PP NP S unary(np, “Alice”) unary(np, “Bob”) unary(p, “with”) ... binary(s, np, vp) binary(vp, v, np) binary(pp, p, np) binary(np, np, pp) ... Fig. 18. An ambiguous sentence that can be parsed two ways in English (left), some of the Chomsky normal form rules for English grammar (center), and the corresponding axioms (right). There would also need to be five axioms of the form string(1, “Alice”), string(2, “saw”), etc. goal1 ⊕= start1 (S) ⊗ length(N) ⊗ c1 (S, 0, N). c1 (X, I − 1, I) ⊕= unary1 (X, W) ⊗ string(I, W). (40) c1 (X, I, K) ⊕= binary1 (X, Y, Z) ⊗ c1 (Y, I, J) ⊗ c1 (Z, J, K). (41) (39) Fig. 19. A weighted logic program for parsing weighted context-free grammars. 1965; Younger 1967), which is a dynamic programming algorithm for parsing using a context-free grammar in Chomsky normal form.6 Figure 19 suggestively has a subscript attached to all but the length and string inputs. In our description of the product of experts framework in §3.2, the axioms length and string correspond to the conditional input sentence I. The unconstrained result of the PRODUCT transformation on the combination of the rules in Figure 19 and a second copy that has “2” subscripts is given in Figure 20. Under the most-probable-path probabilistic interpretation, the value of goal1•2 is the probability of the given string being generated twice, once by each of the two probabilistic grammars, in each case by the most probable tree in that grammar. By constraining Figure 20, we get the more interesting program in Figure 21 that adds the additional requirement that the two parse trees in the two different grammars have the same structure. In particular, in all cases the constraints I1 = I2 , J1 = J2 , K1 = K2 , N1 = N2 are added, so that instead of writing c1•2 (X1 , I1 , J1 , X2 , I2 , J2 ) we just write c1•2 (X1 , X2 , I, J). Lexicalized CFG parsing. An interesting variant of the previous rule involves lexicalized grammars, which are motivated in Figure 22. Instead of describing a gram6 Strictly speaking, the CKY parsing algorithm corresponds to a naı̈ve bottom-up evaluation strategy for this program. Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 19 goal1•2 ⊕= length(N1 ) ⊗ length(N2 ) ⊗ c1•2 (X1 , I1 − 1, I1 , X2 , I2 − 1, I2 ) ⊕= c1•2 (X1 , I1 − 1, I1 , X2 , I2 , K2 ) ⊕= unary1 (X1 , W1 ) ⊗ string(I1 , W1 ) ⊗ c1•2 (X1 , I1 , K1 , X2 , I2 − 1, I2 ) ⊕= unary2 (X2 , W1 ) ⊗ string(I2 , W2 ) ⊗ c1•2 (X1 , I1 , K1 , X2 , I2 , K2 ) ⊕= binary1 (X1 , Y1 , Z1 ) ⊗ binary2 (X2 , Y2 , Z2 ) ⊗ (42) start1 (S1 ) ⊗ start2 (S2 ) ⊗ c1•2 (S1 , 0, N1 , S2 , 0, N2 ). unary1 (X1 , W1 ) ⊗ string(I1 , W1 ) ⊗ (43) unary2 (X2 , W2 ) ⊗ string(I2 , W2 ). (44) binary2 (X2 , Y2 , Z2 ) ⊗ c2 (Y2 , I2 , J2 ) ⊗ c2 (Z2 , J2 , K2 ). (45) binary1 (X1 , Y1 , Z1 ) ⊗ c1 (Y1 , I1 , J1 ) ⊗ c1 (Z1 , J1 , K1 ). (46) c1•2 (Y1 , I1 , J1 , Y2 , I2 , J2 ) ⊗ c1•2 (Z1 , J1 , K1 , Z2 , K2 , J2 ). Fig. 20. The result of the PRODUCT transformation on two copies of Figure 19. goal1•2 ⊕= length(N) ⊗ start1 (S1 ) ⊗ start2 (S2 ) ⊗ c1•2 (S1 , S2 , 0, N) (47) c1•2 (X1 , X2 , I − 1, I) ⊕= unary1 (X1 , W) ⊗ unary2 (X2 , W) ⊗ string(I, W). (48) c1•2 (X1 , X2 , I, K) ⊕= binary1 (X1 , Y1 , Z1 ) ⊗ binary2 (X2 , Y2 , Z2 ) ⊗ (49) c1•2 (Y1 , Y2 , I, J) ⊗ c1•2 (Z1 , Z2 , J, K). Fig. 21. The program in Figure 20 constrained to require internally-identical trees. S NP saw VP PP VP V NP P NP Alice saw Bob with binoculars NP → Alice P → with S → NP VP VP → V NP S-saw saw Alice saw saw Bob VP-saw NP-Alice with with binoculars Alice saw Bob with binoculars Alice → Alice with → with saw → Alice saw saw → saw Bob VP-saw V-saw NP-Bob PP-with P-with NP-binoculars Alice saw Bob with binoculars NP-Alice → Alice P-with → with S-saw → NP-Alice VP-saw VP-saw → V-saw NP-Bob Fig. 22. On the left, the grammar previously shown. In the middle, a context-free dependency grammar, whose derivations can be seen as parse trees (above) or a set of dependencies (below). On the right, a lexicalized grammar. Sample rules are given for each grammar. mar using nonterminals denoting phrases (e.g., NP and VP), which is called a constituent-structure grammar we can define a (context-free) dependency grammar (Gaifman 1965) that encodes the syntax of a sentence in terms of parent-child relationships between words. In the case of the example of Figure 22, the arrows below the sentence in the middle establish “saw” as the root of the sentence; the word “saw” has three children (arguments and modifiers), one of which is the word “with,” which in turn has the child “binoculars.” 20 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith A simple kind of dependency grammar is a Chomsky normal form CFG where the nonterminal set is equivalent to the set of terminal symbols (so that the terminal “with” corresponds to a unique nonterminal with, and so on) and where all rules have the form P → P C, P → C P, and W → w (where P is the “parent” word, C is the “child” word that is dependent on the parent, and W is the nonterminal version of terminal word w). If we encode the constituent-structure grammar in the unary1 and binary1 relations and encode a dependency grammar in the unary2 and binary2 relations, then the product is a lexicalized grammar, like the third example from Figure 22. In particular, it describes a lexicalized context-free grammar with a product of experts probability model (Klein and Manning 2003), because the weight given to any production A-X → B-X C-Y is the semiring-product of the weight given to the production A → B C and the weight given to the dependency based production X → X Y. This was an important distinction for Klein and Manning—they were interested in factored lexicalized grammars that Figure 21 can describe. These are only a small (but interesting) subset of all possible lexicalized grammars. Standard lexicalized CFGs assign weights directly to grammar productions of the form A-X → B-X C-Y, not indirectly (as we do) by assigning weights to a constituentstructure and a dependency grammar. We will return to this point in §6.2 when we consider the “axiom generalization” pattern that allows us to describe general lexicalized CKY parsing (Eisner 1997; Eisner and Satta 1999). Nondeterminism and rule binarization. The result of the PRODUCT transformation shown in Figure 20 was the first time the nondeterminism inherent in lines 6-8 of the description of the PRODUCT transformation (Figure 11) has come into play. Because there were two c1 premises and two c2 premises, they could have been merged in more than one way. For example, the following would have been a potential alternative to rule 46: c1•2 (X1 , I1 , K1 , X2 , I2 , K2 ) ⊕= binary1 (X1 , Y1 , Z1 ) ⊗ binary2 (X2 , Y2 , Z2 ) ⊗ (50) c1•2 (Y1 , I1 , J1 , Z2 , K2 , J2 ) ⊗ c1•2 (Z1 , J1 , K1 , Y2 , I2 , J2 ). However, this would have broken the correspondence between I1 and I2 and made it impossible to constrain the resulting program as we did. An alternative to CKY is the binarized variant of CKY where rule 41 is split into two rules by introducing a new, temporary predicate (rules 51 and 52): temp1 (X, Y, J, K) ⊕= binary1 (X, Y, Z) ⊗ c1 (Z, J, K). (51) c1 (X, I, K) ⊕= c1 (Y, I, J) ⊗ temp1 (X, Y, J, K). (52) In this variant, the nondeterministic choice in the PRODUCT transformation disappears. The choice that we made in pairing was consistent with the choice that is forced in the binarized CKY program. Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 21 goal ⊕= targetlength(M) ⊗ predict(EM−1 , EM , M + 1). predict(EJ−1 , EJ , J + 1) ⊕= predict(EJ−2 , EJ−1 , J) ⊗ trigram(EJ−2 , EJ−1 , EJ ). (54) (53) Fig. 23. A weighted logic program giving a trigram prediction model for a language, which can be generalized to an n-gram model for any n. goal 0 ⊕= sourcelength(N) ⊗ trans(N, []). 0 (55) trans(I , Es) ⊕= trans(I, []) ⊗ phrase(I, I , EJ :: Es). (56) trans(I0 , Es) ⊕= trans(I0 , EJ :: Es). (57) Fig. 24. A weighted logic program that describes monotone decoding—translating a phrase at a time of the input language into the output language without reordering. goal ⊕= sourcelength(N) ⊗ targetlength(M) ⊗ (58) pr•tr(N, M, EM−1 , EM , M + 1). pr•tr(I0 , J + 1, EJ−1 , EJ , Es) ⊕= pr•tr(I, J, EJ−2 , EJ−1 , []) ⊗ (59) trigram(EJ−2 , EJ−1 , EJ ) ⊗ phrase(I, I0 , EJ :: Es). pr•tr(I0 , J + 1, EJ−1 , EJ , Es) ⊕= pr•tr(I0 , J, EJ−2 , EJ−1 , EJ :: Es) ⊗ (60) trigram(EJ−2 , EJ−1 , EJ ). Fig. 25. Phrase translation as the constrained product of Figures 23 and 24. 5.3 Translation Algorithms Another example of two probabilistic models that play the role of experts arises in translation of sentences from one natural language to another. We will summarize how the PRODUCT transformation was applied to a simple form of phrase-to-phrase translation (Koehn et al. 2003) by Lopez (2009). Lopez (2009) suggested a deductive view of algorithms for machine translation, similar to the view of parsing given in Shieber et al. (1995). Lopez used the PRODUCT transformation to derive an algorithm for phrase translation from two different factor programs, one which attempts to enforce fluency (a measure of the grammaticality of a sentence) in the translated sentence and one which attempts to enforce adequacy (a measure of how much of the meaning of an original sentence is preserved in the translation.) If fluency is a measure of the grammaticality of a sentence, then it would seem that the CKY algorithm for parsing context-free grammars would be a candidate. While such models have been used in translation (Charniak et al. 2003), Lopez’s example uses a simpler notion of fluency based on an n-gram language model (Manning and Schütze 1999, Chapter 6). An n-gram model assigns the probability of a sentence to be the product of probabilities of each word following the (n − 1)-word 22 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith sequence immediately preceding it. As a concrete example, let us say that n = 3 (called a “trigram” model) and work with the program in Figure 23. If we were estimating our trigram model based on the relative frequencies of sequences in Shakespeare’s Othello, we would note that the phrase “if it” appears eight times in the text. Three of these are from the sequence “if it be” and one is from the sequence “if it prove,” so the axiom trigram(“if”, “it”, “be”) should have a probability that is three times the probability given to trigram(“if”, “it”, “prove”). If we then stared the program with the initial sentence fragment predict(“if”, “it”, 3), we could derive predict(“it”, “be”, 4) with the aforementioned axiom and then predict(“be”, “demanded”, 5) with the axiom trigram(“it”, “be”, “demanded”), a sequence occurring once in the text. The result so far is a sequence “if it be demanded” that does not appear in Othello, but which perhaps sounds like it could (which is an informal way of describing the criterion for fluency). The weighted logic program Lopez uses to enforce adequacy is the “monotone decoding” logic program presented in Figure 24. The program is slightly contrived in order to interact with the PRODUCT transformation correctly. The atomic proposition trans(I, Es) refers to a particular point, I, in the source-language string and a list Es of unprocessed words in the target language.7 Each deduction consumes a single word (EJ ) in the target language—indeed, this is the only function of rule 57. When there are no words to remove, then either the entire source-language string has been translated (rule 55), or else progress can continue by translating some chunk of the source-language sentence starting from position I and ending at position I0 as the non-empty list of target-language words EJ :: Es and applying rule 56. This translation of a sequence of the source-language words is captured by the premise phrase(I, I0 , Es), corresponding to the source subsequence from position I to position I0 being translated as Es (a target-language phrase). The meaning of phrase could be defined by a set of axioms or by a rule. In the latter case, if we enumerate all the substrings Ds in the source-language sentence as axioms substr(I, I0 , Ds) and provide axioms ptranslate(Ds, Es) describing sourcelanguage to target-language phrase translation, then phrase(I, I0 , Es) may be defined by the following rule: phrase(I, I0 , Es) ⊕= substr(I, I0 , Ds) ⊗ ptranslate(Ds, Es). (61) Note that substr might be provided as an axiom, or derived from axioms encoding the source sentence through another inference rule. Figure 25 displays Lopez’s phrase translation program by constraining the product of the n-gram model and monotone decoding programs. Lopez describes this for any n, but for simplicity we continue using a trigram model (n = 3). The combined predicate simultaneously tracks a position in the source-language sentence I and the target-language sentence J. The word EJ that was discarded at each step in Figure 24 is given relevance by the trigram model. The combination of these two programs uses the monotone decoding program’s capabilities to make sure that the 7 We use a standard syntactic shorthand for lists; “[]” can be read as the constant nil and “E :: Es” can be read as cons(E, Es). Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 23 phrase-by-phrase meaning of the source-language string D1 . . . DN is preserved in the destination language string E1 . . . EM (adequacy) while simultaneously using the trigram model’s capabilities to ensure that the result is a plausible sentence in the destination language (fluency). Our presentation of machine translation algorithms through the PRODUCT transformation is simplistic. Lopez (2009) discusses more powerful translation algorithms that permit, for example, reordering of phrases. 6 Variations on PRODUCT Up to this point, we have viewed our use of the PRODUCT transformation as one that solves a problem of joint optimization: we take two logic programs that describe structures (such as strings, paths, or trees), relate them to one another by adding constraints, and then optimize over the two original structures simultaneously (one instance of this is when we use weighted logic programming to describe a product of experts.) This is a useful pattern, but it is not the only interesting use of the fold/unfold transformation underlying the PRODUCT transformation. In this section we consider two other variants: in the first we only optimize over one of the two structures and fix the other one, and in the second we take the output of PRODUCT as describing not joint optimization over two simple structures but over one complex structure. 6.1 Fixing One of the Factor Structures The usual use of the PRODUCT transformation is to joint optimization on two structures, but general side conditions can be used to take the additional step of fixing one of the two structures and having the weighted logic program perform optimization on the other structure, subject to constraints imposed through the pairing. In the setting where we consider weights to be probabilities, this is useful for solving certain probabilistic inference problems. Using the path-sum semiring (i.e., hR≥0 ∪ {∞}, +, ×, 0, 1i), the result is a program calculating the marginalized quanP tity p(x) = y p(x, y) (where x corresponds to one program’s proof and y to the other program’s proof). This is a useful quantity in learning; for example, the expectation-maximization (EM) algorithm (Dempster et al. 1977) for optimizing the marginalized log-likelihood of observed structures requires calculating sufficient statistics which are based on marginal quantities. Using the mostprobable-path semiring (i.e., h[0, 1], max, ×, 0, 1i), the result is a program for solving argmaxy p(y | x)—that is, for finding the most probable y given the fixed x. The transformation of the constrained result of the PRODUCT transformation to a program with one proof fixed is essentially mechanical. We consider the example of lexicalized parsing from Figure 22. We take the constituent-structure parse as the structure we want to fix in order optimize over the possible matching parses from the dependency grammar. The shape of the constituent-structure parse tree can be represented by a series of new axioms that mirror the structure of the c1 (X, I, J) pred- 24 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith path1•2 (Q1 , Q2 ) ⊕= initial1•2 (Q1 , Q2 ). (63) path1•2 (Q1 , Q2 ) ⊕= path1•2 (P1 , P2 ) ⊗ arc1•2 (P1 , P2 , Q1 , Q2 , A1 , A2 ). (64) Fig. 26. Weighted finite-state transducers as the product of two weighted finitestate machines. 1 3 2 b ax x a y 4 5 (1,3) z ay bx (2,3) (1,4) (1,5) by az (2,4) (2,5) bz Fig. 27. A finite-state transducer that can be expressed as the PRODUCT of two finite-state automata. icate defining the constituent-structure grammar: proof1 (s, 0, 5), proof1 (np, 0, 1), proof1 (vp, 1, 5), proof1 (vp, 1, 3), proof1 (pp, 3, 5), and so on. Then we take the constrained PRODUCT of CKY that we used to describe lexicalized parsing (Figure 21) and, wherever there was a conclusion derived from c1 , we add a matching side condition that references proof1 . The critical rule (49) ends up looking like this: c1•2 (X1 , X2 , I, K) ⊕= binary1 (X1 , Y1 , Z1 ) ⊗ binary2 (X2 , Y2 , Z2 ) ⊗ (62) c1•2 (Y1 , Y2 , I, J) ⊗ c1•2 (Z1 , Z2 , J, K) if proof1 (X1 , I, K). The effect of this additional constraint is to disqualify any proof that does not match the constituent-structure grammar which we have fixed and encoded as proof1 axioms. The idea of partially constraining CFG derivations with some bracketing structure was explored by Pereira and Schabes (1992). 6.2 Axiom Generalization Axiom generalization is another way of manipulating products of weighted logic programs in a way that reveals the simple structures underlying a complex structure. Figure 26, which is intended to describe a weighted finite-state transducer, is close to the weighted logic program in Figure 15 that describes the intersection of two finite-state machines, but there are two differences. First, we have not forced the two symbols to be the same; instead, we wish to interpret A1 from the first expert as the transducer’s input symbol and A2 as the transducer’s output symbol. Second, we have merged initial1 (Q1 ) ⊗ initial2 (Q2 ) to the single product predicate initial1•2 (Q1 , Q2 ), and likewise for arc. As a first approximation, we can just Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 25 define arc1•2 (and, similarly, initial1•2 ) by a single rule of this form: arc1•2 (P1 , P2 , Q1 , Q2 , A1 , A2 ) ⊕= arc1 (P1 , Q1 , A1 ) ⊗ arc2 (P2 , Q2 , A2 ) (65) An example is given in Figure 27. Two finite-state machines, one with two states (a and b) and one with three states (x, y, and z), are shown—we are working over the Boolean semiring, so each arc in the figure corresponds to a true-valued arc axiom. The PRODUCT of these two experts in the manner of Figure 26 is a single finite-state transducer with six states. However, we can only describe a certain subset of finite-state transducers as the direct product of finite-state machines in this way. If we consider all possible Boolean-valued finite-state transducers with two symbols and one state, we have 16 possible transducers, but only 10 that can be “factored” as two independent finite-state machines, such as these three: 0 0 1 (0,1) 1 1 (1,1) 0 0 (0,1) 1 1 (0,0) (1,0) (1,1) (0,1) Six others, like the NOT transducer that outputs 1 given the input 0 and outputs 0 given the input 1, cannot be represented as the product of two FSMs. In many settings, limiting ourselves to the “factorable” finite-state transducers (or lexicalized grammars) can have conceptual or computational advantages. When this does not suffice, we can perform axiom generalization, which amounts to removing the requirement of Eq. 65 that the value of atomic propositions of the form arc1•2 be the product of an atomic proposition of the form arc1 and an atomic proposition of the form arc2 . If we directly define axioms of the form arc1•2 , we can describe transducers in their full generality. This represents a new way of thinking about the PRODUCT transformation. Thus far, we have considered the result of the PRODUCT transformation as a way of describing programs that work over two different structures. Axiom generalization suggests that we can consider the PRODUCT transformation as a way of taking two programs that work over individual structures and deriving a new program that works over a single more complicated structure that, in special cases, can be factored into two different structures. This is particularly relevant in the area of lexicalized grammars and parsing where the general, more complicated structure is what came first and the factored models which we have considered thus far arose later as special cases. Parsing algorithms and the PRODUCT transformation. Many parsing algorithms can be derived by using the PRODUCT transformation as a way of deriving programs that do not neatly factor into two parts. Lexicalized parsing is a simple example; Figure 28 derives a lexicalized parser by performing axiom generalization on Figure 21. The grammar production “P-with → with” can be represented by including the axiom unary1•2 (p, “with”), and the binary production S-saw → NP-Alice VP-saw can be represented by the axiom binary1•2 (s, “saw”, np, “alice”, vp, “saw”). Synchronous grammars are another instance in which the axiom generalization 26 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith goal1•2 ⊕= length(N) ⊗ start(S) ⊗ c1•2 (S, W, 0, N). c1•2 (X, W, I − 1, I) ⊕= unary1•2 (X, W) ⊗ string(I, W). (66) (67) c1•2 (X, W, I, K) ⊕= binary1•2 (X, W, Y, W1 , Z, W2 ) ⊗ c1•2 (Y, W1 , I, J) ⊗ c1•2 (Z, W2 , J, K).(68) Fig. 28. A algorithm for CKY over a general lexicalized grammar derived from Figure 21 by axiom generalization. goal ⊕= length(N) ⊗ start(S) ⊗ c(S, 0, N). (69) c(X, I, I) ⊕= unary(X, ) ⊗ pos(I). (70) c(X, I − 1, I) ⊕= unary(X, W) ⊗ string(I, W). (71) c(X, I, K) ⊕= binary(X, Y, Z) ⊗ c(Y, I, J) ⊗ c(Z, J, K). (72) Fig. 29. A variant of CKY that handles grammar productions of the form X → . goal1•2 ⊕= length1 (M) ⊗ length2 (N) ⊗ start1•2 (S) ⊗ (73) c1•2 (S, 0, N, 0, M). c1•2 (X, I − 1, I, J, J) ⊕= unary1•2 (X, W1 , ) ⊗ string1 (I, W1 ) ⊗ pos2 (J). (74) c1•2 (X, I, I, J − 1, J) ⊕= unary1•2 (X, , W2 ) ⊗ pos1 (I) ⊗ string2 (J, W2 ). (75) c1•2 (X, I − 1, I, J − 1, J) ⊕= unary1•2 (X, W1 , W2 ) ⊗ string1 (I, W1 ) ⊗ string2 (J, W2 ).(76) c1•2 (X1 , I1 , K1 , I2 , K2 ) ⊕= binary1•2 (X, Y, Z) ⊗ (77) c1•2 (Y, I1 , J1 , I2 , J2 ) ⊗ c1•2 (Z, J1 , K1 , J2 , K2 ). Fig. 30. A simple transduction grammar derived from Figure 29. c1•2 (X1 , I1 , K1 , I2 , K2 ) ⊕= inversion1•2 (X, Y, Z) ⊗ (78) c1•2 (Y, I1 , J1 , J2 , K2 ) ⊗ c1•2 (Z, J1 , K1 , I2 , J2 ). Fig. 31. By adding to Figure 30 these rule corresponding to the other way that the c1 and c2 antecedents may be merged in the PRODUCT transformation, we can describe an inversion transduction grammar. view is interesting. A synchronous grammar can be thought of as parsing two different sentences in two different languages with two different grammars using a single parse tree. For example, if X → YZ is a grammar production in one language and A → BC is a grammar production in another language, then X-A → Y-B Z-C is a possible grammar production in the synchronous grammar. A transduction grammar (Wu 1997), is a synchronous grammar which generates two isomorphic derivations with a trivial alignment between the nodes of those two derivations. We can describe a parser for a transduction grammar with the program Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 27 in Figure 30. Synchronous grammars need to be able to deal with situations in which a word in one language does not appear in the matching sentence in the other language; this is done by starting from the enriched CKY program in Figure 29 that can handle grammar productions of the form X → . In practice, transduction grammars do a bad job of aligning two sentences in different natural languages that are translations of each other, because it is often the case that two parts of a pair of sentences need to be in opposite positions relative to one another—in language one, the verb phrase might precede a prepositional phrase, and in language two, the corresponding verb phrase might follow the corresponding prepositional phrase. An inversion transduction grammar describes an alternate form of grammar production, which Wu (1997) writes as X → hYZi. This grammar production declares that if A1 and A2 simultaneously parse as Y in languages one and two (respectively) and B1 and B2 simultaneously parse as Z in languages one and two (respectively), then A1 B1 and B2 A2 simultaneously parse as Z. Somewhat surprisingly, this inversion production rule can be described using the alternate allowable way of merging the premises when the PRODUCT transformation is performed on two copies of the CKY algorithm, as discussed in §5.2 (see rule 50). By adding this alternate form as given in Figure 31, we can describe the algorithm for parsing with inversion transduction grammars described by Wu (1997). 7 The Entropy Semiring and Kullback-Leibler Divergence An important construct in information theory and machine learning is the KullbackLeibler (KL) divergence (Kullback and Leibler 1951). KL divergence is a function of two probability distributions over the same event space. It measures their dissimilarity, though it is not, strictly speaking, a distance (it is not symmetric). For two distributions p and q for random variable X ranging over events x ∈ X , KL divergence is defined as X p(X = x) (79) KL(pkq) = p(X = x) log q(X = x) x∈X X X = p(X = x) log p(X = x) − p(X = x) log q(X = x) (80) x∈X | x∈X {z −H(p) }| {z CE(pkq) } where H(p) denotes the Shannon entropy of the distribution p (Shannon 1948), a measure of uncertainty, and CE(pkq) denotes the cross-entropy between p and q.8 A full discussion of these information-theoretic quantities is out of scope for this paper; we note that they are widely used in statistical machine learning (Koller 8 In brief, the Shannon entropy of distribution p is the expected number of bits required to send a message drawn according to p under an optimal coding scheme. Cross-entropy is the average number of bits required to encode a message in the optimal coding scheme for q when messages are actually distributed according to p. Hence KL(pkq) = CE(pkq)−H(p) is the average number of extra bits required when the true distribution of messages is p but the coding scheme is based on q. Note that KL(pkp) = 0. If there is an event x ∈ X such that p(x) > 0 and q(x) = 0, then KL(pkq) = +∞. 28 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith and Friedman 2009). In this section, we first show how the entropy of p(P ), with P ranging over proofs of goal (the axioms corresponding to random variables A and I are suppressed here, for clarity), can be calculated using a weighted logic program, following Hwa (2004). We then describe a generalization of a result of Cortes et al. (2006) to show how to use PRODUCT to produce a weighted logic program for calculating the KL divergence between the two distributions induced by the WLPs. 7.1 Generalized Entropy Semiring The domain of the generalized entropy semiring is (R ∪ {+∞, −∞})3 . The multiplication and addition operations are defined as follows: hx1 , y1 , z1 i ⊕ hx2 , y2 , z2 i = hx1 + x2 , y1 + y2 , z1 + z2 i (81) hx1 , y1 , z1 i ⊗ hx2 , y2 , z2 i = hx1 x2 , x1 y2 + x2 y1 , z1 z2 i (82) These operations have the required closure, associativity, and commutativity properties previously discussed for semirings. See Cortes et al. (2006) for a proof which can be extended trivially to our generalized semiring. Suppose we have a weighted logic program such that the path-sum (in the hR≥0 ∪ {∞}, +, ×, 0, 1i semiring) is 1 (i.e., the value of the goal theorem is 1). If we map the weights of all axioms in the original program to new values in the generalized entropy semiring, we can use the new semiring to calculate the Shannon entropy of the distribution over proofs of goal: X − p(P = proof ) log p(P = proof ) (83) proof where x ranges over proofs of goal. The mapping is simply w 7→ hw, −w log w, 0i. (The third element of the semiring value is not needed here.) If we solve the new weighted logic program and achieve value hw0 , h0 , 0i for the goal theorem, then under our assumption that w0 = 1 (the value of goal in the original program in the real semiring), h0 is the entropy of the distribution over the proof random variable (given the axioms and goal). The formal result is given as a corollary in §7.2. This semiring can be used, for example, with the CKY algorithm from Figure 19. It makes the derivation of the tree entropy for context-free grammars (i.e., the entropy over the context-free derivations for an ambiguous string) automatic, and obviates the design of a specific algorithm for computing the tree entropy for probabilistic context-free grammars, as described in Hwa (2004). With the CKY algorithm, a proof proof in Eq. 83 represents a derivation in the grammar. Similarly, a weighted logic program describing a finite-state transducer (Figure 16) can be used to compute the entropy of hidden sequences for hidden Markov models as described by Hernando et al. (2005). We now relax the assumption that the sum of all proof scores is 1. Suppose that the value of the goal theorem in the generalized entropy semiring is (w0 , h0 , 0), with w0 6= 1. In this case, h0 is not the entropy of a proper probability distribution. Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 29 We can renormalize the scores of the proofs, u(proof ), by dividing by w0 , treating them as a proper conditional distribution (conditioning on the truth of the goal u(proof ) theorem); then the entropy of this conditional distribution, , is w0 X u(proof ) u(proof ) log (84) − w0 w0 proof   1  X = 0 − u(proof )(log u(proof ) − log w0 ) w proof   X 1  0 1 h0 = 0 h + (log w0 ) u(proof ) = 0 (h0 + w0 log w0 ) = 0 + log w0 w w w proof Therefore, whenever we can use weighted logic programming (in the real semiring) to calculate sums of proof scores, we can use the generalized entropy semiring to find the Shannon entropy of the (possibly renormalized) distribution over proofs. The renormalization uses w0 and h0 , two quantities that are calculated directly when we use the generalized entropy semiring. 7.2 KL Divergence Between Proof Distributions and PRODUCT Cortes et al. (2006) showed how to compute the KL divergence (also called relative entropy) between two distributions over strings defined by probabilistic FSA, using a construct similar to our generalized entropy semiring. We generalize that result to KL divergence over two proof distributions p(P ) and q(P ) given by a weighted logic program P. We assume in this discussion that the set of axioms with non-zero weights are identical under p and under q; the general setting where this does not hold is correctly handled, using a log 0 = −∞ and 0 log a = 0 for all a > 0. We abuse notation slightly and use p and q to denote the values of axioms, theorems, and proofs in the real semiring weighted logic programs used to calculate the sum of proof-scores for goal under axioms weighted according to p and q. Let Proofs(t) denote the set of logical proofs of a theorem t, and for x ∈ Proofs(t), let p(t)—respectively, q(t)—denote the score of the proof x: X p(t) = p(x) (85) x∈Proofs(t) q(t) X = q(x) (86) x∈Proofs(t) (87) We seek the KL divergence: KL(pkq) = X x∈Proofs(goal) p(x) log p(x) q(x) (88) In order to accomplish this calculation, we will first map the weights of axioms 30 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith under p and q into the generalized entropy semiring as follows, for any axiom a: hp(a), q(a)i 7→ hp(a), p(a) log q(a), q(a)i (89) For a theorem t, let X R(t) = p(x) log q(x) (90) x∈Proofs(t) Theorem 2 Solving P in the generalized entropy semiring with weights defined as above results in goal having value hp(goal), R(goal), q(goal)i. Proof: We will treat the weighted logic program as a set of equations with all lefthand-side variables grounded. We will use upper-case to refer to free variables (e.g., Z = hZ1 , . . .i) and lower-case to refer to grounded values (e.g., z = hz1 , . . .i). The range of values that variables Z can get is denoted by Rng(Z). The weighted logic program can be seen as a set of equations: M M c(w) = ai (w0 , z) ⊗ bi (w00 , z) (91) [c(w) ⊕= ai (w0 ,Z)⊗bi (w00 ,Z)]∈P,w0 ⊆w,w00 ⊆w z∈Rng(Z) (Note that any of w, w0 , w00 , and z may be empty.) We now show that the value achieved for c(w) when solving in the semiring is X hp(c(w)), p(x) log q(x), q(c(w))i (92) x∈Proofs(c(w)) where Proofs(c(w)) denotes the set of proofs for c(w). We will show that the solution of Equations 91 is the value in Equation 92 for c(w). For the first and third coordinates, this equality follows naturally because of the definition of the generalized entropy semiring: the first and third coordinates are equivalent to the non-negative real semiring used for summing over proof scores under the two value assignments p and q, respectively. Consider a particular ⊕-addend to the value of c(w), ai (w0 , z) ⊗ bi (w00 , z) (93) 0 0 0 = hp(ai (w , z)), R(ai (w , z)), q(ai (w , z))i ⊗ hp(bi (w00 , z)), R(bi (w00 , z)), q(bi (w00 , z))i * p(a (w0 , z))p(b (w00 , z)), i = i p(ai (w0 , z))R(bi (w00 , z)) + p(bi (w00 , z))R(ai (w0 , z)), q(ai (w0 , z)))q(bi (w00 , z)) (94) + (95) Consider the second coordinate. p(ai (w0 , z))R(bi (w00 , z)) + p(bi (w00 , z))R(ai (w0 , z))   X p(x) log q(x) = p(ai (w0 , z)) x∈Proofs(bi (w00 ,z)) (96) Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 31   X + p(bi (w00 , z)) p(x0 ) log q(x0 ) (97) x0 ∈Proofs(ai (w0 ,z))  =  X X p(x0 )  x0 ∈Proofs(ai (w0 ,z)) p(x) log q(x) x∈Proofs(bi (w00 ,z))   X + p(x) x∈Proofs(bi (w00 ,z)) = X x∈Proofs(bi X (98) x0 ∈Proofs(ai (w0 ,z)) X (w00 ,z)) p(x0 ) log q(x0 ) p(x)p(x0 ) log(q(x)q(x0 )) (99) x0 ∈Proofs(ai (w0 ,z)) Embedding the above in a ⊕-summation over z and a ⊕-summation over inference rule instantiations gives a ⊕-summation over proofs of c(w), X p(x) log q(x) (100) x∈Proofs(c(w)) which is R(c(w)) as desired.  Denote by (p̄, R̄, q̄) the value for goal in the generalized entropy semiring as discussed above, i.e., p̄ = p(goal), R̄ = R(goal), and q̄ = q(goal). If we wish to renormalize p by p̄ and q by q̄ to give proper distributions over proofs of goal (given axioms and goal), then   1 1 R̄ CE p q = − log q̄ (101) p̄ q̄ p̄ Noting that −H(p) = CE(pkp),   1 1 p q = CE(pkp) − CE(pkq) KL p̄ q̄ (102) we can solve for the KL divergence of two (possibly renormalized) distributions p and q using the above results. Alternatively, if the generalized KL divergence between unnormalized distributions is preferred (O’Sullivan 1998), note that (in the notation of the above):   X p(x) p(x) log − p(x) + q(x) = R̄ − p̄ + q̄ (103) q(x) x∈Proofs(goal) Cortes et al. describe how to compute KL divergence between two probabilistic finite-state automata with a single path per string (“unambiguous” automata). The authors make use of finite-state intersection (discussed above in §5.1). This suggests an analogous interpretation of the PRODUCT transformation for computing KL divergence between two weighted logic programs. Let P and Q be two instances of a weighted logic program, with possibly different different axiom weights. Assume we set the values of the axioms of P (ranging over a) to be hp(a), 0, 1i, and for Q we set them to h1, log q(a), q(a)i. If we take a PRODUCT of P and Q, using the “natural” pairing, then we end up with a program that computes hp(goal), R(goal), q(goal)i in the generalized entropy semiring, where 32 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith goal ⊕= path(P, Q) ⊗ final(Q). path(null, Q) ⊕= initial(Q). (105) path(P0 , Q) ⊕= path(P, P0 ) ⊗ biarc(P, P0 , Q, A, B). (106) (104) Fig. 32. Weighted finite-state transducer where arriving at a certain state depends on the last two states. null serves as a place-holder for the non-state prior to the initial state. goal1•2 ⊕= path1•2 (P, Q) ⊗ final1 (Q) ⊗ final2 (Q). path1•2 (null, Q) ⊕= initial1 (Q) ⊗ initial2 (Q). path1•2 (P0 , Q) ⊕= path1•2 (P, P0 ) ⊗ arc1 (P0 , Q, A, B) ⊗ biarc2 (P, P0 , Q, A, B) (107) (108) (109) Fig. 33. The PRODUCT program of Figure 16 with Figure 32, with constraints that match proofs according to states and emissions sequences. R(·) is specified in Eq. 90. These quantities can be used to compute KL divergence as specified in Eq. 102. This is a direct result of Theorem 2. 7.3 KL Divergence and Projections We can use PRODUCT to calculate KL divergence between proof distributions even when P and Q are not two instances of the same program. We consider cases where the proofs of P and the proofs of Q have a shared semantics, that is, each proof of either P or Q maps to an event in some “interpretation space.” As an example, consider the WLP in Figure 16 describing a weighted finite-state transducer. In a more general formulation, where each state depends on the previous N states visited, rather than just the single most recent state. This modification is reflected in Figure 32 for N = 2. The axiom biarc(P0 , Q, P, A, B) is to be interpreted as: “if the last two states were Q and P0 , transfer to state P while reading symbol A and emitting the symbol B.” Since the two programs have different axioms, the spaces of their respective proofs are different. However, both programs have identical semantics to a proof: a proof (in either program) corresponds to a sequence of states that the transducers go through together with the reading of a symbol and the emission of another symbol. Running PRODUCT on the WFST in Figure 16 (we call it P) and the WFST in Figure 32 (we call it Q) with a particular pairing and constraints (such that the paths are identical) yields the program in Figure 33. If we let the axioms a in P have the values hp(a), 0, 1i and the axioms a in Q the values h1, log q(a), q(a)i, then the resulting PRODUCT program in Figure 33, as implied by Theorem 2, calculates the KL divergence between two distributions over the set of state paths: one which is defined using a finite-state transducer with N = 1 and the other with N = 2. We now generalize this idea for two different programs P and Q. We assume Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 33 that PRODUCT is applied in such a way that axioms from P are paired only with axioms from Q, and vice versa. Further, each proof in the PRODUCT program must decompose into exactly one proof in P and one proof in Q.9 For a proof in the PRODUCT program, y, we define πP (y) (πQ (y)) to be the projection of y to a proof in P (Q). The “projection” of a proof is a separation of the proof which uses coupled theorems and axioms into theorems and axioms of only one of the programs. For example, projecting a proof y in the product program in Figure 33 yields two proofs: πP (y) describes a sequence of transitions through the transducer with N = 1 and πQ (y) describes a sequence of transitions through the transducer with N = 2; yet both proofs correspond to the same sequence of states. In the generalized entropy semiring, we set the values of the axioms of P to be hp(a), 0, 1i, and for Q we set them to h1, log q(a), q(a)i. The PRODUCT program computes hp(goal1 ), R(goal1·2 ), q(goal2 )i. This time, the summation in R(goal1·2 ) is over proofs which are implicitly paired: R(goal1·2 ) = X p(πP (y)) log q(πQ (y)) (110) y∈Proofs(goal1·2 ) The quantities p(goal1 ), R(goal1·2 ), and q(goal2 ) can be used as before to compute the KL divergence between the distributions over the shared “interpretation space” of the proofs in the two programs. This technique is only correct when interpretations are in a one-to-one correspondence with the proofs in P and with the proofs in Q, and PRODUCT is applied so that equivalently-interpretable proofs in the two programs are paired. We note that in the general case, the problem of computing KL divergence between two arbitrary distributions is hard. For example, with Markov networks, there are restrictions, which resemble the restrictions we pose, of clique decomposition (Koller and Friedman 2009). 8 Conclusion We have described a framework for dynamic programming algorithms whose solutions correspond to proof values in two constrained weighted logic programs. Our framework includes a program transformation, PRODUCT, which combines the two weighted logic programs that compute over two structures into a single weighted logic program for a joint proof. Appropriate constraints, encoded intuitively as variable unification or side conditions in the weighted logic program, are then added manually. The framework naturally captures and permits generalization of many existing algorithms. We have shown how variations on the the program transformation enable to include a larger set of algorithms as the result of the program transformation. We have concluded by showing how the program transformation 9 Note that these constraints are satisfied in the case of two identical programs with the “natural” pairing, as in §7.2. 34 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith can be used to interpret the computation of Kullback-Leibler divergence for two weighted logic programs which are defined over an identical interpretation space. Acknowledgments The authors acknowledge helpful comments from the two anonymous reviewers, Jason Eisner, Rebecca Hwa, Adam Lopez, Alberto Pettorossi, Frank Pfenning, Sylvia Rebholz and David Smith. This research was supported by an NSF graduate fellowship to the second author and NSF grant IIS-0713265 and an IBM faculty award to the third author. References Charniak, E., Knight, K., and Yamada, K. 2003. Syntax-based language models for machine translation. In Proc. of MT Summit. Chiang, D. 2007. Hierarchical phrase-based translation. Computational Linguistics 32, 2, 201–228. Cocke, J. and Schwartz, J. T. 1970. Programming languages and their compilers: Preliminary notes. Tech. rep., Courant Institute of Mathematical Sciences, New York University. Cohen, S. B. and Smith, N. A. 2007. Joint morphological and syntactic disambiguation. In Proc. of EMNLP-CoNLL. Cortes, C., Mohri, M., Rastogi, A., and Riley, M. D. 2006. Efficient computation of the relative entropy of probabilistic automata. LATIN 2006: Theoretical Informatics, 323–336. Dempster, A., Laird, N., and Rubin, D. 1977. Maximum likelihood estimation from incomplete data via the EM algorithm. Journal of the Royal Statistical Society B 39, 1–38. Durbin, R., Eddy, S., Krogh, A., and Mitchison, G. 1998. Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cambridge University Press. Eisner, J. 1997. Bilexical grammars and a cubic-time probabilistic parser. In Proc. of IWPT. Eisner, J. and Blatz, J. 2007. Program transformations for optimization of parsing algorithms and other weighted logic programs. In Proc. of Formal Grammar. Eisner, J., Goldlust, E., and Smith, N. A. 2005. Compiling Comp Ling: Practical weighted dynamic programming and the Dyna language. In Proc. of HLT-EMNLP. Eisner, J. and Satta, G. 1999. Efficient parsing for bilexical context-free grammars and head automaton grammars. In Proc. of ACL. Felzenszwalb, P. F. and McAllester, D. 2007. The generalized A∗ architecture. Journal of Artificial Intelligence Research 29, 153–190. Gaifman, H. 1965. Dependency systems and phrase-structure systems. Information and Control 8. Goodman, J. 1999. Semiring parsing. Computational Linguistics 25, 4, 573–605. Hernando, D., Crespi, V., and Cybenko, G. 2005. Efficient computation of the hidden Markov model entropy for a given observation sequence. IEEE Transactions on Information Theory 51, 7, 2681–2685. Hinton, G. E. 2002. Training products of experts by minimizing contrastive divergence. Neural Computation 14, 1771–1800. Products of Weighted Logic ProgramsTo appear in Theory and Practice of Logic Programming (TPLP). 35 Hopcroft, J. E. and Ullman, J. D. 1979. Introduction to Automata Theory, Languages, and Computation. Addison-Wesley. Huang, L. and Chiang, D. 2005. Better k-best parsing. In Proc. of IWPT. Hwa, R. 2004. Sample selection for statistical parsing. Computational Linguistics 30, 3. Kasami, T. 1965. An efficient recognition and syntax-analysis algorithm for context-free languages. Tech. Rep. AFCRL-65-758, Air Force Cambridge Research Lab. Klein, D. and Manning, C. D. 2003. Fast exact inference with a factored model for natural language parsing. In Advances in NIPS 15. Klein, D. and Manning, C. D. 2004. Parsing and hypergraphs. New developments in parsing technology, 351–372. Koehn, P., Och, F. J., and Marcu, D. 2003. Statistical phrase-based translation. In Proc. of HLT-NAACL. 127–133. Koller, D. and Friedman, N. 2009. Probabilistic Graphical Models: Principles and Techniques. The MIT Press. Kullback, S. and Leibler, R. A. 1951. On information and sufficiency. Annals of Mathematical Statistics 22, 79–86. Liang, P., Klein, D., and Jordan, M. 2008. Agreement-based learning. In Advances in Neural Information Processing Systems 20, J. Platt, D. Koller, Y. Singer, and S. Roweis, Eds. MIT Press, Cambridge, MA, 913–920. Lopez, A. 2009. Translation as weighted deduction. In Proc. of EACL. Manning, C. and Schütze, H. 1999. Foundations of Statistical Natural Language Processing. MIT Press. McAllester, D. 1999. On the complexity analysis of static analyses. In Proc. of Static Analysis Symposium. Mohri, M. 1997. Finite-state transducers in language and speech processing. Computational Linguistics 23, 2, 269–311. O’Sullivan, J. A. 1998. Alternating minimization algorithms: from Blahut-Armijo to Expectation-Maximization. In Codes, Curves, and Signals: Common Threads in Communications, A. Vardy, Ed. Kluwer, 173–192. Pereira, F. C. N. and Riley, M. D. 1997. Speech recognition by composition of weighted finite automata. In Finite-State Language Processing, E. Roche and Y. Schabes, Eds. MIT Press, 431–453. Pereira, F. C. N. and Schabes, Y. 1992. Inside-outside reestimation from partially bracketed corpora. In Proc. of ACL. 128–135. Pettorossi, A. 1999. Synthesis and transformation of logic programs using unfold/fold proofs. The Journal of Logic Programming 41, 2-3 (December), 197–230. Pettorossi, A. and Proietti, M. 1994. Transformation of logic programs: Foundations and techniques. J. Log. Program. 19, 261–320. Shannon, C. 1948. A mathematical theory of communication. Bell System Technical Journal 27, 379–423. Shieber, S. M., Schabes, Y., and Pereira, F. C. N. 1995. Principles and implementation of deductive parsing. Journal of Logic Programming 24, 1–2, 3–36. Sikkel, K. 1997. Parsing Schemata. Springer-Verlag. Smith, A., Cohn, T., and Osborne, M. 2005. Logarithmic opinion pools for conditional random fields. In Proc. of ACL. Smith, D. A. and Smith, N. A. 2004. Bilingual parsing with factored estimation: Using English to parse Korean. In Proc. of EMNLP. Sutton, C. and McCallum, A. 2005. Piecewise training of undirected models. In Proc. of UAI. 36 Shay B. Cohen, Robert J. Simmons, and Noah A. Smith Tarjan, R. E. 1981. A unified approach to path problems. Journal of the ACM 28, 3, 577–93. Wu, D. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics 23, 3, 377–404. Younger, D. H. 1967. Recognition and parsing of context-free languages in time n3 . Information and Control 10, 2.
6cs.PL
arXiv:1401.3424v4 [math.ST] 1 Apr 2016 Marčenko-Pastur law for Tyler’s M-estimator Teng Zhang Department of Mathematics, University of Central Florida Orlando, Florida 32816, USA [email protected] Xiuyuan Cheng Applied Mathematics Program, Yale University, New Haven, CT 06511, USA [email protected] Amit Singer Department of Mathematics and PACM, Princeton University Princeton, New Jersey 08544, USA [email protected] April 4, 2016 Abstract This paper studies the limiting behavior of Tyler’s M-estimator for the scatter matrix, in the regime that the number of samples n and their dimension p both go to infinity, and p/n converges to a constant y with 0 < y < 1. We prove that when the data samples x1 , . . . , xn are identically and independently generated from the Gaussian distribution N (0, I), ¯ the operator norm Pn of the⊤ difference between a properly scaled Tyler’s Mestimator and i=1 xi xi /n tends to zero. As a result, the spectral distribution of Tyler’s M-estimator converges weakly to the Marčenko-Pastur distribution. 1 Introduction Many statistical estimators and signal processing algorithms require the estimation of the covariance matrix of the data samples. When the underlying distribution of the data samples x1 , . . . , xn ∈ Rp is assumed to have Pnzero mean, a commonly used estimator is the sample covariance matrix Sn = i=1 xi x⊤ i /n. 1 However, the estimator Sn is sensitive to outliers, and performs poorly in terms of statistical efficiency (i.e., it has a large variance) for heavy-tailed distributions, e.g., when the tail decays slower than the Gaussian tail. A popular robust covariance estimator is an M-estimator introduced by Tyler [20], denoted by Σ̂, which is the unique solution to n Σ̂ = p X xi x⊤ i , tr(Σ̂) = 1. −1 x n i=1 x⊤ i i Σ̂ (1) Tyler’s M-estimator gives the “shape” of the covariance, but is missing its magnitude. However, for many applications the “shape” of the covariance suffices, for example, the principal components can be obtained from the “shape”. Compared with the sample covariance estimator, Tyler’s M-estimator is more robust to heavy-tailed elliptical distributions. The density function of elliptical distributions in Rp takes the form f (x; Σ, µ) = |Σ|−1/2 g{(x − µ)⊤ Σ−1 (x − µ)}, R∞ where g is some nonnegative function such that 0 xp−1 g(x) dx is finite. This family of distributions is a natural generalization of the Gaussian distribution by allowing heavier or lighter tails while maintaining the elliptical geometry of the equidensity contours. Elliptical distributions are considered important in portfolio theory and financial data, and we refer to the work by El Karoui [11, Section 4] for further discussion. Besides, elliptical distributions are used by Ollila and Tyler [19] in modeling radar data, where the empirical distributions are heavy-tailed because of outliers. Tyler [20] showed that when a data set follows an unknown elliptical distribution (with mean zero), Tyler’s M-estimator is the most robust covariance estimator in the sense of minimizing the maximum asymptotic variance. This property suggests that Tyler’s M-estimator should be more accurate than the sample covariance estimator for elliptically distributed data. Empirically, it has been shown to outperform the sample covariance estimator in applications such as finance in the work by Frahm and Jaekel [13], anomaly detection in wireless sensor networks by Chen et al. [4], antenna array processing by Ollila and Koivune [18], and radar detection by Ollila and Tyler [19]. 1.1 Asymptotic analysis in a high-dimensional setting Many scientific domains customarily deal with sets of high dimensional data samples, and therefore it is increasingly common to work with data sets where the number of variables, p, is of the same order of magnitude as the number of observations, n. Under this high-dimensional setting, the asymptotic spectral properties of Sn at the limit of infinite number of samples and infinite dimensions have been well studied by Johnstone [15]. A noticeable example is the convergence of the spectral distribution. Denoting the eigenvalues of a matrix A by λ1 (A), . . . , λn (A), its spectral distribution is a discrete probability 2 measure n 1X P = P (·|A) = δλ (A) n i=1 i with δs denoting Dirac measure at s ∈ R. Marčenko and Pastur [16] showed that when the entries of {xi }ni=1 are Gaussian independent identically distributed random variables with mean 0 and variance 1, p, n → ∞ and p/n → y, where 0 < y ≤ 1, the spectral distribution of the eigenvalues of Sn converges weakly to the Marčenko-Pastur distribution defined by p 1 y (y+ − x)(x − y− ) √ ρMP,y (x) = 1[y− ,y+ ] , where y± = (1 ± y)2 . (2) 2π x Tyler’s M-estimator is closely related to and can be considered as a special case of Maronna’s M-estimator, which is defined by n Σ̄ = 1X −1 u(x⊤ xi )xi x⊤ i Σ̄ i n i=1 (3) for a nonnegative function u : [0, ∞) → [0, ∞). The properties of Maronna’s M-estimator in the high-dimensional regime when p, n → ∞, p/n → y and 0 < y < 1 have been analyzed in recent works by Couillet et al. [7, 8], which obtained convergence results for a properly scaled Maronna’s M-estimator under the assumptions that u(x) is nonnegative, nonincreasing and continuous; xu(x) is nondecreasing and bounded and supx xu(x) > 1. Moreover, spiked random matrix models were also studied by Couillet [5]. However, these results do not apply to Tyler’s M-estimator, although Frahm and Jaekel [13] have conjectured that the spectral distribution converges weakly to the Marčenko-Pastur distribution. Some works focused on the performance of Tyler’s M-estimator for the case p, n → ∞ and p/n → 0: p Dümbgen p [10] showed that the condition number p/n), and Frahm and Glombek [12] of Tyler’s estimator is 1 + 4 p/n + o( p showed that the spectral distribution of n/p(Σ̄ − I) converges weakly to a semicircle distribution. 1.2 Main results In this paper, we analyze Tyler’s M-estimator in the high-dimensional setting. Our main results, Theorem 2.3 and Corollary 2.4, show that as p, n → ∞ and p/n → y, 0 < y < 1, the spectral distribution of a properly scaled Tyler’s M-estimator converges weakly to the Marčenko-Pastur distribution ρMP,y (x). Based on the properties of Tyler’s M-estimator, this paper analyzes the spectral distribution when data samples are i.i.d. drawn from other distributions, such as elliptical distributions. When data samples are generated from elliptical distributions, the spectral distribution of the sample covariance estimator has been studied by El Karoui [11, Theorem 2]. Compared to Corollary 2.4, the limiting spectral distribution of Sn is much more complicated, and therefore our result might be more applicable in practice. 3 High-dimensional analysis of Maronna’s M-estimator of the covariance are generally obtained by showing that the operator norm of the difference between M-estimator and a standard Wishart matrix (or sample covariance matrix) tends to 0: Dümbgen [10] proved it by a linear expansion of the M-estimator, and Couillet et al. [7, 8] proved it by representing Maronna’s M-estimator as a weighted sum of xi x⊤ i and prove the uniform convergence of the weights. We follow the same direction while giving an alternate proof for the convergence of the weights, by considering the weights as the solution to an optimization problem, which can handle Tyler’s M-estimator that is not covered by the results in Couillet et al. [7, 8]. We remark that this approach can also be applied to Maronna’s M-estimator to prove some of the results in Couillet et al. [7, 8]. The rest of the paper is organized as follows. In Section 2 we introduce the representation of Tyler’s M-estimator as a linear combination of xi x⊤ i and present the main result that when the data set is i.i.d. sampled from the Gaussian distribution N (0, I), a properly scaled Tyler’s is asymptotically equivalent ¯ to Sn in the sense that kpΣ̂ − Sn k → 0. As a result, the spectral distribution of Tyler’s M-estimator converges weakly to the Marčenko-Pastur distribution. We also extend the result to elliptical distributions. The technical proofs are given in Section 3. While some Lemmas and technical proofs are also used by Couillet et al. [7, 8] (for example, Lemma 3.2 and the analysis in the proof of Theorem 2.3 are similar to Couillet et al. [7, Lemma 2, Theorem 1] and Couillet et al. [8, Lemma 6], we still include them for the completeness of the paper. As for notations, we use c, c′ , C, C ′ to denote any fixed constants as p, n → ∞ (though they may depend on y). Depending on the context, they might denote different values in different equations. 2 Tyler’s M-estimator in the High-dimensional Regime We introduce the representation of Tyler’s M-estimator as a linear combination of xi x⊤ i in Section 2.1, and present the main result in Section 2.2 that when the data set is i.i.d. sampled from the Gaussian distribution N (0, I), kpΣ̂ − Sn k ¯ converges to 0 almost surely. Based on this observation, we prove that the spectral distribution of Tyler’s M-estimator converges weakly the MarčenkoPastur distribution in Section 2.3. The generalization of the results to more general settings is also discussed in Section 2.3. 2.1 Properties of Tyler’s M-estimator The analysis for Tyler’s M-estimator in this paper is based on the following representation, whose proof is deferred to Section 3. We remark that equation (5) in Lemma 2.1 has appeared in the work by Wiesel [21, (27)] and Hardt and Moitra [14, Section A] as “covariance estimation in scaled Gaussian distributions” and “Barthe’s convex program”, but its connection to Tyler’s M-estimator has not been rigorously justified yet. 4 Lemma 2.1. Tyler’s M-estimator can be written as Σ̂ = n X i=1 n  . X ŵi xi x⊤ ŵi xi x⊤ tr i , i (4) i=1 where {ŵi }ni=1 are uniquely defined by (ŵ1 , . . . , ŵn ) = arg min P wi >0, 2.2 n i=1 wi =1 − n X ln wi + i=1 n X  n wi xi x⊤ ln det i . p i=1 (5) Isotropic Gaussian Distribution In this subsection, we assume that {xi }ni=1 ⊂ Rp are i.i.d. drawn from N (0, I). ¯ The main result, Theorem 2.3, characterizes the convergence and convergence rate of Tyler’s M-estimator to Sn in terms of the operator norm. Its proof applies Lemma 2.2, whose proof is rather technical and deferred to Section 3. Tyler’s M-estimator does not exist when p > n (see the argument by Zhang [22, Theorem III.1]) and it is not unique when p = n (one may check that when xi = ei for all 1 ≤ i ≤ p, all diagonal matrices with trace 1 satisfy (1)). As a result, throughout the paper we assume y < 1. Lemma 2.2. If {xi }ni=1 are i.i.d. sampled from N (0, I), then max1≤i≤n |n ŵi − ¯ 1| converges to 0 almost surely as p, n → ∞: There exist C, c, c′ > 0 such that ′ for any ε < c ,   2 Pr max |n ŵi − 1| ≤ ε ≥ 1 − Cne−cε n . (6) 1≤i≤n Theorem 2.3. Suppose that {xi }ni=1 are i.i.d. sampled from N (0, I), p, n → ∞ ¯ and p/n = y, where 0 < y < 1, then the operator norm of the difference between Sn and a scaled Tyler’s M-estimator converges to 0 almost surely, and there exist C, c, c′ > 0 such that for any ε < c′ , ! n 2 1X ⊤ xi xi ≤ ε ≥ 1 − Cne−cε n . (7) Pr p Σ̂ − n i=1 Theorem 2.3 implies that all first order properties of the sample covariance matrix extend to Tyler’s estimator. The strategy of the proof for Theorem 2.3 is as follows. According to Lemma 2.1, a scaled Tyler’s is a linear Pn M-estimator ⊤ combination of xi x⊤ , i.e., it can be written as ŵ x x (up to a scali i=1 i i i ing). Then Lemma 2.2 shows that nŵi converges to 1 uniformly. Based on the following matrix analysis, Theorem 2.3 is concluded. Proof of Theorem 2.3. We first prove that for ε < c′ , ! n n X 2 1X ⊤ ⊤ Pr ŵi xi xi − xi xi ≤ ε ≥ 1 − Cne−cε n . n i=1 i=1 5 (8) Let Bn = Pn i=1 (ŵi − 1/n)xi x⊤ i = kBn k = sup v ⊤ Bn v = sup kvk=1 ≤ sup Pn n X kvk=1 i=1 n X kvk=1 i=1 ŵ − 1 1 n ŵi xi x⊤ i − i=1 ∞ (ŵi − Pn i=1 xi x⊤ i /n, then 1 )(v ⊤ xi )2 n n (v ⊤ xi )2 ≤ knŵ − 1k∞ k 1X xi x⊤ i k. n i=1 Since knŵ − 1k∞ → 0 with probability P estimated in (6), and Davidson and n Szarek [9, Theorem II.13] showed that k i=1 xi x⊤ i /nk is bounded above by √ 2 (1 + 2 y) with probability 1 − C exp(−cn), (8) is proved. Second, since n X ŵi xi x⊤ ≤ i i=1 Pr k n X i=1 n X i=1 ŵi xi x⊤ i − n X xi x⊤ i /n + ′ ŵi xi x⊤ i k< C xi x⊤ i /n , i=1 i=1 ! n X > 1 − Cn exp(−cn). (9) Pn Pn ⊤ Besides, tr( i=1 ŵi xi x⊤ i ) = i=1 ŵi xi xi → p in the same rate as in (9): applying the concentration of high-dimensional Gaussian measure on the sphere by Barvinok [2, Corollary 2.3], we have )# ) ( n " ( n X X ⊤ ⊤ ŵi xi xi > p/(1 − ε) (10) ŵi xi xi < p(1 − ε) , Pr max Pr i=1 i=1  ≤ max Pr     2 min kxi k2 < p(1 − ε) , Pr max kxi k2 > p/(1 − ε) < ne−ε p/4 . 1≤i≤n 1≤i≤n Combining (9), (10) and (4), n X i=1 ŵi xi x⊤ i − p Σ̂ = n X ŵi xi x⊤ i i=1 ( ) n X ŵi xi x⊤ 1 − p/tr( i ) (11) i=1 converges in the same rate as specified in (8). (7) is then proved by combining (8), (11) and the triangle inequality. p From the probabilistic estimation (7) we obtain a convergence rate of O( ln n/n). The logarithmic factor is due to a “max” bound of {ŵi }ni=1 in Lemma 2.2, while in √ fact, an “average” bound is expected. As a result, we conjecture that this ln n factor could be possibly removed by a more rigorous argument. 2.3 More general distributions and spectral distribution We remark that Theorem 2.3 can be extended from the setting of the normal distribution N (0, I) to any elliptical distribution µp , which is characterized ¯ by the probability density function µp (x) = C(gp ) det(Tp )−1/2 gp (x⊤ Tp−1 x), 6 Tp is a positive definite matrix in Rp×p , gp : [0, ∞) → [0, ∞) satisfies Rwhere ∞ p−1 < ∞, and C(gp ) is a normalization parameter that only depends 0 gp (x)x Pn on gp . Then ktr(Tp )Σ̂ − i=1 xi x⊤ i /nk → 0 almost surely as p, n → ∞. The analysis is based on Theorem 2.3, the affine equivariance property of Tyler’s M-estimator, and the fact that Tyler’s M-estimator is unchanged if {xi }ni=1 are replaced by {ci xi }ni=1 . Another direction of generalization of Theorem 2.3 is the model by Couillet et al. [7]: The elements of {xi }ni=1 are i.i.d. sampled from an either real or circularly symmetric complex distribution with E(xij ) = 0, E(x2ij ) = 1, and E(|xij |8+η ) < α for some η, α > 0. Then, following the proof in this paper (while replacing Pn Lemma 3.2 by [7, Lemma 2]), one can show that p Σ̂ − i=1 xi x⊤ i /n → 0 almost surely as p, n → ∞. We have the following results on the weak convergence of the spectral distribution of Tyler’s M-estimator, where the first part proves the conjecture by Frahm and Jaekel [13]. Corollary 2.4. • If {xi }ni=1 are i.i.d. sampled from N (0, I), then the spec¯ tral measure P (·|pΣ̂) converges weakly to the Marčenko-Pastur distribution. • If {xi }ni=1 are i.i.d. sampled from an elliptical distribution C(gp )gp (x⊤ Tp−1 x) such that the spectral measure of Tp converges weakly to a distribution H on R. Then the spectral measure P (·|tr(Tp )Σ̂) converges weakly to a R probabilistic measure ρ whose Stieltjes transform s(z) = 1/(x − z)ρ( dx) (z ∈ C \ R) is given implicitly by Z 1 s(z) = dH(t). t{1 − y − yz s(z)} − z P This corollary can be proved by combining ktr(Tp )Σ̂ − ni=1 xi x⊤ i /nk → 0, the analysis on the perturbation of eigenvalues by Bhatia [3, Corollary III.4.2], Pn /n by Marčenko and Pastur [16], Bai and the spectral measure of i=1 xi x⊤ i Silverstein [16, 1, (6.1.2)] and Slutsky’s Lemma. 3 3.1 Proof of Lemmas Proof of Lemma 2.1 We start with the definition (ẑ1 , . . . , ẑn ) = P arg min ln det n i=1 and Σ̂z = zi =1 n X n X eẑi xi x⊤ i . i=1 7 i=1 e zi xi x⊤ i  (12) (13) The solution to (12) is unique, which follows from the convexity of the objective function (see Wiesel [21, Lemma 4]). Besides, noticing the equivaP lence between (12) and (5) (by plugging wi = ezi /( ni=1 ezi ) and zi = ln wi − Pn ( i=1 ln wi − 1)/n, there exists c1 > 0 such that Σ̂z = c1 Σ̂. Next we will prove that Σ̂z satisfies n X i=1 xi x⊤ i −1 x⊤ i Σ̂z xi = cΣ̂z , for some c > 0. (14) By checking the directional derivative of the objective function in (12), for Pn any (δ1 , . . . , δn ) with i=1 δi = 0, n X −1 δi eẑi x⊤ i Σ̂z xi = 0. i=1 Therefore, there exists c2 such that −1 eẑi x⊤ i Σ̂z xi = c2 , for all 1 ≤ i ≤ n. (15) Therefore, (14) is proved by applying (15) and (13): n X i=1 xi x⊤ i −1 x⊤ i Σ̂z xi = n X eẑi xi x⊤ i /c2 = Σ̂z /c2 , i=1 Since Σ̂z = c1 Σ̂, (14) also holds when Σ̂z is replaced by Σ̂: n X i=1 xi x⊤ i −1 x x⊤ i i Σ̂ = cΣ̂, for some c > 0. (16) At last, we will prove that Σ̂ satisfies the definition of Tyler’s M-estimator in (1), that is, the constant c in (16) is given by c = n/p. For the objective function n X −1 ln(x⊤ xi ) + c ln det(Σ), F (Σ) = i Σ i=1 its derivative with respect to Σ−1 is given by n X i=1 ⊤ −1 xi )−1 xi − cΣ. x⊤ i (xi Σ Therefore, Σ̂ is a stationary point of F (Σ). Since F (Σ) is geodesically convex (argument follows directly from Wiesel [21] and Zhang [22]), Σ̂ is the global minimizer of F (Σ). the minimizer of F (Σ) exists only when c = n/p. Since F (aI) = PnHowever, ⊤ i=1 ln(xi xi ) − n ln a + c p ln a, we have ( as a → 0, if c > n/p F (aI) → −∞ . as a → ∞, if c < n/p Therefore, the constant c in (16) is given by c = n/p, and Lemma 2.1 is proved. 8 3.2 Proof of Lemma 2.2 We start with an outline of the proof, which consists of three parts. First, we rewrite the constrained optimization problem (5) to the problem of finding the root of g(w), which will be defined in (17). Since the root of g(w) is nŵ − 1, we only need to show the convergence of the root of g(w). Second, we will show that g(0) converges to 0, ∇g(0̄) is large and the variation of ∇g(w) is bounded. ¯ ¯ Finally, we will use a perturbation analysis and the observations on g(0) and ¯ ∇g(w) to show that the root of g(w) converges to 0. ¯ The proof depends on Lemma 3.2, Lemma 3.3 and Lemma 3.1, and their proofs are postponed to subsequent sections. Lemma 3.1. For a function f (w) : Rp → Rp , assume that ∇f (0) = I, and ¯ k∇f (w) − ∇f (0)k∞ = maxi≤i≤p k∇fi (w) − ∇fi (0)k∞ < C5 kwk∞ for kwk∞ ≤ ¯ ¯ 1, and kf (0)k∞ < min(1/9C5 , 1/3). Then there exists w̃ such that kw̃k∞ < ¯ 3kf (0)k∞ and f (w̃) = 0. ¯ P Lemma 3.2. If xi ∼ N (0, I) for all 1 ≤ i ≤ n, and S = ni=1 xi x⊤ i /n, then ¯ there exists c, C, c′ > 0 such that for any ε < c′ ,   2 1 −1 S x − 1| < ε ≥ 1 − Cne−cε n . Pr max | x⊤ i i 1≤i≤n p −1 Lemma 3.3. For the n × n matrix A defined by Aij = (x⊤ xj )2 /(n p), (a) i S kAk∞ < 2 with probability 1 − Cn exp(−cn). (b) There exists c = c(p, n) > 0 and C2 = C2 (y) > 0 such that k(I − A + c11⊤ )−1 k∞ < C2 with probability 1 − Cn exp(−cn). We start the first part of the proof with the construction of g(w). We let g(w) = ∇G(w + 1), (17) where G(w) = − n X i=1 ln wi + n n X n c0 X wi xi x⊤ ) + wi − n)2 , ln det( ( i p 2 i=1 i=1 (18) and the constant c0 will be specified later before (34). It is easyP to prove that the minimizer P of G(w) and the zeros of ∇G(w) n n must satisfy i=1 wi = n (otherwise nw/( i=1 wi ) is a better minimizer and ∇G(w) is nonzero). Therefore minimizing (18) is equivalent P P Pn to minimizing − ni=1 ln wi + n/p · ln det( ni=1 wi xi x⊤ ) with constraint i i=1 wi = n, which is the same as (5) except for the constraint. Noticing that a scaling of w increases Pn Pn ) by a constant only depending on the − i=1 ln wi + n/p · ln det( i=1 wi xi x⊤ i scale, the minimizer of (18) is unique and it is nŵ, where ŵ is defined in (5). By the convexity of its equivalent problem (12), the root of g(w) is also unique and it is nŵ − 1. 9 For the second part of the proof, we start by proving that g(0) is small. By ¯ calculation, the i-th component of function g(w) is gi (w) = − n n X X 1 n −1 wi . wi xi x⊤ x i + c0 + x⊤ i ) i (nS + wi + 1 p i=1 i=1 Applying Lemma 3.2, 2 Pr (kg(0)k∞ < ε) ≥ 1 − Cne−cε n . ¯ (19) Now we will prove that ∇g(0̄) is bounded from below. By calculation, its (i, j)th entry is  ∇g(w) i,j = I(i = j) n −1 o2 X 1 n n ⊤ ⊤ x j + c0 . w x x nS + x − i i i i (wi + 1)2 p i=1 Applying Lemma 3.3, k{∇g(0̄)}−1 k∞ < C2 with probability 1 − Cne−cn . (20) Now we bound the variation of ∇g(w) in the region kwk∞ < 1/2. Apply |1/(wi + 1)2 − 1| < 3|wi − 1| ≤ 3kwk∞ and coordinatewise comparison, n |∇i,j g(w) − ∇i,j g(0)| ≤ I(i = j) (3kwk∞ ) + 3kwk∞ · |Aij |. ¯ p Therefore, the variation of ∇g(w) is bounded by k∇g(w) − ∇g(0̄)k∞ < (3 + 3nkAk∞ /p)kwk∞ . (21) At last we finish the third part of the proof of Lemma 2.2 by applying Lemma 3.1 to f (w) = {∇g(0̄)}−1 g(w/2). It is easy to verify that ∇f (0) = I. ¯ Due to (19) and (20), kf (0)k∞ ≤ k(∇g(0̄))−1 k∞ kg(0̄)k∞ → 0 in the same rate as ¯ −cn in (19) and kf (0)k∞ < min(1/9C5 , 1/3) holds with probability 1−Cne . Due ¯ to (19), (21), and the boundedness of kAk∞ (Lemma 3.3), k∇f (w) − ∇(0̄)k∞ < C5 kwk∞ also holds with probability 1 − Cne−cn . Therefore the assumption in Lemma 3.1 holds with probability 1 − Cne−cn and there exists w̃ such that f (w̃) = 0 and kw̃k∞ < 3kf (0)k∞ . (22) ¯ When f (w̃) = 0, we have g(2w̃) = 0 and by previous discussion 2w̃ = nŵ−1. therefore (22) gives knŵ − 1k∞ < 6kf (0)k∞ . ¯ Since kf (0)k∞ converges to 0 in the rate as in (19), knŵ − 1k∞ converges ¯ in the same rate and Lemma 2.2 is proved. 10 3.2.1 Proof of Lemma 3.1 Proof. When kwk∞ ≤ 1, fj (w) − fj (0) = Z 1 t=0 ej w⊤ , ∇f (t w) dt (23) Z 1 ej w⊤ , ∇f (t w) − ∇f (0) + I dt = wj + w⊤ {∇f (t w) − ∇f (0)}ej dt ¯ ¯ t=0 t=0 Z 1 w⊤ {∇f (t w) − ∇f (0)}k∞ ≤ wj + C5 kwk2∞ . ≤wj + k ¯ t=0 = Z 1 Similarly fj (w) − fj (0) ≥ −C5 kwk2∞ + wj . (24) To prove it, we consider the continuous mapping h(w) = w−f (w)/(4+9C5 ) and will prove that h maps A to itself, where A = {w : w ∈ [−3η, 3η]n } and η = kf (0)k∞ . 1. |wi | < 2η. Then apply (23) and (24) (they are applicable since for any w ∈ A, kwk∞ ≤ 1), we have |fi (w)| < |fi (0)| + C5 kwk2∞ + |wi | ≤ η + C5 (3η)2 + 3η < (4 + 9C5 )η (η 2 < η since η < 1). Therefore, |hi (w)| ≤ |wi | + |fi (w)|/(4 + 9C5 ) ≤ 3η. 2. wi > 2η, then applying (24), fi (w) ≥ −|fi (0)| + wi − C5 kwk2∞ ≥ −η + 2η − C5 (3η)2 . ¯ Since η < 1/9C5 , we have fi (w) < 0 and therefore hi (w) ≤ wi ≤ 3η. Similar to case 1 we can prove that hi (w) ≥ −3η. Therefore |hi (w)| < 3η. 3. Similar to case 2, when wi < −2η, |hi (w)| < 3η. Therefore the continuous mapping h maps the convex, compact set A to itself. By Schauder fixed point theorem, h(x) has a fixed point in A and Lemma 3.1 is proved with w̃ being the fixed point. 3.2.2 Proof of Lemma 3.2 Assuming the SVD decomposition of X is X = U ΣV ⊤ , where U ∈ Rn×p and U ⊤ U = I. Since xi ∼ N (0, I) for all 1 ≤ i ≤ n, U is uniformly distributed over ¯ the space of all orthogonal n × p matrices. Since 1 XS −1 X = (U ΣV ⊤ )( V Σ2 V ⊤ )−1 (U ΣV ⊤ ), n (25) 2 if we write the row of U by u1 , . . . , un , then n1 xi S −1 xi = u⊤ i ui = kui k . Since U can be considered as the first p columns of a random n×n orthogonal matrix (with haar measure over the set of all n × n orthogonal matrices), ui can be considered as the first p entries from a random vector of length n that is sampled from the uniform sphere in Rn . 11 Pp Pn Therefore, kui k2 ∼ j=1 gj2 / j=1 gj2 for i.i.d. random variables {gj }nj=1 ∼ N (0, 1). Applying the the concentration result by Barvinok [2, Corollary 2.3], we have ) ( n X 2 n 2 ≤ e−ε n/4 (26) gi ≥ Pr 1 − ε i=1 and ( Pr n X i=1 ) gi2 ≤ n(1 − ε) 2 ≤ e−ε n/4 , (27) therefore ) (   p X p(1 − ε)2 p p 2 2 Pr gi ≤ ≥ Pr p(1 − ε) ≤ ≤ ku1 k ≤ n n(1 − ε)2 1−ε i=1 ( ) n X 2 2 n gi2 ≤ + Pr n(1 − ε) ≤ ≥ 1 − 2e−ε p/4 − 2e−ε n/4 . 1 − ε i=1 For ε ≤ 0.1, we have   n p p o 1 −1 S x − 1| ≤ ε ≥ 1 − n Pr |ku1 k2 − | > ε Pr max | x⊤ i i 1≤i≤n p n n    2 p p(1 − ε/3) ≤ ku1 k2 ≤ (28) ≥1 − n 1 − Pr n n(1 − ε/3)2 2 ≥1 − 2ne−ε p/36 2 − 2ne−ε n/36 , (29) where the second inequality follows from 1−3ε ≤ (1−ε)2 and 1/(1 − ε)2 ≤ 1+3ε. 3.2.3 Proof of Lemma 3.3 P (a) Since kAk∞ = max1≤i≤n ( 1≤j≤n Aij ), and X 1≤j≤n Aij = X 1≤j≤n X 1 ⊤ −1 −1 −1 −1 xi S xj x⊤ xi = x⊤ ( xj x⊤ xi /np j S i S j )S np 1≤j≤n (30) −1 =x⊤ (nS)S −1 xi /np i S = −1 x⊤ xi /p, i S (31) it follows from (29) with ε = 0.1 that kAk∞ < 2 holds with probability 1 − Cn exp(−cn). (b) We first prove that there exists C3 = C3 (y) such that kA − c0 11⊤ k∞ ≤ C3 < 1 with probability 1 − Cn exp(−cn). (32) We start with the proof of (32) with another lemma: Lemma 3.4. There exists a c4 > 0 such that with probability 1 − C exp(−cn), n X √ I(x⊤ 1 xj > c4 p) > 0.75n. j=1 12 Davidson and S. Szarek [9, Theorem II.13] showed that There exists C4 = −1 C4 (y) such that kSk < C4 with probability 1−Cn exp(−cn). Therefore x⊤ xj ≥ i S ⊤ xi xj /C4 and Lemma 3.4 implies that for any 1 ≤ i ≤ n: n X j=1 √ −1 I(x⊤ xj > c4 p/C4 ) > 0.75 with probability 1 − C exp(−cn). (33) i S Let c0 = (c4 /C4 )2 /n, then (33) implies X X −1 |Ai,j − c| ≤ |Ai,j | − 0.25c n ≤ x⊤ xi /p − 0.25(c4 /C4 )2 , (34) i S 1≤j≤n 1≤j≤n where the last step follows from (31). −1 Applying the estimation of x⊤ xi /p in (29) and a union bound argument i S over all 1 ≤ i ≤ n to (34), (32) is proved for C3 = 1 + η − 0.25(c4 /C4 )2 . Lemma 3.3(b) follows from (32) with C2 = 1/(1 − C3 ), where the expansion of (I − A + c11⊤ )−1 exists since kA + c11⊤ k ≤ kA + c11⊤ k∞ < 1. Applying kB1 B2 k∞ ≤ kB1 k∞ kB2 k∞ , we have k(I−A+c11⊤ )−1 k∞ = k 3.2.4 ∞ X k=0 (c11⊤ −A)k k∞ ≤ ∞ X k=0 kc11⊤ −Akk∞ ≤ ∞ X k=0 C3i = 1 . 1 − C3 (35) Proof of Lemma 3.4 We first show that there exists c4 such that for all p, √ E{I(|x⊤ 1 x2 | > c4 p)} ≥ 0.85. (36) WLOG we rotate x1 such that it is nonzero only at the first coordinate, and x2 = (g1 , g2 , ..., gp ) where gi ∼ N (0, 1). Then |x⊤ 1 x2 | = |g1 | kx1 k. Notice that kx1 k2 is the sum of p√independent χ21 distribution and Eχ21 = 1, 1 − Ce−cn . Besides, by central√limit theorem, kx1 k ≤ 2p with probability √ −1 Pr(|g1 | > 2 c4 ) ≥ 0.85 for c4 = Φ (1 − 0.85/2)/ 2. Therefore (36) is proved by combining the estimations on |g1 |, x1 and |x⊤ 1 x2 | = |g1 | kx1 k. To obtain Lemma 3.4 from (36), we apply Hoeffding’s inequality to the √ indicator function I(|x⊤ i xj | > c4 p) over all 1 ≤ j ≤ n, j 6= i. 4 Summary We showed that Tyler’s M-estimator is asymptotically equivalent to Sn in the sense that kpΣ̂ − Sn k → 0 as p, n → ∞ and p/n → y, where 0 < y < 1 and data samples follow the distribution of N (0, I). We also proved the conjecture ¯ that the spectral distribution of Tyler’s M-estimator converges weakly to the Marčenko-Pastur distribution, and extended the results to elliptical distributions. 13 There are several possible future directions of this work. First, it would be interesting to analyze the second order statistics of Tyler’s M-estimator, considering that Couillet [6] has already investigated Maronna’s M-estimators. Second, we would like to theoretically quantify the behavior of Tyler’s M-estimator in the spiked covariance model by Couillet [5], which includes the analysis of the distribution of the top eigenvalue for the null cases and the analysis of the non-null case. A recent work by Morales-Jimenez et al. [17] on the non-null case introduced a mixture model that consists of a Gaussian distribution and some deterministic or random outliers, and analyzed the performance of Maronna’s M-estimator. Analyzing the performance of Tyler’s M-estimator in this model would be another possible future direction. Acknowledgements A. Singer was partially supported by Award Number FA9550-12-1-0317 and FA9550-13-1-0076 from AFOSR, by Award Number R01GM090200 from the NIGMS, and by Award Number LTR DTD 06-05-2012 from the Simons Foundation. References [1] Z. Bai and J. Silverstein. Spectral Analysis of Large Dimensional Random Matrices. Springer series in statistics. Springer, 2009. [2] A. Barvinok. Math 710: Measure concentration. Lecture notes, Department of Mathematics, University of Michigan, 2005. [3] R. Bhatia. Matrix Analysis. Number 169 in Graduate Texts in Mathematics. Springer, New York, 1997. [4] Y. Chen, A. Wiesel, and A. Hero. Robust shrinkage estimation of highdimensional covariance matrices. Signal Processing, IEEE Transactions on, 59(9):4097 –4107, sept. 2011. [5] R. Couillet. Robust spiked random matrices and a robust g-music estimator. Journal of Multivariate Analysis, 140:139 – 161, 2015. [6] R. Couillet, A. Kammoun, and F. Pascal. Second order statistics of robust estimators of scatter. application to GLRT detection for elliptical signals. Journal of Multivariate Analysis, 143:249 – 274, 2016. [7] R. Couillet, F. Pascal, and J. Silverstein. Robust estimates of covariance matrices in the large dimensional regime. Information Theory, IEEE Transactions on, 60(11):7269–7278, Nov 2014. [8] R. Couillet, F. Pascal, and J. W. Silverstein. The random matrix regime of Maronna’s M-estimator with elliptically distributed samples. Journal of Multivariate Analysis, 139:56 – 78, 2015. 14 [9] K. Davidson and S. Szarek. Local operator theory, random matrices and banach spaces. Handbook of the Geometry of Banach Spaces, 1:317, 2001. [10] L. Dümbgen. On Tyler’s M-functional of scatter in high dimension. Annals of the Institute of Statistical Mathematics, 50(3):471–491, 1998. [11] N. El Karoui. Concentration of measure and spectra of random matrices: Applications to correlation matrices, elliptical distributions and beyond. The Annals of Applied Probability, 19(6):2362–2405, 12 2009. [12] G. Frahm and K. Glombek. Semicircle law of Tyler’s M-estimator for scatter. Statistics & Probability Letters, 82(5):959 – 964, 2012. [13] G. Frahm and U. Jaekel. Tyler’s M-estimator, random matrix theory, and generalized elliptical distributions with applications to finance. Technical report, 2007. [14] M. Hardt and A. Moitra. Algorithms and hardness for robust subspace recovery. In S. Shalev-Shwartz and I. Steinwart, editors, COLT, volume 30 of JMLR Proceedings, pages 354–375. JMLR.org, 2013. [15] I. M. Johnstone. High dimensional statistical inference and random matrices. In International Congress of Mathematicians. Vol. I, pages 307–333. Eur. Math. Soc., Zürich, 2007. [16] V. A. Marčenko and L. A. Pastur. Distribution of eigenvalues for some sets of random matrices. Mathematics of the USSR-Sbornik, 1(4):457, 1967. [17] D. Morales-Jimenez, R. Couillet, and M. McKay. Large dimensional analysis of robust M-estimators of covariance with outliers. Signal Processing, IEEE Transactions on, 63(21):5784–5797, Nov 2015. [18] E. Ollila and V. Koivunen. Robust antenna array processing using Mestimators of pseudo-covariance. In Personal, Indoor and Mobile Radio Communications, 2003. PIMRC 2003. 14th IEEE Proceedings on, volume 3, pages 2659–2663 vol.3, 2003. [19] E. Ollila and D. Tyler. Distribution-free detection under complex elliptically symmetric clutter distribution. In Sensor Array and Multichannel Signal Processing Workshop (SAM), 2012 IEEE 7th, pages 413–416, June 2012. [20] D. E. Tyler. A distribution-free M-estimator of multivariate scatter. The Annals of Statistics, 15(1):pp. 234–251, 1987. [21] A. Wiesel. Geodesic convexity and covariance estimation. Signal Processing, IEEE Transactions on, 60(12):6182 –6189, dec. 2012. [22] T. Zhang. Robust subspace recovery by Tyler’s M-estimator. Information and Inference, 2015. 15
10math.ST
Query-adaptive Video Summarization via Quality-aware Relevance Estimation Arun Balajee Vasudevan∗† , Michael Gygli∗†‡ , Anna Volokitin† , Luc Van Gool†§ † ETH § KU Leuven ‡ Gifs.com Zurich {arunv,gygli,anna.volokitin,vangool}@vision.ee.ethz.ch arXiv:1705.00581v2 [cs.CV] 28 Sep 2017 ABSTRACT Although the problem of automatic video summarization has recently received a lot of attention, the problem of creating a video summary that also highlights elements relevant to a search query has been less studied. We address this problem by posing queryrelevant summarization as a video frame subset selection problem, which lets us optimise for summaries which are simultaneously diverse, representative of the entire video, and relevant to a text query. We quantify relevance by measuring the distance between frames and queries in a common textual-visual semantic embedding space induced by a neural network. In addition, we extend the model to capture query-independent properties, such as frame quality. We compare our method against previous state of the art on textual-visual embeddings for thumbnail selection and show that our model outperforms them on relevance prediction. Furthermore, we introduce a new dataset, annotated with diversity and query-specific relevance labels. On this dataset, we train and test our complete model for video summarization and show that it outperforms standard baselines such as Maximal Marginal Relevance. 1 INTRODUCTION Video recording devices have become omnipresent. Most of the videos taken with smartphones, surveillance cameras and wearable cameras are recorded with a capture first, filter later mentality. However, most raw videos never end up getting curated and remain too long, shaky, redundant and boring to watch. This raises new challenges in searching both within and across videos. The problem of making videos content more accessible has spurred research in automatic tagging [2, 39, 51] and video summarization [1, 15, 26, 27, 31, 36, 49, 57, 69]. In automatic tagging, the goal is to predict meta-data in form of tags, which makes videos searchable via text queries. Video summarization, on the other hand, aims at making videos more accessible by reducing them to a few interesting and representative frames [26, 31] or shots [15, 56]. This paper combines the goals of summarising videos and makes them searchable with text. Specifically, we propose a novel method that generates video summaries adapted to a text query (See Fig. 1). ∗ Authors contributed equally Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. MM’17, October 23–27, 2017, Mountain View, CA, USA. © 2017 Copyright held by the owner/author(s). Publication rights licensed to ACM. ISBN 978-1-4503-4906-2/17/10. . . $15.00 DOI: https://doi.org/10.1145/3123266.3123297 Figure 1: Our query-adaptive video summarization model picks frames that are relevant to the query while also giving a sense of entire video. We want to summarise a video of an ironman competition, in which participants swim, bike and run. Query-adapted summaries are representative by showing all three sports, while placing more focus on the frames matching the query. Our approach improves previous works in the area of textual-visual embeddings [28, 34] and proposes an extension of an existing video summarization method using submodular mixtures [15] for creating summaries that are query-adaptive. Our method for creating query-relevant summaries consists of two parts. We first develop a relevance model which allows us to rank frames of a video according to their relevance given a text query. Relevance is computed as the sum of the cosine similarity between embeddings of frames and text queries in a learned visualsemantic embedding space and a query-independent term. While the embedding captures semantic similarity between video frames and text queries, the query-independent term predicts relevance based on the quality, composition and the interestingness of the content itself. We train this model on a large dataset of image search data [20] and our newly introduced Relevance and Diversity dataset (Section 5). The second part of the summarization system is a framework for optimising the selected set of frames not only for relevance, but also for representativeness and diversity using a submodular mixture of objectives. Figure 2 shows an overview of our complete pipeline. We publish our codes and demos † and make the following contributions: • Several improvements on learning a textual-visual embedding for thumbnail selection compared to the work by Liu et al. [34]. These include better alignment of the learning objective to the task at test time and modeling the text queries using LSTMs, fetching significant performance gains. † https://github.com/arunbalajeev/query-video-summary • A way to model semantic similarity and quality aspects of frames jointly, leading to better performance compared to using the similarity to text queries only. • We adapt the submodular mixtures model for video summarization by Gygli et al. [15] to create query-adaptive and diverse summaries using our frame-based relevance model. • A new video thumbnail dataset providing query relevance and diversity labels. As the judgements are subjective, we collect multiple annotations per video and analyse the consistency of the obtained labelling. 2 RELATED WORK The goal of video summarization is to select a subset of frames that gives a user an idea of the video’s content at a glance [60]. To find informative frames for this task, two dominant approaches exist: (i) modelling generic frame interestingness [17, 31] or (ii) using additional information such as the video title or a text query to find relevant frames [33, 34, 56]. In this work we combine the two into one model and make several contributions for query-adaptive relevance prediction. Such models are related to automatic tagging [2, 39, 51], textual-visual embeddings [13, 34, 54] and image description [3, 5, 6, 9, 11, 24, 25, 38] . In the following we discuss approaches for video summarization, generic interestingness prediction models and previous works for obtaining embeddings. Video summarization. Video summarization methods can be broadly classified into abstractive and extractive approaches. Abstractive or compositional approaches transform the initial video into a more compact and appealing representation, e.g. hyperlapses [29], montages [58] or video synopses [50]. The goal of extractive methods is instead to select an informative subset of keyframes [26, 27, 31, 63] or video segments [15, 36] from the initial video. Our method is extractive. Extractive methods need to optimise at least two properties of the summary: the quality of the selected frames and their diversity [14, 15, 52]. Sometimes, additional objectives such as temporal uniformity [15] and relevance [52] are also optimised. The simplest approach to obtain a representative and diverse summary is to cluster videos into events and select the best frame per event [7]. More sophisticated approaches jointly optimise for importance and diversity by using determinantal point process (DPPs) [14, 52, 68] or submodular mixtures [15, 32]. Most related to our paper is the work of Sharghi et al. [52], who present an approach for query-adaptive video summarization using DPPs. Their method however limits to a small, fixed set of concepts such as car or flower. The authors leave handling of unconstrained queries, as in our approach, for future work. In this work, we formulate video summarization as a maximisation problem over a set of submodular functions, following [15]. Frame quality/interestingness. Most methods that predict frame interestingness are based on supervised learning. The prediction problem can be formulated as a classification [49], regression [31, 66], or, as is now most common, as a ranking problem [17, 57, 59, 65]. To simplify the task, some approaches assume the domain of the video given and train a model for each domain [49, 57, 65]. An alternative approach based on unsupervised learning, proposed by Xiong et al. [64], detects “snap points” by using a web image prior. Their model considers frames suitable as keyframes if the composition of the frames matches the composition of the web images, regardless of the frame content. Our approach is partially inspired by this work in that it predicts relevance even in the absence of a query, but relies on supervised learning. Unconstrained Textual-visual models. Several methods exist that can retrieve images given unconstrained text or vice versa [9, 11, 13, 18, 24, 25, 38]. These typically project both modalities into a joint embedding space [13], where semantic similarity can be compared using a measure like cosine similarity. Word2vec [41] and GloVe [48] are popular choices to obtain the embeddings of text. Deep image features are then mapped to the same space via a learned projection. Once both modalities are in the same space, they may be easily compared [13]. A multi-modal semantic embedding space is often used by Zero-shot learning approaches [13, 23, 47] to predict test labels which are unseen in the training. Habibian et al. [18], in the same spirit, propose zero-shot recognition of events in videos by learning a video representation that aligns text, audio and video features. Similarly, Liu et al. [34] use textual-visual embeddings for video thumbnail selection. Our relevance model is based on Liu et al. [34], but we provide several important improvements. (i) Rather than keeping the word representation fixed, we jointly optimise the word and image projection. (ii) Instead of embedding each word separately, we train an LSTM model that combines a complete query into one single embedding vector, thus it even learns multi-word combinations such as visit to lake and Star Wars movie. (iii) In contrast to Liu et al. [34], we directly optimise the target objective. Our experiments show that these changes lead to significantly better performance in predicting relevant thumbnails. 3 METHOD FOR RELEVANCE PREDICTION The goal of this work is to introduce a method to automatically select a set of video thumbnails that are both relevant with respect to a query, but also diverse enough to represent the video. To later optimise relevance and diversity jointly, we first need a way to evaluate the relevance of frames. Our relevance model learns a projection of video frames v and text queries t into the same embedding space. We denote the projection of t and v as t and v, respectively. Once trained, the relevance of a frame v given a query t can be estimated via some similarity measure. As [13], we use the cosine similarity s(t, v) = t·v . ktkkvk (1) While this lets us assess the semantic relevance of a frame w.r.t. a query, it is also possible to make a prediction on the suitability as thumbnails a priori, based on the frame quality, composition, etc. [64]. Thus, we propose to extend above notion of relevance and model the quality aspects of thumbnails explicitly by computing the final relevance as the sum of the embedding similarity and the query-independent frame quality term, i.e. r (t, v) = s(t, v) + qv , (2) where qv is a query-independent score determining the suitability of v as a thumbnail, based on the quality of a frame. Query LSTM E.g.: cooking channel Objectives Semantic embedding space Query similarity Summary frames Similarity Frame level Features Quality score VGG‐19 FC2 space Quality Diversity score Submodular Maximization Diversity Representativeness score Repr Figure 2: Overview of our approach. We show how a summary is created from an example video and the query Cooking channel. We obtain a query adaptive summary by selecting a set of keyframes from the video using our quality-aware relevance model and submodular mixtures, as explained in Sec. 3 and 4. In the following, we investigate how to formulate the task of obtaining the embeddings t and v, as well as qv . 3.1 Training objective Intuitively, our model should be able to answer “What is the best thumbnail for this query?”. Thus, the problem of picking the best thumbnail for a video is naturally formulated as a ranking problem. We desire that the embedding vectors of a query and frame that are a good match are more similar than ones of the same query and a non-relevant frame∗ . Thus, our model should learn to satisfy the rank constraint that given a query t, the relevance score of the relevant frame v + is higher than the relevance score of the irrelevant frame v − : r (t, v + ) > r (t, v − ). (3) Alternatively, we can train the model by requiring that both the similarity score and the quality score of the relevant frame are higher than for the irrelevant frame explicitly, rather than imposing a constraint only on their sum, as above. In this case we would be imposing the two following constraints: s(t, v+ ) > s(t, v− ) qv + > qv − . (4) Experimentally, we find that training with these explicit constraints leads to slightly improved performance (See Tab. 1). In order to impose these constraints and train the model, we define the loss as  loss(t, v + , v − ) = lp max 0, γ − s(t, v+ ) + s(t, v− ) (5) + lp (max (0, γ − qv + + qv − )) , where lp is a cost function and γ is a margin parameter. We follow [17] and use a Huber loss for lp , i.e. the robust version of an l 2 loss. Next, we describe how to parametrize the t, v and qv , so that they can be learned. 3.2 Text and Frame Representation We use a convolutional neural network (CNN) for predicting v and qv , while t is obtained via a recurrent neural network. To jointly learn the parameters of these networks, we use a Siamese ranking network, trained with triplets of (t, v + , v − ) where the weights for the subnets predicting v + and v − are shared. We provide the model architecture in supplementary material. We now describe the textual representation t and the image representations v and qv in more detail. Textual representation. As a feature representation t of the textual query t, we first project each word of the query into a 300dimensional semantic space using the word2vec model [42], which is trained on GoogleNews dataset. We fine-tune the word2vec model using the unique queries from the Bing Clickture dataset [20] as sentences. Then, we encode the individual word representations into a single fixed-length embedding using an LSTM [19]. We use a many-to-one prediction, where the model outputs a fixed length output at the final time-step. This allows us to emphasize visually informative words and handle phrases. Image representation. To represent the image, we leverage the feature representations of a pre-trained VGG-19 network [53] on ImageNet [8]. We replace the softmax layer(1000 nodes) of VGG-19 network with a linear layer M with 301 dimensions. The first 300 dimensions are used as the embedding v, while the last dimension represents the quality score qv . 4 SUMMARIZATION MODEL We use the framework of submodular optimization to create summaries that take into account multiple objectives [32]. In this framework, summarization is posed as the problem of selecting a subset (in our case, of frames) y∗ that maximizes a linear combination of submodular objective functions f(x V , y) = [f 1 (x V , y), ..., fn (x V , y)]T . Specifically, y∗ = arg max wT f(x V , y), (6) y∈YV ∗ Liu et al. [34] does the inverse. It poses the problem as learning to assign a higher similarity to corresponding frame and query than to the same frame and a random query. Thus, the model learns to answer the question “what is a good query for this image?”. where YV denote the set of all possible solutions y and x V the features of video V. In this work, we assume that the cardinality |y| is fixed to some value k (we use k = 5 in our experiments). For non-negative weights w, the objective in Eq. (6) is submodular [30], meaning that it can be optimized near-optimally in an efficient way using a greedy algorithm with lazy evaluations [43, 46]. Objective functions. We choose a small set of objective functions, each capturing different aspects of the summary. Í (1) Query similarity f(·, ·) = v ∈y s(t, v) where t is the query embedding, v is frame embedding and s(·, ·) denotes the cosine similarity defined in Eq. (1). Í (2) Quality score f(·, ·) = v ∈y qv , where qv represents score that is based on the quality of v as a thumbnail. This model scores the image relevance in a query-independent manner based on properties such as contrast, composition, etc. (3) Diversity of the elements in the summary Í f(x V , y) = i ∈y min D x V (i, j), according to some dissimij <i larity measure D. We use the Euclidean distance in of the FC2 features of the VGG-19 network for D † . (4) Representativeness [15]. This objective favors selecting the medoid frames of a video, such that the visually frequent frames in the video are represented in the summary. Weight learning. To learn the weights w in Eq. (6), ground truth summaries for query-video pairs are required. Previous methods typically only optimized for relevance [34] or used small datasets with limited vocabularies [52]. Thus, to be able to train our model, we collected a new dataset with relevance and diversity annotations, which we introduce in the next Section. If relevance and diversity labels are known, we can estimate the optimal mixing weights of the submodular functions through subgradient descent [32]. In order to directly optimize for the F1score used at test time, we use a locally modular approximation based on the procedure of [45] and optimize the weights using AdaGrad [10]. 5 RELEVANCE AND DIVERSITY DATASET (RAD) We collected a dataset with query relevance and diversity annotation to let us train and evaluate query-relevant summaries. Our dataset consists of 200 videos, each of which was retrieved given a different query. Using Amazon Mechanical Turk (AMT) we first annotate the video frames with query relevance labels, and then partition the frames into clusters according to visual similarity. These kind of labels were used previously in the MediaEval diverse social images challenge [21] and enabled evaluation of the automatic methods for creating relevant and diverse summaries. To select a representative sample of queries and videos for the dataset, we used the following procedure: We take the top YouTube queries between 2008 and 2016 from 22 different categories as seed queries‡ . These queries are typically rather short and generic concepts, so to obtain longer, more realistic queries we use YouTube auto-complete to suggest phrases. Using this approach we collect 200 queries. Some examples are brock lesnar vs big show, taylor † Derivation of submodularity of this objective is provided in the suppl. ‡ https://www.google.com/trends/explore swift out of the woods, etc. For each query, we take the top video result with a duration of 2 to 3 minutes. To annotate the videos, we set up two consecutive tasks on AMT. All videos are sampled at one frame per second. In the first task, a worker is asked to label each frame with its relevance w.r.t. the given query. Options for answers are “Very Good”,“Good”, “Not good” and “Trash”, where trash indicates that the frame is both irrelevant and low-quality (e.g. blurred, bad contrast, etc.). After annotating the relevance, the worker is asked to distribute the frames into clusters according to their visual similarity. We obtain one clustering per worker, where each clustering consists of mutually exclusive subsets of video frames as clusters. The number of clusters in the clustering is chosen by the worker. Each video is annotated by 5 different people and a total of 48 subjects participated in the annotation. To ensure high-quality annotations, we defined a qualification task, where we check the results manually to ensure the workers provide good annotations. Only workers who pass this test are allowed to take further assignments. 5.1 Analysis We now analyse the two kinds of annotations obtained through this procedure and describe how we merge these annotations into one set of ground truth labels per video. Label distributions. The distribution of relevance labels is “Very Good”: 17.55%, “Good”: 57.40%, “Not good”: 12.31% and “Trash”: 12.72%. The minimum, maximum and mean number of clusters per video are 4.9, 25.2 and 13.4 respectively over all videos of RAD. Relevance annotation consistency. Given the inherent subjectivity of the task, we want to know whether annotators agree with each other about the query relevance of frames. To do this, we follow previous work [16, 22, 62] and compute the Spearman rank correlation (ρ) between the relevance scores of different subjects, splitting five annotations of each video into two groups of two and three raters each. We take all split combination to find mean ρ for a video. Our dataset has an average correlation of ρ = 0.73 over all videos, where 1 is a perfect correlation while 0 would indicate no consistency in the scores. On the related task of event-specific image importance, using five annotators, consistency is only ρ = 0.4 [62]. Thus, we can be confident that our relevance labels are of high quality. Cluster consistency. To the best of our knowledge, we are the first to annotate multiple clusterings per video and look into the consistency of multiple annotators. MediaEval, for example, used multiple relevance labels but only one clustering [21]. Various ways of measuring the consistency of clusterings exist, e.g. Variation of Information, Normalised Mutual Information or the Rand index (See Wagner and Wagner [61] for an excellent overview). In the following we propose to use Normalised Mutual Information (NMI), an information theoretic measure [12] which is the ratio of the mutual information between two clusterings (I (C, C 0 )) and the sum of entropies of the clusterings (H (C) + H (C 0 )): N MI (C, C 0 ) = 2 · I (C, C 0 ) , H (C) + H (C 0 ) (7) Settings Method Cost LSTM Metrics Quality HIT@1 VG or G Spear Corr. mAP Random - - - 57.17 ± 1.5 - 0.5780 Loss of Liu et al. Ours: L1 Ours: Huber Loss of Liu et al. + LSTM Ours: Huber + LSTM Ours: Frame quality only Q expl i Ours: Huber + LSTM + Q impl i Ours: Huber + LSTM + Q expl i l1 l1 × × × X X × X X × × × × × X X X 68.75 68.09 68.35 70.62 72.63 65.95 70.76 74.76 0.186 0.209 0.279 0.270 0.367 0.236 0.371 0.376 0.6308 0.6348 0.6446 0.6507 0.6685 0.6315 0.6657 0.6712 lhuber l1 lhuber lhuber lhuber lhuber Table 1: Comparison of different model configurations trained on a subset of the Clickture dataset and fine-tuned on our Video Thumbnail dataset (RAD). We report the HIT@1 (fraction of times we select a “Very Good” or “Good” thumbnail), the Spearman correlation of our model predictions with the true candidate thumbnail scores, and mean average precision. The Huber+LSTM+Q expl i model performs best. We chose NMI over the more recently proposed Variation of Information (VI) [40], as NMI has a fixed range ([0, 1]) while still being closely related to VI (see supplementary material). Our dataset has a cluster consistency of 0.54. Since NMI is 0 if two clusterings are independent and 1 iff they are identical, we see that our annotators have a high degree of agreement. Ground truth For evaluation on the test videos, we create a single ground truth annotation for each video. We merge the five relevance annotations as well as the clustering of each query-video pair. For the final ground truth of relevance prediction, we require the labels be either positive or negative for each video frame. We map all “Very Good” labels to 1, “Good” labels to 0.5 and “Not Good” and “Trash” labels to 0. We compute the mean of the five relevance annotation labels and label the frame as positive if the mean is ≥ 0.5 and as negative otherwise. To merge clustering annotations, we calculate NMI between all pairs of clustering and choose the clustering with the highest mean NMI, i.e. the most prototypical cluster. An example of relevance and clustering annotation is provided in Fig. 6. 6 CONFIGURATION TESTING Before comparing our proposed relevance model against state of the art in Sec. 7, we first analyze our model performance using different objectives, cost functions and text representation. For evaluation, we use Query-dependent Thumbnail Selection Dataset (QTS) provided by [34]. The dataset contains 20 candidate thumbnails for each video, each of which is labeled one of the five: Very Good (VG), Good (G), Fair (F), Bad (B), or Very Bad (VB). We evaluate on the available 749 query-video pairs. To transform the categorical labels to numerical values, we use the same mapping as [34]. Evaluation metrics. As evaluation metrics, we are using HIT@1 and mean Average Precision (mAP) as reported and defined in Liu et al. [34], as well as the Spearman’s Rank Correlation. HIT@1 is computed as the hit ratio for the highest ranked thumbnail. Training dataset. For training, we use two datasets: (i) the Bing Clickture dataset [20] and (ii) the RAD dataset (Sec. 5). Clickture is a large dataset consisting of queries and retrieved images from Bing Image search. The annotation is in form of triplets (K, Q, C) meaning that the image K was clicked C times in the search results of the query Q. This dataset is well suited for training our relevance model, since our task is the retrieval of relevant keyframes from a video, given a text query. It is, however, from the image and not the video domain. Thus, we additionally fine-tune the models on the complete RAD dataset consisting of 200 query-video pairs. From each query-video pair, we sample an equi number of positive and negative frames to give equal weight to each video. In total, we use 0.5M triplets (as in Sec. 3.2) from the Clickture and 14K triplets from the RAD for training. Implementation details. We preprocess the images as in [53]. We truncate the number of words in the query at 14, as a tradeoff between the mean and maximum query length in Clickture dataset(5 and 26 respectively) [44]. We set the margin parameter γ in the loss in Eq. (5) to 1 and the tradeoff parameter δ for the Huber loss to 1.5 as in [17]. The LSTM consists of a hidden layer with 512 units. We train the parameters of the LSTM and projection layer M using stochastic gradient descent with adaptive weight updates (AdaGrad) [10]. We add an l 2 penalty on the weights, with a λ of 10−3 . We train for 20 epochs using minibatches of 128 triplets. 6.1 Tested components We discuss three important components of our model next. Objective. We compare our proposed training objective to that of Liu et al. [34]. Their model is trained to rank a positive query higher than a negative query given a fixed frame. In contrast, our method is trained to rank a positive frame higher than a negative frame given a fixed query. Cost function. We also investigate the importance of modeling frame quality. In particular, we compare different cost functions. (i) We enforce two ranking constraints: one for the quality term and one for the embedding similarity, as in Eq.(4) (Q expli ), (ii) We sum the quality and similarity term into one output score, for which we enforce the rank constraint, as in Eq.(3) (Q impli ) or (iii) we don’t model quality at all. HIT @ 1 Method Queries Liu et al. [34] Titles QAR without Q expli QAR (Ours) Method VG VG or G Spear. ρ mAP 40.625 73.83 0.122 0.629 36.71 38.86 72.63 74.76 0.367 0.376 0.6685 0.6712 Table 2: Comparison of thumbnail selection performance against the state of the art, on the QTS evaluation dataset. Note that [34] uses queries for their method which are not publicly available (see text). Precision-Recall curve- QTS Evaluation Dataset Liu et al QAR w/o Qexpli QAR (Ours) 0.75 No textual input Random Video2GIF [17] Ours: Frame quality Q expli Titles Liu et al. [34] +LSTM QAR without Q expli QAR (Ours) Queries Liu et al. [34] +LSTM QAR without Q expli QAR (Ours) HIT@1 Spear. ρ mAP 66.6 ± 3.5 67.0 69.0 0.0 0.167 0.135 0.674 0.708 0.749 70.0 70.0 71.0 0.134 0.182 0.221 0.731 0.743 0.760 72.0 76.0 72.0 0.204 0.268 0.264 0.730 0.752 0.769 Table 3: Performance of our relevance models on the RAD dataset in comparison with previous methods. 0.78 Precision-Recall curve - RAD Dataset Liu et al+LSTM Video2GIF QAR w/o Qexpli QAR (Ours) 0.76 0.65 0.74 Precision Precision 0.70 0.60 0.72 0.70 0.1 0.2 0.3 0.4 0.5 0.6 Recall 0.7 0.8 0.9 1.0 Figure 3: Precision-Recall curve of QTS Evaluation dataset for different methods. Text representation. As mentioned in Sec. 3.2, we represent the words of the query using word vectors. To combine the individual word representations into single vector, we investigate two approaches: (i) averaging the word embedding vectors and (ii) using an LSTM model that learns to combine the individual word embeddings. 6.2 Results We show the results of our detailed experiments in Tab. 1. They give insights on several important points. Text representation. Modeling queries with an LSTM, rather than averaging the individual word representations, improves performance significantly. This is not surprising, as this model can learn to ignore words that are not visually informative (e.g. 2014). Objective and Cost function. The analysis shows that training with our objective leads to better performance compared to using the objective of Liu et al. [34]. This can be explained with the properties of videos, which typically contain many frames that are low-quality or not visually informative [55]. Thus, formulating the thumbnail task in a way that the model can learn about these quality aspects is beneficial. Using the appropriate triplets for training boosts performance substantially (correlation with the 0.68 0.66 0.0 0.2 0.4 0.6 0.8 1.0 Recall Figure 4: Recall-Precision curve of the RAD testet for different methods. Our method (Blue) performs high in terms of mAP. loss of Liu et al. [34] + LSTM: 0.270, Ours: Huber + LSTM 0.367). When including a quality term in the model, performance improves further, where an explicit loss performs slightly better (Ours: Huber + LSTM + Q expli in Tab. 1). Somewhat surprisingly, modeling quality alone already outperforms Liu et al. [34] in terms of mAP, despite not using any textual information. Quality adds a significant boost to performance in the video domain. Interestingly, this is different in the image domain, due to the difference in quality statistics. Images returned by a search engine are mostly of good quality, thus explicitly accounting for it does not improve performance (see supplementary material). To conclude, we see that the better alignment of the objective to the keyframe retrieval task, the addition of an LSTM and modeling quality of the thumbnails improves performance. Together, they provide an substantial improvement compared to Liu et al. ’s model. Our method achieves an absolute improvement of 6.01% in HIT@1, 4.04% in mAP, and an improvement in correlation from 0.186 to 0.376. These gains are even more significant when we consider the possible ranges of these metrics. e.g. for Spearman correlation, Video2GIF [17] Ours Anaconda vs lion Ariana Grande dangerous Liu et al. [34] Diversity − − X − − Quality − X − − X X (33%) MMR [4] X (66%) − Repr X − − − − − Truck tug of war Hecate [55] X (45%) Ours X (43%) X (2%) Baby alive Upper bound Figure 5: Qualitative Results of top ranked keyframes on RAD. a) Liu et al. b) Video2GIF c) Our model (from left). Video titles are shown on the left. Ground truth relevance labels are shown in Blue. P=Positive, N=Negative. human agreement is at 0.73 on the RAD dataset (c.f. Sec. 5.1), thus providing an upper bound. Similarly, HIT@1 and mAP have small effective ranges given their high scores for a random model. 7 Method Similarity − − − X X X (10%) < PR > < CR > < F1 > 0.654 0.671 0.575 0.763 0.775 0.817 0.542 0.808 0.550 0.563 0.672 0.522 0.629 0.578 0.594 0.692 0.825 0.716 0.708 0.787 0.713 0.704 0.825 0.721 0.938 0.925 0.928 Table 4: Performance of summarization methods on the RAD dataset. Repr means Representativeness. X and − depict whether an objective was used or not. MMR and ours learn their corresponding weights. Percentage in parentheses the normalized learnt weights. Upper bound refers to the best possible performance, obtained using the ground truth annotations of RAD. Figure 3 shows the precision-recall curve for the experiment. As can be seen QAR outperforms [34] for all recall ratios. To better understand the effects of using titles or queries, we quantify the value of the two on the RAD dataset below. EXPERIMENTS We evaluate our model (QAR) and compare it to Liu et al. [34] and Video2GIF [17]. Our dataset (RAD) We also evaluate our model on the RAD test set (Tab. 3). QAR (ours) significantly outperforms the previous s-o-a of [17, 34], even when augmenting Liu et al. [34] with an LSTM. QAR improves mAP by 2.9% when using Titles and 3.9% when using Queries over our implementation of Liu et al. [34]+LSTM. We also see that modeling quality leads to significant gains in terms of mAP when using Titles or Queries (+1.7% in both cases). HIT@1 for query relevance, however, is lower when including quality. We believe that the reason for this is that when the query is given, the textual-visual similarity is a more reliable signal to determine the single best keyframe. While including quality improves the overall ranking on mAP, it is solely based on appearance and thus seems to inhibit the fine-grained ranking results at low recall(Fig. 4). However, when only the title is used, the frame quality becomes a stronger predictor for thumbnail selection and improves performance on all metrics. We present some qualitative results of different methods for relevance prediction in Fig. 5. Query-dependent Thumbnail Selection Dataset (QTS) [34] We compare against the s-o-a on the QTS evaluation dataset in Tab. 2. We report the performance of Liu et al. [34] from their paper. Note, however, that the results are not directly comparable, as they use query-video pairs for predicting relevance, while only the titles are shared publicly. Thus, we use the titles instead, which is an important difference. Relevance is annotated with respect to the queries, which often differ from the video titles. We compare the re-implementation of [34] using titles in detail in Tab. 1. Encouragingly, our model performs well even when just using the titles and outperforms them on most metrics. It improves mAP by 4.22% over [34] and correlation by a margin of 0.254 (c.f. Table 2). As mentioned in Sec. 4, we use four objectives for our summarization model. Referring to Tab. 4, we use QAR model to get Similarity and Quality scores while Diversity and Representativeness scores are obtained as described in Sec. 4. We compare the performance of our full model with each individual objective, a baseline based on Maximal Marginal Relevance (MMR) [4] and Hecate [55]. MMR greedily builds a set that maximises the weighted sum of two terms: (i) The similarity of the selected elements to a query and (ii) The dissimilarity to previously selected elements. To estimate the similarity to the query we use our own model (QAR without Q expli ) and for dissimilarity the diversity as defined in Sec. 4. Finally, we In the previous section, we have determined that our objective, embedding queries with an LSTM and explicitly modelling quality performs best. We call this model QAR (Quality-Aware Relevance) in the following and compare against state-of-the-art(s-o-a) models on the QTS and RAD datasets. We also evaluate the full summarization model on RAD. For these experiments, we split RAD into 100 videos for training, 50 for validation and 50 for testing. Evaluation metrics. For relevance we use the same metrics as in Sec. 6. To evaluate video summaries on RAD, we additionally use F1 scores. The F1 score is the harmonic mean of precision of relevance prediction and cluster recall [21]. It is high, if a method selects relevant frames from diverse clusters. 7.1 Evaluating the Relevance Model 7.2 Evaluating the Summarization Model Ours Similarity MMR [4] Hecate [55] Query: Hairstyles for Men 100 131 4 12 1 37 97 129 42 70 26 27 28 93 23 37 129 97 42 112 Relevance Scores Time Similarity MMR Hecate Ours Positive Negative 0 20 40 60 80 100 120 Clusters Figure 6: We show video summaries created by Hecate [55], MMR [4], our similarity model and our full summarization approach. The Green number on the images depicts the frame number. We plot the ground truth relevance scores, marking the selected frames for the shown methods, and cluster annotations over the video in the bottom two rows. For cluster annotation, each color represents a unique cluster. Additional examples are provided in supplementary. compare it to Hecate, recently introduced in [55]. Hecate estimates frame quality using the stillness of the frame and selects representative and diverse thumbnails by clustering the video with k-means and selecting the highest quality frame from the k largest clusters. Results Quantitative results are shown in Tab. 4, while Fig. 6 shows qualitative results. As can be seen, combining all objectives with our model works best. It outperforms all single objectives, as well as the MMR [4] baseline, even though MMR also uses our wellperforming similarity estimation. Similarity alone has the highest precision, but tends to pick frames that are visually similar (c.f. Fig. 6), thus resulting in low cluster recall. Diversification objectives (diversity and representativeness) have a high cluster recall, but the frames are less relevant. Somewhat surprisingly, Hecate [55] is a relatively strong baseline. In particular, it performs well in terms of relevance, despite using a simple quality score. This further highlights the importance of quality for the thumbnail selection task. It also indicates that the used VGG-19 architecture might be suboptimal for predicting quality. CNNs for classification use small input resolutions, thus making it difficult to predict quality aspects such as blur. Finding better architectures for that task is actively researched, e.g. [35, 37], and might be used to improve our method. When analysing the learned weights (c.f. Tab. 4) we find that the similarity prediction is the most important objective, which matches our expectations. Quality gets a lower, but non-zero weight, thus showing that it provides information that is complementary to query-similarity. Thus, it helps predicting the relevance of a frame. The reader should however be aware that differences in the variance of the objectives can affect the weights learned. Thus, they should be taken with a grain of salt and only be considered tendencies. 8 CONCLUSION We introduced a new method for query-adaptive video summarization. At its core lies a textual-visual embedding, which lets us select frames relevant to a query. In contrast to earlier works, such as [52, 67], this model allows us to handle unconstrained queries and even full sentences. We proposed and empirically evaluated different improvements over [34], for learning a relevance model. Our empirical evaluation showed that a better training objective, a more sophisticated text model, and explicitly modelling quality leads to significant performance gains. In particular, we showed that quality plays an important role in the absence of high-quality relevance information, such as queries, i.e. when only the title can be used. Finally, we introduced a new dataset for thumbnail selection which comes with query-relevance labels and a grouping of the frames according to visual and semantic similarity. On this data, we tested our full summarization framework and showed that it compares favourably to strong baselines such as MMR [4] and [55]. We hope that our new dataset will spur further research in query adaptive video summarization. 9 ACKNOWLEDGEMENTS This work has been supported by Toyota via the project TRACEZurich. We also acknowledge the support by the CHIST-ERA project MUSTER. MG was supported by the European Research Council under the project VarCity (#273940). REFERENCES [1] I Arev, HS Park, and Yaser Sheikh. 2014. Automatic editing of footage from multiple social cameras. ACM Transactions on Graphics (TOG) (2014). [2] Lamberto Ballan, Marco Bertini, Giuseppe Serra, and Alberto Del Bimbo. 2015. A data-driven approach for tag refinement and localization in web videos. Computer Vision and Image Understanding (2015). [3] Andrei Barbu, Alexander Bridge, Zachary Burchill, Dan Coroian, Sven Dickinson, Sanja Fidler, Aaron Michaux, Sam Mussman, Siddharth Narayanaswamy, Dhaval Salvi, Lara Schmidt, Jiangnan Shangguan, Jeffrey Mark Siskind, Jarrell Waggoner, Song Wang, Jinlian Wei, Yifan Yin, and Zhiqi Zhang. 2012. Video In Sentences Out. UAI (2012). arXiv:arXiv:1204.2742v1 [4] Jaime Carbonell and Jade Goldstein. 1998. The use of MMR, diversity-based reranking for reordering documents and producing summaries. In ACM SIGIR. [5] Xinlei Chen and C Lawrence Zitnick. 2014. Learning a Recurrent Visual Representation for Image Caption Generation. Proceedings of CoRR (2014). [6] Pradipto Das, Chenliang Xu, Richard F. Doell, and Jason J. Corso. 2013. A thousand frames in just a few words: Lingual description of videos through latent topics and sparse object stitching. CVPR (2013). [7] Sandra E. F. de Avila, Ana P. B. Lopes, A. da Luz, and A. de Albuquerque Araújo. 2011. VSUMM: A mechanism designed to produce static video summaries and a novel evaluation method. Pattern Recognition Letters 32, 1 (2011). [8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In CVPR. [9] J. Donahue, Lisa A. Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, T. Darrell, and K. Saenko. 2015. Long-term recurrent convolutional networks for visual recognition and description. CVPR (2015). [10] John Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive Subgradient Methods for Online Learning and Stochastic Optimization. Journal of Machine Learning Research 12 (2011). [11] Hao Fang, Saurabh Gupta, Forrest Iandola, Rupesh K Srivastava, Li Deng, Piotr Dollár, Jianfeng Gao, Xiaodong He, Margaret Mitchell, John C Platt, et al. 2015. From captions to visual concepts and back. In CVPR. [12] Ana L. N. Fred and Anil K. Jain. 2003. Robust Data Clustering. CVPR (2003). [13] Andrea Frome, Gs Corrado, and Jonathon Shlens. 2013. Devise: A deep visualsemantic embedding model. NIPS (2013). arXiv:arXiv:1312.5650v3 [14] Boqing Gong, Wei-Lun Chao, Kristen Grauman, and Fei Sha. 2014. Diverse sequential subset selection for supervised video summarization. In NIPS. [15] Michael Gygli, Helmut Grabner, and Luc Van Gool. 2015. Video Summarization by Learning Submodular Mixtures of Objectives. In CVPR. [16] M Gygli, H Grabner, H Riemenschneider, F Nater, and L Van Gool. 2013. The Interestingness of Images. In ICCV. [17] Michael Gygli, Yale Song, and Liangliang Cao. 2016. Video2GIF: Automatic Generation of Animated GIFs from Video. CVPR (2016). [18] Amirhossein Habibian, Thomas Mensink, and Cees GM Snoek. 2016. VideoStory embeddings recognize events when examples are scarce. TPAMI (2016). [19] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation (1997). [20] XS Hua, L Yang, M Ye, K Wang, Y Rui, and J Li. 2013. Clickture: A large-scale real-world image dataset. Technical Report. MSR-TR-2013-75. [21] Bogdan Ionescu, Adrian Popescu, Mihai Lupu, Alexandru Lucian Ginsca, and Henning Müller. 2015. Retrieving Diverse Social Images at MediaEval 2015: Challenge, Dataset and Evaluation. MediaEval (2015). [22] Phillip Isola, Devi Parikh, Antonio Torralba, and Aude Oliva. 2011. Understanding the intrinsic memorability of images. In NIPS. [23] M. Jain, Jan C van Gemert, T. Mensink, and C. GM Snoek. 2015. Objects2action: Classifying and localizing actions without any video example. In ICCV. [24] Andrej Karpathy, Armand Joulin, and Fei Fei Li. 2014. Deep fragment embeddings for bidirectional image sentence mapping. NIPS (2014). [25] Andrej Karpathy and Fei Fei Li. 2015. Deep visual-semantic alignments for generating image descriptions. CVPR (2015). [26] Aditya Khosla, Raffay Hamid, CJ Lin, and Neel Sundaresan. 2013. Large-Scale Video Summarization Using Web-Image Priors. CVPR (2013). [27] Gunhee Kim, Leonid Sigal, and Eric P Xing. 2014. Joint summarization of largescale collections of web images and videos for storyline reconstruction. In CVPR. [28] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. 2014. Unifying visualsemantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539 (2014). [29] Johannes Kopf, Michael F Cohen, and Richard Szeliski. 2014. First-person hyperlapse videos. ACM Transactions on Graphics (2014). [30] A. Krause and D. Golovin. 2012. Submodular function maximization. (2012). [31] Yong Jae Lee, Joydeep Ghosh, and Kristen Grauman. 2012. Discovering important people and objects for egocentric video summarization. CVPR (2012). [32] Hui Lin and JA Bilmes. 2012. Learning mixtures of submodular shells with application to document summarization. arXiv preprint arXiv:1210.4871 (2012). [33] Feng Liu, Yuzhen Niu, and Michael Gleicher. 2009. Using Web Photos for Measuring Video Frame Interestingness. IJCAI (2009). [34] Wu Liu, Tao Mei, Yongdong Zhang, C Che, and Jiebo Luo. 2015. Multi-task deep visual-semantic embedding for video thumbnail selection. CVPR (2015). [35] Xin Lu, Zhe Lin, Xiaohui Shen, Radomir Mech, and James Z Wang. 2015. Deep multi-patch aggregation network for image style, aesthetics, and quality estimation. In CVPR. [36] Zheng Lu and Kristen Grauman. 2013. Story-driven summarization for egocentric video. CVPR (2013). [37] Long Mai, Hailin Jin, and Feng Liu. 2016. Composition-preserving deep photo aesthetics assessment. In CVPR. [38] Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L Yuille. 2014. Explain images with multimodal recurrent neural networks. arXiv preprint arXiv:1410.1090 (2014). [39] Masoud Mazloom, Xirong Li, and Cees Snoek. 2016. TagBook: A Semantic Video Representation without Supervision for Event Detection. IEEE Transactions on Multimedia (2016). [40] Marina Meilă. 2003. Comparing clusterings by the variation of information. In Learning theory and kernel machines. Springer. [41] T. Mikolov, K. Chen, G. Corrado, and J. Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013). [42] T Mikolov and J Dean. 2013. Distributed representations of words and phrases and their compositionality. NIPS (2013). [43] M Minoux. 1978. Accelerated greedy algorithms for maximizing submodular set functions. Optimization Techniques (1978). [44] Jonas Mueller and Aditya Thyagarajan. 2016. Siamese Recurrent Architectures for Learning Sentence Similarity. In AAAI. [45] Mukund Narasimhan and Jeff A Bilmes. 2012. A submodular-supermodular procedure with applications to discriminative structure learning. arXiv preprint arXiv:1207.1404 (2012). [46] GL Nemhauser, LA Wolsey, and ML Fisher. 1978. An analysis of approximations for maximizing submodular set functions - I. Mathematical Programming (1978). [47] M. Norouzi, T. Mikolov, S. Bengio, Y. Singer, J. Shlens, A. Frome, G. S. Corrado, and J. Dean. 2013. Zero-shot learning by convex combination of semantic embeddings. arXiv preprint arXiv:1312.5650 (2013). [48] Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global Vectors for Word Representation.. In EMNLP. [49] Danila Potapov, Matthijs Douze, Zaid Harchaoui, and Cordelia Schmid. 2014. Category-specific video summarization. In ECCV. Springer. [50] Yael Pritch, Alex Rav-Acha, and Shmuel Peleg. 2008. Nonchronological video synopsis and indexing. TPAMI 30, 11 (2008). [51] GJ Qi, XS Hua, Y Rui, J Tang, T Mei, and HJ Zhang. 2007. Correlative Multi-Label Video Annotation. (2007). [52] Aidean Sharghi, Boqing Gong, and Mubarak Shah. 2016. Query-Focused Extractive Video Summarization. CoRR abs/1607.05177 (2016). [53] Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014). [54] Richard Socher, Andrej Karpathy, Quoc V Le, Christopher D Manning, and Andrew Y Ng. 2014. Grounded compositional semantics for finding and describing images with sentences. ACL (2014). [55] Y. Song, M. Redi, J. Vallmitjana, and A. Jaimes. 2016. To Click or Not To Click: Automatic Selection of Beautiful Thumbnails from Videos. In CIKM. ACM. [56] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. 2015. TVSUM: Summarizing web videos using titles. In CVPR. [57] Min Sun, Ali Farhadi, and Steve Seitz. 2014. Ranking Domain-Specific Highlights by Analyzing Edited Videos. ECCV (2014). [58] Min Sun, Ali Farhadi, Ben Taskar, and Steve Seitz. 2014. Salient Montages from Unconstrained Videos. ECCV (2014). [59] Min Sun, Kuo-Hao Zeng, Yenchen Lin, and Farhadi Ali. 2017. Semantic Highlight Retrieval and Term Prediction. IEEE Transactions on Image Processing (2017). [60] Ba Tu Truong and Svetha Venkatesh. 2007. Video abstraction. ACM Transactions on Multimedia Computing, Communications, and Applications (2007). [61] Silke Wagner and Dorothea Wagner. 2007. Comparing Clusterings - An Overview. Graph-Theoretic Concepts in Computer Science (2007). [62] Yufei Wang, Zhe Lin, Xiaohui Shen, Radomir Mech, Gavin Miller, and Garrison W Cottrell. 2016. Event-Specific Image Importance. In CVPR. [63] Wayne Wolf. 1996. Key frame selection by motion analysis. Acoustics, Speech, and Signal Processing. ICASSP-96 (1996). [64] Bo Xiong and Kristen Grauman. 2014. Detecting snap points in egocentric video with a web photo prior. In ECCV. [65] Ting Yao, Tao Mei, and Yong Rui. 2016. Highlight detection with pairwise deep ranking for first-person video summarization. In CVPR. [66] Gloria Zen, Paloma de Juan, Yale Song, and Alejandro Jaimes. 2016. Mouse activity as an indicator of interestingness in video. In ICMR. [67] Kuo-Hao Zeng, Yen-Chen Lin, Ali Farhadi, and Min Sun. 2016. Semantic highlight retrieval. In ICIP. [68] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. 2016. Video Summarization with Long Short-term Memory. ECCV (2016). [69] Bin Zhao and Eric P Xing. 2014. Quasi real-time summarization for consumer videos. In CVPR.
1cs.CV
Recognizing hyperelliptic graphs in polynomial time arXiv:1706.05670v1 [cs.DS] 18 Jun 2017 Jelco M. Bodewes* Gunther Cornelissen† Hans L. Bodlaender* Marieke van der Wegen† Abstract Recently, a new set of multigraph parameters was defined, called “gonalities”. Gonality bears some similarity to treewidth, and is a relevant graph parameter for problems in number theory and multigraph algorithms. Multigraphs of gonality 1 are trees. We consider so-called “hyperelliptic graphs” (multigraphs of gonality 2) and provide a safe and complete sets of reduction rules for such multigraphs, showing that for three of the flavors of gonality, we can recognize hyperelliptic graphs in O(n log n + m) time, where n is the number of vertices and m the number of edges of the multigraph. 1 Introduction How complex is a graph G?1 If we start from the premise that the simplest connected graphs are trees, we can assign a complexity to G by stating how much it ‘deviates’, in a quantitative sense, from being a tree. Given that trees have, by definition, no cycles, one might use the first Betti number b1 (G), the number of independent cycles in G, as such a measure. From the algorithmic point of view, however, this is not such a good parameter, since many hard computational problems do not occur in families of graphs in which b1 (G) is bounded. However, another measure of “proximity to a tree”, namely, the treewidth tw(G) of G, is very good for this purpose, witness for example Courcelle’s theorem [8] that in bounded treewith, simple graph properties definable in the monadic second order logic admit linear time verification algorithms. Recently, based on the analogy between graphs and algebraic curves, a new set of graph parameters were defined, all slightly different, but all called “gonality” ([2], [3], [7]). We will consider four versions of gonality, and give them distinct names to avoid confusion. Whereas treewidth is based on subtrees, geometric gonality gon(G) is based on quotient trees. More precisely, the geometric gonality of a graph is the minimal degree of a harmonic morphism onto a tree (cf. infra). Just like treewidth can be defined using a pursuit-evasion game on a graph, a notion of divisorial gonality dgon(G) can be defined using chip firing (or “divisors”; cf. [4]). There also exist stable versions sgon(G) and sdgon(G), given as the minimum gonality over all possible refinements of the graph. We indicate two reasons why studying gonalities is relevant. The first one lies in algorithmic graph theory itself, where gonality is a new natural graph parameter, and it might play a role similar * Department Informatica, Universiteit Utrecht, Postbus 80.089, 3508 TB Utrecht, Nederland, [email protected], [email protected] † Mathematisch Instituut, Universiteit Utrecht, Postbus 80.010, 3508 TA Utrecht, Nederland, [email protected], [email protected] 1 To simplify terminology, we will use the word “graph” for a finite multigraph (i.e. allowing finitely many edges between the same vertices), and use the terminology “simple graph” if there is at most one edge between any pair of vertices. 1 sgon(G) = 2 dgon(G) = 3 sdgon(G) = 2 sgon(G) = 3 dgon(G) = 2 sdgon(G) = 2 Figure 1: The different notions of gonality are not the same. to treewidth for (multi-)graph problems, such as WEIGHTED SPARSTEST EDGE CUT, which is NPhard in bounded treewidth [15]. Secondly, stable gonality is important in number theory. Whereas in general, the existence of solutions to polynomial equations is algorithmically undecidable, if the polynomial equations under consideration define a curve of gonality gon(X), then the union of all solutions in all number fields of degree bounded above by 12 (gon(X) − 1) is finite (cf. [7, §11]). The gonality of a curve is related to the stable gonality of a certain graph related to the equation by reduction modulo prime numbers, and hence our results are relevant in the theory of algorithms for diophantine equations [7]. For more details, see Section 2. We list some known results. Gonality is different from treewidth; for example, treewidth depends only on the underlying simple graph, whereas gonality does not. At the same time, the Betti number, gonalities and treewidth are related by inequalities  1  2 (b1 (G) + 3) ≥ sgon(G); gon(G) ≥ sgon(G) ≥ sdgon(G) ≥ tw(G);  dgon(G) ≥ sdgon(G) (see [7], [11]). Stable gonality sgon(G) (a priori defined using universal quantifiers over three infinite sets) is decidable [14], and computing divisorial gonality dgon(G) is NP-hard [13]. It follows from [10, §5] that computing divisorial gonality is in XP. Treewidth and divisorial gonality are not “tied” in the sense of Norin [18, §2 & 2.6]: there exist G with tw(G) = 2 but arbitrarily high divisorial gonality [17]. Reduction rules are known for treewidth ≤ 3, and this class can be recognized in linear time [1]. A graph G is a tree if and only if either of the gonalities is 1. The next case in terms of this complexity measure is gonality 2, which, contrary to gonality 1, can be different in different flavors (see Figure 1); based on the analogous terminology for algebraic curves, we call graphs of divisorial, stable or stable divisorial gonality 2 respectively divisorial, stable, stable divisorial hyperelliptic graphs (compare [3]). In this paper, we give polynomial time algorithms that recognize dgon(G) ≤ 2, sgon(G) ≤ 2 or sdgon(G) ≤ 2 for a graph G. (We do not consider gonality, which is defined without finiteness conditions.) To obtain our algorithms, we provide safe and complete sets of reduction rules. Similar to recognition algorithms for graphs of treewidth 2 or 3 (see [1]), in our algorithms the rules are applied to the graph until no further rule application is possible; we decide positively if and only if this results in the empty graph. Some of the rules introduce constraints on pairs of vertices, which we model by colored edges. To deal with the fact that some of the rules are not local, we use a data structure that allows us to find an efficient way of applying these rules, leading to a running time of O(n log n + m). 2 .. .. . . .. . .. . (b) Intersection dual graph with finite harmonic morphism of degree p. (a) X modulo p. Figure 2: Picture illustrating Example 2.1 2 Number theoretical applications In this section, we briefly elucidate the relevance of gonality for number theoretical problems, mostly at an intuitive level. There is an analogy between complexity issues in graph algorithms and the difficulty of solving diophantine equations (polynomial equations with integer coefficients). In graph theory, a graph parameter can be very hard to compute, but certain hard problems become fixed parameter tractable with such a fixed parameter. Similarly, the negative solution to Hilbert’s Tenth Problem [9] implies that the existence of solutions to polynomial equations is algorithmically undecidable. However, if certain geometric parameters related to the polynomial equations are fixed, there are positive results about the solution sets of the equations. Geometric concepts such as dimension, Betti numbers and gonality make sense in the theory of diophantine equations. The solutions to a given set of polynomial equations describe a geometric object X; if the set of solutions in the complex numbers C defines a real surface X(C), we say the geometric object is an “algebraic curve” (two-dimensional of the real numbers, but onedimensional over the complex numbers). The first Betti number b1 (X) of X is the number of independent loops on X(C); and the gonality gon(X) of X is the minimal degree of a nonconstant holomorphic map from X(C) to the Riemann sphere (the complex projective line seen as a real surface). A curve of gonality two, i.e., with a degree two map to the Riemann sphere is called a hyperelliptic curve; it can be given by a polynomial equation y 2 = f (x) for some polynomial f , and the degree two map is (x, y) 7→ x. Equivalently, a curve has gonality two if and only if it admits an effective divisor of degree two and rank one. Typical positive results are as follows. A famous result of Faltings [12] in diophantine equations relates to the Betti number as parameter: if X is an algebraic curve with b1 (X) ≥ 2, then X has only finitely many solutions in a given number field K; so the solution set of the corresponding polynomial equations is finite. Another result uses gonality of the curve as parameter: so-called “uniform boundedness” says that the union of all solutions to the equations defining a curve X of gonality gon(X) in all number fields of degree bounded above by 12 (gon(X) − 1) is finite (cf. [7, §11]). There is a direct relation between curve gonality and graph gonality: the gonality of a curve is bounded below by the stable gonality (defined in Section 3 below) of the dual graph of the stable model of any reduction of the curve. Hence lower bounds and recognition of stable gonality provide an algorithmic deduction of a uniform boundedness result for curves with such reductions [7]. We give one example: Example 2.1. For a prime number p, consider the algebraic curve X : (xp − x)(y p − y) = p 3 in the (x, y)-plane. Reducing the curve modulo p, the equation becomes a union of lines x(x − 1) . . . (x − (p − 1)) · y(y − 1) . . . (y − (p − 1)) = 0 mod p (cf. Figure 2). The intersection dual graph is given by a vertex for each component of this reduction, where two vertices are connected by an edge if and only if the corresponding components intersect; in the example, it is the complete bipartite graph Kp,p . The stable gonality of Kp,p is p (since tw(Kp,p ) = p and there isS an obvious map of degree p from Kp,p to a tree). From [7, 4.5 & 11.1] one concludes that the set X(K) is finite, where K runs over all the (infinitely many for p ≥ 5) number fields of degree bounded above by (p − 1)/2. 3 Preliminaries D EFINITIONS Whenever we write “graph” we refer to a multigraph G = (V, E), where V is the set of vertices and E is a multiset of edges. A loop vv will be counted twice in the degree of vertex v. By Gv (u) we denote the induced subgraph of G on v and the component that contains u when we remove v. First we define divisorial gonality, we can consider it intuitively as a chip firing game: we have a graph and some initial configuration that assigns a non-negative number of “chips” to each vertex. We then can fire a vertex by moving a chip from it along each incident edge. If the number of chips on a vertex becomes negative, we consider it to be in debt. Divisorial gonality then asks: what is the minimum number of chips needed for an initial configuration, such that, for any vertex in the graph, there exists a sequence of firings that results in that vertex having at least one chip and no other vertex being in debt? We now give a formal definition for divisorial gonality, based on the concepts of [2] and notation from [10]: L Definition 3.1. Let G be a graph. A divisor D is an element of Z. We use D(v) to denote V (G) the integer assigned to vertex v. We call a divisor D effective if D(v) ≥ 0 for all v ∈ V (G). We denote the set of divisors on G by Div(G) and the set of effective divisors by Div+ (G). The degree deg(D) of a divisor is the sum over D(v) for all v ∈ V (G). By Divk (G) we denote all divisors with degree k. Definition 3.2. Let G be a graph with n vertices. The Laplacian matrix L is given by L = D − A, where D is the diagonal matrix with Dv,v = deg(v) and A is the adjacency matrix of G. We call a divisor P a principal divisor if there exists a divisor D such that P = LD and we denote the set of principal divisors by Prin(G). Definition 3.3. We call two divisors D and D ′ equivalent or, in notation D ∼ D ′ , if there exists a principal divisor P such that D ′ = D − P . Given a divisor D, we have a class of equivalent effective divisors |D| = {D ′ ∈ Div+ (G) | D ∼ D ′ }. Definition 3.4. The rank of a divisor D is denoted by r(D) and defined as follows:  max{k | |D − E| = 6 ∅ for all E ∈ Divk+ (G)} if |D| = 6 ∅; r(D) = −1 if |D| = ∅. Definition 3.5. The divisorial gonality, dgon, of a graph G is the lowest degree for which there exists a divisor of rank greater or equal to one, i.e., dgon(G) = min{deg(D) | D ∈ Div+ (G), r(D) ≥ 1}. 4 Definition 3.6. The stable divisorial gonality of G is sdgon(G) = min{dgon(G′ ) | G′ a refinement of G}. Note that every tree has divisorial gonality 1 and each graph with divisorial gonality 1 is a tree. Note also that for a disconnected graph, the divisorial gonality is equal to the sum of the divisorial gonality of the connected components. We define stable gonality as in [7, Definition 3.6]. Definition 3.7. Let G and H be graphs. A finite morphism is a map φ : G → H such that (i) φ(V (G)) ⊆ V (H), (ii) φ(uv) = φ(u)φ(v) for all uv ∈ E(G), together with, for every e ∈ E(G), an “index” rφ (e) ∈ N. Definition 3.8. We call a finite morphism φ : G → H harmonic if for every v ∈ V (G) it holds that for all e, e′ ∈ Eφ(v) (H) X X rφ (d) = d∈Ev (G),φ(d)=e rφ (d′ ). d′ ∈Ev (G),φ(d′ )=e′ We write mφ (v) for this sum. Definition 3.9. The degree of a finite harmonic morphism φ : G → H is X X rφ (e) = mφ (u), d∈E(G),φ(d)=e u∈V (G),φ(u)=v for e ∈ E(H), v ∈ V (H). This is independent of the choice of e or v ([2], Lemma 2.4). Definition 3.10. A graph G′ is a refinement of G if G′ can be obtained by applying the following operations finitely many times to G. (i) Add a leaf, i.e. a vertex of degree 1; (ii) subdivide an edge by adding a vertex. We call a vertex of G′ \G from which there are two disjoint paths to vertices of G, internal added vertices, we call the other vertices of G′ \G external added vertices. Definition 3.11. The stable gonality of a graph G is sgon(G) = min{deg(φ) | φ : G′ → T a finite harmonic morphism, G′ a refinement of G, T a tree}. For a tree G we can use the identity map to see that sgon(G) = 1. On the other hand, if G is not a tree, then any refinement of G contains a cycle. Such a cycle will be mapped to a path in a tree, and for every edge e in this path there will at least two edges that are mapped to e. Thus sgon(G) > 1 if G is not a tree. For a disconnected graph G its stable gonality is defined to be the sum of the stable gonality of its components. 5 Definition 3.12. We call graphs of stable, divisorial or stable divisorial gonality 2 respectively stable, divisorial, stable divisorial hyperelliptic graphs. In [3], hyperelliptic graphs G are defined by dgon(G) = 2, and it is proven that, for twoconnected graphs G, this is equivalent to the existence of a (not necessarily finite, i.e., possibly mapping edges to vertices) harmonic morphism of degree two from G to a tree. It is known that dgon(G) ≥ tw(G) for all G [11]. Since treewidth is closed under taking minors, we obtain sdgon(G) ≥ tw(G). As mentioned in A.6 of [7], it holds that sgon(G) ≥ sdgon(G) for all G. Thus we see that both divisorial gonality and stable gonality are larger than or equal to treewidth. Lemma 3.13. For any graph G, we have inequalities dgon(G) ≥ tw(G) and sgon(G) ≥ sdgon(G) ≥ tw(G). L EVEL SET DECOMPOSITION In this subsection we introduce a technique to be used in proofs on divisorial and stable divisorial gonality. Using it we can reason about transformations between equivalent divisors. We use the definition from [10]. We recall that two divisors D and D ′ are defined as equivalent if there exists a principal divisor P such that D ′ = D − P . A principal divisor then is a divisor that can be expressed as an image of the Laplacian matrix, in other words there is a divisor f , such that P = Lf . Intuitively this can be understood as follows: P is a vector over all vertices that for each vertex describes the change in number of chips, and f is the vector that for each vertex describes how often it should be fired to produce the result P . Therefore we have some divisor f that describes how often vertices should be fired to produce D ′ from D, D ′ = D − Lf . We now consider the idea of splitting this divisor f into layers, where each layer gives a subset of the graph that should be fired once. We shall see that with the right definition these layers or level sets have certain helpful properties. The level sets are given by the following definition: Definition 3.14. Let D and D ′ be two equivalent divisors, then there must exist a divisor f such that D ′ = D − Lf , where L is the Laplacian matrix. Let m = max{f (v) | v ∈ V (G)} and let k = m − min{f (v) : v ∈ V (G)}. We then define the level sets as follows: Ai = {v ∈ V (G) : f (v) ≥ m − i} for i ∈ {0, . . . , k}. Note that while there are multiple such f , these can only differ by some multiple of the all ones vector and therefore result in the same level set decomposition. We can then also consider the sequence of divisors that is created by firing each level set in the given order: Definition 3.15. A level set decomposition A0 , . . . , Ak belonging to a transformation of D into D ′ produces a sequence of divisors as follows: D0 = D, Di+1 = Di − LIAi for i ∈ {0, . . . , k − 1}. Here, IAi is the divisor with IAi (v) = 1 for v ∈ Ai and IAi (v) = 0 otherwise. 6 Note that for a level set decomposition A0 , . . . , Ak we have that A0 ⊆ A1 ⊆ · · · ⊆ Ak . Also note that the fact that Dk = D ′ follows from the first two parts of the definition. We continue with a useful property of level set decompositions, namely the fact that each divisor in the associated sequence is lower bounded by the pointwise minimum of D and D ′ : Theorem 3.16. Let D and D ′ be two equivalent divisors and A0 , . . . , Ak the level set decomposition of the transformation of D into D ′ . Let D0 , . . . , Dk be the associated sequence of divisors. We then have that: Di (v) ≥ min(D(v), D ′ (v)), ∀i ∈ {0, . . . , k}, ∀v ∈ V (G). Proof. Choose an i ∈ {0, . . . , k} and a vertex v ∈ V (G). If Di (v) ≥ D(v) we are done, so assume that Di (v) < D(v). Since Di (v) < D(v), v must have been fired at least once before Di (the only way a vertex can loose chips is by firing), so there is an Aj with v ∈ Aj and j < i. But since A0 ⊆ · · · ⊆ Ak , we then have that v ∈ Am for all m ≥ j and specifically for all m ≥ i. So v is fired in every subset starting from Ai , but the number of chips on v cannot increase if v is part of the fired subset, so we have that Di (v) ≥ Dm (v) for all m ≥ i and specifically, Di (v) ≥ Dk (v) = D ′ (v). Therefore either Di (v) ≥ D(v) or Di (v) ≥ D ′ (v) for all v ∈ V (G) and i ∈ {0, . . . , k}. We conclude that Di (v) ≥ min(D(v), D ′ (v)) for all v ∈ V (G) and i ∈ {0, . . . , k}. Theorem 3.16 gives a nice result for the transformation between two equivalent effective divisors: Definition 3.17. Given a graph G and an effective divisor D, we call a non-empty firing set A valid if D(a) ≥ outdegA (a) for all a ∈ A. Corollary 3.18. Let D and D ′ be two equivalent effective divisors. Let A0 , . . . , Ak be the level set decomposition of the transformation of D into D ′ and D0 , . . . , Dk the associated sequence of divisors. Then all divisors Di are effective and all firing subsets Ai are valid with respect to Di . Proof. This follows simply from Theorem 3.16: since Di (v) ≥ min(D(v), D ′ (v)), D(v) ≥ 0 and D ′ (v) ≥ 0, we conclude that Di (v) ≥ 0 and so Di is effective. To see that all firing sets are valid, note that a non-valid firing set would result in a non-effective divisor. R EDUCTION RULES We will be talking a lot about reduction rules in this text. By reduction rule we mean a rule that can be applied to a graph to produce a different graph. The following notation shows when a graph can be produced by the application of reduction rules starting from another graph: Definition 3.19. Let G and H be graphs and S be some set of reduction rules. We use GSH to denote that H can be produced by some application of a reduction rule from S. We use GS ∗ H to denote that H can be produced from G by some finite sequence of applications of reduction rules from S. If R is a single rule, we write GRH as shorthand for the application of the singleton set containing R. During the reduction of the graph we will need to keep track of certain restrictions otherwise lost by the removal of vertices and edges. We will maintain these restrictions in the form of a set of pairs on the vertices of the graph: 7 Definition 3.20. Given a graph G = (V, E) a set of constraints C is a set of pairs (v, w), where v, w ∈ V . This set can contain pairs of a vertex with itself, but can contain each pair only once. Though the different forms of gonality use the same concept of a set of constraints, the restrictions given by a constraint differ between them. What a constraint means for each type of gonality will be explained in their respective sections. Our final goal with each set of reduction rules is to show that they can be used to characterize the graphs in a certain class by reduction to the empty graph. For this we need to make sure that membership of the class is invariant under our reduction rules. Definition 3.21. Let R be a rule and S be a set of reduction rules. Let A be a class of graphs. We call R safe for A if from GRH it follows that H ∈ A ⇐⇒ G ∈ A. We call S safe for A if every rule in S is safe for A. Note that if S is safe for a class A then GS ∗ H implies that H ∈ A ⇐⇒ G ∈ A. Apart from our rule sets being safe, we also need to know that, if a graph is in our class, it is always possible to reduce it to the empty graph. Definition 3.22. Let S be a set of reduction rules, A a class of graphs. We call S complete for A if for any graph G ∈ A it holds GS ∗ ∅. For any rule set that is both complete and safe for A it then follows the rule set is suitable for characterization of A. Additionally it is not possible to make a wrong choice early on that would prevent the graph from being reduced to the empty set. Lemma 3.23. Let S be a set of rules that is safe and complete for A, with ∅ ∈ A, then we have the following for all graphs G, H: (i) GS ∗ ∅ if and only if G ∈ A; (ii) if G ∈ A and GS ∗ H, then HS ∗ ∅. Proof. For property i: Let G be a graph, such that GS ∗ ∅. Note that by the safeness of S and the fact that ∅ ∈ A it follows that G ∈ A. Assume on the other hand that G ∈ A, note that by the completeness of S it follows that GS ∗ ∅. For property ii: Let G be a graph in A and H a graph such that GS ∗ H. Note that by the safeness of S we have that H ∈ A, then by completeness of S it follows that HS ∗ ∅. 4 Reduction rules for divisorial gonality We will show that there exists a set of reduction rules that reduces a graph to the empty graph exactly when this graph has divisorial gonality two or lower. Or in other words, exactly when there exists a divisor on the graph with rank greater or equal to one and degree two. This set can therefore be used to test whether the divisorial gonality of a graph is equal to two, since the case of divisorial gonality equal to one can also be easily checked. We will first define an extra layer of structure on top of our normal graph in the form of constraints, after which we will list the required reduction rules. We will first prove that all our reduction rules are safe in the sense that they never change whether a graph has gonality greater than two. After this we proceed by showing that given a graph of divisorial gonality not greater than two, the reduction rules will always result in the empty graph in a finite number of steps. For this section we will assume that our graph is loopless and connected. Cases where this is not the case can be easily handled separately. Loops can simply be removed from the graph since they never impact the divisorial gonality. A disconnected graph has divisorial gonality two or lower exactly when it consists of two trees. 8 C ONSTRAINTS Checking whether a graph has gonality two or lower can be seen as checking whether there exists a divisor with degree two and rank greater or equal to one on our graph. Constraints in this case are used to restrict which divisors and transformations we consider after reduction. The pairs in the constraints place the following restrictions on what divisors and firing sets are allowed: Definition 4.1. Given a constraint r = (v, w) a divisor satisfies r if it is equivalent to an effective divisor after removing one chip from v and one chip from w. In addition any firing set used in transformations should either contain both v and w or neither. Note that in the case that v = w the first part means a divisor should be equivalent to an effective divisor after removing two chips from v and the second condition is fulfilled trivially. It will also be useful to define when constraints are non-conflicting on a cycle. For this we introduce the idea of compatibility on cycles: Definition 4.2. Let C be a cycle. Let CC ⊆ R be the set of constraints that contain a vertex in C. We call the constraints CC compatible if the following hold. (i) If (v, w) ∈ CC then both v ∈ C and w ∈ C. (ii) For each (v, w) ∈ CC and (v ′ , w′ ) ∈ CC , the divisor given by assigning a chip to v and w must be equivalent to the one given by assigning a chip to v ′ and w′ on the subgraph consisting of C. Note that a divisor of degree 2 can only satisfy all constraints on a cycle if they are compatible. Now we are interested in the existence of a divisor that has rank greater or equal to one, while satisfying all constraints. Definition 4.3. Given a graph G = (V, E) and its constraints C, we will call a divisor D suitable if it has degree 2 and r(D) ≥ 1 while also satisfying all constraints in C. Given a graph with constraints we will say that the graph has divisorial gonality 2 or lower if there exists a suitable divisor. Note that for a graph with no constraints this formulation is equivalent to the usual definition of divisorial gonality 2 or lower. We will denote the class of graphs with constraints that has divisorial gonality two or lower as G2d . T HE REDUCTION RULES We are given a graph G = (V, E) and a still empty set of constraints C. We note here that it is assumed we start with a connected graph and all the reduction rules maintain connectivity. The following rules are illustrated in Figure 3, where a constraint is represented by a red dashed edge. We start by covering the two possible end states of our reduction: Rule E1d . Given a graph consisting of exactly one vertex, remove that vertex. Rule E2d . Given a graph consisting of exactly two vertices, u and v, connected to each other by a single edge, and C = {(u, v)}, remove both vertices. Next are the reduction rules to get rid of vertices with degree equal to one. These rules are split by what constraint applies to the vertex: Rule T1d . Let v be a leaf, such that v has no constraints in C. Remove v. 9 Rule T2d . Let v be a leaf, such that its only constraint in C is (v, v). Let u be its neighbor. Remove v and add the constraint (u, u) if it does not exist yet. Rule T3d . Let v1 be a leaf, such that its only constraint in C is (v1 , v2 ), where v2 is another leaf, whose only constraint is also (v1 , v2 ). Let u1 be the neighbor of v1 and u2 be the neighbor of v2 (these can be the same vertex). Then remove v1 and v2 and add the constraint (u1 , u2 ) if it does not exist yet. Finally we have a set of reduction rules that apply to cycles containing at most 2 vertices with degree greater than two. The rules themselves are split by the number of vertices with degree greater than two. Rule C1d . Let C be a cycle of vertices with degree two. If the set of constraints CC on C is compatible, then replace C by a new single vertex. Rule C2d . Let C be a cycle with one vertex v with degree greater than two. If the set of constraints CC on C plus the constraint (v, v) is compatible, then remove all vertices except v in C and add the constraint (v, v) if it does not exist yet. Rule C3d . Let C be a cycle with two vertices v and u of degree greater than two. If there exists a path from v to u that does not share any edges with C and the set of constraints CC on C plus the constraint (v, u) is compatible, then remove all vertices of C except v and u, remove all edges in C and add the constraint (v, u) if it does not exist yet. We shall use Rd for the set consisting of all the above reduction rules: E1d , E2d , T1d , T2d , T3d , C2d and C3d . We will now state the main theorem stating that this set of reduction rules has the desired properties. After this we will build up the proof. C1d , Theorem 4.4. The set of rules Rd is safe and complete for G2d . S AFENESS In this section it is assumed there is a graph G and another graph H that follows from G by applying a rule. Now we first make an observation on the connectivity of our graphs: Lemma 4.5. Let G and H be graphs. If G is connected and GRd∗ H then H is connected. Proof. We observe that the only rule that removes a path between two remaining vertices is C3d . In the case of C3d however we demand that there is a path between v and w outside of C so this path will still exist and it follows that H is still connected. Since we assume our graph G is connected it follows that each produced graph H is also connected. Now we will show for each of the rules in Rd that it is safe. Lemma 4.6. Rules E1d and E2d are safe. Proof. For both rules it should be clear that their starting states as well as the empty graph have divisorial gonality two or lower. From this it follows they both are safe. Lemma 4.7. Rules T1d and T2d are safe. 10 Rule E1d Rule E2d Rule T1d Rule T2d Rule T3d Rule C1d Rule C2d C C Rule C3d C Figure 3: The reduction rules for divisorial gonality Proof. Let v be our vertex with degree 1 and u its neighbor. We know that the only constraint on v can be the constraint (v, v). Note that if H ∈ G2d then there is a divisor on H that puts at least one chip on u. Moving this divisor to G, note that we can move chips to v by firing G − {v}, it follows that this divisor is also suitable for G. Given that G ∈ G2d note that we can find a suitable divisor that has no chips on v by firing v until it contains no chips. This divisor will also be suitable on H. For T2d the proof is analogous, except with two chips on v. Lemma 4.8. Rule T3d is safe. Proof. Let v1 and v2 be the vertices with degree one, such that their only constraint is (v1 , v2 ) and let u1 and u2 be their neighbors. We first assume that H ∈ G2d , then there is a suitable divisor on H with one chip on u1 and another chip on u2 . Note that we can move this divisor to G. Then by firing V (G) − {v1 , v2 } we can move a chip to v1 and v2 . Therefore this divisor is also suitable on G. Assume then that G ∈ G2d , then there is a suitable divisor on G with one chip on v1 and v2 . By firing {v1 , v2 } we can create a divisor with a chip on u1 and u2 (or two on u1 if u1 = u2 ). Note that this divisor is suitable on H. Lemma 4.9. Rule C1d is safe. 11 Proof. We start by assuming that H ∈ G2d . Note that by Lemma 4.5 we have that H is connected. Therefore it follows that H must consist of a single vertex, therefore G consists of a single cycle and it follows that G ∈ G2d , since all constraints are compatible. Assume then that G ∈ G2d instead. Since G is connected it must consist exactly of the cycle C, thus H consists of a single point and H ∈ G2d . Lemma 4.10. Rule C2d is safe. Proof. Let C be our cycle with one vertex v with degree greater than 2. Assume that H ∈ G2d ; then there is a suitable divisor on H with two chips on v. Move this divisor to G. Note that if we fire V (G) − C + {v} then we move the two chips onto the two neighbors of v in C. Since all constraints on C are compatible with the constraint (v, v) it follows that we can move the chips along C while satisfying the constraints on C. From this it follows that our divisor is suitable on G. Assume now that G ∈ G2d . Since all constraints on C are compatible with (v, v), it follows that we can find a suitable divisor with two chips on v. Moving this divisor to H gives a suitable divisor there. Thus, H ∈ G2d . Lemma 4.11. Rule C3d is safe. Proof. Let C be our cycle and v, w the two vertices with degree greater than two in C. We first assume that H ∈ G2d . From this it follows that there exists a suitable divisor on H with a chip on v and a chip on w. Note that in G all constraints on C plus (v, w) are compatible. From this we see that if we move the divisor from H to G it will be able to satisfy all constraints on C. It is also clear that from v and w we can move chips along either of the arcs that form C together with v and w. Therefore the divisor is also suitable on G and thus G ∈ G2d . Let us then assume that instead G ∈ G2d . Clearly there exists a suitable divisor D on G that has a chip on v. We will show that there is a suitable divisor that has a chip on both v and w: Assume that D(w) = 0, then there should be a suitable divisor D ′ with D ′ (w) = 1 and D ∼ D ′ . This implies there is a level set decomposition of the transformation from D to D ′ . If none of the subsets contain v then it follows that D ′ (v) = 1 and we are done. Otherwise let Ai be the first subset that contains v and Di the divisor before firing Ai . Note that we should have Di (a) ≥ outdegAi (a) for all P a ∈ Ai , since all firing sets should be valid by Corollary 3.18. Since deg(Di ) = 2 it follows that a∈Ai outdegAi (a) ≤ 2. This is the same as the cut induced by Ai having size two or lower. Note that the minimum cut between v and w is at least three, since they are both part of C and there exists an additional path outside of C between them. Therefore it follows that Ai can only induce a cut of size two or lower if w ∈ Ai . But this implies that Di (w) ≥ 1, since a vertex can not receive a chip after entering the firing set. We conclude that Di (v) = 1 and Di (w) = 1. Also by the fact that the minimum cut between v and w is at least three it follows that a subset firing can only be valid if the subset contains either both v and w or neither. Since, by Corollary 3.18, any transformation between effective divisors can be done by a series of valid subset firings it follows that any transformation can be done while adhering to the constraint (v, w). Therefore the divisor Di gives us a suitable divisor on H. We conclude that H ∈ G2d . Since we have shown that each of the rules in Rd is safe, we conclude: Theorem 4.12. The ruleset Rd is safe for G2d . 12 C OMPLETENESS Theorem 4.12 implies that membership in G2d is invariant under the reduction rules in Rd . For the reduction rules to be useful, we will also need to confirm that any graph can be reduced to the empty graph by a finite sequence of rule applications. Lemma 4.13. Let G be a graph and v ∈ V (G) a vertex. If there are two different constraints on v, so (v, w), (v, w′ ) ∈ R(G), with w 6= w′ , then G ∈ / G2d . Proof. We first check the possibility where v = w′ . Then any suitable divisor must be equivalent to the divisor D with D(v) = 2, but also equivalent to the divisor D ′ with D ′ (v) = 1 and D ′ (w) = 1. But this means these divisors are equivalent to each other. Note however that since we have the constraint (v, w) any firing set containing v must also contain w. Starting with divisor D however, note that any valid firing set must contain v (it is the only vertex with chips), which means it must also contain w. This implies no level set decomposition from D to D ′ can exist, from which it follows that there is no transformation of D into D ′ , so G ∈ / G2d . The other possibility is that v 6= w and v 6= w′ . This means any suitable divisor should be equivalent to the divisor D with D(v) = 1, D(w) = 1, and equivalent to the divisor D ′ with D ′ (v) = 1, D ′ (w′ ) = 1. Note that any firing set that contains v also contains both w and w′ by our constraints. Moreover any firing set containing w contains w′ by our constraints. Since starting in D any valid firing set must contain either v or w (they are the only vertices with chips), it follows that any valid firing set must contain w′ . Again this implies no level set decomposition from D to D ′ exists, so D and D ′ cannot be equivalent. We conclude no suitable divisor can exist and therefore G ∈ / G2d . Lemma 4.14. Let G ∈ G2d be a graph where none of the rules E1d , E2d , T1d , T2d or T3d can be applied. Then G contains no vertices of degree 1. Proof. Assume on the contrary that G does contain a vertex v with degree 1. By Lemma 4.13 and the fact that G ∈ G2d we have that at most one constraint contains v. If there is no constraint on v, we could apply Rule T1d to it, therefore there is exactly one constraint on v. If this constraint is (v, v) we would be able to apply Rule T2d to v. If the constraint is (v, w), where w is another vertex of degree 1, Rule T3d could be applied to v. The only remaining possibility is that the constraint on v is the constraint (v, w) where w is a vertex with degree greater than 1. We will use D to denote the divisor with D(v) = D(w) = 1. Since we have the constraint (v, w) and G ∈ G2d , D is a suitable divisor. We first consider the case where w is not a cut-vertex. Let u be the neighbor of v. Consider the transformation from D to a divisor D ′ with D ′ (u) = 1. Let A0 be the first firing set in the level decomposition of this transformation. Note that we have v, w ∈ A0 and u ∈ / A0 . Since w is not a cut-vertex, it follows for each neighbor wi of w that there is a path from wi to u that does not contain w or wi = u. Note that if a neighbor wi 6= u is in A0 , then somewhere on its path to u must be an edge that crosses between A0 and its complement Ac0 . But such a crossing edge would imply that the firing set is not valid, since no vertex on this path contains a chip. Since w has degree at least two, and none of its neighbors are in A0 , it follows that the firing set is not valid, since w would lose at least two chips. Since no valid firing set exists to start the transformation, it follows that no transformation from D to D ′ exists, but this implies that r(D) < 1. Since D should be suitable by constraint (v, w), we have a contradiction. We proceed with the case where w is a cut-vertex. Let Cx be a connected component not containing v after removing w. Consider the subset Cx in G. Note that from D we can never 13 obtain an equivalent divisor with two chips on Cx . Since the chip from v would have to move through w to get to Cx , this would require D to be equivalent to a divisor with two chips on w, which is impossible by Lemma 4.13 if G ∈ G2d . Since D has rank greater than zero it then follows that Cx must be a tree. This means Cx must contain a vertex x of degree one, we know however that since we cannot apply rules T1d , T2d or T3d to G, x must have a constraint (x, y) where y is a vertex with degree greater than one. We now consider the possible locations of y. If y ∈ Cx , then D must be equivalent to a divisor with a chip on x and a chip on y. As mentioned before, D cannot be equivalent to a divisor with two chips on Cx , so it follows that y∈ / Cx . Since y ∈ / Cx , D has to be equivalent to the divisor D ′′ with D ′′ (x) = D ′′ (y) = 1. Let Cy be the component containing y. Let A0 be the first subset of the level set decomposition of the transformation of D into D ′′ . Note that v, w ∈ A0 and x, y ∈ / A0 . But this implies that w has at least one neighbor w1 in Cy , with w1 ∈ / A0 , namely the first vertex on the path from w to y. But w also has at least one neighbor w2 in Cx , with w2 ∈ / Cx , namely the first vertex on the path from w to x. This means w has two neighbors that it will send a chip to, but w only has one chip. By / Cy , giving a contradiction. Corollary 3.18, no transformation from D to D ′′ can exist and thus y ∈ We conclude there can be no such constraint (x, y) and from this we conclude that no vertices with degree 1 can exist in G. Lemma 4.15. Let G be a graph with a set of constraints C and let C be a cycle in G with CC the set of constraints that contain a vertex in C. If G ∈ G2d then the constraints CC are compatible. Proof. We start by showing that the first property of a compatible constraint set holds. Let (v, w) ∈ CC be a constraint and let v ∈ C without loss of generality. We show that w ∈ C. Assume on the contrary that w ∈ / C, then let D be the suitable divisor with D(v) = D(w) = 1. Let x be a vertex in C with x 6= v. Let D ′ be any divisor with D ′ (x) ≥ 1. Let A0 be the first firing set of the level set decomposition of the transformation of D into D ′ . Note that v, w ∈ A0 and x ∈ / A0 . Note there are two disjoint paths from v to x, since they are on the same cycle. This implies a chip will be sent along both these paths by A0 , but since w ∈ / C, both these chips must come from v. However, v only has one chip, so we conclude that no transformation can exist and thus D is not suitable, a contradiction. For the second property, let (v, w), (v ′ , w′ ) ∈ CC be two constraints on C. By our first property we have that v, w, v ′ , w′ ∈ C. Let D be the divisor with D(v) = D(w) = 1 and D ′ the divisor with D ′ (v ′ ) = D ′ (w′ ) = 1. Let A0 be the first firing set of the level set decomposition of the transformation of D into D ′ . Note that v, w ∈ A0 and v ′ , w′ ∈ / A0 . We observe that v and w split C into two arcs. Note that both v ′ and w′ must be on the same arc: if they are not on the same arc, there exists disjoint paths from v to v ′ and to w′ that do not contain w. This implies that A0 sends two chips along these paths, but v has only one chip. Now note that C is biconnected, which implies that for a firing set A with w ∈ A and w′ ∈ /A to be valid there must be at least two chips on vertices in C. This follows since there are at least two edges crossing between A and its complement Ac in C. We know D and D ′ must be equivalent, since G ∈ G2d and both correspond to constraints on G. Let A0 , . . . , Ak be the level set decomposition of the transformation of D into D ′ . Since each of these firing sets is valid by Lemma 3.18, it follows this transformation leaves two chips on C at each intermediate divisor. It follows that if we restrain these firing sets to C, we have a sequence of firing sets that transforms D into D ′ on C. Therefore D and D ′ are equivalent on C, so our second property is also fulfilled. Lemma 4.16. Let G be a simple graph of treewidth 2 or lower and containing at least 4 vertices, then G has at least two vertices with degree 2 or lower. 14 Proof. Let T be a tree decomposition of G, such that we can remove no vertex of T while still keeping a valid tree decomposition. Note that T must contain at least two leaves t1 and t2 , and hence t1 must contain at least one vertex v from G that is sent to no vertex of T , since otherwise we would be able to remove t1 . All neighbors of v must be sent to t1 , but t1 contains at most three elements of G, so v has at most two neighbors. Applying the same argument to t2 we find another vertex with at most two neighbors. Lemma 4.17. Given a non-empty graph G ∈ G2d there is a rule in Rd that can be applied to G. Proof. Let G ∈ G2d be such a graph and assume that no rule in Rd can be applied to G. By Lemma 4.14 we have that such a graph can contain no vertices of degree one. Therefore, our graph consists of vertices of degree 2 and of a set T of vertices of degree 3 or greater. Consider the minor H of G created by contracting each path of degree 2 vertices to an edge. Note that any edge in H therefore represents a path consisting of a single edge or a path with any number of vertices with degree 2 in G. Now assume we have a loop in H. Note that this loop corresponds to a path of degree 2 vertices in G going from a degree 3 or greater vertex to itself, so this path plus the vertex forms a cycle with one vertex of degree 3 or greater. By Lemma 4.15 the constraints on this cycle are compatible, so we are able to apply Rule C2d to it. Since we assumed no rules can be applied, it follows H contains no loops. Now we attempt to find a subgraph H ′ of H with no multiple edges. If H contains no multiple edges, simply let H ′ = H. Otherwise let v and w be vertices such that there are at least two edges between v and w. If v and w are still connected to each other after removing two edges e1 , e2 between them, note that these edges correspond to two disjoint paths of degree 2 vertices in G. Thus v, w plus the paths corresponding to e1 and e2 form a cycle C in G with exactly two vertices of degree 3 or greater, where v and w have a path that does not share any edges with C. Again by Lemma 4.15 we have that the constraints on this cycle are compatible and so we are able to apply Rule C2d to C. From this it follows that v and w must be disconnected after removing e1 and e2 . So any multiple edge in H is a double edge, whose removal splits the graph in two connected components. Let H ′ be the connected component of minimal size over all removals of a double edge in H. Note that H ′ cannot contain any double edge, since this would imply a smaller connected component. We now have a minor H ′ of G, where each vertex has degree at least 3 with at most one exception, which has no loops or multiple edges and therefore is a simple graph. Since H ′ only has at most one vertex with degree lower than three, by Lemma 4.16 it follows that tw(H ′ ) ≥ 3 and since treewidth is closed under taking minors we get tw(G) ≥ 3. But then by Lemma 3.13 it follows that dgon(G) ≥ 3, creating a contradiction, since G ∈ G2d . We conclude our assumption must be wrong and there is a rule in Rd that can be applied to G. Now we have everything required to prove our main theorem: Proof of Theorem 4.4. By Theorem 4.12 we have that Rd is safe. It remains to prove that Rd is also complete. Assume that G ∈ G2d . By Lemma 4.17 and Theorem 4.12 we have that we can keep applying rules from Rd to G as long as G has not been turned into the empty graph yet. Now observe that each rule removes at least one vertex or in the case of C3d at least two edges, while never adding more vertices or edges. Since G starts with a finite number of vertices and edges it follows that rules from Rd can be only applied a finite number of times to the graph. When no more rules 15 can be applied to the graph, it follows the graph has been reduced to the empty graph. Therefore GRd∗ ∅ and it follows that Rd is complete. Now by Lemma 3.23 it follows Rd has the properties we want it to have so that we are able to use it for characterization of the graphs with divisorial gonality two or lower. 5 Reduction rules for stable gonality In this section, we give a complete set of safe reduction rules to recognize stable hyperelliptic graphs, i.e. graphs with stable gonality 2. We will first introduce some extra notation and then we will state all rules. Next we will show that all rules are safe for graphs with stable gonality at most 2 and that those graphs can be reduced to the empty graph. It is not hard to see that the set of rules implies a polynomial time algorithm to test if a graph has stable gonality at most 2; in Section 7, we discuss how we can obtain an algorithm with a running time of O(m + n log n). N OTATION For a given graph G, we want to know whether there exists a finite harmonic morphism of degree 2 from a refinement of G to a tree. We will do this by reducing G to the empty graph. During this process we sometimes add constraints to our graph. The set of constraints gives restrictions to which morphisms we allow. Definition 5.1. Let G be a graph, G′ a refinement of G, T a tree. Let φ : G′ → T be a map. We call φ a suitable morphism if it is a finite harmonic morphism of degree 2 and it satisfies the following conditions. (i) For all pairs (v, v) ∈ C it holds that mφ (v) = 2. (ii) For all pairs (u, v) ∈ C with u 6= v it holds that φ(u) = φ(v) and mφ (u) = mφ (v) = 1. We say that a graph with constraints has stable gonality at most 2 if there exists a suitable morphism from a refinement of G to a tree. Let G2s be the class of graphs with constraints that have stable gonality at most 2. We define the empty graph to have stable gonality 0 and thus ∅ ∈ G2s . We will denote the set of constraints that contain a vertex v by Cv . Sometimes it is convenient to think of constraints as an extra set of edges. Now we can prove some lemmas about constraints. Lemma 5.2. Let G be a graph with constraints. If there is a vertex v with |Cv | > 1, then sgon(G) ≥ 3. Proof. Let G be a graph with sgon(G) = 2. Suppose that |Cv | > 1. Let {u, v} and {v, w} be two constraints that contain v. We know that u 6= w. Suppose that φ is a suitable morphism of degree 2. We distinguish two cases. Suppose that u = v. Then we know that mφ (v) = 2. On the other hand we have that mφ (v) = mφ (w) = 1. This yields a contradiction. Now suppose that u 6= v and w 6= v. Notice that φ(u) = φ(v) = φ(w), thus there are at least three vertices mapped to φ(v). We conclude that deg(φ) ≥ 3. This yields a contradiction. We conclude that |Cv | ≤ 1. Lemma 5.3. Let G be a graph, and φ : G → T a finite harmonic morphism of degree 2. If φ(u) = φ(v), then deg(u) = deg(v). 16 Proof. Notice that mφ (u) = mφ (v) = 1. Let e be an edge incident to φ(u). We see that there is exactly one edge e′ such that e′ is incident to u and φ(e′ ) = e. On the other hand every edge that is incident to u is mapped to an edge that is incident to φ(u). So we conclude that degG (u) = degT (φ(u)). Analogously we find that degG (v) = degT (φ(v)). Since φ(u) = φ(v), it follows that deg(u) = deg(v). Lemma 5.4. Let G be a graph where every leaf is incident to a constraint, so if deg(u) = 1 then Cu 6= ∅ for all u. Suppose that (u, v) ∈ C. If deg(u) 6= deg(v), then sgon(G) ≥ 3. Proof. Suppose that deg(u) 6= deg(v). Assume without loss of generality that deg(u) > deg(v). Suppose that sgon(G) = 2. Let G′ be a refinement of G with a minimal number of vertices such that there exists a suitable morphism of degree 2. Let φ : G′ → T be such a morphism. We know that φ(u) = φ(v), thus degG′ (u) = degG′ (v). So there is a neighbor x of v which is an external added vertex. Now we look at φ(x). Notice that there is a neighbor y of u such that φ(x) = φ(y). It is clear that y 6= x, since x is an external added vertex. Thus mφ (x) = mφ (y) = 1. Let x′ be a neighbor of x, not equal to v. Suppose that mφ (x′ ) = 2. We know that the edge e = {x, x′ } has index 1, so there there exists another neighbor of x′ that is mapped to φ(x). We know that y is the unique vertex other than x that is mapped to φ(x), it follows that y is a neighbor of x′ . This yields a contradiction, since x′ is an external added vertex. We conclude that mφ (x′ ) = 1. Inductively we find that mφ (x′′ ) = 1 for all vertices in x′′ ∈ Gv (x). Let x′ 6= v be a leaf in Gv (x); then mφ (x′ ) = 1. Let y ′ be the vertex such that φ(x′ ) = φ(y ′ ). Now it follows that deg(x′ ) = deg(y ′ ), thus y ′ is a leaf. Since x′ is an added vertex, it also follows that Cy = 0. Since every leaf G was incident to a constraint, we conclude that y ′ is added to G. It follows that G′ \{y ′ , x′ } is a refinement of G and that φ′ : G′ \{y ′ , x′ } → T \{φ(y ′ )} is a suitable morphism of degree 2. This yields a contradiction with the minimality of G′ . We conclude that sgon(G) ≥ 3. R EDUCTION RULES We will now state all rules. Figure 4 shows all rules in pictures, constraints are showed as green dashed edges. From now on we will refer to the constraints as green edges. We apply those rules to a given graph G with an empty set of constraints. When a rule adds a constraint uv, and there already exists such a constraint, then the set of constraints does not change. Rule T1s . Let v be a leaf with Cv = ∅. Let u be the neighbor of v. Contract the edge uv. Rule T2s . Let v be a leaf with Cv = {(v, v)}. Let u be the neighbor of v. Contract the edge uv. Rule S1s . Let v be a vertex of degree 2 with Cv = ∅. Let u1 , u2 be the neighbors of v (possibly u1 = u2 ). Contract the edge u1 v. Rule T3s . Let G be a graph where every leaf and every degree 2 vertex is incident to a green edge. Let v1 and v2 be two leaves that are connected by a green edge. Let u1 and u2 be their neighbors (possibly u1 = u2 ). Contract the edges u1 v1 and u2 v2 . Rule S2s . Let G be a graph where every leaf and every degree 2 vertex is incident to a green edge. Let v be a vertex of degree 2 with a green loop, such that there exists a path from v to v in G (possibly containing green edges). Let u1 and u2 be the neighbors of v (possibly u1 = u2 ). Remove v and connect u1 and u2 with a green edge. 17 Rule T1s Rule T2s Rule S1s Rule T3s Rule S2s Rule Ls Rule P1s Rule P2s Rule E1s Rule E2s Rule E3s Figure 4: The reduction rules for recognizing stable hyperelliptic graphs. Rule Ls . Let v be a vertex with a loop. Remove all loops from v and add a green loop to v. Rule P1s . Let uv be an edge such that there also exists a green edge uv. Remove the black edge uv. Rule P2s . Let u, v be vertices, such that |E(u, v)| > 1. Let e and f be two of those edges. If there exists another path, possibly containing green edges, from u to v, then remove e and f and add a green edge from u to v. Rule E1s . Let G be the graph consisting of a single vertex v with Cv = ∅. Remove v. Rule E2s . Let G be the graph consisting of a single vertex v with a green loop. Remove v. Rule E3s . Let G be the graph consisting of a two vertices u and v that are connected by a green edge. Remove u and v. We will write Rs for this set of reduction rules. We can now state the main theorem; in the next sections we will prove this theorem. 18 v v1 v2 u x1 x w y Figure 5: Proof of Lemma 5.7. Theorem 5.5. The set of rules Rs is safe and complete for G2s . S AFENESS Now we will prove that the rules Rs are safe for G2s , i.e., if G a is graph, and H is obtained from G by applying one of the rules, then sgon(G) ≤ 2 if and only if sgon(H) ≤ 2. In all proofs we assume that the original graph is called G and the graph obtained by applying a rule is called H. Lemma 5.6. Rule T1s is safe. Proof. Let v be the leaf in G to which the rule is applied. Suppose that sgon(G) ≤ 2. Since G is a refinement of H, it is clear that sgon(H) ≤ 2. Suppose that sgon(H) ≤ 2. Then there exists a refinement H ′ of H and a suitable morphism φ : H ′ → T . Write u for the neighbor of v in G. We distinguish two cases. Suppose that mφ (u) = 2. Then add a leaf v to u in H ′ to obtain G′ . Now we see that G′ is a refinement of G. Give the edge uv index rφ′ (uv) = 2, and give all other edges e index rφ′ (e) = rφ (e). Add a leaf v ′ to φ(u) in T to obtain T ′ . Then we can extend φ to φ′ : G′ → T ′ , ( φ(x) if x ∈ H ′ ′ φ (x) = v′ if x = v. It is clear that φ′ is a suitable morphism, so we conclude that sgon(G) ≤ 2. Suppose that mφ (u) = 1. Let w be the other vertex such that φ(w) = φ(u). Then add leaves v1 and v2 to u and w in H ′ to obtain G′ . We see that G′ is a refinement of G. Give the edges uv1 and wv2 indices rφ′ (uv1 ) = rφ′ (wv2 ) = 1, and give all other edges e index rφ′ (e) = rφ (e). Add a leaf v ′ to φ(u) in T to obtain T ′ . Then we can extend φ to φ′ : G′ → T ′ , ( φ(x) if x ∈ H ′ ′ φ (x) = v′ if x = v1 , v2 . It is clear that φ′ is a suitable morphism, so we conclude that sgon(G) ≤ 2. Lemma 5.7. Rule T2s is safe. Proof. Let v be the vertex in G to which the rule is applied. Suppose that sgon(G) ≤ 2. Then there exists a refinement G′ of G and a suitable morphism φ : G′ → T . Let u be the neighbor of v in G. We distinguish two cases: Suppose that mφ (u) = 2. Define H ′ as the graph G′ with a green loop at vertex u and without the green loop at v, then H ′ is a refinement of H. Now we see that φ : H ′ → T is a suitable morphism, so sgon(H) ≤ 2. Suppose that mφ (u) = 1. Let v0 = v, v1 , . . . , vk = u be the vertices that are added to the edge uv of G. Let i be the largest integer such that mφ (vi ) = 2. Notice that i < k. Then there 19 exists another vertex x1 in G′ such that φ(vi+1 ) = φ(x1 ). If vi+1 6= u, it follows that there is an edge x1 x2 that is mapped to φ(vi+1 vi+2 ). And since mφ (vi+2 ) = 1, we see that x2 6= vi+2 . It follows that there exists x1 6= vi+1 , . . ., xk−i 6= vk such that φ(vi+j ) = φ(xj ). Write x = xk−i , then φ(x) = φ(u) and mφ (u) = mφ (x) = 1. See Figure 5 for an illustration of this. Notice that x is an external added vertex. Let w be a neighbor of u not equal to vk−1 . Then we see that there exists an vertex y such that φ(uw) = φ(xy). Since x is an external added vertex, we see that w 6= y. We conclude that mφ (w) = 1. Inductively we see that for every vertex w′ in Gvi (vi+1 )\{vi } it holds that mφ (w′ ) = 1. Define H ′ as Gvi (vi+1 )\{vi }, with a green loop at vertex u. Notice that H ′ is a refinement of H. Now we can restrict φ to H ′ and give every edge index rφ′ (e) = 2 to obtain a suitable morphism: φ′ : H ′ → T ′ , where T ′ = φ(Gvi (vi+1 )\{vi }). We conclude that sgon(H) ≤ 2. Suppose that sgon(H) ≤ 2. Then there exists a refinement H ′ of H and a suitable morphism φ : H ′ → T . Write u for the neighbor of v in G. We know that mφ (u) = 2. Then add a leaf with a green loop to u in H ′ to obtain G′ . Now we see that G′ is a refinement of G. Give the edge uv index rφ′ (uv) = 2, and give all other edges e index rφ′ (e) = rφ (e). Add a leaf v ′ to φ(u) in T to obtain T ′ . Then we can extend φ to φ′ : G′ → T ′ , ( φ(x) if x ∈ H ′ φ′ (x) = v′ if x = v. It is clear that φ′ is a suitable morphism, so we conclude that sgon(G) ≤ 2. Lemma 5.8. Rule S1s is safe. Proof. Let v be the vertex in G to which the rule is applied. Suppose that sgon(G) ≤ 2. Let G′ be a refinement of G such that there exists a suitable morphism φ : G′ → T . Since G′ is a refinement of H too, it is clear that sgon(H) ≤ 2. Now suppose that sgon(H) ≤ 2. Let H ′ be a refinement of H such that there exists a suitable morphism φ : H ′ → T . Write u1 and u2 for the neighbors of v in G. We distinguish two cases. Suppose that u1 = u2 . It follows that the edge u1 u2 is subdivided in H ′ , thus H ′ is a refinement of G′ too. We conclude that sgon(G) ≤ 2. Suppose that u1 6= u2 . If the edge u1 u2 is subdivided, we see again that H ′ is a refinement of G, and sgon(G) ≤ 2. So suppose that u1 and u2 are neighbors in H ′ . If r(u1 u2 ) = 2, then add a vertex v on the edge u1 u2 to obtain a graph G′ . Notice that G′ is a refinement of G. Give the edges u1 v and vu2 index 2, and give all other edges e index rφ′ (e) = rφ (e). And add a vertex v ′ on the edge φ(u1 )φ(u2 ) in T to obtain T ′ . Now we see that φ′ : G′ → T ′ given by ( φ(x) if x ∈ H ′ , ′ φ (x) = v′ if x = v, is a suitable morphism. We conclude that sgon(G) ≤ 2. If r(u1 u2 ) = 1, then there exists another edge w1 w2 such that φ(u1 u2 ) = φ(w1 w2 ). Now add a vertex v1 on the edge u1 u2 and a vertex v2 on the edge w1 w2 to obtain a graph G′ . Notice that G′ is a refinement of G. Give the edges u1 v1 , v1 u2 , w1 v2 and v2 w2 index 1, and give all other edges e index rφ′ (e) = rφ (e). Add a vertex v ′ on the edge φ(u1 )φ(u2 ) in T to obtain T ′ . Now we see that φ′ : G′ → T ′ given by ( φ(x) if x ∈ H ′ , φ′ (x) = v′ if x = v1 , v2 , 20 a1 v1 a2 u1 v1 a1 a2 u1 x v2 b1 b2 u2 b3 v2 b1 b2 c u2 b3 w (a) Case 1. (b) Case 2. Figure 6: Proof of Lemma 5.9. is a suitable morphism. We conclude that sgon(G) ≤ 2. Lemma 5.9. Rule T3s is safe. Proof. Let v1 and v2 be the vertices in G to which the rule is applied. “=⇒”: Suppose that sgon(G) ≤ 2. Let G′ be a minimum refinement of G such that there exists a suitable morphism φ : G′ → T , i.e. for every refinement G′′ with less vertices than G′ there is no suitable morphism φ′ : G′′ → T ′ for any tree T ′ . Let u1 and u2 be the neighbors of v1 and v2 in G. We distinguish three cases. Case 1: Suppose that u1 6= u2 , and that there does not exist a path from v1 to v2 , except the green edge v1 v2 . Let a0 = v1 , a1 , . . . , ak = u1 be the subdivision of the edge u1 v1 and b0 = v2 , b1 , . . . , bl = u2 the subdivision of the edge u2 v2 . We know that there exists an edge v2 c such that φ(a0 a1 ) = φ(v2 c). It is clear that c 6= a1 , thus mφ (a1 ) = 1. (See Figure 6a.) Inductively we find that for every vertex a′ in G′a0 (a1 ) it holds that mφ (a′ ) = 1. We conclude that G′a0 (a1 ) is a tree. Analogously we find that G′b0 (b1 ) is a tree. Thus Gv1 (u1 ) and Gv2 (u2 ) are trees. Thus H consists of two black trees connected by a green edge. Now we can construct a refinement H ′ of H, a tree T ′ and a suitable morphism φ′ : H ′ → T ′ . Copy every branch of u1 and add them to u2 and copy every branch of u2 and add them to u1 . Write H ′ for this graph. Now we see that the two trees of H ′ are the same, say T ′ . Now we can define φ′ : H ′ → T ′ as the identity map on each of the components, where φ′ (u1 ) = φ′ (u2 ). Thus φ′ is a suitable morphism. We conclude that sgon(H) ≤ 2. Case 2: Suppose that u1 6= u2 and that there exists a path (possibly containing green edges) from v1 to v2 . Assume that φ(u1 ) 6= φ(u2 ). Let a0 = v1 , a1 , . . . , vk = u1 be the added vertices on the edge v1 u1 and let b0 = v2 , b1 , . . . , bl = u2 be the added vertices on the edge v2 u2 . Assume without loss of generality that k ≤ l. It is clear that all vertices a0 , . . . , ak , b0 , . . . , bl lie on the path from v1 to v2 . If φ(a1 ) 6= φ(b1 ), then there is a vertex x in the path from a1 to b1 that is mapped to φ(v1 ). This yields a contradiction. Thus φ(a1 ) = φ(b1 ). Inductively we find that φ(ai ) = φ(bi ) for all i ≤ k. We conclude that φ(bk ) = φ(u1 ). It follows that deg(bk ) = deg(u1 ). We again distinguish two cases. Suppose that deg(u1 ) > 2. Then bk has an external added neighbor w. We see that u1 has a neighbor x such that φ(bk w) = φ(u1 x). Since w is an external added vertex, it follows that w 6= x. Thus mφ (w) = 1. (See Figure 6b.) Iteratively we see that for every vertex w′ in G′bk (w), it holds that mφ (w′ ) = 1. Notice that G′bk (w) is a tree, since w is an external added vertex. Now let y be a leaf in G′bk (w), and let y ′ be such that φ(y) = φ(y ′ ). Then y ′ is a leaf. It is clear that y ′ has no green edge incident to it, thus y ′ is an added vertex. We conclude that we can remove y and y ′ from G′ and φ(y) from T and still have a suitable morphism. This yield a contradiction with the minimality of G′ . Suppose that deg(u1 ) = 2 in G′ . Then the degree of u1 in G is also 2. It follows that Cu1 6= ∅. Let u1 c be a green edge. If c = u1 , so if u1 has a green loop, then mφ (u1 ) = 2. This yields a 21 contradiction. It is clear that c 6= bl , since bl is an added vertex. It follows that there are 3 distinct vertices that are mapped to φ(u1 ). This yields a contradiction. Altogether we conclude that φ(u1 ) = φ(u2 ). Define H ′ as G′ with a green edge u1 u2 . Now ′ H is a refinement of H, and φ : H ′ → T is a suitable morphism. We conclude that sgon(H) ≤ 2. Case 3: Suppose that u1 = u2 . Analogous to the second case, we prove that mφ (u1 ) = 2. Define H ′ as G′ with a green loop at vertex u1 . Now H ′ is a refinement of H, and φ : H ′ → T is a suitable morphism. We conclude that sgon H ≤ 2. “⇐=”: Suppose that sgon(H) ≤ 2. Then there exists a refinement H ′ of H and a suitable morphism φ : H ′ → T . Write u1 and u2 for the neighbors of v1 and v2 in G. We know that φ(u1 ) = φ(u2 ). Then add a leaves v1 and v2 to u1 and u2 and a green edge v1 v2 in H ′ to obtain G′ . Now we see that G′ is a refinement of G. Give the edges u1 v1 and u2 v2 index rφ′ (u1 v1 ) = rφ′ (u2 v2 ) = 1, and give all other edges e index rφ′ (e) = rφ (e). Add a leaf v ′ to φ(u1 ) in T to obtain T ′ . Then we can extend φ to φ′ : G′ → T ′ , ( φ(x) if x ∈ H ′ ′ φ (x) = v′ if x = v1 , v2 . It is clear that φ′ is a suitable morphism, so we conclude that sgon(G) ≤ 2. Lemma 5.10. Rule S2s is safe. Proof. This proof is analogous to the proof of the second and third case in the proof of Lemma 5.9, so we omit it. Lemma 5.11. Rule Ls is safe. Proof. Let v be the vertex in G to which the rule is applied. Suppose that sgon(G) ≤ 2. Then there exists a refinement G′ of G and a suitable morphism φ : G′ → T . Let u be a vertex that is added to the loop vv. We distinguish two cases. Suppose that mφ (v) = 2. Define H ′ as the graph G′ \G′v (u) with a green loop at vertex v, then ′ H is a refinement of H. Let T ′ = T \φ(G′v (u)) we see that the restricted morphism φ : H ′ → T ′ is a suitable morphism, so sgon(H) ≤ 2. Suppose that mφ (u) = 1. Let v0 = v, v1 , . . . , vk = v be the vertices that are added to the loop vv of G. Let i be the integer such that φ(vi ) = φ(v). Let w be a neighbor of v not equal to v1 or vk−1 . Then there is a neighbor x of vi , not equal to vi−1 and vi+1 , such that φ(w) = φ(x). Notice that x is an external added vertex, thus mφ (w) = mφ (x) = 1. Inductively we see that for every vertex w′ in G′v (w) it holds that mφ (w′ ) = 1. We conclude that G′v (w) is a tree. Define H ′ as G′v (w), with a green loop at vertex v. Notice that H ′ is a refinement of H. Now we can restrict φ to H ′ and give every edge index rφ′ (e) = 2 to obtain a suitable morphism: φ′ : H ′ → T ′ , where T ′ = φ(G′v (w)). We conclude that sgon(H) ≤ 2. Suppose that sgon(H) ≤ 2. Then there exists a refinement H ′ of H and a suitable morphism φ : H ′ → T . We know that mφ (v) = 2. Then add a vertex u to H ′ with two black edges to v to obtain G′ . Now we see that G′ is a refinement of G. Give both edges uv index rφ′ (uv) = 1, and give all other edges e index rφ′ (e) = rφ (e). Add a leaf v ′ to φ(u) in T to obtain T ′ . Then we can extend φ to φ′ : G′ → T ′ , ( φ(x) if x ∈ H ′ φ′ (x) = v′ if x = u. It is clear that φ′ is a suitable morphism, so we conclude that sgon(G) ≤ 2. 22 Lemma 5.12. Rule P1s is safe. Proof. Let uv be the edge in G to which the rule is applied. Suppose that sgon(G) ≤ 2. Let G′ be a refinement of G and φ : G′ → T a suitable morphism. Let Guv be all internal and external added vertices to the edge uv. Now define H ′ = G′ \Guv and T ′ = T \(φ(Guv )). Write φ′ for the restriction of φ to H ′ . Notice that φ′ is a suitable morphism and that H ′ is a refinement of H. Thus sgon(H) ≤ 2. Suppose that sgon(H) ≤ 2. Let H ′ be a refinement of H and φ : H ′ → T a suitable morphism. Add an edge uv and a vertex w on this edge to H ′ , to obtain a refinement G′ of G. Add a vertex w′ to T with an edge to φ(u), to obtain tree T ′ . Give the edges uw and vw index rφ′ (uw) = rφ′ (vw) = 1, and give all other edges e index rφ′ (e) = rφ (e). Look at φ′ : G′ → T ′ , defined as ( φ(x) if x ∈ H ′ ′ φ (x) = w′ if x = w. Notice that φ(uw) = φ(vw), since we there is a green edge uv. We conclude that φ′ is a suitable morphism, thus sgon(G) ≤ 2. Lemma 5.13. Rule P2s is safe. Proof. Let u and v be the vertices in G to which the rule is applied. Suppose that sgon(G) ≤ 2. Let G′ be a refinement of G and φ : G′ → T a suitable morphism. If φ(u) 6= φ(v), then there are at least three paths that are mapped to the path from φ(u) to φ(v) in T . This yields a contradiction. Thus φ(u) = φ(v). Now we see, analogous to the proof of Lemma 5.12, that sgon(H) ≤ 2. Suppose that sgon(H) ≤ 2. Then we find analogous to the proof of Lemma 5.12, that sgon(G) ≤ 2. Lemma 5.14. Rules E1s , E2s and E3s are safe. Proof. All these graphs have stable gonality at most 2, so the statement holds true. Now we have proven that all rules are safe, thus we have the following lemma: Lemma 5.15. The set of rules Rs is safe for G2s . C OMPLETENESS Now we will prove that the set of rules is complete. Let G be a connected graph and apply our algorithm to G, let H be the final graph to which we cannot apply any rule. Notice that H is connected in the sense of black and green edges. We define the graphs H1 , H2 and H3 as a single vertex, a vertex with a green loop and two vertices connected by a green edge respectively. These are exactly the graph that can be reduced to the empty graph by Rules E1s , E2s and E3s . Lemma 5.16. The set of rules Rs is complete for G2s . Proof. Let G be a graph with sgon(G) ≤ 2. Let H be the final graph after reducing G, so we cannot apply a rule to H anymore. By Lemma 5.15 it follows that sgon(H) ≤ 2. Suppose that H 6= ∅. We first say something about the structure of H. If there is a double edge between two vertices u and v, then removing these two edges yields a disconnected graph, 23 otherwise we could apply Rule P2s. Let u1 v1 , . . . , uk vk be all double edges in H. Let Hi,1 , Hi,2 be connected components after removing the edges ui vi . If there is a degree 2 vertex with a green loop, then removing this vertex yields a disconnected graph, otherwise we could apply rule ′ , H ′ be the connected S2s . Let v1 , . . . , vl be all degree two vertices with a green loop. Let Hi,1 i,2 components after removing vi . Let H ′ be the element of ′ {Hi,j | 1 ≤ i ≤ k, j ∈ {1, 2}} ∪ {Hi,j | 1 ≤ i ≤ l, j ∈ {1, 2}} with the minimum number of vertices. Notice that there is at most one vertex v in H ′ with degH ′ (v) 6= degH (v). Now we can say the following about H ′ . • If H ′ contains only one vertex, then we could have applied Rule T1s , T2s , S1s , S2s , E1s or E2s . Thus H ′ contains at least 2 vertices. • If there is a vertex that is incident to more than one green edge, then sgon(H) ≥ 3 by Lemma 5.2. So we can assume that no vertex is incident to more than one green edge. • If H ′ contains a vertex u 6= v of degree 0, then degH (u) = degH ′ (u) = 0. We see that Cu = {(u, w)} with u 6= w, because H ′ is connected and contains at least two vertices. By Lemma 5.4 it follows that degH (u) = degH (w) = 0. Since H is connected it follows that H = H3 , so we can apply Rule E3s . This yields a contradiction. So we can assume that H does not contain vertices with degree 0. • If H contains a leaf u 6= v, then degH ′ (u) = degH (u) = 1. We see that u is incident to a green edge uw, with degH (w) 6= 1, otherwise we could apply Rule T1s , T2s or T3s . By Lemma 5.4, it follows that sgon(H) ≥ 3. So we can assume that H does not contain leaves. • If H contains a vertex u 6= v of degree 2, then we see that Cu = {u, w} with u 6= w, by Rules S1s and S2s and by the choice of H ′ . • We see that H ′ does not contain black loops because of Rule Ls . • By Rules P1s and P2s and by the choice of H ′ it follows that H ′ has no multiple edges. Write H ′′ for the graph obtained from H ′ by removing all green loops and coloring all green edges of H ′ black. Altogether we see that H ′′ is a simple graph with at least two vertices and every vertex, except at most one, has degree at least 3. It follows that H ′′ has treewidth at least 3. If we change the color of all green edges to black, we see that all rules are deletions of vertices or edges, contractions of edges and/or additions of loops. Since the set of graphs with treewidth at most k is closed under these operations, we see that tw(G) ≥ tw(H ′′ ) ≥ 3. But then it follows that sgon(G) ≥ tw(G) ≥ 3. This yields a contradiction. We conclude that sgon(G) ≤ 2, then GRs∗ ∅. Lemma 5.15 says that Rs is safe for G2s and from 5.16 it follows that Rs is complete for G2s . So together this proves Theorem 5.5. We conclude that we can use this set of rules to recognize graphs with stable gonality at most 2. 6 Reduction rules for stable divisorial gonality In this section we show a set of reduction rules to decide whether stable divisorial gonality is at most two. This set is similar to the set of rules for stable gonality, and it uses the concept of constraints of divisorial gonality. 24 sd Rule S1a Rule Lsd Figure 7: The reduction rules for stable divisorial gonality that are different from the rules for sgon N OTATION We use the notion of constraints as in the section about divisorial gonality. We will refer to them as red edges. We call an effective divisor of degree 2 that satisfies all conditions given by the constraints a suitable divisor. Again, let G2sd be the set of all graphs with constraints with stable divisorial gonality at most 2. Lemma 6.1. Let G be a graph. If there is a vertex v with |Cv | > 1, then sdgon(G) ≥ 3. Proof. See the proof of Lemma 4.13. Lemma 6.2. Let G be a graph where every leaf is incident to a red edge, so if deg(u) = 1 then |Cu | > 0 for all u. Suppose that u is a leaf and (u, v) is a red edge. If deg(v) 6= 1, then sdgon(G) ≥ 3. Proof. Analogous to the proof of Lemma 4.14. R EDUCTION RULES We will now state all rules. When a rule adds a red edge uv, and there already exists such a red edge, then the set of constraints does not change. The reduction rules for stable divisorial hyperelliptic graphs are almost the same as the rules for stable hyperelliptic graphs. Instead of sd , S sd and Lsd , green edges we use red edges, and we replace Rule S1s and Ls by new Rules S1a 1b see Figure 7 for the new rules. Rule T1sd (=T1s ). Let v be a leaf with Cv = ∅. Let u be the neighbor of v. Contract the edge uv. Rule T2sd (=T2s ). Let v be a leaf with Cv = {(v, v)}. Let u be the neighbor of v. Contract the edge uv. sd . Let v be a vertex of degree 2 with C = ∅. Let u be the only one neighbor of v. Rule S1a v Remove v and add a red loop to u. sd . Let v be a vertex of degree 2 with C = ∅. Let u and u be the two neighbors of v, Rule S1b v 1 2 with u1 6= u2 . Contract the edge u1 v. Rule T3sd (=T3s ). Let G be a graph where every leaf and every degree 2 vertex is incident to a red edge. Let v1 and v2 be two leaves that are connected by a red edge. Let u1 and u2 be their neighbors. Contract the edges u1 v1 and u2 v2 . Rule S2sd (=S2s ). Let G be a graph where every leaf and every degree 2 vertex is incident to a red edge. Let v be a vertex of degree 2 with a red loop, such that there exists a path from v to v in the black and red graph G. Let u1 and u2 be the neighbors of v. Remove v and connect u1 and u2 with a red edge. 25 Rule Lsd . Let v be a vertex with a loop. Remove all loops from v. Rule P1sd (=P1s ). Let uv be an edge. Suppose that there also exists a red edge from u to v. Remove the black edge uv. Rule P2sd (=P2s ). Let u, v be vertices, such that |E(u, v)| > 1. Let e and f be two of those edges. If there exists another path, possibly containing red edges, from u to v, then remove e and f and add a red edge from u to v. Rule E1sd (=E1s ). Let G be the graph consisting of a single vertex v with Cv = ∅. Remove v. Rule E2sd (=E2s ). Let G be the graph consisting of a single vertex v with a green loop. Remove v. Rule E3sd (=E3s ). Let G be the graph consisting of a two vertices u and v that are connected by a green edge. Remove u and v. We will write Rsd for the set of these reduction rules. We can now state our main theorem. Theorem 6.3. The set of rules Rsd is safe and complete for G2sd . S AFENESS We will show the set Rsd is safe for G2sd . Lemma 6.4. Rule T1sd is safe. Proof. Let v be the vertex in G to which the rule is applied. Suppose that sdgon(G) ≤ 2. Since every refinement of G is a refinement of H, it is clear that sdgon(H) ≤ 2. Suppose that sdgon(H) ≤ 2. Then there exists a refinement H ′ of H and a suitable divisor D. Write u for the neighbor of v in G. There exists a divisor D ′ ∼ D such that D ′ (u) ≥ 1. Now define G′ as H ′ with a leaf v added to u. Look at the divisor D ′ on G′ . For every vertex w ∈ H ′ we can reach a divisor with one chip on w. By adding v to every firing set that contains u, we see that we can still reach every vertex of H ′ in G′ . And we can reach a divisor with a chip on v by firing H ′ in G′ . Thus D ′ is a suitable divisor on G′ . We conclude that sdgon(G) ≤ 2. Lemma 6.5. Rule T2sd is safe. Proof. This proof is analogous to the proof of Lemma 6.4. sd is safe. Lemma 6.6. S1a Proof. Let v be the vertex in G to which the rule is applied. Let u be the neighbor of v. Suppose that sdgon(G) ≤ 2. Let G′ be a refinement of G such that there exists a suitable divisor D. Let C be the cycle through v and u. Notice that D is equivalent to a divisor D ′ with two chips on C. If G\{C} ∪ {u} is a tree, then we are done. Otherwise we see that D ′ is equivalent with D ′′ , where D ′′ (u) = 2. Let H ′ be G′ \C ∪ {u} with a red loop at u. We see that H ′ is a refinement of H and D ′′ is a suitable divisor for H ′ , thus sdgon(H) ≤ 2. Suppose that sdgon(H) ≤ 2. Then there exists a refinement H ′ of H such that D, with D(u) = 2, is a suitable divisor. Let G′ be H ′ without the red loop on u and with a vertex v with two edges to u. Then G′ is a refinement of G. It is clear that D is a suitable divisor for G′ too. We conclude that sdgon(G) ≤ 2. 26 sd is safe. Lemma 6.7. S1b Proof. Let v be the vertex in G to which the rule is applied. Let u1 , u2 be the neighbors of v. We know that u1 6= u2 . Suppose that sdgon(G) ≤ 2. Then there is a refinement G′ of G such that there exists a suitable divisor on G′ . Notice that G′ is a refinement of H too, so sdgon(H) ≤ 2. Now suppose that sdgon(H) ≤ 2. Let H ′ be a refinement of H and D a suitable divisor on H ′ . If the edge u1 u2 is subdivided in H ′ , then H ′ is a refinement of G too, and we are done. Assume that the edge u1 u2 is not subdivided in H ′ . Let D ′ be the divisor with D ′ (u1 ) = D ′ (u2 ) = 1. If D ∼ D ′ , then we can subdivide u1 u2 to obtain a refinement G′ of G. By starting with the divisor D ′ , we can reach all vertices of H ′ as before and we can reach v by firing all vertices of H ′ . Suppose that D ≁ D ′ . Then there exist divisors Du1 and Du2 such that Du1 ∼ D ∼ Du2 and Du1 (u1 ) = 1 and Du2 (u2 ) = 1. It follows that Du1 (u2 ) = 0 and Du2 (u1 ) = 0. Let A0 , . . . Ak be the level set decomposition of the transformation from Du1 to Du2 and D1 , . . . Dk the corresponding divisors. Let Di be the first divisor such that Di (u2 ) > 0, then it is clear that u2 ∈ / Aj for j < i. Since Di ≁ D ′ we see that Di (u1 ) = 0. It follows that u1 ∈ Ai . We conclude that there is a chip fired along the edge u1 u2 . For every vertex w in H ′ we can find a divisor Dw with at least one chip on w, let Bw,1 , . . . Bw,lw be all sets that occur in the level set decomposition of the transformation from Du1 to Dw . For all w, i, let Ew,i be the set of all edges along which a chip is fired by the set Bw,i . Subdivide all edges that occur in some Ew,i to obtain a refinement G′ of G. Let Vw,i be set of vertices that are added on the edges in Ew,i . Define Bw,i as Bw,i together with all added vertices that are on an edge with both endpoints in Bw,i . We can replace every set Bw,i by two sets Bw,i , Bw,i ∪ Vw,i in the level set decomposition Bw,1 , . . . Bw,lw to see that for every vertex w we can still reach a divisor with at least one chip on w. And for every vertex in Vw,i we will encounter a divisor with a chip on that vertex when we transform Du1 in Dw . We conclude that sdgon(G) ≤ 2. Lemma 6.8. T3sd is safe. Proof. Let v1 and v2 be the vertices in G to which the rule is applied. Suppose that sdgon(G) ≤ 2. Let G′ be a minimum refinement of G and D a suitable divisor on G′ . Let u1 and u2 be the neighbors of v1 and v2 in G. We distinguish three cases. Case 1: Suppose that u1 6= u2 , and that there does not exist a path from v1 to v2 , except the red edge v1 v2 . Then we can reach all vertices in Gv1 (u1 ) with only one chip, thus Gv1 (u1 ) is a black tree. Thus Gv1 (u1 ) contains a leaf that is not incident to a red edge. This yields a contradiction with the minimality of G′ . Case 2: Suppose that u1 6= u2 and that there exists a path P , possibly containing red edges, from v1 to v2 . Assume that D ≁ D ′ where D ′ is the divisor such that D ′ (u1 ) = D ′ (u2 ) = 1. Let a0 = v1 , a1 , . . . , ak = u1 be the added vertices on the edge v1 u1 and let b0 = v2 , b1 , . . . , bl = u2 be the added vertices on the edge v2 u2 . Assume that k < l. It is clear that all vertices a0 , . . . , ak , b0 , . . . , bl lie on P . Notice that firing the sets {ai , bi | i ≤ j} for all j ≤ k results in the divisor Dk with Dk (ak ) = Dk (bk ) = 1. Thus Dk (u1 ) = 1. Since bk is an internal added vertex and G′ is a minimum refinement, we see that deg(bk ) = 2. Suppose that u1 is incident to a red edge u1 x. We know that x 6= bk , since bk is an added vertex. Let D ′′ be the divisor with D ′′ (u1 ) = D ′′ (x) = 1. Let A1 , . . . , As be the level set decomposition of the transformation from Dk to D ′′ . We see that u1 cannot lose its chip. Thus bk fires its chip to one of its neighbors when we fire A1 . But then we see that the cut of A1 is at least two, and we can only fire one chip. This yields a contradiction. We conclude that u1 is not incident to a red edge. 27 By the conditions of the rule it follows that deg(u1 ) ≥ 3. Let w ∈ / P be a neighbor of u1 . Now ′ ′ we see that Gu1 (w) is a black tree. It follows that Gu1 (w) contains a leaf that is not incident to a red edge. Since G′ is a minimum refinement, this yields a contradiction. Altogether we conclude that k = l. Let P1 , P2 be the two arcs of P between u1 and u2 . Notice that, if there are two chips on P , then they are either on u1 and u2 or on the same arc Pi . Suppose that there are divisors E, E ′ such that E ∼ E ′ and that there is a set A in the level set decomposition of E ′ − E such that u1 ∈ A and u2 ∈ / A. It follows that there is a chip fired along each of the arcs P1 and P2 . This yields a contradiction. We conclude that for every firing set it holds that either u1 and u2 are both fired or they are both not fired. Now let H ′ be G′ without the red edge v1 v2 and with a red edge u1 u2 . We see that D is a suitable divisor for H ′ as well. Thus sdgon(H) ≤ 2. Case 3: Suppose that u1 = u2 . This case is analogous to case 2. Suppose that sdgon(H) ≤ 2. Then it is clear that sdgon(G) ≤ 2. Lemma 6.9. Rule S2sd is safe. Proof. This proof is analogous to the proof of cases two and three in the proof of Lemma 6.8, so we omit it. Lemma 6.10. Rule Lsd is safe. Proof. There will never be a chip fired over a loop, so loops do nothing for the stable divisorial gonality. Thus sdgon(G) ≤ 2 if and only if sdgon(H) ≤ 2. Lemma 6.11. Rule P1sd is safe. Proof. Let uv be the edge in G to which the rule is applied. Suppose that sdgon(G) ≤ 2. Let G′ be a refinement of G and D a suitable divisor with D(u) = D(v) = 1. Let Guv be all internal and external added vertices to the edge uv. Now define H ′ = G′ \Guv . Look at the divisor D on H ′ and notice that D is a suitable divisor. Observe that H ′ is a refinement of H. Thus sdgon(H) ≤ 2. Suppose that sdgon(H) ≤ 2. Let H ′ be a refinement of H and D a suitable divisor. Add an edge uv to H ′ , to obtain a refinement G′ of G. We see that D ′ is a suitable divisor for G′ , thus sdgon(G) ≤ 2. Lemma 6.12. Rule P2sd is safe. Proof. This proof is analogous to the proof of Lemma 4.11. Lemma 6.13. Rules E1sd , E2sd and E3sd are safe. Proof. All those graphs have stable gonality at most 2, so the statement holds true. Now we have proven that all rules are safe, so we can conclude the following. Lemma 6.14. The set of rules Rsd is safe for G2sd . 28 C OMPLETENESS ∗ Now we will prove that Rsd is complete for G2sd , i.e. if G ∈ G2sd , then GRsd ∅. We define the graphs H1 , H2 and H3 as a single vertex, a vertex with a red loop and two vertices connected by a red edge respectively, these are the graphs that can be reduced to the empty graph by rules E1sd , E2sd and E3sd . Let G be a connected graph and apply our algorithm to G, let H be the final graph to which no rule can be applied. Notice that H is connected as well (in black and red edges). Lemma 6.15. The set of rules Rsd is complete for G2sd . Proof. Let G be a graph with sdgon(G) ≤ 2, and let H the final graphs after reducing G to which no rule can be applied. By 6.14 it follows that sdgon(H) ≤ 2. Suppose that H 6= ∅. As in the proof of Theorem 5.16, if there are two edges between the vertices u and v, then removing these edges lead to H being disconnected. And if there is a degree 2 vertex v with a red loop, then removing v yields a disconnected graph. Let H ′ be the smallest connected component that can be created by removing two parallel edges or a degree 2 vertex, as in the proof of Theorem 5.16. Now we color all red edges black and remove all loops to obtain H ′′ , as in the proof of Theorem 5.16. Then we see that H ′′ is a simple graph that contains at least two vertices and all vertices, except at most one, have degree at least three. Thus H ′′ has treewidth at least three. It follows that sdgon(G) ≥ tw(G) ≥ tw(H ′′ ) ≥ 3. ∗ We conclude that if sdgon(G) ≤ 2, then GRsd ∅. Lemma 6.14 shows that the set of reduction rules Rsd is safe and Lemma 6.15 shows that this set is complete. So together this proves Theorem 6.3. Thus we can use the set Rsd to recognize graphs of stable divisorial gonality at most 2. 7 Algorithms In this section, we discuss how the reduction rules of Sections 5, 4 and 6 lead to efficient algorithms that recognize graphs with stable gonality, divisorial gonality, or stable divisorial gonality 2 or lower. Theorem 7.1. There are algorithms that, given a graph G with n vertices and m edges, decide whether dgon(G) ≤ 2, sgon(G) ≤ 2 or sdgon(G) ≤ 2 in O(n log n + m) time. Each of these algorithms is of the following form: first, with an additional rule, we can ensure that on each pair of vertices, there are at most two parallel edges and at most one red or green edge; then, we verify that the treewidth of G is at most 2; if not, we can directly decide negatively. After that, we repeatedly apply a rule, until none is possible. By Theorems 5.5, 4.4, and 6.3, the final graph after all rule applications is empty, if and only if sgon(G) ≤ 2, dgon(G) ≤ 2 or sdgon(G) ≤ 2, respectively. It is not hard to see that for each rule, deciding whether the rule can be applied, and if so, applying it, can be done in polynomial time; the fact that we work with graphs of treewidth 2 yields an implementation with O(n log n) steps, after the O(m) work to remove the extra parallel edges. Proof of Theorem 7.1. We now present the details of the algorithm, in various steps. S TEP 1. First we introduce a new rule for each of the gonalities, cf. Figure 8. For divisorial gonality we introduce the following rule: 29 Rule M d Rule M s Figure 8: An extra reduction rule for divisorial and stable gonality Rule M d . Let u, v be vertices, such that |E(u, v)| ≥ 3. Let k = ⌊(|E(u, v)| − 1)/2⌋, then remove 2k edges between u and v and add a constraint (u, v). Note that this rule merely shortcuts repeated applications of Rule C3d . Now we introduce a new rule for stable gonality. Rule M s . Let u, v be vertices, such that |E(u, v)| ≥ 3. Remove all edges in E(u, v) and add a green edge from u to v. It is clear that this rule is the same as first applying Rule P2s and then applying Rule P1s to all remaining edges (u, v). For stable divisorial gonality we introduce a similar rule. Rule M sd . Let u, v be vertices, such that |E(u, v)| ≥ 3. Remove all edges in E(u, v) and add a red edge from u to v. This is again the same as first applying Rule P2sd and then applying Rule P1sd to all remaining edges (u, v). All applications of these rules can be done in O(m) at the start of the algorithm, after which we know that no pair of vertices can have more than two edges between them. For stable and stable divisorial gonality, by application of Rule Ls and Lsd we can also ensure in O(m) time that no loops exist (in the case of divisorial gonality loops can be safely ignored). S TEP 2. Recall that by Lemma 3.13 treewidth is a lower bound on divisorial gonality, stable gonality and stable divisorial gonality. Therefore it follows that if tw(G) > 2, the algorithm can terminate. Checking whether treewidth is at most 2 can be done in linear time. Hereafter, we assume our graph has treewidth at most 2. S TEP 3. Recall that graphs of treewidth k and n vertices have at most kn edges. It follows that the underlying simple graph has at most 2n edges. By our previous steps there are at most 2 edges between a pair of vertices and no loops, so there are at most 4n edges left. Lemma 7.2. Let G be a graph of treewidth at most 2, with at most 2 parallel edges between each pair of vertices. For each of the three collections of rules, the number of successive applications of rules in bounded by O(n). Proof. For the reduction rules for divisorial gonality from Section 4 note that all rules except Rule C3d always remove at least one vertex. It follows they can be applied at most n times. For Rule C3d note that the only case where it removes no vertex is when the cycle C consists of a double edge between two vertices. Since in this case we remove 2 edges and there are at most 4n edges left, it follows this rule can also be applied at most 2n times. Therefore at most 3n rules can be applied before we reach the empty graph. 30 For the rules in Section 5, consider the following potential function f : let f (G) = n + 2m + g for a graph G with n vertices, m remaining (black) edges, and g green edges. One easily observes that f (G) = O(n), and each rule decreases f (G) by at least one. The same argument holds for the collection of reduction rules of stable divisorial gonality from Section 6. Thus, for dgon and sgon, the given sets of rules already lead to polynomial time algorithms: for each of the rules, one can test in polynomial time for a given graph (with green edges or constraints) if the rule can be applied to the graph, and if so, apply the rule in polynomial time. S TEP 4. In the remainder of the proof, we will argue that there is an implementation that leads to recognition algorithms running in O(n log n + m) time. For the case of divisorial gonality, by Lemma 4.13, each vertex can only have one constraint that applies to it. Therefore, checking for compatible constraints can be done in time linear in the number of vertices that will be removed by the rule. To maintain this property it is necessary to check for conflicts whenever a rule adds a new constraint, but this can be done in constant time per rule application. By keeping track of the degree of each vertex as we apply rules, vertices with degree one or zero can be found in constant time. It is also easy to directly detect when the graph is one induced cycle; e.g., by keeping track of the number of vertices of degree unequal two or with an incident selfloop or colored edge; this takes care of Rule C1d . The remaining problem then is efficiently finding applications of the following rules: (∗) C2d , C3d , S2s , P2s , S2sd and P2sd . To do this in O(n log n) time in total, we use a technique, also employed by Bodlaender et al. [5, Section 6]. (See also [16] for more results on dynamic algorithms on graphs of bounded treewidth.) Due to the highly technical aspects, the discussion here is not self-contained, and assumes a knowledge of techniques for monadic second order logic formulas on graphs of bounded treewidth. We build a data structure that allows the following operations: deletions of vertices, deletions of edges, contractions of edges, adding a (possible colored) self-loop to a vertex changing the color of an edge (e.g., turning an edge into a green edge), and deciding whether the rules (∗) can be applied, and if so, yielding the pair of vertices to which the rule can be applied. First, by Lemma 2.2 from [6], we can build in O(n) time a tree decomposition of G of width 8, such that the tree T in the tree decomposition is binary and has O(log n) depth. We augment G with a number of labels for edges and vertices. Vertices are labelled with a value that can be ‘selfloop’, ‘deleted’, or ‘usual’ (no selfloop, not deleted). Edges are labelled with a value that can be one of the following: ‘usual’ (black edge, without parallel edge); ‘parallel’; ‘red’; ‘green’; ‘deleted’; ‘contracted’. When we perform an operation that changes the multiplicity of an edge, deletes it, or changes its color, we do not change the tree decomposition, but instead only change the label of the edge. For each of the rules in (∗), there is a sentence φ in Monadic Second Order Logic (MSOL) with two free vertices variables, such that φ(v, w) holds if and only if the corresponding rule in (∗) is applicable to vertices v and w. We can modify these sentences φ to sentences φ′ that apply to graphs where edges can be labelled with labels ‘deleted’ or ‘contracted’, i.e., if G′ is the graph obtained from G by deleting edges with the label ‘deleted’, then φ(v, w) holds for G′ if and only if φ′ (v, w) holds for G. 31 First, when we perform a quantification over edges, we add a condition that the edge is not a deleted edge. A quantification ∀F ⊆ E : ψ(F ) becomes ∀F ⊆ E : (∀e ∈ F : ¬deleted(e)) ⇒ ψ(F ); a quantification ∃F ⊆ E : ψ(F ) becomes ∃F ⊆ E : (∀e ∈ F : ¬deleted(e)) ∧ ψ(F ). Secondly, we modify the sentence to deal with contracted edges, by making the following three changes. For any quantification over sets of edges, we ensure that there are no edges with the label contracted. For any quantification over sets of vertices, we add the condition that for any edge with the label ‘contracted’, either both endpoints are in the set or both endpoints of the contracted edge are not in the set. Finally, elementary predicates like “v = w" or v is incident to e, become a phrase in MSOL: “v = w" is translated to an MSOL sentence that expresses that there is a path from v to w (possibly empty) with all edges on the path labeled as ‘contracted’, and “v incident to e" is translated to a property that expresses that there is a path with contracted edges from v to an endpoint of e. In the same way, we can further modify the sentences to also deal with vertices with the label ‘deleted’. A consequence of Courcelle’s theorem [8] is that for a sentence φ(v, w) with two free vertex variables, we have an algorithm that, given a tree decomposition of a graph G of bounded width, determines if there are vertices v and w for which φ(v, w) holds, and if so, outputs the vertices v and w, and that uses linear time. Using the techniques from [5], we can modify this algorithm such that we can update the graph by changing labels of vertices and edges, and do these queries, such that each update and query costs time that is linear in the depth of the tree of the tree decomposition, i.e., O(log n). (For the details, we refer to [5, Section 6].) We are now ready to wrap up. Each of the rules can be executed by doing O(1) deletions of vertices, edges, contractions of edges, adding or deleting a selfloop, or changing color or multiplicity of an edge. As discussed above, each of such operation costs O(log n) time to the data structure, and finding (if it exists) a pair of vertices to which we can apply a rule from (∗) also costs O(log n) time. It follows that the total time is O(log n) times the number of times we apply a rule; as argued earlier in this section, we have O(n) rule applications, giving a total of O(n log n) time, excluding the O(m) time to remove edges with multiplicity larger than 2. This gives Theorem 7.1. Remark 7.3. The constant factors produced by the heavy machinery of Courcelle’s theorem and tree decompositions of width 8 are very large; a simpler algorithm with a larger asymptotic (but still polynomial) algorithm will be faster in practice. 8 Conclusion We have provided an explicit set of safe and complete reduction rules for (multi-)graphs of divisorial, stable, and stable divisorial gonality ≤ 2, and we have shown that stable, divisorial, and stable divisorial hyperelliptic graphs can be recognized in polynomial time (actually, O(n log n + m)). Some problems left unanswered by our work are: (a) Can hyperelliptic graphs be recognized in linear time? (b) What is the complexity of computing stable gonality? (c) Which problems become fixed parameter tractable with gonality as parameter? (d) Is there an analogue of Courcelle’s theorem for bounded gonality (in any of the meanings)? 32 References [1] Stefan Arnborg and Andrzej Proskurowski. Characterization and recognition of partial 3-trees. SIAM J. Algebraic Discrete Methods, 7(2):305–314, 1986. [2] Matthew Baker and Serguei Norine. Riemann-Roch and Abel-Jacobi theory on a finite graph. Adv. Math., 215(2):766–788, 2007. [3] Matthew Baker and Serguei Norine. Harmonic morphisms and hyperelliptic graphs. Int. Math. Res. Not. IMRN, 15:2914–2955, 2009. [4] Matthew Baker and Farbod Shokrieh. Chip-firing games, potential theory on graphs, and spanning trees. J. Comb. Theory, Ser. A, 120(1):164–182, 2013. [5] Hans L. Bodlaender, Pål Grønås Drange, Markus S. Dregi, Fedor V. Fomin, Daniel Lokshtanov, and Michal Pilipczuk. A ck n 5-approximation algorithm for treewidth. SIAM J. Comput., 45(2):317–378, 2016. [6] Hans L. Bodlaender and Torben Hagerup. Parallel algorithms with optimal speedup for bounded treewidth. SIAM J. Comput., 27(6):1725–1746, 1998. [7] Gunther Cornelissen, Fumiharu Kato, and Janne Kool. A combinatorial Li-Yau inequality and rational points on curves. Math. Ann., 361(1-2):211–258, 2015. [8] Bruno Courcelle. The monadic second-order logic of graphs. I: Recognizable sets of finite graphs. Inf. Comput., 85(1):12–75, 1990. [9] Martin Davis, Yuri Matijasevič, and Julia Robinson. Hilbert’s tenth problem: Diophantine equations: positive aspects of a negative solution. Proc. Sympos. Pure Math., XXVIII:323–378., 1976. [10] Josse van Dobben de Bruyn. Reduced divisors and gonality in finite graphs. Bachelor thesis, Leiden University, 2012. URL: https://www.math.leidenuniv.nl/nl/theses/316/. [11] Josse van Dobben de Bruyn and Dion Gijswijt. Treewidth is a lower bound on graph gonality. arxiv: 1407.7055, 2014. [12] Gerd Faltings. Endlichkeitssätze für abelsche Varietäten über Zahlkörpern. Invent. Math., 73(3):349– 366, 1983. [13] Dion Gijswijt. Computing divisorial gonality is hard. preprint, arxiv:1504.06713, 2015. [14] Ragnar Groot Koerkamp. The stable gonality of finite graphs. Bachelor thesis, Utrecht University, 2016. URL: https://dspace.library.uu.nl/handle/1874/335139. [15] Anupam Gupta, Kunal Talwar, and David Witmer. Sparsest cut on bounded treewidth graphs: algorithms and hardness results. In STOC ’13, pages 281–290. ACM, 2013. [16] Torben Hagerup. Dynamic algorithms for graphs of bounded treewidth. Algorithmica, 27(3):292–315, 2000. [17] Kevin Hendrey. Sparse graphs of high gonality. arxiv:1606.06412, 2016. [18] Sergey Norin. New tools and results in graph minor structure theory. In Surveys in combinatorics 2015, volume 424 of London Math. Soc. Lecture Note Ser., pages 221–260. Cambridge Univ. Press, 2015. 33
8cs.DS
arXiv:1609.05609v1 [math.ST] 19 Sep 2016 Selective sampling after solving a convex problem Xiaoying Tian Harris, Snigdha Panigrahi, Jelena Markovic, Nan Bi, Jonathan Taylor Abstract: We consider the problem of selective inference after solving a (randomized) convex statistical learning program in the form of a penalized or constrained loss function. Our first main result is a change-of-measure formula that describes many conditional sampling problems of interest in selective inference. Our approach is model-agnostic in the sense that users may provide their own statistical model for inference, we simply provide the modification of each distribution in the model after the selection. Our second main result describes the geometric structure in the Jacobian appearing in the change of measure, drawing connections to curvature measures appearing in Weyl-Steiner volume-of-tubes formulae. This Jacobian is necessary for problems in which the convex penalty is not polyhedral, with the prototypical example being group LASSO or the nuclear norm. We derive explicit formulae for the Jacobian of the group LASSO. To illustrate the generality of our method, we consider many examples throughout, varying both the penalty or constraint in the statistical learning problem as well as the loss function, also considering selective inference after solving multiple statistical learning programs. Penalties considered include LASSO, forward stepwise, stagewise algorithms, marginal screening and generalized LASSO. Loss functions considered include squared-error, logistic, and log-det for covariance matrix estimation. Having described the appropriate distribution we wish to sample from through our first two results, we outline a framework for sampling using a projected Langevin sampler in the (commonly occuring) case that the distribution is log-concave. 1. Introduction Based on the explosion of freely available and high quality statistical software, practicing data scientists or statisticians can rely on untold numbers of statistical learning methods that explore their data. Having found an interesting pattern through such methods, in order to report their findings in the scientific literature, the data scientist is confronted in defending the significance of their findings. Naively assessing significance using methods that ignore their exploration is recognized as flawed. A common, though not often used, solution is to use data splitting to evaluate the significance. There has been significant recent research on developing methods of inference that have the same type of guarantees as data splitting but offer more power. These methods are generally referred to as methods for selective inference (Benjamini, 2010; Berk et al., 2013; Barber and Candes, 2014; Fithian et al., 2014; Lee et al., 2016; Tian et al., 2015; Tian and Taylor, 2015). Loosely speaking, selective inference recognizes the in1 imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 2 herent selection biases in reporting the most “significant” results from various statistical models and attempts to adjust for the bias in a rigorous framework. At a high level, selective inference involves two stages: first, query the data by applying some function. Often this function might be the solution of some convex optimization problem. Second, posit a model based on the outcome of the query and perform inference for the parameters or statistical functionals in such models. In order to perform valid inference in this second stage, a common approach espoused above is to condition on the result of the query. An alternative approach is to address the problem via reduction to a problem of simultaneous inference Berk et al. (2013). Under the conditional approach, when the query involves solving a convex problem, we are interested in distributions of the form S|β̂(S) ∈ A (1) where S ∼ F represents our data, β̂ denotes the solution to a convex optimization problem and A is some event of interest, set by the data analyst. In this paper, we address computational problems in the conditional approach to selective inference. In adopting this conditional approach, it quickly becomes apparent distributions of the form (1) can be rather complex. In this work, we propose a unified sampling approach that is feasible for a wide variety of problems. Combined with the randomization idea in Tian and Taylor (2015), this significantly improve both the applicability of selective inference in practice, and the power of the selective tests. 1.1. A canonical example As a concrete example, we consider solving the LASSO Tibshirani (1996a) at some fixed λ, as considered in Lee et al. (2016). Suppose we observe data (X, y), with X ∈ Rn×p and y ∈ Rn . The query by the data analyst returns sign(β̂(X, y; λ)) (with sign(0) = 0) where 1 β̂(X, y; λ) = argmin ky − Xβk22 + λkβk1 . 2 β (2) The sign vector can be expressed in terms of Ê, the support of β̂ and ẑE the signs of the non-zero coefficients of β̂. In principle, after observing (Ê, ẑE ) = (Eobs , zE,obs ), there is no restriction on the choice of models. For simplicity, we posit a parametric normal model that is based on E (dropping the ‘obs′ notation) and fixed X,  2 ME = N (XE bE , σE I) : bE ∈ RE (3) 2 with σE known. Target of inference is naturally bj|E , the parameters of this model. Selective inference essentially takes any distribution F ∈ ME and consider its corresponding selective distribution by conditioning on {(Ê, ẑE ) = (Eobs , zE,obs )}. In this case, our selective distribution is exactly of the form imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 3 of (1), with A being the quandrant specified by the nonzero coefficients E and their signs zE . Despite the simple form of A, the constraint induced a much more intricate set in the space of y. In this special case, Lee et al. (2016) has worked out the constraint set on y and selective tests can be computed explicitly. However, this is not generally true, especially in the context of randomized selection procedures proposed by Tian and Taylor (2015); Fithian et al. (2014). In many cases, the formsof selective distributions are not explicitly computable and we have to use sampling methods to approximate the conditional distribution. Given a large enough sample from such distribution, we can construct the selective tests. In the following paragraph, we first introduce the randomized Lasso procedure which is a special case of randomized selection procedures proposed in Tian and Taylor (2015). Then we would illustrate what is the appropriate conditional distribution to sample and given a large sample from this distribution, how we carry out the tests for bj|E . First, we propose to incorporate randomness into (2). In particular, we consider solving ǫ 1 β̂(X, y, ω; λ) = argmin ky − Xβk22 + λkβk1 − ω T β + kβk22 2 2 β (4) where ω ∼ G is a random vector independent of (X, y), whose distribution is chosen by the data analyst, hence known. We will assume that G is supported on all of Rp with density g. The ridge term with small parameter ǫ ensures the problem above has a solution. This resembles the ridge term in the elastic net proposed by Zou and Hastie (2005). This type of randomized convex program has been considered in Tian and Taylor (2015). It is espoused as it significantly increase the power in the inference stage without much loss in the quality of the selected model. In particular, increasing the scale of the randomization inevitably deteriorates the model selection quality, similar to using a smaller training set to choose a model in the context of data splitting. However, even a small amount of randomization has empirically shown a fairly noticable increase in power. Another benefit of randomization include a form of robustness to rare selection events in the selective CLT of Tian and Taylor (2015). After solving (4), we use its solution β̂(X, y, ω; λ) for model selection. Similar to unrandomized version, the appropriate law after model selection is y | β̂(X, y, ω; λ) ∈ A, y ∼ F, where F is a member of the parametric normal model specified by (3). Various selective tests can be constructed based on this law including the goodness-of-fit tests and the p-values and confidence intervals for bj|E s. We discuss how to construct tests for bj|E s here as they are a special case of testing a single parameter in a multi-parameter exponential family. The general approach is laid out in Fithian et al. (2014). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 4 Selective inference recognizes the fact that interest in bj|E is due to the query returning E as the nonzero variables of the solution to (4). In order to provide valid inference, we therefore consider only the part of sample space that will yield the same model, i.e. conditioning on {(Ê, ẑE ) = (Eobs , zE,obs )}. Conditioning each model in ME yields a new parametric model, the selective model   dF ∗ (·) ∝ 1{(Ê,ẑE )=(Eobs ,zE,obs ),F ∈ME } (·) . (5) M∗(Eobs ,zE,obs ) = F ∗ : dF where (Eobs , zE,obs ) are the observed variables and signs. Note that this approach is not tied to the parametric model in (3). In particular, when the parametric normal model is not appropriate, one might replace M with some other statistical model (i.e. some other collection of distributions) and carry out statistical inference in this new selective model. This approach was laid out in detail in Fithian et al. (2014), with some asymptotic justification for nonparametric models developed in Tian and Taylor (2015). In fact, in the Lee et al. (2016) the (pre-conditioning) statistical model proposed was not ME , rather it was the saturated model  (6) M = N (µ, σ 2 I) : µ ∈ Rn . A fairly simple calculation shows that the hypothesis tests and confidence intervals are exactly the same whether we had started with (6) or (3). However this is not the case for examples like forward stepwise, discussed below and in Fithian et al. (2015); Tibshirani et al. (2014). Let us now consider how we might go about inference in (5). For instance, suppose we are simply interested in a goodness-of-fit test and we will consider the model M = {F0 } so that M∗ consists of F0 restricted to the event that fixes the active set and signs of the LASSO to be (Eobs , zE,obs ). Call this distribution F0∗ . A natural way to test H0 : F = F0 would be follow Fisher’s approach by choosing some test statistic T = T (X, y) and compare our observed value Tobs to the distribution of T under F0∗ . In order to do this, it is sufficient to describe the law β̂λ,∗ (F0 ), i.e. the push forward of F0 and condition this law on the event that fixes the active set and signs to be (Eobs , zE,obs ). Unfortunately, this is no easy task. Of course, a natural alternative is to use Monte Carlo. A large enough sample from F0∗ is sufficient to carry out this goodness-of-fit test. A version of this goodness-of-fit test is considered in Section 4.7 below. The issue of what Monte Carlo method to use remains. Our main contribution in this work is an explicit description of how to sample from distributions such as F0∗ . To be precise, we describe how to sample from conditional distributions where the conditioning depends on the solution to a convex problem. A naive way to sample from this distribution would be to take a Monte Carlo sampler to draw from F0 , retaining only those points where the active set and signs of the LASSO agree with (Eobs , zE,obs ). This requires solving (2) at each sample point, and is clearly infeasible. We might say that this is a direct way to condition the push-forward distribution β̂λ,∗ (F0 ) on the selection event. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 5 Our approach here relies on what is essentially the pull-back of the conditional distribution itself. In concrete terms, we construct an explicit inverse map to problems like (2) on the selection event. Then, instead of sampling (1), we realize the law F0∗ by sampling β̂λ together with some auxillary variables on the conditioning event. After aquiring these samples, we can use the inverse map to reconstruct y that is essentially distributed according to the law F0∗ . The construction for (2) is described in Section 2.2. We will see that the pull-back construction is particularly simple for the randomized lasso (4). The randomization was inspired by the differential privacy literature Dwork et al. (2015), though we will see below that it is already very similar to data splitting. Besides making the pull-back simple to compute, we advocate the use of randomization in the query stage in that conditional inference in the second stage is often more powerful if the selection stage is carried out with randomization than without Fithian et al. (2014); Tian and Taylor (2015). Increasing the scale of the randomization inevitably deteriorates the model selection quality, similar to using a smaller training set to choose a model in the context of data splitting. However, even a small amount of randomization has empirically shown a fairly noticable increase in power. Another benefits of randomization include a form of robustness to rare selection events in the selective CLT of (Tian and Taylor, 2015). Inference for linear functionals is slightly more complicated than the simple goodness-of-fit test. To conclude this section, let us describe the form of the pull-back construction when we have solved (4) and observed the active set and signs (E, zE ) (dropping the ‘obs′ notation). In principle, there is no restriction on the choice of models. For simplicity, we posit a parametric normal model that is based on a subset of variables Ē: o n 2 MĒ = N (XĒ βĒ , σĒ I) : βĒ ∈ RĒ 2 with σĒ known. Often, the data analyst may opt to take Ē = Eobs , though this is not strictly necessary. Based on (E, zE ) she might consult the relevant literature (or data independent of y) and choose to include or delete some variables from the set E. When the variance is unknown, one can either plug in 2 a consistent estimate of σE (c.f. Lemma 14 of Tian and Taylor (2015)) or in2 clude the parameter σĒ in the model and solve instead the square-root Lasso Belloni et al. (2014). If a parametric model is not appropriate then one might use a normal approximation for the pair X T y, and rewrite the selection event in terms of this statistic as opposed to just y. Such an approach is considered in Tian and Taylor (2015), appealing to the selective CLT. Having fixed model MĒ , and knowing the density g for the randomization ω, suppose the data analyst now wants to test H0 : βj|Ē under the assumption F ∈ MĒ . Let PĒ\j denote orthogonal projection onto col(XĒ\j ). In the normal z-test for βj|Ē , we essentially consider the distribution of rj = y − µĒ\j , µĒ\j = PĒ\j y. rj can be recognized as the residual from the model with variables Ē \ j. Selecimsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 6 tive tests for βj|Ē is essentially considering the law of rj under the conditional distributions. An application of our main result Theorem 1 implies that this law (with the appropriate augmentation variables) has density proportional to        1 ẑE β̂E T 2 +ǫ· (rj , β̂E , ẑ−E ) 7→ exp − 2 kµĒ\j + rj k2 ·g X XE β̂E + λ · ẑ−E 0 2σĒ (7) supported on n o col(PĒ\j )⊥ × (β̂E , ẑ−E ) : sign(β̂E ) = ẑE , kẑ−E k∞ ≤ 1 . The augmentation variables (β̂E , ẑ−E ) are variables involved in the optimization problem (4). Their support ensures we stay on the selection event of interest. In fact, their support is recognizable as a subset of n o (β̂, ẑ) : ẑ ∈ ∂(k · k1 )(β̂) , with ∂ denoting the subdifferential with respect to β̂. In particular, this is the subset for which the non-zero coefficients of β̂ are E with signs ẑE . In this simple case, the construction of the inverse map is remarkably simple. If we write out the KKT condition for the optimization problem (4), we have X T X β̂ − X T y + λ · ẑ − ω + ǫ · β̂ = 0 which can be rewritten as ω = X T X β̂ − X T y + λẑ + ǫβ̂. where (β̂, ẑ) are the optimization variables and the subgradient of (4). In this sense, we have inverted the KKT conditions on the selection event of interest, explicitly parameterizing the pairs (y, ω) that yield active set and signs (E, zE ). We call this density the pull-back as this is recognizable precisely as the differential geometric pull-back of a differential form on the interior of the selection event, a subset of the manifold Rn × Rp . As g is assumed to have a density, the boundary of the selection event is ignorable. To those readers who do not recognize the term pull-back, Theorem 1 essentially just applies a standard change of variables to realize a conditional law for (y, ω) via a law for (rj , βE , ẑ−E ). Note that the unselective law of rj is that of y projected onto the residual space form model Ē \ j. The reader may ask what we have gained through this exercise. By constructing an inverse map for (y, ω) on the selection event, we do not have to check whether the pair (y, ω) = (y, X T X β̂ − X T y + λ · ẑ + ǫ · β̂) satisfy the KKT conditions for the pair (E, zE ). That is, we have removed the difficulty of conditioning from the problem. We are still left with a sampling problem, though we have an explicit density as well as a fairly simple support. Further, the density (7) is log-concave in (rj , βE , z−E ) with a simple support hence methods such as imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 7 projected Langevin Bubeck et al. (2015) may be run efficiently as the projection operator is cheap to compute. Section 7 discusses several other examples. Finally, and most importantly, the reader who may also be the data scientist may ask what to do with samples from this density in order to carry out a test of H0 : βj|Ē (F ) = 0. Let β̂j|Ē : Rn → R denote the map that computes OLS coefficient j in the model with variables Ē. Then, given a sufficiently large sample (rj,b , βE,b , z−E,b )B b=1 from density (7), the data analyst will compare the empirical distribution of  B β̂j|Ē (µĒ\j + rj,b ) b=1  B = β̂j|Ē (rj,b ) b=1 to the observed value β̂j|Ē (y). Selective confidence intervals can be constructed by tilting the empirical distribution, though if the true parameter is far from 0, then a reference distribution other than (7) is perhaps more appropriate. One might try replacing rj,⊥ in the density above with rj,⊥ − β̄j|E where β̄j|E is an approximate selective MLE or pseudo MLE Panigrahi et al. (2016). We do not pursue this further here, leaving this for future work. A data analyst concerned about slow mixing in a given MCMC scheme to draw from (7) may take some reversible MCMC algorithm and carry out the exact tests described in Besag and Clifford (1989). 1.2. General approach Having described what might be the canonical example, the LASSO with a parametric Gaussian model for inference, we now lay out our general approach. We consider a randomized version of the optimization problem in (1) formulated as ǫ β̂(S, ω) = argmin ℓ(β; S) + P(β) − ω T β + kβk22 (8) p 2 β∈R where ℓ is some smooth loss involving the data, P is some structure inducing convex function, ǫ > 0 is some small parameter that is sometimes necessary in order to assure the program has a solution and ω ∼ G is a randomization chosen by the data analyst. Our main goal is to sample (S, ω)|β̂(S, ω) ∈ A (9) Our reasons for considering the problem (8) rather than an unrandomized problem are described in the LASSO example above and also hold for the optimization and sampling problems. Namely, we expect an increase in power in the second stage following even a small randomization. Further, as in the LASSO example above, this randomization often allows us to cast the sampling problem as sampling from a distribution on a space that is much simpler than if we had not randomized. That is, the law induced by the pull-back measure in a randomized program is often supported on a simpler region, as opposed to the push-forward measure imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 8 of the non-randomized program. For inference in the non-randomized case, one might take the approach of sending the scale of randomization to 0, though we do not pursue this here. Another feature of the sampling problem related to (8) is that the sampler somewhat decouples the statistical model from the optimization variables. In this sense, our main result provides ways to sample in a model-agnostic fashion: data analysts can supply their own model M, resulting in selective model M∗ . Of course, for inference in M∗ , the analyst may have to use other techniques to reduce their problem to sampling from a particular distribution in M∗ . In the example above, standard exponential family techniques were used to eliminate nuisance parameters βĒ\j|Ē (F ) and the problem was reduced to sampling from only one distribution, constructed by conditioning a distribution in M∗ on the sufficient statistic corresponding to the nuisance parameters. By decoupling, we do not mean statistical independence in any sense. We mean that each distribution in the corresponding selective model is supported on Ω × C where Ω is the original probability space for our data S and C is a set of optimization variables related to the structure inducing function P. Formally, we should note that in the pull-back variables, the selective model describes the distribution of tuples (s, β, z) rather than the original probability space Ω × Rp . Hence, our sampler produces tuples (s, β, z) rather than pairs (s, ω) though ω can always be reconstructed via the map ω = ∇ℓ(β; s) + z + ǫ · β. The two most common examples of interest in statistical learning are P(β) = hK (β) = sup ν T β ν∈K for some convex K ∋ 0, i.e. a seminorm. In this case, C is typically a subset of the normal bundle of K (c.f. Adler and Taylor (2007); Schneider (1993) ) {(β, z) : β ∈ Nz K} where Nz K is the normal cone of K at z, polar to the support cone of K at z. The other common example is a constraint on a seminorm, i.e. ( 0 β∈K P(β) = IK (β) = (10) ∞ β 6∈ K where K = {b : kbk ≤ 1} for some seminorm k·k. In this case C is again typically a subset of N (K) of the form {(β, z) : z ∈ Nβ K} . (11) Such sets arise naturally from the KKT conditions of (8): ω = ∇ℓ(β; S) + z + ǫ · β. Queries related to active sets of variables in the case of the LASSO or group LASSO or the rank of matrices in the case of the nuclear norm correspond imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 9 to smooth subsets of the corresponding normal bundles. When the structure inducing penalty is polyhedral in nature, the sets C are typically polyhedral. For norms with curved unit balls this is no longer the case, and curvature comes in to play. We treat the group LASSO as a canonical example of this in Section 5. 1.3. Related work Most of the theoretical work on high-dimensional data focuses on consistency, either the consistency of solutions Negahban et al. (2010); Van de Geer (2008) or the consistency of the models Wainwright (2009); Zhao and Yu (2006). In the post selection literature, Berk et al. (2013) proposed the PoSI approach, which reduce the problem to a simultaneous inference problem. Because of the simultaneity, it prevents data snooping from any selection procedure, but also results in more conservative inference. In addition, the PoSI method has extremely high computational cost, and is only applicable when the dimension p < 30 or for very sparse models. The authors Meinshausen et al. (2012) proposed a method for computing p-values that controls false discovery rate (FDR) among all variables. The knockoff filter of Barber and Candes (2014) provides similar control of FDR for all variables in the full model. What distinguishes the conditional approach from these simulataneous approaches is that the hypotheses tested, or parameters for which intervals are formed, in selective inference are chosen as a function of the data. Hence, the methods of inference are not always directly comparable. 1.4. Outline of paper We propose the main Theorem 1 of this paper in Section 3, putting forth the sampling density conditional on a selection event. The highlight of this theorem is the reparametrization map that allows to reconstruct the randomization as a function of optimization variables and data and allows us to compute the required conditional density explicitly, up to a normalizing constant. This is followed by a variety of examples of convex optimization programs. The support of the selective density in the first set of examples- LASSO and variants in Section B, graphical models in Section 4.8, forward stepwise in Section 4.4 can be described by polyhedral geometry. We follow this up with a section describing more complex problems where the Jacobian involves a curvature component, the group lasso illustrated as the prototypical example. We extend the selective sampler to the case in which the data analyst considers multiple views, or queries, of the data in Section 6. A selective version of Fisher’s exact test in which the data analyst chooses the sufficient statistics based on the data is described in Section 4.7. Finally, we advocate the projected Langevin sampling technique Bubeck et al. (2015) in Section 7 to sample from a log-concave selective density as described in earlier examples. Each update in such an implementation involves a projection onto a set of constraints induced by the selection event. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 10 The computational cost of each step is often minimal due to a much simpler constraint region using our reparametrization. 2. Inverting the optimization map In this section, we consider the general problem of constructing an explicit inverse to the solution of a convex problem. In this section, we focus on convex problems without additional randomization, deferring randomization to Section 3. Consider a statistical learning problem of the form minimize ℓ(β; S) + P(β), p β∈R S ∼ F, (12) where F is some distribution in some model M and P is some structure inducing convex function, typically of the form (10) or (11). Cone constraints are also easily handled. The subgradient equation for such a problem at a solution β̂(S) reads 0 ∈ α̂(S) + ẑ(S) (13) (S, β̂(S), α̂(S), ẑ(S)) ∈ S F (ℓ, P), (14) with where def S F (ℓ, P) =  (s, β, α, z) : s ∈ supp(F ), ℓ(β; s) < ∞, α ∈ ∂ℓ(β; s), P(β) < ∞,  z ∈ ∂P(β) . (15) Above, and throughout, ∂ and ∇ will denote subdifferentials and derivatives with respect to β unless otherwise noted. The set S F (ℓ, P) can be described by the “base space” supp(F ) and “fibers” {(β, α, z) : ℓ(β; s) < ∞, α ∈ ∂ℓ(β; s), P(β) < ∞, z ∈ ∂P(β)} . We call the map θ̂ s 7→ (s, β̂(s), α̂(s), ẑ(s)) ∈ S F (ℓ, P) (16) the optimization map. In practice, given data S, a computer solves the problem, i.e. produces a point in θ̂(S) ⊂ S F (ℓ, P). Specifically, a solver produces a point in the range of the optimization map:  S0F (ℓ, P) = (s, β, α, z) ∈ S F (ℓ, P) : α + z = 0 . (17) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 11 Formally speaking, the program (12) may have no solutions. On this set θ̂(s) = ∅. In all of our examples except the dual problem considered in Section B.4, we will assume enough so that our convex programs have unique solutions, when they have any. The selection events Lee et al. (2016); Fithian et al. (2014) we are most interested in are typically of the form  SBF (ℓ, P) = (s, β, α, z) ∈ S0F (ℓ, P) : (β, α, z) ∈ B(s) (18) for some nice set-valued function B(s) which could be specified by the zero-set of a function hB : (β, α, z) ∈ B(s) ⇐⇒ hB (s, β, α, z) = 0. In all examples below except basis pursuit in Section B.2, B(s) = B does not depend on s. This set also has the form of a bundle with base space supp(F ) and fibers {(β, α, z) ∈ B(s) : α + z = 0} . We call SBF (ℓ, P) parametrizable if there exists a measurable parametrization ψ defined on some domain D with range SBF (ℓ, P). The map ψ is typically constructed to be an inverse of the optimization map θ̂ on SBF (ℓ, P). Given a parameterization, we will typically  construct a change of measure using φ = πB ◦ ψ to simplify sampling from s : πB−1 (s) 6= ∅ by sampling from SBF (ℓ, P) itself, where πB is the projection onto the base of SBF (ℓ, P). Our construction is similar to what Zhou (2014) called estimator augmentation. By construction, then, πB (s, β, α, z) is such that θ̂ ◦ πB = id|SBF (ℓ,P) with id the identity map. Transforming the probability space and constructing new data vectors that solve (12) as functions of optimization problems is key to our approach of sampling. The explicit parametrization changes with each problem. Typically, in selective inference examples, we condition on some function of β̂, perhaps its support E and possibly the signs zE of the non-zero coefficients. More generally, might condition on something besides (E, zE ) which we might denote by q. The selection event, i.e. the quantity we condition on determines the set of constraints Bq . We shall denote the parametrization map in our problems as ψq , so often we will write ψ(E,zE ) . The map ψq always produces a point (s, β, α, z) ∈ SBFq (ℓ, P), while the domain of this parametrization can vary with the problem at hand, that is ψq : Dq → SBFq (ℓ, P), for a problem specific domain Dq . imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 12 2.1. Inverting the MLE We now begin to describe our approach to inverting the optimization map. Our first example is classical: the density of the MLE of the natural parameters of an exponential family. The formula is not new, going back at least to Fisher (Barndorff-Nielsen, 1983; Efron and Hinkley, 1978). Nevertheless, it serves to illustrate the general approach we take for the general cases later. Our loss function is ℓ(β; S) = Λ(β) − β T S where eΛ(β) = EF0 [eβ T S ] is the moment generating function in the exponential family with reference measure F0 and sufficient statistic S with Fβ the law of S above and in what follows. We also assume that F0 has a density f0 with respect to Lebesgue measure. As we are computing the MLE, our penalty function is P(β) ≡ 0. The convex program we solve is minimize Λ(β) − β T S. (19) p β∈R The KKT conditions or subgradient equation here is just the usual score equation ∇Λ(β̂(S)) = S where ∇Λ(β) = Z sfβ (ds). Rp Our parameterization of S0F (ℓ, P = 0) is ψ(β) = (∇Λ(β), β, 0, 0) which can be interpreted as reconstructing S given β. Standard multivariate calculus then tells us that the density of β̂(S) under the distribution Fβ0 is e∇Λ(β) T β0 −Λ(β0 ) det(∇2 Λ(β))f0 (∇Λ(β)). (20) where ∇2 Λ(β) is the observed information (Efron and Hinkley, 1978). If β0 is the true parameter, then, at the cost of changing the reference measure by a factor of exp(sT β̂(s) − Λ(β̂(s)) det(∇2 Λ(β))1/2 yielding a new Lebesgue density h0 , we can rewrite this as eΛ(β)−Λ(β0 )+∇Λ(β) T (β0 −β) det(∇2 Λ(β))1/2 h0 (∇Λ(β)). (21) As pointed out in (Barndorff-Nielsen, 1983) ignoring the term h0 (∇Λ(β)) (which is the constant 1 in the Gaussian case) yields the usual saddle-point approximation to the density of the MLE, up to the constant of integration. The exponential above can be rewritten as 1 − (β − β0 )T ∇2 Λ(β)(β − β0 ) + R(β; β0 ). 2 imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 13 Note that even ignoring the remainder, this is quadratic in β0 the parameter, and not β the variable of integration in the density. In principle, nothing above really relies on the exponential family structure for the model, though it does rely somewhat on the fact that the loss we use came from an exponential family. It relies on this in that we use the form of the loss to reconstruct data S from optimization variables β. This is similar to what we see in the LASSO example below. Nevertheless, the same argument above shows that if we solve the program (19) then, so long as S ∼ F has a Lebesgue density the density of β̂(S) is f (∇Λ(β)) · det(∇2 Λ(β)) . (22) In this sense, the above display provides an exact recipe to compute the density of the MLE under model misspecification. This is somewhat similar to the general approach taken in Hillier and Armstrong (1999), though we are considering this only in a very restricted setting. 2.2. Pull-Back of the LASSO with fixed design matrix As a second example of a pull-back, we look at the canonical example in the class of regularized convex optimization problems: the LASSO Tibshirani (1996b). The LASSO program is defined for each (X, y, λ) ∈ Rn×p × Rn × (0, ∞) as 1 ky − Xβk22 + λkβk1 . minimize β∈Rp 2 (23) In this example, X is considered fixed and F = L(PC y|X) where PC is projection onto col(X). The law F is supported on col(X) because the optimization map depends only on PC y. In our general notation, we can take S = PC y 1 ℓ(β; y) = ky − Xβk22 2 P(β) = λkβk1 n S F (ℓ, P) = (y, β, α, z) : PC y = y, β ∈ Rp , (24) o α = X T Xβ − X T y, z ∈ ∂(λk · k1 )(β) . The familiar subgradient equations of the LASSO are − α̂(y) = X T (y − X β̂(y)) = ẑ(y), ẑ(y) ∈ ∂(λk · k1 )(β̂(y)). (25) Note also that (25) contains the implicit constraint ẑ(y) ∈ row(X) where row(X) is the rowspace of X. This can be seen from the structure of   F F T S0 (ℓ, P) = (y, β, α, z) ∈ S (ℓ, P) : X (Xβ − y) + z = 0 imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 14 as each point in S0F (ℓ, P) has z ∈ row(X). In Lee et al. (2016), the authors assume general position so the map θ̂ is single valued Tibshirani (2013). The authors then considered the active set and signs of the LASSO solution, conditioning on their value (E, zE ) and the design X. This is the event n o y : (y, β̂(y), α̂(y), ẑ(y)) ∈ S0F (ℓ, P), diag(zE )β̂E (y) > 0, β̂−E (y) = 0 . (26) Note that this event is equivalent to  θ̂(y) ∈ (y, β, α, z) ∈ S0F (ℓ, P) : diag(zE )βE > 0, β−E = 0 . This is our canonical example of a parameterizable set where B = B(E,zE ) = {(β, α, z) : diag(zE )βE > 0, β−E = 0, zE = λ sign(βE ), kz−E k∞ ≤ λ} . The above constraints on β induce further restrictions on the range of the solver SBF(E,z ) (ℓ, P) = {(s, β, α, z) ∈ S0F (ℓ, P) : β ∈ B(E,zE ) } E = {(s, β, α, z) : diag(zE )βE > 0, β−E = 0, α = −X T (s − Xβ) = −z, zE = λ sign(βE ), kz−E k∞ < λ}. The authors in Lee et al. (2016) then carry out selective inference for linear functionals η T µ in the saturated model  (27) M = N (µ, σ 2 ) : µ ∈ Rn with σ 2 > 0 considered known. In this context, selective inference corresponds to taking each F ∈ M and conditioning it on the event (26) which can be rewritten as   πS SBF(E,z ) (ℓ, P) E where πS denotes projection onto the data coordinate. The resulting selective model  ) ( ( F ∗ S (ℓ, P) 1 y ∈ π dF S B (E,zE ) (y) ∝ M∗ = F ∗ : dF 0 otherwise. is an exponential family and sampling is generally not necessary in the saturated model as valid inference typically requires conditioning on sufficient statistic related to nuisance parameters as described in Lee et al. (2016); Fithian et al. (2014). Nevertheless, if sampling were necessary, a naive accept reject sampling scheme for inference in this setting draws vectors y on Rn according to N (µ, σ 2 ) solves the LASSO with the triple (y, X, λ) and checks whether the result lies in (26). Due to the nature of the LASSO, this check can be reduced to verifying whether y satisfies a set of affine inequalities Lee et al. (2016). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 15 2.2.1. Parametrization and pull-back of the LASSO What if we did not have to check these affine inequalities in our sampling scheme? This is the essence of what we propose in this work. We will ultimately see that our approach is essentially equivalent to that of Lee et al. (2016) but the probability space of our sampler is different. The approach we take is similar to (Zhou, 2014). We will see later that, after randomization, the parameter space is generically simpler than if we had not randomized. Our first example of a parameterization is ψ(E,zE ) (β, z) = (Xβ + (X T )† z, β, −z, z) (28) with domain D(E,zE ) = {(β, z) : diag(zE )βE > 0, β−E = 0, z ∈ row(X), zE = λ sign(βE ), kz−E k∞ ≤ λ} (29) and corresponding data reconstruction map φ(E,zE ) : D(E,zE ) → col(X) defined by φ(E,zE ) (β, z) = Xβ + (X T )† z. We see that ψ(E,zE ) is indeed a parameterization of S0F (ℓ, P). Hence, on the range of φ(E,zE ) we do not actually have to check the affine inequalities of (Lee et al., 2016) as they are satisfied by construction. The set D(E,zE ) is a subset of a 2p-dimensional set but is in fact of dimension rank(X) whenever |E| < rank(X). Hence, it has Lebesgue measure 0 in R2p but it is a subset of an affine space so that it inherits a Hausdorff measure Hrank(X) . In a formal sense, the map φ(E,zE ) is the inverse of the optimization map. Lemma 1. On the set of y where that the solution to (23) is unique (X being considered fixed), the following equality holds θ̂−(y,α) ◦ φ(E,zE ) (β, z) = (β, z) for (β, z) ∈ D(E,zE ) , (30) where θ̂−(y,α) (y) = (β̂(y), ẑ(y)). Hence, θ̂−(y,α) ◦ φ(E,zE ) is equivalent to the identity map on D(E,zE ) . More generally, the set (26) is equal to φ(E,zE ) (D(E,zE ) ). Proof. Direct verification of (25). Now choose an element in M, i.e. fix some µ ∈ Rn and consider the N (µ, σ 2 I) density. We define its pull-back to be the law Fµ∗ supported on D(E,zE ) with Hrank(X) density   1 (31) fµ∗ (β, z) ∝ exp − 2 kφ(E,zE ) (β, z) − µk22 · Jφ(E,zE ) (β, z) , 2σ where Jφ(E,zE ) is the Jacobian of the projection of the parameterization, which depends only on X in this case (and can be treated as a constant). We use the imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 16 name pull-back as this density is precisely the pull-back of the measure with Lebesgue density fµ under the reconstruction map φ(E,zE ) . As each density has a pull-back, the model itself has a pull-back M∗ comprising the pull-back of each density. The model also inherits the exponential family structure of M. Taking f0∗ to be the reference measure, we can choose the natural parameter to be µ/σ 2 and sufficient statistic to be φ(E,zE ) (β, z). Finally, note that nothing about our construction so far has anything to do with the parametric model (27). In fact, as long as F has a Lebesgue density the pull-back is well-defined. Lemma 2. Suppose F = L(PC y|X) has density f on C = col(X) and the solution to (23) is F -a.s. unique. Then, the following equality holds   Ly∼F PC y (y, β̂(y), α̂(y), ẑ(y)) ∈ S0F (ℓ, P), diag(zE )β̂E (y) > 0, β̂−E (y) = 0, X  = L(β,z)∼F ∗ φ(E,zE ) (β, z)|X , (32) where F ∗ is supported on relint(D(E,zE ) ) and has Hrank(X) density  f ∗ (β, z) ∝ f φ(E,zE ) (β, z) · Jφ(E,zE ) (β, z) . Proof. Follows from Lemma 1 and standard multivariate calculus combined with the fact that the existence of the density f implies that F (φ(E,zE ) (D(E,zE ) \ relint(D(E,zE ) ))) = 0. Remark 1. One of the consequences of the above is that one can draw response vectors y that have the same active set and signs from a density with relatively simple support when row(X) = Rp . Hence, sampling IID from some density on this support allows one to use importance sampling with the above explicit density as numerator in the importance weight. When row(X) ( Rp , the support is still somewhat complex. For instance, the subgradients z must be in the intersection of a face of the ℓ∞ ball which may be a difficult set to describe. We will see that after randomization, this complexity often disappears. Remark 2. In theory, one might want to drop the uniqueness assumption above. If uniqueness of the solution does not hold then the parameterization is not injective and the event we condition on should be replaced with the event θ̂(y) ∩ SBF (ℓ, P) 6= ∅. In this case, it may still be possible to derive a formula for the law of PC y by invoking the co-area formula Federer (1959); Diaconis et al. (2013). We do not pursue this generalization here. In order to carry out selective inference for a linear functional η T µ, the authors of Lee et al. (2016) conditioned on Pη⊥ y. This is also possible in the pullback model. Fix a linear subspace L ⊂ col(X) considered to be the model imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 17 subspace so that  ML = N (µ, σ 2 I) : µ ∈ L . For each linear function of interest η ∈ L, one generally must condition on (PL − Pη )y to eliminate the nuisance parameter (PL − Pη )µ. If w is the observed value of (PL −Pη )y, a straightforward modification of the argument above shows that the appropriate distribution is supported on the rank(X) − dim(L) + 1 dimensional set D(E,zE ) (w) = {(β, z) ∈ D(E,zE ) : (PL − Pη )φ(E,zE ) (β, z) = w} with Hrank(X)−dim(L)+1 density proportional to fµ∗ . In this case, even if row(X) = Rp , the support is somewhat complex. 3. Inverting the optimization map of a randomized convex program In this section, we describe how to invert the KKT conditions for our family of randomized convex programs. As mentioned in the introduction, randomization comes with advantages like enhanced statistical power in the inference stage, as well as a simplification of the support of the relevant reference distribution. In what follows, all of our convex programs have random variables appearing linearly in the subgradient, so that these random variables can be reconstructed from optimization variables. The random variable is one introduced by a data analyst through additional randomization as described in (Dwork et al., 2015; Tian and Taylor, 2015). Let G be a distribution on Rp and ǫ ≥ 0 a small parameter. Given a canonical problem specified by (F, ℓ, P) we define its randomized version as ǫ minimize ℓ(β; S) + P(β) − ω T β + kβk22 , β∈Rp 2 (S, ω) ∼ F × G. (33) A natural question to ask at this point is: why randomize the program in the above way? One of the inspirations for selective inference after randomization, the topic considered in (Tian and Taylor, 2015) are techniques used in differential privacy (Dwork et al., 2015). The other was the empirical finding in (Fithian et al., 2014) that holding out some data before carrying out selective inference as in Lee et al. (2016) also showed an improvement in selective power. This increase in selective power can be attributed to the fact that choosing a model with a randomized response has more leftover information as defined in Fithian et al. (2014) after model selection than choosing a model with the original response. The choice to add 2ǫ kβk22 − ω T β to the objective is not the only reasonable choice. The main property we want of the perturbation ∆(β, ω) is that ω = φ(β, v) where (β, v) ∈ ∂∆(β, ω). That is, given the value of β and the subgradient v we can reconstruct ω. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 18 Remark 3. Addition of the term 2ǫ kβk22 to the objective ensures that the set of (s, ω) such that (33) has a solution contains the set {(s, ω) : θ̂(s) 6= ∅, ω ∈ supp(G)}. To see this, note that solving (33) is equivalent to computing β̂(s, ω) = prox 1ǫ (ℓ(·;s)+P(·)) (ω/ǫ). Whenever θ̂(s) 6= ∅, the objective ℓ(·; s)+P(β) is a closed proper convex function, hence its proximal mapping is well-defined and β̂(s, ω) is well-defined. If ℓ is strongly convex a.s.-F and P is closed and proper then we can take ǫ = 0. 3.1. Data splitting as randomization The suggested randomization above may strike some readers as lacking motivation. A randomization mechanism probably familiar to most readers is data splitting (Cox, 1975), i.e. randomly partitioning the dataset into two pieces of size (n1 , n − n1 ). We will see that solving a convex program after data splitting is essentially an example of (33). For a specific example, suppose we fit the graphical LASSO Friedman et al. (2008) based on a data matrix X ∈ Rn×p but we first randomly split the data. That is, we form   1 1 T T S1 (g) = X1 (g), In1 − X1 (g) 1n 1 n1 n1 1 n1 with g denoting the random partition of the data into two groups and X1 (g) denoting the data in the first group. Next, we solve the program minimize − log det(Θ) + Tr(S1 (g)Θ) + λP(Θ) Θ:ΘT =Θ,Θ>0 with the usual caveat that the penalty does not charge diagonal elements of Θ. This program is equivalent to solving the problem minimize − log det(Θ) + Tr(SΘ) + Tr((S1 (g) − S)Θ) + λP(Θ), Θ:ΘT =Θ,Θ>0 where S= 1 T X n   1 I − 1n 1Tn X. n This is very close to (33) with ω = ω(g) = S − S1 (g) which is orthogonal to, but not necessarily independent of the full covariance S. As described in Fithian et al. (2014), data splitting can be interpreted as simply conditioning on the value (S1 (g), g) as the only variation in L(S|S1 (g), g) is S2 (g), the second stage data. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 19 3.2. The selective sampler In this section, we describe our main tool for inference after solving a convex program randomized in the above fashion. We call this tool the selective sampler. A solver for the program (33) produces a point in  def S̄ F (G, ǫ, ℓ, P) = (s, ω, β, α, z) : s ∈ supp(F ), ω ∈ supp(G), ℓ(β; s) < ∞, (34) α ∈ ∂ℓ(β; s), P(β) < ∞,  z ∈ ∂P(β) . More precisely, it produces a point in  S̄0F (G, ǫ, ℓ, P) = (s, ω, β, α, z) ∈ S̄ F (G, ǫ, ℓ, P) : ǫ · β + α + z − ω = 0 . (35) Our statistical learning task is typically to infer something about F in some model M. As we are free to choose G however we want, a natural choice is to choose G to have a Lebesgue density supported on all of Rp . In this case, inspection of the KKT conditions or subgradient equation of (33) read ω = α̂(s, ω) + ẑ(s, ω) + ǫ · β̂(s, ω). (36) The subgradient equation determines a canonical map ψ : S F (ℓ, P) → S̄0F (G, ǫ, ℓ, P) (37) ψ(s, β, α, z) = (s, α + z + ǫ · β, β, α, z). (38) defined naturally as As in the non-randomized case, we are typically interested in some selection events of the form S̄BF (G, ǫ, ℓ, P) = {(s, ω, β, α, z) ∈ S̄0F (G, ǫ, ℓ, P) : (β, α, z) ∈ B(s)}. The corresponding randomization reconstruction map is then φ(β, α, z) = α + z + ǫ · β, defined on a domain {(β, α, z) ∈ B(S) : α + z + ǫ · β ∈ supp(G)}, imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 20 for a suitable B(S), defined by the selection event. It turns out that many selection events of interest are such that the restriction of ψ to these events have a simple structure which allows for straightforward description of the selective model. The canonical example of conditioning on the set of active variables and signs of the LASSO as in (Lee et al., 2016) was described in Section 1.1. Theorem 1 (Selective sampler). Suppose that ω is independent of S with distribution G such that supp(G) ⊂ Rp has non-empty interior with Lebesgue density g on supp(G). Then, the map ψ(s, β, α, z) = (s, ǫ · β + α + z, β, α, z) = (s, φ(β, α, z), β, α, z) restricted to  DG = (s, β, α, z) ∈ S F (ℓ, P) : ǫ · β + α + z ∈ supp(G) is onto S̄0F (G, ǫ, ℓ, P). Further, the law LF ×G ((S, ω)|(S, ω, β̂(S, ω), α̂(S, ω), ẑ(S, ω)) ∈ S̄BF (G, ǫ, ℓ, P)) = L((S, ǫβ + α + z)|(s, β, α, z) ∈ DG , (β, α, z) ∈ B(S)) (39) for suitable B(S) and (S, β, α, z) has density proportional to f (s) · g(ǫ · β + α + z) · |Jψ(s, β, α, z)| · 1DG (s, β, α, z) · 1B(s) (β, α, z) = f (s) · g(φ(β, α, z)) · det(D(β,α,z) φ) · 1DG (s, β, α, z) · 1B(s) (β, α, z) (40) with the Jacobian denoting the derivative of the map ψ with respect to (β, α, z) on the fiber over s. Proof. The fact that ψ restricted to DG is onto follows from its construction. Let F (s) = {(β, α, z) : α ∈ ∂ℓ(β; s), z ∈ ∂P(β)} denote the fiber over s and assume that B(s) = B does not vary with s. In this case, in any local coordinates on F (s) standard multivariate calculus yields the density (40) as the derivative of the map (s, β, α, z) 7→ (s, α + z + ǫ · β) takes the form  In×n Ds φ 0n×p  D(β,α,z) φ with determinant det(D(β,α,z) φ). The result then follows by integrating over supp(F ) with density f (s). If S does not have a density then (39) can be derived via the Kac-Rice formula Taylor et al. (2013); Adler and Taylor (2007). If B depends on s then the Kac-Rice formula may often be used to derive the above density given the B can be described as the zero set of some smooth function hB . As most of our examples are such that B does not depend on s we omit the details. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 21 Remark 4. The correct interpretation of (39) has as conditioning event θ̄−(s,ω) (s, ω) ∩ B(s) 6= ∅ where θ̄ is the map θ̄(s, ω) = (s, ω, β̂(s, ω), α̂(s, ω), ẑ(s, ω)). Remark 5. When supp(G) = Rp then DG = S F (ℓ, P). Remark 6. Above, we have constructed ω as a function of (s, β, α, z). This is similar to the LASSO case in Section 2.2 in that we construct new random variables out of optimization variables. Often, it is also possible to construct s from (ω, β, α, z) though there is no canonical embedding unless we provide more structure to the map ℓ. If ℓ is an exponential family negative log-likelihood so that ℓ(β; s) = Λ(β) − sT β then this is certainly possible. However, in order to have a change of measure result as in Theorem 1, the law of S should have a density on Rn . Remark 7. Often, we will want to condition on some functions of S. Theorem 1 formally holds unchanged for any distribution supported on a lower dimensional subset of S. One simply replaces the law F with the appropriate law supported on a lower dimensional set. This device was used in Section 1.1. A further example of this is considered in Appendix A below. Remark 8. The theorem assumes ω is independent of the data S. It is straightforward to see that a similar result holds if we replace throughout the density g above with K(ω; s) a kernel for the conditional density of ω|s. 4. Polyhedral examples We now begin describing several instances of the selective sampler. In this section, the penalties or constraints are polyhedral. In this case, the reconstruction maps are typically affine in the optimization variables but may be non-trivial in the data. Some of these examples were also considered in Tian et al. (2016). We repeat them here, in more explicit detail, as concrete examples of the selective sampler. 4.1. LASSO with Gaussian errors and fixed design matrix As is often the case, the LASSO serves as a canonical example. We denote the parametrization, based on the active set and signs as ψ(E,zE ) on domain D(E,zE ) and the reconstruction map for randomization ω as φ(E,zE ) throughout. In all the below examples, we are implicitly thinking of cases when supp(G) = Rp and G has a Lebesgue density, the canonical example being N (0, σω2 Ip ). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 22 This example was addressed in Section 1.1, though we present it here in the general notation developed so far. The embedding in Theorem 1 plays the role of ψ(E,zE ) in the parametric LASSO example. The randomized LASSO program Tian and Taylor (2015) with randomization G is defined for each (y, X, λ, ω) ∈ Rn × Rn×p × (0, ∞) × Rp as ǫ 1 ky − Xβk22 + kβk22 − ω T β + λkβk1 , minimize p β∈R 2 2 y|X ∼ F, ω ∼ G (41) with ω independent of (X, y). When G = δ0 , we recover the parametric LASSO with ǫ = 0 and the parameteric Elastic Net if ǫ > 0 Zou and Hastie (2005). Supposing then that supp(G) = Rp , and the canonical selection event given by B(E,zE ) = {β : diag(zE )βE > 0, β−E = 0} . A parametrization of {(s, ω, β, α, z) ∈ S̄0F (G, ǫ, ℓ, P) : β ∈ B(E,zE ) } is given by      βE zE T ψ(E,zE ) (y, βE , u−E ) = y, ǫ , − X (y − XE βE ) + λ 0 u−E     βE zE , −X T (y − XE βE ), λ 0 u−E with the domain D(E,zE ) = {(y, βE , u−E ) : diag(zE )βE > 0, ku−E k∞ ≤ 1} . The reconstruction map for ω is given by     β zE φ(E,zE ) (y, βE , u−E ) = ǫ E − X T (y − XE βE ) + λ , 0 u−E (42) (43) again with the same support as the parametrization map. For the canonical event B(E,zE ) , we therefore need to sample from a density proportional to      β zE T fµ (y) · g ǫ E − X T (y − XβE ) + λ · det(XE XE + ǫI) (44) 0 u−E supported on D(E,zE ) , where fµ is the N (µ, σ 2 In ) density. As above, the same expression holds if our model for y|X is not from the normal family. A common variant of the LASSO implemented in Friedman et al. (2010) is minimize p β∈R ky − Xβk22 + λkβk1 . 2kX T yk∞ (45) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 23 A randomized version of this is easily handled, changing the density to be proportional to    T     XE XE X T (y − XE βE ) zE β + λ · det + ǫI fµ (y) · g ǫ E − u−E 0 kX T yk∞ kX T yk∞ and supported on the same set as in (42). Another variant Sabourin et al. (2014) replaces kX T yk∞ with a resampling based quantity median(X T y) resulting in a similar change in the sampling density. Remark 9. Conditioning on Nuisance Statistics: As detailed in Fithian et al. (2014), we condition not just on the selection event B(E,zE ) , but also on the sufficient statistics corresponding to nuisance parameters to obtain optimal UMPU selective tests/ intervals. In such a case, the support for y in the selective sampling density is restricted to a set, denoted as Dobs . This has been sketched for interested readers in A in the appendix. 4.2. The selective sampler is model agnostic: LASSO without parametric assumptions Up this point, we have assumed so far that X is fixed and the law of y|X was from the parametric model {N (µ, σ 2 I) : µ ∈ Rn }. We now remove this assumption, assuming that the law F is now just a law for the pair (X, y) ∈ Rn×p × Rn . A common assumption is the pairs model: IID (xi , yi ) ∼ F̃ for some distribution F̃ on Rp × R, where xTi are the rows of X. While we keep this model in mind, it is not necessary in order to define the appropriate sampler. This section serves as an example of the sense in which our main result Theorem 1 is agnostic to the underlying statistical model. In the notation established so far S = (X, y) ∈ Rn × Rn×p , 1 ℓ(β; (X, y)) = ky − Xβk22 , 2 P(β) = λkβk1 . Therefore, we must sample from a density proportional to      β zE T f (X, y) · g ǫ E − X T (y − XE βE ) + λ · det(XE XE + ǫI) 0 u−E supported on {(X, y, βE , u−E ) : diag(zE )βE > 0, ku−E k∞ ≤ 1} . (46) In order to use this result for particular applications of inference, one may have to reduce the problem to sampling from a particular reference distribution perhaps by conditioning on appropriate sufficient statistics, as described in Appendix A. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 24 4.3. Selective CLT: ℓ1 -penalized logistic regression In this section, we describe an application of the selective CLT of Tian and Taylor (2015), applied to the case of logistic regression with random design matrix X ∈ Rn×p with rows xTi , i = 1, . . . , n. Suppose iid xi ∼ FX , xi ∈ Rp , yi |xi ∼ Bernoulli(π(xTi b)), where π(x) = exp(x)/(1 + exp(x)) and b ∈ Rp is unknown, p fixed, S = (X, y). With logistic loss, # " n 1 X T T yi log π(xi β) + (1 − yi ) log(1 − π(xi β)) , ℓ(β; (X, y)) = − √ n i=1 we solve randomized ℓ1 -penalized logistic regression introduced in Tian and Taylor (2015), ǫ (47) minimize ℓ(β; (X, y)) − ω T β + kβk22 + λkβk1 . β∈Rp 2 On the usual selection event of observing active set and signs (E, zE ), the randomization reconstruction map is     1 T zE βE . φ(E,zE ) (y, βE , u−E ) = ǫ − √ X (y − π(XE βE )) + λ u−E 0 n And, the sampling density becomes proportional to      1 β zE f (X, y) · g ǫ E − √ X T (y − π(XE βE )) + λ 0 u−E n   1 T W (XE βE )XE + ǫI , · det √ XE n (48) where W (Xβ) = diag(π(Xβ)(1 − π(Xβ)))1 is the Hessian matrix of the loss and the density above is supported on the same set as in (46). Sampling (X, y) jointly is not feasible when FX is unknown. Denoting β̄E = argmin − β∈R|E| n X i=1 yi log π(xTE,i β) + (1 − yi ) log(1 − π(xTE,i β)), (49) the MLE for the unpenalized logistic regression with only the variables in E, a Taylor series expansion of ∇ℓ(β; (X, y)) gives2   √ Q(βE − tE ) ∇ℓ(β; (X, y)) ≈ n . −t−E + C(βE − tE ) 1 With slight abuse of notation, we allow π : Rn → Rn , x 7→ π(x) to be the function applied on each coordinate of x ∈ Rn . 2 Alternatively, we might take β̄ to be the one-step estimator in the selected model starting E from β̂E Taylor and Tibshirani (2016). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling Here, t is the observed value of the random vector   β̄E , T = 1 T n X−E (y − π(XE β̄E )) 25 (50) 1 T 1 T XE W (XE β̄E )XE and C = X−E W (XE β̄E )XE . n n Since Q and C converge by the law of large numbers to fixed quantities, we can either treat them as fixed or bootstrap them. For more details, see Tian and Taylor (2015). With p is fixed, pre-selected, T properly scaled is asymptotically normal    √ bE D → N (0, Σ), (51) n T− γ−E Q= where Σ is estimable from the data. When the selected parametric model is correct (E ⊇ supp b) we note that γ−E = 0 and Σ is in fact block-diagonal (Taylor and Tibshirani, 2016). When the selected parametric model is not correct, one can estimate this covariance nonparametrically, using bootstrap (see Section 7). Since asymptotically T is from an exponential family with parameters bE , we could base inference on bE by sampling from the distribution of T instead of (X, y). In this case, the parametrization map becomes        √ Q(βE − tE ) zE β , +λ ψ(E,zE ) (t, βE , u−E ) = t, ǫ E + n u−E −t−E + C(βE − tE ) 0       βE √ Q(βE − tE ) zE , n ,λ , 0 −t−E + C(βE − tE ) u−E (52) |E| with the domain Rp × RzE × [−1, 1]p−|E| . Hence, the density we sample from is proportional to        √ Q(βE − tE ) zE β , (53) +λ f (t) · g ǫ E + n u−E −t−E + C(βE − tE ) 0    bE restricted on the domain of ψ(E,zE ) , where f here is the density of N ,Σ . 0 We note that, in order to construct a valid test about some linear functional of bE we can always condition on the observed u−E and / or βE if we desire. The upside to such conditioning is that the sampling problem becomes easier, with the downside usually being a loss in selective power. 4.4. First step of forward stepwise The Kac-Rice tests, described in Taylor et al. (2013) are based on the solution of maximize η T X T y, where y ∼ F, (54) η∈K imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 26 X ∈ Rn×p is a fixed design, K a convex set that can be stratified into smooth disjoint manifolds and the process η T X T y is Morse for almost every y ∈ Rn . The simplest example of such tests is one step of forward stepwise model selection, in which case K is the ℓ1 ball of radius 1. In this case X will be usually centered and scaled so that X T y corresponds to the marginal Z statistics for p different simple linear regressions. Inference in this broad class of problems, for the global null, is based on the test statistic η ∗T X T y, where η ∗ = argmax η T X T y, (55) η∈K with K being the polar set of convex set C. Since, the null distribution of the above test statistic is intractable, we could instead provide inference based on the conditional law   L η ∗T X T y η ∗ = argmax η T X T y . η∈K With this brief description of inference based on Kac-Rice tests, we focus back on the randomized versions of the Kac-Rice objective in (54). The randomized objective is given by maximize η T (X T y + ω) − IK (η), where y × ω ∼ F × G, p η∈R (56) with ω ∈ Rp and penalty manifests as the characteristic function of convex set K, that is ( 0 if η ∈ K, IK (η) = ∞ otherwise. Here we set ǫ = 0, as the above optimization problem does always have a solution. Having ǫ > 0 would allow for several variables to be selected. As mentioned above, perhaps the simplest example of the randomized optimization problem in (56) is forward stepwise problem with a single step. Specifically, that is consider maximize η T (X T y + ω), η∈K for K = {η ∈ Rp : kηk1 ≤ 1}. The above optimization problem yields the optimal direction ( s∗ if j = j ∗ ∗ ηj = 0 otherwise, j = 1, . . . , p, where j ∗ = argmax |XjT y + wj |, 1≤j≤p imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 27 coordinate with the maximum absolute value, and s∗ = sign(XjT∗ y + wj ∗ ), the corresponding sign. Thus, we condition on the first active direction (both s∗ and j ∗ ), which gives rise to selection event  Ê(s∗ ,j ∗ ) = (y, ω) ∈ Rn × Rp : sign(XjT∗ y + ωj ∗ ) = s∗ , s∗ (XjT∗ y + ωj ∗ ) ≥ max |XjT y + ωj | . 1≤j≤p The subgradient equation yields the reconstruction map φ(j ∗ ,s∗ ) (y, z) = z − X T y, where z ∈ ∂IK (η ∗ ), the set of sub-gradients to K at η ∗ , given by the normal cone ∂IK (η ∗ ) = {c(u1 , . . . , uj ∗ −1 , s∗ , uj ∗ +1 , . . . , up ) : ui ∈ R, |ui | ≤ 1, c > 0}. We can reparametrize the set n o (y, ω, β, α, z) ∈ S̄0F (G, ǫ = 0, ℓ, P) : (y, ω) ∈ Ê(s∗ ,j ∗ ) using ψ(s∗ ,j ∗ ) (y, z) = (y, z − X T y, η ∗ , X T y, z), (ηj∗ = s∗ I{j=j ∗ } , j = 1, . . . , p), with the domain Rn × ∂IK (η ∗ ). With the above reparametrization, we sample (y, z) from a density proportional to f (y) · g(z − X T y), (57) supported on Rn × ∂IK (η ∗ ). Remark 10. The set of sub-gradients at η ∗ for s∗ > 0 can be identified as the epigraph of the ℓ∞ norm (modulo a permutation of the maximum coordinate j ∗ ) and for s∗ < 0, ∂IK (η ∗ ) is the polar cone of the epigraph of ℓ1 norm. Remark 11. Of course, it is usually of interest to take more than one step of forward stepwise. Inference after several steps of forward stepwise is considered in Tibshirani et al. (2014); Fithian et al. (2015). We consider several steps of forward stepwise in Section 6. Remark 12. The data analyst may have some set of variables Ē that she insists on controlling for. In this case, instead of just assuming X is centered, ⊥ we might assume that PĒ X = X and its columns normalized, where PĒ denotes projection onto the column space of XĒ . In this way, centering X corresponds to the common practice controlling for an intercept in the model. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 28 Remark 13. Forward stepwise is also used in generalized linear models such ⊥ as logistic regression. In this case, the quantity X T PE ′ y that appears in the least squares setting can be replaced either by the Wald Z-statistics or the score Z-statistics. For example, for logistic regression, the score Z-statistics take the form XjT (y − π(XE ′ β̄E ′ )) , 1≤j≤p (58) Zj = (XjT W (XE ′ β̄E ′ )Xj )1/2 where β̄E ′ is the unpenalized MLE for the model with variables E ′ and W (XE ′ β̄E ′ ) = diag(π(XE ′ β̄E ′ )(1 − π(XE ′ β̄E ′ )) is a consistent estimate of the variance of y under the model with variables E ′ . While this is the typical parametric estimate of variance forward stepwise would use, one might prefer using a jackknife or bootstrap estimate of this variance if one is unsure whether the model with variables E ′ is correctly specified, as would be the case early on in building a model via forward stepwise. Using Wald type Z statistics would require fitting p different logistic regression models which has some computational burden. Having computed the Zj ’s one might then consider a randomized version of the problem maximize η T Z η:kηk1 ≤1 and proceed as above with S being the vector Z. 4.5. Marginal screening Marginal screening computes marginal Z statistics Zj = XjT y σkXj k2 Tj = XjT y σ̂j kXj k2 or T statistics for each of p centered variables Xj and thresholds their absolute value at some threshold, perhaps z1−α/2 where α is some nominal p-value threshold. This can be expressed in optimization form as 1 minimize kη − T k22 . η:kηk∞ ≤c 2 Selective inference in the nonrandomized setting for this problem was considered in (Lee and Taylor, 2014). A natural randomized version would be 1 minimize kη − T k22 − ω T η. η:kηk∞ ≤c 2 imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 29 Conditioning on the set achieving the threshold c and their signs to be (E, zE ), we see that this event is {(T, η, z) : ηE = c · zE , diag(zE )zE ≥ 0, kη−E k∞ < c, z−E = 0} (59) and the randomization reconstruction map becomes     z c · zE −T + E . φ(E,zE ) (T, η−E , zE ) = 0 η−E We thus sample (y, X, η−E , zE ) (or (T, η−E , zE ) if X is random) from a selective density proportional to     z c · zE −T + E f (T ) · g (60) 0 η−E and supported on the event in (59), where f is the unselective law of T . For logistic regression, one can replace the T statistics above with the score statistics as described in Remark 13. 4.6. Full model screening Another possible way to screen variables is to threshold the coefficients from the output of an unpenalized, randomized selection program given by ǫ 1 ky − Xβk22 + kβk22 − ω T β, ((X, y), ω) ∼ F × G, minimize p β∈R 2 2 (61) Based on the solution β̂((X, y), ω), we define the threshold model (E, zE ) as E = {i : |β̂i ((X, y), ω)| ≥ aσ} and signs zE = sign(β̂E ((X, y), ω)), where a is a constant (e.g. 1 − α quantile of the standard normal or the t-distribution for some nominal level α) and σ is the scaling. The canonical event of interest is B(E,zE ) = {β ∈ Rp : |βi | ≥ aσ ∀i ∈ E, kβ−E k∞ < aσ, diag(zE )βE > 0}. A parametrization of {((X, y), ω, β, α, z) ∈ S̄0F (G, ǫ, ℓ, P) : β ∈ B(E,zE ) } is given by ψ(E,zE ) ((X, y), β) = (X, y), ǫβ − X T (y − Xβ), β, −X T (y − Xβ), 0 with the domain  D(E,zE ) = {((X, y), β) : diag(zE )βE > 0, |βi | ≥ aσ ∀i ∈ E, kβ−E k∞ < aσ} . imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 30 For the canonical event B(E,zE ) , we therefore need to sample ((X, y), β) from a density proportional to  f (X, y) · g ǫβ − X T (y − Xβ) · det(X T X + ǫI) (62) supported on D(E,zE ) , where f and g are the densities of F and G, respectively. If X is random, the sampling of the data simplifies as in Section 4.3. The scaling σ can be estimated as the variance of noise from the selected model with response y and matrix of predictors XE . As long as the estimate of σ is consistent, we can treat it as a constant in the selection event, hence have the selection event as the polyhedral region. 4.7. Selective sampler based on a penalized MLE: Fisher’s exact selective test In Section 2.1, we considered the pull-back of the unpenalized MLE for an exponential family. Allowing for penalization in the problem (19) as well as randomization yields programs of the form ǫ minimize Λ(β) − S T β + kβk22 − ω T β + P(β), (S, ω) ∼ F × G. β∈Rp 2 We describe a version of Fisher’s exact test based on observing some function of the solution to the above program. As we saw in the LASSO, the penalty P is often chosen such that the solution set {(β, z) : z ∈ ∂P(β)} possesses a nice stratification into a discrete collection of bundles C, the canonical example being the ℓ1 norm, or perhaps non-negative constraints. In particular, if P is the support function of some convex set K, then (4.7) is seen to be the normal bundle of K which itself often has a nice stratification. Fixing C to be one of these strata, the subgradient equations of the optimization problem yields a reconstruction map φC (s, β, z) = ∇Λ(β) − s + z + ǫ · β where (β, z) ∈ C and s ∈ supp(F ). In this case, the parametrization map of {(s, ω, β, α, z) ∈ S̄0 (G, ǫ, ℓ, P) : (β, z) ∈ C} becomes ψC (s, β, z) = (s, ∇Λ(β) − s + ǫβ + z, β, ∇Λ(β) − s, z) and the sampling density proportional to f (s) · g(∇Λ(β) − s + ǫβ + z) · |JφC (s, β, z)| and restricted to (s, β, z) ∈ M × C. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 31 We now describe a stylized instance of such a problem used in the context of goodness-of-fit tests. Much of the effort in selective inference has been focused on finding tools for inference about specific parameters in a model, i.e. inference about statistical functionals θ : M → R. Two examples in the literature that employ such goodness-of-fit tests are Choi et al. (2014); Fithian et al. (2015). In this case, we assume the data analyst will use the data to decide what sufficient statistics to use in constructing the goodness-of-fit test. Suppose our data S is a poisson process N with points in X and intensity measure modeled with density   p X dλ(β) βj hj (x) (x) = exp  dµ j=1 with respect to some reference probability measure µ on X . If X = [0, 1] and hj = 1[lj ,uj ] were a collection of indicator functions, this would correspond to a change-point model. We might then use something like the fused LASSO in which we can take the intervals to be the step functions [lj , uj ] = [(j − 1)/p, 1] or perhaps a multiscale approach as in Frick et al. (2014); Chan and Walther (2011); Rivera and Walther (2013) and use the penalty P(β) = p X j=1 λj |βj | a weighted ℓ1 penalty3 . Usually, the background rate is included in such a model by fixing h1 to be the constant function 1. An unpenalized fit for this variable corresponds to its corresponding λ being 0. Depending on our choice of penalty, we might then solve the following randomized program  Z p  X ǫ β minimize Λ(β) − h (x) N (dx) + P(β) + kβk22 − ω T β j j p β∈R 2 X j=1 where  Λ(β) = log  Z X   exp  p X j=1    βj hj (x) − 1 µ(dx) . Having solved the above problem, the data analyst now observes that (β̂(N, ω), ẑ(N, ω)) are in some subbundle C of the set of variable-subgradient pairs. When P is a weighted ℓ1 penalty, the conditioning event n o (N, ω) : Ê(N, ω) = E, zÊ(N,ω) = zE where n o Ê(N, ω) = j : β̂j (N, ω) 6= 0, λj > 0 3 The multiscale approach described in Frick et al. (2014); Chan and Walther (2011); Rivera and Walther (2013) is formally a testing approach. The penalized version above was proposed in Tian et al. (2015) based on the dual of the test statistic in the multiscale literature. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 32 are the selected penalized coordinates. The corresponding subbundle is the set  C = (βE , βU , z−E ) : diag(zE )βE ≥ 0, kz−E k∞ ≤ 1, βU ∈ RU where U = {j : λj = 0} is the set of unpenalized coordinates. With some abuse of notation, we write z = (zE , zU , z−E ) = (zE , 0, z−E ) as the full subgradient where zE are held fixed on C and zU = 0 as these coordinates are unpenalized. The data analyst now decides to test the null hypothesis H0,j|E : βj|E = 0, in the model log λ(β) = X βj Z j 6= 1 hj (x) µ(dx). X j∈E Following the exponential family setup in Fithian et al. (2014), we might condition on the nuisance sufficient statistics Z hi (x)N (dx), i ∈ E \ j. X Assuming we have included a background rate in the model, this fixes the total number of points in the Poisson process to be N (X ) the observed number of points. If we had not used the data to select the intervals, conditioning on these sufficient statistics and resampling points is exactly Fisher’s exact test modulo the choice of test statistic. That is, the appropriate reference measure can be constructed by sampling from a Binomial process Ñ with n = N (X ) points and distribution µ conditioned to satisfy Z Z hi (x) N (dx), i∈E\j hi (x) Ñ (dx) = X X where the right hand side are the values observed in the data. When X is a discrete space, this is a generalization of Fisher’s exact test. Sampling for such problems have a fairly rich literature (c.f. (Diaconis and Sturmfels, 1998)). When X is Euclidean or a manifold, the conditioning event above corresponds to a subset of the configurations of N (X ) points on X and is generally a nontrivial task. See (Diaconis et al., 2013) for further discussion and examples. Having used the data to choose which sufficient statistics to use, we must use an appropriate selective distribution. Beyond just sampling Ñ from the conditional density, we must sample (βE , βU , z−E ). Conditional on Ñ we see that the appropriate density of the joint law with respect to the product of the law of Ñ and p-dimensional Hausdorff measure on C is proportional to   Z (Ñ , βE , βU , z) 7→ g ∇Λ(β) − h(x)dÑ (x) + ǫ · β + diag(λ)z X with ∇Λ(β) = R  P βi hi (x) µ(dx)  P i∈E∪U exp i∈E∪U βi hi (x) µ(dx) XRh(x) exp X imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 33 For computational reasons, to evaluate the integrals over X above, it may be simpler to use a discretization of X as in Lindsey’s method Efron and Tibshirani (1996). Finally, while we have considered using a weighted LASSO to choose the sufficient statistics, one might use a penalty with some curvature as well, requiring the modifications discussed in Section 5. 4.8. Graphical models Gaussian graphical models are a popular way to study network structures. In particular, it has often been used on many types of genome data, e.g. gene expression, metabolite concentrations, etc. Specifically, consider the p-dimensional normally distributed random variable X = (x1 , . . . , xp ) ∼ N (µ, Σ). It is of interest to study the conditional independence structure of the variables {1, 2, . . . , p}. The conditional independence structure is conveniently represented by an undirectional graph (Γ, E), where the nodes are Γ = {1, 2, . . . , p}, and there is an edge (i, j) ∈ E if and only if xi 6⊥ xj conditional on all the other variables Γ\{i, j}. Moreover, assuming the covariance matrix Σ is not singular, we denote the inverse covariance matrix Θ = Σ−1 , then xi ⊥ xj |XΓ\{i,j} ⇐⇒ Θij = 0. In many applications of Gaussian graphical models, we assume the sparse edge structure, where we can hope to uncover the network structure even in the highdimensional setting. We discuss applying the selective sampler to the graphical LASSO Friedman et al. (2008). Another graphical model selection, neighborhood selection, is discussed in Appendix C. The randomized graphical LASSO with randomization ω ∼ G is the convex program minimize − log det(Θ) + Tr((S − ω)Θ) + λP(Θ), Θ:ΘT =Θ,Θ>0 where P(Θ) = X (i,j):i6=j (S, ω) ∼ F × G, (63) |Θ|ij and Usually,  S ∈ Mk = A : AT = A, A ≥ 0, rank(A) = k , S= 1 T X n   1 I − 11T X, n k ≤ p. X ∈ Rn×p for some data matrix so that k is generically min(n − 1, p). Note that Θ, S, ω are p × p matrices. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 34 The subgradient equations read as ω = Θ̂−1 (S, ω) + S + ẑ(S, ω), where ẑ(S, ω) ∈ ∂P(Θ̂(S, ω)). For E ⊂ {(i, j) : i 6= j}, the natural selection event of interest is n o S = (S, ω) : ∃ Θ = Θ̂(S, ω) s.t. sign(ΘE ) = zE , Θ−(E∪{(i,i):1≤i≤p}) = 0 , Owing to symmetry in Θ, on S we have  2λsij ∂P(Θ̂(S, ω))  ẑij (S, ω) = = 0  ∂Θij  2λuij (i, j) ∈ E i=j (i, j) ∈ / E and i 6= j, where and uij = uji satisfies kuk∞ ≤ 1 for (i, j) ∈ / E, i 6= j, sij = sign(Θij ). Due to symmetry, we can parametrize ω by the active elements in Θ along with its diagonal elements (dimension |E|/2 + p) and the inactive elements of the penalty subgradient (dimension (p2 − p − |E|)/2), corresponding to the inactive elements, both restricted to the upper triangular part of Θ. To fix notations, denote E ∗ = {(i, j) ∈ E : i ≤ j} ∪ {(i, j) : i = j}, the set of active coordinates in the upper triangular part of Θ along with its diagonal elements. This leads us to the following parameterization of {(S, ω, Θ, α, z) ∈ S̄0 (G, ǫ = 0, ℓ, P) : (S, ω) ∈ S} ψ(E,zE ) (S, ΘE ∗ , u−E ∗ ) = (S, Θ−1 (E ∗ ) + S + z(u−E ∗ ), Θ(E ∗ ), Θ−1 (E ∗ ) + S, z(u−E ∗ )), (64) where Θ(E ∗ ) is the matrix with E ∗ entries from ΘE ∗ and zeroes elsewhere, zij (u) = 2λ(sij I{(i,j)∈E} + ui,j I{(i,j)6∈E,i6=j} ) and the domain of φ(E,zE ) is restricted to ku−E ∗ k∞ ≤ 1. The randomization reconstruction map can be written as φ(E,zE ) (S, ΘE ∗ , u−E ∗ ) = Θ−1 (E ∗ ) + S + z(u−E ∗ ). With this parametrization, we can sample from the density proportional to f (S) · g(Θ−1 (E ∗ ) + S + z(u−E ∗ )) · |Jψ(S, ΘE ∗ , u−E ∗ )| . and restricted to ku−E ∗ k∞ ≤ 1. The computation of above Jacobian involves (see the Appendix of Taylor and Tibshirani (2016)) ! ∂φ(E,zE ) = 2(σil σjk + σik σjl ) for (i, j) ∈ E ∗ , ∂Θij kl imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling ∂φ(E,zE ) ∂uij ! kl 35 = 2λδik δlj for (i, j) ∈ / E ∗ , i ≤ j, where σij are the elements of Σ and δij is Kronecker delta function. The above computations give the (k, l) coordinate of the (i, j)th upper triangular matrix and as again, we simply restrict attention to {(k, l) : k ≤ l}, the upper triangular half of randomization map. More polyhedral examples, as well as the generalized LASSO are presented in Appendix B. 5. Curvature Up to now, our examples have been polyhedral in nature. In most examples above, the selection event was some part of the normal bundle of such a polyhedral convex set. Not all problems can be described by such polyhedral sets. In this section we give some examples with non-trivial curvature. As above, many of the interesting selection events are still parts of the normal bundle. 5.1. Projection onto a convex set The problem of projection onto a closed convex set K is canonical example of a convex program. Hence, it seems to be a natural place to start to introduce curvature into our discussion. (Indeed two variants of the LASSO above can be expressed in terms of projection onto a convex set.) Let 1 kβ − sk22 2 ( 0 β∈K P(β) = ∞ otherwise. ℓ(β; s) = In this case, with ǫ = 0 the randomized program takes the form 1 minimize kβ − sk22 − ω T β. β∈K 2 with solution β̂(s, ω) = PK (s + ω), where PK is the Euclidean metric projection onto the set K (i.e. the projection with metric induced by the ℓ2 norm). The subgradients to our penalty are ∂P(β) = Nβ (K), the normal cone of K at β. In this case, the subgradient equations take the form s + ω = β + z, β ∈ K, z ∈ Nβ (K). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 36 The normal cone Nβ (K) can be parameterized by its unit vectors S(Nβ (K)) = {u ∈ Nβ (K) : kuk2 = 1} and their length r ∈ [0, ∞). We can therefore draw from the joint distribution of (s, ω) with a density proportional to f (s) · g(β + r · u − s) · |Jψ (β, r, u)|, where4 ψ(s, β, r, z) = (s, β + r · u − s, β, β − s, r · u) β ∈ K, u ∈ S(Nβ (K)), r ≥ 0. Integrating out s yields the density (f ∗ g)(β + r · u) · |Jψ (β, r, u)|. where f ∗ g is the density of the convolution F ∗ G. The Jacobian here is recognizable as that in the Weyl-Steiner volume of tubes formula Weyl (1939); Hotelling (1939); Johnstone and Siegmund (1989); Takemura and Kuriki (2002); Sun (1993); Adler and Taylor (2007) Jψ (β, r, u) = det (I + r · C−u ) = det (I + C−r·u ) , (65) where Cη is the curvature matrix of ∂K at β in the direction η normal to ∂K. Suppose now K is compact and that F ∗ G is the uniform distribution on K ⊕ δB2 . This can be achieved by taking F = δ0 , say and G the uniform distribution itself. In this case, supp(G) = K ⊕ δB2 and we arrive at WeylSteiner’s tube formula Z 1 Jψ (β, r, u) dβ dr du 1= Hp (K ⊕ δB2 ) (β,r,u):β+r·u∈K⊕δB2  Z δ Z = Jψ (β, r, u) dβ du dr. 0 The formula above must be interpreted as a sum over pieces or strata of K of different dimensions so that dβ properly refers to Hausdorff measure of differing dimensions and the matrix in (65) is also of differing dimensions. We refer the readers to Takemura and Kuriki (2002); Adler and Taylor (2007) for further details and points of entry into the volume-of-tubes literature. When K is a smooth body, i.e. its boundary is a differentiable (p − 1)-dimensional hypersurface then there is only stratum and the tube formula reads Z δZ Hp (K ⊕ δB2 ) = Jψ (β, r, ηβ ) dβ dr, 0 ∂K where ηβ is the outward pointing unit normal vector field. 4 This parameterization is slightly different than our usual one, in that it constructs s + ω instead of ω, but the difference is not important. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 37 When F ∗ G is not the uniform distribution on K ⊕ δB2 we might instead try to compute Z (f ∗ g)(β + r · u) · Jψ (β, r, u) dβ dr du (F ∗ G)(K ⊕ δB2 ) = (β,r,u):β+r·u∈K⊕δB2 = Z δ 0 Z  (f ∗ g)(β + r · u) · Jψ (β, r, u) dβ du dr. This problem is considered in some generality when f ∗ g is a smooth density in Chapter 10 of Adler and Taylor (2007). When F ∗ G = N (0, Ip ), their probability, expanded in a power series in δ plays an important role in the Gaussian Kinematic Formula Taylor (2006); Taylor and Adler (2009). Unlike in the Weyl-Steiner formula, while the first steps are similar and are what Weyl said was something any student of calculus could do Weyl (1939), the above probability are not Riemannian invariants of K but depend on how the law F ∗ G relates to K. While the connections to curvature measures and volume-of-tubes formulae are enlightening, we feel concrete examples are also very important. In this section, we consider the group LASSO as a canonical example of a statistical learning problem with curvature. 5.2. Forward stepwise with groups of variables: Kac-Rice with groups This example is a generalization of the first-step of forward stepwise when we allow groups of variables to enter, and is a second example of the Kac-Rice test described above. Non-randomized approaches to this problem are considered in Loftus and Taylor (2014); Yang et al. (2016). The randomized version of the Kac-Rice objective with a group LASSO penalty determined by a partition {1, 2, . . . , p} = ∪g∈G g can be written as maximize η T (X T y + ω) − IK (η), y × ω ∈ F × H p η∈R where K = {η ∈ Rp : X g∈G (66) λg kηg k2 ≤ 1} and IK is the usual characteristic function of set K. In this section we have denoted the law of ω by H with density h to distinguish the density from the group index g. In this case, the optimal η = η ∗ is given by  T   Xg y + ω g if g = g ∗ ηg∗ = λg kXgT y + ωg k2  0 otherwise, where g ∗ = argmax g∈G 1 kXgT y + ωg k2 . λg imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 38 Conditioning on the selection event   Êg∗ = (y, ω) : g ∗ = argmaxkXgT y + ωg k2 /λg , g∈G the subgradient equation leads to the reconstruction map   zg ∗ T ∗ ∗ ∗ ω = φg (y, c, zg , (zg )g6=g ) = −X y + c , (zg )g6=g∗ subject to constraints c > 0, kzg∗ k2 = λg∗ , kzg k2 ≤ λg for g 6= g ∗ . Thus we sample (y, c, zg∗ , (zg )g6=g∗ ) from a density proportional to f (y) · h(cz − X T y) · c|g ∗ |−1 and supported on ∗ supp(F ) × R+ × λg∗ S(R|g | ) × Y λg B2 (R|g| ), g∈G,g6=g∗ ∗ ∗ where λg∗ S(R|g | ) is a sphere with radius λg∗ in R|g | and λg B2 (R|g| ) is an ℓ2 ball in R|g| . The Jacobian here, is the determinant of the derivative of the map  zg∗ c 7→ −XgT∗ y + czg∗ ,  ∗ ∗ ∗ which equals det cVg∗ zg∗ = λg∗ c|g |−1 , where Vg∗ ∈ R|g | × R|g |−1 be an orthonormal basis completion of zg∗ . 5.3. Group LASSO As mentioned above, the LASSO (23) can be expressed in terms of metric projection, as can many problems when ℓ is squared-error loss and P is either a (semi-)norm in or a constraint on a seminorm. When the seminorm is polyhedral the Jacobian in our parameterization typically depends only on the design matrix X. If the unit ball of the seminorm is not polyhedral, the Jacobian of the parameterization involves a curvature term which depends on the design matrix X. We consider the group LASSO Ming and Lin (2005). The group LASSO norm, defined by a partition ∪g∈G g = {1, . . . , p} and weights (λg )g∈G is defined as P(β) = PG,(λg )g∈G (β) = X g∈G λg kβg k2 , (67) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling where βg = β[g] are the coefficients in the g-th group. The group LASSO problem is defined by X 1 2 ky − Xβk + λg kβg k2 . minimize 2 β∈Rp 2 39 (68) g∈G Losses other than squared-error can of course be used with this penalty – it is the penalty that is the group LASSO. The square-root group LASSO Bunea et al. (2013) replaces the squared-error loss above with the ℓ2 loss ky − Xβk2 . The randomized version of the group LASSO is X 1 ǫ minimize ky − Xβk22 − ω T β + kβk22 + λg kβg k2 (69) p β∈R 2 2 g∈G where ω ∼ H with density h. Having solved (69) with solution β̂(X, y, ω), we define the set of active groups as n o Ê(X, y, ω) = g : β̂g (X, y, ω) 6= 0 . The canonical event of interest here is n o (X, y, ω) : Ê(X, y, ω) = E . (70) On this event, the subgradient equation gives the randomization reconstruction map ω = φE (y, (βg )g∈E , (zh )h∈−E )     (λg βg /kβg k2 )g∈E (βg )g∈E , − XT y + = (X T X + ǫI) (zh )h∈−E 0 (71) with the restriction on the inactive subgradients as kzh k2 ≤ λh , h ∈ −E. We may choose to parameterize {(y, ω, β, α, z) ∈ S̄0F (G, ǫ, ℓ, P) : βh = 0 ∀h ∈ −E} in terms of (y, (βg )g∈E , (zh )h∈−E ) as ψE (y, (βg )g∈E , (zh )h∈−E )        (βg )g∈E (βg )g∈E (λg βg /kβg k2 )g∈E = y, (X T X + ǫI) , − XT y + , 0 0 (zh )h∈−E      (βg )g∈E (λg βg /kβg k2 )g∈E − XT y − XT , 0 (zh )h∈−E with the restriction on the inactive subgradient Y (zh )h∈−E ∈ λh B2 (R|h| ), (72) h∈−E imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 40 where B2 (R|h| ) is the ℓ2 ball in R|h| . The density to sample from thus, is proportional to f (y) · g(φE (y, (βg )g∈E , (zh )h∈−E )) · D((βg )g∈E ,(zh )h∈−E ) φE , (73) M = ∪g∈E g, the set of active predictors, and their restricted to (72). Denoting P cardinality as |M | = g∈E |g|, where |g| is size of each group, the non-trivial Jacobian term is precisely given by D((βg )g∈E ,(zh )h∈−E ) φE  T XM XM + ǫI + D1 = det T X−M XM 0|M|×(p−|M|) Ip−|M| T = det(XM XM + ǫI + D1 ),  (74) where D1 is a block diagonal matrices of the form   !! βg βgT λg  ∈ R|M|×|M| . D1 = diag  I− kβg k2 kβg k22 g∈E Remark 14. For groups that are singletons (|g| = 1 for all g ∈ G), the matrix D1 is zero as βg /kβg k2 ∈ {1, −1}. 5.4. Conditioning on active directions in group LASSO If we choose to condition additionally on the active directions ( ) βg ug = , g∈E , kβg k2 writing βg = γg · ug , γg > 0, g ∈ E; kzh k2 ≤ λh , h ∈ −E, we may express the reconstruction in (71) in terms of (y, (γg )g∈E , (zh )h∈−E ) as φ(E,(ug )g∈E ) (y, (γg )g∈E , (zh )h∈−E )     (λg ug )g∈E (γg ug )g∈E T T −X y+ = (X X + ǫI) (zh )h∈−E 0 restricted to (y, (γg )g∈E , (zh )h∈−E ) ∈ Rn × (0, ∞)|E| × Y λh B2 (R|h| ). h∈−E In this case the density we sample from is proportional to f (y) · g(φ(E,(ug )g∈E ) (y, (γg )g∈E , (zh )h∈−E )) · D((γg )g∈E ,(zh )h∈−E ) φ(E,(ug )g∈E ) , (75) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 41 restricted to support above. The Jacobian of the map in this case can be computed as polynomial in (γg )g∈E by taking a geometric approach outlined below. Hence, conditioning further on the active directions allows us to sample from a log-concave density. To see an explicit computation of the Jacobian, let Vg ∈ R|g|×(|g|−1) , g ∈ E, denote a matrix whose columns are orthonormal and orthogonal to ug and set V = diag ((Vg )g∈E ) ∈ R|M|×(|M|−|E|) , U = diag((ug )g∈E ) ∈ R|M|×|E| . Finally set Γ = diag((γg I|g| )g∈E ) ∈ R|M|×|M| . The Jacobian is the derivative of the map (U, γ) 7→ ∇ℓ̃(ΓU ) + ΛU, where T ∇ℓ̃(ΓU ) = (XM XM + ǫI|M| )ΓU, and Λ = diag((λg Ig )g∈E ) ∈ R|M|×|M| . Differentiating first with respect to V (i.e. tangent to U ) and then γ, the derivative can be written as     (∇2 ℓ̃(ΓU )Γ + Λ)V ∇2 ℓ(ΓU )U = ∇2 ℓ̃(ΓU )Γ + ΛV V T V Γ−1 U . T Since ℓ̃ is quadratic, Q = ∇2 ℓ(ΓU ) = XM XM + ǫI|M| does not involve ΓU , the determinant of the above expression is equivalent to   T   5 V −1 (QΓ + Λ)V QU . det Q UT Letting Γ− = diag((γg I|g|−1 )g∈E ), Γ̄ =  − Γ 0 the Jacobian reduces to computing  T −1   ΛV −1 V Q det Γ̄ I|M| + Γ̄ Z T Q−1 ΛV  = det Γ− + V T Q−1 ΛV . 0 I|E| 0 0   , (76) Conditioning on U is now straightforward as V will then be fixed in the above density. 5 Since V U T we condition on the active directions, U and V are fixed matrices hence Q−1 is fixed as well. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 42 5.5. General geometric approach More generally, one can take a geometric approach to obtain the Jacobian, even when we do not condition on the active directions. For the group LASSO, in computing the Jacobian it is convenient to assume that ℓ is somewhat arbitrary but twice-differentiable and to introduce a new Riemannian metric on the fiber of optimization variables ((γg )g∈E , (zg )g∈E ) over the data point s. The convenience is that the structure of the Jacobian is similar for many problems in which the penalty is a seminorm appearing in Lagrange form. We expect a similar formula to hold when the penalty is a seminorm appearing in bound form. In the interest of space, we do not pursue this here. The Riemmanian metric we use is the pull-back of the following metric on Rp : −1 Zβ . (77) hVβ , Zβ i(s,β) = VβT ∇2 ℓ(β; s) + ǫI With ǫ = 0 and ℓ the log-likelihood of an exponential family, in local coordinates the above metric is inverse of the observed information metric Efron and Hinkley (1978) evaluated at the point estimate β. It is perturbed by ǫ times the Euclidean metric above as we have modified our convex program in randomizing it. For the group LASSO, we then now differentiate the parameterization in coordinates γE , (zg )g∈G . The main reason for introducing this metric is that the image tangent vectors in the face   ! Y Y |g| |h| F = λg · S(R ) × λh B2 (R ) g∈E h∈−E P remain orthogonal to the normal vectors g∈E γg zg . The next theorem derives the Jacobian when curvature component is present in the geometry. Theorem 2 (Jacobian meta-theorem). Consider the randomized version of the problem minimize ℓ(β; S) + kβk p β∈R for some F -a.s. twice-differentiable convex ℓ and some seminorm kβk = sup β T z. z∈K Let • F be a smooth face of K; • (ηi )i∈I be an orthonormal (in the Euclidean metric) frame field normal to F; • (Vj )j∈A be an orthonormal (in the Euclidean metric) frame field tangent to F , imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 43 and let ψ denote the parameterization ψ(s, z, γ) = ∇ℓ(β(z, γ); s) + ǫ · β(z, γ) + z P restricted to N (F ) with β(z, γ) = i∈I γi ηi . Then,   ∂ ψ = δil ηi , ∂γl (s,β)   ∂ Vj , ψ =0 ∂γi (s,β) (78) hVj , Vk ψi(s,β) = S−β (Vj , Vk ) + hVj , Vk i(s,β) , where Sη is the shape operator of F as it sits in Rp . Finally, Jψ(s, z, γ) = det(∇2 ℓ(β; s) + ǫ · I) · det(G(s, z, γ) + C−β (z, γ)) = det(I + G(s, z, γ)−1 C−β (z, γ))/ det(H(s, z, γ)), where (79) G(s, z, γ)jk = hVj,z , Vk,z i(s,β) ⊥ ⊥ ηl,z i(s,β) , ηi,z , P(s,β) H(s, z, γ)il = hP(s,β) ⊥ where P(s,β) is projection orthogonal to the tangent space of F in the metric (77) and C−β (z, γ) is the Euclidean curvature matrix of F in the basis Vi,z , 1 ≤ i ≤ dim(F ). Proof. The reconstruction map reads ω = ∇ℓ P X γi ηi ; s i ! +ǫ X γi ηi i ! +z γi ηi is normal to K at z ∈ F . We now compute ! ! X ∂ 2 γi ηi ; s + ǫ · I ηi ψ(s, z, γ) = ∇ ℓ ∂γi i ! ! ! X X 2 γi ηi + Vj . γi ηi ; s + ǫ · I Vj Vj ψ(s, z, γ) = ∇ ℓ with the constraint i i i This is enough to establish (78). Now, the determinant we want to compute is the determinant of the matrix   ∂ ψ V ψ ∂γ where V ψ is the matrix whose columns are Vj ψ, and ∂ columns are ∂γ ψ. i ∂ ∂γ ψ is the matrix whose imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 44 P Factoring out ∇2 ℓ ( i γi ηi ; s) + ǫI and then multiplying by the orthogonal matrix  T V γT shows that the determinant we want to compute is det(∇2 ℓ(β; s) + ǫI) times the determinant of the matrix   S−β (Vj , Vk ) + hVj , Vk i(s,β) 0 . ... I Relation (79) now follows, with the second display simply a familiar formula for the determinant expressed in block form det(∇2 ℓ(β; s) + ǫ · I)−1 = det(G(s, z, γ)) · det(H(s, z, γ)). When the loss is quadratic, or we use a quadratic approximation for the loss, we see that if we condition on z, then the term det(H(s, z, γ)) is constant and only the first term must be computed. Remark 15. We note that while F may not be convex, when ℓ is quadratic (or a quadratic approximation is used for inference) and we condition on z the distribution for inference is log-concave in γ. More generally, if f is logconcave in s as well, then the relevant distribution of (s, γ)|z is jointly logconcave in (s, γ). This general theorem can be applied to compute the Jacobian of the group LASSO with a parametrization in terms of the active and inactive directions and the magnitudes of active coefficients. We refer the readers to Appendix D for a rederivation of this Jacobian using the above geometric perspective. 6. Multiple views of the data Often, an analyst might try fitting several models to a data set. These models might have different numbers of parameters, and different objective functions. Examples include fitting a regularization path Friedman et al. (2010) or stability selection Meinshausen and Bühlmann (2010). Nevertheless, if each model is an instance of (33), then there is a straightforward procedure to construct a selective sampler. Given a collection (Gi , ǫi , ℓi , Pi )i∈I , the appropriate density is proportional to ! Y f (s) · gi (ǫi βi + αi + zi ) · Jψi (s, βi , αi , zi ) (80) i∈I supported on some subset of G s∈supp(F ) Y i∈I pi ! {(βi , αi , zi ) : βi ∈ R , αi ∈ ∂ℓi (βi ; s), zi ∈ ∂Pi (βi )} . imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 45 This set has the form of a bundle with base space supp(F ) and fibers as described above. Typically, each ψi will be restricted to some canonical event Bi . For example, in something similar to stability selection, each Bi might identify the set of variables and signs chosen by the i-th randomized LASSO program. We see that given s, the variables in the fiber are independent. Hence, these can be sampled in parallel. For instance, on different machines initiated with the same random seed, we can sample s IID from density f , then run Gibbs or other samplers to sample (βi , αi , zi ). 6.1. Multiple steps of forward stepwise As an example, we consider taking K steps of forward stepwise. The Kac-Rice test can be extended to K steps of forward stepwise, where the selection event is characterized by a sequence of indices with the corresponding signs that constitute the active set at step K. To make it more explicit, say we consider a sequence of active variables (j1 , . . . , jK ) with corresponding signs (s1 , s2 , . . . , sK ), in the order in which the randomized variables enter the model. The k-th optimization in the set of K optimizations can be written as T ⊥ maximize η T (X−A PA y + ωk ) − IKk (η), where y × ωk ∼ F × Gk , (81) k−1 k−1 η∈Rp−k+1 Ak = {j1 , j2 , . . . , jk } is the active set including the k-th step, X−Ak are the columns of X except for the ones corresponding to the current active set Ak , the characteristic function ( 0 if η ∈ Kk IKk (η) = ∞ otherwise and Kk = {η ∈ Rp−k+1 : kηk1 ≤ 1}. Here, {ωk }K k=1 is a sequence of independent randomization variables with ωk coming from a given distribution Gk in Rp−k+1 and corresponding density gk . ⊥ The projection PAk is onto XAk and PA is the residual after this projection. k The selection event after K steps can be written as Ê{(sk ,jk )}K = k=1  K Y  n ⊥ ∈ R × Rp−k+1 : sign(XjTk PA y + ωk,jk ) = sk , y, {ωk }K k=1 k−1 k=1 ⊥ sk (XjTk PA y k−1 + ωk,jk ) ≥ max c j∈Ak−1 ⊥ |XjT PA y k−1  + ωk,j |, k = 1, . . . , K . The randomization reconstruction map for the k-th step, from the subgradient equation is given by T ⊥ φk (y, zk ) = −X−A PA y + zk , k−1 k−1 imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 46 where, sub-differential zk ∈ Rp−k+1 from the k-th step is restricted to zk ∈ ∂IKk (ηk∗ ), and ηk∗ ∈ Rn−k+1 is the optimal solution for the k-th optimization, as stated in (81). More explicitly, the normal cone is given by ∂IKk (ηk∗ ) = {c · u : u ∈ Rp−k+1 , ujk = sk , |uj | ≤ 1 ∀j ∈ Ack−1 , c > 0}. (82) The sampler density is thus proportional to f (y) · K Y k=1   T ⊥ P y , gk zk − X−A Ak−1 k−1 (83) supported on (y, z1 , . . . , zk ) ∈ Rn × K Y ∂IKk (ηk∗ ). k=1 For logistic regression, one can replace the T statistics above with the score statistics as described in Remark 13. Further examples of algorithms that choose a variable based on multiple views of the data are presented in Appendix E. 7. Selective sampling via projected Langevin In the examples above, we sample from a joint density of data S ∈ Rn and optimization variables T ∈ Rp , from a probability space on (S, T ), achieved via a reparametrization. The selective sampler now samples from a pull back measure on a transformed probability space under a reconstruction map for randomization ω. This idea implemented on randomized convex programs typically allow us to sample from a joint density that is supported on a relatively simpler region. On a canonical selection event B(E,zE ) , defined by the active variables E and their signs/directions zE , the joint sampling density is h(s, t) ∝ f (s) · g(φ(E,zE ) (s, t)) · Dt φ(E,zE ) (s, t) · 1D(E,zE ) (s, t), (84) supported on constraint set D(E,zE ) ⊂ Rn × Rp , under reconstruction map φ(E,zE ) . When we consider f (s) ∝ exp(−f˜(s)) and g(ω) ∝ exp(−g̃(ω)), the negative of logarithm of the sampling density is proportional to h̃(s, φ(E,zE ) (s, t)) = f˜(s)) + g̃(φ(E,zE ) (s, t)) − log Dt φ(E,SE ) (s, t) (85) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 47 supported on set D(E,zE ) . We sample from the target density h̃ using updates from a projected Langevin random walk. Projected Langevin Updates Iterative update [k+1]: The (k + 1) update based on previous update (s(k) , t(k) ) is given by ! !!  (k+1)   (k)  (k) p ∇s h̃(s(k) , φ(E,zE ) (s(k) , t(k) )) s s ξ1 + 2η =P −η (k) t(k+1) t(k) ∇t h̃(s(k) , φ(E,zE ) (s(k) , t(k) )) ξ2 (k) (k) for step-size η, ξ1 × ξ2 ∼ N (0, In ) × N (0, Ip ) and P is projection onto set D(E,zE ) . Computing the (k + 1) update based on previous update (s(k) , t(k) ) thus involves two steps: 1: Computing gradient of the negative of log density w.r.t. (s, t), that is ! ∇s h̃(s(k) , φ(E,zE ) (s(k) , t(k) )) ∇t h̃(s(k) , φ(E,zE ) (s(k) , t(k) )) 2: Compute projection of update from a noisy version of gradient descent of the log density ! !  (k)  (k) p ∇s h̃(s(k) , φ(E,zE ) (s(k) , t(k) )) s ξ1 + − η 2η (k) t(k) ∇t h̃(s(k) , φ(E,zE ) (s(k) , t(k) )) ξ2 onto constraint set D(E,zE ) . When h̃(s, φ(E,zE ) (s, t)) is a convex function on restriction D(E,zE ) and satisfies smoothness properties ∇h̃(s1 , ω1 ) − ∇h̃(s2 , ω2 ) ≤ β |(s1 , ω1 ) − (s2 , ω2 )| ∇h̃(s1 , ω1 ) ≤ C, for all (s1 , ω1 ), (s2 , ω2 ) ∈ D(E,zE ) , and the support D(E,zE ) is convex with nonempty interior and contained in a finite Euclidean ball, the projected Langevin sampler indeed converges to the target density in (84) as proved in Bubeck et al. (2015). Remark 16. For most examples, data S is sampled from a Gaussian density, in which case, ∇f˜ is unbounded. One should be able to remove this condition by considering a restriction of f˜ to a bounded set of probability close to 1 under Gaussian density f . Remark 17. The choices for randomizations typically include a Gaussian, Laplacian or Logistic distribution.6 Under a Logistic density, g̃ is seen to satisfy the smoothness conditions. Qualitatively, our samples do not change much with any of these choices of randomization. 6 Tian and Taylor (2015) use heavy-tailed distributions such as Laplace or logistic as choices for randomization. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 48 Remark 18. Typically, the support set D(E,zE ) is convex, but not bounded. We could again remove the boundedness condition on D(E,zE ) , by considering a compact subset of probability close to 1 under sampling density h. 7.1. Examples We revisit few examples like the LASSO with fixed and random design, the KacRice with forward stepwise, the ℓ1 -penalized Logistic to implement the projected Langevin sampler. We offer inference on the parameters in the selected model E conditional on the canonical selection event B(E,zE ) of selecting active set E and their signs/ directions, obtained upon solving the corresponding randomized programs. We base our tests on the randomized pivot, developed in Tian and Taylor (2015) for selective inference in a randomized setting 7 . Example 7.1. LASSO with fixed design (Section 4.1): The response y is modeled as N (Xb, σ 2 ) with |supp(b)| = s and σ 2 = 1. We test the null hypothesis H0 : bE,j = 0, bE,j denoting the j-th coefficient in the selected model E, conditioning on the selection event B(E,zE ) and sufficient statistics PE\j y, corresponding to nuisance parameters as discussed in A. Thus, the sampling density is proportional to (44) on the support {(y, βE , u−E ) : PE\j y = Yobs , diag(zE )βE > 0, ku−E k∞ ≤ 1}. The updates from projected Langevin sampler can be written as  √   (k)  (k) (k) (k) y − ηy (k) /σ 2 − η∇y g̃(φ(E,zE ) (y (k) , βE , u−E )) + 2ηξ1 ỹ (k+1) √  (k) (k) (k) (k)  (k+1)   , P1 βE − η∇βE g̃(φ(E,zE ) (y (k) , βE , u−E )) + 2ηξ2 = βE     √ (k+1) (k) (k) (k) (k) (k) u−E P2 u−E − η∇u−E g̃(φ(E,zE ) (y , βE , u−E )) + 2ηξ3 (k) (k) (k) where (ξ1 , ξ2 , ξ3 ) ∼ N (0, Ip ) × N (0, I|E| ) × N (0, Ip−|E| ) and independent |E| of everything else and P1 is projection onto the orthant RzE and P2 is the p−|E| projection onto the cube [−1, 1] . The conditioning on nuisance statistic is implemented by fixing PE\j y with the update being ⊥ y (k+1) = Yobs + PE\j ỹ (k+1) , where PE\j is the projection onto the column space of XE\j . Example 7.2. ℓ1 -penalized logistic & LASSO with random design matrix (Section 4.3): In ℓ1 -penalized logistic the response y is modeled as i.i.d. Bernoulli(π(xTi b)), with |supp(b)| = s, with the same null hypothesis as above. In the case of LASSO 7 To construct the pivots from Tian and Taylor (2015) we can either integrate over the null statistic or condition on it. In the case of LASSO or ℓ1 -penalized logistic with random design, the null statistic is T−E defined in (50). In the case of LASSO with fixed X, the null statistic T y. Here we report the results where we integrate over the null statistic (conditioning is X−E on the null is easier computationally since it reduces the size of the sampling space). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 49 with random design, the response is modeled as y ∼ N (Xb, σ 2 In ) with σ 2 = 1. If we denote as β̄E the unpenalized MLE using only selected covariates XE in the respective problems, then we can describe the inference for both problems at once. We sample from the selective density in (53), supported on {(t, βE , u−E ) : PE\j t = Tobs , diag(zE )βE > 0, ku−E k∞ ≤ 1}, conditioning again on selection event B(E,zE ) and nuisance statistics. Here,   Σ̃E,E 0 PE\j is the projection onto the column space of and the nuisance 0 0 statistic equals PE\j T = Σ̃E,E β̄E where ΣE,E is the covariance of β̄E (we use a bootstrap estimate of this covariΣE,E eE,j eT E,j E,j ΣE,E eE,j ance in the shown results) and Σ̃E,E = ΣE,E − eT . The updates in the projected Langevin sampler are  √   (k)  (k) (k) (k) t −ηΣ−1 t(k) − η∇t g̃(φ(E,zE ) (t(k) , βE , u−E )) + 2ηξ1 t̃(k+1) √  (k) (k) (k) (k)  (k+1)   , P β − η∇βE g̃(φ(E,zE ) (t(k) , βE , u−E )) + 2ηξ2 = βE  1 E   √ (k+1) (k) (k) (k) (k) u−E P2 u−E − η∇u−E g̃(φ(E,zE ) (t(k) , βE , u−E )) + 2ηξ3 (k) (k) (k) where (ξ1 , ξ2 , ξ3 ) ∼ N (0, Ip ) × N (0, I|E| ) × N (0, Ip−|E| ) and independent |E| of everything else and P1 is projection onto the orthant RzE and P2 is the p−|E| projection onto the cube [−1, 1] . The conditioning as usual is implemented by keeping PE\j t fixed and updating ⊥ t̃(k+1) . t(k+1) = Tobs + PE\j Example 7.3. Forward stepwise for Kac-Rice (Section 6.1): We perform K steps of randomized forward stepwise as in (81) with the response generated as y ∼ N (Xb, σ 2 In ) and |supp(b)| = s. The null hypothesis is H0 : bjK = 0, where jk is the predictor chosen by the algorithm in the kth step, k = 1, . . . , K. We sample from the density (53) with s = K − 1 supported on  (y, z1 , . . . , zK ) : PAK−1 y = Yobs , zk ∈ ∂IKk (ηk∗ ), k = 1, . . . , K , where we condition on the sufficient statistic for the nuisance parameters in the selected model consisting of predictors XAK . The updates from projected Langevin sampler when one computes the Kac-Rice objective conditional on the selection event determined by forward stepwise with K steps are P  √   (i) (i) K  (i+1) g̃k (φk (y (i) , zk )) + 2ηξ1 y (i) − ηy (i) /σ 2 − η∇y k=1 ỹ     √ (i) (i)   (i) (i)  z1  2ηξ P g̃ (φ (y , z )) + z − η∇   1 z1 1 1 2 1 1   ,  =  ..   ..  .    .     √ zK (i) (i) (i) (i) PK zK − η∇zK g̃K (φK (y , zK )) + 2ηξK+1 imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 50 where Pk is the projection onto ∂IKk (ηk∗ ), given in (82), for all k = 1, . . . , K.8 The conditioning step is then done as above by updating y while keeping PAK−1 y fixed. Example 7.4. Group LASSO with fixed X (Section 5.4): The response y is modeled as N (Xb, σ 2 ) with |supp(b)| = s and σ 2 = 1. We test the null hypothesis H0 : bE,g = 0, bE,g denoting the coefficient corresponding to group g, g ∈ E, conditioning on the selection event B(E,(ug )g∈E ) .9 Additionally, we condition on the sufficient statistics for the nuisance parameters, PE\j y. The sampling density is proportional to (75) on the support ) ( Y n |E| |h| (y, (γg )g∈E , (zh )h∈−E ) ∈ R × (0, ∞) × λh B2 (R ) : PE\g y = Yobs . h∈−E The updates for   ỹ (k+1) (k+1)    (γg )g∈E  (k+1) (zh )h∈−E from projected Langevin sampler can be written as   √ (k) (k) (k) 2 (k) y − ηy /σ − η∇ g̃(φ ) + 2ηξ y 1  √   (k) (k)  P1 (γg )(k) − η∇(γg )g∈E g̃(φ(k) ) + η(Tr(Dg ))g∈E + 2ηξ2  , g∈E     √ (k) (k) P2 (zh )h∈−E − η∇(zh )h∈−E g̃(φ(k) ) + 2ηξ3 where (k) (k) φ(k) = φ(E,(ug )g∈E ) (y (k) , (γg )g∈E , (zh )h∈−E ), (k) (k) (k) (ξ1 , ξ2 , ξ3 ) ∼ N (0, In ) × N (0, I|E| ) × N (0, Ip−|M| ), independent of everything else, P1 is projectionQonto the orthant (0, ∞)|E| and P2 is the projection onto the product of balls h∈−E λh B2 (R|h| ). Here, term (k) (Tr(Gg ))g∈E comes from differentiating the logarithm of the Jacobian (76) with respect to (λg )g∈E at iteration k:  −1 Dg(k) = (Γ− )(k) + V T Q−1 ΛV Gg , where Gg =   ∂V T Q−1 ΛV and (Γ− )(k) = diag (γg(k) I|g|−1 )g∈E . ∂λg Note that Gg is a fixed matrix, thus at every iteration we just need to invert  (k) (Γ− )(k) + V T Q−1 ΛV in order to compute Dg . The conditioning on PE\g y is done as in the above examples. 8 This projection is easily done using Remark 10. includes conditioning on the active directions as well to get a log-concave density (Section 5.4) 9 This imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 51 Acknowledgements Jonathan Taylor was supported in part by National Science Foundation grant DMS-1208857 and Air Force Office of Sponsored Research grant 113039. Jelena Markovic was supported by Stanford Graduate Fellowship. Jonathan Taylor would like to thank the Berkeley Institute of Data Science, where part of this manuscript was written while on sabbatical in Fall 2015. References Adler, R. J. and Taylor, J. (2007), Random Fields and Geometry, 1 edn, Springer. Barber, R. F. and Candes, E. (2014), ‘Controlling the false discovery rate via knockoffs’, arXiv:1404.5609 [math, stat] . arXiv: 1404.5609. URL: http://arxiv.org/abs/1404.5609 Barndorff-Nielsen, O. (1983), ‘On a formula for the distribution of the maximum likelihood estimator’, Biometrika 70(2), 343–365. URL: http://biomet.oxfordjournals.org/content/70/2/343 Becker, S., Candes, E. and Grant, M. (2010), Templates for convex cone problems with applications to sparse signal recovery, Technical report, Stanford University. URL: http://arxiv.org/abs/1009.2065 Belloni, A., Chernozhukov, V. and Wang, L. (2014), ‘Pivotal estimation via square-root Lasso in nonparametric regression’, The Annals of Statistics 42(2), 757–788. arXiv: 1105.1475. URL: http://arxiv.org/abs/1105.1475 Benjamini, Y. (2010), ‘Simultaneous and selective inference: Current successes and future challenges’, Biometrical Journal. Biometrische Zeitschrift 52(6), 708–721. Berk, R., Brown, L., Buja, A., Zhang, K. and Zhao, L. (2013), ‘Valid postselection inference’, The Annals of Statistics 41(2), 802–837. URL: http://projecteuclid.org/euclid.aos/1369836961 Besag, J. and Clifford, P. (1989), ‘Generalized monte carlo significance tests’, Biometrika 76(4), 633–642. Bien, J., Taylor, J. and Tibshirani, R. (2013), ‘A lasso for hierarchical interactions’, Annals of statistics 41(3), 1111. Bubeck, S., Eldan, R. and Lehec, J. (2015), ‘Sampling from a logconcave distribution with Projected Langevin Monte Carlo’, arXiv preprint arXiv:1507.02564 . URL: http://arxiv.org/abs/1507.02564 Bunea, F., Lederer, J. and She, Y. (2013), ‘The Group Square-Root Lasso: Theoretical Properties and Fast Algorithms’, arXiv preprint arXiv:1302.0261 . arXiv: 1302.0261. URL: http://arxiv.org/abs/1302.0261 Chan, H. P. and Walther, G. (2011), ‘Detection with the scan and the average imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 52 likelihood ratio’, arXiv preprint arXiv:1107.4344 . URL: https://arxiv.org/abs/1107.4344 Chen, S., Donoho, D. and Saunders, M. (1998), ‘Atomic decomposition for basis pursuit’, SIAM Journal on Scientific Computing 20(1), 33–61. URL: http://epubs.siam.org/doi/abs/10.1137/S1064827596304010 Choi, Y., Taylor, J. and Tibshirani, R. (2014), ‘Selecting the number of principal components: estimation of the true rank of a noisy matrix’, arXiv preprint arXiv:1410.8260 . arXiv: 1410.8260. URL: http://arxiv.org/abs/1410.8260 Cox, D. (1975), ‘A note on data-splitting for the evaluation of significance levels’, Biometrika 62(2), 441–444. URL: http://biomet.oxfordjournals.org/content/62/2/441.full.pdf Diaconis, P., Holmes, S. and Shahshahani, M. (2013), Sampling from a Manifold, in ‘Institute of Mathematical Statistics Collections’, Institute of Mathematical Statistics, Beachwood, Ohio, USA, pp. 102–125. URL: http://projecteuclid.org/euclid.imsc/1379942050 Diaconis, P. and Sturmfels, B. (1998), ‘Algebraic algorithms for sampling from conditional distributions’, The Annals of Statistics 26(1), 363–397. URL: http://projecteuclid.org/euclid.aos/1030563990 Dwork, C., Feldman, V., Hardt, M., Pitassi, T., Reingold, O. and Roth, A. (2015), ‘The reusable holdout: Preserving validity in adaptive data analysis’, Science 349(6248), 636–638. URL: http://www.sciencemag.org/content/349/6248/636 Efron, B. and Hinkley, D. V. (1978), ‘Assessing the Accuracy of the Maximum Likelihood Estimator: Observed Versus Expected Fisher Information’, Biometrika 65(3), 457–482. URL: http://www.jstor.org/stable/2335893 Efron, B. and Tibshirani, R. (1996), ‘Using specially designed exponential families for density estimation’, The Annals of Statistics 24(6), 2431–2461. URL: http://projecteuclid.org/euclid.aos/1032181161 Federer, H. (1959), ‘Curvature measures’, Transactions of the American Mathematical Society 93(3), 418–491. URL: http://www.ams.org/journals/tran/1959-093-03/S0002-9947-19590110078-1/S0002-9947-1959-0110078-1.pdf Fithian, W., Sun, D. and Taylor, J. (2014), ‘Optimal Inference After Model Selection’, arXiv preprint arXiv:1410.2597 . arXiv: 1410.2597. URL: http://arxiv.org/abs/1410.2597 Fithian, W., Taylor, J., Tibshirani, R. and Tibshirani, R. (2015), ‘Selective sequential model selection’, arXiv preprint arXiv:1512.02565 . URL: https://arxiv.org/abs/1512.02565 Frick, K., Munk, A. and Sieling, H. (2014), ‘Multiscale change point inference’, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 76(3), 495–580. URL: http://onlinelibrary.wiley.com/doi/10.1111/rssb.12047/suppinfo Friedman, J., Hastie, T. and Tibshirani, R. (2008), ‘Sparse inverse covariance estimation with the graphical lasso’, Biostatistics 9(3), 432–441. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 53 URL: http://biostatistics.oxfordjournals.org/content/9/3/432.short Friedman, J., Hastie, T. and Tibshirani, R. (2010), ‘Regularization paths for generalized linear models via coordinate descent’, Journal of Statistical Software 33(1), 1–22. URL: https://core.ac.uk/download/files/153/6287975.pdf Hillier, G. and Armstrong, M. (1999), ‘The Density of the Maximum Likelihood Estimator’, Econometrica 67(6), 1459–1470. URL: http://www.jstor.org/stable/2999568 Hotelling, H. (1939), ‘Tubes and spheres in n-spaces and a class of statistical problems’, American Journal of Mathematics 61, 440–460. Hyun, S., G’Sell, M. and Tibshirani, R. J. (2016), ‘Exact Post-Selection Inference for Changepoint Detection and Other Generalized Lasso Problems’, arXiv preprint arXiv:1606.03552 . URL: https://arxiv.org/abs/1606.03552 Johnstone, I. and Siegmund, D. (1989), ‘On Hotelling’s formula for the volume of tubes and Naiman’s inequality’, The Annals of Statistics 17(1), 184–194. URL: https://projecteuclid.org/euclid.aos/1176347010 Lee, J. D., Sun, D. L., Sun, Y. and Taylor, J. E. (2016), ‘Exact post-selection inference with the lasso’, The Annals of Statistics 44(3), 907–927. URL: http://projecteuclid.org/euclid.aos/1460381681 Lee, J. D. and Taylor, J. E. (2014), Exact Post Model Selection Inference for Marginal Screening, in Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence and K. Q. Weinberger, eds, ‘Advances in Neural Information Processing Systems 27’, Curran Associates, Inc., pp. 136–144. URL: http://papers.nips.cc/paper/5292-exact-post-model-selectioninference-for-marginal-screening.pdf Loftus, J. R. and Taylor, J. E. (2014), ‘A significance test for forward stepwise model selection’, arXiv preprint arXiv:1405.3920 . URL: https://arxiv.org/abs/1405.3920 Meinshausen, N. and Bühlmann, P. (2006), ‘High-dimensional graphs and variable selection with the lasso’, The Annals of Statistics 34(3), 1436–1462. URL: https://projecteuclid.org/euclid.aos/1152540754 Meinshausen, N. and Bühlmann, P. (2010), ‘Stability selection’, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 72(4), 417–473. URL: http://onlinelibrary.wiley.com/doi/10.1111/j. 1467-9868.2010.00740.x/abstract Meinshausen, N., Meier, L. and Bühlmann, P. (2012), ‘P-values for highdimensional regression’, Journal of the American Statistical Association . Ming, Y. and Lin, Y. (2005), ‘Model selection and estimation in regression with grouped variables’, Journal of the Royal Statistical Society: Series B 68(1), 49–67. URL: http://onlinelibrary.wiley.com/doi/10.1111/j. 1467-9868.2005.00532.x/abstract Negahban, S., Ravikumar, P., Wainwright, M. J. and Yu, B. (2010), ‘A unified framework for high-dimensional analysis of $m$-estimators with decomposable regularizers’, arXiv:1010.2731 . imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 54 URL: http://arxiv.org/abs/1010.2731 Panigrahi, S., Taylor, J. and Weinstein, A. (2016), ‘Bayesian post-selection inference in the linear model’, arXiv preprint arXiv:1605.08824 . URL: https://arxiv.org/abs/1605.08824 Rivera, C. and Walther, G. (2013), ‘Optimal detection of a jump in the intensity of a poisson process or in a density with likelihood ratio statistics’, Scandinavian Journal of Statistics 40(4), 752–769. URL: http://onlinelibrary.wiley.com/doi/10.1111/sjos.12027/full Sabourin, J., Valdar, W. and Nobel, A. (2014), ‘A Permutation Approach for Selecting the Penalty Parameter in Penalized Model Selection’, arXiv preprint arXiv:1404.2007 . arXiv: 1404.2007. URL: http://arxiv.org/abs/1404.2007 Schneider, R. (1993), Convex Bodies: the Brunn-Minkowski Theory, Vol. 44 of Encyclopedia of Mathematics and its Applications, Cambridge University Press, Cambridge. Sun, J. (1993), ‘Tail probabilities of the maxima of Gaussian random fields’, The Annals of Probability 21(1), 34–71. URL: https://projecteuclid.org/euclid.aop/1176989393 Sun, T. and Zhang, C.-H. (2011), ‘Scaled Sparse Linear Regression’, arXiv preprint arXiv:1104.4595 . arXiv: 1104.4595. URL: http://arxiv.org/abs/1104.4595 Takemura, A. and Kuriki, S. (2002), ‘Maximum of Gaussian field on piecewise smooth domain: Equivalence of tube method and Euler characteristic method.’, Ann. of Appl. Prob. 12(2), 768–796. URL: https://projecteuclid.org/euclid.aoap/1026915624 Taylor, J. E. (2006), ‘A Gaussian kinematic formula’, The Annals of Probability 34(1), 122–158. URL: https://projecteuclid.org/euclid.aop/1140191534 Taylor, J. E. and Adler, R. J. (2009), ‘Gaussian processes, kinematic formulae and Poincares limit’, The Annals of Probability 37(4), 1459–1482. URL: https://projecteuclid.org/euclid.aop/1248182144 Taylor, J., Loftus, J. and Tibshirani, R. (2013), ‘Tests in adaptive regression via the Kac-Rice formula’, The Annals of Statistics 44(2), 743–770. URL: http://projecteuclid.org/euclid.aos/1458245734 Taylor, J. and Tibshirani, R. (2016), ‘Post-selection inference for l1-penalized likelihood models’, arXiv preprint arXiv:1602.07358 . URL: http://arxiv.org/abs/1602.07358 Tian, X., Bi, N. and Taylor, J. (2016), ‘MAGIC: a general, powerful and tractable method for selective inference’, ArXiv e-prints . URL: https://arxiv.org/abs/1607.02630 Tian, X., Loftus, J. R. and Taylor, J. E. (2015), ‘Selective inference with unknown variance via the square-root LASSO’, arXiv preprint arXiv:1504.08031 . arXiv: 1504.08031. URL: http://arxiv.org/abs/1504.08031 Tian, X. and Taylor, J. E. (2015), ‘Selective inference with a randomized response’, arXiv preprint arXiv:1507.06739 . arXiv: 1507.06739. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 55 URL: http://arxiv.org/abs/1507.06739 Tibshirani, R. (1996a), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal Statistical Society: Series B 58(1), 267–288. Tibshirani, R. (1996b), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal Statistical Society: Series B 58(1), 267–288. URL: http://onlinelibrary.wiley.com/doi/10.1111/j. 1467-9868.2011.00771.x/abstract Tibshirani, R. J. (2013), ‘The Lasso Problem and Uniqueness’, Electronic Journal of Statistics 7, 1456–1490. URL: http://projecteuclid.org/euclid.ejs/1369148600 Tibshirani, R. J. (2015), ‘A General Framework for Fast Stagewise Algorithms’, Journal of Machine Learning Research 16, 2543–2588. URL: http://jmlr.org/papers/v16/tibshirani15a.html Tibshirani, R. J. and Taylor, J. (2011), ‘The solution path of the generalized lasso’, The Annals of Statistics 39(3), 1335–1371. URL: http://projecteuclid.org/euclid.aos/1304514656 Tibshirani, R. J. and Taylor, J. (2012), ‘Degrees of freedom in lasso problems’, The Annals of Statistics pp. 1198–1232. Tibshirani, R., Taylor, J., Lockhart, R. and Tibshirani, R. (2014), ‘Post-selection adaptive inference for Least Angle Regression and the Lasso’, arXiv preprint arXiv:1401.3889 . URL: http://arxiv.org/abs/1401.3889 Van de Geer, S. A. (2008), ‘High-dimensional generalized linear models and the lasso’, The Annals of Statistics pp. 614–645. Wainwright, M. J. (2009), ‘Sharp thresholds for high-dimensional and noisy sparsity recovery using-constrained quadratic programming (lasso)’, Information Theory, IEEE Transactions on 55(5), 2183–2202. Weyl, H. (1939), ‘On the volume of tubes’, American Journal of Mathematics 61, 461–472. Yang, F., Foygel Barber, R., Jain, P. and Lafferty, J. (2016), ‘Selective Inference for Group-Sparse Linear Models’, ArXiv e-prints . URL: https://arxiv.org/abs/1607.08211 Zhao, P. and Yu, B. (2006), ‘On model selection consistency of lasso’, The Journal of Machine Learning Research 7, 2541–2563. Zhou, Q. (2014), ‘Monte Carlo Simulation for Lasso-Type Problems by Estimator Augmentation’, Journal of the American Statistical Association 109(508), 1495–1516. arXiv: 1401.4425. URL: http://arxiv.org/abs/1401.4425 Zou, H. and Hastie, T. (2005), ‘Regularization and variable selection via the elastic net’, Journal of the Royal Statistical Society: Series B 67(2), 301–320. URL: http://onlinelibrary.wiley.com/doi/10.1111/j. 1467-9868.2005.00503.x/abstract imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 56 Appendix A: Nuisance parameters & conditional distributions In forming selective hypothesis tests or intervals when the model M is an exponential family, it was noted in Fithian et al. (2014) that the classical approach of conditioning on appropriate sufficient statistics can be used to eliminate nuisance parameters. In this section we describe selective samplers that can be used to sample from such conditional distributions. Specifically, suppose we use the randomized Lasso (41) with X considered fixed and Ē is a subset of features that we will use to form the selected model n o 2 2 MĒ = N (XĒ βĒ , σĒ ) : βĒ ∈ R|Ē| , σĒ >0 . (86) Often, for a given (E, sE ) observed after fitting the randomized Lasso, we will 2 choose Ē = E but this is not strictly necessary. As σĒ is part of the indexing 2 set for MĒ , we are assuming σĒ is unknown. Suppose we want a selective test of H0,j|Ē : βj|Ē = 0 in the model MĒ . Standard exponential family calculations detailed in Fithian et al. (2014) tell us that we can construct such a test in the presence of the nuisance 2 parameters (βĒ\j|Ē , σĒ ) by conditioning on the appropriate sigma-algebra:    T σ XĒ\j y, kyk22 = σ PĒ\j y, k(I − PĒ\j )yk22 . For any fixed values of the sufficient statistics, say (Wobs , SSEobs ) (with PĒ\j Wobs = Wobs ) the conditional distribution for any F ∈ MĒ is supported on the set  Dobs = y : y = Wobs + r, PĒ\j r = 0, krk2 = SSEobs . Under H0,j|Ē it is uniformly distributed over the above set. We must therefore sample from the set {(y, βE , u−E ) : y ∈ Dobs , diag(sE )βE > 0, kuk−E ≤ 1} with a density proportional to      βE sE T f (y) · g ǫ . − X (y − XE βE ) + λ 0 u−E Appendix B: LASSO and variants In this section, we describe two common variants of the LASSO. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 57 B.1. LASSO in bound form The LASSO program in bound form is defined as minimize β∈Rp :kβk1 ≤δ 1 ky − Xβk22 2 (87) with its corresponding randomized version minimize β∈Rp :kβk1 ≤δ 1 ǫ ky − Xβk22 + kβk22 − ω T β. 2 2 (88) In the notation established so far S=y 1 ℓ(β; y) = ky − Xβk22 2 ( 0 kβk1 ≤ δ P(β) = ∞ otherwise. Typically, we will be interested in doing inference when the constraint above is tight. In this case, for the canonical event B(E,zE ) the set S̄0F (G, ǫ, ℓ, P) can be parametrized by  (y, βE , u−E , c) : y ∈ Rn , βE ∈ R|E| , kβE k1 = δ, diag(zE )βE > 0, u−E ∈ Rp−|E| , ku−E k∞ ≤ 1,  c ∈ R, c > 0 . Another possible reparametrization, can be stated as  (y, βE\1 , u−E , c) : y ∈ Rn , βE\1 ∈ R|E|−1 , sE,1 βE,1 = δ − sTE\1 βE\1 , diag(sE\1 )βE\1 > 0, u−E ∈ Rp−|E| , ku−E k∞ ≤ 1,  c ∈ R, c > 0 , where βE\1 = (βE,2 , . . . , βE,|E| ) and similarly sE\1 . The above parameterization can be expressed as      βE zE T ψ(E,zE ) (y, βE\1 , u−E , c) = y, ǫ , − X (y − XE βE ) + c 0 u−E     βE zE , −X T (X − XE βE ), c , 0 u−E imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 58 where βE is expressed in terms of βE\1 as above and the corresponding randomization reconstruction map is     zE β , φ(E,zE ) (y, βE\1 , u−E , c) = ǫ E − X T (y − XE βE ) + c u−E 0 where   (δ − sTE\1 βE\1 )sE,1 . βE =  βE\1 0 The sampling density is proportional to  f (y) · g φ(E,zE ) (y, βE\1 , u−E , c) · det(D(βE\1 ,u−E ,c) φ(E,zE ) ) , where the last determinant equals −c p−|E| −ǫsTE\1 sE,1 + X1T (−sE,1 X1 sTE\1 + XE\1 ) sE,1 det T ǫI + XE\1 (−sE,1 X1 sTE\1 + XE\1 ) sE\1 ! . B.2. Basis pursuit The basis pursuit problem Chen et al. (1998) is defined as minimize β∈Rp :ky−Xβk2 ≤δ kβk1 . (89) In the notation established so far S=y ( ℓ(β; y) = 0 ky − Xβk2 ≤ δ ∞ otherwise P(β) = kβk1 . The set on which the problem (89) has a solution is n o y : θ̂(y) 6= 0 = {y : k(I − PC )yk2 ≤ δ} , where PC denotes projection onto col(X). Its randomized version is minimize β∈Rp :ky−Xβk2 ≤δ ǫ kβk22 − ω T β + kβk1 . 2 (90) This problem, with non-random choices of ω is considered in Becker et al. (2010). We see that n o  (y, ω) : θ̄(y, ω) 6= ∅ = (y, ω) : θ̂(y) 6= ∅ . imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 59 Generically, when a solution exists, the constraint will be tight, hence our events of interest will typically condition on ky − Xβk2 = δ. We see then  S̄0F (G, ǫ, ℓ, P) = (y, ω, β, α, z) : y ∈ supp(F ), ky − Xβk2 = δ, ω ∈ supp(G), ω = ǫβ + α + z, α = cX T (Xβ − y), c > 0, z ∈ ∂P(β) . by For the canonical event B(E,zE ) , the set S̄BF (G, ǫ, ℓ, P) can be parameterized  (y, βE , u−E , c) : k(I − PC )yk2 ≤ δ, βE ∈ R|E| , diag(zE )βE > 0, ky − XE βE k2 = δ, u−E ∈ Rp−|E| , ku−E k∞ ≤ 1,  c ∈ R, c > 0 . (91) This set has the form of a bundle over a subset of Rn , with fibers that are the product a |E| − 1 dimensional ellipse, a p − |E| dimensional cube and a half-line. The parameterization is      β zE ψ(E,zE ) (y, βE , u−E , c) = y, ǫ E − cX T (y − XE βE ) + , 0 u−E     βE zE . , −cX T (y − XE βE ), 0 u−E     zE βE T , the − cX (y − XE βE ) + Denoting φ(E,zE ) (βE , u−E , c) = ǫ u−E 0 sampling density is proportional to  f (y) · g φ(E,zE ) (βE , u−E , c) · det(D(βE ,u−E ,c) φ(E,zE ) ) and restricted to (91), where det(D(βE ,u−E ,c) φ(E,zE ) ) is the Jacobian of the parameterization in the coordinates described above, i.e. coordinates on the product of an |E| − 1 dimensional ellipse, an p − |E| dimensional cube and a half-line. B.3. Square-root LASSO Using the LASSO with an unknown noise level is somewhat of a chicken-and-egg problem, as knowing which value of λ to choose is somewhat difficult. Suppose that instead we use the square-root LASSO. The square-root LASSO program Belloni et al. (2014); Sun and Zhang (2011) is defined as minimize ky − Xβk2 + λkβk1 . p β∈R (92) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 60 Selective inference for this program was considered in (Tian et al., 2015). This program has the advantage that it is possible to choose a reasonable value of λ without knowing anything about the noise level. Its randomized version is ǫ minimize ky − Xβk2 + kβk22 − ω T β + λkβk1 . p β∈R 2 (93) We can take the law of ω to be independent of any particular noise level, though we probably should choose ǫ = O(n−1/2 ) if thinking of something like the pairs model. In the notation established so far S = y, ℓ(β; y) = ky − Xβk2 , P(β) = λkβk1 . We are most interested in inference when y − X β̂(y, ω) 6= 0 on which ℓ is differentiable and for the canonical event B(E,zE ) = {β : diag(zE )βE > 0, β−E = 0} we can parametrize {(y, ω, β, α, z) ∈ S̄0 (G, ǫ, ℓ, P) : y ∈ Dobs , β ∈ B(E,zE ) } as ψ(E,zE ) (y, βE , u−E ) =     X T (y − XE βE ) zE βE , +λ y, ǫ − u−E 0 ky − XE βE k2 !    X T (y − XE βE ) zE βE . ,λ ,− u−E 0 ky − XE βE k2 |E| with the domain Dobs × RzE × [−1, 1]p−|E|. We therefore must sample from a density proportional to        T X T (y − XE βE ) XE R(y, βE )XE βE zE f (y)·g ǫ − +λ + ǫI · det 0 u−E ky − XE βE k2 ky − XE βE k2 and restricted to to the domain of ψ(E,zE ) , where R(y, βE ) = In − (y − XE βE )(y − XE βE )T . ky − XE βE k22 B.4. More complex penalties: generalized LASSO and other quadratic programs Clearly, not every statistical learning problem of interest can be expressed in terms of the LASSO or forward stepwise. For example, the generalized LASSO Tibshirani and Taylor (2011) considers a penalty of the form P(β) = λkDβk1 . Selective inference for the entire solution path of the generalized LASSO, analgous to the solution path of Least Angle Regression and LASSO in Tibshirani et al. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 61 (2014) has been considered in Hyun et al. (2016). We consider a fixed value of λ combined with randomization. For some D, the programs can be re-expressed as LASSO problems but not all D (c.f. Tibshirani and Taylor (2011)). More generally, we might be confronted with solving a problem of the form minimize ℓ(β; S) + P̄(Dβ − γ). β (94) That is, we consider our usual problem with P(β) = P̄(Dβ − γ). For example, a linearly constrained quadratic program might take the form 1 minimize β T Qβ − S T β β 2 (95) subject to Dβ ≤ γ with S ∼ F . In this case, P̄ is the non-positive cone constraint. Sometimes, the Fenchel conjugate P ∗ is simple in the sense that its naturally associated selection events are easy to parametrize. In this case, we can proceed as before and simply consider randomized programs of the form ǫ 1 minimize β T Qβ − (S + ω)T β + P̄(Dβ − γ) + kβk22 . β 2 2 (96) However, when P ∗ is complex, then our usual approach may be computationally expensive. In this section, we describe an alternative randomization scheme that may yield simpler sampling algorithms as described in Section 7. For concreteness, we consider the quadratic program (96). Similar calculations hold for the generalized LASSO by simply replacing one convex conjugate with another. The issue of parameterization of selection events arises quickly upon inspection of the subgradient equation for the randomized program (96). It will often be of interest to condition on the set of tight constraints o n Ê(S, ω) = j : DjT β̂(S, ω) = γj . On this event, the subgradient equation reads T ω = Qβ − S + DE zE T where zE ≥ 0 and DE β = γE . We see the sub-gradient DE zE lies in the cone generated by the rows of DE . If we were to parametrize this selection event, we might write it as ω = Qβ − S + z̄E where z̄E ∈ cone(DE ). For arbitrary D and E, this cone may be rather complex. Each step of the projected Langevin implementation described in Section 7 requires projection onto this set. Projection onto this set can be achieved via a imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 62 non-negative least squares problem, which would in principal need to be re-run for every step of the sampler. We shall instead solve the corresponding randomized dual program. Introducing variable v = Dβ − γ, the dual of (95) which solves   ǫ 1 maximize minimize β T Qβ − S T β + kβk22 + uT (Dβ − γ − v) , u β,v:v≤0 2 2 subject to u ≥ 0 is equivalent to solving 1 minimize (DT u − S)T (Q + ǫI)−1 (DT u − S) + uT γ u 2 subject to u ≥ 0 (for the generalized LASSO, the non-negative cone constraint is replaced with the constraint kuk∞ ≤ λ). The set of tight constraints can be thought of as corresponding to non-zero u’s in this dual problem E. When the dual problem is strongly convex this is a one-to-one correspondence, though if it is not strongly convex there may be more than one dual solution Tibshirani and Taylor (2012). The randomized dual problem solves the program 1 minimize (DT u − S)T (Q + ǫI)−1 (DT u − S) + uT (γ − ω) u 2 (97) subject to u ≥ 0, which is dual to 1 ǫ minimize β T Qβ − S T β + kβk22 + IK (β) β 2 2 with K = {β : Dβ − γ + ω ≤ 0} and IK denoting its characteristic function. However, for the above randomized QP that we propose to solve, the law of ω has to supported on the set   ω : ∩j β : DjT β ≤ γj − ωj 6= 0 . to ensure feasibility of the primal problem, though the dual problem always has a solution. Remark 19. Generally, this is not a problem in examples like the generalized LASSO, where P is finite everywhere. Further, in the applications of selective inference we have in mind, a data analyst is typically solving statistical learning programs to suggest parameters of interest. For example, Heirnet Bien et al. (2013) is a statistical learning method that seeks hierarchical interactions that can be expressed as a linearly constrained quadratic program. In this case, a randomization which makes the primal problem infeasible is not catastrophic. Denoting Qǫ = (Q + ǫI)−1 , imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 63 one solution to the above problem that ensures feasibility of the primal problem is to instead consider ǫ′ 1 minimize (DT u − S)T Qǫ (DT u − S) + uT (γ − ω) + kuk22 u 2 2 (98) subject to u ≥ 0 for ǫ′ > 0. This always yields a feasible primal 1 minimize z T (DQǫ DT + ǫ′ I)−1 z − z T (DQǫ DT + ǫ′ I)−1 (DQǫ S − γ + ω), z≤0 2 as DQǫ DT + ǫ′ I is a positive definite matrix with added positive ǫ′ . The reconstruction map in this case is given by     uE 0 T ω = φE (S, u, z−E ) = DQǫ (DE uE − S) + ǫ′ , (99) +γ+ 0 z−E which allows us to sample (S, uE , z−E ) from a sampling density proportional to f (S) · g(φE (S, uE , z−E )) |E| p−|E| and supported on supp(F ) × R+ × R− , where R+ = {x ∈ R : x ≥ 0} and R− = {x ∈ R : x ≤ 0}. Another possibility is to consider the dual of the randomized objective (96) given by 1 minimize (DT u − S − ω)T Qǫ (DT u − S − ω) + uT γ, u≥0 2 which lead to reconstruction map T DQǫ ω = DQǫ DE uE − DQǫ S + γ +  0 z−E  where we have conditioned on the value of the set Ē(S, ω) = {j : ûj (S, ω) > 0} of being E. As noted above, the set Ē(S, ω) may not be unique even when E(S, ω) is (c.f. Tibshirani and Taylor (2012)). In this sense, we are conditioning on a particular value of the boundary set, the boundary set determined by the solver we use to solve this dual problem. Denoting the density of DQǫ ω as g̃, the sampling density of (S, uE , z−E ) is now proportional to    0 T f (S) · g̃ DQǫ DE uE − DQǫ S + γ + z−E and restricted to z−E ≤ 0, uE ≥ 0 and z−E + γ−E ∈ col(D−E ). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 64 ⊥ To avoid enforcing the implicit constraint of PD (z−E + γ−E ) being 0 at each −E step of the sampler, we can sample from an approximate density for (S, uE , z−E ) ⊥ that substitutes the Dirac delta operator δ0 (PD (z−E +γ−E )) with a smoothed −E version, an example being !    2 1 0 ⊥ T ·exp − ′ PD−E (z−E + γ−E ) f (S)·g̃ DQǫ DE uE − DQǫ S + γ + z−E 2ǫ 2 |E| p−|E| supported on supp(F ) × R+ × R− , a much simpler constraint set. Appendix C: Neighborhood selection We now consider neighborhood selection (a selective sampler for which appears in Tian et al. (2016)). Meinshausen and Bühlmann (2006) proposed neighborhood selection with the LASSO to achieve this goal. The algorithm can be formulated as the following optimization problem, for any node i β̂ i,λ = argmin kxi − Xβk22 + λkβk1 , (100) β∈Rp ,βi =0 where with slight abuse of notation, X ∈ Rn×p is the data matrix observed from n i.i.d observations, and xi is the i-th column of X. Choice of λ is discussed in Chapter 3 of Meinshausen and Bühlmann (2006). Denote B̂ = (β̂ 1 , β̂ 2 , . . . , β̂ p ), we propose the randomized version of (100), B̂ = ǫ kX − XBk2F + λkBk1 + kBk2F − ΩB, 2 B∈Rp×p ,Bii =0 ∀i∈{1,...,p} argmin (101) i.i.d where Ω = (ω 1 , . . . , ω p ) ∼ G. This is the matrix form of (33), and the KKT conditions are decomposable across the nodes. Suppose for node i, E i is the active set for (100), si is the corresponding signs, λ · ui is the subgradient corresponding to the inactive variables except the i-th and X−i is the columns of X except the i-th column. For every node i, the i-th coordinate of β i is held to be zero, and (100) is in fact a regression of dimension p − 1, thus T i αi = −X−i (xi − XE i βE (102) i ), and the reparametrization map, ψ(E,zE ) (X, B − , u) = (X, α + z + ǫB − , B − , α, z), where α = (α1 , α2 , . . . , αp ), αi from (102),  i s 1 2 p i z = (z , z , . . . , z ), z = λ i , u B − ∈ R(p−1)×p is p × p matrix without the diagonal elements. imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 65 Since ω i , 1 ≤ i ≤ p, are independent, and the Jacobian Y T Jψ(E,zE ) (X, B − , u) = det(XE i XE i + ǫI), i∈Γ conditioning on (E i , si ), the distribution for selective inference has the following density,  i  Y  β i i  s T i g ǫ E + λ i − X−i (xi − XE i βE f (X) · i) 0 u i∈Γ (103) Y T · det(XE i XE i + ǫI) . i∈Γ After seeing the active set E = (E 1 , . . . , E p ), it is natural to choose the selected model (i.e. the edge set E) to be E. However, since the active set E is not necessarily symmetric, we choose the edge set E to be E ∨ , where E ∨ = {(i, j) | Eij = 1 or Eji = 1}. Under this model, the distribution of X is an exponential family. More specifically, conditioning on the set of edges E ∨ ,   1 T f (X) ∝ exp − Tr(ΘX X) 2   X X 1 = exp − Θij xTi xj − Θii kxi k2  . 2 ∨ (i,j)∈E i∈Γ Note that this is an exponential family with sufficient statistics {xTi xj , (i, j) ∈ E ∨ , kxi k2 , i ∈ Γ}. Therefore, the law for selective inference (103) is also an exponential family with the same sufficient statistics. To construct the UMPU tests as in Fithian et al. (2014) for the null hypothesis H0,ij : Θij = 0, we condition on the sufficient statistics corresponding to the nuisance parameter and sample from the law (103). Appendix D: Recomputing Jacobian for group lasso We apply the Jacobian meta theorem 2 to the group Lasso to re-derive the Jacobian in the selective sampler density from a geometric perspective. For the tangential frame Vi,z we can split this over groups as (ΠTg (Vg,j,zg )1≤j≤|g|−1 )g∈E = (Vg,zg )g∈E where Πg : Rp → Rg is projection onto the g coordinates represented by the matrix Πg ∈ R|g|×p . The the vectors Vg,j,zg are chosen to be are orthonormal imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 66 T z = 0. For the normal frame within Tg (λg S(Rg )) additionally satisfying Vg,j,z g g we can take ηg = zg /λg and X β= γg zg /λg . g∈G In the standard basis of Tβ Rp , the metric (77) has matrix X T X + ǫI. The curvature matrix C−β is 0 except on the tangent spaces coming from active groups, on which it is block diagonal with blocks γg I|g|−1 , g ∈ E. λg Hence the matrix of interest is −1 G(z, γ)   GE (z, γ)−1 CE,−β (z, γ) 0 C−β (z, γ) = 0 0 where the blocks of GE are T Πg (X T X + ǫI)−1 ΠTh Vh,zh , GE,gh (z, γ) = Vg,z g g, h ∈ E. The eigenvalues of the matrix which show up in the determinant are the same as the reciprocal of the eigenvalues of the matrix with blocks ḠE,gh (z, γ) = λg 1/2 T λh Vg,zg (X T X + ǫI)−1 Vh,zh , γg γh g, h ∈ E. (104) Finally, the matrix H(z, γ) can be ignored if we consider X to be fixed. Appendix E: Multiple views of the data In this section, we consider three more algorithms that choose variables by querying the data several times. E.1. Top K screening A simple way of variable selection called top K screening selects the K most correlated features Xj with response vector y. For a given K, the unrandomized version of such a screening selects most correlated variables (j1 , j2 , . . . , jK ) with corresponding signs (s1 , s2 , . . . , sK ), such that s1 XjT1 y ≥ s2 XjT2 y ≥ · · · ≥ sK XjTK y ≥ max j6∈{j1 ,...,jK } |XjT y|. Selective inference in the nonrandomized setting for this problem was considered in (Lee and Taylor, 2014). imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 67 Here, we consider a randomized version of the optimization above yielding the k-th most correlated variable for 1 ≤ k ≤ K, with added independent randomization variable ωk ∼ Gk ∈ Rp−k+1 and corresponding density gk is given by T maximize η T (X−A y + ωk ) − IKk (η), where y ∼ F. k−1 η∈Rp−k+1 (105) Here, Ak = {j1 , j2 , . . . , jk } is the active set including the k-th step, X−Ak are the columns of X except for the ones corresponding to the current active set Ak , ωk is a sequence of randomizations and IKk (η) is the characteristic function of Kk = {η ∈ Rp−k+1 : kηk1 ≤ 1}, that is, IKk (η) = ( 0 if η ∈ Kk ∞ otherwise. Denote the optimal solution of (105) as ( sk if j = jk ∗ ηk,j = 0 otherwise, (106) where j ∈ {1, . . . , p} \ Ak−1 = Ack−1 10 and jk = argmax XjT y + ωk,j . The j∈Ack−1 subgradient equation in k-th step leads to a reconstruction map for the randomization given by T ωk = φk (y, zk ) = −X−A y + zk , k−1 constraining sub-differential zk ∈ Rp−k+1 to zk ∈ ∂IKk (ηk∗ ) = {c · u : u ∈ Rp−k+1 , ujk = sk , |uj | ≤ 1 ∀j ∈ Ack−1 , c > 0}. Conditioning on the selection event of choosing the K most correlated variables with their corresponding signs  K Y  n K Rp−k+1 : sign(XjTk y + ωk ) = sk , Ê{(sk ,jk )}K = ∈ R × y, {ω } k k=1 k=1 k=1  T sk (XjTk y + ωk ) ≥ max |X y + ω |, k = 1, . . . , K . k j c j∈Ak−1 we sample (y, z1 , . . . , zk ) from the selective sampling density proportional to f (y) · K Y k=1   T y , gk zk − X−A k−1 (107) 10 η ∗ ∈ Rp−k+1 is indexed by this set and all p − k + 1-dimensional vectors in this and the k following section will be indexed by Ack−1 . imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 68 supported on Rn × K Y ∂IKk (ηk∗ ). k=1 For logistic regression, one can replace the T statistics above with the score statistics as described in Remark 13. E.2. Stagewise algorithms Instead of fully projecting out the current variables at each step as in forward stepwise, one can use an incremental approach as in a stagewise algorithm Tibshirani (2015). The first step of a randomized version of such an algorithm might consist of solving the problem maximize η T (X T (y − Xα0 ) + ω1 ) η:kηk1 ≤1 (108) with α0 = η0∗ = 0. We update α1 = α0 + δ · η0∗ for some learning rate δ > 0. Subsequent problems are given by maximize η T (X T (y − Xαk ) + ωk ) η:kηk1 ≤1 with solution ηk∗ and αk = δ · k−1 X (109) ηj∗ . j=0 After K steps, the sampler density is thus proportional to f (y) · K Y k=1  gk zk − X T (y − Xαk−1 ) , (y, z1 , . . . , zk ) ∈ Rn × p with K = {η ∈ R : kηk1 ≤ 1}. K Y (110) ∂IK (ηk∗ ). k=1 E.3. Screening via thresholding randomized Post-LASSO An alternative way to screen is through the following two-stage procedure where we use randomized LASSO as in (41) in the first stage to select the model (E, zE ). In the second stage, we solve an unpenalized, randomized program as in (61) with the selected predictors XE 1 ǫ1 minimize ky − XE γk22 + kγk22 − ω1T γ, ((X, y), ω1 ) ∼ F × G1 , |E| 2 2 γ∈R (111) imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016 Tian Harris et al./Selective sampling 69 where G1 is a known distribution on R|E| and ω1 is independent from the randomization variable ω used in the first stage. We perform a second stage of selection based on the output of the above convex program to threshold the resulting coefficients of γ̂((X, y), ω1 ) resulting in a further selected model Ẽ = {i : |γ̂i ((X, y), ω1 )| > aσ} with their signs z̃Ẽ = sign(γ̂Ẽ ((X, y), ω1 )), where a is a constant and σ is again the scaling which can be estimated by the noise variance of the selected model as in Section 4.6. The canonical event of interest from the two-step procedure becomes B(E,zE ,Ẽ,z̃ Ẽ ) = {(βE ,u−E , γ) : diag(zE )βE > 0, ku−E k∞ ≤ 1, diag(z̃Ẽ )γẼ > 0, |γi | ≥ aσ ∀ i ∈ Ẽ, kγ−Ẽ k∞ < aσ}. Now the sampling density on ((X, y), βE , u−E , γ) becomes proportional to       zE β · g1 ǫ1 γ − X T (y − Xγ) f (X, y) · g ǫ E − X T (y − XE βE ) + λ u−E 0 where the optimization variables (βE , u−E , γ) are restricted to B(E,zE ,Ẽ,z̃ ) . Ẽ imsart-generic ver. 2008/08/29 file: selective_sampler.tex date: September 20, 2016
10math.ST
ON DELTA INVARIANTS AND INDICES OF IDEALS arXiv:1705.05042v1 [math.AC] 15 May 2017 TOSHINORI KOBAYASHI Abstract. Let R be a Cohen-Macaulay local ring with a canonical module. We consider Auslander’s (higher) delta invariants of powers of certain ideals of R. Firstly, we shall provide some conditions for an ideal to be a parameter ideal in terms of delta invarints. As an application of this result, we give upper bounds for orders of Ulrich ideals of R when R has Gorenstein punctured spectrum. Secondly, we extend the definition of indices to the ideal case, and generalize the result of Avramov-Buchweitz-Iyengar-Miller on the relationship between the index and regularity. 1. Introduction Let (R, m, k) be a Cohen-Macaulay local ring with a canonical module. The Auslander δ-invariant δR (M) for a finitely generated R-module M is defined to be the rank of maximal free summand of the a minimal Cohen-Macaulay approximation of M. For an integer n ≥ 0, the n-th δ-invariant is also defined by Auslander, Ding and Solberg [2] as δRn (M) = δR (ΩnR M), where ΩnR M denotes the n th syzygy module of M in the minimal free resolution. On these invariants, combining the Auslander’s result (see[2, Corollary 5.7]) and Yoshino’s one[13], we have the following theorem. Theorem 1.1 (Auslander, Yoshino). Let d > 0 be the Krull dimension of R. Consider the following conditions. (a) R is a regular local ring. (b) There exists n ≥ 0 such that δ n (R/m) > 0. (c) There exists n > 0 and l > 0 such that δ n (R/ml ) > 0. Then, the implications (a) ⇔ (b) ⇒ (c) hold. The implication (b) ⇒ (a) holds if depth grm (R) ≥ d − 1. Here we denote by grI (R) the associated graded ring of R with respect to an ideal I of R. In this paper, we characterize parameter ideals in terms of (higher) δ-invariants as follows. Theorem 1.2. Let (R, m) be a Cohen-Macaulay local ring with a canonical module ω, having a infinite residue field k and Krull dimension d > 0. Put I be an m-primary ideal of R such that I/I 2 is a free R/I-module. Consider the following conditions. (a) δ(R/I) > 0. (b) I is a parameter ideal of R. (c) ∃n ≥ 0 such that δ n (R/I) > 0. 2010 Mathmatics subject Classification. 13A30, 13C14, 13H10. Key words and phrases. Cohen-Macaulay ring, Cohen-Macaulay approximation, Ulrich ideal. 1 2 TOSHINORI KOBAYASHI (d) ∃n > 0 and l > 0 such that δ n (R/I l ) > 0. Then, the implications (a) ⇒ (b) ⇔ (c) ⇒ (d) hold. The implication (d) ⇒ (c) holds if depth grI (R) ≥ d − 1 and I i /I i+1 is a free R/I-module for any i > 0. The implication (b) ⇒ (a) also holds if I ⊂ tr(ω). Here tr(ω) is the trace ideal or ω. that is, the image of the natural homomorphism ω ⊗R HomR (ω, R) → R mapping x ⊗ f to f (x) for x ∈ ω and f ∈ HomR (ω, R). This result recovers the Theorem 1.1 by letting I = m. On the other hand, Ding studies the δ-invariant of R/ml with l ≥ 1 and defines the index index(R) of R to be the smallest number l such that δ(R/ml ) = 1. Extending this, we define the index for an ideal. Definition 1.3. For an ideal I of R, we define the index index(I) of I to be the infimum of integers l ≥ 1 such that δR (R/I l ) = 1. The definition above recovers the ordinary index by taking the maximal ideal. Using the argument of Ding [5] concerning on indices of rings, Avramov, Buchweitz, Iyengar and Miller[3, Lemma 1.5] showed the following equality. Theorem 1.4 (Avramov-Buchweitz-Iyengar-Miller). Assume that R is a Gorenstein local ring and grm (R) is a Cohen-Macaulay graded ring. Then index(R) = reg(grm (R)) + 1. The other main aim of this paper is to prove the following result. Theorem 1.5. Let R be a Cohen-Macaulay local ring having a canonical module and Krull dimension d > 0, and I be an ideal of R such that grI (R) is a Cohen-Macaulay graded ring and I l /I l+1 is R/I-free for 1 ≤ l ≤ index I. Then we have index I ≥ reg(grI (R)) + 1. The equality holds if I ⊂ tr(ω). Note that this theorem recovera Theorem 1.4 by letting I = m. There are some examples of ideals which satisfy the whole conditions in Theorem 1.2 and 1.5. One of them is the maximal ideal m in the case that grm (R) is Cohen-Macaulay (for example, R is a hypersurface or a localization of a homogeneous graded ring.) Other interesting examples are Ulrich ideals. These ideals are defined in [6] and many examples of Ulrich ideals are given in [6] and [7]. We shall show in Section 3 that Ulrich ideals satisfy the assumption of Theorem 1.2 and 1.5. We have an application of Theorem 1.2 concerning Ulrich ideals as follows. Corollary 1.6. Let I be an Ulrich ideal of R that is not a parameter ideal. Assume that R is Gorenstein on the punctured spectrum. Then I 6⊂ mindex(R) . In particular, sup{n | I ⊂ mn for any Ulrich ideal I that is not a parameter ideal} is finite. We prove this result in Section 3. 2. Proofs Throughout this section, let (R, m, k) be a Cohen-Macaulay local ring of dimension d > 0 with a canonical module ω, and assume that k is infinite. We recall some basic properties of the Auslander δ-invariant. 3 For a finitely generated R-module M, a short exact sequence (2.0.1) p 0→Y →X − →M →0 is called a Cohen-Macaulay approximation of M if X is a maximal Cohen-Macaulay Rmodule and Y has finite injective dimension over R. We say that the sequence (2.0.1) is minimal if each endomorphism φ of X with p ◦ φ = p is an automorphism of X. It is known (see [1], [8]) that a minimal Cohen-Macaulay approximation of M exists and is unique up to isomorphism. If the sequence (2.0.1) is a minimal Cohen-Macaulay approximation of M, then we define the (Auslander) δ-invariant δ(M) of M as the maximal rank of a free direct summand of X. We denote by δ n (M) the δ-invariant of n-th syzygy Ωn M of M in the minimal free resolution for n ≥ 0. Lemma 2.1. Let N be a maximal Cohen-Macaulay R-module. Then δ 1 (N) = 0. In particular, δ n (M) = 0 for n ≥ d + 1 and any finitely generated R-module M. Proof. Suppose that δ 1 (N) > 0. Then ΩN has a free direct summand. Let ΩN = X ⊕ R. (σ,τ )T π There is a short exact sequence 0 → X ⊕ R −−−→ R⊕m − → N → 0. According to [12, Lemma 3.1], there exist exact sequences τ (2.1.1) 0→R− → R⊕m → B → 0, (2.1.2) 0 → R⊕m → A ⊕ B → N → 0 for some R-modules A, B. By the sequence (2.1.2), B is a maximal Cohen-Macaulay R-module. In view of (2.1.1), B is a free R-module provided that B has finite projective dimension from (2.1.1). Then, the sequence (2.1.1) splits and τ has a left inverse map. This contradicts that the map π is minimal.  We now remark on δ-invariants under reduction by a regular element. The following lemma is shown in [9, Corollary 2.5]. Lemma 2.2. Let M be a finitely generated R-module and x ∈ m be a regular element on M and R. If 0 → Y → X → M → 0 is an minimal Cohen-Macaulay approximation of M, then 0 → Y /xY → X/xX → M/xM → 0 is a minimal Cohen-Macaulay approximation of M/xM over R/(x). In particular, it holds that δR (M) ≤ δR/(x) (M/xM). In the proofs of our theorems, the following lemma plays a key role. We remark that in the case I = m, similar statements are shown in [5] and [13]. Lemma 2.3. Let l > 0 be an integer, I be an m-primary ideal of R and x ∈ I \ I 2 be an R-regular element. Assume that I i /I i+1 is a free R/I-module for any 1 ≤ i ≤ l and the multiplication map x : I i−1 /I i → I i /I i+1 is injective for any 1 ≤ i ≤ l, where we set I 0 = R. Then the following hold. (1) xI i = (x) ∩ I i+1 for all 0 ≤ i ≤ l. (2) I i /I i+1 ∼ = I i−1 /I i ⊕ I i /(xI i−1 + I i+1 ) for all 1 ≤ i ≤ l. (3) I i /xI i ∼ = I i−1 /I i ⊕ I i /xI i−1 for all 1 ≤ i ≤ l. 4 TOSHINORI KOBAYASHI (4) (I i + (x))/xI i ∼ = R/I i ⊕ I i /xI i−1 for all 1 ≤ i ≤ l. (5) (I i + (x))/x(I i + (x)) ∼ = R/(I i + (x)) ⊕ I i /xI i−1 for all 1 ≤ i ≤ l. Proof. (1): We prove this by induction on i. If i = 0, there is nothing to prove. Let i > 0. The injectivity of x : I i−1 /I i → I i /I i+1 shows that xI i−1 ∩ I i+1 = xI i . By the indcution hypothesis, xI i−1 = (x) ∩ I i . Thus it is seen that xI i = (x) ∩ I i (x) ∩ I i ∩ I i+1 = (x) ∩ I i+1 . (2): As R/I is an Artinian ring, the injective map x : I i−1 /I i → I i /I i+1 of free R/Imodules is split injective. We can also see that the cokernel of this map is I i /(xI i−1 +I i+1 ). Therefore we have an isomorphism I i /I i+1 ∼ = I i−1 /I i ⊕ I i /(xI i−1 + I i+1 ). (3): We have the following natural commutative dialgram with exact rows: 0 I i−1 /I i / x I i /xI i / / I i /xI i−1 0 / =  0 I i−1 /I i /  x  I i /I i+1 / / I i /(xI −1 + I i+1 ) 0 / We already saw in (2) that the second row is a split exact sequence, and thus the first row is also a split exact sequence. Therefore we have an isomorphism I i /xI i ∼ = I i−1 /I i ⊕ I i /xI i−1 . (4): The cokernel of the multiplication map x : R/I i → (I i +(x))/xI i is (I i +(x))/(x) = I i /((x) ∩ I i ), which coincides with I i /xI i−1 by (1). Consider the following commutative diagram with exact rows: 0 x I i−1 /I i / / I i /xI i ι1  0 / R/I i I i /xI i−1 / ι2 / 0 =  x /  (I i + (x))/xI i / I i /xI i−1 / 0 Here ι1 , ι2 are the natural inclutions. The first row is a split exact sequence as in (3). Therefore the second row is also a split exact sequence and we have an isomorphism (I i + (x))/xI i ∼ = R/I i ⊕ I i /xI i−1 . (5): The cokernel of the multiplication map x : R/(I i + (x)) → (I i + (x))/x(I i + (x)) is (I i + (x))/(x) = I i /xI i−1 . We can get the following commutative diagram with exact rows: 0 / R/I i x / (I i + (x))/xI i π1  0 / R/(I i + (x)) I i /xI i−1 / π2 x (I i + (x))/x(I i + (x)) 0 =  / / /  I i /xI i−1 / 0 Here π1 , π2 are the natural surjections. Then we can prove (5) in a manner similar to (4).  In the case that the dimension d is at most 1, the δ-invariants mostly vanish. Lemma 2.4. Assume d ≤ 1 and I is an m-primary ideal of R. If δ(I) > 0, then I is a parameter ideal of R. Proof. Since d ≤ 1, the m-primary ideal I is a maximal Cohen-Macaulay R-module. Therefore the condition δ(I) > 0 provides that I has a free direct summand. We have 5 I = J + (x) and J ∩ (x) = 0 for some ideal J and R-regular element x ∈ I. Let y ∈ J. Then xy ∈ J ∩ (x) = 0. Since x is R-regular, the equality xy = 0 implies y = 0. This shows that J = 0 and I = (x).  Now we can prove Theorem 1.2. Proof of Theorem 1.2. (b) ⇒ (c): If I is a parameter ideal, then Ωd (R/I) = R and hence δ d (R/I) = 1 > 0. (a), (c) ⇒ (b): Assume that δ(R/I) > 0. Then the inequality δ(I) > 0 also holds because I/I 2 is a free R/I-module and thus there is a surjective homomorphism I → R/I. Therefore we only need to prove the implication (c) ⇒ (b) in the case n > 0. We show the implication by induction on the dimension d. If d = 1, then n = 1 by Lemma 2.1. Using Lemma 2.4, it follows that I is a parameter ideal. Now let d > 1. Take x ∈ I \ mI to be an R-regular element. Then the image of x in the free R/I-module I/I 2 forms a part of a free basis over R/I. This provides that the map x : R/I → I/I 2 is injective. We see from Lemma 2.2 that (2.4.1) n−1 n−1 δR/(x) (I/xI) = δR/(x) (ΩR/(x) (I/xI)) n−1 = δR/(x) (ΩR (I) ⊗R R/(x)) n−1 ≥ δR (ΩR I) = δRn (R/I) > 0. Applying Lemma 2.3 (3) to i = 1, we have an isomorphism I/xI ∼ = R/I ⊕I/(x) and hence n−1 n−1 n−1 we obtain an equality δR/(x) (I/xI) = δR/(x) (R/I) + δR/(x) (I/(x)). It follows from (2.4.1) n−1 n−1 that δR/(x) (R/I) > 0 or δR/(x) (I/(x)). Note that the ideal I := I/(x) of R := R/(x) 2 satisfies the same condition as (c), that is, the module I/I is free over R/I = R/I, 2 because I/I = I/((x) + I 2 ) is a direct summand of I/I 2 by Lemma 2.3 (2). By the induction hypothesis, the ideal I is a parameter ideal of R. Then we see that I is also a parameter ideal of R. (c) ⇒ (d): This implication is trivial. Next we prove by induction on d the implication (d) ⇒ (b) when depth grI (R) ≥ d − 1 and I i /I i+1 is a free R/I-module for any i > 0. If d = 1, then δ(I l ) > 0 by Lemma 2.1. By Lemma 2.4, it follows that I l is a parameter ideal. Set (y) := I l . Taking a minimal reduction (t) of I, we have I m+1 = tI m for any m ≫ 0. Setting m = pl, we obtain that I∼ = y p I = I m+1 = tI m = (ty p ). This shows that I is a parameter ideal. Assume d > 1. Since k is infinite, there is an element x ∈ I \I 2 such that the initial form n−1 x∗ ∈ G is a non-zerodivisor of G. We see from Lemma 2.2 that δR/(x) (I l /xI l ) ≥ δRn (R/I l ) > 0 in the same way as (2.4.1). Applying Lemma 2.3 (3), we get an isomorphism I l /xI l ∼ = n−1 n−1 n−1 I l−1 /I l ⊕ I l /xI l−1 and then we see that δR/(x) (I l /xI l ) = δR/(x) (I l−1 /I l ) + δR/(x) (I l /xI l−1 ). n−1 n−1 Since I l−1 /I l is a free R/I-module, we have δR/(x) (R/I) > 0 or δR/(x) (I l /xI l−1 ) > 0. n−1 In the case that δR/(x) (R/I) > 0, we already showed that I is a parameter ideal. So n−1 we may assume that δR/(x) (I l /xI l−1 ) > 0. The equality xI l−1 = I l ∩ (x) in Lemma 2.3 (1) shows that the image I l of I l in R/(x) coinsides with I l /xI l−1 . Thus it holds that l i i+1 n−1 n−1 δR/(x) (I ) = δR/(x) (I l ) > 0. We also note that I /I is free over R/I by Lemma 2.3 (3). 6 TOSHINORI KOBAYASHI By the induction hypothesis, I is a parameter ideal of R/(x). This implies that I is also a parameter ideal of R. Finaly, the implication (b) ⇒ (a) follows from the proof of [10, Theorem 11.42].  Next, to prove Theorem 1.5, we start by recalling the definition of regularity; see [11, Definition 3]. Definition 2.5. Let A be a positively graded homogeneous ring and M be a finitely generated graded A-module. Then the (Castelnuovo-Mumford) regularity of M is defined by regA (M) = sup{i + j|HAi + (M)j 6= 0}. Here we state some properties of regularity. Remark 2.6. Let A and M be the same as in the definition above. (1) Let a ∈ A be a homogeneous M-regular element of degree 1. Then we have regA/(a) (M) = regA (M). (2) If A is an artinian ring, then reg(M) = max{p | Mp 6= 0}. Now let us state the proof of Theorem 1.5. Proof of Theorem 1.5. Since k is infinite, there exists a regular sequence x1 , . . . , xd of R in I such that the sequence of initial forms x∗1 , . . . , x∗d makes a homogeneous system of parameters of grI (R). Then the eqularity grI (R)/(x∗1 , . . . , x∗d ) = grI ′ (R′ ) holds, where R′ = R/(x1 , . . . , xd ) and I ′ = I/(x1 , . . . , xd ). It holds that reg(grI (R)) = reg(grI ′ (R′ )) = max{p| grI ′ (R′ )p 6= 0} = max{p| grI (R)p 6⊂ (x∗1 , . . . , x∗d )} = max{p|I p 6⊂ (x1 , . . . , xd )}. To show the inequality index(I) ≥ regI (R) + 1, it is enough to check that I p ⊂ (x1 , . . . , xd ) if δ(R/I p ) > 0. We prove this by induction on d. Let R be the quotient ring R/(x1 ) and I be the ideal I/(x1 ) of R. Suppose δR (R/I p ) > 0. Then, since there is a surjection from J := I p + (x) to R/I p by Lemma 2.3 (4), δR (J) is greater than 0 . Lemma 2.2 yields that δR (J/x1 J) ≥ δR (J) > 0. Using Lemma 2.3 (5), we obtain an isomorphism J/x1 J ∼ = R/J ⊕ I p /x1 I p−1, and hence δR (J/x1 J) = δR (R/J) + p p−1 δR (I /x1 I ). Therefore we see that δR (R/J) > 0 or δR (I p /x1 I p−1 ) > 0. Now assume that d = 1. If δR (I p /x1 I p−1) > 0, then I p /x1 I p−1 = R since I p /x1 I p−1 = I p /(x1 ) ∩ I p is an ideal of the Artinian ring R and we apply Lemma 2.4. Therefore I p = R and this is a contradiction. So we get δR (R/J) > 0. In this case, R/J must have an R-free summand. This shows that J = (x1 ) and I p ⊂ (x1 ). Next we assume that d > 1. By Theorem 1.2, δR (I p /x1 I p−1 ) = 0. So we have p δR (R/J) > 0. Then R/J = R/(I p + (x1 )) = R/I hold. By the induction hypothep sis, I ⊂ (x2 , . . . , xd )/(x1 ). Hence we get I p ⊂ (x1 , . . . , xd ). It remains to show that index(I) = reg(grI (R)) + 1 if I ⊂ tr(ω). We only need to prove that I p ⊂ (x1 , . . . , xd ) implies δ(R/I p ) > 0. This immidiately follows from the inequalities δ(R/I p ) ≥ δ(R/(x1 , . . . , xd )) and δ(R/(x1 , . . . , xd )) > 0 by applying Theorem 1.2 (b)⇒ (a) to the ideal (x1 , . . . , xd ).  3. examples In this section, R, m, k), and d are the same as in the previous section. Let I be an m-primary ideal of R. To begin with, let us recall the definition of Ulrich ideals. 7 Definition 3.1. We say that I is an Ulrich ideal of R if it satisfies the follwing. (1) grI (R) is a Cohen-Macaulay ring with a(grI (R)) ≤ 1 − d. (2) I/I 2 is a free R/I-module. Here we denote by a(grI (R)) the a-invariant of a(grI (R)). Since k is infinite, the condition (1) of Definition 3.1 is equivalent to saying that I 2 = QI for some/any minimal reduction Q of I. Next, we prove that Ulrich ideals satisfies the assumption of Theorem 1.2 and 1.5. Proposition 3.2. Let I be an Ulrich ideal of R. Then I l /I l+1 is a free R/I-module for any l ≥ 1. Proof. We use induction on l. By definition, I/I 2 is free over R/I. Let l > 1, and take a minimal reduction Q of I. Consider the canonical exact sequence 0 → I l /Ql → Ql−1 /Ql → Ql−1 /I l → 0 of R/Q-modules. Then Ql−1 /Ql is a free R/Q-module and Ql−1 /I l = Ql−1 /IQl−1 = R/I ⊗R/Q Ql−1 /Ql is a free R/I-module. Therefore I l /Ql = ΩR/Q ((R/I)⊕m ) = ΩR/Q (R/I)⊕m = (I/Q)⊕m for some m. Since I/Q is free over R/I, I l /Ql is also a free R/I-module. We now look at the canonical exact sequence 0 → Ql /I l+1 → I l /I l+1 → I l /Ql → 0 of R/I-modules. Then as we already saw, I l /Ql Ql /I l+1 are both free over R/I. Thus the sequence is split exact and I l /I l+1 is a free R/Imodule.  Now we give the proof of Corollary 1.6. Proof of Corollary 1.6. It follows from [4, Theorem 1.1] that index(R) is finite number. Since I is not a parameter ideal, we have δ(R/I) = 0 by Theorem 1.2. If I ⊂ mindex(R) , then we have a surjective homomorphism R/I → R/mindex(R) and thus δ(R/I) ≥ δ(R/mindex(R) ) > 0. This is a contradiction.  To end this section, we give an example of an ideal showing that the condition I 2 /I 3 is free over R/I does not implies that I l /I l+1 is free over R/I for any l ≥ 1. Example 3.3. Let S = k[[x, y]] be the formal power series ring in two variables, n be the maximal ideal of S, L = (x4 )S, J = (x2 , y)S, R = S/L be the quotient ring of S by L and I be the ideal J/L of R. Then I/I 2 is free over R/I but I 2 /I 3 is not so. Proof. We note that J is a parameter ideal of S and therefore J l /J l+1 is free over S/J for any l ≥ 1. Since I 2 = (J 2 + L)/L = J 2 /L, we have I/I 2 = (J/L)/(J 2 /L) = J/J 2 which is free over S/J = R/I. On the other hand, J 3 6⊃ L so I 2 /I 3 6= J 2 /J 3 . But L ⊂ nJ 2 , and hence the minimal number of generators µR (I 2 ) is equal to µS (J 2 ). If I 2 /I 3 is a free R/I-module (hence free over S/J), then I 2 /I 3 and J 2 /J 3 have the same rank µR (I 2 ). However, there is a natural surjective homomorphism φ : J 2 /J 3 → I 2 /I 3 , so φ must be an isomorphism. This contradicts that I 2 /I 3 6= J 2 /J 3 .  Acknowlegments. The author is grateful to his supervisor Ryo Takahashi for giving him helpful advice throughout the paper. 8 TOSHINORI KOBAYASHI References [1] M. Auslander and R. O. Buchweitz, The homological theory of maximal CohenMacaulay approximations, Mem. Soc. Math. France 38 (1989), 5-37. [2] M. Auslander, S. Ding, Ø. Solberg, Liftings and weak liftings of modules, J. Algebra 156 (1993), 273–317. [3] L. L. Avramov, R.-O. Buchweitz, S. B. Iyengar, and C. Miller, Homology of perfect complexes, Adv. in Math. 223 (2010), 1731-1781. [4] S. Ding, A note on the index of Cohen-Macaulay local rings, Comm. Algebra 21 (1993), 53-71. [5] S. Ding, The associated graded ring and the index of a Gorenstein local ring, Proc. Amer. Math. Soc. 120 (1994), 1029-1033. [6] S. Goto, K.Ozeki, R. Takahashi, K. Watanabe, K. Yoshida, Ulrich ideals and modules, Math. Proc. Cambridge Philos. Soc. 156 (2014), no. 1, 137–166. [7] S. Goto, K.Ozeki, R. Takahashi, K. Watanabe, K. Yoshida, Ulrich ideals and modules over two-dimensional rational singularities, Nagoya Math. J. 221 (2016), no. 1, 69–110. [8] M. Hashimoto and A. Shida, Some remarks on index and generalized Loewy length of a Gorenstein local ring, J. Algebra 187 (1997), no. 1, 150-162. [9] K.Yoshida, A note on minimal Cohen-Macaulay approximations, Communications in Algebra 24(1) (1996), 235–246. [10] G. J. Leuschke; R. Wiegand, Cohen-Macaulay Representations, Mathematical Surveys and Monographs, vol. 181, American Mathematical Society, Providence, RI, 2012. [11] A. Ooishi, Castelnuovo’s regularity of graded rings and modules, Hiroshima Math. J. 12 (1982), 627–644. [12] R. Takahashi, Syzygy modules with semidualizing or G-projective summands J. Algebra 295 (2006), no. 1, 179–194. [13] Y. Yoshino, On the higher delta invariants of a Gorenstein local ring, Proc. Amer. Math. Soc. 124 (1996), 2641–2647. Graduate School of Mathematics, Nagoya University, Furocho, Chikusaku, Nagoya, Aichi 464-8602, Japan E-mail address: [email protected]
0math.AC
Strong Algorithms for the Ordinal Matroid Secretary Problem∗ José A. Soto† Abner Turkieltaub‡ Victor Verdugo§ arXiv:1802.01997v1 [cs.DS] 6 Feb 2018 Abstract In contrast with the standard and widely studied utility variant, in the ordinal Matroid Secretary Problem (MSP) candidates do not reveal numerical weights but the decision maker can still discern if a candidate is better than another. We consider three competitiveness measures for the ordinal MSP. An algorithm is α ordinal-competitive if for every weight function compatible with the ordinal information, the expected output weight is at least 1/α times that of the optimum; it is α intersection-competitive if its expected output includes at least 1/α fraction of the elements of the optimum, and it is α probability-competitive if every element from the optimum appears with probability 1/α in the output. This is the strongest notion as any α probability-competitive algorithm is also α intersection, ordinal and utility (standard) competitive. Our main result is the introduction of a technique based on forbidden sets to design algorithms with strong probability-competitive ratios on many matroid classes. In fact, we improve upon the guarantees for almost every matroid class considered in the MSP literature: we achieve probability-competitive ratios of e for transversal matroids (matching Kesselheim et a stronger notion); of 4 for graphic matroids (improving on 2e by Korula al. [29], but under √ and Pál [33]); of 3 3 ≈ 5.19 for laminar matroids (improving on 9.6 by Ma et al. [39]); and of k k/(k−1) for a superclass of k column sparse matroids, improving on the ke result by Soto [44]. We also get constant ratios for hypergraphic matroids, for certain gammoids and for graph packing matroids that generalize matching matroids. The forbidden sets technique is inspired by the backward analysis of the classical secretary problem algorithm and by the analysis of the e-competitive algorithm for online weighted bipartite matching by Kesselheim et al. [29]. p 1/ρ) utility-competitive algorithm for uniform maAdditionally, we modify Kleinberg’s 1 + O( p troids of rank ρ in order to obtain a 1 + O( log ρ/ρ) probability-competitive algorithm. Our second contribution are algorithms for the ordinal MSP on arbitrary matroids. We devise an O(1) intersection-competitive algorithm, an O(log ρ) probability-competitive algorithm and an O(log log ρ) ordinal-competitive algorithm for matroids of rank ρ. The last two results are based on the O(log log ρ) utility-competitive algorithm by Feldman et al. [19]. 1 Introduction In the classical secretary problem (see, e.g., [21] for a survey) an employer wants to select exactly one out of n secretaries arriving in random order. After each arrival, the employer learns the relative merits of the new candidate (i.e., he can compare the candidate with previous ones but no numerical quantity is revealed), and must reject or accept immediately. Lindley [36] and Dynkin [16] show that the strategy of sampling 1/e fraction of the candidates and then selecting the first record has a probability of at least 1/e of selecting the best secretary and that no algorithm can beat this constant. During the last decade, generalizations of this problem have attracted the attention of researchers, specially due to applications in online auctions and online mechanism design. ∗ This work was partially supported by FONDECYT project 11130266, Conicyt PCI PII 20150140 and Núcleo Milenio Información y Coordinación en Redes ICM/FIC RC130003 † Departamento de Ingenierı́a Matemática & CMM, Universidad de Chile. Email: [email protected] ‡ Departamento de Ingenierı́a Matemática, Universidad de Chile. Email: [email protected] § Departamento de Ingenierı́a Industrial, Universidad de Chile & Départment d’Informatique CNRS UMR 8548, École normale supérieure, PSL Research University. Email: [email protected] 1 Arguably the most natural extension is the generalized secretary problem by Babaioff et al. [5]. In their setting, a set R of candidates of known size is presented to an algorithm on uniform random order. On arrival, each element r reveals its hidden weight w(r) and the algorithm must irrevocably decide whether to select it or not while preserving the independence of the set of selected elements on a given independence system1 (R, I). The objective is to maximize the total weight of the selected independent set. Mainly motivated by applications and by the richness of their structural properties, Babaioff et al. focuses on the case in which (R, I) is a matroid 2 , where this problem takes the name of Matroid Secretary Problem (MSP). An α competitive algorithm (which we call α utility-competitive algorithm to differentiate it from other measures) is one that returns an independent set whose expected weight is at least 1/α times the weight of an optimum independent set. Babaioff et al. posts the well-known, and still open, Matroid Secretary Conjecture which in its weak form states that there must be a constant competitive algorithm for the MSP on any matroid, and on its strong form, claims that the constant is e. They also provide a Threshold Price Algorithm (TPA) achieving an O(log ρ) utility-competitive ratio on matroids of rank ρ. Better √ algorithms have improved this ratio to O( log ρ) by Chakraborty and Lachish [11] and later to the current best ratio of O(log log ρ) by Lachish [34] and independently by Feldman et al. [19]. The generalized secretary problem has also been studied on non-matroidal systems such as knapsack [4], online matchings on graphs and hypergraphs [13, 33, 29] and LP packings [30]. Babaioff et al. [5] show that for general independence systems every algorithm on n elements must be Ω(log n/ log log n) utility-competitive and Rubinstein [42] provides an O(log n log ρ) competitive algorithm. Many works consider alternative objective functions, such as minimum sum of ranks [1], time discounted variants [3], convex costs under knapsack type restrictions [6] and submodular objectives [7, 18, 20]. Another rich line of research focuses on relaxing the random order condition and the power of the adversary setting the weights. There are O(1) competitive algorithms for the MSP on the random order and assignment model [44], on the adversarial order random assignment model [41, 44] and on the free order model [25]. Non-uniform arrival orders have also been considered [28]. Related to this line of research is the design of order-oblivious algorithms, which can sample a constant fraction of the elements but afterwards, the arrival order may be arbitrary; and results related to prophet inequality settings on which the weight of each element is a random variable with known distribution but the arrival order is adversarial, see [32, 2, 15, 43]. 1.1 Ordinal MSP versus Utility MSP A common requirement for the mentioned algorithms for the MSP on general matroids is that they require to use numerical weights: they all treat elements with similar weights more or less equivalently. In contrast, in the definition of the classical secretary problem, the decisions made by an algorithm only rely on the ordinal preferences between candidates. This is a very natural condition as for many applications it is difficult to determine (for instance, via an interview) a numerical weight representing each candidate, but it is often simple to compare any two of them. For this reason we restrict our study to the ordinal MSP in which the decision maker can only compare seen elements. This is, it can check which element is higher in an underlying hidden total order ≻. In particular, the algorithm cannot really compute the weight of a set. However, matroids 1 An independence system is a pair (R, I) where R is finite, and I is a nonempty family of subsets of R that is closed under inclusion. The sets in I are called independent sets. 2 A matroid is an independence system (R, I) satisfying the next augmentation property: whenever X, Y ∈ I, and |X| < |Y |, there must be an element r ∈ Y \ X such that X + r ∈ I. 2 have the following extremely nice feature: it is possible to find a maximum weight independent set using only ordinal information by greedily constructing an independent set using the order ≻. The set OPT found this way is the unique lexicographically3 maximum independent of the matroid. So it is quite natural to ask what can be done without numerical weights in the secretary setting. In order to measure the performance of an algorithm we introduce three notions of competitiveness for the ordinal MSP. We say that an algorithm is α ordinal-competitive if for every nonnegative weight function compatible with the total order, the expected weight of the output independent set ALG is at least 1/α the weight of OPT. It is α intersection-competitive if the expected fraction of elements in OPT that is included into ALG is at least 1/α, and it is α probability-competitive if for every element r ∈ OPT, the probability that r is included into ALG is at least 1/α. It is not hard to see that any α ordinal-competitive algorithm is also α competitive in the classical (utility) sense and also that any α probability-competitive algorithm is also α competitive in every other setting. All the mentioned algorithms [5, 11, 34, 19] for the standard or utility MSP do not work in the ordinal model, and so the existence of an α-utility competitive algorithm does not imply the existence of an α probability/intersection/ordinal competitive algorithm. To the authors’ knowledge the only known algorithms for ordinal MSP on general matroids (apart from the trivial O(ρ) probabilitycompetitive that selects the top independent singleton using a classical secretary algorithm) are Bateni et al.’s O(log2 ρ) ordinal-competitive algorithm [7] (which also works for submodular objective functions) and Soto’s variant of TPA [44] which is O(log ρ) ordinal-competitive. Independently and simultaneously from this work, Hoefer and Kodric [23] study the ordinal secretary problem obtaining constant ordinal-competitive algorithms for bipartite matching, general packing LP and independent sets with bounded local independence number. They also show that Feldman and Zenklusen’s reduction [20] from submodular to linear MSP works in the ordinal model. The strong matroid secretary conjecture may hold even for the probability-competitive notion. We note, however, that for some matroid classes the notion of competitiveness matters. In the J-choice K-best secretary problem by Buchbinder et al. [10], one has to select J elements from a randomly ordered stream getting profit only from the top K elements. In particular, every α competitive algorithm for the ρ-choice ρ-best secretary problem that only uses ordinal information, is an α intersection-competitive algorithm for uniform matroids of rank ρ. Chan et al. [12] show that the best competitive ratio achievable by an algorithm that can only use ordinal information on the 2-choice 2-best secretary problem is approximately 0.488628. In contrast, by using numerical information one can achieve a higher competitive ratio of 0.492006 that also works for the sumof-weights objective function (i.e., for utility-competitiveness). Thus showing a gap between the optimal probability ratio and the optimal utility ratio achievable on uniform matroids of rank 2. 1.2 MSP on specific matroids An extensive amount of work has been done on the last decade on the MSP on specific matroid classes including unitary [36, 16, 22, 8, 9], uniform [4, 31], transversal [5, 13, 33], graphic [5, 3, 33], cographic [44], regular and MFMC [14], k-column sparse [44] and laminar matroids [24, 25, 39]. Even though the algorithms are stated for the utility MSP, many of the proofs work directly on either the ordinal MSP under ordinal or even probability competitiveness. We include in Table 1 a summary of all known results together with the improved bounds obtained in this paper. 3 A set A = {a1 , . . . , ak } ⊆ R is lexicographically larger than a set B = {b1 , . . . , bk } ⊆ R of the same size, with respect to the order ≻ on A ∪ B if ai ≻ bi for the first i on which ai and bi differ. 3 Matroid Class Transversal µ exch. gammoids (type of hypermatching) Matching matroids µ exch. matroidal packings Graphic Hypergraphic k-sparse matroids k-framed matroids Semiplanar gammoids Laminar Uniform U (n, ρ) Cographic Regular, MFMC o: u: o: o: o: p: p: o: p: o: Previous guarantees (u,o,p competitive) 16 [13], 8 [33], e [29] O(µ2 ) [33], eµ [29] 16 [5], 3e [3], 2e [33] ke [44] √ 177.77 [24], 3 3e ≈ 14.12 [25] 9.6 [39] e [4] p 1 + O( 1/ρ) 3e [44] 9e [14] New algorithms p-guarantee ref. forb. size e Alg. 1 1 µµ/(µ−1) Alg. 2 µ 4 µ µ/(µ−1) 4 4 k k/(k−1) k k/(k−1) 4/3 √ 4 3 3 ≈ 5.19615 p 1 + O( log ρ/ρ) - Alg. Alg. Alg, Alg, Alg. Alg. Alg. Alg. 3 3 4 4 5 5 6 7 2 µ 2 2 k k 4 3 Alg. 8 - - - Table 1: State of the art competitive ratios for all known matroid classes, including our results. 1.3 Our results and techniques We first formalize our new competitiveness notions for ordinal MSP and study their interrelations. We say that a performance notion is stronger than another if any algorithm that is α competitive for the former is also α competitive for the latter. Under this definition, we show that probability is stronger than ordinal and intersection, and that ordinal is stronger than utility. On the first part of our article we focus on a powerful technique to define strong algorithms for the MSP on many classes of matroids. Informally (see the exact definition in Section 3), we say that an algorithm has forbidden set of size k if it samples without selecting s ∼ Bin(n, p) elements and the following condition holds. Suppose that an element r ∗ of OPT arrives in a position t > s and let Rt be the set of elements that arrived on or before time t. For each time step i between s + 1 and t − 1 there is a random set Fi of at most k forbidden elements such that if for every i, the element arriving at time i is not forbidden (i 6∈ Fi ) then r ∗ is sure to be selected. The following is the key lemma that shows why algorithms with small forbidden sets are useful for the MSP. Lemma 1 (Key Lemma). By setting the right sampling probability p = p(k), every algorithm with forbidden sets of size k is α(k) probability competitive, where ( (1/e, e) if k = 1, (p(k), α(k)) = 1 k (k− k−1 , k k−1 ) if k ≥ 2. Obtaining algorithms with small forbidden sets is simple for many matroid classes. In fact, it is easy to see that the variant of the standard classical secretary algorithm which samples s ∼ Bin(n, p) elements and then selects the first element better than the best sampled element, has forbidden sets of size 1. Suppose that the maximum element r ∗ arrives at time t > s and denote by x the second best element among those arrived up to time t. If x does not arrive at any time between s + 1 and t − 1 then for sure, x will be used as threshold and thus r ∗ will be selected. In the 4 above notation, all forbidden sets Fi are equal to the singleton {x}. Using the key lemma, by setting p = 1/e, this algorithm is e competitive. We provide new algorithms that beat the state of the art guarantees for transversal, graphic, k-sparse and laminar matroids. We also provide new algorithms for other classes of matroids such as matching matroids, certain matroidal graph packings (which generalize matching matroids), hypergraphic matroids, k-framed matroids (which generalize k-sparse matroids), semiplanar gammoids p and low exchangeability gammoids. As an interesting side result, we revisit Kleinberg’s 1 + O( 1/ρ) ordinal-competitive algorithm. We show that its probability-competitiveness is bounded away from 1 and propose an algorithm that achieves p a probability-competitive ratio of 1 + O( log ρ/ρ). Our new results for specific classes of matroids are summarized on Table 1, which, for completeness includes all matroid classes ever studied on the MSP, even those for which we couldn’t improve the state of the art. In the references within the table u, o and p stand for utility, ordinal and probability competitiveness respectively. On the second part of this paper we obtain results for the ordinal MSP on general matroids. Theorem 2. There exists a ln(2/e) intersection-competitive algorithm for the MSP. Theorem 3. There exist an O(log log ρ) ordinal-competitive algorithm and an O(log ρ) probability competitive algorithm for the MSP. Even though Theorem 2 is attained by a very simple algorithm, we note that standard ideas such as thresholding do not work for the intersection notion since elements outside OPT that are higher than actual elements from OPT do not contribute to the objective. The algorithms mentioned in Theorem 3 are based on the recent O(log log ρ) utility-competitive algorithm by Feldman et al. [19]. Their algorithm samples a fraction of the elements so to classify most of the non-sampled ones into h = O(log ρ) weight classes consisting on elements whose weights are off by at most a factor of 2. They implement a clever random strategy to group consecutive weight classes together into buckets, each one containing roughly the same random number of weight classes. On each bucket they define a single random matroid with the property that if one picks an independent set from each one of these matroids, their union is independent in the original one. Their algorithm then selects a greedy independent set on each bucket matroid and output their union. Due to the random bucketing, the expected number of elements selected on each weight class is at least Ω(1/ log h) times the number of elements that OPT selects from the same class. This means that the global algorithm is actually O(log h) = O(log log ρ) utility-competitive. In the ordinal setting we cannot implement this idea in the same way, since basically we do not have any weights. However, we can still partition the elements of the matroid into ordered layers. The idea is to select a collection of thresholds obtained from the optimum of a sample, and use them as separators to induce the layers. This motivates the definition of the Layered-MSP (see Section 4.2). For both ordinal and probability competitiveness we provide a reduction from the Layered-MSP to the ordinal MSP. For the ordinal notion, we select as thresholds a geometrically decreasing subset of the sample optimum according to the value order, so we can partition the matroid into h = O(log ρ) layers. For the probability notion we use all the elements of the sample optimum as thresholds. The crucial result in this part is that our reduction allows to go from any g(h) competitive algorithm for the Layered-MSP to a g(O(1+log h)) ordinal-competitive algorithm, and to a g(O(h)) probability-competitive algorithm. In particular, by applying Feldman et al.’s algorithm, interpreting these layers as weight classes, we get an O(log log ρ) ordinal-competitive algorithm and an O(log ρ) probability-competitive algorithm for the original matroid. 5 1.4 Organization In Section 2 we fix some notation and formally describe the performance guarantees for the ordinal MSP, studying their relations. In Section 3 we prove our key lemma for algorithms with small forbidden sets. We then devise simple algorithms for all the matroid classes mentioned in Table 1. In Section 4 we describe our new algorithms for general matroids, and prove Theorems 2 and 3. To keep the discussion clear and simple, we defer some of the proofs to the Appendix. 2 Preliminaries Let M = (R, I, ≻) be a matroid with ground set R = {r 1 , r 2 , . . . , r n }, and ≻ a total order. We call ≻ the value order and say that r 1 is the highest valued element, r 2 is the second one, and so on, then r 1 ≻ r 2 ≻ · · · ≻ r n . By matroid properties, for every subset Q ⊆ R, there is a unique lexicographically optimum base4 OPT(Q) obtained by applying the greedy algorithm in the order ≻, over the set Q. We say that a nonnegative weight function w : E → R+ is compatible with the value order if r i ≻ r j =⇒ w(r i ) ≥ w(r j ). Note that for every compatible weight function the set OPT = OPT(E) is a maximum weight independent set. We reserve the use of superscripts k ∈ N on a set Q to denote the subset of the highest min{k, |Q|} valued elements of Q. In particular Rk and OPTk denote the set of the top k elements of the matroid and of OPT respectively. We also reserve n and ρ to denote the number of elements of R and its rank respectively. In the (utility/ordinal) MSP, the elements of a (nonnegatively weighted/totally ordered matroid) are presented in uniform random order to an online algorithm that does not know a priori the (weights/value order) of unrevealed elements. At any moment, the algorithm can (view the weight of/compare in the total order) any pair of revealed element. When a new element r is presented, the algorithm must decide whether to add r to the solution and this decision is permanent. The algorithm must guarantee that the set of selected elements is at all times independent5 in the matroid. The objective of the algorithm is to return a set ALG as close as OPT as possible according to certain competitiveness metric. To make notation lighter, we use + and − for the union and difference of a set with a single element respectively. That is, Q+r −e = (Q∪{r})\{e}. The rank of a set S is the cardinality of its bases, ρ(Q) = max{|I| : I ∈ I, I ⊆ Q}, and the span of Q is span(Q) = {r ∈ R : ρ(Q + r) = ρ(Q)}. In matroids, OPT has the property of improving any subset in the following sense. Lemma 4. Let Q ⊆ R. Then, OPT ∩ Q ⊆ OPT(Q). Proof of Lemma 4. Let r = r k be an element of OPT ∩ Q. Since r k is selected by the Greedy algorithm we have that r 6∈ span(Rk−1 ). But then r 6∈ span(Rk−1 ∩ Q) and so it is also selected by the Greedy algorithm applied only on the set Q. Therefore, r ∈ OPT(Q). Recall that the algorithm considered for the MSP have access to the elements in an online and uniformly at random fashion. We denote by r1 the first element arriving, r2 the second, and so on. In general, Rt = {r1 , r2 , . . . , rt } is the set of elements arriving up to time t. 4 A base of a set Q is a maximal independent subset X ⊆ Q. For particular classes of matroids, we may assume that M is known beforehand by the algorithm, or alternatively that it is discovered by the algorithm via an independence oracle that allows it to test any subset of revealed elements. In any case, it is a standard assumption that the algorithm at least know the number of elements n in the matroid. 5 6 Utility competitiveness. competitive if An algorithm for the utility MSP returning a set ALG is α ≥ 1 utilityE[w(ALG)] ≥ w(OPT)/α. 2.1 (1) Measures of competitiveness for the ordinal MSP We introduce three measures of competitiveness or the ordinal MSP. Recall that the algorithm only learns ordinal information about the elements but it cannot access numerical weights. In this sense, they are closer to the classical secretary problem than the utility variant (for a discussion about this aspect in the original secretary problem, see [21]). Since the weight function remains completely hidden for the algorithm, the first measure of competitiveness we consider is the following. An algorithm is α ordinal-competitive if for every weight function w compatible with the value order, condition (1) holds. An equivalent characterization of competitiveness is obtained by the following lemma. Lemma 5. An algorithm is α ≥ 1 ordinal-competitive if and only if for every k ∈ [n], E|ALG ∩ Rk | ≥ E|OPT ∩ Rk |/α. (2) Proof. Consider an α ordinal-competitive algorithm returning a set ALG and let k ∈ [n]. Define the weight function w(r) = 1 if r ∈ Rk and zero otherwise, which is compatible with the value order. Then, E[|ALG ∩ Rk |] = E[w(ALG)] ≥ α1 w(OPT) = α1 |OPT ∩ Rk |. Now suppose that (2) holds. Then, for any compatible function w, and defining w(r n+1 ) := 0, E[w(ALG)] = n X k=1 ≥ n X k=1 (w(r k ) − w(r k+1 )) · E[|ALG ∩ Rk |] (w(r k ) − w(r k+1 )) · 1 1 E[|OPT ∩ Rk |] = E[w(OPT)]. α α In the second measure we consider, we want to make sure that every element of the optimum is part of the output with large probability. We say that an algorithm is α ≥ 1 probability-competitive if for every e ∈ OPT, Pr(e ∈ ALG) ≥ 1/α. (3) Finally, in the third measure we want to maximize the number of elements in the optimum that the algorithm outputs. We say an algorithm is α ≥ 1 intersection-competitive if E[|OPT ∩ ALG|] ≥ |OPT|/α. (4) Relation between variants. The ordinal and probability measures are in fact stronger than the standard utility notion. That is, any α ordinal/probability-competitive algorithm yields to an α utility-competitive algorithm. Furthermore, the probability is the stronger of them all. Lemma 6. If an algorithm is α ordinal-competitive then it is α utility-competitive. If an algorithm is α probability-competitive then it is also α ordinal, utility and intersection-competitive. 7 Proof of Lemma 6. If an algorithm is α ordinal-competitive then by definition it is α utilitycompetitive. Now consider an α probability-competitive algorithmPreturning a set ALG. For any k ∈ [n], we have E[|ALG ∩ Rk |] ≥ E[|ALG ∩ OPT ∩ Rk |] = r∈OPT∩Rk Pr(r ∈ ALG) ≥ |OPT∩Rk | . α which means that the algorithm is α ordinal-competitive, and therefore, it is also α utility-competitive. To see that the algorithm is also α intersection-competitive we note that E[|ALG ∩ OPT|] = X e∈OPT Pr(e ∈ ALG) ≥ |OPT| . α An algorithm for the utility variant, which recall is able to use the elements’ weights, may not be adapted to the ordinal MSP, since in this model it can not use any weight information. In other words, the existence of an α utility-competitive for a matroid (or matroid class) does not imply the existence of an α-competitive algorithm for any of the other three measures. It is worth noting that the intersection-competitive measure is incomparable with the other measures. It is not hard to find fixed families of instances for which a given algorithm is almost 1 intersection-competitive but has unbounded utility/ordinal/probability-competitiveness. There are also examples achieving almost 1 ordinal-competitiveness but unbounded intersection-competitive ratio. 3 Improved algorithms for specific matroids In this section we describe a powerful forbidden sets technique to analyze algorithms for the ordinal MSP. Thanks to this technique we devise algorithms for many matroid classes previously studied in the context of MSP and to other matroids that have not been studied in this context. Our results improve upon the best known competitive ratios for almost all studied matroid classes. The only classes in which we do not find an improvement are cographic matroids for which there is already a 3e probability-competitive algorithm [44], and regular-and-MFMC matroids for which there is a 9e this section we briefly study uniform matroids, ordinal-competitive algorithm [14]. At the end of p for which Kleinberg’s algorithm [31] yields a 1 + O( 1/ρ) ordinal-competitive guarantee. We show that the probability-competitiveness of Kleinberg’s algorithm is at least 4/3 and we propose a p variant that achieves a slightly weaker ratio of 1 + O( log ρ/ρ) in the probability notion. 3.1 Forbidden sets technique The following is the key definition that allows us to devise constant probability-competitive algorithms for specific classes of matroids. Recall that |R| = n, and for t ∈ [n], rt is the random element arriving at time t, and Rt is the random set of elements arriving at or before time t. Definition 7. An algorithm has forbidden sets of size k if it has the following properties. 1. (Correctness) The algorithm returns an independent set ALG. 2. (Sampling property) It chooses a sample size s at random from Bin(n, p) for some fixed sampling probability p, and it does not accept any element from the first s arriving ones. 3. (k-forbidden property) For every triple (X, Y, r ∗ ) with Y ⊆ R, r ∗ ∈ OPT(Y ) and X ⊆ Y − r ∗ , one can define a set F(X, Y, r ∗ ) ⊆ X of at most k forbidden elements of X such that the following condition holds. Let t ≥ s + 1 be a fixed time. If rt ∈ OPT(Rt ) and for every j ∈ {s + 1, . . . , t − 1}, rj 6∈ F(Rj , Rt , rt ) then rt is selected by the algorithm. 8 To better understand the k-forbidden property suppose that a fixed element r ∗ ∈ OPT arrives at step t ≥ s + 1, that is rt = r ∗ . Note that the set Rt−1 of elements arriving before rt is a random subset of size t − 1 of R − rt , and no matter the choice of Rt−1 , r ∗ is always part of OPT(Rt ) = OPT(Rt−1 + rt ). Inductively, for j = t − 1 down to j = 1, once Rj is specified, rj is a uniform random element of Rj , and Rj−1 is defined as Rj − rj . Moreover, this choice is independent of the previous random experiments (i.e., the choices of {rj+1 , . . . , rt−1 , Rt−1 }). The k-forbidden property (3.) says that if for every j ∈ {s + 1, . . . , t − 1} element rj is not a forbidden element in F(Rj , Rt , rt ) then r ∗ is guaranteed to be selected by the algorithm. Designing algorithms with small forbidden sets is the key to achieve constant competitiveness as our key Lemma (that we restate below) shows. Lemma 1 (Key Lemma). By setting the right sampling probability p = p(k), every algorithm with forbidden sets of size k is α(k) probability competitive, where ( (1/e, e) if k = 1, (p(k), α(k)) = k 1 (k− k−1 , k k−1 ) if k ≥ 2. Proof. Fix an element r ∗ from OPT and condition on the realization of s ∼ Bin(n, p), on the time t on which r ∗ = rt arrives and on the set Y = Rt ∋ rt of the first t elements arriving. Abbreviate Pt (·) = Pr(·|rt = r ∗ , Rt = Y, s). By Lemma 4, r ∗ ∈ OPT(Rt ) and by the k-forbidden property, Pt (r ∗ ∈ ALG) ≥ Pt (For all j ∈ {s + 1, . . . , t − 1}, rj ∈ Rj \ F(Rj , Y, r ∗ )) t−1 t−1 Y j − k Y ∗ Pr(rj ∈ Rj \ F(Rj , Y, r )) ≥ , = j + j=s+1 j=s+1 where x+ = max{0, x}. The equality above holds because of the independence of the random experiments defining iteratively rt−1 , rt−2 , down to rs+1 as mentioned before the statement of this lemma. By removing the initial conditioning we get n 1 X Pr(r ∈ ALG) ≥ Es∼Bin(n,p) n ∗ t−1 Y t=s+1 j=s+1  k 1− j  . (5) + To compute the right hand side we use the following auxiliary process. Suppose that n people participate in a game. Each player x arrives at a time τ (x) chosen uniformly at random from the interval [0, 1]. Each person arriving after time p selects a subset of k partners from the set of people arriving before them, without knowing their actual arrival times (if less than k people have arrived before her, then all of them are chosen as partners). A player wins if she arrives after time p and every one of her partners arrived before time p. Since the arrival times are equally distributed and the event that two people arrive at the same time has zero probability the arrival order is uniform among all possible permutations. Furthermore, the number of people arriving before time p distributes as Bin(n, p). Using these facts, the probability that a given person x wins is exactly the right hand side of (5). But we can also compute this probability using its arrival time τ (x) as Z 1 p Pr( all partners of x arrived before time p | τ (x) = τ ) dτ ≥ 9 Z 1 p (p/τ )k dτ, which holds since the arrival time of each partner of x is a uniform random variable in [0, τ ], and conditioned on τ , each partner arrives before time p with probability p/τ . Since x may have less than k partners, we don’t necessarily have equality. We conclude that for every r ∗ ∈ OPT,  Z 1 −p ln(p), if k = 1, k ∗ (p/τ ) dτ = p − pk Pr(r ∈ ALG) ≥  p , if k ≥ 2. k−1 By optimizing the value of p as a function of k, we obtain that the probability of r ∗ ∈ ALG is 1/α(k), with α(k) as in the statement of the lemma and the probability achieving it is p = p(k). The idea of analyzing an algorithm as a series of stochastically independent experiments which defines the reverse arrival sequence appears very early in the history of the secretary problem. As mentioned in the introduction, one can prove that the algorithm for the classical secretary algorithm that samples s ∼ Bin(n, p(1)) elements and then selects the first element better than all the sampled ones is e probability-competitive showing that it has forbidden sets of size 1. In our notation, for each (X, Y, r ∗ ) with r ∗ the maximum element of Y and X ⊆ Y − r ∗ , define the forbidden set F(X, Y, r ∗ ) as the singleton OPT(X). The 1-forbidden condition states that if the element rt arriving at time t is a record (it is in OPT(Rt )), and if for every time j ∈ {s + 1, . . . , t − 1}, rj 6∈ OPT(Rj ) (i.e., the j-th arriving element is not a record), then rt will be chosen by the algorithm. Since all forbidden sets have size at most 1, setting the sampling probability to be p(1) = 1/e guarantees the probability-competitive ratio of α(1) = e. Kesselheim et al.’s algorithm for online bipartite matching [29] uses a technique very similar to that in the previous lemma to compute the expected weight contribution of the k-th arriving vertex, which is enough to prove asymptotically e + O(1/n) utility-competitiveness for online bipartite matchings. In fact, for transversal matroids, this yields to an e + O(1/n) ordinal-competitive algorithm. But their analysis does not imply any guarantee on the probability notion. In the next section we show by using a different analysis, that their algorithm has forbidden sets of size 1 and so it is actually e probability-competitive (always, not just asymptotically). It turns out that many classes of matroids behave very similarly to transversal matroids in the following sense. A set X is independent on a given matroid if and only if each element r ∈ X can be mapped to an object (e.g., an edge covering r, a path ending in r, a subgraph covering r, etc.) r̃ such that the set X̃ = {r̃ : r ∈ X} satisfies a combinatorial property (e.g., a matching covering X, a collection of edge/node disjoint paths connecting X with some source, a collection of disjoint subgraphs covering X). We call X̃ a witness for the independence of X. A set X may have multiple witnesses, but we will always assume that there is a canonical witness, witness(X), that can be computed by the algorithm and furthermore, the choice of the witness cannot depend on the set of elements seen so far nor on its arrival order. The fact that the witnesses do not depend on the arrival order makes them amenable to the analysis by the reverse arrival sequence analysis above, which in turn, will help us to devise algorithms with constant-size forbidden sets. 3.2 Transversal matroids and Gammoids Transversal matroids. Let G = (L ∪ R, F ) be a bipartite graph with color classes L and R, where the elements of R are called the terminals of G. The transversal matroid T [G] associated to G is the matroid with ground set R whose independent sets are those X ⊆ R that can be covered by a 10 matching in G. We call G the transversal presentation of T [G]. Gammoids. Let G = (V, E) be a digraph and two subsets S, R ⊆ V called sources and terminals respectively, which do not need to be disjoint. The gammoid Γ(G, S, R) is the matroid over the terminals where X ⊆ R is independent if X is linked to S, that is, if there are node-disjoint directed paths starting from S and ending on each element of X. We say that (G, S, R) is the gammoid presentation of the matroid. Note that transversal matroids are gammoids by declaring all the nonterminals as sources and directing the arcs in the transversal presentation from sources to terminals. Transversal and Gammoid MSP. In the Transversal MSP, a transversal presentation G for an unknown ordered matroid M = T [G] is either revealed at the beginning of the process, or it is revealed online in the following way. Initially, the algorithm only knows the number of terminals. Terminals arrive in random order and whenever r ∈ R arrives, it reveals its ordinal value information and the set of its neighbors in L. In the gammoid MSP, a gammoid presentation (G, S, R) for an unknown gammoid is either revealed at the beginning or it is revealed online as elements from R arrive: when a terminal r ∈ R arrives all possible S-r paths are revealed. At that time the algorithm has access to the subgraph Gt ⊆ G only containing the arcs belonging to every possible S-Rt path and can test whether a vertex is in S or not. Most of the known algorithms for transversal MSP [5, 13, 33] work with ordinal information. The best algorithm so far, by Kesselheim et al. [29] achieves an asymptotically optimal utilitycompetitive ratio of e + O(1/n) for the more general (non-matroidal) vertex-at-a-time bipartite online matching problem, in which edges incident to the same arriving vertex may have different weights and the objective is to select a matching of maximum total weight. For the specific case of transversal matroid this algorithm can be implemented in the ordinal model, meaning that is e + O(1/n) ordinal-competitive. Interestingly, for the broader bipartite matchings case, Kesselheim’s algorithm does not work in the ordinal model, but the previous algorithm by Korula and Pál [33] does, achieving 8 ordinal-competitiveness. A recent result by Hoefer and Kodric [23] improves this factor to 2e ordinal-competiive for bipartite matchings. Exchangeability parameter for gammoids. Let us define a parameter to control the competitiveness of our algorithm for the gammoid MSP. Let X be an independent set and let Q be a path linking a terminal r ∈ R \ X outside X to S. The exchangeability µ of the presentation (G, S, R) is the maximum number of paths in PX that Q intersects. The intuition behind is the following: in order to include Q into PX while keeping disjointness we have to remove or exchange at least µ paths from PX . For instance, if we define the diameter d of the gammoid presentation as the maximum number of nodes in any source-terminal path, then µ is at most d. If furthermore the terminals are sinks, that is out-degree 0, then µ is at most d − 1, since paths ending at different terminals cannot intersect on a terminal. Remark 8. This is the case for transversal matroids: their diameter in the gammoid presentation is 2 and their exchangeability is 1. For our results, we assume the algorithm also knows an upper bound µ for the exchangeability parameter, and in this case we call the problem µ-gammoid MSP or bounded exchangeability gammoid MSP. The µ-gammoid MSP problem is a special case of the (non-matroidal) hypergraph vertex-at-atime matching (HVM) with edges of size at most µ + 1 studied by Korula and Pál [33] and later by 11 Kesselheim et al.’s [29] online hypermatching problem (see the discussion in those papers for precise definitions). They achieve O(µ2 ) and eµ utility-competitiveness respectively. Below we propose an algorithm for the µ-gammoid MSP that has forbidden sets of size µ. Provided we know µ upfront we get an α(µ) probability-competitive algorithm. Note that for µ ≥ 2, α(µ) = µ1+1/(µ−1) < eµ, so our guarantee strictly improves on that of previous algorithms for HVM and hypermatching, on the special case of µ-gammoids. Our algorithms. We use the convention that for every vertex v covered by some matching M , M (v) denotes the vertex matched with v in M . Furthermore, for every independent set X ⊆ R, we select canonically a witness matching MX := witness(X) that covers X. In the case of gammoids, for any set P of node-disjoint paths linking some set X to S, and for every v ∈ X, P(v) denotes the unique path in P linking v to S. We also say that P covers a vertex u if u is in the union of the vertices of all paths in P. Furthermore, for every independent set X ⊆ R, we canonically select a fixed collection of node-disjoint S-X directed paths PX := witness(X) linking X to S, and we assume that this choice does not depend on the entire graph but only on the minimum subgraph containing all arcs in every S-X path. We also recall that on step i, Ri = {r1 , . . . , ri } denotes the set of revealed terminals and Gi denotes the subgraph of the presentation currently revealed. Algorithm 1 for transversal matroids. Algorithm 2 for µ-bounded gammoids. Input: Presentation of a transversal matroid T [G] Input: Presentation of a gammoid Γ := Γ(G, S, R) whose terminals arrive in random order. whose terminals arrive in random order. ⊲M and ALG are the currently chosen matching ⊲P and ALG are the currently chosen collection and right vertices respectively. of node-disjoint paths and terminals selected respectively. 1: ALG ← ∅, s ← Bin(n, p), M ← ∅ 1: ALG ← ∅, s ← Bin(n, p), P ← ∅. 2: for i = s + 1 to n do 2: for i = s + 1 to n do 3: if ri ∈ OPT(Ri ) and ℓi := MOPT(Ri ) (ri ) is 3: if ri ∈ OPT(Ri ) and no vertex in the path not covered by M then POPT(Ri ) (ri ) is covered by P then 4: ALG ← ALG + ri , M ← M ∪ {ℓi ri } 4: ALG ← ALG+ri , P ← P∪{POPT(Ri ) (ri )} 5: Return ALG. 5: Return ALG. The algorithms above can compute OPT(Ri ) without knowing the terminal weights, by just using the greedy algorithm. This requires that one is able to check independence algorithmically in each case. Indeed, for the transversal MSP algorithm, a set X ⊆ Ri is independent if and only if the maximum cardinality matching on Gi [NG (X) ∪ X] has size |X|. In the case of gammoids, one can check if X ⊆ Ri is independent, by a standard reduction to a flow problem on Gi . Theorem 9. Algorithm 2 has forbidden sets of size equal to the exchangeability µ of the gammoid presentation. If µ is known, we can set p = p(µ) to get an α(µ) probability-competitive algorithm. Proof. By construction, the set P contains node-disjoint paths covering ALG at every time step, hence the algorithm is correct. The sampling condition is also satisfied by design. Let r ∗ ∈ OPT(Y ) where Y is a fixed set of terminals of size t ≥ s + 1, and suppose that Rt = Y and rt = r ∗ . Note that rt is selected by the algorithm if all vertices in the path POPT(Rt ) (rt ) are not covered by the collection P prior to that iteration. In other words, by defining the forbidden sets to be F(X, Y, r ∗ ) = {v ∈ OPT(X) : POPT(X) (v) intersects POPT(Y ) (r ∗ )}, 12 the element rt is selected if rj 6∈ F(Rj , Rt , rt ) for all j ∈ {s + 1, . . . , t − 1}. By definition, each forbidden set has size at most µ. Algorithm 1 is essentially Algorithm 2 applied over the gammoid presentation of the transversal matroid T [G]. Together with Remark 8, it follows the result for the transversal MSP. Theorem 10. Algorithm 1 has forbidden sets of size 1, and therefore, by choosing p = 1/e, it is an (optimal) e probability-competitive for transversal matroids. We remark that every constant utility-competitive algorithm so far for transversal MSP requires to learn a bipartite presentation online. It is an open problem to find constant competitive algorithms for transversal matroids that only access the matroid via an independence oracle. 3.3 Matching Matroids and Matroidal Graph Packings Matroidal graph packings. Let H be a finite set of graphs. An H-packing of a host graph G = (V, E) is a collection Q = {Hi }i=1...k of node-disjoint subgraphs of G such that each H ∈ Q is isomorphic to some graph in H. A vertex in G is said to be covered by Q if it belongs to some graph of Q. Let R ⊆ V be a set of vertices called terminals and consider the independence system M(R, G, H) over R whose independent sets are all X ⊆ R for which there is an H-packing covering X in G. We say that H is matroidal if M(V (G), G, H) defines a matroid for every graph G. Note that in this case, if M(R, G, H) is the restriction of M(V (G), G, H) to a subset R, then it is also a matroid. We call every such M(R, G, H) a graph packing matroid. Matroidal families. For an extensive treatment see Loebl and Poljak [37], Janata [27] and the references therein. In the following examples, Kn denotes the complete graph on n vertices and Sn denotes the star with n legs. If H = {K1 } then M(V, G, H) is the free matroid over V where all sets are independent. If H = {K2 } then M(V, G, H) is the matching matroid of G, whose independent sets are all vertex sets that can be covered by a matching. If for some k the family H = {S1 , S2 , . . . , Sk } is a sequential sets of stars, then M(V, G, H) is matroidal. It is, in fact, the matroid union of many matching matroids. The family M(V, G, H) is also matroidal if H = {K2 , H} where H is either a factor-critical graph or a 1-propeller.6 For most known matroidal classes there are polynomial time algorithms available to check independence. This is the case for all classes above. Observe that transversal matroids are instances of matching matroids restricted to one side of the bipartition of the host graph. As we did for gammoids, we also define an exchangeability parameter µ to control the competitiveness of our algorithm. Consider an H-packing Q of G, and a subgraph H ⊆ G from the class H covering a terminal r ∈ R that is not covered by Q. The exchangeability µ of M(R, G, H) is the maximum number of terminals over all such Q and H that would become uncovered if we removed all graphs from Q that intersect H, namely,     X µ := max |V (H ′ ) ∩ R| : H covers a terminal not covered by Q .  ′  ′ H ∈Q : V (H)∩V (H )6=∅ 6 A factor-critical graph is one such that H − x admits a perfect matching for all x ∈ V (H). A 1-propeller is a graph having a leaf r, and a vertex c such that for every x ∈ V (H) − c, H − x admits a perfect matching. 13 This parameter may be complicated to compute but there is a simple upper bound: let h be the maximum number of vertices of a graph from H. Then the worst possible situation occurs when Q contains only graphs of size h, and H is also a graph of size h intersecting every graph in Q on exactly one vertex (different from r). In this case, all graphs from Q must be removed, so the number of newly uncovered vertices becomes h · (h − 1). This means that µ ≤ h(h − 1). In the H-Packing MSP, the algorithm receives a collection H of graphs. A host graph G = (V, E) with terminals R ⊆ V is either revealed at the beginning or it is revealed online as elements from R arrive: when a terminal r ∈ R arrives, all possible edges that belong to a graph H ⊆ G with r ∈ V (H) with H ∈ H (via isomorphism) are revealed. More precisely, let Rt denote the set of terminals revealed up to time t. At that time the algorithm has access to the subgraph Gt ⊆ G induced by all vertices belonging to every possible subgraph H ⊆ G, with H ∈ H that intersects Rt . The algorithm can also test whether a vertex is a terminal or not. We also assume that an upper bound µ for the exchangeability parameter is available, and in this case we call the problem bounded H-packing MSP. Analogously to previous sections, for every independent X we select a canonical packing QX := witness(X) that does not depend on the arrival order. In the description of the algorithm we use the convention that for every H-packing Q of a set X, and for every v ∈ V covered by Q, Q(v) denotes the unique graph in Q covering v. We also recall that on step i, Ri = {r1 , . . . , ri } denotes the set of revealed terminals and Gi denotes the subgraph of the presentation currently revealed. Algorithm 3 for µ bounded H-packing matroids. Input: A matroidal family H and a host graph G = (V, E) whose terminals R ⊆ V arrive in random order. ⊲ Q and ALG are the currently chosen H-packing and terminals selected respectively. 1: ALG ← ∅, s ← Bin(n, p), Q ← ∅. 2: for i = s + 1 to n do 3: if ri ∈ OPT(Ri ) and ri is already covered by Q. then 4: ALG ← ALG + ri . 5: else if ri ∈ OPT(Ri ) and Q ∪ {QOPT(Ri ) (ri )} is an H-packing. then 6: ALG ← ALG + ri , Q ← Q ∪ {QOPT(Ri ) (ri )}. 7: Return ALG. The algorithm can compute OPT(Ri ) without knowing the terminal weights, by applying the greedy algorithm for M(R, G, H). Theorem 11. Algorithm 3 has forbidden sets of size equal to the exchangeability µ of the Hpresentation. If µ is known beforehand, we can set p = p(µ) to obtain an α(µ) probabilitycompetitive algorithm for µ-bounded graph packing matroids. Proof. Correctness and the sampling condition for forbidden sets are satisfied by design. Now let r ∗ ∈ OPT(Y ) where Y is a fixed set of terminals of size t ≥ s + 1, and suppose that Rt = Y and rt = r ∗ . Terminal rt is selected by the algorithm if either Q already covers it on arrival, or if all vertices in the graph QOPT(Rt ) (rt ) were not covered by graphs in Q prior to that iteration. In any case by defining as forbidden sets F(X, Y, r ∗ ) = {v ∈ OPT(X) : V (QOPT(X) (v)) intersects V (QOPT(Y ) (r ∗ )) \ {r ∗ }}, we have that rt is selected if rj 6∈ F(Rj , Rt , rt ) for all j ∈ {s + 1, . . . , t − 1}. By definition, each forbidden set has size at most µ. 14 We remark that the competitiveness achievable by an algorithm heavily depends on the way the matroid is presented. For instance, consider a matching matroid M with host graph G = (V, E). Note that the exchangeability of this matroid is µ ≤ 2(2 − 1) = 2. If the graph is presented online then we can use the previous algorithm to obtain an α(2) = 4 probability-competitive algorithm. If on the other hand the graph is presented upfront, then we can use the fact that every matching matroid is transversal [17] to construct a transversal presentation of M and then apply our e probability-competitive algorithm using that presentation. 3.4 Graphic and Hypergraphic Matroids Graphic and hypergraphic matroids. The graphic matroid M[G] = (R, I) associated to a graph G = (V, R) is the one whose independent sets are all the subsets of edges X ⊆ R such that (V, X) is a forest. The hypergraphic matroid M[G] associated to a hypergraph G = (V, R), whose edges may be incident to any number of vertices (if all the edges have size 1 or 2 we are back in the graphic case), is the matroid over R whose independent sets X ⊆ R are those, for which one  can canonically choose for every r ∈ X an edge denoted by edge(r, X) = u(r)v(r) in KV = (V, V2 ) with both endpoints in r in such a way that all edge(r, X), for r ∈ X are different, and the collection edge(X) = {edge(r, X) : r ∈ X} is a forest [38]. One can check that the hypergraphic matroid M[G] = (R, I) is the matroid induced from the graphic matroid M[KV ] via the bipartite graph (R ∪ V2 , Ẽ) with ef ∈ Ẽ if f ⊆ e. In other words, X is independent in the hypergraphic matroid M[G] if edge(X) is independent in the graphic matroid M[KV ]. Moreover, if G is already a graph, edge(X) = X and edge(r, X) = r for all r ∈ X. In the graphic MSP /hypergraphic MSP we assume that the underlying graph/hypergraph G of a matroid M[G] is either revealed at the beginning or revealed online in the natural way: we learn edges as they arrive. Let X ⊆ R be an independent set. By orienting each connected component of the forest edge(X) from an arbitrary root, we obtain a canonical orientation arc(X) of edge(X) (for convenience, we denote by arc(e, X) the oriented version of edge(e, X)) with indegree deg− arc(X) (v) ≤ 1 for every vertex v. The converse is almost true in the following sense. If A is a set of arcs (maybe including loops) such that deg− A (v) ≤ 1 for every vertex then the underlying graph is not necessarily a forest in KV , but a pseudoforest: every connected component contains at most 1 cycle, which is directed. In fact, the edge sets of pseudoforest of a given graph J are exactly the independent set of the so called bicircular matroid of J. This matroid is transversal with presentation H, where V (H) = V (J) ∪ E(J) and ve ∈ E(H) if and only if e is incident to v. This is the starting point for our algorithm for graphic matroids. The algorithm. The plan is to only consider edges that belong to the current optimum. Furthermore, if we select an edge, then we orient it and include it into an arc set A with the property that each vertex has maximum in-degree 1. Instead of using a random orientation (as in the algorithms by Korula and Pál [33] or Soto [44]), at every step we use the canonical orientation of the current optimum forest. In order to avoid closing a cycle we also impose that an arc (u, v) can not be − added to A if deg− A (u) = 1 or degA (v) = 1. The same algorithm works on hypergraphic matroids if we replace each independent set X on the hypergraphic matroid by its associated forest edge(X). We also recall that on step i, Ri = {r1 , . . . , ri } denotes the set of revealed edges. The algorithm is fully described below. 15 Algorithm 4 for graphic or hypergraphic matroids. Input: A hypergraphic matroid M[G] with underlying hypergraph G = (V, R), whose edges arrive in random order. ⊲ ALG and A are the currently selected independent set and the orientation of its associated forest. 1: ALG ← ∅, s ← Bin(n, p), A ← ∅. 2: for i = s + 1 to n do 3: if ri ∈ OPT(Ri ) then 4: Let ai = (ui , vi ) = arc(ri , OPT(Ri )) be the canonical orientation of edge(ri , OPT(Ri )). − 5: if deg− A (ui ) = 0 = degA (vi ) then 6: ALG ← ALG + ri , A ← A + ai 7: Return ALG. Theorem 12. Algorithm 4 has forbidden sets of size 2. By setting p = p(2) = 1/2, we get an α(2) = 4 probability-competitive algorithm for both graphic and hypergraphic matroids. Proof. We first prove that the edge set  obtained from A by removing its orientation is acyclic. Suppose by contradiction that at the end of some step i, A contains for the first time a set C such that its unoriented version Ĉ is an undirected cycle. Since this is the first time a cycle appears, ri must be selected and ai = (ui , vi ) must be contained in C. Since ai is included in A, we know that − after its inclusion, deg− A (vi ) = degC (vi ) = 1 (before its inclusion, the indegree of vi was 0) and that − + − deg− A (ui ) = degC (ui ) = 0. But since Ĉ is a cycle the outdegree of ui is degC (ui ) = 2− degC (ui ) = 2. But then, there must be another vertex x in C with indegree 2. This cannot happen because at every moment the indegree of each vertex is at most 1. The proof above guarantees correctness of the algorithm: for the graphic case  = ALG and for the hypergraphic case, each edge ri of ALG is mapped to edge(ri , OPT(Ri )) ∈  which form a forest. In both cases we conclude ALG is independent. Since the sampling condition is satisfied by design, we only need to prove the 2-forbidden condition. Let r ∗ ∈ OPT(Y ) where Y is an arbitrary set of t ≥ s + 1 edges, and suppose that Rt = Y and rt = r ∗ . The algorithm would then define an arc at = (ut , vt ) and will proceed to add rt to ALG provided that no arc aj = (uj , vj ) considered before has head vj equal to ut or vt . In other words, by defining n o arc(f,OPT(X)) is not oriented F(X, Y, r ∗ ) = f ∈ OPT(X) : towards ∗ any endpoint of edge(r ,OPT(Y )) then rt is selected if rj 6∈ F(Rj , Rt , rt ) for all j ∈ {s + 1, . . . , t − 1}. Moreover, since each arc set arc(OPT(X)) = {arc(f, OPT(X)) : f ∈ OPT(X)} has maximum indegree 1, there are at most 2 arcs in arc(OPT(X)) oriented towards an endpoint of edge(r ∗ , OPT(Y )). So each forbidden set has size at most 2. 3.5 Column Sparse Representable Matroids and Multi-Framed matroids Column sparse matroids. An interesting way to generalize graphic matroids is via their matrix representation. We say that a matroid M is represented by a m × n matrix M with coefficients in a field F if we can bijectively map the elements of its ground set to the columns of M in such a way that the independent sets of M are in correspondence with the sets of columns that are linearly independent in Fm . Graphic matroids are representable by their adjacency matrix interpreted in GF (2). In fact they can be represented in any field. Matroids that have this property are called 16 regular. Note that each graphic matroid is representable by a very sparse matroid: each column has only 2 non-zero elements. Following [44] we say that a matroid is k column sparse representable if it admits a representation whose columns have at most k nonzero entries each. These matroids include many known classes such as graphic matroids (k = 2), rigidity matroids [47] on dimension d (k = 2d) and more generally matroids arising from rigidity theory from d-uniform hypergraphs. These matroids are called (k, ℓ)-sparse matroids (where 0 ≤ ℓ ≤ kd − 1), and they are, using our notation, kd column sparse [45]. Interesting cases include generic bar-joint framework rigidity in the plane [35], which are characterized by (2, 3)-sparse matroids in dimension d = 2 (they are 4 column  sparse) and generic body-bar framework rigidity in Rd which are characterized by ( d+1 , d+1 )2 2 d+1 sparse matroids [46] (they are d 2 column sparse). Let M be a matroid with k sparse representation M and let X be an independent set of columns. It is easy to show (see e.g., [44]) that we can select one nonzero coordinate from each column in X such that no two selected entries lie on the same row. In other words, each independent set in M is also independent in the transversal matroid whose bipartite representations has color classes the rows and columns of M and where a column i is connected to a row j if entry Mij is nonzero. Even though the converse is not true, we can use the intuition obtained from graphic matroids to extend the algorithm to k column sparse representable matroids with only few changes. Instead of doing that, we are going to further generalize this class of matroids in a different direction. Multiframed matroids. A matroid M is called a frame matroid [48] if it can be extended to a second matroid M′ (i.e. M is a restriction of M′ ) which possesses a frame B, that is, a base such that every element of M is spanned by at most 2 elements of B. For instance, take a graphic matroid M[G] and consider the graph H = (V (G) + v0 , E(G) ∪ {v0 v : v ∈ V (G)} where v0 is a new vertex. Then, M[H] is an extension of M[G] containing the star centered at v0 as basis B := δH (v0 ) = {v0 v : v ∈ V (G)}. Since every edge uv in G is spanned by the set {v0 u, v0 v} of (at most) 2 elements, we conclude that B is a frame for M[G]. We define a new class of matroids, called multiframed of k-framed matroids, in a natural way. A matroid M = (R, I) is a k-framed matroid if it admits an extension MB = (R′ , I) having a k-frame B, i.e., a base such that each element of M is spanned by at most k elements of B. Without loss of generality we assume that B ∩ R = ∅ (by adding parallel elements) and R′ = B ∪ R. In other words, MB is obtained by adjoining the k-frame B to the original matroid M. Observe that if M is represented by a k column sparse matrix M , then the matrix [I|M ] obtained by adjoining an identity (in the field F) represents an extension MB of M where the columns of I form a base B such that each column in M is spanned by at most k elements from B (exactly those elements associated to the k nonzero rows of the column). This means that k-framed matroids generalizes k column sparse matroids. This generalization is strict since there are frame matroids that are nonrepresentable. We define the k-framed MSP as the variant of the MSP in which the k-framed matroid M, and its extension MB is either fully known beforehand or we simply have access to B and an independence oracle for MB (in the case of k column sparse matroid it is enough to receive the columns of the representation in an online fashion). We need some notation. For every r in the ground set R, we define the set C(B, r) = {y ∈ B : B +r −y is independent} which is also the minimal subset of B spanning r. It is easy to see that C(B, r) + r is the unique circuit in B + r, often called the fundamental circuit of r with respect to the base B in MB . Define also for each y ∈ B, the set K(B, y) = {r ∈ R : B +r−y is independent}. It is easy to see that K(B, y) + y is the unique cocircuit inside R + y in the matroid MB , often 17 called the fundamental cocircuit of y with respect to the base B. Observe that by definition, r ∈ K(B, y) ⇐⇒ y ∈ C(B, r). Furthermore, by definition of k-framed matroids, C(B, r) has at most k elements. Before presenting the algorithm we need the following result. Lemma 13. Let X be an independent set of a k-framed matroid M with k-frame B. There is a (canonical) injection πX : X → B such that B + x − πX (x) is independent for all x ∈ X. Proof. Extend X to a base X ′ of M′ . By the strong basis exchange axiom there is a bijection π : X ′ → B such that B + x − π(x) is a base for all x ∈ X ′ . The restriction of π to X yields the desired injection. Algorithm 5 for k-framed matroids. Input: A k-frame matroid M, with independence oracle access to MB and to B. The elements of M arrive in random order. ⊲ ALG is the set currently selected and B ′ is the set of elements of the frame B that have been marked. 1: ALG ← ∅, s ← Bin(n, p), B ′ ← ∅. 2: for i = s + 1 to n do 3: if ri ∈ OPT(Ri ) and C(B, ri ) ∩ B ′ = ∅ then 4: ALG ← ALG + ri , B ′ ← B ′ + πOP T (Ri ) (ri ) 5: Return ALG. Theorem 14. Algorithm 5 has forbidden sets of size k. By setting p = p(k), we get an α(k) probability-competitive algorithm for k-framed matroids. Proof. Suppose that the algorithm is not correct, and let Z = {ri(1) , ri(2) , . . . , ri(ℓ) } be a circuit in ALG with s + 1 ≤ i(1) ≤ i(2) ≤ · · · ≤ i(ℓ). When ri(1) arrived, y := πOPT(Ri(1) ) (ri(1) ) ∈ C(B, ri(1) ) was marked. Furthermore, our algorithm guarantees that for every element ri ∈ ALG with i > i(1), y 6∈ C(B, ri ). In particular, y ∈ C(B, ri(j) ) ⇐⇒ j = 1, or equivalently, K(B, y) ∩ Z = {ri(1) }. But this implies that the circuit Z and the cocircuit K := K(B, y) + y intersect only in one element, which cannot happen in a matroid. Therefore, the algorithm is correct. Since the sampling condition is satisfied by design, we only need to prove the k-forbidden condition. Let r ∗ ∈ OPT(Y ) where Y is an arbitrary set of t ≥ s + 1 elements, and suppose that Rt = Y and rt = r ∗ . The algorithm accepts r ∗ if and only if no element of C(B, rt ) was marked before. Let y ∈ C(B, rt ) be an arbitrary element. A sufficient condition for y not being marked at step j is that πOPT(Rj ) (rj ) 6= y. Therefore, if we define the forbidden sets F(X, Y, r ∗ ) = {f ∈ OPT(X) : πOPT(X) (f ) ∈ C(B, r ∗ )} it is clear that rt is selected if rj 6∈ F(Rj , Rt , r ∗ ) for all j ∈ {s + 1, . . . , t − 1}. Moreover, since πOPT(X) is injective we conclude that |F(X, Y, r ∗ )| ≤ |C(B, r ∗ )| ≤ k. 3.6 Laminar Matroids and Semiplanar Gammoids Arc-capacitated Gammoids. In this section we define special classes of gammoids amenable for our techniques. An arc-capacitated gammoid (ACG) M[N ] is defined by a directed network N = (G, s, R, c) where G = (V, E) is a digraph, s ∈ V is a single source, R ⊆ V \ {s} is a collection of terminals and c : E → Z+ is a strictly positive integer capacity function on the arcs of G. The 18 ground set of M[N ] is R and every set J ⊆ R is independent if and only if there exist an s-R flow on N satisfying the capacity constraints on the arcs and where each j ∈ J receives one unit of flow. Without loss of generality we assume that every terminal is reachable from s via a directed path. It is easy to see that ACGs are equivalent to gammoids without loops, but they are more closely related to transportation or distribution applications. A semiplanar gammoid is an ACG whose digraph G admits a semiplanar drawing, which is a planar drawing where all terminals are on the x-axis, the source node is on the positive y-axis, and the rest of the graph (arcs and nodes) are strictly above the x-axis, in a way that they do not touch the infinite ray starting upwards from the source node. It is easy to see that G admits a semiplanar drawing if and only if it is planar and all the terminals and the source are contained in the same face (which could be a cycle, a tree or a pseudoforest). Laminar Matroids. An important example of semiplanar gammoids are laminar matroids. A collection L of non-empty subsets of a finite set of terminals R is called laminar if for every L, H ∈ L, L ∩ H ∈ {L, H, ∅}. The laminar matroid M[R, L, c], where L is a laminar family over R and c : L → Z+ is a positive integer capacity function on the sets of L, is the matroid with ground set R and whose independent sets are those X ⊆ R for which the number of elements that X contains in any set of the laminar family does not exceed its capacity, i.e. |X ∩ L| ≤ c(L). Observe that, since c is positive, every singleton is independent. Furthermore, without loss of generality we assume that R ∈ L (with c(R) equal to the rank of the matroid) and that all singletons are in L (with c({r}) = 1 for each r ∈ R). Adding those sets does not destroy laminarity. The Hasse diagram of the containment order of L forms a tree T ′ where every set in L is a child of the smallest set in L that strictly contains it. Note that the leaves of T ′ are in bijection with the terminals and the root vR that represents R. Consider the directed graph G obtained by adding an auxiliary vertex s to T ′ connected to vR and orienting the tree away from s. If we put s on the upper semiaxis, we draw G with all arcs pointing downwards, and we assign capacities to each arc equal to the capacity of the laminar set associated to the arc’s head we obtain a semiplanar representation of M[R, L, c]. Furthermore, the terminals in R appear in the x-axis in the left-to-right order induced by the drawing of the tree. Semiplanar drawings and neighbours. In what follows we fix a semiplanar drawing G of a semiplanar gammoid M[N ]. We identify R with the set [n] by labeling the terminals from left to right as they appear in the x-axis. For convenience in the rest of the presentation, we add to G two auxiliary nodes on the x-axis, 0 and n + 1, where 0 is located to the left of 1, and n + 1 is located to the right of n, together with the arcs s0 and s(n + 1). Observe that by our assumptions on the drawing, it is possible to add those arcs without destroying semiplanarity. For any set J ⊆ [n] and any y ∈ J ∪ {0, n + 1}, we denote by LeftJ (y) the closest element to y in J ∪ {0} that is located strictly to its left (understanding LeftJ (0) = 0). Similarly, we denote by RightJ (y) the first element in J ∪ {n + 1} located to the right of y (understanding RightJ (n + 1) = n + 1). For y ∈ [n] \ J, we call LeftJ+y (y) and RightJ+y (y) its left and right neighbors in J (note that they maybe equal to 0 or n+1 and thus, they are not necessarily elements of J). Ancestors, tree-order and representatives. In the future we will map each terminal outside an independent set J to its neighbors in J. For the case in which G is a tree (i.e., for laminar matroids) we want to consistently assign each element to just one of them. We do this as follows. For every pair of nodes x, y in G, let xGy be the unique undirected x-y path in G. We say that x is 19 an ancestor of y (and y is a descendant of x) if sGy contains sGx; in that case we denote x ⊒ y (and y ⊑ x). Note that (V, ⊑) is a partial order, and in fact, it is a join-semilattice where x ∨ y is the lowest common ancestor of x and y. Note that by our drawing choice, for any i ∈ [j, j ′ ] ⊆ [0, n + 1] we have i ⊑ j ∨ j ′ . In particular, if [j, j ′ ] ⊆ [k, k ′ ] ⊆ [0, n + 1] then j ∨ j ′ ⊑ k ∨ k′ . For every nonempty set J ⊆ [n], and for every terminal y ∈ [n], we define its representantive πJ (y) in J such that   if y ∈ J, y πJ (y) = LeftJ+y (y) if y ∈ [n] \ J and y ∨ LeftJ+y (y) ⊏ y ∨ RightJ+y (y),   RightJ+y (y) if y ∈ [n] \ J and y ∨ LeftJ+y (y) ⊒ y ∨ RightJ+y (y). This element is well defined since for all y ∈ [n] \ J, both y ∨ LeftJ+y (y) and y ∨ RightJ+y (y) belong to sGy and so one is an ancestor of the other. Observe that πJ (y) is never equal to 0 (because then s = 0 ∨ y ⊏ RightJ+y (y) ∨ y which is a contradiction as s is the root), nor n + 1 (because then s = (n + 1) ∨ y ⊑ LeftJ+y (y) ∨ y ⊑ vR ⊏ s). In particular, πJ (y) ∈ {LeftJ+y (y), y, RightJ+y (y)} ∩ J. A graphical way to understand the definition of the representative of an element y outside J is the following: let j and j ′ be the left and right neighbors of y in J respectively and call sGj and sGj ′ the left and right paths respectively. The representative of y is its left neighbor (respectively, its right neighbor) if and only if by starting from y and walking up on G against its orientation, the first path hit is the left path (respectively the right path). In case of a tie, the representative is the right neighbor (see Figure 1). We claim also that for every j ∈ J, the set of elements to its right having j as representative is an interval of the form [j, k] with k < RightJ (j). Indeed, this is true if RightJ (j) = n + 1. Suppose now that j ′ := RightJ (j) ≤ n, and that the claim does not hold. Then, there must be two consecutive terminals i, i + 1 ∈ (j, j ′ ) with πJ (i + 1) = j and πJ (i) = j ′ . But then, we get the following contradiction: (i + 1) ∨ j ′ ⊑ i ∨ j ′ ⊑ i ∨ j ⊑ (i + 1) ∨ j ⊏ (i + 1) ∨ j ′ where the first inequality holds since [i + 1, j ′ ] ⊆ [i, j ′ ], the second, by the definition of πJ (i), the third since [j, i] ⊆ [j, i + 1] and the fourth by definition of πJ (i + 1). Since every element in (j, j ′ ) either has j or j ′ as representative, we conclude, in fact, that the entire set πJ−1 (j) of elements with j as representative is an interval of terminals enclosing j but strictly contained in [LeftJ (j), RightJ (j)]. In other words (πJ (j))j∈J is a partition of [n] into |J| intervals. The algorithms. In what follows we describe our algorithms for semiplanar gammoids and laminar matroids. Apart from some special case occurring when the sample is empty, our algorithms are extremely simple. Each one computes the optimal solution OPT(Rs ) of the sample and leave their elements unmarked. When a terminal ri that is part of the current optimum arrives, it checks if the (closest neighbors / representative) of ri in OPT(Rs ) (are / is) unmarked. If so, it marks (them / it) and selects ri as part of the solution. 20 5 4 3 3 2 0 1 2 2 2 3 4 5 6 7 8 9 2 10 11 12 13 2 14 15 16 17 18 Figure 1: Semiplanar drawing of a laminar matroid with ground set [17]. Nonunit arc capacities are shown. Below we show the partition (πJ (j))j∈J induced by the independent set J = {3, 5, 10, 13, 15}. Algorithm 6 for semiplanar gammoids. Algorithm 7 for laminar matroids. Input: An semiplanar gammoid with a fixed semiplanar drawing 1: ALG ← ∅, s ← Bin(n, p), 2: if s = 0 then 3: ALG ← {r1 } 4: else B ← ∅. 5: for i = s + 1 to n do 6: if ri ∈ OPT(Ri ), LeftOPT(Rs )+ri (ri ) 6∈ B and RightOPT(Rs )+ri (ri ) 6∈ B then 7: ALG ← ALG + ri , and B ← B ∪ {LeftOPT(Rs )+ri (ri ), RightOPT(Rs )+ri (ri )} Input: A laminar matroid with a fixed semiplanar drawing 1: ALG ← ∅, s ← Bin(n, p), 2: if s = 0 then 3: ALG ← {r1 } 4: else B ← ∅. 5: for i = s + 1 to n do 6: if ri ∈ OPT(Ri ) and πOPT(Rs ) (ri ) 6∈ B then 7: ALG ← ALG + ri , and B ← B ∪ {πOPT(Rs ) (ri )}. 8: Return ALG. 8: Return ALG. Theorem 15. Algorithm 6 has forbidden sets of size 4. By setting p = p(4) = α(4) = 44/3 ≈ 6.3496 probability-competitive algorithm for semiplanar ACGs. p 3 1/4 we get an The previous result applied to laminar matroids already beat the best guarantees known for √ that class (9.6 [39] and 3 3e [25, 26]). But we can do better for laminar matroids since for those, there is a unique path from the source to each terminal. p Theorem √ 16. Algorithm 7 has forbidden sets of size 3. By setting p = p(3) = 1/3 we get an α(3) = 3 3 probability-competitive algorithm for laminar matroids. √ We observe that this algorithm is very similar to the 3 3e competitive algorithm of Jaillet et al. [25, 26]. Note that the partition of the terminals given by πOPT(Rs ) induces a unitary partition matroid P ′ . After the sample, Algorithm 7 simply selects on each part, the first arriving element that is part of the current optimum. It can be shown that the partition matroid P ′ we define is the same as the one defined in [26, Section 3.2]. The main algorithmic difference is that in [26], the authors use the algorithm for the classic secretary problem to select one terminal on each part that has constant probability of being the largest element. Instead, we select the first arriving element 21 on each part that is part of the optimum at the time of arrival. This small change makes the competitive ratio of our algorithm e times smaller than theirs, but the analysis is more involved. To prove Theorems 15 and 16 we need to develop some extra tools. The unit capacity semiplanar gammoid associated to an independent set and its standard arc-covering. For every nonempty independent set J ⊆ [n] in the semiplanar gammoid M[N ] we chose an arbitrary but fixed s-J flow fJ satisfying the arc capacity constraints. Define the unit capacity network N 1 (J) = (G1 (J), s, R, 1), where G1 (J) is obtained from G by splitting each arc uv with fJ (uv) ≥ 2 into fJ (uv) parallel arcs (we keep all arcs uv with fJ (uv) ∈ {0, 1}, we also keep the arcs s0 and s(n + 1)). We do this in such a way that G1 (J) is still semiplanar. The matroid M[N 1 (J)] is the unit capacity semiplanar gammoid associated to J. Observe that every path in G1 (J) corresponds canonically to an unsplitted path in G going through the same nodes. Furthermore, every arc-disjoint collection P of s-R paths in G1 (J) can be regarded, by unsplitting parallel arcs, as an s-R flow on G satisfying the capacity constraints. Therefore, we have the following important lemma. Lemma 17. Any independent set in M[N 1 (J)] is also independent in the original matroid M[N ]. We say that two paths P and Q in the drawing of a graph cross if P enters Q on the left, shares zero or more arcs with Q and then exits Q on the right, or viceversa. A collection of paths is mutually noncrossing if no pair of them cross.7 The s-J flow fJ in the network N can be mapped to an s-J flow fJ1 in N 1 (J) in a natural way. By applying flow-decomposition on fJ1 we get a collection {P k }k∈J of arc-disjoint paths. Define also P 0 and P n+1 as the paths consisting of a single arc s0 and s(n + 1) respectively. By a standard planar uncrossing argument, we can assume that all paths in {P k }k∈J∪{0,n+1} are mutually noncrossing (but they can still share internal nodes, in fact all paths P k with k ∈ J contain s and vR , see Figure 2). For j ∈ J + 0, call j ′ = RightJ+0 (j) and define the collection of arcs Aj ⊆ E(G1 (J)) as those that are drawn in the closed planar region ′ Rj bounded by P j , P j and the x-axis. Since paths {P k }k∈J∪{0,n+1} form a topological star whose tips are in the x-axis we conclude that {Rk }k∈J+0 is a division of the region bounded by P 0 , P n+1 and the x-axis, and thus {Ak }k∈J+0 is a covering of all arcs in G1 (J). In fact, if 1 ≤ j ′ 6= n + 1, ′ ′ then every arc of P j belongs to pieces Aj and Aj , while each arc in G1 (J) \ {P k }k∈J belongs to a single piece of the covering. Furthermore, the only terminals contained in region Rj are those in [j, j ′ ]. Lemma 18. Let y ∈ [n] \ J, and j = LeftJ+y (y), j ′ = RightJ+y y be its neighbors in J. Then there is a directed s-y path DJ (y) in G1 (J) whose arc set is completely contained in Aj . Furthermore, DJ (y) can be chosen so that it arcs-intersects at most one path P̄ in (P k )k∈J . In the semiplanar ′ case, P̄ must be one of P j and P j , and in the laminar case, P̄ = P πJ (ℓ) . Proof. Note that y ∈ [j, j ′ ] is in the planar region Rj . Since y is reachable from s in G, there is a ′ path Q from s to y in G1 (J). Let v be the last vertex in Q contained in the vertices of P j and P j ′ (v maybe in one or both paths), say v is in P̄ ∈ {P j , P j }. By concatenating the initial piece of P̄ 7 An alternative way to understand this definition is to imagine that in the drawing nodes and arcs have positive area (they are circles and thick lines). A drawing of a path is any continuous non-self-intersecting curve drawn inside the union of all circles and thick lines associated to the nodes and arcs of P visiting them in the correct order. A collection of paths are mutually noncrossing if we can find drawings such that no pair of them intersects. Note that two noncrossing paths can still share arcs and nodes. 22 3 2 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 0 1 2 3 4 5 6 7 8 9 10 11 12 13 Figure 2: On the left, a flow fJ associated to the independent set J = {1, 4, 6, 11} on a semiplanar gammoid M[N ] with ground set R = [13] (0 and 14 are auxiliary nodes). On the right, M[N 1 (J)] and the partition of the drawing into regions R0 , R1 , R4 , R6 and R11 induced by the paths {P j }j∈J∪{0,14} . The regions R0 , R4 and R11 are shaded. from s to v and the final piece of Q from v to y we get a path DJ (y) that is completely contained in Aj and that intersects the arcs of at most one path in {Pj , Pj ′ }. ′ Consider the same situation in the laminar case. All arcs in P j , P j and Q are splitted versions of arcs in sGj, sGj ′ and sGy respectively. The vertex v defined above satisfies v = y ∨ j ∨ j ′ . If ′ y ∨ j ⊏ y ∨ j ′ then v = y ∨ j, πJ (y) = j and we can construct DJ (y) avoiding all arcs in P j by selecting P̄ = P j in the previous argument. Analogously, if y ∨ j ′ ⊑ y ∨ j then v = y ∨ j ′ , πJ (y) = j ′ , ′ and we can choose DJ (ℓ) to avoid P j by setting P̄ = P j . Lemma 19. Let J ⊆ [n] be a nonempty independent set in M[N ] and I ⊆ [n] \ J. Consider the following properties. (P1) For every x, y ∈ I with x 6= y, {LeftJ+x (x), RightJ+x (x)} ∩ {LeftJ+y (y), RightJ+y (y)} = ∅. (P2) M[N ] is laminar and for every x, y ∈ I, πJ (x) 6= πJ (y). If either (P1) or (P2) holds then I is independent in M[N 1 (J)] Proof. Suppose that (P 1) is satisfied. Then for every pair of distinct x, y ∈ I, the paths DJ (x) and DJ (y) constructed in Lemma 18 must lie on non-consecutive (and hence disjoint) regions in {Rk }k∈J+0 . Therefore the paths {DJ (z)}z∈I are mutually arc-disjoint from which we conclude that I is an independent set in M[N 1 (J)]. Suppose now that (P 2) is satisfied. Let x < y be distinct terminals in I. We claim that the paths DJ (x) and DJ (y) are arc disjoint. If x and y belong to different regions then the only possibility for them to share an arc is that x is in Rj , y ∈ Rj ′ with j ′ = RightJ (j), and both share ′ arcs in P j . But then, by Lemma 18, πJ (x) = πJ (y) = j ′ which is a contradiction. If x and y are in the same region Rj , then x, y ∈ [j, j ′ ] with j = LeftJ+x (x) = LeftJ+y (y) and ′ j = RightJ+x (x) = RightJ+y (y) ∈ J + (n + 1). Since the function πJ partitions [n] into intervals and x < y, we must have πJ (x) = j, πJ (y) = j ′ . Suppose now that DJ (x) and DJ (y) had an arc a in common and let w be its head. Since DJ (x) and DJ (y) are split versions of sGx and sGy we get that w ⊒ x ∨ y. Since w and x ∨ j are both in sGx, one is an ancestor of the other. Note that w cannot be an ancestor of x ∨ j since above the latter DJ (x) coincides with P j which is arc-disjoint from DJ (y) by Lemma 18 (and a is a common arc). It follows that x ∨ j ⊐ w ⊒ x ∨ y ⊒ y, and thus, 23 14 x ∨ j ⊒ y ∨ j. But since πJ (y) = j ′ we have y ∨ j ⊒ y ∨ j ′ and we conclude that x ∨ j ⊒ y ∨ j ′ ⊒ j ′ . From the last expression we get x ∨ j ⊒ x ∨ j ′ which contradicts the fact that πJ (x) = j. We have thus shown that all paths (DJ (z))z∈I are arc-disjoint, thence I is independent in M[N 1 (J)] Now we are ready to define the forbidden sets of sizes at most 4 and 3 respectively for Algorithms 6 and 7. For (X, r ∗ ) with r ∗ ∈ [n], X ⊆ [n] \ {r ∗ } define the sets F4 (X, r ∗ ) = {LeftX (LeftX+r∗ (r ∗ )), LeftX+r∗ (r ∗ ), RightX+r∗ (r ∗ ), RightX (RightX+r∗ (r ∗ ))} I4 (X, r ∗ ) = [LeftX (LeftX+r∗ (r ∗ )), RightX (RightX+r∗ (r ∗ ))]. The set F4 (X, r ∗ ) contains the 2 closest terminals to the left of r ∗ in X and the 2 closest terminals to its right. I4 (X, r ∗ ) is the enclosing interval containing F4 (X, r ∗ ). Similarly, for the case of laminar matroids, define F3 (X, r ∗ ) = {LeftX (πX (r ∗ )), πX (r ∗ ), RightX (πX (r ∗ ))} I3 (X, r ∗ ) = [{LeftX (πX (r ∗ )), RightX (πX (r ∗ ))}]. where F3 (X, r ∗ ) consists of the representative of r ∗ in X and its two neighbors, and I3 (X, r ∗ ) is its enclosing interval. We need one last technical lemma to analize the algorithms performance. Lemma 20. Let X and r ∗ as above, and let x ∈ X. (a) If x ∈ F4 (X, r ∗ ) then I4 (X, r ∗ ) ⊆ I4 (X − x, r ∗ ). (b) If x 6∈ F4 (X, r ∗ ) then F4 (X, r ∗ ) = F4 (X − x, r ∗ ) and I4 (X, r ∗ ) = I4 (X − x, r ∗ ). If the matroid is laminar, the following properties also hold (c) If x ∈ F3 (X, r ∗ ) then I3 (X, r ∗ ) ⊆ I3 (X − x, r ∗ ). (d) If x 6∈ F3 (X, r ∗ ) then F3 (X, r ∗ ) = F3 (X − x, r ∗ ) and I3 (X, r ∗ ) = I3 (X − x, r ∗ ). Proof of Lemma 20. Part (a): Let F4 (X, r ∗ ) = {a, b, c, d} from left to right (note that near the borders we could have a = b or c = d), in particular a ≤ b ≤ r ∗ ≤ c ≤ d. Denote by a− = LeftX (a) and d+ = RightX (d). (i) If x ∈ {a, b} then F4 (X − x, r ∗ ) = (F4 (X, r ∗ ) − x) + a− and so I4 (X, r ∗ ) = [a, d] ⊆ [a− , d] = I4 (X − x, r ∗ ). (ii) If x ∈ {c, d} then F4 (X − x, r ∗ ) = (F4 (X, r ∗ ) − x) + d+ and so I4 (X, r ∗ ) = [a, d] ⊆ [a, d+ ] = I4 (X − x, r ∗ ). Part (b): Direct. For parts (c) and (d) let F3 (X, r ∗ ) = {a, b, c} from left to right (in particular b = πX (r ∗ ), note that we could be in the cases a = b = 0 or b = c = n + 1). Denote by a− = LeftX (a) and c+ = RightX (c). Part (c): We have many possibilities to analyze. (i) If x = b, then the closest neighbors of r ∗ in X − x are a and c. In particular πX−x (r ∗ ) is either a or c. In both cases, I3 (X, r ∗ ) = [a, c] ⊆ I3 (X − x, r ∗ ). (ii) If x = a 6= b and r ∗ ∈ [b, c], then b and c are still the closest neighbors of r ∗ in X − x, and in particular, the representative in X − x is the same as that in X, i.e., πX−x (r ∗ ) = b. Note that since we removed a, LeftX−x (b) = a− and so, I3 (X, r ∗ ) = [a, c] ⊆ [a− , c] = I3 (X − x, r ∗ ) 24 (iii) The case x = c 6= b and r ∗ ∈ [a, b] is analogous to the previous one (iv) If x = a 6= b, and r ∗ ∈ [a, b], then the closest neighbors of r ∗ in X − x are a− and b. We have that r ∗ ∨ b ⊑ r ∗ ∨ a ⊑ r ∗ ∨ a− where the first inequality holds since πX (r ∗ ) = b and the second one since [a, r] ⊆ [a− , r ∗ ]. We conclude that πX−x (r ∗ ) = b. In particular I3 (X, r ∗ ) = [a, c] ⊆ [a− , c] = I3 (X − x, r ∗ ). (v) If x = c 6= b and r ∗ ∈ [b, c] then the closest neighbors of r ∗ in X − x are b and c+ . We have that r ∗ ∨ b ⊏ r ∗ ∨ c ⊑ r ∗ ∨ c+ where the first inequality holds since πX (r ∗ ) = b and the second holds since [r ∗ , c] ⊆ [r ∗ , c+ ]. We conclude that πX−x (r ∗ ) = b. In particular I3 (X, r ∗ ) = [a, c] ⊆ [a, c+ ] = I3 (X − x, r ∗ ). Part (d): Since x 6∈ F3 (X, r ∗ ) the neighbors of r ∗ in X−x are the same as those in X, it follows that πX−x (r ∗ ) = πX (r ∗ ) = b, LeftX−x (b) = a, RightX−x (b) = c. Therefore F3 (X − x, r ∗ ) = F3 (X, r ∗ ) and I3 (X − x, r ∗ ) = I3 (X, r ∗ ). Now we are ready to prove the guarantees for Algorithms 6 and 7. Proofs of Theorems 15 and 16. For both algorithms, if s = 0 then |ALG| = 1 and so it is independent. In the following assume that s ≥ 1. Line 6 in the algorithms guarantees that the set ALG satisfies the conditions of Lemma 19, hence ALG is independent in M[N 1 (J)], and by Lemma 17, ALG is independent in the original matroid. This proves correctness. Since the sampling condition holds by construction, we only need to check the forbidden property. For the rest of the proof, define F(X, r ∗ ) = Fi (X, r ∗ ), I(X, r ∗ ) = Ii (X, r ∗ ) where i = 4 on the semiplanar case, and i = 3 for the laminar case. We will show that the sets F(X, Y, r ∗ ) := F(X, r ∗ ) ∩ [n] are forbidden sets of size at most 4 for Algorithm 6 and of size 3 for Algorithm 7. Let r ∗ ∈ OPT(Y ) where Y is an arbitrary set of t ≥ s + 1 elements, and suppose that Rt = Y and rt = r ∗ . Assume now that the condition for every i ∈ {s + 1, . . . , t − 1}, ri 6∈ F(OPT(Ri ), rt ), (⋆) holds. We have to show that rt = r ∗ is chosen by the algorithm. Claim. The intervals I(OPT(Ri ), r ∗ ) are non-decreasing in i, namely, for all i ≤ t − 2, I(OPT(Ri ), r ∗ ) ⊆ I(OPT(Ri+1 ), r ∗ ). Indeed, let i ≤ t − 2. If OPT(Ri ) = OPT(Ri+1 ) then the claim is trivial, so assume otherwise. In particular, we have ri+1 ∈ OPT(Ri+1 ) \ OPT(Ri ). To simplify notation, let A = OPT(Ri+1 ) and A′ = A − ri+1 . By condition (⋆), ri+1 6∈ F(A, r ∗ ), and then by Lemma 20 using X = A, F(A, r ∗ ) = F(A′ , r ∗ ) and I(A, r ∗ ) = I(A′ , r ∗ ). (6) By the matroid exchange axiom we have two cases: either OPT(Ri ) = A′ or OPT(Ri ) = A′ + r̃ for some r̃ 6= ri+1 . In the first case we have by (6) that I(OPT(Ri ), r ∗ ) = I(A′ , r ∗ ) = I(A, r ∗ ) = I(OPT(Ri+1 ), r ∗ ), which ends the proof, and so we focus on the the second case. If r̃ ∈ F(A′ + r̃, r), then by (6) and Lemma 20 applied to X = A′ + r̃, we have I(OPT(Ri ), r ∗ ) = I(A′ + r̃, r ∗ ) ⊆ I(A′ , r ∗ ) = I(A, r ∗ ) = I(OPT(Ri+1 ), r ∗ ). On the other hand, if r̃ 6∈ F(A′ + r̃, r), then, again by Lemma 20 and and (6) we have I(OPT(Ri ), r ∗ ) = I(A′ + r̃, r ∗ ) = I(A′ , r ∗ ) = I(A, r ∗ ) = I(OPT(Ri+1 ), r ∗ ). This concludes the proof of the claim. 25 We show how to finish the proof of the lemma using the claim. Suppose that ri is selected by the algorithm for some i ∈ {s+1, . . . , t−1}. By (⋆) and the claim we deduce that ri 6∈ I(OPT(Ri ), r ∗ ) ⊇ I(OPT(Rs ), r ∗ ). In particular ri is far away from r ∗ : (a) In the semiplanar case, there are at least 2 terminals of OPT(Rs ) between r ∗ and ri . In particular, {LeftOPT(Rs )+ri (ri ), RightOPT(Rs )+ri (ri )} and {LeftOPT(Rs )+r∗ (r ∗ ), RightOPT(Rs )+r∗ (r ∗ )} do not intersect and so, at iteration i, neither LeftOPT(Rs )+r∗ (r ∗ ) nor RightOPT(Rs )+r∗ (r ∗ ) are added into B. (b) In the laminar case, there is at least one terminal of OPT(Rs ) between πOPT(Rs ) (r ∗ ) and ri . In particular, πOPT(Rs ) (ri ) 6= πOPT(Rs ) (r ∗ ), and so, at iteration i, πOPT(Rs ) (r ∗ ) is not added into B. Since the statements above are satisfied for every i ≤ t − 1, we conclude that at time t, rt = r ∗ satisfies the conditions in line 6 of the algorithms, and so it is selected. This concludes the proof that Algorithms 6 and 7 have forbidden sets of sizes 4 and 3 respectively. 3.7 Uniform matroids We devise a variant of Kleinberg’s algorithm [31] for uniform matroids whose probability competitiveness tends to 1 as the rank ρ goes to infinity. Kleinberg’s algorithm is better described when both the rank ρ = 2k and the number of elements n = 2N are powers of 2. It was originally presented in a recursive manner, but it is illustrative to describe it without using recursion: Kleinberg’s algorithm. For every i ∈ N, let Ii be the interval of the first n/2i elements. Then the intervals Jk := Ik , Jk−1 := Ik−1 \ Ik , . . . , J1 := I1 \ I2 and J0 = I0 \ I1 partition the ground set R. The algorithm treats each interval in {J0 , J1 , . . . , Jk } separately. For 0 ≤ i ≤ k − 1, it selects at most bi = ρ/2i+1 elements from Ji and at most one element from Jk , so that in total at most ρ elements are selected. The way the algorithm selects an element or not is determined by a threshold for each interval. Namely, it selects the first arriving element from Jk and for i ∈ {0, 1, . . . , k − 1}, the algorithm uses as threshold in the interval Ji the (ρ/2i )-th highest element seen strictly before the interval, i.e., the (ρ/2i )-th element of Ii+1 . It selects every element better than this threshold until the budget bi of the interval is depleted, ignoring all the elements arriving later in this interval. The probability-ratio of Kleinberg’s algorithm is at least 4/3. Kleinberg p [31] shows that the previous algorithm is guaranteed to obtain an expected p fraction 1 − O( p 1/ρ) of the optimum weight, which in our notation means to be 1/(1 − O( 1/ρ)) = 1 + O( 1/ρ) utility-competitive. This guarantee also holds for the ordinal notion since the algorithm does not need weights. However, as we show next, its probability competitiveness is bounded away from 1. Since J0 and I1 have the same cardinality, with probability tending to 1/4 as ρ goes to infinity, we simultaneously have that |I1 ∩ Rρ−1 | < |J0 ∩ Rρ−1 | and r ρ ∈ J0 . Given that, the threshold for J0 , which is the ρ/2-th element of I1 , will be attained by an element of Rρ−1 which is strictly better than r ρ . Thus, r ρ will not be selected. Therefore, the algorithm selects r ρ (which is in OPT) with probability at most 1−1/4 = 3/4. An asymptotically 1 probability-competitive algorithm. The next algorithm is a simple modification of Kleinberg’s, which we write in a continuous setting. Every element is associated with a uniform random variable with support [0, 1). It is useful for this part to imagine [0, 1) as a time interval, and identify the realization of the uniform random variable as the arrival time of the element. For each j ∈ N, let Ij be the interval [0, 2−j ) and Jj = [2−j−1 , 2−j ) to be its second half. The 26 sequence {Jj }j≥0 partitions the interval [0, 1). For convenience, let Kj = [2−j−1 , 2−j−1 (2 − 4εj )) be the left (1−4εj ) fraction of the interval Jj , for some parameter εj , depending on ρ, to be chosen later. Algorithm 8 for uniform matroids. Input: A uniform matroid U (n, ρ) with ground set R and rank ρ. 1: Get a sample of size n, independently and uniformly at random from [0, 1). Sort it from smallest to largest as t1 < t2 < · · · < tn . Thus ti is interpreted as the arrival time of ri . 2: ALG ← ∅ 3: for i = 1 to n do 4: Compute the index j such that ti ∈ Jj = [2−j−1 , 2−j ). 5: Compute the threshold fj equal to the ⌈( 21 )j+1 (1 + εj )ρ⌉-th highest element in Ri with arrival time in Ij+1 = [0, 2−j−1 ). 6: if less than ( 12 )j+1 ρ elements with arrival time in Jj have been selected and ri ≻ fj then 7: ALG ← ALG + ri 8: Return ALG. .. . I3 J2 I2 J1 I1 I0 J0 0 · · · 0.25 0.5 1 Figure 3: Definition of the sets Ij and Jj p Theorem 21. Algorithm 8 is 1 + O( log ρ/ρ) probability-competitive. √ By going from utility to probability we incur in a factor of log ρ on the error probability. It √ remains open whether we can achieve Kleinberg’s optimal utility-competitiveness of 1 + O(1/ ρ) for the stronger probability notion. Proof of Theorem the algorithm selects at most (1/2)j+1 ρ elements from interval Jj , in P∞21. Since j+1 total at most j=0 (1/2) ρ = ρ elements are included into ALG. Therefore, the algorithm is correct. Since the matroid is uniform of rank ρ, the optimal base is OPT = {r 1 , r 2 , . . . , r ρ }. Let r ∗ = r k ∈ OPT, for some k ∈ {1, . . . ,p ρ}. In what follows, let j be an integer such that 1 ∗ 0 ≤ j ≤ ⌊ 2 log(ρ/96)⌋ := j and let εj = 12 · 2j ln ρ/ρ. We first find a lower bound on the probability that r ∗ is selected, given that its arrival time is in Jj . (j) (j) For each r i ∈ R − r ∗ , let Xi and Yi be the indicator variables that the arrival time of r i is in Ij+1 and Kj respectively. Let σ(fj ) be the global ranking of the computed threshold fj , that is, r σ(fj ) = fj . Consider the following three events, X (j) Xi i∈[ρ+1]\{k} < (1 + εj ) ρ, 2j+1 X (j) Xi  1+εj   i∈ 1−εj ρ +1 \{k} 27 ≥ 1 + εj ρ, 2j+1 X  1+εj   i∈ 1−εj (j) Yi ρ +1 \{k} < 1 2j+1 ρ, that we call U1 , U2 and U3 respectively. Consider a fourth event, U4 , that is tk ∈ Kj where tk is the arrival time of r k . We claim that provided tk ∈ Jj , the intersection of these four events guarantees that r ∗ is selected by the algorithm. Conditional on the event that tk ∈ Jj , event U1 ∩ U2 implies that ρ + 1 < σ(fj ) ≤ ⌈(1 + εj )ρ/(1 − εj )⌉ + 1 and r ∗ ≻ fj . Event U2 ∩ U3 implies that the number of elements strictly higher than the threshold fj and arriving on interval Kj is strictly less than ( 21 )j+1 ρ. In particular, this implies that the algorithm has not selected enough elements in the interval Jj . Therefore, conditional on the event that tk ∈ Jj , the event U1 ∩ U2 ∩ U3 ∩ U4 implies that r ∗ ∈ ALG. Calling Ui to the negation of event Ui , by union bound and noting that events from 1 to 3 are independent of the arrival time of r ∗ = r k , we have P 4 Pr(r k 6∈ ALG | tk ∈ Jj ) ≤ i=1 Pr(Ui | tk ∈ Jj ) = Pr(U1 ) + Pr(U2 ) + Pr(U3 ) + 4εj . (j) The random variables {Xi }i∈[n]\{k} are independently and identically Bernoulli distributed of (j) parameter equal to the length of Ij+1 , that is 2−j−1 . Similarly, the random variables {Yi }i∈[n]\{k} are independently and identically Bernoulli distributed of parameter equal to the length of Kj that is (1 − 4εj )2−j−1 . In the following we upper bound the probabilities in the sum above by using the Chernoff bound [40, p. 64-66, Theorems 4.4 and 4.5].   12 · 2j ln ρ    1  ε2 j Pr(U1 ) ≤ exp − 2−j−1 ρ = exp − 2−j−1 ρ = exp −2 ln ρ ≤ . 3 3ρ ρ (j) (j) Let µX and µY be the expected sums of the random variables Xi , and respectively Yi , for l m   1+εj   1+ε 1+εj ρ 1 j+1 . The choice of j i ∈ 1−εj ρ + 1 \ {k}. For µX we have µX = 1−εj ρ 2 ≥ 1−εjj · 2j+1 guarantees that 1/ρ < ǫj < 1/8, and therefore     1 + εj 1 − 4εj 1 + εj (1 − 4εj )(1 + 2εj ) ρ 1 − 4εj 1 ρ µY = ρ ≤ + εj ρ · j+1 ≤ ≤ · . 1 − εj 2j+1 1 − εj 2 1 − εj 2j+1 1 + εj 2j+1 In the last inequality we used that εj < 1/8. By Chernoff bound on events U2 and U3 , we obtain  ε2   12 · 2j ln ρ 1 + ε X (j) ρ  1 j j Xi < µX (1 − εj )) ≤ exp − µX ≤ exp − · · j+1 ≤ , Pr(U2 ) ≤ Pr( 2 2ρ 1 − εj 2 ρ i  ε2   12 · 2j ln ρ 1 + ε (1 − 4ε )ρ  1 X (j) j j j Yi ≥ µY (1 + εj )) ≤ exp − µY ≤ exp − Pr(U3 ) ≤ Pr( · · ≤ . 3 3ρ 1 − εj 2j+1 ρ i Putting all together, it follows that ∗ k Pr(r 6∈ ALG) = ≤ X Pr(r 6∈ ALG|t ∈ Jj ) X X j>j ∗ j>j ∗ k 1 2j+1 + 1 k j≥0 3 + 4εj ρ 2j+1  + 1 2j+1 j X j=0 = Pr(r k 6∈ ALG|tk ∈ Jj ) 1 2j ∗ +1 +  1 2j+1 s  3 12 ln ρ X j/2 1 2 +4 , ρ ρ 2j+1 j≥0 q q  q  p log ρ log ρ 1 1 which is O + + = O . Therefore, the algorithm is 1/(1 − O( log ρ/ρ)) = ρ ρ ρ ρ p (1 + O( log ρ/ρ)) probability-competitive. 28 4 4.1 Algorithms for the ordinal MSP on general matroids O(1) intersection-competitiveness: Proof of Theorem 2 Our algorithm for the intersection notion works as follows. We first sample s elements. After that, we select an element as long as it is part of the optimum of the subset of elements seen so far, and it preserves the independence of the current solution. Recall that we denote by Ri = {r1 , r2 , . . . , ri } the first i elements seen by the algorithm. Algorithm 9 Improving Greedy Input: Matroid M(R, I) in random order r1 , r2 , . . . , rn , and a fixed integer s ∈ {1, . . . n}. 1: ALG ← ∅ 2: for i = s + 1 to n do 3: if ri ∈ OPT(Ri ) and ALG + ri ∈ I then 4: ALG ← ALG + ri . 5: Return ALG. The idea of only considering elements that belong to the current optimum is not new. Ma et al. [39] consider an algorithm that after sampling a fraction, selects an element as long as it belongs to the current offline optimum, and they prove this to be 9.6 utility-competitive (and the analysis holds for probability as well) for laminar matroids. The same algorithm was suggested by Babaioff et al. [5], and they showed how this algorithm fails to be even constant utility-competitive. In contrast, we show this algorithm to be O(1) intersection-competitive. Lemma 22. Let B = {ri : ri ∈ OPT(Ri ), i ∈ {s + 1, . . . , n}}. Then E[|B|] = (Hn − Hs )ρ, where Hj denotes the j-th harmonic number.  Proof. For every i ∈ [n] if we condition the set Ri to be some fixed set F ∈ Ri , we have Pr(ri ∈ OPT(Ri )|Ri = F ) = Pr(ri ∈ OPT(F )) ≤ ρ . i Therefore, by linearity of expectation we conclude that E[|B|] = n X i=s+1 n X ρ Pr(ri ∈ OPT(Ri )) ≤ = (Hn − Hs )ρ. i i=s+1 Proof of Theorem 2. We study the competitiveness of the algorithm when the sample size is s, and then we optimize over this value to conclude the theorem. For any random ordering, |ALG| = ρ(R \ Rs ) ≥ ρ(OPT \ Rs ) = |OPT \ Rs |. Then, E[|ALG ∩ OPT|] equals E[|ALG|] + E[|OPT \ Rs |] − E[|ALG ∪ (OPT \ Rs )|] ≥ 2E[|OPT \ Rs |] − E[|ALG ∪ (OPT \ Rs )|]. Furthermore, ALG ∪ (OPT \ Rs ) ⊆ B, where B is the set defined in Lemma 22. Since the elements arrive in uniform random order, for r ∈ OPT we have that Pr(r ∈ / Rs ) = (n − s)/n = 1 − s/n. Therefore, the right hand side of the previous inequality is at least     Z n 1 2s 2s − dx ρ = 2 − + ln(s/n) ρ. 2E[|OPT \ Rs |] − E[|B|] ≥ 2 − n n s x 29 This quantity is maximized in s = n/2. So, by assuming n even (which can be done by adding an extra dummy element if n is odd), and setting the algorithm for s = n/2, we obtain E[|ALG ∩ OPT|] ≥ ρ(1 − ln(2)) = |OPT|/ ln(e/2). 4.2 Ordinal/Probability-competitiveness: Proof of Theorem 3 We introduce a variant of the MSP that helps us to leverage existing algorithms for the utility version of the MSP, in order to get competitive algorithms for the ordinal and probability variants. We need a concept similar to the aided sample-based MSP introduced by Feldman et al .[19]. In the Layered-MSP the input is a tuple (M, F, C, ≻) where M = (R, I, ≻) is a totally ordered matroid, C = {c1 , c2 , . . . , ck } is a finite set with C ∩ R = ∅, ≻ is a total order over R ∪ C with c1 ≻ c2 ≻ · · · ≻ ck , and F is a random subset of R in which every element is present with probability 1/2. The set C defines a partition of R in the following way. We call C0 = {r ∈ R : r ≻ c1 } the highest layer and Ck = {r ∈ R : ck ≻ r} the lowest layer. For j ∈ {1, . . . , k − 1}, the j-th layer is the set Cj = {r ∈ R : cj ≻ r ≻ cj+1 }. By construction, the layers {C0 , C1 , . . . , Ck } induced by C form a partition of R. In the following we call a tuple (M, F, C, ≻) as a layered matroid. An algorithm for the Layered-MSP first sees F but is unable to select any of its elements. The rest of the elements of M arrive in uniform random order. At time step t, the algorithm can get the full value order in Rt by using only ordinal information, it can use an independence oracle to test any subset of Rt and it can also check membership of any element to each layer induced by C. Definition 23. We say that an algorithm for the Layered-MSP is α-competitive if it returns an independent set ALG ∈ I, and for each j ∈ {0, 1, . . . , |C|}, E|ALG ∩ Cj | ≥ α1 |OPT ∩ Cj |, where the expectation is taken both over the distribution of F and the internal algorithm randomness. Theorem 24 (Feldman et al., [19, Corollary 4.1]). There exists an 8⌈log(|C| + 1) + 1⌉-competitive algorithm for the Layered-MSP. For the sake of completeness we include the mentioned algorithm for Layered-MSP of Feldman et al. We remark that the algorithm was introduced in the context of the utility version. Nevertheless, the result above follows in the absence of weights. Algorithm 10 (Feldman et al. [19]) for Layered-MSP Input: A layered matroid (M, F, C, ≻) 1: Let τ be a uniformly at random number from {0, 1, . . . , ⌈log(|C| + 1)⌉}. 2: Let ∆ be a uniformly at random number from {0, 1, . . . , 2τ − 1}. Smin{|C|,2τ i−∆} 3: Let B = {B1 , B2 , . . . , B⌈(∆+|C|)/2τ ⌉ } where Bi = j=max{0,2τ (i−1)−∆+1} Cj . 4: With probability 1/2, set H = odd(|C|) or even(|C|) otherwise. 5: For each i ∈ H let Ti ← ∅. 6: for each element in r ∈ R \ F do 7: Let i be such that r ∈ Bi . 8: if i ∈ H and r ∈ Ni and Ti + r ∈ Ii then 9: Ti ← Ti + r. 10: Return ALG = ∪i∈H Ti . 30 About the algorithm of Feldman et al. We denote by odd(k) and even(k) the odd and even numbers, respectively, in the set {0, 1, . . . , k}. The set Ii is the independent sets family of a matroid Mi = (Ni , Ii ) defined as follows. Let B≥i = ∪j∈{i,...,⌈(∆+|C|)/2τ ⌉} Bj . The matroid M1 is obtained from M by contracting 8 F ∩ B≥2 and then restricting9 to B1 . For i > 1, Mi is obtained from M by contracting F ∩ B≥i+1 and then restricting it to Bi ∩ span(F ∩ B≥i−1 ). 4.2.1 Reduction from Layered-MSP to ordinal and probability variants of the MSP The main result of this section corresponds to the lemma below. Theorem 3 follows directly using this lemma, and the rest of this section is devoted to prove the lemma. Lemma 25. Suppose there exists a g(|C|)-competitive algorithm for the Layered-MSP, where g is a non-decreasing function. Then, (i) there exists an O(g(1 + log ρ)) ordinal-competitive algorithm for the MSP, and (ii) there exists an O(g(1 + ρ)) probability-competitive algorithm for the MSP. Proof of Theorem 3. We observe that g(x) = 8⌈log(x + 1) + 1⌉ is non-decreasing, so we can apply Lemma 25 using Algorithm 10 and Theorem 24. In the following, let Alayer be a g(|C|)-competitive algorithm for the layered-MSP. Our algorithm for Lemma 25 (i), depicted as Algorithm 11, first gets a sample from R with expected size n/2, and constructs a partition C using the optimum of the sample. By sampling a set F over the remaining elements it feeds Alayer with a layered matroid. Algorithm 11 O(g(1 + log ρ)) ordinal-competitive algorithm Input: Matroid M(R, I, ≻) in random order r1 , r2 , . . . , rn . 1: Let s ∼ Bin(n, 1/2) and compute OPT(Rs ) = {s(1), . . . , s(ℓ)}, where s(1) ≻ s(2) ≻ · · · ≻ s(ℓ). 2: Let C = {s(1), s(2), s(4), . . . , s(2k−1 )}, where k = ⌊log ℓ⌋ + 1. 3: Let t ∼ Bin(n − s, 1/2) and let F = {rs+1 , . . . , rs+t } be the next t elements from R \ Rs . 4: Return ALG = Alayer (M|R\Rs , F, C, ≻). Proof of Lemma 25 (i). Let OPT = {f (1), . . . , f (ρ)} be such that f (1) ≻ f (2) ≻ · · · ≻ f (ρ). Consider the function T : N → N given by T (0) = 0, T (i) = |{r ∈ R : r ≻ f (i)}| if i ∈ [1, ρ] and T (i) = n if i > ρ. In particular, for i ∈ [1, ρ], T (i) is the number of elements in M that are at least as high as the i-th element of OPT, so f (i) = r T (i) . Observe that for all k such that T (i) ≤ k < T (i + 1) we have |OPT ∩ Rk | = i. In the following we study the expected number of elements from Rk that the algorithm selects, so we can conclude using Lemma 5. If T (0) ≤ k < T (1) it holds E[|ALG ∩ Rk |] = 0 = |OPT ∩ Rk |, so this case is done. Let k ∈ N be such that T (1) ≤ k < T (8). Let f be the highest non-loop element in the value order with f (1) ≻ f . With probability at least 1/4, it holds that f (1) ∈ R \ Rs and f ∈ Rs . In this case, f ∈ OPT(Rs ), since f (1) = s(1), and the only non-loop element of C0 is f (1). Thus, C0 = {f (1)} = OPT(R \ Rs ) ∩ C0 . Therefore, Pr(C0 = {f (1)}) ≥ Pr(f (1) ∈ R \ Rs , f ∈ Rs ) ≥ 1/4. 8 9 The contraction of M = (R, I) by Q, M/Q, has ground set R−Q and a set I is independent if ρ(I∪Q)−ρ(Q) = |I|. The restriction of M = (R, I) to Q, M|Q , has ground set Q and a set I is independent if I ∈ I and I ⊆ Q. 31 Since g is non-decreasing and |C| ≤ 1+log ℓ ≤ 1+log ρ, Algorithm Alayer is g(1+log ρ)-competitive. Furthermore, the event C0 = {f (1)} depends only on steps 1 and 2 of the algorithm before executing Alayer . It follows that 1 1 Pr(f (1) ∈ ALG|C0 = {f (1)}) = E[|ALG ∩ C0 | | C0 = {f (1)}] 4 4 1 1 ≥ E[|OPT(R \ Rs ) ∩ C0 | | C0 = {f (1)}] = . 4g(1 + log ρ) 4g(1 + log ρ) Pr(f (1) ∈ ALG) ≥ Since |OPT ∩ Rk | ≤ 8, we conclude in this case that E[|ALG ∩ Rk |] ≥ Pr(f (1) ∈ ALG) ≥ 1 1 ≥ |OPT ∩ Rk |. 4g(1 + log ρ) 32g(1 + log ρ) Let j ≥ 3 and k be such that T (2j ) ≤ k < T (2j+1 ). We denote q = 2j−3 . Let Aj be the event where |{f (1), . . . , f (2q)}∩R\Rs | ≥ q and Bj is the event where |{f (2q+1), . . . , f (6q)}∩Rs | ≥ 2q. We have that Pr(Aj ∩ Bj ) ≥ 1/4, since in our algorithm the probability for an element to be sampled equals the probability of not being sampled. Observe that any subset of elements strictly better than f (t) has rank at most t − 1, and therefore f (2q)  s(2q). If Bj holds, then s(2j−2 ) = s(2q)  f (6q). Since f (6q)  f (8q) = f (2j ), it follows that j−3 [ i=0 j Ci ⊆ {r ∈ R \ Rs : r  f (2j )} = RT (2 ) ∩ R \ Rs . This implies that j−3 i 1 h [ 1 h T (2j ) Ci | Aj ∩ Bj ≥ E ALG ∩ E[|ALG ∩ R |] ≥ E |ALG ∩ R 4 4 k i=0 i Aj ∩ Bj . Furthermore, if Aj holds, then OPT(R \ Rs ) ∩ j−3 [ i=0 Ci = |{f ∈ OPT ∩ R \ Rs : f  s(2q)}| ≥ |{f ∈ OPT ∩ R \ Rs : f  f (2q)}| ≥ q. Since g is non-decreasing and |C| ≤ 1+log ℓ ≤ 1+log ρ, Algorithm Alayer is g(1+log ρ)-competitive. Since events Aj and Bj depend only on the sampling at line 1 of the algorithm (before executing Alayer ) and by using linearity of the expectation and the observation above we have that j−3 [ 1 h Ci E ALG ∩ 4 i=0 i Aj ∩ Bj ≥ ≥ j−3 h [ 1 Ci E OPT(R \ Rs ) ∩ 4g(1 + log ρ) i=0 Aj ∩ Bj 1 1 q≥ |OPT ∩ Rk |, 4g(1 + log ρ) 64g(1 + log ρ) i where the last inequality holds since |OPT ∩ Rk | ≤ 2j+1 = 16q. By using Lemma 5 we conclude that the algorithm is O(g(1 + log ρ)) ordinal-competitive. 32 Algorithm 12 O(g(1 + ρ)) probability-competitive algorithm Input: Matroid M(R, I, ≻) in random order r1 , r2 , . . . , rn . 1: Let s ∼ Bin(n, 1/2) and compute OPT(Rs ) = {s(1), . . . , s(ℓ)}, where s(1) ≻ s(2) ≻ · · · ≻ s(ℓ). 2: Let t ∼ Bin(n − s, 1/2) and let F = {rs+1 , . . . , rs+t } be the next t elements from R \ Rs . 3: Return ALG = Alayer (M|R+ , F ∩Rs+ , OPT(Rs ), ≻), where Rs+ = {r ∈ R\Rs : r ∈ OPT(Rs +r)}. s To prove Lemma 25 (ii), consider Algorithm 12 depicted above. Before the analysis, it will be useful to consider the next process. Let (Xt )t∈N be a sequence of Bernoulli independent random variables such that Pr(Xt = 0) = Pr(Xt = 1) = 1/2 for every t ∈ N. We create two sets V, W ⊆ R iteratively using the following procedure. Algorithm 13 Coupling procedure Input: Matroid M(R, I, ≻). 1: Initialize V ← ∅, W ← ∅ and θ ← 0. 2: for i = 1 to n do 3: if r i ∈ OPT(V + r i ) then 4: θ ← θ + 1 and Yθ = Xi . 5: if Yθ = 0 then 6: V ← V + ri 7: else W ← W + r i . The value θ represents a counter on the elements that improve over the current set V . When the counter is updated, we say that the element considered on that iteration is assigned coin Yθ . Lemma 26. (V, W ) has the same distribution as (OPT(Rs ), Rs+ ) in Algorithm 12. Proof. Let (Vi , Wi )ni=1 be the states of the coupling process at the end of each iteration. Let Z = {r i : Xi = 0}. Observe that Z and the sample Rs of Algorithm 12 have the same distribution. Since the coupling procedure checks from highest to lowest element, it follows that Vi = OPT(Z∩Ri ) for every i ∈ {1, . . . , n}, and therefore Vn = OPT(Z). To conclude the lemma it suffices to check that Wn = {r ∈ R \ Z : r ∈ OPT(Z + r)}. In fact, it can be proven by induction that Wi = {r ∈ Ri \ Z : r ∈ OPT(Vi + r)} for every i ∈ {1, . . . , n}, and so the lemma follows, since Rn = R and Vn = OPT(Z). Proof of Lemma 25 (ii). Thanks to Lemma 26 we assume that (OPT(Rs ), Rs+ ) is generated by the process described in Algorithm 13. In what follows, fix f (j) = r T (j) ∈ OPT. We know that at step T (j) of the coupling procedure, f (j) ∈ OPT(Rs + f (j)) no matter the trajectory of (Xt )t∈N . Let θ be such that r T (j) is assigned coin Yθ , that is, Yθ = XT (j) . Then, with probability at least 1/8, the event E defined as Yθ−1 = Yθ+1 = 0 and Yθ = 1, holds. If E happens, let s(h) be the element of OPT(Rs ) who was assigned coin Yθ−1 . In particular, the element s(h + 1) is assigned coin Yθ+1 . Thus, Ch = {r ∈ Rs+ : s(h) ≻ r ≻ s(h + 1)} = {f (j)} = OPT(Rs+ ) ∩ Ch . 33 Therefore by using that the occurrence of E is decided before executing Alayer which is g(1 + ℓ) ≤ g(1 + ρ) competitive, we get that 1 Pr(f (j) ∈ ALG|E) 8 1 1 1 = E[|ALG ∩ Ch ||E] ≥ E[|OPT(Rs+ ) ∩ Ch ||E] = , 8 8g(ρ + 1) 8g(ρ + 1) Pr(f (j) ∈ ALG) ≥ and we conclude that Algorithm 12 is O(g(1 + ρ)) probability-competitive. 4.3 Comparison between ordinal measures In this section we discuss about some incomparability results for the competitiveness measures previously introduced. We show that an algorithm that is utility-competitive is not necessarily competitive for the rest of the measures. In particular, we provide an instance where the O(log ρ) utility-competitive algorithm by Babaioff, Immorlica and Kleinberg [5] has a poor competitiveness for the other three. Regarding the notions of the ordinal MSP, we show that the intersection and the ordinal measures are incomparable. More specifically, we show the existence of an algorithm and an instance where it is arbitrarily close to 1 intersection-competitive, but have an unbounded ordinal/probability-competitiveness. And on the other hand, we also show the existence of an algorithm and an instance where it is arbitrarily close to 1 ordinal-competitive, but have an unbounded intersection/probability-competitiveness. Recall that probability is the strongest in the sense that it implies competitiveness for all the other measures considered (see Lemma 6). In the utility variant, the weight w(r) of an element is revealed to the algorithm when arrived. Suppose for simplicity that the rank ρ of the matroid is known10 and that the weights the algorithm sees are all different. In the above mentioned algorithm, called by the authors the Threshold Price Algorithm (TPA), it is taken a sample Rs of s ∼ Bin(n, 1/2) elements11 and it records the top weight w∗ of a non-loop12 element seen in Rs . It chooses uniformly at random a number τ in the set {0, 1, 2, . . . , ⌈log2 ρ⌉}, and then it selects greedily any non-sampled element whose weight is at least T = w∗ /2τ . Theorem 27 (Babaioff et al. [5]). The algorithm TPA is O(log ρ) utility-competitive. 4.3.1 TPA is Ω(ρ) intersection, ordinal and probability-competitive We show in this section that TPA is Ω(ρ)-competitive in the intersection, ordinal and probability measures. We first revisit the proof by [5] for the O(log ρ) utility-competitiveness of TPA. Proof of Theorem 27. Let OPT = {f (1), . . . , f (ρ)} be the optimal base such that w1 > w2 > · · · > wρ , where wi = w(f (i)) for each i ∈ {1, . . . , ρ}. Suppose that f (1) ∈ / Rs . Then, if the second non-loop element of the matroid is sampled and if τ = 0, the element f (1) will be selected in the second phase of TPA. Hence Pr(f (1) ∈ ALG) ≥ 1/4 · 1/(⌈log ρ⌉ + 1) = Ω(1/ log ρ). 10 This assumption can be removed by using standard arguments; we can set ρ equal to twice the rank of the sampled part. 11 The original analysis uses half of n, but the analysis gets simpler if one uses Bin(n, 1/2) since one can assume that each element is in the sample with probability 1/2 independent of the rest. 12 A loop in a matroid is an element that belongs to no basis. 34 Let B = {i ∈ {2, . . . , ρ} : wi ≥ w1 /ρ}, and let Ei be the event where f (1) ∈ Rs and wi /2 < T ≤ wi . For each i ∈ B, we have log(w1 /wi ) < log ρ and therefore Pr(Ei ) = 1/2·Pr(τ = ⌈log(w1 /wi )⌉) = Ω(1/ log ρ). The random order assumption implies that in expectation (i − 1)/2 ≥ i/4 elements in {f (2), . . . , f (i)} are non-sampled, hence the expected rank of the matroid restricted to non-sampled elements of weight at least wi = w(f (i)) is Ω(i). Therefore, given i ∈ B and conditioned on Ei , the algorithm selects Ω(i) elements of value at least wi /2. It follows that for each i ∈ B,   i . E[|ALG ∩ {r : w(r) ≥ wi /2}|] ≥ Pr(Ei )E[|ALG ∩ {r : w(r) ≥ wi /2}||Ei ] = Ω log ρ In addition, observe that the elements P in OPT \ {f (i) : i ∈ B ∪ {1}} have total weight less than w1 (ρ − |B| − 1)/ρ < w1 , and therefore 2 i∈B∪{1} wi ≥ w(OPT). Putting all together, we have that the expected weight of the output is ρ 1X (wi − wi+1 )E[|ALG ∩ {r : w(r) ≥ wi /2}|] 2 i=1  X  1 (wi − wi+1 ) · i =Ω log ρ i∈B∪{1}    X  1 1 =Ω wi = Ω w(OPT). log ρ log ρ E[w(ALG)] ≥ i∈B∪{1} We study in the following the competitiveness of TPA for the ordinal measures. Since probability is the strongest (Lemma 6), it is enough to check that TPA is Ω(ρ) ordinal and intersection competitive. Let R = {r 1 , . . . , r n } with n = 2ρ3 . Consider the laminar family {Rn/2 , R}, where Rn/2 = {r 1 , . . . , r n/2 }, and take the laminar matroid (R, I) where a set I ∈ I is independent if |I ∩ Rn/2 | ≤ 1 and |I| ≤ ρ. In particular, the matroid rank is ρ. Given ε > 0, the weights are given by w(r i ) = 8 − εi for i ∈ {1, . . . n/2}, and w(r i ) = 7 − εi for i ∈ {n/2 + 1, . . . , n}. Observe that the optimal basis is given by OPT = {r 1 , r n/2+1 , . . . , r n/2+ρ−1 }. If we run TPA on this instance, with probability p = 1 − 2−n/2 the top weight w∗ in the sample is from Rn/2 , and thus, 7 < w∗ < 8. Let E be this event. No matter the value of T , the algorithm always select at most ρ elements in the non-sampled part, and therefore E[|ALG ∩ OPT|] = (1 − 2−n/2 )E[|ALG ∩ OPT||E] + 2−n/2 E[|ALG ∩ OPT||E] 3 ≤ E[|ALG ∩ OPT||E] + ρ · 2−ρ ≤ E[|ALG ∩ OPT||E] + 1. Conditional on E, we have that either T > 7 or T < 4. In the first case, TPA will select at most one element, which would come from Rn/2 , and so |ALG ∩ OPT| ≤ 1. Otherwise, if T < 4, the algorithm will select at most one element from Rn/2 and at most the first ρ non-sampled elements from R \ Rn/2 . The expected number of elements in {r n/2+1 , . . . , r n/2+ρ−1 } appearing in {rs+1 , . . . , rs+ρ }, that is the first ρ elements appearing after the sample, is (ρ − 1)ρ/n < 1. It follows that E[|ALG ∩ OPT||E] is upper bounded by 2, and therefore E[|ALG ∩ OPT|] ≤ 3. The discussion above implies as well that E[|ALG ∩ Rn/2+ρ−1 |] ≤ 3, and then TPA is Ω(ρ) intersection and ordinal competitive. The conclusion for the ordinal case follows by using the characterization for ordinal competitiveness in Lemma 5. 35 4.3.2 The ordinal and intersection measures are not comparable We show in the following that there is no competitiveness dominance between the intersection and the ordinal notions. Let m and M be two positive integers. Consider R = {r 1 , r 2 , . . . , r (M +1)m } and let L be the partition of R given by L= m [ i {{r }} ∪ i=1 M [ {{r jm+1 , . . . , r (j+1)m }}. j=1 In particular, we consider the partition matroid Mm,M = (R, I) where I ∈ I if |I ∩ L| ≤ 1 for every L ∈ L. The matroid rank is m + M . The algorithm we choose is greedy: we initialize ALG ← ∅, and when an element r arrives it is selected if ALG + r ∈ I. Lemma 28. Suppose M ≥ m2 . Then, the greedy algorithm over instance Mm,M is (1 + 1/m) ordinal-competitive and Ω(m) intersection-competitive. In fact, the ordinal competitiveness holds no matter what the value of M is. We adjust the value of M in order to get a poor competitiveness for the intersection notion. Proof of Lemma 28. The optimal base of Mm,M is the highest element of each part in L, that is, OPT = {r 1 , r 2 , . . . , r m } ∪ {r jm+1 : j ∈ {1, . . . , M }}. In particular, we have ( k |OPT ∩ Rk | = m+j if k ∈ {1, . . . , m}, if k ∈ {jm + 1, . . . , (j + 1)m} and j ∈ {1, . . . , M }. Observe that an element r ∈ Q, with Q ∈ L, is selected by the algorithm if and only if r is the first element of Q that arrives. Therefore, Pr(r i ∈ ALG) = 1 if i ∈ {1, . . . , m} and Pr(r i ∈ ALG) = 1/m if i ∈ {m + 1, . . . , (m + 1)m}. It follows that E[|ALG ∩ Rk |] = k if k ∈ {1, . . . , m}. If k > m, then k E[|ALG ∩ R |] = m X i=1 i Pr(r ∈ ALG) + k X i=m+1 Pr(r i ∈ ALG) = m + 1 (k − m). m Thus, when k ∈ {1, . . . , m}, we have |OPT ∩ Rk |/E[|ALG ∩ Rk |] = 1. Suppose k = jm + r with j ≥ 1 and 0 ≤ r ≤ m. Then, m+j |OPT ∩ Rk | = 1 k E[|ALG ∩ R |] m+ m (k − m) m−r 1 1 =1+ 2 ≤1+ ≤ , m + mj + r − m m+j−1 m and thus the greedy algorithm is (1 + 1/m) ordinal-competitive for Mm,M . In the first inequality we used that φ(x) = (m − x)/(m2 + mj + x − m) is a decreasing function in the interval [0, m]. Observe that the competitiveness result holds no matter the value of M . It remains to study the intersection competitiveness. By the observations above, we have that m + m2 m+M m |OPT| ≥ = ≥ , 1 E[|ALG ∩ OPT|] m+m 2 m+ m ·M and so the algorithm is at least Ω(m) intersection-competitive. 36 Although not mentioned explicitly in the proof, since Pr(r i ∈ ALG) = 1/m for i > m it follows that the algorithm is m probability competitive for Mm,M , and for every M . In the following we construct an instance for which the intersection competitiveness is close to 1, but the ordinal competitiveness is poor. Let m be a positive integer and R = {r 1 , . . . , r 2m−1 }. Consider the partition L given by  L = {r 1 , . . . , r m }, {r m+1 }, {r m+2 }, . . . , {r 2m−1 } . Let Nm = (R, I) be the partition matroid where I ∈ I if |I ∩ L| ≤ 1 for every L ∈ L. The matroid rank is m. Lemma 29. The greedy algorithm over instance Nm is (1 + 1/m) intersection-competitive and Ω(m) ordinal-competitive. Proof. An element r ∈ Q, with Q ∈ L, is selected by the algorithm if and only if r is the first element of Q that arrives. Therefore, Pr(r i ∈ ALG) = 1/m if i ∈ {1, . . . , m} and Pr(r i ∈ ALG) = 1 if i ∈ {m + 1, . . . , 2m − 1}. Since OPT = {r 1 , r m+1 , . . . , r 2m−1 }, 1 |OPT ∩ R1 | = = 1/(1/m) = m, 1 1 E[|ALG ∩ R |] Pr(r ∈ ALG) hence the algorithm is Ω(m) ordinal-competitive. Finally, we have m 1 |OPT| = ≤1+ , E|ALG ∩ OPT| 1/m + m − 1 m and we conclude that the greedy algorithm is (1 + 1/m) intersection-competitive over Nm . References [1] M. Ajtai, N. Megiddo, and O. Waarts. Improved Algorithms and Analysis for Secretary Problems and Generalizations. SIAM Journal on Discrete Mathematics, 14(1):1–27, 2001. [2] P. D. Azar, R. Kleinberg, and S. M. Weinberg. Prophet inequalities with limited information. In Proc. of SODA 2014, pages 1358–1377, 2014. [3] M. Babaioff, M. Dinitz, A. Gupta, N. Immorlica, and K. Talwar. Secretary Problems: Weights and Discounts. In Proc. of SODA 2009, pages 1245–1254, 2009. [4] M. Babaioff, N. Immorlica, D. Kempe, and R. Kleinberg. A knapsack secretary problem with applications. In Proc. of APPROX-RANDOM 2007, volume 4627 of LNCS, pages 16–28, 2007. [5] M. Babaioff, N. Immorlica, and R. Kleinberg. Matroids, secretary problems, and online mechanisms. In Proc. of SODA 2007, pages 434–443, 2007. [6] S. Barman, S. Umboh, S. Chawla, and D. L. Malec. Secretary problems with convex costs. In Proc. of ICALP 2012, volume 7391 of LNCS, pages 75–87, 2012. [7] M. Bateni, M. Hajiaghayi, and M. Zadimoghaddam. Submodular secretary problem and extensions. ACM Transactions on Algorithms, 9(4):1–23, 2013. 37 [8] M. Beckmann. Dynamic programming and the secretary problem. Computers & Mathematics with Applications, 19(11):25–28, 1990. [9] F. T. Bruss. Sum the odds to one and stop. The Annals of Probability, 28(3):1384–1391, 2000. [10] N. Buchbinder, K. Jain, and M. Singh. Secretary Problems via Linear Programming. Mathematics of Operations Research, 39(1):190–206, 2014. [11] S. Chakraborty and O. Lachish. Improved Competitive Ratio for the Matroid Secretary Problem. In Proc. of SODA 2012, pages 1702–1712, 2012. [12] T.-H. H. Chan, F. Chen, and S. H.-C. Jiang. Revealing Optimal Thresholds for Generalized Secretary Problem via Continuous LP: Impacts on Online K-Item Auction and Bipartite KMatching with Random Arrival Order. In Proc. of SODA 2015, pages 1169–1188, 2015. [13] N. B. Dimitrov and C. G. Plaxton. Competitive Weighted Matching in Transversal Matroids. Algorithmica, 62(1-2):333–348, 2012. [14] M. Dinitz and G. Kortsarz. Matroid Secretary for Regular and Decomposable Matroids. SIAM Journal on Computing, 43(5):1807–1830, 2014. [15] P. Dütting and R. Kleinberg. Polymatroid prophet inequalities. In Proc. of ESA 2015, volume 9294 of LNCS, pages 437–449, 2015. [16] E. B. Dynkin. The optimum choice of the instant for stopping a Markov process. Soviet Math. Dokl, 4:627–629, 1963. [17] J. Edmonds and D. Fulkerson. Transversals and matroid partition. Journal of Research of the National Bureau of Standards Section B Mathematics and Mathematical Physics, 69B(3):147– 153, 1965. [18] M. Feldman, J. Naor, and R. Schwartz. Improved competitive ratios for submodular secretary problems (extended abstract). In APPROX-RANDOM, volume 6845 of LNCS, pages 218–229, 2011. [19] M. Feldman, O. Svensson, and R. Zenklusen. A Simple O (log log(rank)) -Competitive Algorithm for the Matroid Secretary Problem. In Proc. of SODA 2015, pages 1189–1201, 2015. [20] M. Feldman and R. Zenklusen. The submodular secretary problem goes linear. In Proc. of FOCS 2015, pages 486–505, 2015. [21] T. S. Ferguson. Who Solved the Secretary Problem? Statistical Science, 4(3):282–289, 1989. [22] J. P. Gilbert and F. Mosteller. Recognizing the Maximum of a Sequence. Journal of the American Statistical Association, 61(313):35, 1966. [23] M. Hoefer and B. Kodric. Combinatorial Secretary Problems with Ordinal Information. In I. Chatzigiannakis, P. Indyk, F. Kuhn, and A. Muscholl, editors, Proc. of ICALP 2017, volume 80 of Leibniz International Proceedings in Informatics (LIPIcs), pages 133:1–133:14, Dagstuhl, Germany, 2017. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik. 38 [24] S. Im and Y. Wang. Secretary Problems: Laminar Matroid and Interval Scheduling. In Proc. of SODA 2011, pages 1265–1274, 2011. [25] P. Jaillet, J. A. Soto, and R. Zenklusen. Advances on Matroid Secretary Problems: Free Order Model and Laminar Case. In Proc. of IPCO 2013, volume 7801 of LNCS, pages 254–265, 2013. [26] P. Jaillet, J. A. Soto, and R. Zenklusen. Advances on Matroid Secretary Problems: Free Order Model and Laminar Case (Full version). http://arxiv.org/abs/1207.1333v2, 2014. [27] M. Janata. Matroids induced by packing subgraphs. SIAM J. Discrete Math., 18(3):525–541, 2005. [28] T. Kesselheim, R. Kleinberg, and R. Niazadeh. Secretary Problems with Non-Uniform Arrival Order. In Proc. of STOC 2015, pages 879–888, 2015. [29] T. Kesselheim, K. Radke, A. Tönnis, and B. Vöcking. An Optimal Online Algorithm for Weighted Bipartite Matching and Extensions to Combinatorial Auctions. In Proc. of ESA 2013, volume 8125 of LNCS, pages 589–600, 2013. [30] T. Kesselheim, K. Radke, A. Tönnis, and B. Vöcking. Primal beats dual on online packing LPs in the random-order model. In Proc. of STOC 2014, pages 303–312, 2014. [31] R. Kleinberg. A Multiple-Choice Secretary Algorithm with Applications to Online Auctions. In Proc. of SODA 2005, pages 630–631, 2005. [32] R. Kleinberg and S. M. Weinberg. Matroid prophet inequalities. In Proc. of STOC 2012, pages 123–136, 2012. [33] N. Korula and M. Pál. Algorithms for Secretary Problems on Graphs and Hypergraphs. In Proc. of ICALP 2009, volume 5556 of LNCS, pages 508–520, 2009. [34] O. Lachish. O(log log Rank) Competitive Ratio for the Matroid Secretary Problem. In Proc. of FOCS, pages 326–335, 2014. [35] G. Laman. On graphs and rigidity of plane skeletal structures. Journal of Engineering Mathematics, 4(4):331–340, 1970. [36] D. V. Lindley. Dynamic Programming and Decision Theory. Applied Statistics, 10(1):39–51, 1961. [37] M. Loebl and S. Poljak. On matroids induced by packing subgraphs. Journal of Combinatorial Theory, Series B, 44(3):338–354, 1988. [38] M. Loréa. Hypergraphes et matroı̈des. Cahiers Centre Etud. Rech. Oper, 17:289–291, 1975. [39] T. Ma, B. Tang, and Y. Wang. The Simulated Greedy Algorithm for Several Submodular Matroid Secretary Problems. Theory of Computing Systems, 58(4):681–706, 2016. [40] M. Mitzenmacher and E. Upfal. Probability and computing: Randomized algorithms and probabilistic analysis. Cambridge University Press, 2005. 39 [41] S. Oveis Gharan and J. Vondrák. On Variants of the Matroid Secretary Problem. Algorithmica, 67(4):472–497, 2013. [42] A. Rubinstein. Beyond matroids: secretary problem and prophet inequality with general constraints. In Proc. of STOC 2016, pages 324–332, 2016. [43] A. Rubinstein and S. Singla. Combinatorial prophet inequalities. In Proc. of SODA 2017, pages 1671–1687. SIAM, 2017. [44] J. A. Soto. Matroid Secretary Problem in the Random-Assignment Model. SIAM Journal on Computing, 42(1):178–211, 2013. [45] I. Streinu and L. Theran. Natural realizations of sparsity matroids. Ars Mathematica Contemporanea, 4(1):141–151, 2011. [46] T. S. Tay. Rigidity of multi-graphs. I. Linking rigid bodies in n-space. Journal of Combinatorial Theory, Series B, 36(1):95–112, 1984. [47] W. Whiteley. Some matroids from discrete applied geometry. Contemporary Mathematics, 197:171–311, 1996. [48] T. Zaslavsky. Frame Matroids and Biased Graphs. European Journal of Combinatorics, 15(3):303–307, 1994. 40
8cs.DS
On the Universal Approximability of Quantized ReLU Neural Networks Yukun Ding 1 Jinglan Liu 1 Yiyu Shi 1 arXiv:1802.03646v1 [cs.LG] 10 Feb 2018 Abstract Compression is a key step to deploy large neural networks on resource-constrained platforms. As a popular compression technique, quantization constrains the number of distinct weight values and thus reducing the number of bits required to represent and store each weight. In this paper, we study the representation power of quantized neural networks. First, we prove the universal approximability of quantized ReLU networks. Then we provide upper bounds of storage size given the approximation error bound and the bit-width of weights for function-independent and functiondependent structures. To the best of the authors’ knowledge, this is the first work on the universal approximability as well as the associated storage size bound of quantized neural networks. 1. Introduction Various deep neural networks deliver state-of-the-art performance on many tasks such as target tracking, autonomous driving, etc. Their prevalence has extended to embedded or mobile devices for edge intelligence, where security, reliability or latency constraints refrain the networks from running on servers or in clouds. However, large network sizes and the associated expensive computation and memory consumption make edge intelligence a challenging task. In response, as will be detailed in Section 2, substantial effort has been made to reduce the memory consumption of neural networks while minimizing the accuracy loss. The memory consumption of neural networks can be reduced by either directly reducing the number of weights or decreasing the number of bits (bit-width) needed to represent and store each weight, which can be employed on top of each other (Choi et al., 2016). The number of weights can be reduced by pruning (Han et al., 2015b), weight sparsifying (Liu et al., 2015), structured sparsity learning (Wen et al., 2016) and low rank approximation (Denton et al., 2014). The 1 Department of Computer Science and Engineering, University of Notre Dame, IN, USA. Correspondence to: Yukun Ding <[email protected]>. bit-width is reduced by quantization that maps data to a smaller set of distinct levels (Sze et al., 2017). Note that while quantization may stand for linear quantization only (Li et al.; Gysel et al., 2016) or nonlinear quantization only (Han et al., 2015a; Choi et al., 2016) in different works, our discussion will cover both cases. However, as of today quantization is still only empirically shown to be robust and effective to compress various neural network architectures (Hubara et al., 2016; Zhou et al., 2017b; Zhuang et al., 2017). Its theoretical foundation still remains mostly missing. Specifically, many important questions remain unanswered. For example: • Why even binarized networks, those most extremely quantized with bit-width down to one, still work well? • How small can a quantized network be to achieve a desired accuracy? In this paper, we provide some insights for these questions from a theoretical perspective. We focus on deep networks with the ReLU activation function, which is among the most widely used in deep neural networks (Xu et al., 2015). The conclusions on ReLU networks can be extended to any other networks that use piecewise linear activation functions with finite breakpoints such as leaky ReLU immediately, as one can replace a ReLU network by an equivalent one using these activation functions while only increasing the number of units and weights by constant factors (Yarotsky, 2017). The main contributions of this paper include: • We prove the universal approximation property of quantized ReLU networks. We prove that even the most extremely quantized ReLU networks using two distinct weight values are capable of representing a wide class of functions with arbitrary accuracy. • Given the number of distinct weights and the desired approximation error, we further provide an upper bound of the storage size which is the total number of bits needed to store all weights for the quantized ReLU network. The remainder of the paper is organized as follows. Section 2 reviews related works. Section 3 lays down the models and On the Universal Approximability of Quantized ReLU Neural Networks assumptions of our analysis. We prove the universal approximability and the upper bounds with function-independent structure in Section 4 and extend it to function-dependent structure in Section 5. Concluding remarks are given in Section 6. 2. Related Works Linear quantization does mapping with same distance between contiguous quantization levels and is usually implemented by storing weights as fixed-point numbers with reduced bit-width (Li et al.; Gysel et al., 2016). Nonlinear quantization clusters weights to different groups and all weights in one group share a single value. Then the weights are stored using lossless binary coding (the index to a lookup table) instead of the actual values (Han et al., 2015a; Choi et al., 2016). The most important difference between linear quantization and nonlinear quantization is that the latter does not require weights to be low precision values. There are rich literatures on how to obtain quantized networks, either by linear quantization or nonlinear quantization (Zhou et al., 2017a; Leng et al., 2017; Shayar et al., 2017). It is reported that a pruned AlexNet can be quantized to eight bits and five bits in convolutional layers and fully connected layers respectively without any loss of accuracy. Similar result is also observed in LENET-300-100, LENET-5, and VGG-16 (Han et al., 2015a). One may argue that some of these benchmark networks are known to have redundancy. However, recent works show that quantization works well even on networks that are designed to be extremely small and compact. The bit-width of SqueezeNet, which is a stateof-the-art compact network, can be reduced to eight while preserving the original accuracy (Gysel et al., 2016; Iandola et al., 2016). There are some representative works that can achieve little accuracy loss on ImageNet classification even using binary or ternary weights (Courbariaux et al., 2015; Rastegari et al., 2016; Li et al., 2016; Zhu et al., 2016). More aggressively, some works also reduce the precision of activations, e.g. (Hubara et al., 2016; Rastegari et al., 2016). Although the classification accuracy loss can be minimized, the universal approximation property is apparently lost, as with limited output precision the network cannot achieve arbitrary accuracy. Accordingly, we do not include them in the discussion of this paper. Motivated by the great empirical success, the training of quantized neural networks has been analyzed theoretically (Li et al., 2017b; Li et al.; Choi et al., 2016; Li et al., 2017a). The universal approximability of ReLU networks is proved in (Mhaskar & Micchelli, 1992) and revisited in (Sonoda & Murata, 2017). Recently, Hanin discusses the expressive power of ReLU networks with bounded width and proves that a ReLU network with width d + 1 can approximate any continuous convex function of d variables arbitrarily well (Hanin, 2017). Shaham et al. construct a sparselyconnected depth-4 ReLU network and prove its error bound (Shaham et al., 2016). Liang et al. prove that, for a large class of piecewise smooth functions, a network with ReLU and binary step units can provide an error bound  with O( 1 ) layers and O(poly log 1 ) neurons (Liang & Srikant, 2016). Zhao et al. prove the universal approximation property of low displacement rank (LDR) neural networks with a mild condition on the displacement operator, which is the result of another effective technique of neural network compression (Zhao et al., 2017). Our work is based on (Yarotsky, 2017) which provides approximation error bound for deep ReLU networks through construction. 3. Models and Assumptions Throughout this paper, we define ReLU networks as feedforward neural networks with the ReLU activation function σ(x) = max(0, x). The ReLU network considered includes several input units, a number of hidden units, and one output unit. Without loss of generality, each unit can only connect to units in the next layer. We denote the number of distinct weight values as λ (λ ≥ 2), for both linear quantization and nonlinear quantization. To store each weight, we only need log(λ) 1 bits to encode the index i.e., the bit-width is log(λ). The overhead to store sparse structures can be ignored because it varies depending on the implementation and can be easily reduced to the same order as the weight storage using techniques such as compressed sparse row (CSR) for nonlinear quantization. The number of bits needed to store the codebook can also be ignored, because it has lower order of complexity. We consider any function f in the Sobolev space: f ∈ Wn,∞ ([0, 1]d ) and max ess sup |Dn f (x)| ≤ 1. n:|n|≤n x∈[0,1]d (1) We denote this function space as Fd,n in this paper. The space Wn,∞ consists of all locally integrable function f : Ω → R such that Dn f ∈ L∞ (Ω), where |n| ≤ n. Thus we are actually considering a broad range of functions since any f that has bigger domain or derivative can be simply scaled to satisfy given restrictions. 4. Function-independent Structure We start our analysis with function-independent structure, where the topology of the network is fixed for any function considered. The main idea to prove the universal approximability is to construct a network that can achieve arbitrary accuracy given the number of distinct weight values λ. The 1 tion Throughout this paper, we omit base 2 for clarity of presenta- On the Universal Approximability of Quantized ReLU Neural Networks storage size of this network then naturally serves as an upper bound. We start with nonlinear quantization, and first present the approximation of some basic functions by sub-networks in Section 4.1. We then present the sub-network that approximates any weight in Section 4.2, and finally the whole network composed of these sub-networks in Section 4.3. We extend the results to linear quantization in Section 4.4. 4.1. Approximation of squaring/multiplication Proposition 1. The function f (x) = x2 on the segment [0, 1] can be approximated by a ReLU network fsr with only two weight values 12 and − 21 , so that (i) if x = 0, fsr (x) Since = 0;  xy = 2  (ii) the approximate error s = O 2−2r ; (iii) the depth is O (r); (iv) the width is a constant; (v) the number of weight is O (r). Proof. For fs (x) = x2 , let fsr be the piecewise linear interpolation of f with 2r + 1 uniformly distributed breakpoints k r 2r , k = 0, . . . , 2 . We have k fsr ( r ) 2 Figure 1. ReLU Implementation of fsr (x) where b and w denote bias and weight respectively. Details of block Bm and block Bg are depicted on the left. k = ( r )2 , 2 k = 0, . . . , 2 r (2) s = ||fsr (x) − fs (x)||∞ ≤ 2−2(r+1) . Consider the function g : [0, 1] 7→ [0, 1]: ( 2x x < 12 g(x) = 2(1 − x) x ≥ 12 . fsr (x) = x − 2 −  x 2 2 −  y 2 2 ! , (6) we can implement multiplication with the help of Proposition 1. Proposition 2. Given x ∈ [−1, 1], y ∈ [−1, 1], and only two weight values 12 and − 12 , there is a ReLU network with two input units that implements a function ×0 : R2 7→ R, so that (i) if x = 0 or y = 0, then ×0 (x, y) = 0; 0 (ii) for any  x, y, the error ×0 = | × (x, y) − xy| = −2r O 2 ; (iii) the depth is O (r); and the approximation error Note that x+y 2 r X 2−2i g ◦i (x) (3) (4) (5) i=1 where g ◦i (x) is the i-th iterate of g(x). Since g(x) can be implemented by a ReLU network as g(x) = 2σ(x) − 4σ(x − 12 ), g ◦r (x) can be obtained by concatenating such implementation of g(x) for r times. Now, to implement fsr (x) based on g ◦r (x), all we need are weights {2−2 , 2−4 , · · · , 2−2(r−1) , 2−2r }, which can be easily constructed with additional 2r layers and the weight 1 r 2 . Then, fs (x) can be implemented by a ReLU network as in Figure 1. A connection from or to a block indicates connections to all units in the block. Some units are copied to compensate the scaling caused by 12 and to facilitate the construction of higher level networks. (iv) the width is a constant; (v) the number of weights is O (r). Proof. Build a sub-network ξs that computes fxr as in Proposition 1 and let        |x + y| |x| |y| 0 r r r × (x, y) = 2 fs − fs − fs . 2 2 2 (7) Then the statement (i) is followed by property (i) of Proposition 1. Inserting Equation (3) to Equation (7), we get the error bound of ×0 : ×0 ≤ 6 · 2−2(r+1) . (8) Since we can easily construct a sub-network Babs that computes the absolute value, ×0 (x, y) can be constructed as a linear combination of three parallel ξs ’s and feed them with |x| |y| |x+y| 2 , 2 , and 2 . Then claims of statement (iii), (iv), and (v) are also obtained. 4.2. Approximation of weights Proposition 3. A connection with any weight w ∈ [−1, 1] in a ReLU network can be approximated by a sub-network On the Universal Approximability of Quantized ReLU Neural Networks that has only λ ≥ 2 distinct weights, so that (i) the sub-network is equivalent to a connection with weight w0 where |w0 − w| < 2−t ;   1 (ii) the depth is O λt λ−1 ; (iii) the width is O(t);   1 (iv) the number of weights is O λt λ−1 +1 ; Proof. Consider that we need a weight w to feed the input x to a unit in the next layer as wx. With limited number of distinct weight values, we can construct the weight we need by cascade and combination. For clarity, we first consider w ≥ 0 and x ≥ 0, and relax these assumptions later. The connections with w = 0 can be seen as an empty network while w = 1 can be easily implemented by 4 units with weight 12 . Now we show how to represent all integral multiples of 2−t from 2−t to 1−2−t . 1 Without loss of generality, we assume t λ−1 is an integer. 1 We use λ weights that include − 2 and W : W , {2−1 , 2−t 1 λ−1 , 2−t 2 λ−1 , · · · , 2−t λ−2 λ−1 }. (9) We first construct all w from Wc which is defined as Wc , {2−1 , 2−2 , · · · , 2−(t−1) }. (10) the network is solely determined by w and works for any input x. 4.3. Approximation of general smooth functions Theorem 1. For any f ∈ Fd,n , given λ distinct weights, there is a ReLU network with fixed structure that can approximate f with any error  ∈ (0, 1), so that     1  1 n2d+3 λ−1 + log  ; (i) the depth is O λ log  (ii) the number of weights is    1 +1  d  1 n n2d+3 λ−1 ; O λ log   (iii) the number of bits needed to store the network is   1  λ−1  +1 d d+3 1 n . O λ(log λ) log n2  Proof. The proof is composed of four parts. We first approximate f by Taylor polynomial as f2 and prove the approximation error bound. Then we approximate f2 by a ReLU network that is denoted as f 0 with bounded error. After that, we present the ReLU implementation the network f 0 . At last, we show the complexity of the network. Consider a function: 1 Similar to a numeral system with radix equal to t λ−1 , any wi ∈ Wc can be obtained by concatenating weights from W 1 while every weights in W is used no greater than t λ−1 − 1 times. After that, all integral multiples of 2−t from 2−t to 1 − 2−t can be represented by a binary expansion on Wc . Note that connections in the last layer for binary expansion use weight 12 , thus additional 2−1 is multiplied to scale the resolution from 2−(t−1) to 2−t . Since for any in  weight  Wc we need to concatenate no more than 1 λ−1 λ t − 1 weights in a straight line, the sub-network  1  has no greater than λ t λ−1 − 1 + 1 layers, and no greater  1  than 4tλ t λ−1 − 1 + 8t + 4 weights. We now relax the assumption w ≥ 0. When w < 0, the sub-network can be constructed as w = |w|, while we use − 12 instead of 12 in the last layer. To relax the assumption x ≥ 0, we can make a duplication of the sub-network. Let all the weights in the first layer of the sub-network be 12 for one and − 12 for the other. Here we are utilizing the gate property of ReLU. In this way, one sub-network is activated only when x > 0 and the other is activated only when x < 0. The sign of the output can be adjusted by flipping the sign of weights in the last layer. Note that the configuration of ψm (x) = d Y h(3N xk − 3mk ), (11) k=1 where m = (m1 , · · · , md ) ∈ {0, 1, · · · , N }d , and h(x) is defined as follows:   |x| ≤ 1 1 h(x) = 2 − |x| 1 < |x| < 2 . (12)   0 |x| ≥ 2 Note that X ψm (x) ≡ 1, x ∈ [0, 1]d (13) m and mk 1 (14) < ∀k}. N N For all m, the order n−1 Taylor polynomial for the function f is  X Dn f m n Tm = x− . (15) m n! x= N N supp ψm ⊂ {x : xk − n:|n|<n According to Tm , we define an approximation to f by Taylor polynomial using Tm0 as X f2 , ψm Tm0 , (16) m∈{0,··· ,N }d On the Universal Approximability of Quantized ReLU Neural Networks where Tm0 X = βm,n  n:|n|<n and βm,n is ple of 1 n Dn f n! m x= N  d n−|n| . N m n x− , N (17) rounded to the nearest integral multi- Proposition 2. Consider that if we construct our function ×0 with | ×0 (x, y) − xy| < ×0 = 2−2(r+1) , then | ×0 (x, y) − xz| ≤ |x(y − z)| + ×0 . (21) 0 Applying Equation (21) to |fm,n (x) − fm,n (x)| repeatedly, we bound it to Equation (22) on the next page. Then the approximation error of f2 is bounded by Equation (18). X |f (x) − f2 (x)| = | ψm (x)(f (x) − Tm0 (x))| m X ≤ m m:|xk − Nk ≤ 2d |f (x) − Tm0 (x)| 1 |< N ∀k max m m:|xk − 2d k N max m m:|xk − 2d dn ≤ n! 1 |< N ∀k  1 N X k N Finally, we define our approximation to f (x) as f 0 (x): f 0 (x) , X X 0 βm,n fm,n (x). (23) m∈{0,··· ,N }d n:0<|n|<n Using Equation (22), we get the error bound of the approximation to f2 (x) as in Equation (24). |f (x) − Tm (x)|+ 1 |< N ∀k |Tm (x) − Tm0 (x)| |f 0 (x) − f2 (x)| = | X X  0 (x) − fm,n (x) | βm,n fm,n m∈{0,··· ,N }d n:|n|<n n n max ess sup |D f (x)|+ n:|n|=n x∈[0,1]d =| X X  0 βm,n fm,n (x) − fm,n (x) |  m:x∈suppψm n:|n|<n Dn f m n X 0 2 max |βm,n − | (x − ) ≤ 2d max |fm,n (x) − fm,n (x)| n! N m:x∈suppψ m n:|n|<n n:|n|<n  n d n 2d dn 1 ≤ 2 d (d + n − 1) ×0 . + ≤ n! N (24)  n  1  n−1 ! The second line follows again the support property and d 2 d d d + ··· + statement (i) of Proposition 2. The third line uses the bound n N N N |βm,n | ≤ 1.  n   d 1 ≤ 2d 1+ Then the final approximation error bound is as follows: N n! (18) |f 0 (x) − f (x)| ≤ |f 0 (x) − f2 (x)| + |f (x) − f2 (x)| The second step follows ψm (x) = 0 when x ∈ / suppψm . In  n d d n d+1 the third step we turn the sum to multiplication, because for ≤ 2 d (d + n − 1) ×0 + 2 . d N any x there are up to 2 terms ψm (x) that are not equal to (25) zero. The fourth step uses a Lagrange’s form of the Taylor Using Equation (8) and choosing r as remainder. The fifth step follows different round precision of βm,n in different order and the fact that the number of log (6N n (d + n − 1)) terms with order i is not greater than di . r= − 1, (26) 2 We rewrite f2 as the approximation error turns to X X f2 = βm,n fm,n (x), (19)  n d m∈{0,··· ,N }d n:|n|<n 0 d (27) . |f (x) − f (x)| ≤ 3 · 2 N where   n m fm,n (x) = ψm x − . (20) For any given  > 0, we need N Note that βm,n is a constant and thus f2 is a linear combi 1 nation of at most dn (N + 1)d terms of fm,n (x). Note that 3 · 2d dn n N≥ . (28) when d = 1, the number of terms should be n(N + 1)d  instead; but for simplicity of presentation we loosely use the same expression as they are on the same order. We now present the construction of the network for f 0 (x). 0 0 We define an approximation to fm,n (x) as fm,n (x). The only If every fm,n (x) can be computed by a sub-network, then 0 0 0 difference between fm,n (x) and fm,n (x) is that all multiplif (x) is simply a weighted sum of all outputs of fm,n (x). 0 cation operations are approximated by × as discussed in By Proposition 3, we can implement the needed weights d  On the Universal Approximability of Quantized ReLU Neural Networks     m|n|  m1  0  0 fm,n (x) − fm,n (x) = ×0 h(3N x1 − 3m1 ), · · · , ×0 h(3N xd − 3md ), ×0 xi1 − , × · · · , xi|n| − N N      mi|n|  mi1  − h(3N x1 − 3m1 ) h(3N x2 − 3m2 ) · · · h(3N xd − 3md ) xi1 − · · · xi|n| − N N ≤ (d + |n|)×0 i1 , · · · , i|n| ∈ {1, 2, · · · , d} (22) Table 1. Block configuration for function-independent structure Figure 2. ReLU implementation of f 0 (x). The connections of all Bfmn are the same. Every connection from BN to other blocks has no greater than two weights. βm,n by choosing t = log 0 to constructing fm,n (x). nN n dn . Then we simplify the task ×0 can be implemented as discussed in Proposition 2. For  i h( 3N xi − m , noticing that h(x) = h(−x), we can N mi mi i first compute |xi − m | as σ(x N  i − N )+σ( N −xi ) and then mi scale it to 3N |xi − N | . The implementation of h(x) can thus be simplified as 1−σ(x−1)+σ(x−2) since the input is nonnegative. Furthermore, by choosing N as cd2 where c ∈ n if n > 1. N and c > 1, N1 is an integral multiple of n1 Nd    n 2 When n = 1, N1 is an integral multiple of n1 Nd . As discussed in Proposition 3, we build a weight construction network Bw in the way that all integral multiples of the i minimal precision can be obtained. Therefore, all m N can be obtained in the same way as βm,n , except that we need to concatenate two weight construction sub-networks. So far we have proved that, for any f ∈ Fd,n there exists a network that approximates f arbitrary well with as few as 2 distinct weight values. This proves the universal approximability of qunatized ReLU networks, and answers the first question in Section 1. Now we analyze the complexity of the network. The implementation of f 0 (x) is shown in Figure 2. The function and size of blocks are listed in Table 1. While we can write the complexity of the network in an explicit expression, here we use the O notation for clarity. Let Nd , Nw , Nb be the depth, the number of weights, and the number of bits required respectively. The weight construction blocks Bw have the highest order of number of weights block function width Bw construct weights t depth  1  λ t λ−1 − 1 + 1 BN Babs Bs Bh B×0 construct N1 , · · · , NN−1 get absolute values scale by 3N implement h(x) implement ×0 (x, y) 2N 4 4 12 60 1 1 log 3N 2 1 3r + 1   1 and we have Nw = O λt λ−1 +1 N d . Meanwhile, we get   1 n Nd = O λt λ−1 + log N . Inserting t = log nN dn and  1 N = O 1 n , we get  1  !  1 n2d+3 λ−1 + log Nd = O λ log   (29) and n2d+3 Nw = O λ log   1 d !  λ−1 +1   n 1 .  (30) Multiplying Nw by log λ, Nb is obtained. This answers the second question in Section 1 and concludes the proof of Theorem 1. 4.4. Extension to Linear quantization With λ distinct values, a linearly quantized network has a minimal resolution of λ1 . The proof for the approximability of linear quantization can be done similarly except for a different sub-network for weight approximation with width t and depth logt λ +1. Here we give the conclusion for brevity: Given weight maximum precision λ1 , there exists a ReLU network that can approximate any f ∈ Fd,n with  any error 1 1  ∈ (0, 1) so that the depth is O log λ + 1 log  , the   2 1  nd number of weights is O log1 λ log 1 , and the   d    2 1 n number of bits is O log 1 , which are asymptot ically worse than those in the nonlinear quantization. On the Universal Approximability of Quantized ReLU Neural Networks 5. Function-dependent Structure The network complexity can be reduced if the network topology can be set according to a specific target function, i.e. function-independent structure. In this section, we provide an upper bound for function-dependent structure when d = 1 and n = 1, which is asymptotically better than that of fixed structure. Specifically, we first define an approximation to f (x) as fe(x) that has special properties to match the peculiarity of quantized networks. Then we use piecewise linear interpolation and “cached” functions (Yarotsky, 2017) to approximate fe(x) by ReLU networks. Figure 3. An example to illustrate the relationship between f (x), f + and fe(x). 5.1. Function Transformation While simply using piecewise linear interpolation at the scale of  can satisfy the error bound with O () weights, the complexity can be reduced by first doing interpolation at a coarser scale and then fill the details in the intervals to make the error go down to . By assigning “cached function” to every interval depending on specific functionand proper  1 scaling, the number of weights is reduced to O  log(1/) without constraints on weights (Yarotsky, 2017). The key difficulty in applying this approach to quantized ReLU networks is that the required linear interpolation at i T exactly where i = 1, 2, · · · , T is not feasible because of the constraint on weight selection. To this end, we transform f (x) to fe such that the approximation  error is bounded; the Lipschitz constant is preserved; i e f T are reachable under the constraints of weight selection without increasing the requirement on weight precision. We first divide [0, 1] into T intervals uniformly and define f + (x) as       T f dTTxe −t dT xe 2  , f + (x) , f (x) +    T −f 2−t T   (31) where t ∈ N. Note that f + ( Ti ) =  T f ( Ti ) 2−t  −t 2 T and + df (x) dx (x) = dfdx on( Ti , i+1 T ) where , i = 1, 2, · · · , T . Then, e we define f (x): & ' T f (0) 2−t , (32) fe(0) , −t 2 T dfe(x) , dx ( sgn(f + (x) − fe(x)) df (x) dx fe(x) 6= f + (x) (33) fe(x) = f + (x). f + (x) is parallel with f (x) on any given interval ( Ti , i+1 T ] −t and is shifted to let f + ( Ti ) be an integral multiple of 2T . −t fe( Ti ) is f ( Ti ) rounded up to an integral multiple of 2T . Meanwhile, fe(x) approaches f (x) with fixed slope whenever they are not equal. An example is shown in Figure 3. Proposition 4. fe(x) defined above has the following three properties. (i)fe(x) is a continuous function with Lipschitz constant 1;   T f ( Ti ) 2−t (ii) fe( Ti ) = 2−t T ; (iii) |fe(x) − f (x)| < 2−t T ; Proof. Statement (i) follows the definition directly. We give a proof for statement (ii) by contradiction and induction. If fe( Ti ) = f + ( Ti ), then fe(x) = f + (x) on ( Ti , i+1 T ]. The proof idea for fe( Ti ) > f + ( Ti ) and fe( Ti ) < f + ( Ti ) are basically the same thus we present the  proofof the latter for T f ( Ti ) 2−t brevity. We first assume fe( Ti ) = 2−t T , which is satisfied by Equation (32) when i = 0. Since fe(x) = f + (x) + e after the first intersection on ( Ti , i+1 T ], if f (x) and f (x) + i+1 e i+1 have no intersection on ( Ti , i+1 T ], then f ( T ) < f ( T ) e (x) because fe( Ti ) < f + ( Ti ). Meanwhile, we have dfdx = 1, and     Z i+1 e T i+1 i df (x) fe = fe + dx i T T dx T & ' T f Ti 2−t 1 = + −t 2 T T &  ' T f Ti + T1 2−t (34) = 2−t T & ' T f i+1 2−t T ≥ −t 2 T   i+1 ≥ f+ . T On the Universal Approximability of Quantized ReLU Neural Networks + i+1 This contradicts the fact that fe( i+1 T ) < f ( T ). Thus i + fe(x) and f (x)  intersect  on (0, T ] and in turn guarantee i+1 T f ( T ) 2−t that fe( i+1 T )= 2−t T . By induction, we prove that fe(x) and f + (x) intersect on every interval ( Ti , i+1 T ]. This implies statement (ii). Now we prove the statement (iii). Note that we have −t 0 ≤ fe( Ti ) − f ( Ti ) ≤ 2T by statement (ii). In every + e interval ( Ti , i+1 T ), f (x) = f (x) after the their intersec−t tion. Therefore we have 0 ≤ |f + (x) − f (x)| ≤ 2T by Equation (31). Before the intersection, if fe(0) < f + (0), + d(fe(x)−f + (x)) (x) ≥ 0. Since dfdx(x) = dfdx on ( Ti , i+1 dx T ), d(fe(x)−f (x)) i i e we have ≥ 0, thus 0 ≤ f ( T ) − f ( T ) ≤ dx −t fe(x) − f (x) ≤ f + (x) − f (x) ≤ 2T . If fe( Ti ) ≥ f + ( Ti ), + apply the same logic and we obtain 0 ≤ f (x) − f (x) ≤ −t fe(x)−f (x) ≤ fe( Ti )−f ( Ti ) ≤ 2T . This implies statement (iii) and concludes the proof. Table 2. Block configuration for function-dependent structure. Terms with lower order are omitted. block BT Bδ Bm Bf T BΦ Bs1 Bs2 Bs3 Bf1∗ Bf2∗ Bf3∗ Bf 00 function construct 1, 2, · · · , T construct δ 1 , · · · , m−1 construct m m implement f T (x) implement Φ(x) scale by T scale by T1 scale by 1δ first layer of f ∗ second layer of f ∗ third layer of f ∗ implement f 00 (x) width log T 1 t T 4T 4 1 4 T 3m m3m 4 depth log T 1 λ(t λ−1 ) 1 λ(t λ−1 ) 1 1 log T log T log m 1 1 1 1 1 8 t = log m, δ = 8m , T =  log 1 , and denote the output of  00 the network as f (x), we have |f (x) − f 00 (x)| = |f (x) − fe(x)| + |fe(x) − f 00 (x)| ≤ . The approximation network is shown in Figure 4. The sizes of blocks are given in Table 2 where f T is the uniform linear interpolation function of f with T − 1 breakpoints, f ∗ is the sum of the selected cached functions, Φ(x) is a filtering function. Figure 4. ReLU implementation of f 00 (x). Illustration only, details are omitted. For example, arrows between blocks can represent more than one connection in the network and there are short-cut connections that allow scaling only part of the inputs. 5.2. Implementation by ReLU Networks Theorem 2. For any f ∈ F1,1 , given λ distinct weights, there is a ReLU network with function-dependent structure that can approximate f with any error  ∈ (0, 1), so that   1  (i) the depth is O λ log log 1 λ−1 + log 1 ; (ii) the number of weights is    1 +1 O λ log log 1 λ−1 + 1 (iii) the number of bits needed to store the network is     1 +1 O log λ λ log log 1 λ−1 + 1 . Proof. We use the construction idea from (Yarotsky, 2017) and the weight construction method described in Section  4.2. By choosing the hyper-parameters as m = 12 log 1 , The inputs connections to Bf1∗ and the connections inside Bm have higher order to the number of weights than others.  The total depth is  1  O λ log log 1 λ−1 + log 1 . The total number of    1 +1 weights required is O λ log log 1 λ−1 + 1 and the     1 +1 number of bits is O log λ λ log log 1 λ−1 + 1 . 5.3. Extension to Linear Quantization Using the same construction method, we obtain the following conclusion: For any f ∈ F1,1 , given weight maximum precision λ1 , there is a function-dependent ReLU network that can approximate  f with any error  ∈ (0, 1) so that  the 1 depth is O log 1 , the number of weights is O  , and   the number of bits is O log λ  . 6. Conclusion In this paper, we have presented an analysis of the representation power of quantized neural networks. The proof for the universal approximability of quantized ReLU networks and upper bounds for the number of parameters and storage size with given approximation accuracy and bit-width are given for the first time. On the Universal Approximability of Quantized ReLU Neural Networks References Choi, Yoojin, El-Khamy, Mostafa, and Lee, Jungwon. Towards the limit of network quantization. arXiv preprint arXiv:1612.01543, 2016. Courbariaux, Matthieu, Bengio, Yoshua, and David, JeanPierre. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems, pp. 3123–3131, 2015. Denton, Emily L, Zaremba, Wojciech, Bruna, Joan, LeCun, Yann, and Fergus, Rob. Exploiting linear structure within convolutional networks for efficient evaluation. In Advances in Neural Information Processing Systems, pp. 1269–1277, 2014. Gysel, Philipp, Motamedi, Mohammad, and Ghiasi, Soheil. Hardware-oriented approximation of convolutional neural networks. arXiv preprint arXiv:1604.03168, 2016. Han, Song, Mao, Huizi, and Dally, William J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015a. Han, Song, Pool, Jeff, Tran, John, and Dally, William. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems, pp. 1135–1143, 2015b. Hanin, Boris. Universal function approximation by deep neural nets with bounded width and relu activations. arXiv preprint arXiv:1708.02691, 2017. Hubara, Itay, Courbariaux, Matthieu, Soudry, Daniel, ElYaniv, Ran, and Bengio, Yoshua. Quantized neural networks: Training neural networks with low precision weights and activations. arXiv preprint arXiv:1609.07061, 2016. Iandola, Forrest N, Han, Song, Moskewicz, Matthew W, Ashraf, Khalid, Dally, William J, and Keutzer, Kurt. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and¡ 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. Langley, P. Crafting papers on machine learning. In Langley, Pat (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stanford, CA, 2000. Morgan Kaufmann. Leng, Cong, Li, Hao, Zhu, Shenghuo, and Jin, Rong. Extremely low bit neural network: Squeeze the last bit out with admm. arXiv preprint arXiv:1707.09870, 2017. Li, Fengfu, Zhang, Bo, and Liu, Bin. Ternary weight networks. arXiv preprint arXiv:1605.04711, 2016. Li, Hao, De, Soham, Xu, Zheng, Studer, Christoph, Samet, Hanan, and Goldstein, Tom. Towards a deeper understanding of training quantized neural networks. Li, Hao, De, Soham, Xu, Zheng, Studer, Christoph, Samet, Hanan, and Goldstein, Tom. Training quantized nets: A deeper understanding. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 30, pp. 5807–5817. 2017a. Li, Hao, De, Soham, Xu, Zheng, Studer, Christoph, Samet, Hanan, and Goldstein, Tom. Training quantized nets: A deeper understanding. In Advances in Neural Information Processing Systems, pp. 5813–5823, 2017b. Liang, Shiyu and Srikant, R. Why deep neural networks for function approximation? 2016. Liu, Baoyuan, Wang, Min, Foroosh, Hassan, Tappen, Marshall, and Pensky, Marianna. Sparse convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 806–814, 2015. Mhaskar, Hrushikesh N and Micchelli, Charles A. Approximation by superposition of sigmoidal and radial basis functions. Advances in Applied mathematics, 13(3):350– 373, 1992. Rastegari, Mohammad, Ordonez, Vicente, Redmon, Joseph, and Farhadi, Ali. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pp. 525–542. Springer, 2016. Shaham, Uri, Cloninger, Alexander, and Coifman, Ronald R. Provable approximation properties for deep neural networks. Applied and Computational Harmonic Analysis, 2016. Shayar, Oran, Levi, Dan, and Fetaya, Ethan. Learning discrete weights using the local reparameterization trick. arXiv preprint arXiv:1710.07739, 2017. Sonoda, Sho and Murata, Noboru. Neural network with unbounded activation functions is universal approximator. Applied and Computational Harmonic Analysis, 43(2): 233–268, 2017. Sze, Vivienne, Chen, Yu-Hsin, Yang, Tien-Ju, and Emer, Joel. Efficient processing of deep neural networks: A tutorial and survey. arXiv preprint arXiv:1703.09039, 2017. Wen, Wei, Wu, Chunpeng, Wang, Yandan, Chen, Yiran, and Li, Hai. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems, pp. 2074–2082, 2016. On the Universal Approximability of Quantized ReLU Neural Networks Xu, Bing, Wang, Naiyan, Chen, Tianqi, and Li, Mu. Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853, 2015. Yarotsky, Dmitry. Error bounds for approximations with deep relu networks. Neural Networks, 94:103–114, 2017. Zhao, Liang, Liao, Siyu, Wang, Yanzhi, Tang, Jian, and Yuan, Bo. Theoretical properties for neural networks with weight matrices of low displacement rank. arXiv preprint arXiv:1703.00144, 2017. Zhou, Aojun, Yao, Anbang, Guo, Yiwen, Xu, Lin, and Chen, Yurong. Incremental network quantization: Towards lossless cnns with low-precision weights. arXiv preprint arXiv:1702.03044, 2017a. Zhou, Yiren, Moosavi-Dezfooli, Seyed-Mohsen, Cheung, Ngai-Man, and Frossard, Pascal. Adaptive quantization for deep neural network. arXiv preprint arXiv:1712.01048, 2017b. Zhu, Chenzhuo, Han, Song, Mao, Huizi, and Dally, William J. Trained ternary quantization. arXiv preprint arXiv:1612.01064, 2016. Zhuang, Bohan, Shen, Chunhua, Tan, Mingkui, Liu, Lingqiao, and Reid, Ian. Towards effective lowbitwidth convolutional neural networks. arXiv preprint arXiv:1711.00205, 2017.
1cs.CV
Asymptotic generalized bivariate extreme with random index arXiv:1701.04682v1 [math.ST] 17 Jan 2017 M. A. Abd Elgawada,b , A. M. Elsawaha,c,d,∗, Hong Qina and Ting Yana a Faculty of Mathematics and Statistics, Central China Normal University, Wuhan 430079, China b c d Department of Mathematics, Faculty of Science, Benha University, Benha 13518, Egypt Department of Mathematics, Faculty of Science, Zagazig University, Zagazig 44519, Egypt Division of Science and Technology, BNU-HKBU United International College, Zhuhai 519085, China Abstract In many biological, agricultural, military activity problems and in some quality control problems, it is almost impossible to have a fixed sample size, because some observations are always lost for various reasons. Therefore, the sample size itself is considered frequently to be a random variable (rv). The class of limit distribution functions (df’s) of the random bivariate extreme generalized order statistics (GOS) from independent and identically distributed rv’s are fully characterized. When the random sample size is assumed to be independent of the basic variables and its df is assumed to converge weakly to a non-degenerate limit, the necessary and sufficient conditions for the weak convergence of the random bivariate extreme GOS are obtained. Furthermore, when the interrelation of the random size and the basic rv’s is not restricted, sufficient conditions of the convergence and the forms of the limit df’s are deduced. Illustrative examples are given which lend further support to our theoretical results. Keywords: Weak convergence; Random sample size; Generalized order statistics; Generalized bivariate extreme. 1 Introduction The concept of generalized order statistics (GOS) have been introduced by Kamps (1995). It’s enable a unified approach to ascendingly ordered random variables (rv’s) as ordinary order statistics (oos), sequential order statistics (sos), order statistics with non integral sample size, progressively type II censored order statistics (pos), record values, kth record ∗ Corresponding author. E-mail: [email protected], [email protected], a [email protected] 1 values and Pfeifer’s records. Let γn = k > 0, γr = k+n−r+ Pn−1 j=r mj > 0, r = 1, 2, ..., n−1, and m̃ = (m1 , m2 , · · · , mn−1 ) ∈ ℜn−1 . Then the rv’s Xr:n ≡ X(r, n, m̃, k), r = 1, 2, ..., n, are called GOS based on the distribution function (df) F with density function f which are defined by their probability density function (pdf)    n n−1 Y Y (m̃,k) f1,2,...,n:n(x1 , x2 , ..., xn ) =  γj   (1 − F (xj ))γj −γj+1 −1 f (xj ) j=1 j=1 × (1 − F (xn ))γn −1 f (xn ), where F −1 (0) ≤ x1 ≤ ... ≤ xn ≤ F −1 (1). In this work, we consider a wide subclass of GOS , by assuming γj − γj+1 = m + 1 > 0. This subclass is known as m−GOS. Clearly many important practical models of m−GOS are included such as oos, order statistics with non integer sample size and sos. The marginal df’s of the rth and r̀th m−GOS (Nasri- Roudsari, 1996 and Barakat, 2007) are (m,k) (m,k) represented by Φr:n (x) = ILm (x) (r, N − r + 1) and Φr̀:n (x) = ILm (x) (N − Rr + 1, Rr ), R x a−1 1 (1 − respectively, where r̀ = n − r + 1, Lm (x) = 1 − (1 − F (x))m+1 , Ix (a, b) = β(a,b) 0 t t)b−1 dt denotes the incomplete beta ratio function, N = ℓ + n − 1, Rr = ℓ + r − 1 and ℓ= k m+1 . Moreover, by using the results of Kamps (1995), we can write explicitly the joint df’s of the rth and sth m−GOS, m 6= −1, 1 ≤ r < s ≤ n, as: Z F (x) Z F (y) m m+1 r−1 m+1 (m,k) Φr,s:n (x, y) = Cn⋆ ξ η γs −1 (1 − ξ ) (ξ − η m+1 )s−r−1 dηdξ, x ≤ y, 0 ξ (1.1) where Cn⋆ = (m+1)2 Γ(N +1) Γ(N −s+1)(r−1)!(s−r−1)! and Γ(.) is the usual gamma function. Recently Barakat et al. (2014a) studied the limit df’s of joint extreme m−GOS, for a fixed sample size. Moreover, the asymptotic behavior for bivariate df of the lower-lower (l-l), upper-upper (u-u) and lower-upper (l-u) extreme m−GOS in Barakat et al. (2014b). In the last few years much efforts had been devoted to investigate the limit df’s of independent rv’s with random sample size. The appearance of this trend is naturally because many applications require the consideration of such problem. For example, in many biological, agricultural and in some quality control problems, it is almost impossible to have a fixed sample size because some observations always get lost for various reasons. Therefore, the sample size n itself is considered frequently to be a rv νn , where νn is independent of the basic variables (i.e., the original random sample) or in some applications the interrelation of the basic variables and the random sample size is not restricted. Limit theorems for extremes with random sample size indexes have been thoroughly studied in the above mentioned two particular cases : 2 1. The basic variables and sample size index are independents (see, Barakat, 1997). 2. The interrelation of the basic variables and the random sample size is not restricted (see, Barakat and El Shandidy, 1990, Barakat, 1997 and Barakat et al., 2015a). Our aim in this paper is to characterize the asymptotic behavior of the bivariate df’s of the (l-l), (u-u) and (l-u) extreme m− GOS with random sample size. When the random sample size is assumed to be independent of the basic variables and its df is assumed to converge weakly to a non-degenerate limit, the necessary and sufficient conditions for the weak convergence of the random bivariate extreme m−GOS are obtained. Furthermore, when the interrelation of the random size and the basic rv’s is not restricted, sufficient conditions of the convergence and the forms of the limit df’s are deduced. An illustrative examples are given which lend further support to our theoretical results. Throughout this paper the convergence in probability and the weak convergence, as n → ∞, respectively, p w ”. denoted as ” −→ ” and ” −→ n n 2 Asymptotic random bivariate extreme under m−GOS 2.1 Random sample size and basic rv’s are independents In this subsection we deal with the weak convergence of bivariate df’s of the (u-u), (l-l) and (l-u) extreme m−GOS are fully characterized in Theorems 2.1, 2.2 and 2.3, respectively. When the sample size itself is a rv νn , which is assumed to be independent of the basic variables Xr:n , r = 1, 2, ..., n. Theorem 2.1. Consider the following three conditions : (m,k) (n) (n) (n) w (m,k) Φ̂r,s (x, y), x ≤ y, Φr̀,s̀:n (xn , yn ) = P (Zr̀,s̀:n < x) = P (Zr̀:n < x, Zs̀:n < y) −→ n Hn (nz) = P ( (m,k) νn w H(z), < z) −→ n n (n) (n) (n) (i) (ii) w (m,k) Ψ̂r,s (x, y) Φr̀,s̀:νn (xn , yn ) = P (Zr̀,s̀:νn < x) = P (Zr̀:νn < x, Zs̀:νn < y) −→ n Z ∞ (m,k) = (iii) Ωr,s (zκ1 , zκ2 )dH(z). 0 Then any two of the above conditions imply the remaining one, where xn = x1n = an x1 + bn , yn = x2n = an x2 + bn , an > 0, bn are suitable normalizing constants, x = (x1 , x2 ) = (n) (x, y), r̀ = n − r + 1 < n − s + 1 = s̀, Zr`i :n = Xr` :n −bn i ,i an 3 (m,k) = 1, 2, (r`1 , r`2 ) = (r̀, s̀), Φ̂r,s (x, y) is a non-degenerate df, H(z) is a df with H(+0) = 0, (m,k) Ωr,s (κ1 , κ2 ) Γr (x) = 1 Γ(r) = Rx 0    1 − ΓRs (κm+1 ), 2 m+1   1 − ΓRr (κ1 ) − 1 Γ(Rr ) x ≥ y, R∞ κm+1 1 I κm+1  (Rs , Rr − Rs )uRr −1 e−u du, x ≤ y, 2 u θ r−1 e−θ dθ denotes the incomplete gamma ratio function, κi = Uj;α(xi ), i = α 1, 2, j ∈ {1, 2, 3}, U1;α (xi ) = x−α i , xi ≤ 0; U2;α (xi ) = (−xi ) , xi > 0, α > 0 and U3 (xi ) = U3;α (xi ) = e−xi , ∀ xi . (m,k) Remark 2.1. The continuity of the limit df Φ̂r,s (m,k) the limit Ψ̂r,s (x, y) in (i) implies the continuity of (x, y). Hence the convergence in (iii) is uniform with respect to x and y. Remark 2.2. It is natural to look for the limitations on νn , under which we get the (m,k) relation Φ̂r,s (m,k) (x, y) ≡ Ψ̂r,s (x, y) ∀ x, y. In view of Theorem 2.1, the last equation is satisfied if and only if the df H(z) is degenerate at one, which means the asymptotically almost randomlessness of νn . We assume, due to Remark 2.2, that H(z) is a non-degenerate df and H(+0) = 0, i.e., continuous at zero. (m,k) Proof of the implication (i) + (ii) ⇒ (iii): First, we note that Φr̀,s̀:n (xn , yn ) can be written in the form (see, Theorem 2.3 in Barakat et al., 2014b) Z N 1 (m,k) Φr̀,s̀:n (xn , yn ) = 1 − ΓRr (N Lm (xn )) − I NLm (yn ) (Rs , Rr − Rs )uRr −1 e−u du, Γ(Rr ) N Lm (xn ) u (2.1) where Lm (.) = 1 − Lm (.). Now by using the total probability rule we get, (m,k) Φr̀,s̀:νn (xn , yn ) = ∞ X (m,k) Φr̀,s̀:t (xn , yn )P (νn = t). (2.2) t=r Assume that Hn (z) = P t≤z P (νn = t) = P (νn ≤ z) and z = [ nt ], where [θ] denotes the greatest integer part of θ. Thus, the relation (2.1) show that the sum term in (2.2) is a Riemann sum of the integral (m,k) Φr̀,s̀:νn (xn , yn ) = Z ∞ 0 (m,k) Φr̀,s̀:n (xn , yn , z)dHn (nz), (2.3) where, for sufficiently large n, we have (m,k) Φr̀,s̀:n (xn , yn , z) 1 = 1−ΓRr (z Ǹ Lm (xn ))− Γ(Rr ) Z z Ǹ z ǸLm (xn ) I zǸLm (yn ) (Rs , Rr −Rs )uRr −1 e−u du, u where Ǹ = ( ℓ−1 z + n) ∼ n. Appealing to the condition (i), Theorem 2.3 in Barakat et al. (2014b) and Remark 2.2, we get (m,k) w Ω(m,k) (zκ , zκ ), Φr̀,s̀:n (xn , yn , z) −→ 1 2 r,s n 4 (2.4) where the convergence is uniform with respect to x and y, over any finite interval of z. Now, let ξ be a continuity point of H(z) such that 1 − H(ξ) < ǫ, (ǫ is an aribtary small value). Then, we have Z ∞ Z (m,k) Ωr,s (zκ1 , zκ2 )dH(z) ≤ ξ ∞ ξ dH(z) = 1 − H(ξ) < ǫ. (2.5) In view of the condition (ii) we get, for sufficiently large n, that Z ∞ (m,k) Φr̀,s̀:n (xn , yn , z)dHn (nz) ≤ 1 − Hn (nξ) ≤ 2(1 − H(ξ)) < 2ǫ. (2.6) ξ On the other hand, by the triangle inequality, we get Z ξ Z ξ (m,k) (m,k) Φr̀,s̀:n (xn , yn , z)dHn (nz) − Ωr,s (zκ1 , zκ2 )dH(z) 0 ≤ + Z 0 ξ 0 Z ξ 0 (m,k) Φr̀,s̀:n (xn , yn , z)dHn (nz) − Z (m,k) Ωr,s (zκ1 , zκ2 )dHn (nz) − ξ 0 Z ξ 0 (m,k) Ωr,s (zκ1 , zκ2 )dHn (nz) (m,k) Ωr,s (zκ1 , zκ2 )dH(z) , (2.7) where the convergence in (2.4) is uniform over the finite interval 0 ≤ z ≤ ξ. Therefore, for arbitrary ǫ > 0 and for sufficiently large n, we have Z ξh i (m,k) (m,k) Φr̀,s̀:n (xn , yn , z) − Ωr,s (zκ1 , zκ2 ) dHn (nz) ≤ ǫHn (nξ) ≤ ǫ. (2.8) 0 In order to estimate the second difference on the right hand side of (2.7), we construct Riemann sums which are close to the integral there. Let T be a fixed number and 0 = ξ0 < ξ1 < ... < ξT = ξ be continuity points of H(z). Furthermore, let T and ξi be such that Z ξ 0 (m,k) Ωr,s (zκ1 , zκ2 )dHn (nz) − T X (m,k) Ωr,s i=1 (ξi κ1 , ξi κ2 )(Hn (nξi ) − Hn (nξi−1 )) < ǫ and Z ξ 0 (m,k) Ωr,s (zκ1 , zκ2 )dH(z) − T X (m,k) Ωr,s i=1 (ξi κ1 , ξi κ2 )(H(ξi ) − H(ξi−1 )) < ǫ. w H(ξ ), 0 ≤ i ≤ T, the two Riemann sums are closer Since, by the assumption Hn (nξi ) −→ i n to each other than ǫ for all n sufficiently large. Thus, once again by the triangle inequality, the absolute value of the difference of the integrals is smaller than 3ǫ. Combining this fact with (2.8), the left hand side of (2.7) becomes smaller than 4ǫ for all large n. Therefore, in view of (2.5), (2.6) and (2.4), we have Z Z ξ (m,k) (m,k) (m,k) (x, y) < Φr̀,s̀:νn (xn , yn ) − Ψ̂r,s Φr̀,s̀:n (xn , yn , z)dHn (nz) − 0 5 ξ 0 (m,k) Ωr,s (zκ1 , zκ2 )dH(z) + Z ∞ ξ (m,k) Φr̀,s̀:n (xn , yn , z)dHn (nz) + Z ∞ ξ (m,k) Ωr,s (zκ1 , zκ2 )dH(z) < 7ǫ. This completes the proof of the first part of the theorem. Proof of the implication (i) + (iii) ⇒ (ii): Starting with the relation (2.3), we select a subsequence {n′ } of {n} for which Hn′ (n′ z) converges weakly to an extended df H ′ (z) (i.e., H ′ (∞) − H ′ (0) ≤ 1 and such a subsequence exists by the compactness of df’s). Then, by repeating the first part of the theorem for the subsequence {n′ }, (m,k) with the exception that we choose ξ so that H ′ (∞) − H ′ (ξ) < ǫ, we get Ψ̂r,s (x, y) = R ∞ (m,k) (m,k) (m,k) ′ 0 Ωr,s (zκ1 , zκ2 )dH (z). Since the function Ψ̂r,s (x, y) is a df, we get Ψ̂r,s (∞, ∞) = R∞ 1 = 0 dH ′ (z) = H ′ (∞) − H ′ (0), which implies that H ′ (z) is a df. Now, if Hn (nz) did not converge weakly, then we can select two subsequences {n′ } and {n′′ } such that w w H ′′ (z), where H ′ (z) and H ′′ (z) are df’s. In this H ′ (z) and Hn′′ (n′′ z) −→ Hn′ (n′ z) −→ ′′ ′ n n case, we get (m,k) Ψ̂r,s (x, y) = Z ∞ 0 Z (m,k) Ωr,s (zκ1 , zκ2 )dH ′ (z)= ∞ 0 (m,k) Ωr,s (zκ1 , zκ2 )dH ′′ (z). Thus, let (y → ∞), we get Z 0 ∞ ΓRr (zκm+1 )dH ′ (z) 1 = Z ∞ 0 ΓRr (zκm+1 )dH ′′ (z). 1 (2.9) Appealing to equation (2.9) and by using the same argument which is applied in the proof of the second part of Theorem 2.1 in Barakat, 1997, we can easily prove H ′ (z) = H ′′ (z). This complete the proof of the second part. Proof of the implication (ii) + (iii) ⇒ (i): For proving this part, we need first present the following lemma. Lemma 2.1. For all xi , i = 1, 2, we have (n) (2.10) (n) (2.11) [1 − ΓRri (N Lm (xin ))] − σi,N ≤ P (Zr̀i :n < xi ) ≤ [1 − ΓRri (N Lm (xin ))] + ρi,N and ΓRri (N Lm (xin )) − ρi,N ≤ P (Zr̀i :n ≥ xi ) ≤ ΓRri (N Lm (xin )) + σi,N , where 0 < ρi,N , σi,N −→ 0 (or equivalently, as n −→ ∞), (Rr1 , Rr2 ) = (Rr , Rs ) and N (r1 , r2 ) = (r, s). Proof. Since 0 ≤ ΓRri (x) ≤ 1 ∀ x, the proof of the lemma will immediately follow from the result of Smirnov (1952) (Theorem 3, p. 133, or Lemma 2.1 in Barakat, 1997). We now turn to the proof of the last part of Theorem 2.1. In view of Remark 2.1, we can (m,k) assume, without any loss of generality, that the df Ψ̂r,s 6 (x, y) is continuous. Therefore, (m,k) the condition (iii) will be satisfied for all univariate marginals of Ψ̂r,s (x, y), i.e., we have (m,k) w Ψ̂(m,k) (x ), i = 1, 2, Φr`i :νn (xin ) −→ i ri n (m,k) where Ψ̂ri (xi ) = We shall now prove R∞ 0 (2.12) (m,k) )]dH(z), i = 1, 2, is the marginals df’s of Ψ̂r,s [1−ΓRri (zκm+1 i (m,k) w Φ̂(m,k) (x ) = 1 − Γ (κm+1 ), i = 1, 2. Φr`i :n (xin ) −→ i Rri i ri n (x, y). (2.13) (n) In view of Lemma 2.1, we first show that the sequence {Zr`i :n }n , i = 1, 2, is stochastically bounded (see, Feller, 1979). If we assume the contrary, we would find εi,1 , εi,2 > 0 such that at least one of the two following relations (a) (b) lim n→∞ hni P (Zr`i :n ≥ xi ) ≥ εi,1 > 0, ∀ xi > 0, i = 1, 2, hni lim P (Zr`i :n < xi ) ≥ εi,2 > 0, ∀ xi < 0, i = 1, 2 n→∞ (n) is satisfied. The assertions (a) and (b) mean that the sequence {Zr`i :n }n , is not stochastically bounded at the left (−∞) and at the right (+∞), respectively. Let the assumption (a) be true. Since H(z) is non-degenerate df, we find ε0 > 0 and β > 0 such that P ν n n  ≥ β ≥ ε0 , for sufficiently large n. (2.14) Using the following well known inequality, for i = 1, 2,     hni (n) P Zr`i :n1 ≥ xi ≥ P Zr`i :n2 ≥ xi , ∀ n1 ≥ n2 . (2.15) We thus get the following inequalities, for sufficiently large n,     X (n) (n) P Zr`i :t ≥ xi P (νn = t) P Zr`i :νn ≥ xi ≥ t≥[nβ]     (n) hni ≥ P Zr`i :[nβ] ≥ xi P (νn ≥ [nβ]) ≥ ε0 P Zr`i :[nβ] ≥ xi , i = 1, 2, (note that P (νn ≥ [nβ]) ≥ P (νn ≥ nβ)). Therefore, lim n→∞ hni (n) P (Zr`i :νn ≥ xi ) ≥ ε0 lim P (Zr`i :[nβ] ≥ xi ). n→∞ hni Now, if we find ε′i,1 > 0 such that lim P (Zr`i :[nβ] ≥ xi ) ≥ ε′i,1 > 0, we get n→∞ hni lim P (Zr`i :νn ≥ n→∞ xi ) ≥ ǫ0 ε′i,1 > 0, which contradicts the right stochastic boundedness of the sequence (n) {Zr`i :νn }n and consequently contradicts the relation (2.12). However, if such an ε′i,1 > 0 does not exist we have lim n→∞ hni P (Zr`i :[nβ] ≥ xi ) = 0, which in view of Lemma 2.1 (relation hni (2.10)) leads to the following chain of implications (∀ xi > 0) P (Zr`i :[nβ] ≥ xi ) → 0 ⇒ ΓRri ([N β]Lm (xin )) → 0 ⇒ [N β]Lm (xin ) → 0 ⇒ N Lm (xin ) → 0 (since N Lm (xin ) → 0) 7 (n) ⇒ ΓRri (N Lm (xin )) → 0 ⇒ P (Zr`i :n ≥ xi ) → 0, which contradicts the assumption (a). Consider the assumption (b). Since H(z) is a df we can find a positive integer δ and real number α > 0 such that P( νn ≤ δ) ≥ α, n for sufficiently large n. (2.16) Therefore, in view of (2.16) and the inequality (2.15), we have (n) P (Zr`i :νn < xi ) ≥ δn X t=r (n) (n) P (Zr`i :t < xi )P (νn = t) ≥ P (Zr`i :δn < xi )P ( νn ≤ δ) n (n) ≥ αP (Zr`i :δn < xi ), i = 1, 2. hni hni lim P (Zr`i :νn < xi ) ≥ α n→∞ lim P (Zr`i :δn < xi ). By using Lemma 2.1 Hence, we get n→∞ (relation (2.11)) and applying the same argument as in the case (a), it is easy to show that the last inequality leads to a contradiction (the last inequality, in view of the assumption (n) (b)), which yields that the sequences {Zr`i :n }n , i = 1, 2, is not stochastically bounded at (n) the left. This completes the proof that the sequences {Zr`i :n }n , i = 1, 2, are stochastically (m,k) bounded. Now, if Φr`i :n (xin ), i = 1, 2, did not converge weakly, then we could select two ′ (m,k) (m,k) subsequences {n′ } and {n′′ } such that Φr`i :n′ (xin′ ) would converge weakly to Φ̂ri and ′′ (m,k) (m,k) Φr`i :n′′ (xin′′ ) to another limit df Φ̂ri (xi ) (xi ). In this case we get (by repeating the ′ ′′ first part of Theorem 2.1 for the univariate case and for the two subsequences {n }, {n }) (m,k) Ωri (xi ) = Z 0 ∞h ′ (m+1) 1 − ΓRri (zκi Z i ) dH(z) = 0 ∞h ′′ (m+1) 1 − ΓRri (zκi i ) dH(z). However, Lemma 3.2 in Barakat (1997) shows that the last equalities, cannot hold unless ′ ′′ κi ≡ κi , i = 1, 2. Hence the relation (2.13) is proved. Hence, the proof of Theorem 2.1 is completed. Let G and Gν be the classes of all possible limit df’s in (i) and (iii), respectively. The class G is fully determined by Barakat et al. (2014b). Furthermore, let S and Sν be the necessary and sufficient conditions for the validity of the relations (i) and (iii), respectively. The following corollary characterizes the class Gν . Corollary 2.1. For every df Ψ̂(m,k) (x, y) in Gν there exists a unique df Φ̂(m,k) (x, y) in G , such that Ψ̂(m,k) (x, y) is uniquely determined by the representation (iii). Moreover, S = Sν . ′ (m,k) Proof of corollary 2.1. Let us first prove the implication {Φ̂r,s ′ (m,k) =⇒ {Ψ̂r,s ′′ (m,k) Ψ̂r,s ′′ (m,k) (x, y) 6= Ψ̂r,s ′ (m,k) (x, y), while Φ̂r,s ′′ (m,k) (x, y) 6= Φ̂r,s ′ (m,k) (x, y)}. If we assume the contrary, we get Ψ̂r,s ′′ (m,k) (x, y) 6= Φ̂r,s (x, y)} (x, y) = (x, y). Appealing to the first part of Theorem 8 2.1 , we get R∞h 0 ′ (m+1) 1 − ΓRri (zκi i i R∞h ′′ (m+1) ) dH(z), i = 1, 2. ) dH(z) = 0 1 − ΓRri (zκi The last equalities, as we have seen before, from Lemma 3.2 in Barakat (1997), cannot hold unless κ′i = κ′′i , i = 1, 2. Therefore, Corollary 2.1 is followed as a consequence of Theorem 2.1 and the last implication. This completes the proof of Corollary 2.1. Theorem 2.2. Consider the following three conditions : w (m,k) (n) (n) (n) (m,k) Φr,s:n (xn , yn ) = P (Zr,s:n < x) = P (Zr:n < x, Zs:n < y) −→ Φr,s (x, y), x ≤ y, n Hn (nz) = P ( νn w H(z), < z) −→ n n (i) (ii) w (m,k) (n) (n) (n) (m,k) < y) −→ Ψr,s (x, y) < x, Zs:ν < x) = P (Zr:ν (xn , yn ) = P (Zr,s:ν Φr,s:ν n n n n n Z ∞ (m,k) Ωr,s = (zρ1 , zρ2 )dH(z). (iii) 0 Then any two of the above conditions imply the remaining one, where xn = cn x + dn , yn = (n) cn y +dn , cn > 0, dn are suitable normalizing constants, 1 ≤ r < s ≤ n, Zri :n = (m,k) 1, 2, Φr,s Xri :n −dn ,i cn = (x, y) is a non-degenerate df, H(z) is a df with H(+0) = 0,   Γs (ρ2 ), x ≥ y, (m,k) Ωr,s (ρ1 , ρ2 ) = R ρ1  1 r−1 e−u du, x ≤ y, (r−1)! 0 Γs−r (ρ2 − u)u ρi = Vj;β (xi ), i = 1, 2, j ∈ {1, 2, 3}, V1;β (xi ) = (−xi )−β , xi ≤ 0; V2;β (xi ) = xβi , xi > 0, β > 0 and V3 (xi ) = V3;β (xi ) = exi , ∀ xi . Theorem 2.3. Consider the following three conditions : (n) (n) (m,k) w (n) Φr(m,k) (x)Φ̂s(m,k) (y) < x, Zs̀:n < y) −→ Φr,s̀:n (xn , yn ) = P (Zr,s̀:n < x) = P (Zr:n n = Γr (ρ1 ))[1 − ΓRs (κm+1 ), 1 ≤ r, s ≤ n, 2 Hn (nz) = P ( νn w H(z), < z) −→ n n (n) (n) (m,k) (i) (ii) (m,k) w (n) Ωr(m,k) (zρ1 )Ωs < x, Zs̀:νn < y) −→ Φr,s̀:νn (xn , yn ) = P (Zr,s̀:νn < x) = P (Zr:ν n n (zκ2 ). (iii) Then any two of the above conditions imply the remaining one, where xn = cn x+dn , yn = (n) an y + bn , an , cn > 0, bn , dn are suitable normalizing constants, Zr:n = (n) Xr:n −dn , Zs̀:n cn = Xs̀:n −bn (m,k) (m,k) , Φr (x), Φ̂s (y) an are non-degenerate df’s, H(z) is a df with H(+0) = 0, Z ∞ Z ∞   (m,k) (m,k) Ωr Γr (zρ1 )dH(z) and Ωs (zρ1 ) = (zκ2 ) = 1 − ΓRs (zκm+1 ) dH(z). 2 0 0 Proof of Theorems 2.2 and 2.3. Without significant modifications, the method of the proof of Theorems 2.2 and 2.3 are the same as that Theorem 2.1, except only the obvious changes. Hence, for brevity the details of the proof are omitted. 9 2.2 The interrelation of νn and the basic rv’s is not restricted When the interrelation between the random index and the basic variables is not restricted, parallel theorem of Theorem 2.1 may be proved by replacing the condition (ii) by a stronger one. Namely, the weak convergence of the df Hn (nz) must be replaced by the convergence νn n in probability of the rv to a positive rv T . However, the key ingredient of the proof of this parallel result is to prove the mixing property, due to Rényi (see, Barakat and Nigm, 1996) of the sequence of order statistics under consideration. In the sense of Rényi a sequence {un } of rv’s is called mixing if for any event E of positive probability, the conditional df of {un }, under the condition E, converges weakly to a non-degenerate df, which does not depend on E, as n → ∞. The following lemma proves the mixing property (n) for the sequence {Zr̀,s̀:n }n . (n) Lemma 2.2. Under the condition (i) in Theorem 2.1 the sequence {Zr̀,s̀:n }n is mixing. (l) (n) Proof. The lemma will be proved if one shows the relation P (Zr̀,s̀:n < x | Zr̀,s̀:l < w Φ̂(m,k) (x, y), for all integers l = r, r + 1, .... The sufficiency of the above relation x) −→ r,s n can easily be proved as a direct multivariate extension of Lemma 6.2.1, of Galambos (1987). However, this relation is equivalent to hli (n) w Φ(m,k) (x, y), P (Zr̀,s̀:n ≥ x | Zr̀,s̀:l ≥ x) −→ r,s n (m,k) where Φr,s (m,k) (x, y) is the survival function of the limit df Φ̂r,s (m,k) Φr,s (2.17) (x, y), i.e., (m,k) (x, y) = 1 − Φ̂r(m,k) (x) − Φ̂s(m,k) (y) + Φ̂r,s (x, y). Therefore, our lemma will be established if one proves the relation (2.17). Now, we can write (l) (n) (n) (l) (n) P (Zr̀,s̀:n ≥ x | Zr̀,s̀:l ≥ x) = P (Zr̀,s̀:n ≥ x, Zr̀,s̀:l < x | Zr̀,s̀:l ≥ x) (n) (l) (n) + P (Zr̀,s̀:n ≥ x, Zr̀,s̀:l ≥ x | Zr̀,s̀:l ≥ x). (2.18) Bearing in mind that all Xr:n , r = 1, 2, ..., n, are i.i.d rv’s, the first term in (2.18) can be written in the form (n) (n) (l) (n) ∗(n) (l) P (Zr̀,s̀:n ≥ x, Zr̀,s̀:l < x | Zr̀,s̀:l ≥ x) = P (Zr̀,s̀:(n−l) ≥ x, Zr̀,s̀:l < x | Zr̀,s̀:l ≥ x) (n) ∗(n) ∗(n) (l) = P (Zr̀,s̀:(n−l) ≥ x) − P (Zr̀,s̀:(n−l) ≥ x, Zr̀,s̀:l ≥ x | Zr̀,s̀:l ≥ x), where ∗(n) ∗(n) ∗(n) Zr̀,s̀:(n−l) = (Zr̀:(n−l) , Zs̀:(n−l) ), 10 ∗(n) Zr̀:(n−l) = ((rth largest of X1,l+1:n−l , X1,l+2:n−l , ..., X1,n:n−l ) − bn )/an , ∗(n) Zs̀:(n−l) = ((sth largest of X2,l+1:n−l , X2,l+2:n−l , ..., X2,n:n−l ) − bn )/an . Therefore, in view of (2.18), we have ∗(n) (l) (n) P (Zr̀,s̀:n ≥ x | Zr̀,s̀:l ≥ x) = P (Zr̀,s̀:(n−l) ≥ x) − ∆n (x), (l) (n) ∗(n) (l) (l) (n) (2.19) where ∆n (x) = P (Zr̀,s̀:n ≥ x, Zr̀,s̀:l ≥ x | Zr̀,s̀:l ≥ x) − P (Zr̀,s̀:(n−l) ≥ x, Zr̀,s̀:l ≥ x | Zr̀,s̀:l ≥ T T (n) ∗(n) x). By using the well-known inequalities Zr̀,s̀:(n−l) ≤ Zr̀,s̀:n and P (B C) − P (A C) ≤ P (B) − P (A), for any three events A, B and C, for which A ⊆ B, we get ∗(n) (n) (l) 0 ≤ ∆n (x)P (Zr̀,s̀:l ≥ x) ≤ P (Zr̀,s̀:n ≥ x) − P (Zr̀,s̀:(n−l) ≥ x). (2.20) On the other hand, by virtue of the condition (i) in Theorem 2.1, it is easy to prove that ∗(n) (n) (m,k) lim P (Zr̀,s̀:(n−l) ≥ x) = lim P (Zr̀,s̀:(n−l) ≥ x) = Φr,s n→∞ n→∞ (x, y) (2.21) m+1 m+1 ⇒ (N − l)Lm (xin ) −→ , ∀ x′i s for which κi < (note that N Lm (xin ) −→ n κi n κi ∞, i = 1, 2 ). By combining the relations (2.19)-(2.21), the proof of the relation (2.17) follows immediately. Hence the required result. Considering the facts that the normalizing constants, which may be used in the bivari(m,k) ate extreme case are the same as those for the univariate case, and the limit df Φ̂r,s (x, y) is continuous, we can easily by using Lemma 2.2, show that the proof of the following theorem follows without any essential modifications as a direct multivariate extension of the proof of Theorem 2.1 in Barakat and El Shandidy (1990), except only the obvious changes. Theorem 2.4. Consider the condition νn p −→ T , n n (ii)′ where T is a positive rv. Under the conditions of Theorem 2.1 , we have the implication (i) + (ii)′ ⇒ (iii). 3 Illustrative examples The range and midrange are widely used, particularly in statistical quality control as an estimator of the dispersion tendency and in setting confidence intervals for the population standard deviation as well as in Monte Carlo methods. In fact, the range itself is a very simple measure of dispersion, gives a quick and easy to estimate indication about the 11 spread of data. Let us defines the random generalized ranges Rνn (m, k) = Xνn :νn −X1:νn and the random generalized midranges Vνn (m, k) = X1:νn +Xνn :νn , X1:νn 2 = X(1, νn , m, k) and Xνn :νn = X(νn , νn , m, k). The normalized generalized ranges and the normalized gen(n) (n) eralized midranges are defined by Rνn (m, k) = A−1 n:r (Rνn (m, k) − Bn:r ) and Vνn (m, k) = A−1 n:υ (Vνn (m, k) − Bn:υ ), respectively, where An:r = 2An:υ = an > 0, Bn:r = bn − dn and Bn:υ = 1 2 (bn + dn ) are suitable normalizing constants. In this section, some illustrative examples for the most important distribution functions are obtained, which lend further support to our theoretical results. In the following examples we consider an important practical situation when νn has a geometric distribution with mean n. In this case we can w H(z) = 1 − e−z (z ≥ 0). easily show that P ( νnn < z) −→ n Example 3.1 (standard Cauchy distribution). Let F (x) = 1 2 + 1 π tan−1 x. Then α = β = 1. In view of Theorems 1.1 and 2.1, Part 1, in Barakat et al. (2015b), we get, after some algebra,    1,   an = η = lim ηn = lim 0, n→∞ n→∞ cn     ∞, if m = 0, if m > 0, if − 1 < m < 0. The random generalized ranges and midranges, for standard Cauchy distribution are given by, if m = 0, P (R(n) νn (0, k) ≤ r) w −→ n Z 0 and (0, k) P (Vν(n) n ≤ v) w −→ n Z 0 ∞Z ∞ 0 ∞Z ∞ −∞  −1 1 − Γk (z(r − y)−1 ) y −2 ze−z(1+y ) dydz   −1 1 − Γk (z(v − y)−1 ) y −2 ze−z(1−y ) dydz, respectively, with An:r = 2An:υ = an , Bn:r = Bn:υ = 0. Moreover, if m > 0, w 1 − e−r P (R(n) νn (m, k) ≤ r) −→ n −1 and w −1 e−(−v) , (m, k) ≤ v) −→ P (Vν(n) n n respectively, with An:r = 2An:υ = cn , Bn:r = Bn:υ = 0. Finally, when −1 < m < 0, Z ∞h i w −(m+1) 1 − Γ (zr ) e−z dz P (R(n) (m, k) ≤ r) −→ ℓ νn n 0 (n) and the df of Vνn (m, k) converge weakly to the same limit, with An:r = 2An:υ = an , Bn:r = Bn:υ = 0. 12 Example 3.2 (Pareto distribution). It can be shown that, for the Pareto distribution F (x) = (1 − x−σ )I[1,∞) (x), σ > 0. Therefore, in view of Theorems 1.1 and 2.1, Part 1, in cn an Barakat et al. (2015b), since −→ n 0, ∀ m. Then w P (R(n) νn (m, k) ≤ r) −→ n Z 0 ∞h i 1 − Γℓ (zr −σ(m+1) ) e−z dz (n) and the df of Vνn (m, k) converge weakly to the same limit, where An:r = 2An:υ = an , Bn:r = Bn:υ = 0. Example 3.3 (uniform distribution). For the uniform (−θ, θ) distribution, by using Theorem 2.1, Part 1, in Barakat et al. (2015b), since (0, k) P (Rν(n) n Z w −→ n ≤ r) 0 and Z w (0, k) ≤ v) −→ P (Vν(n) n n 0 ∞Z 0 −∞ ∞Z ∞ −∞ an cn −→ n 1, if m = 0. Then [1 − Γk (z(y − r))] zez(y−1) dydz [1 − Γk (z(y − v))] ze−z(y+1) dydz, respectively, with An:r = 2An:υ = an , Bn:r = Bn:υ = 0. Example 3.4 (Beta(α, β) distribution). For the beta distribution F (x; α, β), 0 ≤ x ≤ 1, α, β > 0. Therefore, in view of Theorem 2.2, Part 5, in Barakat et al. (2015b), if α = (m + 1)β. Then (m, k) P (Rν(n) n ≤ r) w −→ n Z 0 ∞Z 0 α −∞ [1 − Γℓ (z(y − r)α )] zηα(−ηy)α−1 e−z(1+(−ηy) ) dydz and P (Vν(n) (m, k) n ≤ v) Z w −→ n 0 ∞Z ∞ −∞ 1 α [1 − Γℓ (z(y − v)α )] zηα(ηy)α−1 e−z(1+(ηy) ) dydz, 1 α respectively, where η = ( βc ) β ( m+1 cα ) , c = Γ(α+β) Γ(α)Γ(β) . Clearly, the same result holds for the power distribution F (x; α, 1). Example 3.5 (standard normal, logistic, Laplace, and log-normal distributions). After some algebra, we get,   √ 1 , for the normal distribution,   m+1  c n η −1 = lim = 1, for the logistic and Laplace distribution, n→∞ an     0, for the log-normal distribution. Moreover, for the standard normal, logistic, and Laplace distributions, we get w P (R(n) νn (m, k) ≤ r) −→ n Z 0 ∞Z ∞ −∞ h i −ηy 1 − Γℓ (ze(y−r)(m+1) ) zηe−ηy e−z(1+e ) dydz 13 and w P (Vν(n) (m, k) ≤ v) −→ n n Z 0 ∞Z ∞ −∞ h i ηy 1 − Γℓ (ze(y−v)(m+1) ) zηeηy e−z(1+e ) dydz, for the standard normal distribution (m = 0, k = 1). Then    f (r),  Z ∞  1 2 y dy w 2 P (Rνn (0, 1) ≤ an r + 2bn ) −→ = n 2 + y + e−r )2 3,  (y 0    f (r), 2 r < ln 4, r=ln 4, r > ln 4, where  4e−r 1 −1 √ f1 (r) = (4e − 1) cot ( √ )−1 , 4e−r − 1 4e−r − 1 # " √ 1 + 1 − 4e−r 2e−r −r −1 √ ln , 1− √ f2 (r) = (1 − 4e ) 1 − 4e−r 1 − 1 − 4e−r −r −1  √ 1 an = √ , bn = 2 ln n − 2 ln n and P (Vνn (0, 1) ≤ an v) w −→ n 1− Z 0 ∞ (y(e2v ln ln n + ln 4π √ 2 2 ln n dy = (1 + e−2v )−1 . + 1) + 1)2 Finally, for the log-normal distribution,  R∞ w (n) (n) 1 − Γℓ (ze−r(m+1) ) e−z dz and the df of Vνn (m, k) converge P (Rνn (m, k) ≤ r) −→ 0 n weakly to the same limit. Example 3.6 (exponential (σ) and Rayleigh (σ) distributions). In view of Theorem 2.2, Part 4 in Barakat et al. (2015b). Then R∞  w (n) (n) 1 − Γℓ (ze−r(m+1) ) e−z dz and the df of Vνn (m, k) converge P (Rνn (m, k) ≤ r) −→ 0 n weakly to the same limit, for exponential (σ) and Rayleigh (σ) distributions. Acknowledgements Elsawah’s work was partially supported by the UIC GRANT R201409 and the Zhuhai Premier Discipline Grant and Qin’s work was partially supported by the National Natural Science Foundation of China (Nos. 11271147, 11471135, 11471136). References [1] Barakat, H. M. (1997). Asymptotic properties of bivariate random extremes. J. Stat. Plann. Inference, 61, 203-217. [2] Barakat, H. M. (2007). Limit theory of generalized order statistics. J. Statist. Plann. Inference. Vol. 137, No. 1,1-11. 14 [3] Barakat, H. M. and El Shandidy, M. A. (1990). On the limit distribution of the extremes of a random number of independent random variables. J. Statist. Plann. Inference, 26, 353-361. [4] Barakat, H. M. and Nigm, E. M. (1996). The mixing property of order ststistics with some applications. Bull. Malaysian Math. Soc. (Second Series) 19, 39-52. [5] Barakat, H. M., Nigm, E. M. and Abd Elgawad, M. A. (2014a). Limit theory for joint generalized order statistics, REVSTAT Statistical Journal 12(3), 199-220. [6] Barakat, H. M., Nigm, E. M. and Abd Elgawad, M. A. (2014b). Limit theory for bivariate extreme generalized order statistics and dual generalized order statistics, ALEA, Lat. Am. J. Probab. Math. Stat. 11 (1), 331-340. [7] Barakat, H. M., Nigm, E. M. and Al-Awady, M. A. (2015a). Asymptotic properties of multivariate order statistics with random index. Bull. Malays. Math. Sci. Soc. 38(1), 289-301. [8] Barakat, H. M., Nigm, E. M. and Elsawah, A. M. (2015b). Asymptotic distributions of the generalized range, midrange, extremal quotient, and extremal product, with a comparison study. Comm. Statist. Theory Methods, 44, 900-913. [9] Cramer, E. (2003). Contributions to Generalized Order Statistics. Habililationsschrift, Reprint, University of Oldenburg. [10] Feller, W. (1979). An introduction to probability theory and its applications. Vol. 2, John Wiley & Sons. Inc. (Wiley Eastern University edition). [11] Galambos, J. (1978, 1987). The asymptotic theory of extreme order statistics, New York, Wiley, (1st ed.) Kreiger. FI (2nd ed.). [12] Kamps, U. (1995). A Concept of Generalized Order Statistics. Teubner, Stuttgart. [13] Nasri-Roudsari, D. (1996). Extreme value theory of generalized order statistics, J. Statist. Plann. Inference, 55, 281-297. [14] Smirnov, N.V. (1952). Limit distribution for terms of a variational series. Trans. Amer. Math. Soc. Ser. I. 11, 82-143. 15
10math.ST
arXiv:1403.0026v2 [math.GR] 22 Feb 2016 COMMENSURATIONS AND METRIC PROPERTIES OF HOUGHTON’S GROUPS JOSÉ BURILLO, SEAN CLEARY, ARMANDO MARTINO, AND CLAAS E. RÖVER Abstract. We describe the automorphism groups and the abstract commensurators of Houghton’s groups. Then we give sharp estimates for the word metric of these groups and deduce that the commensurators embed into the corresponding quasi-isometry groups. As a further consequence, we obtain that the Houghton group on two rays is at least quadratically distorted in those with three or more rays. Introduction The family of Houghton groups Hn was introduced by Houghton [7]. These groups serve as an interesting family of groups, studied by Brown [2], who described their homological finiteness properties, by Röver [10], who showed that these groups are all subgroups of Thompson’s group V , and by Lehnert [9] who described the metric for H2 . Lee [8] described isoperimetric bounds, and de Cornulier, Guyot, and Pitsch [4] showed that they are isolated points in the space of groups. Here, we classify automorphisms and determine the abstract commensurator of Hn . We also give sharp estimates for the word metric which are sufficient to show that the map from the abstract commensurator to the group of quasi-isometries of Hn is an injection. 1. Definitions and background Let N be the set of natural numbers (positive integers) and n ≥ 1 be an integer. We write Zn for the integers modulo n with addition and put Rn = Zn × N. We interpret Rn as the graph of n pairwise disjoint Date: March 20, 2018. The authors would like to thank Charles Cox for his remarks in an earlier version of this paper. The authors are also grateful for the hospitality of NUI Galway. The first author acknowledges support from MEC grant #MTM2011-25955. The second author acknowledges support from the National Science Foundation and that this work was partially supported by a grant from the Simons Foundation (#234548 to Sean Cleary). 1 2 BURILLO, CLEARY, MARTINO, AND RÖVER rays; each vertex (i, k) is connected to (i, k + 1). We denote by Symn , FSymn and FAltn , or simply Sym, FSym and FAlt if n is understood, the full symmetric group, the finitary symmetric group and the finitary alternating group on the set Rn , respectively. The Houghton group Hn is the subgroup of Sym consisting of those permutations that are eventually translations (of each of the rays). In other words, the permutation σ of the set Rn is in Hn if there exist integers N ≥ 0 and ti = ti (σ) for i ∈ Zn such that for all k ≥ N, (i, k)σ = (i, k + ti ); throughout we will use right actions. Note that necessarily the sum of the translations ti must be zero because the permutation needs of course to be a bijection. This implies that H1 ∼ = FSym. For i, j ∈ Zn with i 6= j let gij ∈ Hn be the element which translates the line obtained by joining rays i and j, given by (i, n)gij = (i, n − 1) if n > 1, (i, 1)gij = (j, 1), (j, n)gij = (j, n + 1) if n ≥ 1 and (k, n)gij = (k, n) if k ∈ / {i, j}. We also write gi instead of gi i+1 . It is easy to see that {gi | i ∈ Zn }, as well as {gij | i, j ∈ Zn , i 6= j}, are generating sets for Hn if n ≥ 3 as we can simply check that the commutator [g0 , g1] = g0−1g1−1 g0 g1 transposes (1, 1) and (2, 1). In the special case of H2 , g1 is redundant as g1 = g0−1. Further, an additional generator to g0 is required to generate the group; we choose τ which fixes all points except for transposing (0, 1) and (1, 1). It is now clear that the commutator subgroup of Hn is given by  FAlt, if n ≤ 2 ′ Hn = FSym, if n ≥ 3 For n ≥ 3, we thus have a short exact sequence (1) π 1 −→ FSym −→ Hn −→ Zn−1 −→ 1 where π(σ) = (t0 (σ), . . . , tn−2 (σ)) is the abelianization homomorphism. We note that as the sum of all the eventual translations must be zero, we have the last translation is determined by the preceding ones: (2) tn−1 (σ) = − n−2 X i=0 ti (σ). METRIC PROPERTIES OF HOUGHTON’S GROUPS 3 We will use the following facts freely throughout this paper, see Dixon and Mortimer [5] or Scott [11]. Lemma 1.1. The group FAlt is simple and equal to the commutator subgroup of FSym, and Aut(FAlt) = Aut(FSym) = Sym. 2. Automorphisms of Hn Here we determine the automorphism group of Hn . First we establish that we have to look no further than Sym. We let NG (H) denote the normalizer, in G, of the subgroup H of G. Proposition 2.1. Every automorphism of Hn , n ≥ 1, is given by conjugation by an element of Sym, that is to say Aut(Hn ) = NSym (Hn ). Proof. From the above, the finitary alternating group FAlt is the second derived subgroup of Hn , and hence characteristic in Hn . So every automorphism of Hn restricts to an automorphism of FAlt. Since Aut(FAlt) = Sym, this restriction yields a homomorphism Aut(Hn ) −→ Sym and we need to show that it is injective with image equal to NSym (Hn ). In order to see this let ψ ∈ Aut(Hn ) be an automorphism. Compose this with an inner automorphism (of Sym) so that the result is an (injective) homomorphism α : Hn → Sym whose restriction to FAlt is trivial. We let k ∈ N and consider the following six consecutive points aℓ = (i, k + ℓ) of Rn for ℓ ∈ {0, 1, . . . , 5}. We denote by giα the image of gi under α, and by (x y z) the 3-cycle of the points x, y and z. Using the identities gi−1 (a1 a2 a3 )gi = (a0 a1 a2 ) and gi−1 (a3 a4 a5 )gi = (a2 a3 a4 ) and applying α, which is trivial on FAlt, we get (giα )−1 (a1 a2 a3 )giα = (a0 a1 a2 ) and (giα)−1 (a3 a4 a5 )giα = (a2 a3 a4 ). Hence, we must have that giα maps {a1 , a2 , a3 } to {a0 , a1 , a2 }, and then also {a3 , a4 , a5 } to {a2 , a3 , a4 }. The conclusion is that it maps a3 to a2 . Applying a similar argument to all points in the branches i and i + 1, it follows that giα = gi , and since i was arbitrary, this means that α is the identity map.  With Lemma 1.1 in mind we now present the complete description of Aut(Hn ). Theorem 2.2. For n ≥ 2, the automorphism group Aut(Hn ) of the Houghton group Hn is isomorphic to the semidirect product Hn ⋊ Sn , where Sn is the symmetric group that permutes the n rays. 4 BURILLO, CLEARY, MARTINO, AND RÖVER Proof. By the proposition, it suffices to prove that every α ∈ Sym which normalizes Hn must map (i, k + m) to (j, l + m) for some k, l ≥ 1 and all m ≥ 0. To this end, we pick α ∈ NSym (Hn ) and σ ∈ Hn . Since σ α (= α−1 σα) is in Hn and maps the point xα to (xσ)α, these two points must lie on the same ray for all but finitely many x ∈ Rn . Similarly, x and xσ lie on the same ray for all but finitely many x ∈ Rn , as σ ∈ Hn . In fact, given a ray, we can choose σ so that whenever x lies on that ray, then x and xσ are successors on the same ray. Combining these facts, we see that α maps all but finitely many points of ray i to one and the same ray, say ray j. This defines a homomorphism from Aut(Hn ) onto Sn , which is obviously split, since given a permutation of the n rays, it clearly defines an automorphism of Hn . So assume that α is now in the kernel of that map, so it does not permute the rays, and take σ a gji generator of Hn , i.e., an infinite cycle inside Sym. Then, since conjugating inside Sym preserves cycle types, the element σ α ∈ Hn is also a single infinite cycle. This means that σ α has nonzero translations in only two rays, and these translations are 1 and −1. For all points in the support of σ, we have that σ k sends this point into the i-th ray for all sufficiently large k. Therefore, as α sends almost all points in the i-th ray into the i-th ray, the same is true for ασ . Hence ti (σ α ) is positive, so it must be ti (σ α ) = 1. It is quite clear now that α translates by an integer in the ray i, sufficiently far out. This finishes the proof since this could be done for any i, and hence α ∈ Hn .  3. Commensurations of Hn First, we recall that a commensuration of a group G is an isomorφ phism A −→ B, where A and B are subgroups of finite index in G. Two commensurations φ and ψ of G are equivalent if there exists a subgroup A of finite index in G, such that the restrictions of φ and ψ to A are equal. The set of all commensurations of G modulo this equivalence relation forms a group, known as the (abstract) commensurator of G, and is denoted Com(G). In this section we will determine the commensurator of Hn . For a moment, we let H be a subgroup of a group G. The relative commensurator of H in G is denoted ComG (H) and consists of those g ∈ G such that H ∩ H g has finite index in both H and H g . Similar to the homomorphism from NG (H) to Aut(H), there is a homomorphism from ComG (H) to Com(H); Its kernel consists of those elements of G that centralize a finite index subgroup of H. METRIC PROPERTIES OF HOUGHTON’S GROUPS 5 In order to pin down Com(Hn ), we first establish that every commensuration of Hn can be seen as conjugation by an element of Sym, and then characterize ComSym (Hn ). Since a commensuration φ and its restriction to a subgroup of finite index in its domain are equivalent, we can restrict our attention to the following family of subgroups of finite index in Hn , in order to exhibit Com(Hn ). For every integer p ≥ 1, we define the subgroup Up of Hn by Up = hFAlt, gip | i ∈ Zn i. We collect some useful properties of these subgroup first, where A ⊂f B means that A is a subgroup of finite index in B. Lemma 3.1. Let n ≥ 3. (i) For every p, the group Up coincides with Hnp , the subgroup generated by all pth powers in Hn , and hence is characteristic in Hn .   n−1 2p , p even FAlt, p even ′ . and |Hn : Up | = (ii) Up = pn−1 , p odd FSym, p odd (iii) For every finite index subgroup U of Hn , there exists a p ≥ 1 with FAlt = Up′ ⊂ Up ⊂f U ⊂f Hn . The same is essentially true for the case n = 2, except that Up′ is always equal to FAlt in this case, with the appropriate change in the index. Proof. First we establish (ii). We know that [gi , gj ] is either trivial, when j ∈ / {i − 1, i + 1}, or an odd permutation. So the commutator identities [ab, c] = [a, c]b [b, c] and [a, bc] = [a, c][a, b]c imply the first part, and the second part follows immediately using the short exact sequence (1) from Section 1. Part (i) is now an exercise, using that FAltp = FAlt. In order to show (iii), let U be a subgroup of finite index in Hn . The facts that FAlt is simple and U contains a normal finite index subgroup of Hn , imply that FAlt ⊂ U. Let p be the smallest even integer such that (pZ)n−1 is contained in the image of U in the abelianisation of Hn . It is now clear that Up is contained in U.  Noting that Com(H1 ) = Aut(H1 ) = Sym, we now characterize the commensurators of Houghton’s groups. Theorem 3.2. Let n ≥ 2. Every commensuration of Hn normalizes Up for some even integer p. The group Np = NSym (Up ) is isomorphic 6 BURILLO, CLEARY, MARTINO, AND RÖVER to Hnp ⋊ (Sp ≀ Sn ), and Com(Hn ) is the direct limit of Np with even p under the natural embeddings Np −→ Npq for q ∈ N. Proof. Let φ ∈ Com(Hn ). By Lemma 3.1, we can assume that Up is contained in the domain of both φ and φ−1 for some even p. Let V be the image of Up under φ. Then V has finite index in Hn and so contains FAlt, by Lemma 3.1. However, the set of elements of finite order in V equals [V, V ], whence [V, V ] = FAlt, as FAlt and FSym are not isomorphic. This means that the restriction of φ to FAlt is an automorphism of FAlt, and hence yields a homomorphism ι : Com(Hn ) −→ Sym. That ι is an injective homomorphism to ComSym (Hn ) follows from a similar argument to the one in Proposition 2.1 applied to gi p and six points of the form aℓ = (i, k + pℓ) with ℓ ∈ {0, 1, . . . , 5}. Since the centralizer in Sym of FAlt, and hence of any finite index subgroup of Hn , is trivial, the natural homomorphism from ComSym (Hn ) to Com(Hn ) mentioned above is also injective, and we conclude that Com(Hn ) is isomorphic to ComSym (Hn ), and that ι is the isomorphism: ι : Com(Hn ) −→ ComSym (Hn ). From now on, we assume that φ ∈ ComSym (Hn ). In particular, the action of φ is given by conjugation, and our hypothesis is that Upφ ⊂ Hn . Now we can apply the argument from the proof of Theorem 2.2 to σ ∈ U and σ φ (instead of σ α ). Namely, consider the ith ray, and choose p a σ = gji ∈ Up . So we have ti (σ) = p, tj (σ) = −p and zero translation elsewhere. Now, except for finitely many points, σ φ preserves the rays and sends xφ to xσφ. Thus there is an infinite subset of the ith ray which is sent to the same ray by φ, say ray k. The infinite subset should be thought of as a union of congruence classes modulo p, except for finitely many points. We claim that no infinite subset of ray j can be mapped by φ to ray k. This is because, if it were, the infinite subset would contain a congruence class modulo p (except for finitely many points) from which we would be able to produce two points, x, y, in the support of σ φ such that all sufficiently large positive powers of σ φ send x into ray k and all sufficiently large negative powers of σ φ send y into ray k, and this is not possible for an element of Hn . This means that φ maps an infinite subset of ray i onto almost all of ray k (observe that ray k is almost contained in the support of σ φ so must be almost contained in the image of the unions of rays i and j). Now applying similar arguments to φ−1 we get that φ is a bijection between rays i and k except for finitely many points. In fact, φ must induce METRIC PROPERTIES OF HOUGHTON’S GROUPS 7 bijections between the congruence classes modulo p (except for finitely many points) inside the two ray systems. Thus φ induces a permutation of the ray system. Again, looking at large positive powers we deduce that tk (σ φ ) > 0 and since the support of σ φ is almost equal to the union of two rays, we must have that tk (σ φ ) = p, as σ and hence σ φ have exactly p orbits. In particular, we may deduce that φ normalises Up . So Upφ = Up . In order to proceed, it will be useful to change the ray system. Specifically, each ray can be split into p rays, preserving the order. This realises Up as a (normal) subgroup of Hnp . We say two of these new rays are equivalent if they came from the same old ray. Thus there are n equivalence classes, each having p elements. The group Up acts on this new ray system as the subgroup of Hnp consisting of all σ ∈ Hnp such that ti (σ) = tj (σ) whenever the ith and j th rays are equivalent. Note that because we have split the rays, these translation amounts can be arbitrary in Up (as a subgroup of Hnp ) and not just multiples of p. In particular, we have that Up = Upφ ⊂ Hnp and the previous arguments imply that φ induces a bijection on the new ray system and sends equivalent rays to equivalent rays (since it is actually permuting the old ray system). Since φ permutes the rays, but must preserve equivalence classes, we get a homomorphism from NSym (Up ) to the subgroup of Snp which preserves the equivalence classes - this is easily seen to be (Sp ≀ Sn ) and the above homomorphism is split. As in Theorem 2.2 we now conclude that the kernel of this homomorphism is Hnp and hence we get that Np = NSym (Up ) is Hnp ⋊(Sp ≀Sn ).  We note that Com(Hn ) is not finitely generated, for if it were, it would lie in some maximal Np . 4. Metric estimates for Hn In this section we will give sharp estimates for the word length of elements of Houghton’s groups. This makes no sense for H1 which is not finitely generated. As mentioned in the introduction, the metric in H2 was described by Lehnert [9]. In order to deal with Hn for n ≥ 3, we introduce the following measure of complexity of an element. Given σ ∈ Hn , we define pi (σ), for i ∈ Zn , to be the largest integer such that (i, pi (σ))σ 6= (i, pi (σ) + ti (σ)). Note that if ti (σ) < 0, then necessarily pi (σ) ≥ |ti (σ)|, as the first element in each ray is numbered 1. 8 BURILLO, CLEARY, MARTINO, AND RÖVER The complexity of σ ∈ Hn is the natural number P (σ), defined by X pi (σ). P (σ) = i∈Zn And the translation amount of σ is 1X T (σ) = |ti (σ)|. 2 i∈Z n The above remark combined with (2) immediately implies P (σ) ≥ T (σ). It is easy to see that an element with complexity zero is trivial, and only the generators gij have complexity one. Theorem 4.1. Let n ≥ 3 and σ ∈ Hn , with complexity P = P (σ) ≥ 2. Then the word length |σ| of σ with respect to any finite generating set satisfies P/C ≤ |σ| ≤ KP log P, where the constants C and K only depend on the choice of generating set. Proof. Since the word length with respect to two different finite generating sets differs only by a multiplicative constant, we can and will choose { gij | i, j ∈ Zn , i 6= j} as generating set to work with, and show that the statement holds with C = 1 and K = 7. The lower bound is established by examining how multiplication by a generator can change the complexity. Suppose σ has complexity P and consider σgij . It is not difficult to see that  p (σ) + 1, if k = i and (i, pi (σ) + 1)σ = (i, 1)    k pk (σ) − 1, if k = j, (j, pj (σ) + 1)σ = (j, 1) and (3) pk (σgij ) = (j, pj (σ))σ = (i, 1)    p (σ), otherwise k where the first two cases are mutually exclusive, as i 6= j. Thus |P (σgij ) − P (σ)| ≤ 1, which establishes the lower bound. The upper bound is obtained as follows. Suppose σ ∈ Hn has complexity P . First we show by induction on T = T (σ) that there is a word ρ of length at most T ≤ P such that the complexity of σρ is P̄ with P̄ ≤ P and T (σρ) = 0. The case T = 0 is trivial. If T > 0, then there are i, j ∈ Zn with ti (σ) > 0 and tj (σ) < 0. So T (σgij ) = T − 1. Moreover, P (σgij ) ≤ P , because the first case of (3) is excluded, as it implies that ti (σ) = −pi (σ) ≤ 0, contrary to our assumption. This completes the induction step. We are now in the situation that σρ ∈ FSym and loosely speaking we proceed as follows. METRIC PROPERTIES OF HOUGHTON’S GROUPS 9 Q (1) We push all irregularities into ray 0, i.e. multiply by gi0 pi(σρ) . (2) We push all points back into the ray to which they belong, except for those from ray 0 which we mix into ray 1, say. (3) We push out of ray 1 separating the points belonging to rays 0 and 1 into ray 0 and any other ray, say ray 2, respectively. (4) We push the points belonging to ray 1 back from ray 2 into it. These four steps can be achieved by multiplying by an element µ of length at most 4P̄ , such that σρµ is an element which, for each i, permutes an initial segment Ii of ray i. Notice that σρµ is now an element of Hn which maps each ray to itself, and hence ti (σρµ) = 0 for all i. It is clear Pthat µ can be chosen so that the total length of the moved intervals, |Ii | ≤ P̄ . Finally, we sort each of these intervals using a recursive procedure, modeled on standard merge sort. In order to sort the interval I = I2 say, we push each of its points out of ray 2 and into either ray 0 if it belongs to the lower half, or to ray 1 if it belongs to the upper half of I. If each of the two halves occurs in the correct order, then we only have to push them back into ray 2 and are done, having used 2|I| generators. If the two halves are not yet sorted, then we use the same “separate the upper and lower halves” approach on each of them recursively in order to sort them. In total this takes at most 2|I| log2 |I| steps. Altogether we have used at most X |Ii | log2 |Ii | ≤ 7P log2 P P + 4P̄ + 2 i∈Zn generators to represent the inverse of σ; we used the hypothesis P ≥ 2 in the last inequality.  We note that because there are many permutations, the fraction of elements which are close to the lower bound goes to zero in much the same way as shown for Thompson’s group V by Birget [1] and its generalization nV by Burillo and Cleary [3]. Lemma 4.2. Let n ≥ 3. For Hn take the generating set g1 , . . . , gn−1 with n − 1 elements. Consider the following sets: • Bk is the ball of radius k, • Ck is the set of elements in Hn which have complexity P = k, • Dk ⊂ Ck is the set of elements of Ck which have word length at most k log2n−2 k Then, we have that |Dk | lim =0 k→∞ |Ck | 10 BURILLO, CLEARY, MARTINO, AND RÖVER An element of complexity P , according to the metric estimates proved above, has word length between P and P log P . What this lemma means is that most elements with complexity P will have word length closer to P log P than to P . Proof. Observe that |Dk | |Bk log2n−2 k | ≤1 because it is a subset. Now, introduce the Ck as |Ck | |Dk | |Ck | |Bk log2n−2 k | and the proof will be complete if we show that lim k→∞ |Ck | |Bk log2n−2 k | = ∞. In Ck there are at least (nk − 2)! elements. This is because we can take a transposition involving a point at distance k down one of the rays, with another point. Since this already ensures P = k, we are free to choose any permutation of the other nk − 2 points which are at one of the first k positions in each ray. And inside |Bk log2n−2 k |, counting grossly according to the number of generators, there are at most (2n − 2)k log2n−2 k = k k elements. Now the limit becomes (nk − 2)! k→∞ kk which is easily seen that it approaches infinity using Stirling’s formula and the fact that n ≥ 3.  lim Consequentially, these estimates give an easy way to see that the group has exponential growth. We note that exponential growth also follows easily from the fact that g01 and g02 generate a free subsemigroup. Proposition 4.3. Let n ≥ 3. Then Hn has exponential growth. Proof. Consider a finitary permutation of complexity P , and observe that there are at least P ! of those. By the metric estimate, its word length is at most KP log P . Using again as in the previous lemma the notation Bk for a ball, we will have that the group has exponential growth if log |Bk | lim > 0. k→∞ k METRIC PROPERTIES OF HOUGHTON’S GROUPS 11 In our case, this amounts to log |BKP log P | log P ! 1 ≥ lim = . P →∞ P →∞ KP log P KP log P K lim  5. Subgroup embeddings We note that each Hn is a subgroup of Hm for n < m and that our estimates together with work of Lehnert are enough to give at least quadratic distortion for some of these embeddings. Theorem 5.1. The group H2 is at least quadratically distorted in Hm for m ≥ 3. Proof. We consider the element σn of H2 which has T (σn ) = 0 and transposes (0, k) and (1, k) for all k ≤ n. Then σn corresponds to the word gn defined in Theorem 8 of [9], where it is shown to have length of the order of n2 with respect to the generators of H2 in Lemma 10 there, which are exactly the generators for H2 given in the introduction. One easily checks that σn = g02 n g12 n g02 −n g12 −n in H3 . Thus a family of words of quadratically growing length in H2 has linearly growing length in H3 , which proves the theorem.  A natural, but seemingly difficult, question is whether Hn is distorted in Hm for 3 ≤ n < m. Another question, which also seems difficult, is to ask whether Hn is distorted in Thompson’s group V , under the embeddings mentioned in the introduction, [10]. 6. Some quasi-isometries of Hn Commensurations give rise to quasi-isometries and are often a rich source of examples of quasi-isometries. Here we show that the natural map from the commensurator of Hn to the quasi-isometry group of Hn , which we denote by QI(Hn ), is an injection. That is, we show that each commensuration is not within a bounded distance of the identity. That this is an injection also follows from the more general argument of Whyte which appears as Proposition 7.5 in Farb-Mosher [6]. Theorem 6.1. The natural homomorphism from Com(Hn ) to QI(Hn ) is an embedding for n ≥ 2. Proof. We will show that for each non-trivial φ ∈ Com(Hn ) and every N ∈ N we can find a σ ∈ Hn such that d(σ, σ φ ) ≥ N, so none of the non-trivial images are within a bounded distance of the identity. 12 BURILLO, CLEARY, MARTINO, AND RÖVER By Theorem 3.2, we can and will view φ as a non-trivial element of Np ⊂ Sym for some even p. If φ eventually translates a ray i non-trivially to a possibly different ray j, then we let σ = ((i, N) (i, N + 1)), a transposition in the translated ray. The image of σ under conjugation by φ is the transposition ((j, N + t), (j, N + t′ + 1)), and the distance d(σ, σ φ ) is the length of σ −1 σ φ , which is at least N since it moves at least one point at distance N down one of the rays. If φ does not eventually translate a ray but eventually non-trivially permutes ray i with another ray j, then we can show boundedness away from the identity by taking σ = ((j, N) (j, N + 1)). The point (i, N) is fixed by σ but is moved to (i, N + 1) under σ φ ensuring that the length of σ −1 σ φ is at least N. Finally, if φ does not have the preceding two properties, then φ is a non-trivial finitary permutation. Since Houghton’s group is ktransitive, for every k, we can find a σ ∈ Hn such that φσ has support disjoint from that of φ, and at distance at least N down one of the rays. Hence σ −1 φ−1 σφ = σ −1 σ φ has length at least N.  7. Co-Hopficity Houghton’s groups are long known to be Hopfian although they are not residually finite, see [4]. In this section we will prove that Hn is not co-Hopfian, by exhibiting a map which is injective but not surjective. The map is the following: f : Hn −→ Hn s 7→ f (s) defined by: if (i, n)s = (j, m), then: (i, 2n − 1)f (s) = (j, 2m − 1) and (i, 2n)f (s) = (j, 2m). It is straightforward to show that f is a homomorphism. It is injective, because if s is not the identity with (i, n)s 6= (i, n), then (i, 2n)f (s) 6= (i, 2n). And clearly the map is not surjective, because the permutation always sends adjacent points (i, 2n − 1), (i, 2n) to adjacent points, and a permutation which does not do this cannot be in the image. In fact, Hn has many proper subgroups isomorphic to the whole group. The following argument was pointed out to us by Peter Kropholler. One can well-order the ray system by taking a lexicographic order. The group Hn is then the group of all almost order preserving bijections of the well-ordered ray system. It is then clear that the ray system METRIC PROPERTIES OF HOUGHTON’S GROUPS 13 minus a point is order isomorphic to the original ray system, which demonstrates that a point stabiliser is a subgroup isomorphic to Hn . Theorem 7.1. Houghton’s groups Hn are not co-Hopfian. References [1] Jean-Camille Birget. The groups of Richard Thompson and complexity. Internat. J. Algebra Comput., 14(5-6):569–626, 2004. International Conference on Semigroups and Groups in honor of the 65th birthday of Prof. John Rhodes. [2] Kenneth S. Brown. Finiteness properties of groups. In Proceedings of the Northwestern conference on cohomology of groups (Evanston, Ill., 1985), volume 44, pages 45–75, 1987. [3] José Burillo and Sean Cleary. Metric properties of higher-dimensional Thompson’s groups. Pacific J. Math., 248(1):49–62, 2010. [4] Yves de Cornulier, Luc Guyot, and Wolfgang Pitsch. On the isolated points in the space of groups. J. Algebra, 307(1):254–277, 2007. [5] John D. Dixon and Brian Mortimer. Permutation groups, volume 163 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1996. [6] B. Farb and L. Mosher. The geometry of surface-by-free groups. Geom. Funct. Anal., 12(5):915–963, 2002. [7] C. H. Houghton. The first cohomology of a group with permutation module coefficients. Arch. Math. (Basel), 31(3):254–258, 1978/79. [8] Sang Rae Lee. Geometry of Houghton’s groups, 2012. arXiv:1212.0257. [9] Jörg Lehnert. Some remarks on depth of dead ends in groups. Internat. J. Algebra Comput., 19(4):585–594, 2009. [10] Claas E. Röver. Subgroups of finitely presented simple groups. PhD thesis, University of Oxford, 1999. [11] W. R. Scott. Group theory. Dover Publications Inc., New York, second edition, 1987. Departament de Matemàtica Aplicada IV, EETAC – UPC, Esteve Terrades 5, 08860 Castelldefels (Barcelona), Spain E-mail address: [email protected] Department of Mathematics R8133, The City College of New York, Convent Ave & 138th, New York, NY 10031, USA E-mail address: [email protected] Mathematics, University of Southampton, Highfield, Southampton, SO17 1BJ, England E-mail address: [email protected] Department of Mathematics, National University of Ireland, Galway, University Road, Galway, Ireland E-mail address: [email protected]
4math.GR
arXiv:1801.02347v2 [math.GR] 3 Apr 2018 GROUPS ACTING ON ROOTED TREES AND THEIR REPRESENTATIONS ON THE BOUNDARY STEFFEN KIONKE Abstract. We consider groups that act on spherically symmetric rooted trees and study the associated representation of the group on the space of locally constant functions on the boundary of the tree. We introduce and discuss the new notion of locally 2-transitive actions. Assuming local 2-transitivity our main theorem yields a precise decomposition of the boundary representation into irreducible constituents. The method can be used to study Gelfand pairs and enables us to answer a question of Grigorchuk. To provide examples, we analyse in detail the local 2-transitivity of GGS-groups. Moreover, our results can be used to determine explicit formulae for zeta functions of induced representations defined by Klopsch and the author. 1. Introduction The objects of interest in this article are groups which act on spherically symmetric rooted trees. This means, they act on a tree T with a distinguished vertex ε, called the root, such that every two vertices having same distance from the root have the same finite number of neighbours. The set Ln of vertices which have distance n from the root is called the n-th level of the tree. Whenever a group G acts on T by root-preserving automorphisms, then G also acts on the boundary ∂T of T . The boundary is the set of all infinite rays starting at the root and it is equipped in a natural way with a profinite topology. The action of G on ∂T gives rise to representations of G on spaces of functions on the boundary. These representations should contain information about the action of G on T and, conversely, it would be desirable to understand such representations via the geometric properties of the action of G on T . An interesting and natural example of such a representation is the unitary representation ρ(2) of G on L2 (∂T , µ), where µ is the Aut(T )-invariant probability measure µ on the boundary. This representation decomposes as a Hilbert space direct sum of infinitely many finite dimensional irreducible representations; c.f. [3]. By contrast, for certain weakly branch groups the associated Koopman representation of G on L2 (∂T , ν) is irreducible for most Bernoulli measures ν 6= µ; see [16]. In a similar vein, the representation of a weakly branch group G on ℓ2 (G/ StG (ξ)) for a point ξ ∈ ∂T is irreducible; see [3]. Date: April 5, 2018. 2010 Mathematics Subject Classification. Primary 20E08; Secondary 20E18, 43A65, 22D30. Key words and phrases. Groups acting on rooted trees, GGS-groups, Gelfand pairs, representation zeta functions. The research was supported by DFG grant KL 2162/1-1. 1 2 S. KIONKE In L2 (∂T , µ), however, the locally constant complex-valued functions C ∞ (∂T , C) form a dense subrepresentation and the structure of the unitary representation ρ(2) is completely determined by the structure of the representation ρ∂,C on C ∞ (∂T , C). We will review this close relationship in Section 2.3 below and we will use it to translate results concerning ρ(2) to the representation ρ∂,C . Our objective is to describe a decomposition of ρ∂,C into irreducible constituents in terms of geometric data. One way to approach this problem is via distance transitive (or 2-point homogeneous) actions on metric spaces; these are isometric actions which are transitive on equidistant pairs of points. It is well-known that a distance transitive action of a finite group H on a finite metric space (X, d) yields a multiplicitiy-free representation of H on C[X]; see [10, Ex. 2.5]. In other words, (H, StH (x)) is a Gelfand pair for every x ∈ X. We refer to [9, 15, 22] for an introduction to Gelfand pairs. Suffice it to mention that finding and understanding Gelfand pairs is an important issue in the representation theory of groups with applications to statistics and probability theory; see for instance [10, 14]. Distance transitive actions of groups on the boundary of finite trees were frequently used to construct Gelfand pairs of finite groups; see [12, 13, 29]. Five examples of Gelfand pairs arising from groups acting on infinite rootes trees (e.g. the Grigorchuk group) are discussed in [4]. In the appendix of [8] Bekka, de la Harpe and Grigorchuk refined the distancetransitivity method and obtained an explicit description of the irreducible constituents of the boundary representation ρ∂,C under the assumption that G acts distance transitively on every level. In this case the boundary representation ρ∂,C contains, apart from the trivial representation, exactly one irreducible summand of dimension |Ln | − |Ln−1 | for all n ∈ N. The account in [8] is based on unitary representations and can be translated via the remarks in Section 2.3 to apply to ρ∂,C . A direct argument is contained in [25, Thm. 6.3]. As a consequence, the representation is multiplicity-free and gives rise to a Gelfand pair (G, Pξ ) of profinite groups where G is the closure of G in the automorphism group G = Aut(T ) and Pξ is a parabolic subgroup, i.e. the stabiliser in G of a point ξ ∈ ∂T . Here we generalise the result of Bekka, de la Harpe and Grigorchuk by assuming only a weak version of distance transitivity and by working over a general field F of suitable characteristic. In fact, the representation on locally constant functions can be defined over an arbitrary base field F . Indeed, the space C ∞ (∂T , F ) of locally constant F -valued functions on ∂T carries a representation of G by imposing ρ∂,F (g)(f )(ξ) = f (g −1 ξ) for all g ∈ G, f ∈ C ∞ (∂T , F ) and ξ ∈ ∂T . The nature of the representation depends strongly on the characteristic of the underlying field and on whether it divides the Steinitz order of G. Here we say that the characteristic of F divides the Steinitz order of G, if it divides the order of some finite continuous quotient of G. One extremal case, where T is the p-regular tree, F is of characteristic p and G is contained in a Sylow pro-p subgroup of Aut(T ) is studied in [21]. Here, however, we will consider the opposite case and assume that the characteristic of F does not divide the Steinitz order of G. In this case, the representation ρ∂,F is semi-simple, i.e., decomposes as a direct sum of finite dimensional irreducible representations. A spherically transitive action of G on T will be called locally 2-transitive, if for all distinct vertices u, v ∈ Ln the common stabiliser StG (u)∩StG (v) acts transitively REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 3 on D(u) × D(v). Here D(u) denotes the set of descendants of u, that is, the neighbours of u which lie in the next level Ln+1 . In fact, every distance transitive action is locally 2-transitive; see Lemma 4. For a vertex v let Irr0F (v) denote the set of non-trivial irreducible F -representations π of StG (v) which occur with non-zero multiplicity m(π, θv,F ) in the permutation representation θv,F of StG (v) on F [D(v)]. Our main result provides a precise description of the irreducible constitutens of ρ∂,F . Theorem 1. Let G ≤ G act locally 2-transitively on T . Assume that F is a field whose characteristic does not divide the Steinitz order of G. 0 The induced representation IndG StG (v) (π) is irreducible for every π ∈ IrrF (v). The decomposition of ρ∂,F into distinct irreducible constituents is given by M M ρ∂,F ∼ m(π, θvn ,F ) IndG = 11F ⊕ StG (vn ) (π), n∈N0 π∈Irr0F (vn ) where vn denotes an arbitrary vertex of level n and 11F is the trivial representation on F . In particular, the representation ρ∂,F is multiplicity-free if and only if θv,F is multiplicity-free for every vertex v. For F = C this implies the following localglobal principle. Theorem 2 (Local-global principle). Let G ≤c G be a closed subgroup which acts locally 2-transitively on T . Let Pξ ≤ G be a parabolic subgroup. Then (G, Pξ ) is a Gelfand pair if and only if (StG (v), StG (u)) is a Gelfand pair for every vertex v and descendant u ∈ D(v). On one hand, the local-global principle can be used to construct Gelfand pairs of profinite groups. Indeed, if the number of descendants of every vertex is a prime or the square of a prime, then the local pairs (StG (v), StG (u)) are Gelfand pairs; for details we refer to Corollary 1. On the other hand, we use the local-global principle to construct branch groups G ≤c G such that (G, Pξ ) is not a Gelfand pair for every parabolic subgroup Pξ of G; see Example 3. This answers a question of Grigorchuk; see Problem 10.1 in [20]. In Section 4 we study the local 2-transitivity of GGS-groups, named after the famous finitely generated, infinite torsion groups of Grigorchuk [19] and Gupta and Sidki [23]. These groups attracted considerable interest recently and the reader is refered to the monograph [5, Sec. 2.3] and the articles [17, 18, 32] for more information. Here we consider (generalised) GGS-groups which act on a pk -regular rooted tree. In this case every vertex in the tree T has exactly pk descendants for a prime number p. Theorem 4 characterises the defining vectors with locally 2-transitive associated GGS-groups. It follows from Remark 3 that our criterion is always satisfied if p is an odd prime and k = 1. The following result is a consequence of Theorem 4. Theorem 3. Let p be an odd prime number and let T be the p-regular rooted tree. Every GGS-group Ge ≤ Aut(T ) acts locally 2-transitively. Moreover, (Ge , Pξ ) is a Gelfand pair for every parabolic subgroup Pξ ≤c Ge . This result confirms a conjecture mentioned by Bartholdi and Grigorchuk in the very last sentence of [4]. We remark that the unique GGS-group which acts on the rooted binary tree does not act locally 2-transitively (see Remark 4). However it is still true that (Ge , Pξ ) is a Gelfand pair. 4 S. KIONKE 1.1. Applications to zeta functions of representations. Let G be a profinite group and consider, for every n, the number rn (G) of irreducible n-dimensional complex continuous representations of G. We assume that G is representation rigid, this means, that the numbers rn (G) are all finite. The paradigm of representation growth postulates that the asymptotic behaviour of the sequence (rn (G))n∈N contains information on the structure of the group G. In recent years this line of investigation stimulated a considerable amount of new results; see e.g. [1, 2, 24, 30, 31] and the survey [26] as well as references therein. Here once again groups which act on regular rooted trees provide a class of examples with remarkable properties – for self-similar branched groups this was discussed by Bartholdi in [7]. His results apply, in particular, to the Gupta-Sidki p-groups which will also be considered in Section 4 below. The special case of iterated wreath products was studied already in [6]. In many examples the numbers rn (G) grow at most polynomially in n and it is natural to investigate the asymptotic properties using the Dirichlet series ζG (s) = ∞ X rn (G) , ns n=1 which is called the representation zeta function. For certain self-similar branch groups Bartholdi’s method [7] allows the efficient computation of the coefficients rn (G) using a functional equation of ζG . However, it is worth noting, that in general explicit formulae for ζG are quite difficult to obtain; see [1, 2] for an extensive treatment of compact p-adic Lie groups of type A2 (such as SL3 (Zp )). Facing this difficulty it might be fruitful to break the problem into smaller pieces by counting only certain subsets of all representations. In the recent preprint [25] B. Klopsch and the author developed such a more flexible approach to representation zeta functions which is based on the following observation. The representation zeta function at s − 1 is ζG (s − 1) = ∞ X rn (G)n = ns n=1 X π∈IrrC (G) dim(Vπ ) dim(Vπ )s and the dimension dimC (Vπ ) of an irreducible complex representation (π, Vπ ) of G is equal to the multiplicity of π in the regular representation of G on the space of locally constant functions C ∞ (G, C). Equivalently, it is the multiplicity in the regular representation of G on L2 (G) with respect to the Haar measure; see Section 2.3. Let (ρ, Vρ ) be an admissible smooth representation of G; see Section 2.2 for the definition of smooth representations. Such a representation decomposes into a direct sum M m(π, ρ)Vπ Vρ = π∈IrrC (G) over the finite dimensional irreducible complex representations of G which occur with finite multiplicities m(π, ρ) ∈ N0 . Under suitable assumptions it is natural to investigate the Dirichlet series X m(π, ρ) , ζρ (s) = dimC (Vπ )s π∈Irr(G) REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 5 which will be called the zeta function of the representation (ρ, Vρ ); for the details we refer to [25]. As observed above, the zeta function of the regular representation equals ζG (s − 1). Here the results of [25] will not be used. It is sufficient to know that this setting provides a wealth of examples, where explicit formulae for zeta functions of representations can actually be determined; see [25, Sect. 7]. Theorem 1 provides a new way to obtain explicit formulae for zeta functions of representations. We return to the situation where G ≤c G is a closed locally 2transitive subgroup of the automorphism group of a spherically symmetric rooted tree T . Theorem 1 yields a formula for the zeta function of the representation ρ∂,C on the boundary X X m(π, θvn ,C ) (1) ζρ∂,C (s) = 1 + . s s |L n | dim(Vπ ) 0 n∈N0 π∈IrrC (vn ) Since ∂T ∼ = G/P for a parabolic subgroup P , the representation ρ∂,C is isomorphic to the smoothly induced representation IndG P (11C ); see Section 2.2 for the definition of smooth induction. In certain cases, this makes it possible to use formula (1) to determine the zeta function of a smoothly induced representation. Let G be any profinite group with a closed subgroup P ≤c G. If one can realise the homogeneous space G/P as the boundary of a spherically symmetric rooted tree T with a locally 2-transitive action of G, then formula (1) describes the zeta function of IndG P (11C ). We give an example to illustrate the usefulness of this approach in Section 3.4. 2. Preliminaries from representation theory Here we review some preliminary results from representation theory. Let F be a field. Recall that for an abstract group H, a subgroup S ≤ H and an F representation (π, Vπ ) of S, the representation induced from π is the representation of H on the space IndH S (π) = {f : H → Vπ | f (sx) = π(s)f (x) for all s ∈ S}. by the right regular action, i.e. ρ(h)(f )(x) := f (xh) for all f ∈ IndH S (π) and all x, h ∈ H. We remark that sometimes this is also called the coinduced representation. 2.1. Representation theory of finite groups. In this short subsection H denotes a finite group and F denotes a field whose characteristic does not divide the order of H. Under these assumptions Maschke’s theorem implies that the group algebra F [H] is a semisimple F -algebra (see e.g. [28, XVIII Thm. 1.2]). In particular, every representation of H on an F -vector space decomposes as a direct sum of finite dimensional irreducible representations. Moreover, we recall that every action of H on a finite set A gives rise to a representation ρA,F of H on the vector space nX o F [A] = ca a | ca ∈ F a∈A of formal F -linear combinations of the elements in A by imposing ρA,F (g)(a) = ga. We collect two preliminary results from the representation theory of finite groups. 6 S. KIONKE Lemma 1. Assume that H acts on finite sets A and B. If the diagonal action of H on A × B is transitive, then dimF HomF [H] (F [A], F [B]) = 1. In particular, the trivial representation is the only irreducible F -representation of H which occurs in both F [A] and F [B]. Proof. The space of F -linear maps HomF (F [A], F [B]) is equipped with an H-action via (h·ϕ)(a) := hϕ(h−1 a) for all a ∈ A. We note that the H-invariants in this space are exactly the H-equivariant linear maps. Since A is finite, there is an isomorphism of F [H]-modules HomF (F [A], F [B]) ∼ = F [A × B] We conclude that HomF [H] (F [A], F [B]) ∼ = F [A × B]H = HomF (F [A], F [B])H ∼ where F [A × B]H denotes the space of H-invariants. Since H acts transitively on A × B, the space of H-invariants in F [A × B] is one-dimensional.  Lemma 2. Let S ≤ H be a subgroup. Given an irreducible F -representation (π, Vπ ) of S so that dimF EndF [S] (π) = dimF EndF [H] (IndH S (π)), then the induced representation IndH S (π) is irreducible. Proof. Since π is irreducible, the endomorphism algebra E = EndF [S] (π) is a division F -algebra. The induction functor provides a homomorphism of F -algebras J : E → EndF [H] (IndH S (π)). Since E is a division algebra the homomorphism J is injective. The assumption on the dimensions implies that J is actually an isomorphism. In particular, EndF [H] (IndH S (π)) is a division algebra and, hence, does not contain non-trivial idempotent elements. By Maschke’s theorem the representation EndF [H] (IndH S (π)) is semi-simple, i.e. every subrepresentation is a direct summand. Therefore the non-existence of idempotent elements implies that IndH S (π) is irreducible.  2.2. Representation theory of profinite groups. Let H denote a profinite group and let F be a field. A representation (ρ, Wρ ) of H on some (possibly infinite dimensional) F -vector space Wρ is called smooth if every w ∈ Wρ has an open stabilizer in H. This means, that Wρ is the union of finite dimensional subrepresentations on which the action of H factors through some finite continuous quotient of H. In particular, a finite dimensional representation is smooth exactly if it factors over some finite continuous quotient of H. We will impose a condition on the characteristic of F , which ensures that smooth representations decompose into direct sums of finite dimensional irreducible representations. Definition 1. We say that the characteristic of F does not divide the Steinitz order of H, if char(F ) does not divide the order of any finite continuous quotient of H. This definition alludes to the Steinitz order of a profinite group in terms of Steinitz numbers. We refer to [34, Chapter 2] for the definition of the Steinitz order of a profinite group. Here this connection will not be important. Assume now that char(F ) does not divide the Steinitz order of H. The key observation is that, since smooth representations of H are unions of representations REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 7 of finite quotients, Maschke’s theorem implies that every smooth representation (ρ, Wρ ) of H over F is a direct sum of isotypic components, i.e. M Wρ (π) Wρ = π∈IrrF (H) where the sum runs over all smooth irreducible F -representations of H. The isotypic component Wρ (π) isomorphic to a direct sum of copies of π: the number m(π, ρ) of copies is called the multiplicity of π in ρ. The representation ρ is said to be admissible, if all multiplicities are finite; in other words, all isotypic components are finite dimensional. For convenience we recall the concept of smooth induction. Let S ≤c H be a closed subgroup and let (π, Vπ ) be a smooth representation of S. The representation smoothly induced from π is a representation of H on the space IndH S (π) = {f : H → Vπ | f loc. constant and f (sx) = π(s)f (x) for all s ∈ S}. The representation of H is given by the right regular action, i.e. ρ(h)(f )(x) := f (xh) for all f ∈ IndH S (π) and all x, h ∈ H. In particular, if π is the trivial representation on a one dimensional F -vector space, then IndH S (π) is the space C ∞ (H/S, F ) of locally constant F -valued functions on the homogeneous space H/S. The representation of H on C ∞ (H/S, F ) is always admissible. 2.3. Unitary versus smooth representations. Let G be a profinite group and let G ≤ G be a dense subgroup. We consider the homogeneous space X = G/P for a closed subgroup P ≤c G. The pushforward measure µ of the normalized Haar measure on G is a regular G-invariant Borel probability measure on X. We consider the representation of G on the complex Hilbert space L2 (X, µ) defined by (ρ(g)h)(x) = h(g −1 x) for all h ∈ L2 (X, µ) and g ∈ G. In this situation, the properties of the representation ρ|G of G on L2 (X, µ) are essentially determined by the representation of G (or G) on the space C ∞ (X, C) of locally constant C-valued functions on X. This relation seems to be well-known, however, we could not find a good reference. We include a short discussion, since this method can be used to interpret our results in the framework of unitary representations. Proposition 1. In the situation above the following statements hold. (1) The subspace C ∞ (X, C) is dense in L2 (X, µ). (2) Every closed irreducible G-subrepresentation of L2 (X, µ) is G-stable and contained in C ∞ (X, L C). (3) Let C ∞ (X, C) = ϑ∈Irr(G) V (ϑ) be the decomposition of the G-representation on C ∞ (X, C) into isotypic components. The isotypic components V (ϑ) are finite dimensional and pairwise orthogonal in L2 (X, µ). In particular, L L2 (X, µ) is isomorphic to the Hilbert space direct sum c ϑ∈Irr(G) V (ϑ). Proof. First we consider (1). Every f ∈ C ∞ (G, R C) provides a convolution operator Tf ∈ B(L2 (X, µ)) given by Tf (h)(x) = G f (g)h(gx)dg with respect to the normalized Haar measure on G. Since f is locally constant, the function Tf (h) is locally constant for every h ∈ L2 (X, µ). Let ε > 0 and h ∈ L2 (X, µ) be given. The multiplication map G × L2 (X, µ) → L2 (X, µ) is continuous, hence there is an 8 S. KIONKE open normal subgroup K Eo G such that kh − ρ(k)hk2 < ε for all k ∈ K. Let f ∈ C ∞ (G, R) be a positive function supported in K with total mass one, then Z f (k)kh − ρ(k −1 )hk2 dk < ε kTf (h) − hk2 ≤ G and we conclude that C ∞ (X, C) is dense in L2 (X, µ). Let V ⊆ L2 (X, µ) be a closed irreducible G-subrepresentation. It follows from the continuity of the map G× L2 (X, µ) → L2 (X, µ), that V is also G-stable. By the Peter-Weyl theorem every irreducible unitary representation of the compact group G is finite dimensional; see [27, Corollary 5.4.2]. Moreover, G is profinite and the no-small-subgroup argument shows that the representation of G on V factors over ∞ some finite continuous quotient. L We deduce that V lies in C (X, C). ∞ Finally, let C (X, C) = ϑ∈Irr(G) V (ϑ) be the decomposition into isotypic components as in (3). Since the representation is admissible, the isotypic components are finite dimensional. For the fact that distinct isotypic components are orthogonal we refer to [27, Lemma 3.4.21].  3. Locally 2-transitive actions on rooted trees 3.1. Notation. Let (Xi )i∈N be a sequence of finite sets. Throughout, mi = |Xi | denotes the cardinality of Xi and we assume mi ≥ 2 for all i ∈ N. For every n ∈ N0 we define n Y Xi , Ln := i=1 this is the set of finite sequences (x1 , . . . , xn ) with xi ∈ Xi . Here we use the convention that L0 consists of exactly one element: the empty sequence ε. The spherically symmetric rooted tree F∞T = (V (T ), E(T )) (based on (Xi )i∈N ) is a graph on the set of vertices V (T ) = n=0 Ln and there is an edge between u ∈ Ln and v ∈ Ln+1 exactly if u is a prefix of v. The distinguished element ε ∈ L0 is called the root of T . In addition, if mi = m for all i, then we say that T is the m-regular rooted tree. The set Ln is called the n-th level of T . For a vertex u ∈ Ln the set of descendants D(u) consists of all vertices in Ln+1 which are incident to u. If u = (x1 , . . . , xn ) and y ∈ Xn+1 , then we write uy for the descendant (x1 , . . . , xn , y) of u. The group of root-preserving automorphisms G = Aut(T ) of T is a profinite group. A neighbourhood base of open normal subgroups is given by the pointwise stabilisers G(n) of the n-th level, i.e., G(n) = {g ∈ Aut(T ) | ∀u ∈ Ln gu = u}. A subgroup G ≤ G is said to act spherically transitive on T , if G acts transitively on the level Ln for every n ∈ N. Clearly, every root-preserving automorphism also preserves the levels setwise. For g ∈ G and a vertex u ∈ Ln , we define the label of g at u to be the unique permutation g(u) ∈ Aut(Xn+1 ) such that g(ux) = g(u)g(u) (x) for all x ∈ Xn+1 . The boundary ∂T of T is the profinite set Y ∂T := Xi , i∈N which can be considered as the set of infinite, non-backtracking paths in T starting at the root. The automorphism group G acts continuously on ∂T . We observe that REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 9 a subgroup G ≤ G acts spherically transitively on T if and only if its closure G ≤c G acts transitively on ∂T . The stabiliser of u ∈ Ln in G will be denoted by StG (u). The stabiliser Pξ of a point ξ ∈ ∂T in G is called a parabolic subgroup of G. All parabolic subgroups are conjugated in G, whenever the action of G is spherically transitive. 3.2. Local 2-transitivity: Definition and examples. Definition 2. A subgroup G ≤ G acts locally 2-transitively on T , if G acts spherically transitive and for all n ∈ N and every pair u, v ∈ Ln of distinct vertices, the group StG (u) ∩ StG (v) acts transitively on D(u) × D(v) via the diagonal action. Remark 1. It is worth noting, that a group G acts locally 2-transitively if and only if its closure G does. We give two examples of actions that are locally 2-transitive. First, we discuss iterated wreath products. For every i ∈ N, let Gi ⊆ Aut(Xi ) be a transitive subgroup. We denote the collection (Gi | i ∈ N) of these subgroups by C. The iterated wreath product defined by C is the closed subgroup GC of G which consists of all elements which have only labels in C, i.e. GC = {g ∈ G | ∀n ∈ N0 ∀u ∈ Ln g(u) ∈ Gn+1 }. Lemma 3. Let C be a collection of transitive subgroups. The action of the iterated wreath product GC on T is locally 2-transitive. Proof. It follows immediately from an inductive argument and the fact that every Gi ⊆ Aut(Xi ) is transitive, that the action of GC is spherically transitive. Now, let u, v ∈ Ln be distinct elements of the n-th level. The group StG (u) ∩ StG (v) acts on D(u) × D(v) like Gn+1 × Gn+1 , thus the action is transitive.  Second, we argue that the notion of locally 2-transitive actions weakens the notion of distance transitive actions. The shortest-path distance on T induces a metric d on every level Ln . A group G ≤ G is said to act distance transitively on T , if for all n ∈ N and all u1 , u2 , v1 , v2 ∈ Ln with d(u1 , u2 ) = d(v1 , v2 ), there is some g ∈ G so that gui = vi for all i ∈ {1, 2}. Lemma 4. Let G ≤ G. If G acts distance transitively on T , then the action of G is locally 2-transitive. Proof. Clearly, a distance transitive action is spherically transitive. Let u, v ∈ Ln be distinct vertices. Every pair of descendants u′ ∈ D(u) and v ′ ∈ D(v) satisfies d(u′ , v ′ ) = d(u, v) + 2. Therefore, if G acts distance transitively on T , then StG (u) ∩ StG (v) acts transitively on D(u) × D(v).  Example 1. Here we construct, using iterated wreath products, groups which act locally 2-transitively, but not distance transitively. Let H be a finite group with |H| > 2. We define Xi = H for all i. The associated tree T is the rooted |H|-regular tree. Let Gi = H ⊆ Aut(Xi ), where H acts on itself by multiplication from the left. The collection C = (H|i ∈ N) defines the iterated wreath product GC = . . . H ≀ H ≀ H group. By Lemma 3 the group GC acts locally 2-transitively on T . We verify that GC does not act distance transitively. Since |H| ≥ 3 we find pairwise distinct elements x, y, z ∈ H. We obtain three vertices in L2 : u1 = v1 = 10 S. KIONKE (x, x), u2 = (y, x) and v2 = (z, x) with d(u1 , u2 ) = 4 = d(v1 , v2 ). Every element h ∈ H which satisfies hx = x is already the identity element. Therefore, an element g ∈ GC with gu1 = u1 has label g(ε) = Id and acts trivially on the first level. In particular, gu2 6= v2 and we deduce that GC does not act distance transitively. Example 2. We discuss an example of a group which acts spherically transitively but not locally 2-transitively. Let T be the rooted binary tree, i.e. Xi = {0, 1} for all i. We consider the group G which is generated by two elements a, b ∈ G. The automorphism a simply swaps the two elements of the first level, i.e. a(u1 x) = (1 − u1 )x. for all u1 ∈ {0, 1} and x ∈ Ln . The automorphism b is defined recursively by the rule ( (0)a(x) if u1 = 0 b(u1 x) = (1)b(x) if u1 = 1 for u1 ∈ {0, 1} and x ∈ Ln . The group G is infinite and generated by two elements of order two, which implies that G is isomorphic to the infinite dihedral group D∞ . Clearly, G acts transitively on the first level L1 . Using induction it follows that G acts indeed spherically transitive. We will see that the action of G on T is not locally 2-transitive. Consider the vertices on the second level and their descendants. 1 5 3 7 4 8 2 6 Figure 1. A numbering of the vertices on the third level. If we number the vertices on the third level as in Figure 1, then the generators a and b act on the third level like reflections of the regular octagon as indicated in Figure 2. In particular, the action factors through the dihedral group D8 . The 1 2 a 8 3 7 4 6 5 b Figure 2. The action on the third level. vertices of the second level correspond to pairs of opposite vertices in the octagon. Now, if we choose two neighbouring pairs of opposite vertices, say {1, 5} and {2, 6}, then the intersection of the two stabilisers consists merely of the identity and the point reflection in the center. Clearly, this group with two elements cannot act transitively on the 4-element set {1, 5} × {2, 6}. REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 11 Later the following characterisation of locally 2-transitive actions will be useful. Lemma 5. Let G ≤ G be a spherically transitive group. Choose a vertex un ∈ Ln for every n ∈ N0 . The group G acts locally 2-transitively on T if and only if for every n ∈ N and every v ∈ Ln with un+1 ∈ / D(v) the group StG (un+1 ) ∩ StG (v) acts transitively on D(v). Proof. Assume first that G acts locally 2-transitively and let v ∈ Ln (n ≥ 1) be given such that un+1 ∈ / D(v). Let u ∈ Ln be the predecessor of un+1 and observe that u 6= v. By assumption the group StG (u) ∩ StG (v) acts transitively on D(u) × D(v), thus StG (un+1 ) ∩ StG (v) acts transitively on D(v). Conversely, let x, y ∈ Ln with x 6= y be given. Let (a, b), (c, d) ∈ D(x) × D(y). Since the action of G is spherically transitive, there is g ∈ G such that ga = un+1 . By assumption there is an element h ∈ StG (un+1 ) ∩ StG (gy) such that hgb = gd. Thus γ1 = g −1 hg ∈ StG (a) ∩ StG (y) satisfies γ1 b = d. By the same argument, we find γ2 ∈ StG (d) ∩ StG (x) such that γ2 a = c. Now γ2 γ1 ∈ StG (x) ∩ StG (y) satisfies γ2 γ1 a = c and γ2 γ1 b = d. We deduce that the action of StG (x) ∩ StG (y) on D(x) × D(y) is transitive.  3.3. The representation on the boundary. Let G ≤ G. We fix a field F whose characteristic does not divide the Steinitz order of G. In the case where G acts spherically transitive, this assumption implies that char(F ) ∤ mn for every n ∈ N. For every n ∈ N the group G (and G) acts on the level Ln and we obtain, as above, an associated representation ρn,F of G on the space nX o F [Ln ] = cu u | cu ∈ F . u∈Ln Similarly, for every vertex u ∈ Ln the stabiliser StG (u) acts on the descendants D(u) and we obtain an associated representation θu,F of StG (u) on F [D(u)]. As G acts on the boundary ∂T , we obtain a natural representation ρ∂,F of G on the space C ∞ (∂T , F ) of F -valued locally constant functions on ∂T . More precisely, we define ρ∂,F (g)(f )(ξ) := f (g −1 ξ) for all g ∈ G, f ∈ C ∞ (∂T , F ) and ξ ∈ ∂T . In fact, the boundary representation can be described in various ways. First, it is the direct limit of the level representations ρn,F , i.e., ρ∂,F = lim ρn,F . −→ n∈N Moreover, if G is spherically transitive, then ∂T ∼ = G/P for any parabolic subgroup P of G. It follows from the definition of smooth induction (see Section 2.2) that ρ∂,F = IndG P (11F ). 0 Let v ∈ V (T ) and let (θv,F , F [D(v)]0 ) denote the complement of the onedimensional trivial subrepresentation inside (θv,F , F [D(v)]). As a vector space o n X X cw = 0 . cw w | F [D(v)]0 = w∈D(v) w∈D(v) Let Irr0F (v) denote the set of irreducible F -representations of StG (v) which occur 0 with non-zero multiplicity in θv,F . 12 S. KIONKE Lemma 6. Assume that G acts locally 2-transitively on T . Let π ∈ Irr0F (v) and σ ∈ Irr0F (w) for vertices v, w ∈ V (T ). Then the following statements hold. (1) The induced representation Iv (π) = IndG StG (v) (π) is irreducible. (2) If v ∈ Ln , then dimF Iv (π) = |Ln | dimF π. (3) Iw (σ) ∼ = Iv (π) if and only if v = tw and π ∼ = t σ for some t ∈ G. Here t σ is the representation of t StG (w)t−1 given by t σ(g) = σ(t−1 gt). Proof. Since G acts spherically transitively, we have |G : StG (v)| = |Ln | and assertion (2) follows. For t ∈ G we write Ht = StG (v) ∩ t StG (w). We apply Mackey’s intertwining number theorem (see (10.23) in [11]): dimF HomF [G] (Iv (π), Iw (σ)) = X dimF HomF [Ht ] (π|Ht , t σ|Ht ) t∈R where the sum runs over a set R ⊆ G of representatives for the double cosets StG (v)\G/ StG (w). Assume first that v and w lie in the same level of T . Fix t ∈ R and define u = tw and H = Ht . Clearly, if u = v, then H = StG (v) and dimF HomF [H] (π|H , t σ|H ) is non-zero exactly, when t σ ∼ = π. In this case the dimension is the dimension of the endomorphism F -algebra EndF [H] (π). Assume now that u 6= v. The action of G is locally 2-transitive, this means that the group H = StG (v)∩StG (u) acts transitively on D(v)×D(u). The representation π|H is a subrepresentation of (θv,F |H , F [D(v)]) and t σ|H is a subrepresentation of (θu,F |H , F [D(u)]). By Lemma 1 the trivial representation is the unique irreducible representation of H which occurs in both F [D(v)] and in F [D(u)]. However, by assumption π is non-trivial and since H acts transitively on D(v), we deduce that π|H does not admit a non-zero H-fixed vector. We deduce that π|H and t σ|H do not share an irreducible constituent; this implies dimF HomF [H] (π|H , t σ|H ) = 0. In particular, this yields dimF EndF [G] (Iv (π))) = dimF EndStG (v) (π) and we deduce (1) from Lemma 2. Now the computation above implies (3) provided v and w lie in the same level. Finally, assume that v and w lie in different levels, say v ∈ Ln and w ∈ Lm with n < m. We observe that dimF Iv (π) = |Ln | dimF π ≤ |Ln |(mn+1 − 1) < |Ln+1 | ≤ dimF Iw (σ), and thus the representations are not isomorphic.  Proof of Theorem 1 and Theorem 2. Recall that the representation ρ∂,F on the locally constant functions on the boundary ∂T is the direct limit of the representations ρn,F of G on F [Ln ]. We will proceed by induction on n to show that (2) ρn,F ∼ = 11F ⊕ n−1 M M m(π, θvj ,F ) Ivj (π). j=0 π∈Irr0F (vj ) Clearly, ρ0,F = 11F is the trivial representation. Assume that formula (2) holds and consider ρn+1,F . Let u ∈ D(vn ) be a descendant of vn . Then the following short REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 13 calculation completes the induction step: StG (vn ) G ∼ ρn+1,F ∼ = IndG StG (u) (11F ) = IndStG (vn ) IndStG (u) (11F ) G G 0 ∼ ∼ = IndG StG (vn ) (θvn ,F ) = IndStG (vn ) (11F ) ⊕ IndStG (vn ) (θvn ,F ) M ∼ m(π, θvn ,F )Ivn (π). = ρn,F ⊕ π∈Irr0F (vn ) By Lemma 6 the irreducible representations Ivn (π) are pairwise distinct, therefore the multiplicities of irreducible representations in ρ∂,F are exactly the multiplicities m(π, θvn ,F ). We deduce that ρ∂,F is multiplicity-free if θv,F is multiplicity-free for every v ∈ V (T ). The converse holds as well since vn ∈ Ln was arbitrary.  Now we specialise to the case F = C. The following result provides a simple method to construct Gelfand pairs of profinite groups. Corollary 1. Assume that for every i ∈ N the number mi is either a prime or the square of a prime. Let G ≤c G be a closed subgroup which acts locally 2-transitively on T . Then (G, Pξ ) is a Gelfand pair for every parabolic subgroup Pξ ⊆ G. Proof. By the local-global principle, we need to verify that (StG (v), StG (u)) is a Gelfand pair for every vertex v and descendant u ∈ D(v). Let H denote the image of StG (v) in the symmetric group Sym(D(v)). We observe that it suffices to verify that H contains an abelian transitive subgroup A. Indeed, the representation of A on C[D(v)] is multiplicity-free and thus the representation of H on C[D(v)] cannot contain an irreducible constituent of multiplicity exceeding 1. Case 1 : |D(v)| = p is a prime number. Then H contains an element g of order p, i.e. a p-cycle. We set A = hgi. Case 2 : |D(v)| = p2 is the square of a prime number. Let K ≤ H be a Sylow p-subgroup, say |K| = pr . We note that K acts transitively on D(v). Indeed, suppose that |K/K ∩ StH (u)| ≤ p, then pr−1 divides | StH (u)| and so pr+1 divides | StH (u)|p2 = |H|. This contradicts the fact that K is a Sylow p-subgroup. Let g ∈ Z(K) be a non-trivial element in the center of K. If hgi acts transitively, we can choose A = hgi. Otherwise, we find and element w ∈ D(v) such that O = {g k w | k ∈ Z} has exactly p elements. The action of K on D(v) is transitive, so there is some h ∈ K such that hw ∈ / O. We conclude that the group A = hg, hi is abelian and acts transitively on D(v).  Example 3. We return to the setting of Example 1 where GC = . . . H ≀ H ≀ H is the iterated wreath product of some finite group H with |H| ≥ 3. Note that the group GC is a branch group; see [5, §1.3]. If the group H is non-abelian, then (GC , Pξ ) is not a Gelfand pair for every ξ ∈ ∂T . For simplicity, we discuss only for the symmetric group H = S3 . In this case Xi = S3 is the symmetric group on 3-letters and let Gi = S3 ⊆ Aut(Xi ), where S3 simply acts on itself by multiplication from the left. Therefore T is the 6-regular rooted tree. For every vertex u, the action of StG (u) on D(u) factors through the regular representation of S3 . The regular representation of S3 on C[S3 ] decomposes into two one-dimensional and the sign characrepresentations, namely the trivial representation 11 = ter , and a two-dimensional irreducible representation τ = which occurs twice. By Theorem 1 the irreducible representations Iv (τ ) occur with multiplicity two in 14 S. KIONKE C ρ∂,C = IndG Pξ (11) for every vertex v where Pξ denotes a parabolic subgroup of GC . We conclude that (GC , Pξ ) is not a Gelfand pair. 3.4. An application to zeta functions of representations. The aim of this section is to use Theorem 1 to compute the zeta function of an induced representation. The definition given in the introduction suffices for our purposes. For more information on zeta functions of induced representations the reader should consult [25]. Let R be a compact discrete valuation ring with maximal ideal p ⊆ R and uniformiser π ∈ p. The residue field k = R/p is finite, say of cardinality q. We note that the ring R is the ring of integers of some non-archimedean local field (i.e., Qp , Fp ((t)) or finite extensions thereof). The general linear group GLn (R) is a profinite group. For every r ∈ N, we define the principal congruence subgroup GLrn (R) = ker (GLn (R) → GLn (R/pr )) of level r as the kernel of the reduction-mod-pr homomorphism. Let P denote the parabolic subgroup of GLn (R) which consists of all matrices A = (ai,j ) ∈ GLn (R) with ai,1 = 0 for all i ≥ 2. We put P r = P ∩ GLrn (R). Proposition 2. In the above setting let G = GL1N +1 (R). The group G acts locally 2-transitively on a q N -regular rooted tree T such that P 1 is the stabiliser of a point ξ ∈ ∂T . The pair (G, P 1 ) is Gelfand and the zeta function of the smoothly induced representation ρ = IndG P 1 (11C ) is (3) ζρ (s) = q N 1 − q −(s+1)N . 1 − q −sN Remark 2. We observe that formula (3) only depends on the residue field cardinality q and is, in particular, independent of the characteristic of R. This phenomenon was observed in several different cases and calls for a deeper investigation; see also [25, Problem 1.1]. Moreover, we note that formula (3) was known to hold if R is of characteristic 0 [25, Prop. 7.7]. Proof. Let S be a commutative ring. An element (x1 , . . . , xN +1 ) ∈ S N +1 is called primitive, if the ideal generated by x1 , . . . , xN +1 is S. We denote by PN (S) the S-rational points of the N -dimensional projective space, that is PN (S) = {x ∈ S N +1 | x is primitive }/S ×. The equivalence class of (x1 , . . . , xN +1 ) in PN (S) will be denoted using projective coordinates as (x1 : x2 : . . . : xN +1 ). Consider the tree T whose vertices of level n ∈ N0 are the elements of Ln = {v ∈ PN (R/pn+1 ) | v ≡ (1 : 0 : . . . : 0) mod p}. There is an edge between u ∈ Ln−1 and v ∈ Ln if and only if v ≡ u mod pn . We define un ∈ Ln to be the point un = (1 : 0 : . . . : 0) ∈ PN (R/pn+1 ). The root of T is the vertex u0 . The boundary ∂T can be (and will be) identified with ∂T = {x ∈ PN (R) | x ≡ u0 mod p}. We fix ξ = (1 : 0 : . . . : 0) ∈ ∂T . For every n ∈ N the group GLN +1 (R) acts on PN (R/pn ) and the restrictions of these actions to GL1N +1 (R) provide an action on T . Observe that P 1 is the stabiliser of ξ. REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 15 Step 1: The action of G on T is spherically transitive. Let v ∈ Ln , say v = (1 : v2 + pn+1 : . . . : vN +1 + pn+1 ) with certain vi ∈ p. Then gun = v where   1 0 ··· 0  v2  1   g= .  ∈ GL1N +1 (R). ..  ..  . vN +1 1 Step 2: The action of G on T is locally 2-transitive. Let v ∈ Ln−1 with v 6= un−1 for some n ≥ 2. Since G acts spherically transitive, it is sufficient to show that StG (un ) ∩ StG (v) acts transitively on the descendants D(v); see Lemma 5. Let x, y ∈ D(v) ⊆ Ln . There are 1 ≤ ℓ < n and primitive vectors x̃ = (x2 , . . . , xN +1 ) and ỹ = (y2 , . . . , yN +1 ) ∈ RN such that x = (1 : π ℓ x2 + pn+1 : . . . : π ℓ xN +1 + pn+1 ) and y = (1 : π ℓ y2 + pn+1 : . . . : π ℓ yN +1 + pn+1 ). Indeed, the exponent ℓ is the largest number such that v (as well as x and y) hangs below uℓ−1 . Since the primitive vectors x̃ and ỹ satisfy x̃ ≡ ỹ mod pn−ℓ , there is a matrix n−ℓ A ∈ GLN (R) such that Ax̃ = ỹ. Moreover, n − ℓ ≥ 1 and thus the matrix   1 0 ··· 0 0    g = .   ..  A 0 is an element of StG (un ) ∩ StG (v) such that gx = y. Step 3: The representation of StG (un ) on C[D(un )] decomposes into pairwise distinct 1-dimensional irreducible representations. Let n ∈ N0 . We consider the bijective map D(un ) → k N which maps an element (1 : π n+1 x2 : . . . : π n+1 xN +1 ) to (x2 + p, . . . , xN +1 + p) ∈ k N . A short calculation shows that via this identification the stabiliser StG (un ) acts like the additive group k N on itself by translations. More precisely, StG (un+1 ) is normal in StG (un ) and the quotient is isomorphic to k N . In particular, C[D(un )] decomposes like the regular representation of the abelian group k N . Finally, we deduce from Theorem 1 that (G, P 1 ) is a Gelfand pair. Moreover, the representation on the boundary ρ∂,C is isomorphic to the induced representation IndG P 1 (11C ). It decomposes into a direct sum of the trivial representation and, for every n ∈ N0 , exactly q N − 1 irreducible representations of dimension q nN . Using the geometric series we obtain ∞ X 1 − q −(s+1)N ζρ (s) = 1 + (q N − 1)q −snN = q N .  1 − q −sN n=0 4. Local 2-transitivity of GGS-groups In this section we look at a famous family of examples: the GGS-groups. For more information about these groups the reader may consult [5, Sec. 2.3]. One aim of this section is to prove that GGS-groups which act on a p-regular rooted tree are locally 2-transitive provided that p is odd (see Theorem 3). In fact, we will work in greater generality and study generalised GGS-groups which act on pk -regular trees. Theorem 4 provides a criterion for local 2-transitivity of such groups. 16 S. KIONKE Let p be a prime number and let k ≥ 1 be an integer. We define X = Z/pk Z. In this section we define Xi = X for all i ∈ N. The associated tree T is the pk -regular rooted tree. For convenience, we put m = pk . We note that the tree hanging from a vertex u is isomorphic to T itself. For every g ∈ StG (u) the restriction of g to the subtree hanging from u is an element gu ∈ G. In particular, we obtain an isomorphism ψ : G(1) → G × · · · × G {z } | m times defined as ψ(g) = (g1+mZ , . . . , gm+mZ ). Let α ∈ Sym(X) be the cyclic permutation α(u) = u + 1. Let a ∈ G be the automorphism which permutes the vertices on the first level, and consequently also the trees hanging from them, cyclically according to α, i.e. the label at the root is a(ε) = α. Let e = (e1 , . . . , em−1 , em ) ∈ (Z/mZ)m . We say that e is a defining vector if em = 0 and ei 6≡ 0 mod p for some index i. In most references on GGS-groups the defining vector is of length m − 1 and the last entry em = 0 is ignored. For our purposes, however, it is very convenient to consider the additional entry em = 0 to be part of the defining vector. In addition, it is convenient to define ei+mℓ = ei for all ℓ ∈ Z, such that we are able to consider the indices of e1 , . . . , em modulo m. We use a defining vector e to define recursively an automorphism be ∈ G(1) ⊆ G by imposing ψ(be ) = (ae1 , ae2 , . . . , aem−1 , be ). Definition 3. Let e be a defining vector. The subgroup Ge of G generated by a and be is called the GGS-group defined by e. Whenever the defining vector is clear from the context, we write G instead of Ge and b instead of be . 4.1. Aperiodic defining vectors. We say that a defining vector e is periodic modulo p, if there is a t ∈ {1, . . . , m − 1} such that ei ≡ ei+t mod p for all i. Otherwise, e is said to be aperiodic modulo p. For every t ∈ {1, . . . , m − 1} we define the (2 × m)-matrix   e1 ... em At = e1+t . . . em+t with entries in the local ring Z/mZ. The reduction of At modulo p will be denoted by At . Lemma 7. Let e ∈ (Z/mZ)m be a defining vector. The following statements are equivalent. (1) e is aperiodic modulo p. (2) At ∈ M2×m (Fp ) has rank 2 for all t ∈ {1, . . . , m − 1}. (3) Left multiplication with At defines a surjective Z/mZ-linear map from (Z/mZ)m onto (Z/mZ)2 for all t ∈ {1, . . . , m − 1}. Proof. (1) =⇒ (2): Assume that At does not have rank 2. Then the two rows are linearly dependent and there is λ ∈ Fp such that ei ≡ λei+t mod p for all i. k We infer that ei ≡ λm ei+mt ≡ λp ei mod p for all i ∈ {1, . . . , m}. By assumption ej 6≡ 0 mod p for some j and we conclude λ = 1, i.e., e is periodic modulo p. REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 17 (2) =⇒ (3): Since m = pk , the ring Z/mZ is a local ring. In particular, it follows from Nakayama’s lemma that the linear map defined by At is surjective onto (Z/mZ)2 if and only if the linear map defined by At is surjective onto F2p . (3) =⇒ (1): Let t ∈ {1, . . . , m P− 1}. By assumption Pm there is a vector c = (c1 , . . . , cm ) ∈ (Z/mZ)m such that m e c = 1 and i i i=1 i=1 ei+t ci = 0. In particular, since 0 6≡ 1 mod p, we deduce that ej 6≡ ej+t mod p for some j ∈ {1, . . . , m}.  Remark 3. Assume that k = 1, that is m = p, where p is a prime number. In this case every defining vector e is aperiodic (modulo p). Indeed, suppose that ei = ei+t for all i ∈ {1, . . . , p} and some t ∈ {1, . . . , p − 1}. The element t + pZ generates the cyclic group Z/pZ and we conclude that ei = ej for all i and j. This yields a contradiction since ep = 0, whereas by assumption ei 6= 0 for some i ∈ {1, . . . , p−1} 4.2. A criterion for local 2-transitivity of GGS-groups. Let p be a prime number, let m = pk and let e ∈ (Z/mZ)m be a defining vector. We consider the GGS-group G = Ge and its subgroup H = [bk , aℓ ] | k, ℓ ∈ {1, . . . , m − 1} . Here [bk , aℓ ] = bk aℓ b−k a−ℓ denotes the commutator of bk and aℓ . We write bi = b , ae1 , . . . , aem−i ). In particular, we have ai ba−i , so that ψ(bi ) = (ae1−i , . . . , ae−1 , |{z} i the identity ψ([bk , aℓ ]) = ψ(bk b−k ℓ ) (4) = (ak(e1 −e1−ℓ ) , . . . , ak(eℓ−1 −e−1 ) , akeℓ b−k , ak(eℓ+1 −e1 ) , . . . , bk a−kem−ℓ ). Lemma 8. Let u ∈ Ln with n ≥ 1. The group H ∩ StG (u) acts transitively on the descendants D(u). Proof. We first verify that the assertion holds for u ∈ L1 . Say u = t + mZ. We consider the group of labels H(u) = {g(u) | g ∈ H}. Formula (4) implies that H(u) = hαet −ej | j ∈ {1, . . . , m}i If et generates Z/mZ, then H(u) contains the element αet −em of order m. On the other hand, if et is no generator of Z/mZ, then some ej is and thus also et − ej . We deduce that H(u) is the cyclic group generated by α which acts transitively on D(u). As a next step, we will show that Hu = {gu | g ∈ H} contains H for u ∈ L1 as above. Consider the following two cases. Case 1 : et does not generate the group Z/mZ. In this case et−j generates Z/mZ for some j ∈ {1, . . . , m − 1}. Clearly, then also et − et−j is a generator. By (4) the group Hu is generated by aet −et−j and the elements aket b−k for all k. We conclude that Hu = ha, bi = G and thus H ⊆ Hu . Case 2: et is a generator of Z/mZ. In this case, we use (4) to deduce that the group Hu contains all elements of the form aket b−k for k ∈ Z. A short calculation yields Hu ∋ a−ℓet bℓ bk−ℓ a(ℓ−k)et aket b−k = bk−ℓet b−k = [a−ℓet , bk ] for all k, ℓ ∈ {1, . . . , m − 1}. Since et is a generator of Z/mZ, we deduce H ⊆ Hu . Finally, it follows by induction on n ≥ 1, that StG (u) ∩ H acts transitively on D(u) for all u ∈ Ln .  18 S. KIONKE For p = 2 we need to exclude a problematic case. We say that the defining vector e is centered, if em/2 is the only entry of e which generates Z/mZ. Lemma 9. Let p be a prime. If p = 2, we assume that e is not centered. Let u, v ∈ L1 with u 6= v. The image of the homomorphism Ψu,v : [G, G] → G×G defined as Ψu,v (g) = (gu , gv ) contains the subgroup H × H. Proof. Say u = i0 + mZ and v = j0 + mZ with 1 ≤ i0 , j0 ≤ m. Let i, j ∈ {1, . . . , m} \ {i0 , j0 }. Is is easy to verify that for all k, ℓ ∈ Z the following identities hold:   (5) Ψu,v [bki , bℓj0 ] = 1, [akej0 −i , bℓ ] ,   (6) Ψu,v [bki0 , bℓj ] = [bk , aℓei0 −j ], 1 ,   (7) Ψu,v [bki0 , bℓj0 ] = [bk , aℓei0 −j0 ], [akej0 −i0 , bℓ ] . Suppose that ei0 −j is a generator of Z/mZ for some j 6= j0 . Then equation (6) shows that the image of Ψu,v contains H ×1. However, by (5) and (7) the projection of Ψu,v ([G, G]) ∩ (H × H) onto the second factor contains H, therefore the claim follows. By the same argument as above we can conclude if ej0 −i is a generator of Z/mZ for some i 6= i0 . Finally, we observe that one of the two cases above applies. Indeed, if p is odd, then at least one entry of e is a generator and we have i0 − j0 6≡ j0 − i0 mod m. On the other hand, suppose that p = 2 and i0 − j0 ≡ j0 − i0 mod m. In this case i0 − j0 = ±m/2. Since e is not centered, there is some j 6= j0 such that ei0 −j generates Z/mZ.  Theorem 4. Let p be a prime number and recall that m = pk . Let e ∈ (Z/mZ)m be a defining vector. If p = 2, we assume that e is not centered. The GGS-group Ge acts locally 2-transitively on T if and only if the vector e is aperiodic modulo p. In this case (Ge , Pξ ) is a Gelfand pair for every parabolic subgroup Pξ ≤ Ge Proof. As a first step, we show that StG (u)∩StG (v) acts transitively on D(u)×D(v) for all u 6= v in L1 exactly if e is aperiodic modulo p. Assume that e is aperiodic modulo p. Say u = i + mZ and v = i + t + mZ with i ∈ {1, . . . , m} and t ∈ {1, . . . , m − 1}. By Lemma 7 (3) left multiplication with the two row circulant matrix   ei . . . ei+m−1 ei+t . . . ei+m−1+t defines a surjective map onto (Z/mZ)2 . In particular, we find c1 , . . . , cm ∈ Z/mZ such that m−1 m−1 X X cm−j ei+t+j = 0. cm−j ei+j = 1 and j=0 j=0 We deduce that the element g = bc11 bc22 · · · bcmm ∈ G ∩ G(1) satisfies g(v) = Id, whereas g(u) is the cyclic permutation α. Since the same argument applies with u and v interchanged, it follows that G ∩ G(1) = StG (u) ∩ StG (v) acts transitively on D(u) × D(v). Conversely, let t ∈ {1, . . . , m − 1} be arbitrary. Choose u = 1 + mZ and v = 1+t+mZ. Note that an element g ∈ StG (v) which fixes one descendant of v already acts trivially on D(v). Since the action is locally 2-transitive, we deduce that there REPRESENTATIONS ON THE BOUNDARY OF ROOTED TREES 19 is an element g ∈ StG (u) ∩ StG (v) = G ∩ G(1) such that g(v) = Id but g(u) has order m = pk . Since G ∩ G(1) = hb1 , . . . , bm i and (G ∩ G(1))/(G ∩ G(2)) is abelian, we see k−1 k−1 that there is some j ∈ {1, . . . , m} such that (bpj )(u) 6= (bpj )(v) ; in other words, em−j+1 6≡ em−j+t+1 mod p. We deduce that e is aperiodic modulo p. Assume now that e is aperiodic modulo p. The final step is to verify that for distinct vertices u, v ∈ Ln the action of StG (u) ∩ StG (v) on D(u) × D(v) is transitive. We proceed by induction on n. The case n = 1 was already discussed above. Assume now that n ≥ 2. We distinguish two cases. Case 1 : u and v have the same predecessor w ∈ L1 on the first level. Since G is fractal (c.f. [33]), we have G = Gw = {gw | g ∈ StG (w)} and the claim follow from the induction hypothesis. Case 2 : u and v have distinct predecessors u1 6= v1 on the first level L1 . This means, u = u1 ũ and v = v1 ṽ. It follows from Lemma 9 that H × H is contained in the image of the restriction homomorphism Ψu1 ,v1 : [G, G] → G × G. Now, by Lemma 8 the group H ∩ StG (ũ) acts transitively on D(ũ) and similarly H ∩ StG (ṽ) acts transitively on D(ṽ). This proves the assertion. Finally, assume that G acts locally 2-transitively and let Pξ ≤ G be a parabolic subgroup. Since G is fractal, the action of the stabiliser StG (u) on D(u) factors through the regular representation of a cyclic group of order m and thus the representation on C[D(u)] is multiplicity-free. By the local-global principle (G, Pξ ) is a Gelfand pair.  Remark 4. The assumption that the defining vector is not centered for p = 2 is neccessary. Indeed, consider the case p = 2 and k = 1. There is precisely one GGS-group G acting on the rooted binary tree T ; it is defined by the vector (1, 0) ∈ (Z/2Z)2 . Note that the defining vector is centered and aperiodic. This is exactly the group considered in Example 2, where we have seen that G does not act locally 2-transitively on T . It seems possible that groups with centered defining vector in general do not act locally 2-transitively. The closure G of G ∼ = D∞ in G is isomorphic to the pro-2 completion of the infinite dihedral group D∞ . The group P = hbi is a parabolic subgroup of G and one can verify by direct calculation that (G, P ) is nevertheless a Gelfand pair. It seems possible, that a weaker assumption still suffices the prove a local-global principle for Gelfand pairs. References [1] N. Avni, B. Klopsch, U. Onn, and C. Voll, Representation zeta functions of compact p-adic analytic groups and arithmetic groups, Duke Math. J. 162 (2013), 111–197. [2] , Similarity classes of integral p-adic matrices and representation zeta functions of groups of type A2 , Proc. London Math. Soc. 112 (2016), 267–350. [3] L. Bartholdi, R. Grigorchuk, On the spectrum of Hecke type operators related to some fractal groups, Proc. Steklov Inst. Math. 231, (2000), 1–41. [4] L. Bartholdi, and R. Grigorchuk, On parabolic subgroups and Hecke algebras of some fractal groups, Serdica Math. J. 28 (2002), 47–90. [5] L. Bartholdi, R. Grigorchuk, Z. Šuniḱ, Branch groups, Handbook of Algebra, Vol. 3, Elsevier/North-Holland, Amsterdam, 2003. [6] L. Bartholdi, P. de la Harpe, Representation zeta functions of wreath products with finite groups, Groups Geom. Dyn. 4 (2010), no. 2, 209–249. [7] L. Bartholdi, Representation zeta functions of self-similar branched groups, Groups Geom. Dyn. 11 (2017), no. 1, 29–56. 20 S. KIONKE [8] M. B. Bekka, and P. de la Harpe, Irreducibility of Unitary Group Representations and Reproducing Kernels Hilbert Spaces, Expo. Math. 21 (2003), 115–149. [9] D. Bump Lie Groups, 2nd Edition, Springer-Verlag, New York, 2013. [10] T. Ceccherini-Silberstein, F. Scarabotti, F. Tolli, Finite Gelfand pairs and their applications to probability and statistics, J. Math. Sci. 141 (2007), 1182–1229. [11] C. W. Curtis, and I. Reiner, Methods of Representation Theory, Volume I, John Wiley and Sons, New York, 1981. [12] D. D’Angeli, A. Donno, Self-similar groups and finite Gelfand pairs, Algebra Discrete Math. (2007), 54–69. [13] , Appendix: Gelfand pairs associated with the action of G, European J. Combin. 33 (2012), no. 7, 1422–1426. [14] P. Diaconis, Group Representations in Probability and Statistics, Institute of Mathematical Statistics, Hayward, CA, 1988. [15] J. Dieudonné, Éléments d’analyse 6, Chapitre XXII, Gauthier-Villars, Paris, 1975. [16] A. Dudko, R. Grigorchuk, On irreducibility and disjointness of Koopman and quasi-regular representations of weakly branch groups, Modern theory of dynamical systems, 51–66, Amer. Math. Soc., Providence, RI, 2017. [17] G. A. Fernández-Alcober, A. Zugadi-Reizabal, GGS-groups: order of congruence quotients and Hausdorff dimension, Trans. Amer. Math. Soc. 366 (2014), 1993–2017. [18] G. A. Fernández-Alcober, A. Garrido, J. Uria-Albizuri, On the congruence subgroup property for GGS-groups Proc. Amer. Math. Soc. 145 (2017), 3311–3322. [19] R. Grigorchuk, On Burnside’s problem on periodic groups (Russian), Funktsional. Anal. i Prilozhen. 14 (1980), 53–54. , Solved and unsolved problems around one group, in: Infinite groups: geometric, [20] combinatorial and dynamical aspects, Progr. Math. 248, Birkhäuser, Basel, 2005. [21] R. Grigorchuk, Y. Leonov, V. Nekrashevych, V. Sushchansky, Self-similar groups, automatic sequences, and unitriangular representations, Bull. Math. Sci. 6 (2016), 231–285. [22] B. H. Gross, Some applications of Gel’fand pairs to number theory, Bull. Amer. Math. Soc. 24 (1991), 277–301. [23] N. Gupta, S. Sidki, On the Burnside problem for periodic groups, Math. Z. 182 (1983), 385–388. [24] A. Jaikin-Zapirain, Zeta function of representations of compact p-adic analytic groups, J. Amer. Math. Soc. 19 (2006), 91–118. [25] S. Kionke and B. Klopsch, Zeta functions associated to admissible representations of compact p-adic Lie groups, preprint arXiv:1707.08485 (26 Jul 2017). [26] B. Klopsch, Representation growth and representation zeta functions of groups Note Mat. 33 (2013), 107–120. [27] E. Kowalski, An introduction to the representation theory of groups, American Mathematical Society, Providence, RI, 2014. [28] S. Lang, Algebra, 3rd Ed., Springer-Verlag, New York, 2002. [29] G. Letac, Les fonctions sphériques d’un couple de Gelfand symétrique et les chaı̂nes de Markov, Adv. Appl. Prob. 14 (1982), 272–294. [30] M. Larsen and A. Lubotzky, Representation growth of linear groups, J. Eur. Math. Soc. 10 (2008), 351–390. [31] A. Lubotzky and B. Martin, Polynomial representation growth and the congruence subgroup growth, Israel J. Math. 144 (2004), 293–316. [32] E. Pervova, Profinite completions of some groups acting on trees J. Algebra 310 (2007), 858–879. [33] J. Uria-Albizuri, On the concept of fractality for groups of automorphisms of a regular rooted tree, preprint arXiv:1604.05950 (20 Apr 2016). [34] J. S. Wilson, Profinite Groups, Clarendon Press, Oxford, 1998. Karlsruher Institut für Technologie, Fakultät für Mathematik, Institut für Algebra und Geometrie, Englerstr.2, 76131 Karlsruhe, Germany. E-mail address: [email protected]
4math.GR
The null hypothesis of common jumps in case of irregular and asynchronous observations Ole Martin and Mathias Vetter∗ arXiv:1712.07159v1 [math.ST] 19 Dec 2017 Christian-Albrechts-Universität zu Kiel December 21, 2017 Abstract This paper proposes novel tests for the absence of jumps in a univariate semimartingale and for the absence of common jumps in a bivariate semimartingale. Our methods rely on ratio statistics of power variations based on irregular observations, sampled at different frequencies. We develop central limit theorems for the statistics under the respective null hypotheses and apply bootstrap procedures to assess the limiting distributions. Further we define corrected statistics to improve the finite sample performance. Simulations show that the test based on our corrected statistic yields good results and even outperforms existing tests in the case of regular observations. Keywords and Phrases: Asynchronous observations; common jumps; high-frequency statistics; Itô semimartingale; stable convergence AMS Subject Classification: 62G10, 62M05 (primary); 60J60, 60J75 (secondary) 1 Introduction A key issue of the statistical analysis in continuous time is to understand the fine structure of the underlying processes based on discrete observations. Tools are needed in order to answer fundamental questions about suitable models for these processes, and probably the most basic ones are concerned with whether the processes evolve continuously in time or exhibit jumps. Within the class of semimartingales several statistical tests on this topic have been proposed in recent years, both in one and in several (usually two) dimensions and under various null hypotheses. For an overview see for example Sections 11.4 and 15.3 in Jacod and Protter (2012) or Chapters 10 and 14 in Aı̈t-Sahalia and Jacod (2014). In this work we are interested in the null hypothesis that common jumps of a bivariate process exist. In fact this question was tackled and solved prominently in Jacod and Todorov (2009), but only under the additional assumption that observations come in at equidistant and synchronous times. Both assumptions are rarely justified in practice, which is why practitioners usually ignore a lot of observations (and then work with one-minute observations, for example) in order to artificially ∗ Christian-Albrechts-Universität zu Kiel, Mathematisches Seminar, Ludewig-Meyn-Str. 4, 24118 Kiel, Germany. E-mail: [email protected]/[email protected] 1 construct a sampling scheme which is close enough to justify the assumptions of regularity. Both from a theoretical and a practical point of view this is unsatisfactory, and one would like to understand what happens when the underlying observations come in randomly at irregular and asynchronous times. This is precisely what we aim at in this paper, and our study complements the previous work Martin and Vetter (2017) in which a test for the presence of common jumps (so under the opposite null hypothesis of no common jumps) was constructed. It has turned out that the construction of a test in two dimensions is difficult unless one properly understands the univariate situation. For this reason we provide two statistical procedures: First, we want to decide whether or not jumps are present in a realized path of a stochastic process in one dimension, based on discrete irregular observations of the process and under the null hypothesis that jumps exist. Second, we want to decide whether common jumps are present or not in two realized processes, based on irregular and asynchronous observations and under the null hypothesis that there exist common jumps. Thus, this work can not only be understood as a generalization of the results in Jacod and Todorov (2009), but it also complements the univariate discussion from Aı̈t-Sahalia and Jacod (2009). e be a one-dimensional process observed at times t̃i,n and let In the following let X (l) (1) (2) X = (X , X ) be a bivariate process where X (l) is observed at times ti,n , l = 1, 2, all over the time interval [0, T ]. Then for some k ≥ 2 our tests are based on the ratio statistics P 4 e e i≥k:t̃i,n ≤T (Xt̃i,n − Xt̃i−k,n ) P 4 e −X e k i≥1:t̃i,n ≤T (X t̃i,n t̃i−1,n ) and P (1) (2) i,j≥k:ti,n ∧tj,n ≤T k2 P (1) (1) ti,n ti−k,n (X (1) − X (1) (1) (2) i,j≥1:ti,n ∧tj,n ≤T (1) (2) (2) tj,n tj−k,n )2 (X (2) − X (2) (1) (2) (2) (X (1) − X (1) )2 (X (2) − X (2) ti,n ti−1,n tj,n )2 1{(t(1) tj−1,n (1) (2) (2) i−k,n ,ti,n ]∩(tj−k,n ,tj,n ]6=∅} )2 1{(t(1) . (1) (2) (2) i−1,n ,ti,n ]∩(tj−1,n ,tj,n ]6=∅} The common feature of both statistics is that they have a different asymptotic e has jumps in [0, T ] or not, behaviour, depending on whether, for the first one, X (1) (2) or whether, for the second one, X and X have common jumps in [0, T ] or not. In this paper we investigate the asymptotics of these statistics both under the null hypothesis that (common) jumps are present and under the alternative that (common) jumps do not exist. We will see that the limits are similar in structure to those in the setting of equidistant observation times. Further we will develop central limit theorems using techniques from Bibinger and Vetter (2015) and Martin and Vetter (2017) for both statistics under the null hypothesis that (common) jumps are present. Unlike in the setting of equidistant observation times the limiting variables in the central limit theorem are not mixed normal but have a more complicated distribution. We use the bootstrap method introduced in Martin and Vetter (2017) to estimate the quantiles of this limit distribution and to construct a feasible testing procedure. Finally we conduct a simulation study to check the finite sample performance of our tests. As in Martin and Vetter (2017) we use the models from Jacod and Todorov (2009) to compare our results with the results from similar tests in the setting of equidistant observations. Compared to these results we find that the finite 2 sample performance of our tests is rather poor, especially under the null hypothesis, which is due to a rather large contribution of terms which vanish asymptotically in the central limit theorem. To improve the performance of our tests we therefore construct an estimator which (partially) corrects for those asymptotically vanishing terms. Using this estimator we get a huge improvement in the performance of our tests. In particular, our results become even better than the corresponding results for the tests from Aı̈t-Sahalia and Jacod (2009), Jacod and Todorov (2009) and Aı̈t-Sahalia and Jacod (2014) in the setting of equidistant observation times. The remainder of the paper is organized as follows: As the structure of the results and the formal arguments are simpler in the setting where we test for jumps in a onedimensional process we first derive in Section 2 two statistical tests for jumps within (J) a stochastic process. The formal setting and the estimator Φk,T,n is introduced in Section 2.1. In Section 2.2 we derive the consistency of this estimator under both hypotheses, we cover the asymptotics of the estimator in the form of a central limit theorem in Section 2.3, and in Section 2.4 we use a bootstrap method to derive (J) two feasible tests, one using the original statistic Φk,T,n and one using a corrected e (J) (ρ) based on the central limit theorem. Building on the results estimator Φ k,T,n from Section 2 we proceed similarly in Section 3 to derive two statistical tests for deciding whether or not two processes jump at a common time. In Section 4 we examine the finite sample properties of the tests derived in Sections 2 and 3 by means of a simulation study. The appendix containing the proofs is split into two parts: Section 5 contains the main structure of the proofs, leaving out most technical arguments. Section 6, available online, contains all proofs in details and thus fills the gaps left in Section 5. 2 A test for the absence of jumps In this section we will derive a statistical test based on high-frequency observations which allows to decide whether an observed path of a process contains jumps or not. 2.1 Settings and test statistic First we specify the mathematical model for the process and the observation times. Let X be a one-dimensional Itô semimartingale on the probability space (Ω, F, P) of the form Zt Xt = X0 + Zt bs ds + 0 Zt Z σs dWs + 0 δ(s, z)1{|δ(s,z)|≤1} (µ − ν)(ds, dz) 0 R Zt Z + δ(s, z)1{|δ(s,z)|>1} µ(ds, dz). (2.1) 0 R Here, W denotes a standard Brownian motion, µ is a Poisson random measure on R+ × R whose predictable compensator satisfies ν(ds, dz) = ds ⊗ λ(dz) for some σ-finite measure λ on R endowed with the Borelian σ-algebra, b and σ are adapted processes and δ is a predictable function on Ω × R+ × R. For a more detailed discussion of the components of X consult Section 2.1 of Jacod and Protter (2012). We write ∆Xs = Xs − Xs− with Xs− = limt%s Xt for a possible jump of X in s. 3 Further we define a sequence of observation schemes (πn )n∈N via πn = (ti,n )i∈N0 , n ∈ N, where the (ti,n )i∈N0 are increasing sequences of stopping times with t0,n = 0. By |πn |T = sup {ti,n ∧ T − ti−1,n ∧ T |i ≥ 1} we denote the mesh of the observation times up to some fixed time horizon T ≥ 0. Formally we will develop a statistical test which allows to decide whether a realized path X(ω) contains jumps in a given time interval [0, T ] or not. Specifically we want to decide based on the observations (Xti,n (ω))i∈N0 to which of the following two subsets of Ω (J) ΩT = {∃t ∈ [0, T ] : ∆Xt 6= 0} , (C) (J) c ΩT = ΩT (J) ω belongs. Here, ΩT is the set of all ω for which the path of X up to T has at least (C) one jump and ΩT is the set of all ω for which the path of X is continuous on [0, T ]. All our test statistics are based on the increments ∆i,k,n X = Xti,n − Xti−k,n , ∆i,n X = ∆i,1,n X. and we denote by Ii,k,n = (ti−k,n , ti,n ] , i ≥ k ≥ 1, Ii,n = Ii,1,n , the corresponding observation intervals. For convenience we set Ii,k,n = ∅ and accordingly ∆i,k,n X = 0 for i < k. Further we define for k ∈ N and a function h : R → R the functionals X V (h, [k], πn )T = h(∆i,k,n X), i≥k:ti,n ≤T V (h, πn )T = V (h, [1], πn )T . Considering these functionals for g : x 7→ x4 and k ≥ 2 we build the statistic (J) Φk,T,n = V (g, [k], πn )T kV (g, πn )T whose asymptotics we use to construct a statistical tests for the absence of jumps. Remark 2.1. In the setting of equidistant observation times ti,n = i/n our statistic becomes PbnT c g(∆i,k,n X) (J) . (2.2) Φk,T,n = Pi=k bT /nc k i=1 g(∆i,n X) On the other hand in Aı̈t-Sahalia and Jacod (2009) a test is constructed based on the statistic PbnT /kc g(∆ik,k,n X) (J) i=1 Φk,T,n = P (2.3) bT /nc i=1 g(∆i,n X) where at the lower observation frequency n/k only increments over certain observation intervals Iik,k,n enter the estimation. Intuitively it seems that using the statistic (2.2) should be better than using (2.3), because in (2.2) we utilize the available data more exhaustively by using all increments at the lower observation frequency. This intuition is confirmed by Proposition 10.19 in Aı̈t-Sahalia and Jacod (2014) where central limit theorems are developed for both (2.2) and (2.3) and it is shown that (2.2) has a smaller asymptotic variance. 4 2.2 Consistency (J) In order to derive results on the asymptotic behaviour of Φk,T,n we need to impose certain structural assumptions on the Itô semimartingale X and the observation scheme. Further we introduce the notation X n |Ii,k,n |2 Gk,n (t) = 2 k i≥k:ti,n ≤t and abbreviate Gn (t) = G1,n (t). Condition 2.2. The process bs is locally bounded and the process σs is càdlàg. Furthermore, there exists a locally bounded process Γs withR |δ(ω, s, z)| ≤ Γs (ω)γ(z) for some deterministic bounded function γ which satisfies (γ(z)2 ∧ 1)λ(dz) < ∞, RT and the process σ fulfills 0 |σs |ds > 0 almost surely. Additionally the following assumptions on the observation scheme hold: (i) The sequence of observation schemes (πn )n is exogenous, i.e. independent of the process X and its components, and fulfills P |πn |T −→ 0. (ii) The functions Gn (t), Gk,n (t) convergence pointwise on [0, T ] in probability to strictly increasing functions G, Gk : [0, T ] → [0, ∞). The structural assumptions on b, σ, δ are not very restrictive and occur elsewhere in the literature in similar form. The assumption that σ almost surely does not vanish on [0, T ] and Condition 2.2(ii) are only needed to derive the asymptotic behaviour (J) (C) (J) (J) of Φk,T,n on ΩT . Φk,T,n converges on ΩT also without these assumptions. We will see in the proof of Theorem 2.3 that (J) P V (g, πn )T −→ BT = X (∆Xt )4 , t≤T P V (g, [k], πn )T −→ (J) kBT (J) (J) (C) (J) which yields the asymptotic behaviour of Φk,T,n on ΩT . On ΩT we have BT = 0 and we expand the fraction by n to get an asymptotic result. To describe the limit in that case we define Z T Z T (J) (J) CT = 3σs4 dG(s), Ck,T = 3σs4 dGk (s). 0 0 Theorem 2.3. Under Condition 2.2 it holds  (J) 1, on ΩT , P (J) (J) Φk,T,n −→ kCk,T  (J) , on Ω(C) T . CT Remark 2.4. We obtain Gn (t) − Gn (s) + O(|πn |t ) ≤ kGk,n (t) − kGk,n (s) ≤ kGn (t) − kGn (s) 5 (2.4) for all t ≥ s ≥ 0 from the series of elementary inequalities k X a2i ≤ i=1 k X ai 2 ≤k i=1 k X a2i (2.5) i=1 which holds for any a1 , . . . , ak ≥ 0, k ∈ N. Here, the second inequality follows from the Cauchy-Schwarz inequality. Equality in (2.5) holds for a1 ≥ 0, a2 = . . . = ak = 0 respectively a1 = . . . = ak > 0. The relations of Gn and Gk,n are preserved in the limit which yields G(t) − G(s) ≤ kGk (t) − kGk (s) ≤ kG(t) − kG(s) for all t ≥ s ≥ 0. Hence we get (J) kCk,T ∈ [1, k]. (J) CT (J) (J) P Based on the fact that Φk,T,n −→ 1 on ΩT (J) a random variable which is strictly greater than 1 if we will construct a test with critical region (J) (C) and that Φk,T,n converges on ΩT (J) (J) kCk,T > (J) CT to by Remark 2.4, (J) Ck,T,n = {Φk,T,n > 1 + ck,T,n } (2.6) (J) for an appropriate series of decreasing random positive numbers ck,T,n , n ∈ N. In the following we illustrate the result from Theorem 2.3 by looking at two prominent observation schemes: first Poisson sampling which is truely random and asynchronous and second equidistant sampling for which results exist in the literature and which will serve as a kind of benchmark. Example 2.5. Consider the observation scheme where ti,n −ti−1,n are i.i.d. Exp(nλ), λ > 0, distributed. We will call this observation scheme Poisson sampling as the observation times ti,n correspond to the jump times of a Poisson process with intensity nλ. In this setting Condition 2.2(ii) is fulfilled as shown in Proposition 1 from Hayashi and Yoshida (2008) with G(t) = 2 t, λ and as proved in Section 6.3 with Gk (t) = k+1 t. kλ (2.7) This yields that the limit under the alternative is (k + 1)/2. Remark 2.6. In the case of equidistant synchronous observations, i.e. ti,n = i/n, it holds G(t) = Gk,n (t) = t which yields (J) kCk,T (J) CT 6 = k. |Iin (s)+2,3,n | |Iin (s)+1,3,n | |Iin (s),3,n | s tin (s)−3,n tin (s)−2,n tin (s)−1,n tin (s),n tin (s)+1,n tin (s)+2,n 3|Iin (s),1,n | Figure 1: Illustrating the origin of ξk,n,− (s), ξk,n,+ (s) for k = 3. (J) (C) Hence in this setting Φk,T,n converges to a known deterministic limit on ΩT as well (J) Φk,T,n which also allows to construct a test using for the null hypothesis of no jumps (compare Section 10.3 in Aı̈t-Sahalia and Jacod (2014)). This is not immediately possible in the irregular setting, unless the law of the generating mechanism is known to the statistician. 2.3 Central limit theorem (J) (J) In this section we derive a central limit theorem for Φk,T,n which holds on ΩT . Denote by in (s) the index of the interval Ii,n associated with s ∈ Ii,n and define k−1 X ξk,n,− (s) = n (k − j)2 |Iin (s)−j,1,n |, ξk,n,+ (s) = n j=1 k−1 X (k − j)2 |Iin (s)+j,1,n |. j=1 ξk,n,− (s) + ξk,n,+ (s) is the σ(πn : n ∈ N)-conditional variance of n k−1 X |k − j|∆in (s)+j,1,n W = n k−1 X ∆in (s)+j,k,n W − nk∆in (s),1,n W. j=0 j=−(k−1),j6=0 This identity is illustrated in Figure 1. The following condition summarizes the assumptions we need additionally to Condition 2.2 to derive a central limit theorem. Condition 2.7. The process X and the sequence of observation schemes (πn )n fulfill Condition 2.2. Further the following additional assumptions on the observation schemes hold: (i) We have |πn |T = oP (n−1/2 ). (ii) The integral Z [0,T ]P P hY i g(x1 , . . . , xP )E hp (ξk,n,− (xp ), ξk,n,+ (xp )) dx1 . . . dxP p=1 7 (2.8) converges for n → ∞ to Z g(x1 , . . . , xP ) [0,T ]P P Z Y hp (y) Γ(xp , dy)dx1 . . . dxP p=1 R for all bounded continuous functions g : RP → R, hp : R2 → R, p = 1, . . . , P , and any P ∈ N. Here Γ(·, dy) is a family of probability measures on [0, T ] RT with uniformly bounded first moments and 0 Γ(x, {(0, 0)})dx = 0. Part (i) of Condition 2.7 guarantees that |πn |T vanishes sufficiently fast, while part (ii) of 2.7 yields that the (ξk,n,− (s), ξk,n,+ (s)) converge in law in a suitable sense. Because of the exogeneity of the observation times we may assume in the following that the probability space has the form (Ω, F, P) = (ΩX × ΩS , X ⊗ S, PX ⊗ PS ) (2.9) where X denotes the σ-algebra generated by X and its components and S denotes the σ-algebra generated by the observation scheme (πn )n . To describe the limit in the upcoming central limit theorem we define q X (J) Fk,T = 4 (∆XSp )3 (σSp − )2 ξk,− (Sp ) + (σSp )2 ξk,+ (Sp )USp . (2.10) p:Sp ≤T Here, (Sp )p≥0 denotes an enumeration of the jump times of X and the (ξk,− (s), ξk,+ (s)), s ∈ [0, T ], are independent random variables which are distributed according to Γ(s, dy) and the Us , s ∈ [0, T ], are i.i.d. standard normal distributed random variables. Both the (ξk,− (s), ξk,+ (s)) and the Us are independent of X and its come Note that F (J) is e F, e P). ponents and defined on an extended probability space (Ω, k,T well-defined because the sum in (2.10) is almost surely absolutely convergent and independent of the choice for the enumeration (Sp )p≥0 ; compare Proposition 4.1.3 in Jacod and Protter (2012). Theorem 2.8. If Condition 2.7 holds, we have the X -stable convergence (J)  √  (J) L−s Fk,T n Φk,T,n − 1 −→ (J) kBT (2.11) (J) on ΩT . (J) (J) Here, the limit Fk,T /(kBT ) in (2.11) is defined on the extended probability space e Further the statement of the X -stable convergence on Ω(J) means that e e P). (Ω, F, T we have     √   (J) e g F (J) /(kB (J) ) Y 1 (J) E g n(Φk,T,n − 1) Y 1Ω(J) → E T k,T Ω T T for all bounded and continuous functions g and all X -measurable bounded random variables Y . For more background information on stable convergence in law we refer to Jacod and Protter (2012), Jacod and Shiryaev (2002) and Podolskij and Vetter (2010). 8 Example 2.9. Condition 2.7 is fulfilled in the setting of Poisson sampling introduced in Example 2.5. Part (i) is fulfilled by Lemma 8 from Hayashi and Yoshida (2008) which states E [(|πn |T )q ] = o(n−α ) (2.12) for all q ≥ 1 and α < q. That 2.7(ii) is fulfilled is proved in Section 6.3.1. 2.4 Testing procedure In this section we develop a statistical test for testing the null hypothesis that (J) t 7→ Xt (ω) has jumps in [0, T ] (i.e. ω ∈ ΩT ) against the alternative that t 7→ Xt (ω) (C) is continuous on [0, T ] (i.e. ω ∈ ΩT ). To employ Theorem 2.8 for this purpose (J) we have to estimate the distribution of the limiting variable Fk,T and therefore the distribution of ξk (Sp ) for the jump times Sp . Because the distribution of ξk (Sp ) depends on the unknown observation scheme around Sp , of which we observe only a single realization, we use a bootstrap method to estimate the distribution from the realization of the observation scheme close to Sp . For this method to work we need some sort of local homogeneity which will be guaranteed by Condition 3.11(i). To formalize the bootstrap method let Ln and Mn be sequences of natural numbers which tend to infinity. Set ξˆk,n,m,− (s) = n k−1 X (k − j)2 |Iin (s)+Vn,m (s)−j,1,n |, j=1 ξˆk,n,m,+ (s) = n (2.13) k−1 X (k − j)2 |Iin (s)+Vn,m (s)+j,1,n | j=1 for m = 1, . . . , Mn where the random variable Vn,m (s) attains values in {−Ln , . . . , Ln } with probabilities Ln  X −1 P(Vn,m (s) = l|S) = |Iin (s)+l,n | |Iin (s)+j,n | , l ∈ {−Ln , . . . , Ln }. (2.14) j=−Ln Here, (ξˆk,n,m,− (s), ξˆk,n,m,− (s)) is chosen from the (ξk,n,− (tin (s)+i,n ), ξk,n,− (tin (s)+i,n )), i = −Ln , . . . , Ln , which make up the 2Ln +1 realizations of (ξˆk,n,− (t), ξˆk,n,− (t)) which lie ’closest’ to s, with probability proportional to the interval length |Iin (s)+i,n |. This corresponds to the probability with which a random variable which is uniformly distributed on the union of these intervals Iin (s)+i,n , i = −Ln , . . . , Ln , but otherwise independent from the observation scheme would fall into the interval Iin (s)+i,n . Due to the structure of the predictable compensator ν the jump times Sp of the Itô semimartingale X are also evenly distributed in time. This explains why we choose such a random variable Vn,m (s) for the estimation of the law of (ξˆk,n,m,− (Sp ), ξˆk,n,m,− (Sp )). Using the estimators (2.13) for realizations of ξk (s) we build the following esti(J) mators for realizations of Fk,T X (J) Fbk,T,n,m = 4 (∆i,n X)3 1{|∆i,n X|>β|Ii,n |$ } i:ti,n ≤T q × (σ̂n (ti,n , −))2 ξˆk,n,m,− (ti,n ) + (σ̂n (ti,n , +))2 ξˆk,n,m,+ (ti,n )Un,i,m , 9 (2.15) m = 1, . . . , Mn , where β > 0 and $ ∈ (0, 1/2). Here an increment which is large compared to a given threshold is identified as a jump and the local volatility is estimated by (σ̂n (s, −))2 = (σ̂n (s, +))2 = 1 bn 1 bn X (∆i,n X)2 , i:Ii,n ⊂[s−bn ,s) X (2.16) (∆i,n X)2 i:Ii,n ⊂[s,s+bn ] for a sequence (bn )n with bn → 0 and |πn |T /bn → 0. Further the Un,i,m are i.i.d. standard normal distributed random variables which are independent of F and defined e e F, e P). on the extended probability space (Ω, Denote by b (J) (α) = Q bα Q k,T,n  (J) Fbk,T,n,m |m = 1, . . . , Mn   (J) the bαMn c-th largest element of the set Fbk,T,n,m |m = 1, . . . , Mn . b (J) (α) converges on Ω(J) under appropriate conditions to the We will see that Q T k,T,n (J) (J) X -conditional α-quantile Qk (α) of Fk,T which is defined via  e F (J) ≤ Q(J) (α) X (ω) = α, ω ∈ Ω(J) , P T k,T k (2.17)  (J) (J) (J) and we set Qk (α) (ω) = 0, ω ∈ (ΩT )c . Such a random variable Qk (α) exists (J) if Condition 2.10 is fulfilled because the X -conditional distribution of Fk,T will be (J) almost surely continuous on ΩT under Condition 2.10. Condition 2.10. Assume that the process X and the sequence of observation schemes (πn )n satisfy Condition 2.7 and that the set {s ∈ [0, T ] : σs = 0} is almost P surely a Lebesgue null set. Further, let the sequence (bn )n fulfill |πn |T /bn −→ 0 and suppose that (Ln )n and (Mn )n are sequences of natural numbers converging to infinity and Ln /n → 0. Additionally,  (i) e P(( e ξˆk,n,1,− (sp ), ξˆk,n,1,+ (sp )) ≤ xp , p = 1, . . . , P |S) P − P Y  e k,− (sp ), ξk,+ (sp )) ≤ xp ) > ε → 0 P((ξ p=1 as n → ∞, for all ε > 0 and any x ∈ R2×P , P ∈ N, and sp ∈ (0, T ), p = 1, . . . , P . (ii) The volatility process σ is itself an Itô semimartingale, i.e. a process of the form (2.1). (C) (iii) On ΩT (J) (J) we have kCk,T > CT almost surely. Part (i) of Condition 2.10 guarantees that the bootstrapped realizations (ξˆk,n,m,− (s), ξˆk,n,m,+ (s)) 10 b (J) (α) consistently estimate the distribution of (ξk,− (s), ξk,+ (s)) and thereby that Q k,T,n (J) (J) yields a valid estimator for Qk (α) on ΩT . Part (ii) is needed for the convergence of the volatility estimators σ̂n (Sp , −), σ̂n (Sp , +) for jump times Sp , and part (iii) (J) guarantees that Φk,T,n converges under the alternative to a value different from 1, which is the limit under the null hypothesis. Theorem 2.11. If Condition 2.10 is fulfilled, the test defined in (2.6) with (J) ck,T,n b (J) (1 − α) Q k,T,n =√ , α ∈ [0, 1], nkV (g, πn )T (2.18) has asymptotic level α in the sense that we have  e Φ(J) > 1 + c (J) F (J) → α P k,T,n k,T,n (2.19) (J) for all F (J) ⊂ ΩT with P(F (J) ) > 0. The test is consistent in the sense that we have  e Φ(J) > 1 + c (J) F (C) → 1 P k,T,n k,T,n (C) for all F (C) ⊂ ΩT (2.20) with P(F (C) ) > 0. Note that to carry out the test introduced in Theorem 2.11 the unobservable √ (J) variable n is not explicitly needed, even though n occurs in the definition of ck,T,n . b (J) (1 − α). What This factor actually cancels as it also enters as a linear factor in Q k,T,n remains is the dependence of bn and Ln on n, though, but for these auxiliary variables only a rough idea of the magnitude of n usually is sufficient. Similar observations hold for all tests constructed later on as well. The simulation results in Section 4.1 show that the convergence in (2.19) is rather √ (J) slow, because certain terms in n(Φk,T,n − 1) which vanish in the limit contribute significantly in the small sample. Our goal is to diminish this effect by including estimates for those terms in the testing procedure. The asymptotically vanishing terms stem from the continuous part which is mostly captured in the small increments. To estimate their contribution we define (J) Ak,T,n = n X (∆i,k,n X)4 1{|∆i,k,n X≤β|Ii,k,n |$ } i≥k:ti,n ≤T X − kn (∆i,n X)4 1{|∆ i≥1:ti,n ≤T (l) $ i,n X|≤β|Ii,n | } . (J) using the same β, $ as in (2.15). It can be shown that Ak,T,n is a consistent estimator (J) (J) for k 2 Ck,T − kC1,T . We then define for ρ ∈ (0, 1) the adjusted estimator (J) e (J) (ρ) = Φ(J) − ρ Φ k,T,n k,T,n n−1 Ak,T,n kV (g, πn )T where we partially correct for the contribution of the asymptotically vanishing terms. 11 Corollary 2.12. Let ρ ∈ (0, 1). If Condition 2.10 is fulfilled, it holds with the notation from Theorem 2.11  e Φ e (J) (ρ) > 1 + c (J) F (J) → α P k,T,n k,T,n (J) for all F (J) ⊂ ΩT (2.21) with P(F (J) ) > 0 and  e Φ e (J) (ρ) > 1 + c (J) F (C) → 1 P k,T,n k,T,n (C) for all F (C) ⊂ ΩT (2.22) with P(F (C) ) > 0. The closer ρ is to 1 the faster is the convergence in (2.21), but also the slower is the convergence in (2.22). Hence an optimal ρ should be chosen somewhere in between. Our simulation results in Section 4.1 show that it is possible to pick a ρ very close to 1 without significantly worsening the power compared to the test from Theorem 2.11. Example 2.13. The assumptions on the observation scheme in Condition 2.10 are fulfilled in the Poisson setting. That part (iii) is fulfilled has been shown in Example 2.5 and that part (i) is fulfilled is proved in Section 6.3.2. In fact for our testing procedure to work in the Poisson setting we do not need the weighting from (2.14). All intervals could also be picked with equal probability. This is due to the fact that the interval lengths (n|Iin (s)+Vn,m (s)+j,n |)j=−(k−1),...,−1,1,...,k−1 and hence ξk,n,m,− (s), ξk,n,m,+ (s) are (asymptotically) independent of n|Iin (s)+Vn,m (s),n |. However the weighting is important if the interval lengths of consecutive intervals are dependent as illustrated in the following example. Example 2.14. Define an observation scheme by t2i,n = 2i/n and t2i+1,n = (2i + 1 + α)/n, i ∈ N0 , with α ∈ (0, 1) (compare Example 33 in Bibinger and Vetter (2015)). Let us consider the case k = 2. The observation scheme is illustrated in Figure 2. It can be easily checked that Condition 2.2 holds with G(t) = (1 + α2 )t and G2 (t) = t. Further it can be shown similarly as in Bibinger and Vetter (2015) that Condition 2.7 is fulfilled for Γ defined via Γ(s, {(1 + α, 1 + α)}) = 1−α , 2 Γ(s, {(1 − α, 1 − α)}) = 1+α 2 for all s > 0. Hence in order for the distribution of ξˆk,n,1 (s) to approximate Γ(s, ·) the variable in (s) + Vn,m (s) has to pick the intervals of length (1 + α)/n with higher probability than those with length (1 − α)/n, because it holds n|Iin (s)+Vn,m (s),n | = 1 + α ⇒ ξˆ2,n,m,− (s) = ξˆ2,n,m,+ (s) = 1 − α, n|Ii (s)+V (s),n | = 1 − α ⇒ ξˆ2,n,m,− (s) = ξˆ2,n,m,+ (s) = 1 + α. n n,m 12 (1 + α)/n (1 − α)/n (1 + α)/n (1 − α)/n (1 + α)/n (1 − α)/n 1/n 3/n t1,n t0,n = 0 5/n t3,n t2,n = 2/n t4,n = 4/n t5,n t6,n = 6/n Figure 2: The sampling scheme from Example 2.14 3 A test for the absence of common jumps In this section we will derive a statistical test based on high-frequency observations which allows to decide whether two processes jump at a common time or not. The methods we use are similar to those in Section 2. However, the form of the occuring variables and the proofs will be different because of special effects due to the asynchronicity of the data. 3.1 Settings and test statistic As a model for the stochastic process we again consider an Itô semimartingale X = (X (1) , X (2) )∗ , which is now two-dimensional, defined on the probability space (Ω, F, P) and of the form Zt Xt = X0 + Zt bs ds + 0 Zt Z σs dWs + 0 δ(s, z)1{kδ(s,z)k≤1} (µ − ν)(ds, dz) 0 R2 Zt Z δ(s, z)1{kδ(s,z)k>1} µ(ds, dz), (3.1) + 0 R2 where W = (W (1) , W (2) )∗ is a two-dimensional standard Brownian motion and µ is a Poisson random measure on R+ × R2 whose predictable compensator satisfies ν(ds, dz) = ds ⊗ λ(dz) for some σ-finite measure λ on R2 endowed with the Borelian σ-algebra. b is a two-dimensional adapted process, ! (1) σs 0 σs = (3.2) (2) p (2) ρs σs 1 − ρ2s σs is a (2 × 2)-dimensional process and δ is a two-dimensional predictable process on (1) (2) Ω × R+ × R2 . σs , σs ≥ 0 and ρs ∈ [−1, 1] are all univariate adapted. The observation scheme  (1) (2) πn = (ti,n )i∈N0 , (ti,n )i∈N0 , n ∈ N, (3.3) here consists of two (in general different) increasing sequences of stopping times (l) (l) (ti,n )i∈N0 , l = 1, 2, with t0,n = 0. By  (l) (l) |πn |T = sup ti,n ∧ T − ti−1,n ∧ T i ≥ 1, l = 1, 2 we again denote the mesh of the observation times up to T ≥ 0. 13 Formally we will develop a statistical test which allows to decide to which of the following two subsets of Ω the ω which generated the observed path t 7→ Xt (ω) belongs (CoJ)  (1) (2) = ∃t ∈ [0, T ] : ∆Xt ∆Xt 6= 0 , (nCoJ) (CoJ) c ΩT = ΩT . ΩT (l) (l) (l) We denote the observation intervals of X (l) by Ii,k,n = (ti−k,n , ti,n ], i ≥ k ≥ 1, (l) (l) l = 1, 2, Ii,n = Ii,1,n , and by (l) (l) (l) ∆i,k,n X (l) = Xti,n − Xti−k,n , (l) (l) ∆i,n X (l) = ∆i,1,n X (l) (l) the increments of X (l) over those intervals. As in Section 2.1 we set Ii,k,n = ∅ and (l) ∆i,k,n X (l) = 0 for k < i. Further we define the following functionals for k ∈ N and h : R2 → R X (1) (2) V (h, [k], πn )T = h(∆i,k,n X (1) , ∆j,k,n X (2) )1{I (1) ∩I (2) i,k,n (1) (2) i,j≥k:ti,n ∧tj,n ≤T j,k,n 6=∅} , V (h, πn )T = V (h, [1], πn )T which are generalizations of the famous estimator for the quadratic covariation based on asynchronous observations from Hayashi and Yoshida (2005). Computing these functionals for f : (x1 , x2 ) 7→ (x1 x2 )2 we build the statistic (CoJ) Φk,T,n = V (f, [k], πn )T k 2 V (f, πn )T whose asymptotic behaviour for k ≥ 2 is the foundation for the upcoming statistical test. By X (CoJ) (1) (2) BT = (∆Xt )2 (∆Xt )2 t≤T we denote the sum of the squared co-jumps of X (1) and X (2) . (l) Remark 3.1. In the setting of equidistant observation times ti,n = i/n, l = 1, 2, (CoJ) the statistic Φk,T,n is equal to PbnT c (1) i,j=k (Xi/n (1) (2) (2) − X(i−k)/n )2 (Xj/n − X(j−k)/n )2 1{((i−k)/n,n]∩((j−k)/n,j/n]6=∅} . PbnT c (1) (1) (2) (2) k 2 i=1 (Xi/n − X(i−1)/n )2 (Xi/n − X(i−1)/n )2 (3.4) In Jacod and Todorov (2009) a test for common jumps is constructed based on the statistic PbnT /kc (1) (1) (2) (2) (Xki/n − Xk(i−1)/n )2 (Xki/n − Xk(i−1)/n )2 i=1 (3.5) PbnT c (1) (1) 2 (X (2) − X (2) 2 (X − X ) ) i=1 i/n (i−1)/n i/n (i−1)/n 14 where at the lower observation frequency n/k only increments over the intervals (l) Iki,k,n , l = 1, 2, enter the estimation. Further in Section 14.1 of Aı̈t-Sahalia and Jacod (2014) a test for common jumps based on PbnT c (1) (1) (2) (2) 2 2 i=k (Xi/n − X(i−k)/n ) (Xi/n − X(i−k)/n ) PbnT c (1) (1) (2) (2) k i=1 (Xi/n − X(i−1)/n )2 (Xi/n − X(i−1)/n )2 (3.6) is discussed. As argued in Remark 2.1 it seems advantegeous to use the statistic (3.6) over (3.5). However, in the asynchronous setting it is a priori not clear which observation intervals should be best paired, because there is no one-to-one correspondence of observation intervals in one process to observation intervals in the other process as there is in the synchronous situation. To use the available data as exhaustively as possible we therefore decided to include products of squared increments over all overlapping observation intervals at the lower observation frequency (CoJ) in the numerator of Φk,T,n . 3.2 Consistency (CoJ) In this section we investigate under which conditions Φk,T,n converges to a certain limit. The following structural assumptions which are similar to those in Condition (CoJ) (CoJ) 2.2 are needed to obtain an asymptotic result for Φk,T,n on ΩT . (1) (2) Condition 3.2. The process bs is locally bounded and the processes σs , σs , ρs are càdlàg. Furthermore, there exists a locallyR bounded process Γs and a deterministic bounded function γ which satisfies (γ(z)2 ∧ 1)λ(dz) < ∞ such that kδ(ω, s, z)k ≤ Γs (ω)γ(z). The sequence of observation schemes (πn )n fulfills P |πn |T −→ 0. (CoJ) Theorem 3.3. Under Condition 3.2 we have on ΩT (CoJ) P Φk,T,n −→ 1. (3.7) (CoJ) (nCoJ) As in Section 2.3 the asymptotic behaviour of Φk,T,n on ΩT cated than on (CoJ) ΩT . e k0 ,n (t) = G Hk0 ,n (t) = is more compli- First we introduce the following functions n (k 0 )3 n (k 0 )3 (1) X (1) (2) |Ii,k0 ,n ∩ Ij,k0 ,n |2 , (2) i,j≥k0 :ti,n ∧tj,n ≤t (1) X (2) |Ii,k0 ,n ||Ij,k0 ,n |1{I (1) (1) (2) i,j≥k0 :ti,n ∧tj,n ≤t i,k0 ,n (2) ∩Ij,k0 ,n 6=∅} , p (l) (1) (1) (2) for k 0 = 1, k. Set W t = (Wt , ρt Wt + 1 − (ρt )2 Wt )∗ such that W is the (l) Brownian motion driving X (l) , l = 1, 2. Further denote by in (s) the index of the 15 (l) (l) interval Ii,n associated with s ∈ Ii,n , l = 1, 2, and set (l) X ηk0 ,n,− (s) = j,k0 ,n (3−l) j≥k0 :tj,n ≤T (l) X ηk0 ,n,+ (s) = j,k0 ,n (l) (l) W (l) in (s),n (∆i,k0 ,n W (l) , . . . , ∆ (l) (l) (l) (l) 2 ) 1{i<i(l) (s)} , n (l) 2 ) 1{i≥i(l) (s)+k0 } , n (l) (3−l) i:Ii,k0 ,n ∩Ij,k0 ,n 6=∅ in (s)−k0 +1,n t (l) (l) X (l) Λk0 ,n (s) = ηk0 ,n,− (s), ∆ (l) (∆i,k0 ,n W (l) (3−l) i:Ii,k0 ,n ∩Ij,k0 ,n 6=∅ 1{s∈I (3−l) } (3−l) j≥k0 :tj,n ≤T (l) X 1{s∈I (3−l) } − s, ∆ (l) in (s)+1,n W W in (s)−1,n (l) (l) (l) (l) , s − t (l) , in (s)−1,n  (l) (l) , . . . , ∆ (l) W , ηk0 ,n,+ (s) 0 i (s)+k −1,n n for k 0 = 1, k. Condition 3.4. The assumptions from Condition 3.2 hold and σ almost surely R T (1) (2) fulfills 0 |σs σs |ds > 0. Further: (i) It holds |πn |T = oP (n−1/2 ). e 1,n (t), G e k,n (t), H1,n (t), Hk,n (t) converge pointwise on [0, T ] in (ii) The functions G e1 , G e k , H1 , Hk : [0, ∞) → [0, ∞). probability to strictly increasing functions G (iii) The integral Z [0,T ]P1 +P2 P1 hY  n (1) (1) g(x1 , . . . , xP1 , x01 , . . . , x0P2 )E h(1) nΛ1,n (xp ), 2 Λk,n (xp ) p k p=1 × P2 Y (2) h(2) nΛ1,n (x0p ), p p=1 n (2) 0 i Λ (x ) dx1 . . . dxP1 dx01 . . . dx0P2 k 2 k,n p converges for n → ∞ to Z [0,T ]P1 +P2 g(x1 , . . . , xP1 , x01 , . . . , x0P2 ) × P2 Z Y 2 p=1 R P1 Z Y 2 p=1 R  h(1) y Γ(1) (xp , dy) p  h(2) y 0 Γ(2) (x0p , dy 0 )dx1 . . . dxP1 dx01 . . . dx0P2 p (l) for all bounded continuous functions g : RP1 +P2 → R, hp : R2(k+1) → R, p = 1, . . . , Pl , and any Pl ∈ N, l = 1, 2. Here Γ(l) (·, dy), l = 1, 2, are fami(l) (l) lies of probability measures on [0, T ] where (Λ1 (x), Λk (x)) ∼ Γ(l) (·, x) has first moments which are uniformly bounded in x. Further the components (l) (l) (l) W have uniformly bounded of Λk (x) which correspond to the ∆ (l) in (x)+j,n second moments. 16 (CoJ) (nCoJ) In order to describe the limit of Φk,T,n on ΩT we define Z T Z T (CoJ) e k0 (s) + (σs(1) σs(2) )2 dHk0 (s), k 0 = 1, k, 2(ρs σs(1) σs(2) )2 dG Ck0 ,T = 0 0 X  (CoJ) (1) 2 (2) (2) (1) Dk0 ,T = (∆XSp ) Rk0 (Sp ) + (∆XSp )2 Rk0 (Sp ) , k 0 = 1, k. p:Sp ≤T (l) Here (Sp )p≥0 is an enumeration of the jump times of X and Rk0 (s) is defined by 0 (l) Rk0 (s) (l) (l) (σs− )2 ηk0 ,− (s) = + k  X + (l) χj + σs− p δ− (s)U− (s) j=−k0 +i i=1 σs(l) −1 X (l) σs− i−1 2 X p (l) (l) δ+ (s)U+ (s) + σs χj + (σs(l) )2 ηk0 ,+ (s), l = 1, 2, k 0 = 1, k, j=1 where (l) (l) (l) (l) (Λ1 (s), Λk (s)) = (η1,− (s), δ− (s), δ+ (s), η1,+ (s)),  (l) (l) (ηk,− (s), χ−k+1 , . . . , χ−1 , δ− (s), δ+ (s), χ1 , . . . , χk−1 , ηk,+ (s)) e whose dise F, e P) are random variables defined on an extended probability space (Ω, tribution is given by (l) e(Λ1 P (l) (l) (s),Λk (s)) (dy) = Γ(l) (s, dy), l = 1, 2. (l) The (Λ1 (s), Λk (s)), s ∈ [0, T ], are independent of each other and independent of the process X and its components. The random variables U− (s), U+ (s) are i.i.d. e as well. They are also independent of F e F, e P) standard normal and defined on (Ω, (l) and of (Λ(l) (s), Λk (s)). Theorem 3.5. Under Condition 3.4 we have the X -stable convergence (CoJ) (CoJ) L−s Φk,T,n −→ (nCoJ) on ΩT Dk,T (CoJ) D1,T (CoJ) + kCk,T (3.8) (CoJ) + C1,T . (CoJ) From Theorem 3.3 we know that Φk,T,n converges to 1 under the null hypothesis (CoJ) and from Theorem 3.5 we can conclude that Φk,T,n converges under the alternative (CoJ) (CoJ) to a limit which is almost surely different from 1 if kCk,T 6= C1,T , or under mild additional conditions if there is at least one jump in one of the components of X on (CoJ) (CoJ) (CoJ) [0, T ]. Indeed if we have kCk,T 6= C1,T , Φk,T,n has to be different from 1 because (CoJ) Ck,T (CoJ) , C1,T (CoJ) are F-conditionally constant and Dk,T or their F-conditional distributions admit densities. If (CoJ) Dk,T (CoJ) D1,T (CoJ) and D1,T (CoJ) kCk,T = are either zero (CoJ) C1,T holds mild (CoJ) conditions guarantee 6= almost surely which also yields Φk,T,n 6= 1. Hence we will construct a test with critical region  (CoJ) (CoJ) (CoJ) Ck,T,n = |Φk,T,n − 1| > ck,T,n (3.9) 17 (1) I (1) (1) n−1 Ln (s) (1) n−1 Rn (s) in (s),n X (1) s n−1 Ln (s) n−1 Rn (s) X (2) (2) (2) (2) n−1 Ln (s) n−1 Rn (s) I (2) in (s),n (l) (l) Figure 3: Illustration of the terms Ln (s), Rn (s), l = 1, 2, and Ln (s), Rn (s). (CoJ) for a (possibly random) sequence (ck,T,n )n∈N . As in Section 2 we will consider the situation where the observation times are generated by Poisson processes as an example for a random and irregular sampling scheme which fulfills the conditions we need for the testing procedure to work. Example 3.6. Consider the extension of the setting from Example 2.5 to two dimensions where the observation times of X (1) and X (2) originate from the jump times (l) (l) of two independent Poisson processes, i.e. the ti,n − ti−1,n are Exp(nλl )-distributed with λl > 0 and independent for i, n ≥ 1 and l = 1, 2. That Condition 3.4(i) is fulfilled in this setting follows from (2.12). The convergences of G1,n and H1,n have been shown in Proposition 1 of Hayashi and Yoshida (2008), the convergences of Gk,n and Hk,n are proved in Lemma 6.5. That Condition 3.4(iii) is fulfilled in the Poisson setting is proved in Section 6.3.1. 3.3 Central limit theorem (CoJ) (CoJ) Before we start to derive a central limit theorem for Φk,T,n on ΩT , we restrict ourselves to the case k = 2. This simplifies notation, and also from the simulation results in Section 4.1 for the test from Section 2 it seems to be optimal to choose k as small as possible. First we introduce the following notation to describe intervals around some time s at which a common jump might occur (l) (l) L(l) n (s) = n|I (l) |, R(l) n (s) = n|I (l) Ln (s) = n|I (1) ∩ I (2) in (s)−1,n (1) in (s)−1,n in (s)+1,n (2) in (s)−1,n |, l = 1, 2, (1) |, Rn (s) = n|I (1) in (s)+1,n (2) ∩ I (2) in (s)+1,n | which are illustrated in Figure 3 and  (1) (2) (2) Zn (s) = L(1) n , Rn , Ln , Rn , Ln , Rn (s). Limits of these variables will occur in the central limit theorem. To ensure convergence of the Zn (s) we need to impose the following assumption on the observation scheme. 18 Condition 3.7. The process X and the sequence of observation schemes (πn )n fulfill Condition 3.2, 3.4(i)-(ii). Further the integral Z P i hY hp (Zn (xp )) dx1 . . . dxP g(x1 , . . . , xP )E [0,T ]P p=1 converges for n → ∞ to Z P Z Y  g(x1 , . . . , xP ) hp y Γ(xp , dy)dx1 . . . dxP [0,T ]P p=1 R for all bounded continuous functions g : RP → R, hp : R6 → R and any P ∈ N. Here Γ(·, dy) is a family of probability measures on [0, T ] with uniformly bounded RT first moments and 0 Γ(x, {0}6 )dx = 0. Using the limit distribution of Zn (s) implicitly defined in Condition 3.7 we set X (CoJ) (1) (2) F2,T = 4 ∆XSp ∆XSp p:Sp ≤T × h (2) ∆XSp  (1) σSp − q q (1),− (1) (1),+ L(Sp )USp + σSp R(Sp )USp q  (1) (1) (2) + (σSp − )2 (L(1) − L)(Sp ) + (σSp )2 (R(1) − R)(Sp )USp q q  (1),− (1),+ (1) (2) (2) + ∆XSp σSp − ρSp − L(Sp )USp + σSp ρSp R(Sp )USp q (2) (2) (3) + (σSp − )2 (1 − (ρSp − )2 )L(Sp ) + (σSp )2 (1 − (ρSp )2 )R(Sp )USp q i (2) (2) (4) + (σSp − )2 (L(2) − L)(Sp ) + (σSp )2 (R(2) − R)(Sp )USp . Here, (Sp )p≥0 is an enumeration of the common jump times of X (1) and X (2) ,  (1) (1) (2) (2) L , R , L , R , L, R (s) is distributed according to Γ(s, ·) and the vectors (1),− (1),+ (2) (3) (4)  Us , Us , Us , Us , Us are standard normally distributed and independent for different values of s. Similarly as for (2.10) we obtain that the infinite sum in (CoJ) (CoJ) the definition of F2,T has a well-defined limit. Using F2,T we are able to state the following central limit theorem. Theorem 3.8. If Condition 3.7 is fulfilled, we have the X -stable convergence (CoJ)  √  (CoJ) L−s F2,T n Φ2,T,n − 1 −→ (CoJ) 4BT (CoJ) on the set ΩT (3.10) . Example 3.9. Condition 3.7 is fulfilled in the Poisson setting. The convergence of the integrals follows as Condition 3.4(iii) from Lemma 6.6 which is stated and proved in Section 6.3.1. Example 3.10. Condition 3.7 is also fulfilled in the setting of equidistant obser(1) (2) vation times ti,n = ti,n = i/n. In that case we have  L(1) , R(1) , L(2) , R(2) , L, R (s) = (1, 1, 1, 1, 1, 1) 19 for any s ∈ (0, T ]. Hence we get h X (CoJ) (1) (2) (2) (1) (1),− (1) (1),+  F2,T = 4 ∆XSp ∆XSp ∆XSp σSp − USp + σSp USp p:Sp ≤T (2) (1) (1),− (2) (1),+ + ∆XSp σSp − ρSp − USp + σSp ρSp USp q i (2) (2) (3)  + (σSp − )2 (1 − (ρSp − )2 ) + (σSp )2 (1 − (ρSp )2 )USp (CoJ) and F2,T is X -conditionally normally distributed with mean zero and variance h X 2 (1) (2) (2) (1) (1) (2) 16 (∆XSp )2 (∆XSp )2 ∆XSp σSp − + ∆XSp σSp − ρSp − p:Sp ≤T (2) (1) (1) (2) + ∆XSp σSp + ∆XSp σSp ρSp 2 i (1) (2) (2) + (∆XSp )2 (σSp − )2 (1 − (ρSp − )2 ) + (σSp )2 (1 − (ρSp )2 ) which is similar to the result in Theorem 4.1(a) of Jacod and Todorov (2009). 3.4 Testing procedure In this section we will develop a testing procedure based on Theorem 3.8 for which (CoJ) we have to estimate the law of the limiting variable F2,T . Therefore we need to estimate the law of  Z(Sp ) = L(1) , R(1) , L(2) , R(2) , L, R (Sp ). for a time Sp at which a common jump occurs. As in Section 2.4 we will apply a bootstrap method for this purpose. Let (Ln )n and (Mn )n be sequences of natural numbers which tend to infinity. Set (l) Lb(l) n,m (s) = n I (l) (l) in (s)+Vn,m (s)−1,n b (l) (s) = n I (l) , R (l) n,m (l) in (s)+Vn,m (s)+1,n , (3.11) b n,m (s) as the lengths of the overlapping parts of Lb(l) and define Lbn,m (s), R n (s), l = 1, 2, (l) (1) (2) b n (s), l = 1, 2, where Vn,m (s) = (Vn,m (s), Vn,m (s)) ∈ {−Ln , . . . , Ln } × respectively R {−Ln , . . . , Ln } and P(Vn,m (s) = (l1 , l2 )|S) (1) = |I (1) in (s)+l1 ,n (2) ∩ I (2) in (s)+l2 ,n Ln X  | j1 ,j2 =−Ln (1) |I (1) in (s)+j1 ,n (2) ∩ I (2) in (s)+j2 ,n −1 | , l1 , l2 ∈ {−Ln , . . . , Ln }. Because we are in a two-dimensional setting, we have to simultaneously choose observation intervals of X (1) and X (2) around which the in(1) (2) tervals used for simulating Z(s) lie. Here we choose the intervals (Ii,n , Ij,n ) by a probability which is proportional to the length of the overlapping parts of those two intervals. As in Section 2.4 this corresponds to the probability with which a jump (1) (2) time Sp uniformly distributed in time would fulfill Sp ∈ Ii,n and Sp ∈ Ij,n . Based on (3.11) we define via  bn,m (s) = Lb(1) , R b (1) , Lb(2) , R b (2) , Lbn,m , R b n,m (s), m = 1, . . . , Mn , Z n,m n,m n,m n,m 20 estimators for realizations of Z(s). Further we define using the components of bn,m (s) and by considering increments which are larger than a certain threshold Z as jumps X (1) (2) (CoJ) ∆i,n X (1) ∆j,n X (2) 1{|∆(1) X (1) |>β|I (1) |$ } 1{|∆(2) X (2) |>β|I (2) |$ } FbT,n,m = 4 (1) i,n (2) i,n j,n j,n i,j:ti,n ∧tj,n ≤T × h (2) ∆j,n X (2)  σ̂ (1) q (1,2),− Lbn,m (τi,j,n )Un,(i,j),m q (1) b n,m (τi,j,n )U (1,2),+ + σ̂ (1) (ti,n , +) R n,(i,j),m (1) (ti,n , −)  (1) b + (σ̂ (1) (ti,n , −))2 (Lb(1) n,m − Ln,m )(τi,j,n )  1/2 (1) (1) b (1) − R b n,m )(τi,j,n ) Un,i,m + (σ̂ (1) (ti,n , +))2 (R n,m q  (1) (2) (1,2),− + ∆i,n X (1) σ̂ (2) (tj,n , −)ρ̂(τi,j,n , −) Lbn,m (τi,j,n )Un,(i,j),m q (2) (2) b n,m (τi,j,n )U (1,2),+ + σ̂ (tj,n , +)ρ̂(τi,j,n , +) R n,(i,j),m  (2) (2) 2 2 b + (σ̂ (tj,n , −)) (1 − (ρ̂(τi,j,n , −)) )Ln,m (τi,j,n ) 1/2 (2) (2) b n,m (τi,j,n ) + (σ̂ (2) (tj,n , +))2 (1 − (ρ̂(τi,j,n , +))2 )R Un,j,m  (2) b + (σ̂ (2) (tj,n , −))2 (Lb(2) n,m − Ln,m )(τi,j,n ) 1/2 i (3) (2) (2) 2 b (2) b + (σ̂ (tj,n , +)) (Rn,m − Rn,m )(τi,j,n ) Un,j,m × 1{I (1) ∩I (2) 6=∅} i,n j,n (1) (2) (l) (l) with τi,j,n = ti,n ∧ tj,n . Here, we use (2.16) for the estimation of σs− , σs estimators for ρs− , ρs defined by ρ̂n (s, −) = κ̂n (s, −) , (1) (2) σ̂n (s, −)σ̂n (s, −) ρ̂n (s, +) = and κ̂n (s, +) , (1) (2) σ̂n (s, +)σ̂n (s, +) where κ̂n (s, −) = κ̂n (s, +) = 1 bn 1 bn (1) X (1) (2) ∆i,n X (1) ∆i,n X (2) 1{I (1) ∩I (2) 6=∅} i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂[s−bn ,s) (1) X (1) (2) ∆i,n X (1) ∆i,n X (2) 1{I (1) ∩I (2) 6=∅} , i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] (1,2),− (1,2),+ (1) for a sequence (bn )n with bn → 0 and |πn |T /bn → 0. The Un,(i,j),m , Un,(i,j),m , Un,i,m , (2) (3) Un,j,m , Un,j,m are i.i.d. standard normal and we have β > 0 and $ ∈ (0, 1/2). Similarly as in Section 2.4 we denote by b (CoJ) (α) = Q bα Q T,n  (CoJ) |FbT,n,m | m = 1, . . . , Mn (CoJ)  the bαMn c-th largest element of the set {|FbT,n,m ||m = 1, . . . , Mn } and we will 21 b (CoJ) (α) converges on Ω(CoJ) under appropriate conditions to the X see that Q T,n T (CoJ) conditional α-quantile Q(CoJ) (α) of |F2,T | which is defined via  e |F (CoJ) | ≤ Q(CoJ) (α) X (ω) = α, ω ∈ Ω(CoJ) , P T 2,T (3.12) (CoJ) and we set Q(CoJ) (α)(ω) = 0, ω ∈ (ΩT )c . Such a random variable Q(CoJ) (α) exists because Condition 3.11 will guarantee that the X -conditional distribution of (CoJ) (CoJ) F2,T is almost surely continuous on ΩT . Condition 3.11. The process X and the sequence of observation schemes (πn )n fulfill Condition 3.4 and 3.7. Further the set {s ∈ [0, T ] : kσs k = 0} is almost surely a Lebesgue null set. (Ln )n and (Mn )n are sequences of natural numbers converging to P infinity with Ln /n → 0 and the sequence (bn )n fulfills |πn |T /bn −→ 0. Additionally, (i) P   Y e e e b P(Z(sp ) ≤ xp ) > ε → 0 (3.13) P P(Zn,1 (sp ) ≤ xp , p = 1, . . . , P |S) − p=1 as n → ∞ for all ε > 0 and any xp ∈ R6 , sp ∈ (0, T ), p = 1, . . . , P , P ∈ N, with si 6= sj for i 6= j. (ii) It holds X √ (l) (3−l) n |Ii,n |1/2 |Ij,n |1{I (l) ∩I (3−l) 6=∅} = OP (1), l = 1, 2. (l) (3−l) i,j:ti,n ∧tj,n i,n j,n ≤T (iii) The process σ is a 2 × 2-dimensional matrix-valued Itô semimartingale. (CoJ) (iv) It holds 2C2,T (CoJ) 6= C1,T (CoJ) or D2,T (CoJ) 6= D1,T almost surely. b (CoJ) (α) consistently As for Condition 2.10 part (i) of Condition 3.11 yields that Q T,n estimates Q(CoJ) (α). Part (ii) is a technical condition needed in the proof of Theorem (l) (l) 3.12. Further part (iii) provides that σ̂n (Sp , −), σ̂n (Sp , +), ρ̂n (Sp , −), ρ̂n (Sp , +) (l) (l) are consistent estimators for σSp − , σSp , ρSp − , ρSp , and part (iv) guarantees that the limit under the alternative is almost surely different from 1. Denote by (DisJ) ΩT (CoJ) = {∃s ∈ [0, T ] : ∆Xs(1) 6= 0 ∨ ∆Xs(2) 6= 0} ∩ ΩT the subset of Ω where either X (1) or X (2) has a jump in [0, T ] but there exists no (nCoJ) (CoJ) (CoJ) common jump. This is the subset of ΩT on which D1,T , D2,T do not vanish under Condition 3.11(i)–(ii). Theorem 3.12. Let Condition 3.11 be fulfilled. The test defined in (3.9) with (CoJ) c2,T,n b (CoJ) (1 − α) Q T,n , α ∈ [0, 1], =√ n4V (f, πn )T (3.14) has asymptotic level α in the sense that we have  e |Φ(CoJ) − 1| > c (CoJ) F (CoJ) → α P 2,T,n 2,T,n 22 (3.15) (CoJ) for all F (CoJ) ⊂ ΩT with P(F (CoJ) ) > 0. Further the test is consistent in the sense that we have  e |Φ(CoJ) − 1| > c (CoJ) F → 1 P 2,T,n 2,T,n (3.16) (DisJ) (nCoJ) for all F ⊂ ΩT with P(F ) > 0, and even for all F ⊂ ΩT (CoJ) (CoJ) we have 2C2,T > CT almost surely. with P(F ) > 0 if To improve the performance of our test in the finite sample we adjust the estimator similarly as for the test in Corollary 2.12 by (partially) subtracting terms that contribute in the finite sample but vanish asymptotically. Therefore we define for β, $ as above the quantity X (1) (2) (CoJ) (∆i,2,n X (1) )2 (∆j,2,n X (2) )2 1{I (1) ∩I (2) 6=∅} AT,n = n (CoJ) Φ2,T,n (1) i,2,n (2) j,2,n i,j:ti,n ∧tj,n ≤T × 1{|∆(1) i,2,n X X − 4n (1) (1) (1) |≤β|I (1) |$ ∨|∆(2) X (2) |≤β|I (2) |$ } i,2,n j,2,n j,2,n (2) (∆i,n X (1) )2 (∆j,n X (2) )2 1{I (1) ∩I (2) 6=∅} i,n (2) j,n i,j:ti,n ∧tj,n ≤T × 1{|∆(1) X (1) |≤β|I (1) |$ ∨|∆(2) X (2) |≤β|I (2) |$ } . i,n i,n j,n j,n Using this expression we then define for ρ ∈ (0, 1) by (CoJ) e (CoJ) (ρ) = Φ(CoJ) − ρ Φ 2,T,n 2,T,n n−1 AT,n 4V (f, πn )T the adjusted estimator. Corollary 3.13. Let ρ ∈ (0, 1). If Condition 3.11 is fulfilled, it holds with the notation of Theorem 3.12  e |Φ e (CoJ) (ρ) − 1| > c (CoJ) F (CoJ) → α P 2,T,n 2,T,n (CoJ) for all F (CoJ) ⊂ ΩT (3.17) with P(F (CoJ) ) > 0 and  e |Φ e (CoJ) (ρ) − 1| > c (CoJ) F → 1 P 2,T,n 2,T,n (DisJ) (3.18) (nCoJ) for all F ⊂ ΩT with P(F ) > 0, and even for all F ⊂ ΩT (CoJ) (CoJ) we have 2C2,T > CT almost surely. with P(F ) > 0 if Example 3.14. The assumptions on the observation scheme from Condition 3.11 are fulfilled in the setting of Poisson sampling introduced in Example 2.9. That part (i) holds is shown in Section 6.3.2 and that Condition 3.7 is fulfilled is proven in Section 6.3.1. That part (ii) of Condition 3.11 is fulfilled can be proven in the same P e P e e k,n (t) −→ e k,n (t) −→ way as G Gk and H Hk were shown; compare Lemma 6.5. Similarly as discussed in Section 2.4 after Example 2.13 we can omit the weighting in (3.12) for obtaining a working testing procedure also within the two-dimensional 23 Poisson setting. However, in the setting where both processes X (1) and X (2) are observed at the observation times introduced in Example 2.14 with different α1 6= α2 it can be easily verified that the weighting in (3.12) is necessary and leads to a correct (CoJ) estimation of the distribution of F2,T . 4 Simulation results We conduct simulation studies to verify the effectivity and to study the finite sample properties of the developed tests. 4.1 Testing for jumps In our simulation the observation times originate from a Poisson process with intensity n which corresponds to λ = 1 in Example 2.5 and yields on average n observations in the time interval [0, 1]. We simulate from the model Z dXt = Xt σdWt + α Xt− xµ(dt, dx) R where X0 = 1 and the Poisson measure µ has the predictable compensator ν(dt, dx) = κ 1[−h,−l]∪[l,h] (x) dtdx. 2(h − l) This model is a one-dimensional version of the model used in Section 6 of Jacod and Todorov (2009). We consider the parameter settings displayed in Table 1 with σ 2 = 8 × 10−5 in all cases. We choose n = 100, n = 1,600 and n = 25,600. In a trading day of 6.5 hours Case I-j II-j III-j Cont α 0.01 0.01 0.01 0.00 Parameters κ l h 1 0.05 0.7484 5 0.05 0.3187 25 0.05 0.1238 Table 1: Parameter settings for the simulation. this corresponds to observing X (1) and X (2) on average every 4 minutes, every 15 √ seconds and every second. Further we set β = 0.03, ω = 0.49 and use bn = 1/ n √ for the local interval in the estimation of σs and Ln = bln(n)c, Mn = b10 nc in the simulation of the ξˆk,n,m,− (s), ξˆk,n,m,+ (s). For the choice of these parameters see Section 5 of Martin and Vetter (2017). The cases I-j to III-j correspond to the presence of jumps of diminishing size. When there are smaller jumps we choose a situation where there are more jumps such that the overall contribution of the jumps to the quadratic variation is roughly the same in all three cases. The fraction of the quadratic variation which originates from the jumps matches the one estimated in real financial data from Todorov and Tauchen (2011). In all three cases where the model allows for jumps we only use paths in the simulation study where jumps were realized. In the fourth case Cont we consider a purely continuous model. The 24 parameter values are taken from Jacod and Todorov (2009). We applied the two testing procedures from Theorem 2.11 and Corollary 2.12 for k = 2, 3, 5 and the results are displayed in Figures 5 and 6. In Figure 5 the results from Theorem 2.11 are presented. In the left column the empirical rejection rates are plotted against the theoretical value of the test and in the right column we show (J) estimated density plots based on the simulated values of Φk,T,n . In Figure 6 we present the results from the test in Corollary 2.12 for ρ = 0.9 (for the choice of ρ = 0.9 see Figure 4) in the same way. The density plots here show the estimated e (J) (ρ). density of Φ k,T,n In Figure 5 we observe in the case Cont that the power of the test from Theorem 2.11 is very good for n = 1,600 and n = 25,600. Further the empirical rejection rates match the asymptotic values rather well for all considered values of k in the cases I-j and II-j at least for the highest observation frequency corresponding to n = 25,600. However in the case III-j there is a severe over-rejection even for n = 25,600. In general we observe over-rejection in all cases. The empirical rejection rates match the asymptotic values better in the cases where there are on average larger jumps. Further the results are better the smaller k is. Note that for n = 100 the cases III-j and Cont are not distinguishable using our test as the rejection curves for n = 100 in those two cases are almost identical. The density plots show (J) the convergence of Φk,T,n to 1 in the presence of jumps and to (k + 1)/2 under the alternative as predicted from Example 2.5. In Figure 6 we immediately see that the observed rejection rates from the test from Corollary 2.12 match the asymptotic values much better than those from Theorem 2.11. Hence adjusting the estimator has a huge effect on the finite sample performance of the test. Here the observed rejection rates match the asymptotic values quite well in the case III-j at least for n = 25,600 and in the cases I-j and II-j we get already for n = 1,600 very good results. The results in the case Cont show that the power remains to be very good. The density plots show that under e (J) (ρ) is more centered around 1 than Φ(J) . Under the the presence of jumps Φ k,T,n k,T,n (J) e alternative Φ (ρ) clusters around the value 1 + (1 − ρ)(k − 1)/2 which is much k,T,n e (J) (ρ) still seem to be large closer to 1 than (k + 1)/2, but the observed values of Φ k,T,n enough such that they can be well distinguished from 1 as can be seen from the high empirical rejection rate. Figure 4 illustrates how the performance of the test from Corollary 2.12 depends on the choice of the parameter ρ. For this purpose we investigate for k = 2 the empirical rejection rates in the cases III-j and Cont with n = 25,600 for the test with level α = 5% in dependence of ρ. We plot for ρ ∈ [0, 1] the empirical rejection rate under the null hypothesis in the case III-j which serves as a proxy for the type-I error of the test together with one minus the empirical rejection rate under the alternative hypothesis in the case Cont which serves as a proxy for the type-II error of the test. Finally we plot the sum of both error proxies to obtain an indicator for the overall performance of the test in dependence of ρ. As expected we observe a decrease in the type-I error as ρ increases and an increase in the type-II error. While we observe an approximately linear decrease in the type-I error, the type-II error is equal to zero until 0.8, then slightly increases and starts to steeply increase at ρ = 0.9. As expected from the theory the type-II error converges to 1 − α for 25 1.0 ρ → 1. In this example the overall error is minimized for a relatively large value of ρ close to ρ = 0.9. Further we carried out simulations for the same four parameter settings based on equidistant observation times ti,n = i/n. In this specific setting the test from Theorem 2.11 coincides with tests discussed in Aı̈t-Sahalia and Jacod (2009) and in Chapter 10.3 of Aı̈t-Sahalia and Jacod (2014). The simulation results are presented in Figures 7, 8 and 9 in the same fashion as in Figures 5, 6 and 4 for the irregular observations. We observe that the results both from Theorem 2.11 and Corollary 2.12 based on irregular observations are not significantly worse than those obtained in the simpler setting of equidistant observation times. Especially we can conclude that the adjustment technique introduced for Corollary 2.12 cannot only be used to improve the finite sample performance of our test based on irregular observations, but also can be used to improve existing tests in the literature which are based on equidistant observations. 0.2 0.4 0.6 0.8 Level 1-Power (Level+(1-Power)) 0.0 0.05 0.0 0.2 0.4 0.6 0.8 1.0 ρ Figure 4: This graphic shows for k = 2, α = 5% and n = 25,600 the empirical rejection rate in the case Cont (dotted line) and 1 minus the empirical rejection rate in the case III-j (dashed line) from the Monte Carlo simulation based on Corollary 2.12 as a function of ρ ∈ [0, 1]. 26 1.0 0.0 0.5 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 0 0.0 0.0 2 4 0.5 0.5 6 8 10 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 10 8 6 4 2 0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 0.5 Cont 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 2.5 3.0 c(MIN, MAX) Cont 8 0.0 6 1.0 4 0.5 III-j II-j 2 0.0 0 0.0 0.0 2 4 6 0.5 0.5 8 10 12 14 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) II-j 1.0 I-j 3.0 0.0 0.5 1.0 1.5 2.0 0.5 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 0 0.0 0.0 1 2 3 0.5 0.5 4 5 6 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 6 5 4 3 2 1 0 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 c(MIN, MAX) Cont 5 0.0 4 1.0 3 0.5 Cont 2 0.0 1 1.0 III-j II-j 0 8 6 0.5 0 0.0 2 4 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) I-j 1.0 II-j 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) (a) Rejection curves from the Monte (b) Density estimation of Φ(J) from the k,T,n Carlo for k = 2. Monte Carlo for k = 2. 3.0 0 1 2 3 4 0.5 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.0 1 2 0.5 0.5 3 4 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 0.0 1.0 0.5 1.0 1.5 2.0 2.5 4 3 2 1 0 3.0 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 5 6 c(MIN, MAX) Cont 2.5 0.0 1.0 2.0 0.5 Cont 1.5 0.0 1.0 2 1 1.0 0 0.0 0.5 III-j II-j 0.5 3 4 5 1.0 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) I-j II-j 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) (c) Rejection curves from the Monte (d) Density estimation of Φ(J) from the k,T,n Carlo for k = 3. Monte Carlo for k = 3. 0 1 2 3 4 (e) Rejection curves from the Monte (f) Density estimation of Φ(J) k,T,n from the Carlo for k = 5. Monte Carlo for k = 5. Figure 5: These graphics show the simulation results for the test from Theorem 2.11. The dotted lines correspond to n = 100, the dashed lines to n = 1,600 and the solid lines to n = 25,600. In all cases N = 10,000 paths were simulated. 27 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 6 0 0.0 0.0 2 4 0.5 0.5 8 10 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 10 8 6 4 2 0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 0.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 2.5 3.0 c(MIN, MAX) Cont 60 0.0 50 1.0 40 0.5 Cont 30 0.0 20 1.0 10 0.5 III-j II-j 0 0.0 0 0.0 0.0 2 4 6 0.5 0.5 8 10 12 14 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) II-j 1.0 I-j 3.0 0.0 0.5 1.0 1.5 2.0 2.0 2.5 7 6 4 5 0.5 1.0 0 0.0 1 2 3 0.5 0.5 0.0 0.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 7 6 5 4 3 2 1 0 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 c(MIN, MAX) Cont 35 1.5 c(MIN, MAX) III-j 1.0 1.0 1.0 30 0.5 25 0.0 20 1.0 15 0.5 Cont 10 0.0 5 1.0 III-j II-j 0 8 6 0.5 0 0.0 2 4 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) I-j 1.0 II-j 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) e (J) (0.9) (a) Rejection curves from the Monte (b) Density estimation of Φ k,T,n Carlo for k = 2. from the Monte Carlo for k = 2. 3.0 0 1 2 3 4 0.0 0.5 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 0 0.0 0.0 1 2 0.5 0.5 3 4 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 4 3 2 1 0 3.0 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 5 6 c(MIN, MAX) Cont 20 1.0 15 0.5 Cont 10 0.0 5 1.0 III-j II-j 0 5 4 0.5 0 0.0 1 2 3 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) I-j 1.0 II-j 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) e (J) (0.9) (c) Rejection curves from the Monte (d) Density estimation of Φ k,T,n Carlo for k = 3. from the Monte Carlo for k = 3. 0 1 2 3 4 e (J) (0.9) from (e) Rejection curves from the Monte (f) Density estimation of Φ k,T,n Carlo for k = 5. the Monte Carlo for k = 5. Figure 6: These graphics show the simulation results for the test from Corollary 2.12. The dotted lines correspond to n = 100, the dashed lines to n = 1,600 and the solid lines to n = 25,600. In all cases N = 10,000 paths were simulated. 28 8 6 1.0 0.0 (a) Rejection curves for k = 2. 1.0 1.5 2.0 2.5 2.0 2.5 6 5 4 2 1 0 1.0 0.0 1.0 2.0 2.5 I-j 3 4 5 1.0 0.5 1.0 0 0.0 1 2 0.5 0.5 0.0 0.0 0.0 0.5 0.0 1.0 0.5 1.0 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 2 0 0.0 0.0 1 0.5 0.5 3 1.0 Cont 1.0 III-j 0.0 0.5 8 6 4 2 0 0 1 2 3 4 5 6 II-j 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 c(MIN, MAX) Cont 3.0 0 1 2 3 4 (d) Density estimation of Φ3,T,n . II-j 1.0 1.5 3.0 (J) (c) Rejection curves for k = 3. I-j 0.5 2.5 0.0 1.0 (e) Rejection curves for k = 5. 0.5 1.0 1.5 2.0 2.5 II-j 4 0.5 2.0 (J) Φ2,T,n . 3 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 1.5 1.5 5 1.0 3 0.5 1.0 1.0 4 0.5 c(MIN, MAX) III-j 0.0 0.5 0.5 3 0.0 1.0 1.0 0.5 0.0 0.0 0.0 2 1.0 3.0 2 0.5 Cont 2.5 1 0.0 2.0 1 4 2 1.0 0 0.0 0.5 III-j 1.5 c(MIN, MAX) Cont 3.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 6 1.0 I-j 0.5 0.5 0.0 0.0 1.0 (b) Density estimation of II-j 1.0 I-j 0.5 0.5 0 0.5 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 0.0 0.0 3.0 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 5 6 c(MIN, MAX) Cont 2.0 1.0 3.0 1.5 0.5 0 0.0 2 4 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 2.5 12 2.0 10 1.5 c(MIN, MAX) III-j 1.0 1.0 1.0 8 0.5 6 0.0 4 1.0 1.0 0.5 Cont 2 6 0.0 0.5 1.0 II-j 0.0 0.5 III-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 0.0 0 0.0 0.0 2 4 0.5 0.5 8 10 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) II-j 1.0 I-j 0 1 2 (f) Density estimation of 3 4 (J) Φ5,T,n . Figure 7: Simulation results for the test from Theorem 2.11 based on equidistant observations ti,n = i/n. The dotted lines correspond to n = 100, the dashed lines to n = 1,600 and the solid lines to n = 25,600. In all cases N = 10,000 paths were simulated. 29 1.0 0.0 0.5 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 60 0 0.0 0.0 20 40 0.5 0.5 80 100 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 12 10 8 6 4 2 0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 0.5 Cont 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 2.5 3.0 c(MIN, MAX) Cont 150 6 0.0 100 1.0 50 0.5 III-j II-j 0 0.0 0 0.0 0.0 2 4 0.5 0.5 8 10 12 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) II-j 1.0 I-j 3.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 30 0 0.0 0.0 10 20 0.5 0.5 40 50 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 7 6 5 4 3 2 1 0 3.0 0.0 0.5 3.0 0 1 1.5 2.0 2.5 c(MIN, MAX) III-j 0.0 0.5 1.0 0 0.0 0.0 5 10 0.5 0.5 15 20 1.0 1.0 1.0 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 1 2 3 4 II-j 0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) 0.5 3.0 3.0 4 0.0 0.5 1.0 1.5 2.0 2.5 3.0 5 6 c(MIN, MAX) Cont 25 0.0 3 20 1.0 2 e (J) (0.99). Φ 3,T,n 15 0.5 Cont 3.0 10 0.0 2.5 5 1.0 III-j 2.0 0 5 4 0.5 0 0.0 1 2 3 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) I-j 1.0 1.0 II-j 1.5 c(MIN, MAX) Cont (c) Rejection curves from the Monte (d) Density estimation of Carlo for k = 3. I-j 1.0 80 1.0 60 0.5 Cont 40 0.0 20 1.0 III-j II-j 0 8 6 0.5 0 0.0 2 4 0.5 0.5 0.0 0.0 range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) I-j 1.0 II-j 1.0 I-j range(dens[[1]][[”y”]], dens[[length(n)]][[”y”]]) e (J) (0.99). (a) Rejection curves from the Monte (b) Density estimation of Φ 2,T,n Carlo for k = 2. 0 1 2 3 4 e (J) (0.99). (e) Rejection curves from the Monte (f) Density estimation of Φ 5,T,n Carlo for k = 5. Figure 8: Simulation results for the test from Corollary 2.12 based on equidistant observations ti,n = i/n. The dotted lines correspond to n = 100, the dashed lines to n = 1,600 and the solid lines to n = 25,600. In all cases N = 10,000 paths were simulated. 30 1.0 0.2 0.4 0.6 0.8 Level 1-Power (Level+(1-Power)) 0.0 0.05 0.0 0.2 0.4 0.6 0.8 1.0 ρ Figure 9: This graphic shows for k = 2, α = 5%, n = 25,600 and equidistant observations ti,n = i/n the empirical rejection rate in the case Cont (dotted line) and 1 minus the empirical rejection rate in the case III-j (dashed line) from the Monte Carlo simulation based on Corollary 2.12 as a function of ρ ∈ [0, 1]. We achieve a minimal overall error for approximately ρ = 0.99. 4.2 Testing for disjoint jumps We simulate according to the model used in Section 6 of Jacod and Todorov (2009), since by using the same configurations as in their paper we can compare the performance of our approach to the performance of their methods in the case of equidistant and synchronous observations. The model for the process X is given by Z Z (1) (1) (1) (1) (1) dXt = Xt σ1 dWt + α1 Xt− x1 µ1 (dt, dx1 ) + α3 Xt− x3 µ3 (dt, dx3 ), ZR ZR (2) (2) (2) (2) (2) dXt = Xt σ2 dWt + α2 Xt− x2 µ2 (dt, dx2 ) + α3 Xt− x3 µ3 (dt, dx3 ), R R where [W (1) , W (2) ]t = ρt and the Poisson measures µi are independent of each other and have predictable compensators νi of the form νi (dt, dxi ) = κi 1[−hi ,−li ]∪[li ,hi ] (xi ) dtdxi 2(hi − li ) where 0 < li < hi for i = 1, 2, 3, and the initial values are X0 = (1, 1)T . We consider the same twelve parameter settings which were discussed in Jacod and Todorov (2009) of which six allow for common jumps and six do not. In the case where common jumps are possible, we only use the simulated paths which contain 31 common jumps. For the parameters we set σ12 = σ22 = 8 × 10−5 in all scenarios and choose the parameters for the Poisson measures such that the contribution of the jumps to the total variation remains approximately constant and matches estimations from real financial data; see Huang and Tauchen (2006). The parameter settings are summarized in Table 2; compare Table 1 in Jacod and Todorov (2009). The first six cases in Table 2 describe situations where common jumps might be present, the other six cases situations where there exist no common jumps. In the cases I-j, II-j and III-j there exist only common jumps and no disjoint jumps and the Brownian motions W (1) and W (2) are uncorrelated. In the cases I-m, II-m and III-m we have a mixed model which allows for common and disjoint jumps and also the Brownian motions are positively correlated. In the cases I-d0, II-d0 and III-d0 the Brownian motions W (1) , W (2) are uncorrelated while in the cases I-d1, II-d1 and III-d1 the processes are driven by the same Brownian motion W (1) = W (2) . The prefixes I, II and III indicate an increasing number of jumps present in the observed paths. Since our choice of parameters is such that the overall contribution of the jumps to the quadratic variation is roughly the same in all parameter settings, this corresponds to a decreasing size of the jumps. Hence in the cases I-* we have few large jumps while in the cases III-* we have many small jumps. As a model for the observation times we use the Poisson setting discussed in Examples 2.5 and 2.9 for λ1 = λ2 = 1 and we set T = 1. As in Section 4.1 we choose n = 100, n = 1,600 and n = 25,600. We set β = 0.03 and $ = 0.49 for all √ occuring truncations. We use bn = 1/ n for the local interval in the estimation of √ (l) (l) σs− , σs , ρs and Ln = bln(n)c, Mn = b10 nc in the simulation of the Zbn,m (s). For an explanation for the choice of these parameters see again Section 5 of Martin and Vetter (2017). We only use paths in the simulation where µi ([0, 1], R) 6= 0 whenever αi 6= 0, i = 1, 2, 3. In Figures 10–13 we display the results from the simulation for the testing procedures from Theorem 3.12 and Corollary 3.13. First we plot in Figure 10 for all twelve cases the empirical rejection rates from Theorem 3.12 in dependence of α ∈ [0, 1] as for the plots in the left column of Figure 5. The six plots on the left show the results for the cases where the hypothesis of the existence of common jumps is true. Similarly as in Figure 5 we observe that the empirical rejection rates match the Case I-j II-j III-j I-m II-m III-m I-d0 II-d0 III-d0 I-d1 II-d1 III-d1 ρ 0.0 0.0 0.0 0.5 0.5 0.5 0.0 0.0 0.0 1.0 1.0 1.0 α1 0.00 0.00 0.00 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 κ1 l1 h1 1 5 25 1 5 25 1 5 25 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.7484 0.3187 0.1238 0.7484 0.3187 0.1238 0.7484 0.3187 0.1238 Parameters α2 κ2 l1 0.00 0.00 0.00 0.01 1 0.05 0.01 5 0.05 0.01 25 0.05 0.01 1 0.05 0.01 5 0.05 0.01 25 0.05 0.01 1 0.05 0.01 5 0.05 0.01 25 0.05 h1 0.7484 0.3187 0.1238 0.7484 0.3187 0.1238 0.7484 0.3187 0.1238 α3 0.01 0.01 0.01 0.01 0.01 0.01 κ3 1 5 25 1 5 25 Table 2: Parameter settings for the simulation. 32 l3 0.05 0.05 0.05 0.05 0.05 0.05 h3 0.7484 0.3187 0.1238 0.7484 0.3187 0.1238 I-d0 1.0 1.0 0.0 0.5 1.0 1.0 0.0 0.5 0.5 1.0 1.0 1.0 1.0 0.0 0.5 0.5 0.0 0.5 0.5 III-d1 1.0 1.0 0.0 0.0 0.0 III-d0 0.5 1.0 0.5 1.0 1.0 0.0 0.0 III-m 0.0 0.5 0.5 1.0 1.0 0.5 0.0 II-d1 0.5 0.0 III-j 0.0 1.0 0.0 0.0 1.0 0.5 II-d0 0.5 1.0 0.5 0.0 0.5 0.0 II-m 1.0 II-j 0.0 0.5 0.5 0.0 0.5 0.0 0.5 1.0 0.5 0.0 0.0 I-d1 1.0 I-m 1.0 I-j 0.0 0.5 1.0 0.0 0.5 1.0 Figure 10: Empirical rejection curves from the Monte Carlo simulation for the test derived from Theorem 3.12. The dotted lines represent the results for n = 100, the dashed lines for n = 1,600 and the solid lines for n = 25,600. In each case N = 10,000 paths were simulated. postulated asymptotic level of the test better if n is larger or if common jumps are larger on average. In all six cases where common jumps are present and for all values of n the test overrejects. This is due to the fact that the asymptotically negligible (CoJ) (CoJ) terms in Φ2,T,n (terms which are contained in Φ2,T,n but do not contribute in the (CoJ) limit) tend to be positive, hence Φ2,T,n is on average systematically larger than 1 (see Figure 11) which yields the bias. However, at least for n = 25,600 the observed rejection rates match the asymptotic level quite well. Only in the cases III-j and III-m where the jumps are on average very small the empirical rejection rates are still far higher than the asymptotic level. The results are worse in the mixed model than in the model where there are only common jumps, because idiosyncratic jumps contribute to the asymptotically vanishing error. The test has very good power against the alternative of idiosyncratic jumps as can be seen in the six plots on the right hand side which correspond to the cases where there are no common jumps. (CoJ) Figure 11 shows density estimations for Φ2,T,n in all twelve cases. If there are (CoJ) common jumps it is visible in the density plots that Φ2,T,n converges to 1 as n → ∞. However for n = 100, n = 1,600 in the cases II-j, II-m and for n = 100, n = 1,600, n = 25,600 in the cases III-j, III-m the density peaks at a value significantly larger than 1 which corresponds to the overrejection in Figure 10. Under the alternative of 33 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.5 2.5 3.5 1 2 3 4 5 I-d1 0 3.0 2.0 1.0 0.0 15 10 5 0 3.0 I-d0 range(dens[[1]]y, dens[[length(n)]]y) 1.0 I-m range(dens[[1]]y, dens[[length(n)]]y) 0.5 range(dens[[1]]y, dens[[length(n)]]y) 0 5 10 15 I-j 1 2 3 4 1.5 2.0 2.5 0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.5 2.5 3.5 4 3 2 1 0 3.0 2.0 1.0 0.0 10 12 8 6 4 2 0 3.0 range(dens[[1]]y, dens[[length(n)]]y) 1.0 range(dens[[1]]y, dens[[length(n)]]y) 0.5 range(dens[[1]]y, dens[[length(n)]]y) 0 2 4 6 8 10 12 14 c(min(min(dens[[1]][[”x”]],II-j dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), 0.8), II-m II-d0 II-d1 min(max(dens[[1]][[”x”]], dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), MAX)) 0.5 1.5 2.5 3.5 1.5 2.0 2.5 0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.5 2.5 3.5 7 6 5 4 3 2 1 0 5 4 3 2 1 0 12 10 8 6 4 2 0 3.0 range(dens[[1]]y, dens[[length(n)]]y) 1.0 range(dens[[1]]y, dens[[length(n)]]y) 0.5 range(dens[[1]]y, dens[[length(n)]]y) 0 2 4 6 8 10 12 14 c(min(min(dens[[1]][[”x”]],III-j dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), 0.8), III-m III-d0 III-d1 min(max(dens[[1]][[”x”]], dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), MAX)) 0.5 1.5 2.5 3.5 (CoJ) Figure 11: Density estimations for Φ2,T,n from the Monte Carlo. The dotted lines correspond to n = 100, the dashed lines to n = 1,600 and the solid lines to n = 25,600. In all cases N = 10,000 paths were simulated. (CoJ) disjoint jumps Φ2,T,n tends to cluster around 1.5 which corresponds to the results obtained in Example 2.5 for the one-dimensional setting. Our simulation results from Theorem 3.12 are worse than the results in the equidistant setting displayed in Figure 4 of Jacod and Todorov (2009) while the power of our test is much better. This effect is partly due to the fact that, contrary to our approach, in Jacod and Todorov (2009) idiosyncratic jumps, although their contribution is asymptotically negligible, are included in the estimation of the asymptotic variance in the central limit theorem. Hence they consistently overestimate the asymptotic variance which yields lower rejection rates. Further the asymptotically (CoJ) negligible terms in Φ2,T,n are larger relative to the asymptotically relevant terms in the asynchronous setting than in the setting of synchronous observation times which increases the rejection rates in the asynchronous setting. The test from Corollary 3.13 outperforms the test from Theorem 3.12 in the simulation study. This can be seen in Figures 12 and 13 which show the results from the Monte Carlo simulation for the test from Corollary 3.13 with ρ = 0.75 (for the choice of ρ = 0.75 see Figure 14) in the same fashion as for Theorem 3.12 in Figures 10 and 11. In the cases where common jumps are present we observe that the empirical rejection rates match the asymptotic level much better than in Figure 10. In Figure 12 we see that in the cases I-j, I-m, II-j, II-m we get good results already for n = 1,600 and in the cases III-j, III-m at least for n = 25,600. 34 I-d0 1.0 1.0 0.0 0.5 1.0 1.0 0.0 0.5 0.5 1.0 1.0 1.0 1.0 0.0 0.5 0.5 0.0 0.5 0.5 III-d1 1.0 1.0 0.0 0.0 0.0 III-d0 0.5 1.0 0.5 1.0 1.0 0.0 0.0 III-m 0.0 0.5 0.5 1.0 1.0 0.5 0.0 II-d1 0.5 0.0 III-j 0.0 1.0 0.0 0.0 1.0 0.5 II-d0 0.5 1.0 0.5 0.0 0.5 0.0 II-m 1.0 II-j 0.0 0.5 0.5 0.0 0.5 0.0 0.5 1.0 0.5 0.0 0.0 I-d1 1.0 I-m 1.0 I-j 0.0 0.5 1.0 0.0 0.5 1.0 Figure 12: Empirical rejection curves from the Monte Carlo simulation for the test derived from Corollary 3.13. The dotted lines represent the results for n = 100, the dashed lines for n = 1,600 and the solid lines for n = 25,600. In each case N = 10,000 paths were simulated. The power of the test from Corollary 3.13 is practically as good as for the test from e (CoJ) (ρ) instead of Φ(CoJ) allows Theorem 3.12. Hence using the adjusted statistic Φ 2,T,n 2,T,n to get far better level results while the power of the test remains almost the same. e (CoJ) (ρ) is much more centered Figure 13 shows that the adjusted estimator Φ 2,T,n (CoJ) around 1 than Φ2,T,n if there exist common jumps which can be seen e.g. in the e (CoJ) (ρ) clusters around a value very close to 1 cases III-j and III-m. Further Φ 2,T,n also if there exist no common jumps. However, in the cases *-d0 and *-d1 the peak of the density still occurs at a value which is noticeably larger than 1. Figure 14 illustrates similarly as Figure 4 for the test for jumps the performance of the test from Corollary 3.13 in dependence of ρ. We choose the cases III-j and III-d0 as representatives for the null hypothesis and the alternative. As expected the level as well as the power of the test decrease as ρ increases. Here we get the lowest overall error for a value of ρ close to 0.75. 35 1.5 2.0 2.5 0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.5 2.5 3.5 10 20 30 40 50 I-d1 0 15 10 5 0 15 10 5 0 3.0 I-d0 range(dens[[1]]y, dens[[length(n)]]y) 1.0 I-m range(dens[[1]]y, dens[[length(n)]]y) 0.5 range(dens[[1]]y, dens[[length(n)]]y) 0 5 10 15 I-j 1 2 3 4 1.5 2.0 0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.5 2.5 3.5 50 40 30 20 10 0 30 25 20 15 10 5 0 10 12 8 6 4 2 0 2.5 range(dens[[1]]y, dens[[length(n)]]y) 1.0 range(dens[[1]]y, dens[[length(n)]]y) 0.5 range(dens[[1]]y, dens[[length(n)]]y) 0 2 4 6 8 10 12 14 c(min(min(dens[[1]][[”x”]],II-j dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), 0.8), II-m II-d0 II-d1 min(max(dens[[1]][[”x”]], dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), MAX)) 0.5 1.5 2.5 3.5 1.5 2.0 0.5 1.0 1.5 2.0 0.5 1.0 1.5 2.0 2.5 20 15 10 5 0 15 10 5 0 10 12 8 6 4 2 0 2.5 range(dens[[1]]y, dens[[length(n)]]y) 1.0 range(dens[[1]]y, dens[[length(n)]]y) 0.5 range(dens[[1]]y, dens[[length(n)]]y) 0 2 4 6 8 10 12 14 c(min(min(dens[[1]][[”x”]],III-j dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), c(min(min(dens[[1]][[”x”]], 0.8), dens[[length(n)]][[”x”]]), 0.8), III-m III-d0 III-d1 min(max(dens[[1]][[”x”]], dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), min(max(dens[[1]][[”x”]], MAX)) dens[[length(n)]][[”x”]]), MAX)) 0.5 1.0 1.5 2.0 2.5 1.0 e (CoJ) (ρ) from the Monte Carlo. The dotted Figure 13: Density estimations for Φ 2,T,n lines correspond to n = 100, the dashed lines to n = 1,600 and the solid lines to n = 25,600. In all cases N = 10,000 paths were simulated. 0.2 0.4 0.6 0.8 Level 1-Power (Level+(1-Power)) 0.0 0.05 0.0 0.2 0.4 0.6 0.8 1.0 ρ Figure 14: This graphic shows for α = 5% and n = 1,600 the empirical rejection rate in the case III-j and 1 minus the empirical rejection rate in the case III-d0 from the Monte Carlo simulation based on Corollary 3.13 as a function of ρ ∈ [0, 1]. 36 5 Structure of the proofs Throughout the proofs we will assume that the processes bs , σs and Γs are bounded on [0, T ]. By Conditions 2.2 and 3.2 they are all locally bounded. A localization procedure then shows that the results for bounded processes can be carried over to the case of locally bounded processes (see e.g. Section 4.4.1 in Jacod and Protter (2012)). Further we introduce the decomposition Xt = X0 + B(q)t + Ct + M (q)t + N (q)t of the Itô semimartingale (2.1) or (3.1) with Z Z t  bs − (δ(s, z)1{kδ(s,z)k≤1} − δ(s, z)1{γ(z)≤1/q} )λ(dz) ds, B(q)t = 0 Z t Ct = σs dWs , 0 Z tZ M (q)t = δ(s, z)1{γ(z)≤1/q} (µ − ν)(ds, dz), 0 Z tZ N (q)t = δ(s, z)1{γ(z)>1/q} µ(ds, dz). 0 Here q is a parameter which controls whether jumps are classified as small jumps or big jumps. As we have already applied most of the techniques used in the upcoming proofs in Martin and Vetter (2017), we keep the discussion of similar parts brief and add references to Martin and Vetter (2017) for more detailed arguments. Parts in the proofs which are new or specific to the tests introduced in this paper will be discussed in full detail. 5.1 Proofs for Section 2 (J) Proof of Theorem 2.3. We obtain the following two representations for Φk,T,n Pk−1 P (J) Φk,T,n (J) Φk,T,n = = i≥1:tki+l,n ≤T l=0 (∆ki+l,k,n X)4 kV (g, πn )T n k P i≥k:ti,n ≤T (∆i,k,n X)4 nV (g, πn )T , . (5.1) (5.2) Theorem 3.3.1 from Jacod and Protter (2012) yields X P (J) (∆ki+l,k,n X)4 −→ BT , 0 ≤ l ≤ k − 1, i≥1:tki+l ≤T P (J) (J) (J) and V (g, πn )T −→ BT . Because of BT 6= 0 on ΩT plugging these into (5.1) (J) yields the convergence to 1 on ΩT . Proposition A.2 in Martin and Vetter (2017) applied to the Itô semimartingale et = (Xt , Xt )∗ and the observation scheme defined by t̃(l) = ti,n , l = 1, 2, yields X i,n n X (J) (∆i,n C)4 −→ CT P i:ti,n ≤T 37 (5.3) (l) e e by Condition 2.2 because of ρes = 1, σ es = σs , l = 1, 2, and H(t) = G(t) = G(t). Similarly we obtain n X P (J) (∆i,k,n C)4 −→ kCk,T . k (5.4) i:ti,n ≤T (C) We then get the convergence on ΩT nV (g, [k 0 ], πn )T − n X k0 from (5.2) using (5.3), (5.4) and ∆i,k0 ,n C 4 k 0 = 1, k, P 1Ω(C) −→ 0, (5.5) T i:ti ≤T which will be proved in Section 6.1. Note that because of surely and because G is strictly increasing we have (J) CT RT 0 |σs |ds > 0 almost > 0 almost surely. Proof of Theorem 2.8. We prove √ L−s (J) n (V (g, [k], πn )T − kV (g, πn )T ) −→ Fk,T (5.6) from which (2.11) easily follows. Step 1. First we discretize σ and restrict ourselves to the big jumps ∆N (q)s of which there are only finitely many. We define σ(r), C(r) for r ∈ N by σ(r)t = σj2−r Rt if t ∈ [(j − 1)2−r , j2−r ), C(r)t = 0 σ(r)s dWs . Further let (Sq,p )p be an enumeration of the jump times of N (q). By dxe = min{n ∈ Z : n ≥ x} we denozr for x ∈ R the smallest integer which is greater or equal to x. Using the notion of the jump times we modify the discretized processes σ(r), C(r) further via ( Z t σSq,p if s ∈ [Sq,p , dSq,p /2r e/2r ) e σ̃(r, q)s = , C(r, q)t = σ̃(r, q)s ds. σ(r)s otherwise 0 Using this notation we then define k−1 X √ X e q) + ∆i−j,n C(r, e q)). R(n, q, r) = 4 n (∆i,n N (q))3 (k − j)(∆i+j,n C(r, i:ti,n ≤T j=1 and show √ lim lim sup lim sup P(| n (V (g, [k], πn )T − kV (g, πn )T ) − R(n, q, r)| > ε) → 0 q→∞ r→∞ n→∞ (5.7) for all ε > 0. The proof of (5.7) is given in Section 6.1. Step 2. Next we prove L−s (J) R(n, q, r) −→ Fk,T (q, r) := 4 X (∆N (q)Sq,p )3 Sq,p ≤T × (σ̃(r, q)Sq,p − )2 ξk,− (Sq,p ) + (σ̃(r, q)Sq,p )2 ξk,+ (Sq,p ) for all q > 0 and r ∈ N. 38 1/2 USq,p (5.8) To this end note that on the set Ω(n, q, r) where two different jumps of N (q) are further apart than |πn |T and any jump time of N (q) is further away from j2−r than k|πn |T for any j ∈ {1, . . . , bT 2r c} it holds X R(n, q, r)1Ω(n,q,r) = 4 (∆N (q)Sq,p ) 3√ Sq,p ≤T n k−1 X (k − j) j=1 × (σ̃(r, q)Sq,p − ∆in (Sq,p )−j,n W + σ̃(r, q)Sq,p ∆in (Sq,p )+j,n W )1Ω(n,q,r) . As in the proof of Proposition A.3 of Martin and Vetter (2017) it can be shown that Condition 2.7(ii) yields the X -stable convergence k−1 k−1  √ X   √ X n (k − j)∆in (Sq,p )−j,n W, n (k − j)∆in (Sq,p )+j,n W Sq,p ≤T j=1 j=1 L−s −→  q ξk,− (Sq,p )USq,p ,− , q   ξk,+ (Sq,p )USq,p ,+ Sq,p ≤T for standard normally distributed random variables (Us,− , Us,+ ) which are independent of F and of the ξk,− (s), ξk,+ (s). Using this stable convergence, Proposition 2.2 in Podolskij and Vetter (2010) and the continuous mapping theorem we then obtain 4 X 3√ (∆N (q)Sq,p ) n k−1 X (k − j) j=1 Sq,p ∈P (q,T ) L−s (J) × (σ̃(r, q)Sq,p − ∆in (Sq,p )−j,n W + σ̃(r, q)Sq,p ∆in (Sq,p )+j,n W ) −→ Fk,T (q, r). Because of P(Ω(n, q, r)) → 1 as n → ∞ for any q, r this convergence yields (5.8). Step 3. Finally we need to show e (J) − F (J) (q, r)| > ε) = 0, lim lim sup P(|F k,T k,T q→∞ (5.9) r→∞ for all ε > 0. (5.9) can be proven using that Γ(·, dy) has uniformly bounded first moments together with the boundedness of the jump sizes of X respectively N (q). Combining the results from Steps 1 to 3 we obtain (5.6). The structure of the proof of Theorem 2.11 and especially of (5.10) therein is identical to the structure of the proof of Theorem 4.2 and (A.27) in Martin and Vetter (2017). Proof of Theorem 2.11. For proving (2.19) we need to show  √ e n(V (g, [k], πn )T − kV (g, πn )T ) > Q b k,T,n (1 − α) F (J) → α P which follows from Theorem 2.8 and  e {|Q b (J) (α) − Q(J) (α)| > ε} ∩ Ω(J) → 0 lim P T k,T,n k n→∞ (5.10) for all ε > 0 and any α ∈ [0, 1]. Hence it remains to prove (5.10) for which we give a proof in Section 6.1. 39 (J) (C) For proving (2.20) we observe that Φk,T,n converges on ΩT to a limit strictly greater than 1 by Theorem 2.3 and Condition 2.10(ii). On the other hand we will show (J) ck,T,n 1Ω(C) = oPe (1) (5.11) T in Section 6.1. Hence we obtain (2.20). As a prerequisite for the proof of (5.10) we need the following Lemma which will be proven in Section 6.1. Lemma 5.1. Suppose Conditions 2.7 and 2.10(ii) are fulfilled and Sp is a stopping time with ∆XSp 6= 0 almost surely. Then it holds P P σ̂n (Sp , −)1{Sp <T } −→ σSp − 1{Sp <T } , σ̂n (Sp , +)1{Sp <T } −→ σSp 1{Sp <T } . Proposition 5.2. Suppose Condition 2.10 is fulfilled. Then it holds e P Mn  1 X  e (J) ≤ Υ|X ) > ε ∩ Ω(J) → 0 1{Fb(J) ≤Υ} − P(F T k,T Mn k,T,n,m (5.12) m=1 for any X -measurable random variable Υ and all ε > 0. The proof of Proposition 5.2 requires Lemma 5.1 and will be given in Section 6.1. Proof of (5.10). We have e Q b (J) (α) > Q(J) (α) + ε} ∩ Ω(J) ) P({ T k,T,n k e =P e ≤P Mn  1 X Mn − (bαMn c − 1) (J)  1{Fb(J) >Q(J) (α)+ε} > ∩ ΩT Mn Mn k,T,n,m k  1 Mn m=1 Mn X m=1 1{Fb(J) (J) k,T,n,m >Qk (α)+ε} (J)  − Υ(α, ε) > (1 − α) − Υ(α, ε) ∩ ΩT e (J) > Q(J) (α) + ε|X ). Because the X -conditional distribution of with Υ(α, ε) = P(F k,T k (J) (J) Fk,T is continuous on ΩT , it holds (1 − α) − Υ(α, ε) > 0 almost surely. (5.12) then yields e Q b (J) > Q(J) (α) + ε} ∩ Ω(J) ) → 0 P({ T k,T,n k because Mn 1 X 1{Fb(J) >Q(J) (α)+ε} − Υ(α, ε) Mn k,T,n,m k m=1 (J) converges on ΩT in probability to zero by (5.12). Analogously we derive e Q b (J) < Q(J) (α) − ε} ∩ Ω(J) ) → 0 P({ T k,T,n k which together with (5.13) yields (5.10). 40 (5.13) (J) (J) P (J) Proof of Corollary 2.12. From Ak,T,n −→ k 2 Ck,T −kCT , which can be shown using techniques from the proof of (A.26) in Martin and Vetter (2017), we deduce √ (J) on ΩT (J) nρ n−1 Ak,T,n kV (g, πn )T (J) 1Ω(J) = ρn−1/2 T Ak,T,n P 1Ω(J) −→ 0 kV (g, πn )T (5.14) T and (J) (J) ρ Ak,T,n P knV (g, πn )T 1Ω(C) −→ ρ (J) k 2 Ck,T − kCT (5.15) 1Ω(C) (J) kCT T T (C) on ΩT . (5.14) together with Theorem 2.8 yields the X -stable convergence √ √ L−s (J) e (J) (ρ) − 1) = n(Φ(J) − ρ(kn)−1 A(J) /V (g, πn )T − 1) −→ n(Φ Fk,T k,T,n k,T,n k,T,n (J) on ΩT and hence (2.21) follows as in the proof of (2.19). From (5.15) we derive (J) (J) (J) Φk,T,n −ρ n−1 Ak,T,n kV (g, πn )T P 1Ω(J) − 1 −→ (1 − ρ) (J) k 2 Ck,T − kCT 1Ω(C) (J) kCT T T which is almost surely larger than 0 by Condition 2.10(ii) and hence (2.22) follows as in the proof of (2.20). 5.2 Proofs for Section 3 Proof of Theorem 3.3. From (2.2) in Martin and Vetter (2017) we obtain (CoJ) P V (f, πn )T −→ BT , k−1 X V (f, [k], πn )T = (1) X (2) f (∆ki+l1 ,k,n X (1) , ∆kj+l2 ,k,n X (2) ) l1 ,l2 =0 i,j≥k:t(1) (2) ki+l1 ,n ∧tkj+l2 ,n ≤T × 1{Iki+l1 ,k,n ∩Ikj+l2 ,k,n 6=∅} P −→ k−1 X (CoJ) BT (CoJ) = k 2 BT , l1 ,l2 =0 which yields (3.7). Proof of Theorem 3.5. It holds (CoJ) Φk,T,n = n/k 2 V (f, [k], πn )T nV (f, πn )T and the proof of L−s (CoJ) (n/k 2 V (f, [k], πn )T , nV (f, πn )T ) −→ (Dk,T (CoJ) + kCk,T (CoJ) , D1,T (CoJ) + C1,T ) is similar to the proof of Theorem 3.2 in Martin and Vetter (2017). This yields (3.8) R T (1) (2) because by Condition 3.7 we have 0 |σs σs |ds > 0 which together with the fact (CoJ) that H1 is strictly increasing guarantees C1,T 41 (CoJ) + D1,T (CoJ) ≥ C1,T > 0. The proof of Theorem 3.8 has the same structure as the proof of Theorem 2.8. Proof of Theorem 3.8. We prove √ L−s (CoJ) n (V (f, [2], πn )T − 4V (f, πn )T ) −→ F2,T (5.16) from which (3.10) easily follows. e q) of σ, C as in Step 1 in Step 1. We introduce discretized versions σ̃(q, r), C(r, the proof of Theorem 2.8 with the only difference that by (Sq,p )p we denote an enumeration of the jump times of N (1) (q)N (2) (q) and not of N (q). Thereby we turn the focus to the common jumps. We then show √ e q, r)| > ε) = 0 lim lim sup lim sup P(| n (V (f, [2], πn )T − 4V (f, πn )T ) − R(n, q→∞ r→∞ n→∞ (5.17) for all ε > 0 where e q, r) = R e(1) (n, q, r) + R e(2) (n, q, r), R(n, X √ (l) (l) e (l) e(l) (n, q, r) = n 2 ∆i−1,1,n N (l) (q)∆i,1,n C (r, q) R (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T   (l) e (l) (r, q)∆(l) N (l) (q) ∆(3−l) N (3−l) (q) 2 1 (l) + ∆i−1,1,n C i,1,n j,2,n {I (3−l) i,2,n ∩Ij,2,n 6=∅} , l = 1, 2. The proof for (5.17) will be given in Section 6.2. Step 2. Next we will show L−s (CoJ) e q, r) −→ R(n, FT (q, r) (5.18) for any q > 0 and r ∈ N where (CoJ) FT (q, r) := 4 (1) X (2) ∆XSq,p ∆XSq,p Sq,p ∈P (q,T ) q q (2) (1),− (1),+ × ∆XSq,p σ̃ (1) (r, q)Sq,p − L(Sq,p )USq,p + σ̃ (1) (r, q)Sq,p R(Sq,p )USq,p q  (2) + (σ̃ (1) (r, q)Sq,p − )2 (L(1) − L)(Sq,p ) + (σ̃ (1) (r, q)Sq,p )2 (R(1) − R)(Sq,p )USq,p q q  (1) (1),− (1),+ + ∆XSq,p σ̃ (2) (r, q)Sq,p − ρ̃(r, q)Sq,p − L(Sq,p )USp + σ̃ (2) (r, q)Sq,p ρ̃(r, q)Sq,p R(Sp )USq,p h  + (σ̃ (2) (r, q)Sq,p − )2 (1 − (ρ̃(r, q)Sq,p − )2 )L(Sq,p ) + (σ̃ (2) (r, q)Sq,p )2 (1 − (ρ̃(r, q)Sq,p )2 )R(Sq,p ) + 1/2 (3) USq,p q i (4) (σ̃ (2) (r, q)Sq,p − )2 (L(2) − L)(Sq,p ) + (σ̃ (2) (r, q)Sq,p )2 (R(2) − R)(Sq,p )USq,p . Here σ̃ (1) (r, q), σ̃ (2) (r, q), ρ̃(r, q) are discretized versions of σ (1) , σ (2) , ρ which are obe q) above. The proof of (5.18) will also be tained in the same way as σ̃(r, q), C(r, sketched in Section 6.2. Step 3. Finally we consider e (CoJ) − F (CoJ) (q, r)| > ε) = 0 lim lim sup P(|F T T q→∞ (5.19) r→∞ for all ε > 0 which can be proven using that the first moments of Γ(·, dy) are uniformly bounded together with the boundedness of the jump sizes of X respectively N (q). Combining (5.17), (5.18) and (5.19) yields (5.16). 42 Proof of Theorem 3.12. For proving (3.15) we will prove √ e P  b (CoJ) (1 − α) F (J) → α n |V (f, [2], πn )T − 4V (f, πn )T | > Q T,n which follows from Theorem 3.8 and  (CoJ) (CoJ)  (CoJ) e |Q b lim P (α)| > ε ∩ ΩT =0 T,n (α) − Q (5.20) n→∞ for all ε > 0 and any α ∈ [0, 1]. Hence it remains to prove (5.20) for which we give a proof in Section 6.2. (CoJ) For proving (3.16) we observe that Φ2,T,n converges on the given F to a random variable which is under Condition 3.11(iv) almost surely different from 1 by Theorem (CoJ) (CoJ) 3.5. Note to this end that Ck,T , C1,T are F-conditional constant while the F(CoJ) conditional distribution of Dk,T (CoJ) or Dk,T (CoJ) 6= D1,T (CoJ) , D1,T (CoJ) is continuous. Hence kCk,T (CoJ) 6= C1,T almost surely imply (CoJ) kCk,T (CoJ) + Dk,T (CoJ) 6= C1,T (CoJ) + D1,T a.s. Hence (3.16) follows from (CoJ) c2,T,n 1Ω(nCoJ) = oP (1) (5.21) T which will be proved in Section 6.2. As a prerequisite for the proof of (5.20) we need the following Lemma which will be proven in Section 6.2. Lemma 5.3. Suppose Conditions 3.7 and 3.11(ii) are fulfilled and Sp is a stopping time with ∆XSp 6= 0 almost surely. Then for l = 1, 2 it holds (l) σ̂n(l) (Sp , −)1{Sp <T } −→ σSp − 1{Sp <T } , P P ρ̂n (Sp , −)1{Sp <T } −→ ρSp − 1{Sp <T } , (l) σ̂n(l) (Sp , +)1{Sp <T } −→ σSp 1{Sp <T } , (5.22) P P ρ̂n (Sp , +)1{Sp <T } −→ ρSp 1{Sp <T } . (5.23) Proposition 5.4. Suppose Condition 3.11 is fulfilled. Then it holds e P Mn   1 X e (CoJ) ≤ Υ|X ) > ε ∩ Ω(J) → 0 1{Fb(CoJ) ≤Υ} − P(F 2,T T Mn T,n,m (5.24) m=1 for any X -measurable random variable Υ and all ε > 0. The proof of Proposition 5.4 requires Lemma 5.3 and will be given in Section 6.2. The proof for (5.20) based on Proposition 5.4 is then identical to the proof of (5.10) based on Proposition 5.2 and is therefore omitted here. Proof of Corollary 3.13. Using arguments from the proof of Theorem 3.5 and the proof of (A.26) in Martin and Vetter (2017) we derive (CoJ) L−s AT,n (CoJ) −→ 4(2C2,T (CoJ) − C1,T 43 (CoJ) ) + 4(D2,T (CoJ) − D1,T ). (5.25) (CoJ) Hence on ΩT it holds (CoJ) √ n−1 AT,n P 1Ω(CoJ) −→ 0 n 4V (f, πn )T T and combining this with (3.10) yields the X -stable convergence (CoJ) (CoJ)  n−1 AT,n √  (CoJ) L−s F2,T − 1 −→ n Φ2,T,n − ρ (CoJ) 4V (f, πn )T 4BT (5.26) (CoJ) on ΩT . Replacing (3.10) with (5.26) in the proof of (3.15) yields (3.17). Under the alternative we obtain using Theorem 3.5 and (5.25) (CoJ) (CoJ) Φ2,T,n −ρ n−1 AT,n 4V (f, πn )T  − 1 1Ω(nCoJ) T (CoJ) L−s −→ (1 − ρ) (2C2,T (CoJ) − C1,T (CoJ) C1,T (CoJ) ) + (D2,T (CoJ) − D1,T (CoJ) + D1,T ) 1Ω(nCoJ) T where the limit is almost surely different from zero by Condition 3.11(iv). We then (CoJ) obtain (3.18) as in the proof of Theorem 3.12 because of c2,T,n 1Ω(nCoJ) = oP (1); T compare (5.21). References Aı̈t-Sahalia, Y. and J. Jacod (2009). Testing for jumps in a discretely observed process. Ann. Statist. 37 (1), 184–222. Aı̈t-Sahalia, Y. and J. Jacod (2014). High-Frequency Financial Econometrics. Princeton University Press. ISBN: 0-69116-143-3. Bibinger, M. and M. Vetter (2015). Estimating the quadratic covariation of an asynchronously observed semimartingale with jumps. Annals of the Institute of Statistical Mathematics 67, 707–743. Billingsley, P. (1999). Convergence of probability measures (2 ed.). Wiley series in probability and statistics. Probability and statistics section. Wiley. Brockwell, P. J. and R. A. Davis (1991). Time Series: Theory and Methods (2 ed.). Springer Series in Statistics. Springer. ISBN: 1-441-90319-4. Hayashi, T. and N. Yoshida (2005). On covariance estimation of non-synchronously observed diffusion processes. Bernoulli 11 (2), 359–379. Hayashi, T. and N. Yoshida (2008). Asymptotic normality of a covariance estimator for non-synchronously observed processes. Annals of the Institute of Statistical Mathematics 60 (2), 367–406. Huang, X. and G. Tauchen (2006). The relative contribution of jumps to total price variance. J. Financial Econometrics 4, 456–499. 44 Jacod, J. and P. Protter (2012). Discretization of Processes. Springer. ISBN: 364224-126-3. Jacod, J. and A. Shiryaev (2002). Limit Theorems for Stochastic Processes (2 ed.). Springer. ISBN: 3-540-43932-3. Jacod, J. and V. Todorov (2009). Testing for common arrivals of jumps for discretely observed multidimensional processes. The Annals of Statistics 37 (1), 1792–1838. Martin, O. and M. Vetter (2017). Testing for simultaneous jumps in case of asynchronous observations. To appear in: Bernoulli . Podolskij, M. and M. Vetter (2010). Understanding limit theorems for semimartingales: a short survey. Statistica Neerlandica 64, 329–351. Todorov, V. and G. Tauchen (2011). Volatility jumps. Journal of Business and Economic Statistics 29, 356–371. 45 6 Online-Appendix: Proofs for the arguments omitted in Section 5 We will make repeatedly use of the following estimates. Throughout the upcoming proofs K and Kq will denote generic constants, the latter dependent on q, to simplify notation. Lemma 6.1. If Condition 2.2 or Condition 3.2 is fulfilled and the processes bt , σt , Γt e p,q , eq ≥ 0 such that are bounded there exist constants Kp , Kp0 , Kp,q , K kB(q)s+t − B(q)s kp ≤ Kp,q tp ,   E kCs+t − Cs kp |Fs ≤ Kp tp/2 ,   0 E kM (q)s+t − M (q)s kp |Fs ≤ Kp0 teq ,   e p,q t + Kp,q tp , E kN (q)s+t − N (q)s kp |Fs ≤ K   0 E kXs+t − Xs kp |Fs ≤ Kp0 t, (6.1) (6.2) (6.3) (6.4) (6.5) for all s, t ≥ 0 with s + t ≤ T and all q > 0, p ≥ 1, p0 ≥ 2. Here, eq can be e p,q may be chosen chosen such that eq → 0 for q → ∞. For p ≥ 2 the constant K independently of q. Proof of Lemma 6.1. The inequalities (6.1)–(6.5) follow from Condition 2.2 respectively 3.2 and inequalities (2.1.33), (2.1.34), (2.1.37), (2.1.41) in Jacod and Protter (2012). The following elementary statement will be used frequently in the upcoming proofs. Lemma 6.2. Let (Yn )n∈N be a sequence of real-valued integrable random variables on (Ω, F, P) and G ⊂ F a sub-σ-algebra. Then it holds   P P E |Yn | G −→ 0 ⇒ Yn −→ 0 Proof. For ε < 1 it holds by Markov’s inequality and Jensen’s inequality        E E |Yn | ∧ 1 G E E |Yn | G ∧ 1 E[|Yn | ∧ 1] P(|Yn | > ε) = P(|Yn | ∧ 1 > ε) ≤ = ≤ . ε ε ε This estimate yields the claim as a sequence of random variables (Yen )n∈N converges in probability to zero if and only if E[|Yen | ∧ c] → 0 for any c > 0. 6.1 Proofs for Section 2 Proof of (5.5). We will give the proof for k 0 = 1 only as the proof for general k 0 = k is identical. It holds Xt 1Ω(C) = (Bt + Ct )1Ω(C) with T Z Bt = RT R 0 t Z bs − 0 since T R2  δ(s, z)1{kδ(s,z)k≤1} λ(dz) ds (C) δ(s, z)µ(ds, dz) ≡ 0 on ΩT . 1 (6.6) (C) On ΩT we obtain using (6.1) and (6.2) X   (∆i,n C)4 1Ω(C) S E nV (g, πn )T − n T i:ti,n ≤T X      E (∆i,n B)4 S + E |∆i,n B|3 |∆i,n C| S ≤ nK i:ti,n ≤T     + E (∆i,n B)2 (∆i,n C)2 S + E |∆i,n B| |∆i,n C|3 S X 4 7/2 3 5/2  ≤ nK K Ii,n + K Ii,n + K Ii,n + K Ii,n i:ti,n ≤T X ≤ nK max{T 3/2 , 1}(|πn |T )1/2 Ii,n 2 i:ti,n ≤T which vanishes as n → ∞ by Condition 2.2(i) and (ii). (5.5) then follows from Lemma 6.2. Proof of (5.7). To simplify notation we set ∆i,k0 ,n X = 0, Ii,k0 ,n = ∅, k 0 = 1, k, whenever ti,n > T in this proof. Step 1. We first prove √ lim sup P(| n (V (g, [k], πn )T − kV (g, πn )T ) − R(n)| > ε) → 0 (6.7) n→∞ where R(n) = √ X n 3 4 (∆i,n X) i:ti,n ≤T k−1 X  (k − j) ∆i+j,n X + ∆i−j,n X . j=1 Using the identity 4 (x1 + . . . + xk ) = k X 4 (xi ) + 4 i=1 + 12 k X X k X X (xi ) xj + 6 (xi )2 (xj )2 3 i=1 j>i i=1 j6=i k XX X (xi )2 xj xj 0 + 24 i=1 j6=i j 0 >j k XX X X x i x i0 x j x j 0 (6.8) i=1 i0 >i j>i0 j 0 >j which is a specific case of the multinomial theorem, we derive √ n (V (g, [k], πn )T − kV (g, πn )T ) = √ k−1 X X  (k − j) 4 (∆i,n X)3 (∆i+j,n X) + 4 (∆i,n X)3 (∆i−j,n X) n i:ti,n ≤T j=1 + OP √ n k−1 X X Kk,j (∆i,n X)2 (∆i+j,n X)2  + OP √ n|πn |T  (6.9) i:ti,n ≤T j=1 where we used the inequalities |xi x0i | ≤ (xi )2 +(xi0 )2 , |xj xj 0 | ≤ (xj )2 +(xj 0 )2 to include terms with powers (2, 1, 1, 0) and (1, 1, 1, 1) from (6.8) into the first summand of the last line of (6.9). The last term in (6.9) is due to boundary effects at T . 2 By iterated expectations and inequality (6.5) we get √ k−1 h X X i √ nE Kk,j (∆i,n X)2 (∆i+j,n X)2 S ≤ nKT |πn |T . i:ti,n ≤T j=1 √ Hence those terms and the OP ( n|πn |T )-term vanish for n → ∞ by Condition 2.7(i). This yields (6.7). Step 2. Next we prove lim lim sup P(|R(n) − R(n, q)| > ε) → 0 (6.10) q→∞ n→∞ where R(n, q) = k−1 X  √ X n 4 (∆i,n N (q))3 (k − j) ∆i+j,n C + ∆i−j,n C . i:ti,n ≤T j=1 Therefore we first consider lim lim sup P(|R(n) − R0 (n, q))| > ε) → 0 (6.11) q→∞ n→∞ for all ε > 0 with 0 R (n, q) = √ n X 3 4 (∆i,n N (q)) i:ti,n ≤T k−1 X (k − j) ∆i+j,n X + ∆i−j,n X  j=1 Using 3 |a3 − b3 | = |a − b||a2 + ab + b2 | ≤ |a − b|(a2 + b2 ) 2 we derive |R(n) − R0 (n, q)| X √ ≤ nK ∆i,n X − ∆i,n N (q) (∆i,n X)2 + (∆i,n N (q))2  i:ti,n ≤T × k−1 X |∆i+j,n X + ∆i−j,n X|. (6.12) j=1 The S-conditional expectation of (6.12) is after using iterated expectations, the Hölder inequality with p1 = 3, p2 = 3/2 on ∆i,n (B(q) + C)((∆i,n X)2 + (∆i,n N (q))2 ), the Cauchy-Schwarz inequality on ∆i,n M (q)((∆i,n X)2 + (∆i,n N (q))2 ) and Lemma 3 6.1 bounded by X  √ nK (Kq |Ii,n |3 + |Ii,n |3/2 )1/3 (|Ii,n | + K|Ii,n | + Kq |Ii,n |3 )2/3 i:ti,n ≤T + (eq |Ii,n |)1/2 (|Ii,n | + K|Ii,n | + Kq |Ii,n |4 )1/2 k−1 X  |Ii+j,n |1/2 j=−(k−1),j6=0 ≤ K(Kq (|πn |T ) 1/6 ≤ K(Kq (|πn |T ) 1/6 3 1/2 + (eq + Kq (|πn |T ) ) + (eq ) 1/2 √ X |Ii,n | ) n k−1 X i:ti,n ≤T j=−(k−1),j6=0 |Ii+j,n |1/2 √ X ) n |Ii,k,n ||Ii,k,n |1/2 i:ti,n ≤T  ≤ K(Kq (|πn |T )1/6 + (eq )1/2 ) nT X |Ii,k,n |2 1/2 (6.13) i:ti,n ≤T where we used |Ii,n ||Ii+j,n |1/2 ≤ |Imax{i,i+j},k,n |3/2 for |j| ≤ k − 1 and the CauchySchwarz inequality for sums to obtain the last two inequalities. The last bound vanishes as n → ∞, q → ∞ by Condition 2.2. Hence by Lemma 6.2 we have proved (6.11). Further we prove lim lim sup P(|R(n, q) − R0 (n, q))| > ε) → 0 (6.14) q→∞ n→∞ for all ε > 0. Denote by Ω(n, q) the set where two jumps of N (q) are further apart than k|πn |T . By iterated expectations and Lemma 6.1 it holds E[|R(n, q) − R0 (n, q))|1Ω(n,q) |S] k−1 X  √ X ≤K n E |∆i,n N (q)|3 i:ti,n ≤T |∆i+j,n (B(q) + M (q))| S  j=−(k−1),j6=0 √ X ≤K n (|Ii,n | + Kq |Ii,n |3 ) i:ti,n ≤T k−1 X (Kq |Ii,n |2 + eq |Ii,n |)1/2 j=−(k−1),j6=0  ≤ K(1 + Kq (|πn |T )2 )(Kq |πn |T + eq )1/2 nT X |Ii,k,n |2 1/2 i:ti,n ≤T where the last inequality follows as in (6.13). The last bound vanishes as n → ∞, q → ∞ by Condition 2.2. Hence by Lemma 6.2 we have proved (6.14) because of P(Ω(n, q)) → 1 as n → ∞ for any q > 0 which together with (6.11) yields (6.10). Step 3. Finally we consider lim lim sup lim sup P(|R(n, q) − R(n, q, r)| > ε) → 0. q→∞ r→∞ (6.15) n→∞ Using iterated expectations, inequality (6.4) and (2.1.34) from Jacod and Protter 4 (2012) we obtain E[|R(n, q) − R(n, q, r)||S] k−1 X X √ |Ii,n | ≤ (K + Kq (|πn |T ) )E n 2 i:ti,n ≤T Z ( ti+j,n kσs − σ̃(r, q)s k2 ds)1/2 S  ti+j−1,n j=−(k−1),j6=0 2 ≤ (K + Kq (|πn |T ) ) k−1 X X √ ×E n |Ii,n |(2k − 2)1/2 i:ti,n ≤T Z j=−(k−1),j6=0 ti+j,n kσs − σ̃(r, q)s k2 ds 1/2  S . ti+j−1,n This quantity is using the Cauchy-Schwarz inequality for sums further bounded by X Z ti,n p  1/2  2 2 (K + Kq (|πn |T ) ) Gn (T )E (2k − 2) kσs − σ̃(r, q)s k2 ds S i:ti,n ≤T p  = (K + Kq (|πn |T )2 ) Gn (T )E Z T ti−1,n 1/2  kσs − σ̃(r, q)s k2 ds S 0 T Z p  P −→ K G(T )E kσs − σ̃(r, q)s k2 ds 1/2 S  0 where the convergence as n → ∞ follows from Condition 2.2. Here the limit vanishes as r → ∞ for any q > 0 since the expectation vanishes as r → ∞ by dominated convergence because σ, σ̃(r, q) are bounded by assumption and because (σ̃(r, q))r∈N is a sequence of right continuous elementary processes approximating σ. Hence we get (6.15) by Lemma 6.2. Combining (6.7), (6.10) and (6.15) we obtain (5.7). Proof of (5.11). Note that it holds (J) ck,T,n = √ b (J) nQk,T,n (1 − α) (6.16) nkV (g, πn )T (J) (C) where the denominator in (6.16) converges to kCT > 0 on ΩT as shown in the √ b (J) proof of Theorem 2.3. Hence it remains to show that the numerator nQ n,T (1 − α) (C) is oPe (1) on ΩT . We have √ √ b (J) (1 − α) ≤ nQ n,T Mn n X (J) Fbk,T,n,m . bαMn c (6.17) m=1 Further we get using K|Ii,n |p + Kp |Ii,n |p/2 ≤ Kp |Ii,n |p(1/2−$) β p |Ii,n |p$ √ is defined as in (6.6), the inequality a + b ≤ P(|∆i,n (B + C)|p > β p |Ii,n |p$ |S) ≤ for p ≥ 1 where the process (Bt )t≥0 5 √ √ a + b for a, b ≥ 0, iterated expectations and twice the Cauchy-Schwarz inequality √  (J) E n|Fbk,T,n,m |1Ω(C) S T √ X ≤ n E[|∆i,n (B + C)|3 1{|∆i,n (B+C)|>β|Ii,n |$ } i:ti,n ≤T q q  × σ̂n (ti,n , −) E[ξˆk,n,m,− (ti,n )|F] + σ̂n (ti,n , +) E[ξˆk,n,m,+ (ti,n )|F] |S] 1/4 √ X ≤ n E[(∆i,n (B + C))12 |S]P(|∆i,n (B + C)|p > β p |Ii,n |p$ |S) i:ti,n ≤T 1/2 p × E[2(σ̂n (ti,n , −))2 + 2(σ̂n (ti,n , +))2 |S] K n|πn |T X p 12 6 1/4 p(1−2$)/4 ≤ Kp n |πn |T (Kq |Ii,n | + |Ii,n | ) |Ii,n | (4bn /bn )1/2 i:ti,n ≤T √ √ X ≤ Kp n(|πn |T )1/2+p(1−2$)/4 n |Ii,n |3/2 . i:ti,n ≤T If we pick p such that p(1 − 2$)/4 ≥ 1/2 the final bound vanishes as n → ∞ by Condition 2.7(i) and 2.2(ii); compare (6.13). Hence by Lemma 6.2 we obtain √ b (J) nQn,T (1 − α) = oPe (1). Proof of Lemma 5.1. First we prove the claim for a constant stopping time s ∈ (0, T ). We will only give the proof for σ̂n (s, +) as the proof for σ̂n (s, −) is identical. Without loss of generality we assume s + bn ≤ T . Using the elementary inequality |x2 − y 2 | ≤ 1+ε |x − y|2 + εy 2 ε which holds for all x, y ∈ R and ε > 0 (compare Step 2 in the proof of Theorem 9.3.2 of Jacod and Protter (2012)) we obtain X 1 (σ̂n(l) (s, +))2 − (∆i,n C)2 bn i:Ii,n ⊂[s,s+bn ] ≤ ≤ 1 bn 1 bn X Kε (∆i,n (B(q) + M (q) + N (q)))2 + ε(∆i,n C)2  i:Ii,n ⊂[s,s+bn ] X  3Kε ((∆i,n B(q))2 + (∆i,n M (q))2 + (∆i,n N (q))2 ) + ε(∆i,n C)2 . i:Ii,n ⊂[s,s+bn ] P 2 Here we have i:Ii,n ⊂[s,s+bn ] (∆i,n N (q)) = 0 with probability approaching 1 as n → ∞ because the probability for the event that there exists a jump of N (q) in [s, s+bn ] vanishes as bn → 0. Further the S-conditional expectation of the remaining terms is using (6.1)–(6.3) bounded by  K Kε eq + Kε,q |πn |T + ε which vanishes as first n → ∞, then q → ∞ and finally ε → 0. Hence it remains to show X 1 P (∆i,n C)2 −→ σs2 . bn i:Ii,n ⊂[s,s+bn ] 6 (6.18) Therefore we apply Lemma 2.2.12 from Jacod and Protter (2012) with ζin = 1 1 (∆in (s)+i,n C)2 − (σs )2 |Iin (s)+i,n |, bn bn Gin = σ(Ftin (s)+i,n , S). (6.19) We obtain using the conditional Itô isometry and the fact that σ is bounded X   n |E[ζin |Gi−1 ]| S (6.20) E i:Iin (s)+i,n ⊂[s,s+bn ] 1 = bn ≤ 1 bn Z  E E[ X ti−1,n i:Ii,n ⊂[s,s+bn ] X ti,n  |(σu )2 − (σs )2 |du|Fti−1,n , S] S + OP (|πn |T /bn )  E E[ε|Ii,n | + K|Ii,n |1{supu∈I i,n |(σu )2 −(σs )2 |>ε} |Fti−1,n , S] S  i:Ii,n ⊂[s,s+bn ] + OP (|πn |T /bn ) ≤ (ε + KE[1{supu∈[s,s+bn ] |(σu )2 −(σs )2 |>ε} |S]) + OP (|πn |T /bn ) ≤ ε + KP( sup u∈[s,s+bn ] |(σu )2 − (σs )2 | > ε|S) + OP (|πn |T /bn ) (6.21) where the first term vanishes as ε → 0, the second term vanishes as n → ∞ for all ε > 0 because σ is right-continuous and the third term vanishes as n → ∞ by Condition 2.10. Further we get X 4|πn |T bn P n K −→ 0 (6.22) E[(ζin )2 |Gi−1 ]≤ 2 bn i:Iin (s)+i,n ⊂[s,s+bn ] P from (a − b)2 ≤ 2a2 + 2b2 , inequality (6.2), the boundedness of σ and |πn |T /bn −→ 0. P P P P n ] −→ n ] −→ 0 from (6.20) and 0 and i E[(ζin )2 |Gi−1 Hence we obtain i E[ζin |Gi−1 (6.22) which yields X P ζin −→ 0 i:Iin (s)+i,n ⊂[s,s+bn ] by Lemma 2.2.12 from Jacod and Protter (2012) which implies (6.18). The proof for the convergence of σ̂n (Sp , +) for a stopping time Sp is identical to the proof for σ̂n (s, +) above. For σ̂n (Sp , −) however the situation is different. There we use the sum over all i ∈ N with ti,n ∈ [Sp − bn , Sp ) which can be written as tin (Sp −bn )+i−1,n , i = 1, . . . , in (Sp ) − in (Sp − bn ) + 1. Here, in general tin (Sp −bn )+i−1,n is no stopping time and we cannot use conditional expectations as above in that case. This problem can be circumvented by observing that jump times Sp are asymptotically independent of (Cu )u∈[Sp −δ,Sp +δ] for δ → 0. To formalize this idea we define new filtrations (Ftε )t≥0 , ε > 0, such that jump times Sp with µ(Sp , (−∞, ε] ∪ [ε, ∞)) = 1 are F0ε -measurable. Then X is also an Itô semimartingale with respect to the filtration (Ftε )t≥0 and we have C ε = C for the continuous martingale part C ε of X under the filtration (Ftε )t≥0 . The convergence of σ̂n (Sp , −) for any jump time Sp with µ(Sp , (−∞, ε] ∪ [ε, ∞)) = 1 then follows because Sp is F0ε -measurable and therefore practically deterministic with respect to the new filtration. The statement for general jump times Sp is obtained by letting ε → 0. Consult Step 6 in the proof of Theorem 9.3.2 from Jacod and Protter (2012) for more details. 7 Proof of Proposition 5.2. Denote by Sp , p = 1, . . . , P , the jump times of the P largest jumps of X in [0, T ] and introduce the notation Rk (P, n, m) = P X (∆in (Sp ),n X)3 1{|∆in (Sp ),n X|>β|Iin (Sp ),n |$ } p=1 1/2 × (σ̂n (tin (Sp ),n , −))2 ξˆk,n,m,− (tin (Sp ),n ) + (σ̂n (tin (Sp ),n , +))2 ξˆk,n,m,+ (tin (Sp ),n ) × Un,in (Sp ),m , Rk (P ) = P X 1/2 (∆XSp )3 (σSp − )2 ξk,− (Sp ) + (σSp )2 ξk,+ (Sp ) USp . p=1 Step 1. From a result similar to Lemma A.6 in Martin and Vetter (2017), which follows from Condition 2.10(i), and Lemma 5.1 we derive e P Mn   1 X e k (P ) ≤ Υ|X ) > ε ∩ Ω(J) → 0 1{Rk (P,n,m)≤Υ} − P(R T Mn (6.23) m=1 as n → ∞ for any fixed P . Step 2. Next we show e lim lim sup P P →∞ n→∞ Mn  1 X (J)  (1{Rk (P,n,m)≤Υ} − 1{Fb(J) ≤Υ} ) > ε ∩ ΩT = 0 Mn k,T,n,m m=1 (6.24) for all ε > 0 which follows from Mn 1 X e k (P, n, m) − Fb(J) P(|R lim lim sup k,T,n,m | > ε) = 0 P →∞ n→∞ Mn (6.25) m=1 for all ε > 0; compare Step 4 in the proof of Proposition A.7 from Martin and Vetter (2017). On the set Ω(q, P, n) on which the jumps of N (q) are among the P largest jumps and two different jumps of N (q) are further apart than |πn |T it holds (J) |Rk (P, n, m) − Fbk,T,n,m |1Ω(q,P,n) 2 X ≤ |∆i,n (X − N (q))|3 1{|∆i,n X|>β|Ii,n |$ } bn i:ti,n ≤T ≤K X j6=i:Ij,n ⊂[ti,n −bn ,ti,n +bn ] 1/2 × ξˆk,n,m,− (ti,n ) + ξˆk,n,m,+ (ti,n ) |Un,i,m | 1 X  3 3 |∆i,n (B(q) + C)| + |∆i,n M (q)| bn i:ti,n ≤T 1/2 2 (∆j,n X) X (∆j,n X) 2 1/2 j6=i:Ij,n ⊂[ti,n −bn ,ti,n +bn ] 1/2 × ξˆk,n,m,− (ti,n ) + ξˆk,n,m,+ (ti,n ) |Un,i,m |. (6.26) For the continuous parts in (6.26) we get using the Cauchy-Schwarz inequality, 8 Lemma 6.1 and ξˆk,n,m,− (ti,n ), ξˆk,n,m,+ (ti,n ) ≤ nK|πn |T h X 1 E |∆i,n (B(q) + C)|3 bn X i:ti,n ≤T (∆j,n X)2 1/2 j6=i:Ij,n ⊂[ti,n −bn ,ti,n +bn ] i 1/2 × ξˆk,n,m,− (ti,n ) + ξˆk,n,m,+ (ti,n ) |Un,i,m | S 1/2 X   6 3 2bn ≤K Kq |Ii,n | + |Ii,n | n|πn |T bn i:ti,n ≤T √ = K n|πn |T Kq (|πn |T )3 + 1 1/2 T (6.27) which vanishes as n → ∞ by Condition 2.7(i). Denote ξˆk,n,m (i) = ξˆk,n,m,− (ti,n ) + ξˆk,n,m,+ (ti,n ). We get for the remaining terms containing |∆i,n M (q)|3 in (6.26) using the Jensen inequality and inequalities (6.3), (6.5) 1 h X 3 |∆i,n M (q)| E bn i:ti,n ≤T X (∆j,n X)2 j6=i:Ij,n ⊂[ti,n −bn ,ti,n +bn ] × ξˆk,n,m,− (ti,n ) + ξˆk,n,m,+ (ti,n ) 1 X  h E E[|∆i,n M (q)|3 |S, Fti−1,n ] bn ≤K i:ti,n ≤T X i:ti,n ≤T h  h1 + E |∆i,n M (q)|3 E bn X i:ti,n ≤T × Ln  X (∆j,n X)2 1/2 S i 2 (∆j,n X) 1/2 S, Fti,n i i  q  S E 2 ξˆk,n,m (i) S X (∆j,n X)2 S i1/2 j:Ij,n ⊂[ti,n −bn ,ti,n ) X (∆j,n X)2 S, Fti,n i1/2 S i 1/2  E ξˆk,n,m (i) S j:Ij,n ⊂[ti,n ,ti,n +bn ] Ln −1  k−1 1/2   X X |Ii+ζ,n | n (k − j)2 (|Ii+l+j,n | + |Ii+l−j,n |) |Ii+l,n | i:ti,n ≤T X i:ti,n ≤T j=1 ζ=−Ln X √ ≤ Keq n |Ii,n | = Keq n X i  b 1/2 n |Ii,n | K bn l=−Ln √ |Un,i,m | S j:Ij,n ⊂[ti,n ,ti,n +bn ]  h1 X  |Ii,n | E bn ≤ Keq 1/2 j:Ij,n ⊂[ti,n −bn ,ti,n ) h h 1 + E |∆i,n M (q)|3 E bn ≤ Keq 1/2 Ln X Ln  X −1 |Ii+l,n | |Ii+ζ,n | l=−Ln |Ii,n | √ + OP ( n|πn |T ) k−1 X j=−(k−1) ζ=−Ln |Ii+j,n | 1/2 Ln X l=−Ln k−1 X |Ii+l+j,n |1/2 j=−(k−1) Ln  X −1 |Ii+l,n | |Ii+l+ζ,n | ζ=−Ln (6.28) where we changed the index i → i+l to derive the last equality. Then (6.28) vanishes 9 as in (6.13) because of Ln X |Ii+l,n | −1 |Ii+l+ζ,n | ζ=−Ln l=−Ln ≤ Ln X 0 X Ln X |Ii+l,n | Ln −1 X |Ii+l+ζ,n | + |Ii+l,n | ζ=−Ln l=−Ln l=0 Ln X −1 |Ii+l+ζ,n | ≤ 2. (6.29) ζ=−Ln Hence by Lemma 6.2 we get (6.25) because of P(Ω(q, P, n)) → ∞ as P, n → ∞ for any q > 0. Step 3. Finally we show P e (J) e k (P ) ≤ Υ|X )1 (J) −→ P(R P(Fk,T ≤ Υ|X )1Ω(J) Ω T (6.30) T (J) P as P → ∞. We have Rk (P ) −→ Fk,T as P → ∞ and it can be shown that this convergence yields (6.30); compare Step 3 in the proof of Proposition A.7 in Martin and Vetter (2017). Combining (6.23), (6.24) and (6.30) yields (5.12). 6.2 Proofs for Section 3 Proof of (5.17). Step 1. We first prove √ lim P(| n (V (f, [2], πn )T − 4V (f, πn )T ) − R(n)| > ε) = 0 (6.31) n→∞ for all ε > 0 where R(n) = R(1) (n) + R(2) (n) and X √ (l) (l) R(l) (n) = n 2∆i−1,1,n X (l) ∆i,1,n X (l) (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T 2 (3−l) × ∆j,2,n X (3−l) 1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} , l = 1, 2. Therefore note that it holds √ n (V (f, [2], πn )T − 4V (f, πn )T ) h X 2 (2) 2 √ (1) (1) (2) = n ∆i−1,1,n X (1) + ∆i,1,n X (1) ∆j−1,1,n X (2) + ∆j,1,n X (2) (1) (2) i,j≥2:ti,n ∧tj,n ≤T − (1) X ∆i−l1 ,1,n X (1) 2 (2) i−l1 ,1,n ∩Ij−l2 ,1,n 6=∅} l1 ,l2 =0,1 √ + OP ( n|πn |T ) X √ = n (1) h (1) i 2 (2) ∆j−l2 ,1,n X (2) 1{I (1) (1) (2) 1{I (1) (2) i,2,n ∩Ij,2,n 6=∅} (2) 2∆i−1,1,n X (1) ∆i,1,n X (1) ∆j−1,1,n X (2) + ∆j,1,n X (2) 2 (2) i,j≥2:ti,n ∧tj,n ≤T 2 (2) (1) (1) (2) + 2 ∆i−1,1,n X (1) + ∆i,1,n X (1) ∆j−1,1,n X (2) ∆j,1,n X (2) X 2 (2) 2 (1) + ∆i−l1 ,1,n X (1) ∆j−l2 ,1,n X (2) 1{I (1) (2) ∩I i−l1 ,1,n l1 ,l2 =0,1 √ + OP ( n|πn |T ), j−l2 ,1,n i 1{I (1) =∅} (2) i,2,n ∩Ij,2,n 6=∅} (6.32) 10 √ where the OP ( n|πn |T )-terms are due to boundary effects. Because of the indicator 1{I (1) we may use iterated expectations and inequality (6.5) to bound (2) ∩I =∅} i−l1 ,1,n j−l2 ,1,n (1) (2) the S-conditional expectation of the sum over the (∆i−l1 ,1,n X (1) )2 (∆j−l2 ,1,n X (2) )2 in (6.32) by √ 4K n (1) X (1) (2) |Ii,2,n ||Ij,2,n |1{I (1) (2) i,2,n ∩Ij,2,n 6=∅} (2) = 32KH2,n (T ) √ n i,j≥2:ti,n ∧tj,n ≤T which converges to zero in probability as n → ∞ because of Condition 3.4(ii). Hence we obtain (6.31) by Lemma 6.2. Step 2. Next we will show lim lim sup P(|R(l) (n) − R(l) (n, q)| > ε) → 0 (6.33) q→∞ n→∞ for all ε > 0 and l = 1, 2 with R(l) (n, q) = √ X n 2 (l) (l) (3−l) 2∆i−1,1,n X (l) ∆i,1,n X (l) ∆j,2,n N (3−l) (q) 1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T (3−l,l) (l) , (3−l) l = 1, 2. Denote by ∆(j,i),(k2 ,k1 ),n X the increment of X over the interval Ii,k1 ,n ∩Ij,k2 ,n (3−l\l) (3−l) (l) and by ∆(j,i),(k1 ,k2 ),n X the increment of X over Ij,k1 ,n \ Ii,k2 ,n (which might be the sum of the increments over two separate intervals). Using the elementary inequality |a2 − b2 | = (ρ|a + b|)(ρ−1 |a − b|) ≤ ρ2 (a + b)2 + ρ−2 (a − b)2 (2) (2) which holds for any a, b ∈ R, ρ > 0 for a = ∆j,2,n X (2) and b = ∆j,2,n N (2) (q) yields E[|R(l) (n) − R(l) (n, q)||S]  X  (l) 2  √ (l) (3−l) ≤2 n ρ2 E |∆i−1,1,n X (l) ∆i,1,n X (l) | ∆j,2,n (X (3−l) + N (3−l) (q)) S (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T  (l) 2  (l) (3−l) + ρ−2 E |∆i−1,1,n X (l) ∆i,1,n X (l) | ∆j,2,n (X (3−l) − N (3−l) (q)) S 1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} . (6.34) Using (l) (3−l,l) (l\(3−l)) |∆ι,1,n X (l) | ≤ |∆(j,ι),(2,1),n X (l) | + |∆(ι,j),(1,2),n X (l) |, ι = i − 1, i, 2 2 2 (3−l) 2 (3−l,l) (3−l,l) (3−l\l) ∆j,2,n Y ≤ 3 ∆(j,i),(2,1),n Y + 3 ∆(j,i−1),(2,1),n Y + 3 ∆(j,i),(2,2),n Y , for Y = X (3−l) +N (3−l) (q) to treat increments over overlapping and non-overlapping 11 intervals differently we obtain that the first summand in (6.34) is bounded by 6ρ 2 √   (l) 2  (l) (3−l\l) E |∆i−1,1,n X (l) ∆i,1,n X (l) | ∆(j,i),(2,2),n (X (3−l) + N (3−l) (q)) S X n (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T X   (l) 2  (3−l,l) (3−l,l) + E |∆2i−1−ι,1,n X (l) ∆(j,ι),(2,1),n X (l) | ∆(j,ι),(2,1),n (X (3−l) + N (3−l) (q)) S ι=i−1,i  (l) 2  (l\3−l) (3−l,l) + E |∆2i−1−ι,1,n X (l) ∆(ι,j),(1,2),n X (l) | ∆(j,ι),(2,1),n (X (3−l) + N (3−l) (q)) S  × 1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} √ ≤ Kρ2 n  (l) (l) (3−l) (3−l) (|Ii−1,1,n ||Ii,1,n |)1/2 (|Ij,2,n | + Kq |Ij,2,n |2 ) X (3−l) (l) i,j≥2:ti,n ∧tj,n + X  ≤T (l) (l) |I2i−1−ι,1,n |1/2 |Iι,1,n (3−l) (l) (3−l) (l) (3−l) ∩ Ij,2,n |1/2 (|Iι,1,n ∩ Ij,2,n | + Kq |Iι,1,n ∩ Ij,2,n |4 )1/2 ι=i−1,i + (l) (l) (l) (|Ii−1,1,n ||Ii,1,n |)1/2 (|Iι,1,n ∩ (3−l) Ij,2,n | + (l) Kq |Iι,1,n ∩  (3−l) Ij,2,n |2 )  1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} √ ≤ Kρ2 (1 + Kq |πn |T )H2,n (T )/ n √ + Kρ2 (1 + Kq (|πn |T )3 )1/2 n (l) X (l) (3−l) |Ii,2,n |1/2 |Ii,2,n ∩ Ij,2,n |1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T (6.35) where we used iterated expectations, the Cauchy-Schwarz inequality for the second summand and Lemma 6.1 repeatedly. The first term in the last bound vanishes as n → ∞ because of Condition 3.2 and 3.4(ii) while the second term vanishes after letting n → ∞, q → ∞ and then ρ → 0 by Condition 3.4(i) and (ii) because of X √ (l) (l) (3−l) n |Ii,2,n |1/2 |Ii,2,n ∩ Ij,2,n |1{I (l) ∩I (3−l) 6=∅} (l) (3−l) i,j≥2:ti,n ∧tj,n ≤ √ n i,2,n j,2,n ≤T X (l) (l) X |Ii,2,n |1/2 |Ii,2,n | ≤ nT (l) (l) |Ii,2,n |2 1/2 (l) i≥2:ti,n ≤T √ ≤ (23 T H2,n (T ))1/2 + OP ( n|πn |T ) i≥2:ti,n ≤T (6.36) where the inequality in the second to last line follows from the Cauchy-Schwarz inequality for sums. Analogously we can bound the second summand in (6.34) by √ Kρ−2 (Kq |πn |T + 1 + eq )H2,n (T )/ n X 1/2 √ (l) + Kρ−2 (Kq (|πn |T )3 + |πn |T + eq )1/2 nT |Ii,2,n |2 + OP ( n|πn |T ) (l) i≥2:ti,n ≤T where the first term vanishes as n → ∞ and the second term vanishes as n → ∞ and then q → ∞ for any ρ > 0. Hence using Lemma 6.2 we have proved (6.33). Step 3. Further we will show e(l) (n, q)| > ε) → 0 lim lim sup P(|R(l) (n, q) − R q→∞ n→∞ 12 (6.37) for all ε > 0 and l = 1, 2 with e(l) (n, q) = R √ n (l) X (l) (l) 2 ∆i−1,1,n N (l) (q)∆i,1,n C (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T  (3−l) 2 (l) (l) + ∆i−1,1,n C (l) ∆i,1,n N (l) (q) ∆j,2,n N (3−l) (q) 1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} , l = 1, 2. Using 2 (3−l,l) (3−l,l) E[∆(j,ι),(2,1),n (X (l) − N (l) (q)) ∆(j,ι),(2,1),n N (3−l) (q) |S] (l) (3−l) (l) (3−l) ≤ (Kq |Iι,1,n ∩ Ij,2,n |1/6 + (eq )1/2 )|Iι,1,n ∩ Ij,2,n | which can be derived as in (6.13) we get analogously to (6.35) X √ (l) (l) E n 2∆i−1,1,n (X (l) − N (l) (q))∆i,1,n (X (l) − N (l) (q)) (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T 2  (3−l) × ∆j,2,n N (3−l) (q) 1{I (l) ∩I (3−l) 6=∅} S i,2,n j,2,n h X (l) (l) ((Kq |Ii−1,1,n | + 1 + eq )|Ii−1,1,n |)1/2 √ ≤K n (l) (3−l) i,j≥2:ti,n ∧tj,n ≤T (l) (l) (3−l) (3−l) × ((Kq |Ii,1,n | + 1 + eq )|Ii,1,n |)1/2 (1 + Kq |Ij,2,n |)|Ij,2,n | + X  (l) (l) ((Kq |I2i−1−ι,1,n | + 1 + eq )|I2i−1−ι,1,n |)1/2 ι=i−1,i (l) (3−l) (l) (3−l) × (Kq |Iι,1,n ∩ Ij,2,n |1/6 + (eq )1/2 )|Iι,1,n ∩ Ij,2,n | (l) (l) (l) (l) + ((Kq |I2i−1−ι,1,n | + 1 + eq )|I2i−1−ι,1,n |)1/2 ((Kq |Iι,1,n | + 1 + eq )|Iι,1,n |)1/2 i (l) (3−l) (l) (3−l) × (1 + Kq |Iι,1,n ∩ Ij,2,n |)|Iι,1,n ∩ Ij,2,n | 1{I (l) ∩I (3−l) 6=∅} i,2,n j,2,n √ ≤ K(Kq |πn |T + 1 + eq )(1 + Kq |πn |T )H2,n (T )/ n X 1/2 (l) + (Kq |πn |T + 1 + eq )1/2 (Kq (|πn |T )1/6 + (eq )1/2 ) nT |Ii,2,n |2 (l) i≥2:ti,n ≤T (6.38) which vanishes as n → ∞ and then q → ∞. Furthermore we get also analogously to (6.35) X √ (l) (l) 2∆i−1,1,n (B (l) (q) + M (l) (q))∆i,1,n (N (l) (q)) E n (l) (3−l) i,j≥2:ti,n ∧tj,n × √ ≤K n ≤T 2 (3−l) ∆j,2,n N (3−l) (q) 1{I (l) ∩I (3−l) 6=∅} i,2,n j,2,n h X (l) (3−l) i,j≥2:ti,n ∧tj,n (l) S  (l) (l) (l) ((Kq |Ii−1,1,n | + eq )|Ii−1,1,n |)1/2 ((1 + Kq |Ii,1,n |)|Ii,1,n |)1/2 ≤T (3−l) (3−l) × (1 + Kq |Ij,2,n |)|Ij,2,n | (l) (l) (l) (3−l) (l) (3−l) + ((Kq |Ii−1,1,n | + eq )|Ii−1,1,n |)1/2 ((1 + Kq |Ii,1,n ∩ Ij,2,n |)|Ii,1,n ∩ Ij,2,n |)1/2 13 (l) (3−l) (l) (3−l) × ((1 + Kq |Ii,1,n ∩ Ij,2,n |3 )|Ii,1,n ∩ Ij,2,n |)1/2 (l) (3−l) (l) (3−l) (l) (l) + ((Kq |Ii−1,1,n ∩ Ij,2,n | + eq )|Ii−1,1,n ∩ Ij,2,n |)1/2 ((1 + Kq |Ii,1,n |)|Ii,1,n |)1/2 (l) (3−l) (l) (3−l) × ((1 + Kq |Ii−1,1,n ∩ Ij,2,n |3 )|Ii−1,1,n ∩ Ij,2,n |)1/2 (l) (l) (l) (l) + ((Kq |Ii−1,1,n | + eq )|Ii−1,1,n |)1/2 ((1 + Kq |Ii,1,n |)|Ii,1,n |)1/2 (l) (3−l) (l) (3−l) × (1 + Kq |Ii,1,n ∩ Ij,2,n |)|Ii,1,n ∩ Ij,2,n | (l) (l) (l) (l) + ((Kq |Ii−1,1,n | + eq )|Ii−1,1,n |)1/2 ((1 + Kq |Ii,1,n |)|Ii,1,n |)1/2 i (l) (3−l) (l) (3−l) × (1 + Kq |Ii−1,1,n ∩ Ij,2,n |)|Ii−1,1,n ∩ Ij,2,n | 1{I (l) ∩I (3−l) 6=∅} i,2,n j,2,n √ 1/2 3/2 ≤ K(Kq |πn |T + eq ) (1 + Kq |πn |T ) H2,n (T )/ n X 1/2 (l) |Ii,2,n |2 + (Kq |πn |T + eq )1/2 (1 + Kq |πn |T ) nT (6.39) (l) i≥2:ti,n ≤T which vanishes as n → ∞ and then q → ∞. The same obviously also holds if we switch the roles of i − 1 and i. Hence using Lemma 6.2 the estimates (6.38) and (l) (l) (6.39) yield (6.37) because ∆i−1,1,n N (q)∆i,1,n N (q) = 0 eventually for all i as there are only finitely many big jumps. Step 4. Finally it remains to show e q) − R(n, e q, r)| > ε) = 0 lim lim sup lim sup P(|R(n, q→∞ r→∞ (6.40) n→∞ e q) = R e(1) (n, q) + R e(2) (n, q). The proof for (6.40) is identical for all ε > 0 with R(n, to the proof of (6.15) because we have (l) (3−l) |∆i−ι,1,n N (l) (q)|(∆j,2,n N (3−l) (q))2 1{I (l) (3−l) i,2,n ∩Ij,2,n 6=∅} (l) (3−l) (l) = |∆i−ι,1,n N (l) (q)|(∆i−ι,1,n N (3−l) (q))2 ≤ Kk∆i−ι,1,n N (q)k3 for ι = 0, 1 and l = 1, 2 on the set Ω(n, q) where two different jump times of N (q) are further apart than 4|πn |T and because of P(Ω(n, q)) → 1 for any q > 0. Combining (6.31), (6.33), (6.37) and (6.40) then yields (5.17). Proof of (5.18). On the set Ω(n, q, r) where any jump times Sq,p in [0, T ] are further apart than 4|πn |T from each other and from all j2−r with 1 ≤ j ≤ bT 2r c we have √ X e q, r)1Ω(n,q,r) = 1Ω(n,q,r) 4 n R(n, ∆N (2) (q)Sq,p ∆N (1) (q)Sq,p Sq,p ≤T  × ∆N (2) + ∆N (1) (q)Sp (1) (1) σ̃Sq,p − (r, q)∆ (1) i (S (q)Sq,p n h q,p )−1,n (1) in (Sq,p )+1,n W (1)  (2) σ̃Sq,p − (r, q) (2) × ρ̃Sq,p − (r, q)∆ (2) in (Sq,p )−1,n (2) (1) W (1) + σ̃Sq,p (r)∆ (1) (2) + σ̃Sq,p (r, q) ρ̃Sq,p (r, q)∆ (2) in (Sq,p )+1,n q  (2) 1 − ρ̃Sq,p − (r, q)2 ∆ (2) W (2) in (Sq,p )−1,n q i (2) W (1) + 1 − (ρ̃Sq,p (r, q))2 ∆ (2) W (2) W (1) + in (Sq,p )+1,n where the factor 4 stems from the fact that any jump of N (l) (q) is observed in two (l) (l) consecutive intervals Ii−1,2,n , Ii,2,n , l = 1, 2. 14 Similarly as in the proof of Proposition A.3 in Martin and Vetter (2017) it can be shown that Condition 3.7 yields the X -stable convergence (1) (1) ∆ (1) W (1) , ∆ (1) (2) W (2) , ∆ (2) in (Sq,p )−1,n (2) in (Sq,p )+1,n (2) ∆ (2) in (Sq,p )−1,n (2) W (1) , ∆ (2) W (1) , ∆ (2) W (1) , in (Sq,p )−1,n in (Sq,p )+1,n   W (2) Sq,p ≤T in (Sq,p )+1,n q q L−s (1) (2) −→ L(Sq,p )USq,p ,− + (L(1) − L)(Sq,p )USq,p ,− , q q (1) (2) R(Sq,p )USq,p ,+ + (R(1) − R)(Sq,p )USq,p ,+ , q q (1) (3) L(Sq,p )USq,p ,− + (L(2) − L)(Sq,p )USq,p ,− , q q (1) (3) R(Sq,p )USq,p ,+ + (R(2) − R)(Sq,p )USq,p ,+ , q q   (4) (4) L(2) (Sq,p )USq,p ,− , R(2) (Sq,p )USq,p ,+ (6.41) Sq,p ≤T (i) (i) where Us,− , Us,+ for i = 1, . . . , 4 are standard normally distributed random variables which are independent of F and of the random vectors (L1 , R1 , L2 , R2 , L, R)(s). (6.41) together with Proposition 2.2 in Podolskij and Vetter (2010) and the continuous mapping theorem then yields X √ 4 n ∆N (2) (q)Sq,p ∆N (1) (q)Sq,p Sq,p ∈P (q,T )  h i (1) (1) (1) (1) × ∆N (2) (q)Sp σ̃Sq,p − (r, q)∆ (1) W (1) + σ̃Sq,p (r, q)∆ (1) W (1) in (Sq,p )−1,n in (Sq,p )+1,n h (2) (2) + ∆N (1) (q)Sq,p σ̃Sq,p − (r, q) ρ̃Sq,p − (r, q)∆ (2) W (1) in (Sq,p )−1,n q  (2) W (2) + 1 − (ρ̃Sq,p − (r, q))2 ∆ (2) in (Sq,p )−1,n q i (2) (2) (2) + σ̃Sq,p (r, q) ρ̃Sq,p ∆ (2) W (1) + 1 − ρ̃Sq,p (r, q))2 ∆ (2) W (2) in (Sq,p )+1,n L−s −→ in (Sq,p )+1,n (CoJ) FT (q, r). Because of P(Ω(n, q, r)) → 1 as n → ∞ for any q, r this yields (5.18). Proof of (5.21). From the proof of Theorem 3.8 we know that nV (f, πn )T converges (nCoJ) on ΩT stably in law to a non-negative random variable. Comparing (6.17) in the proof of Theorem 2.11 it then suffices to show √ (CoJ) nFbT,n,m 1Ω(nCoJ) = oP (1) (6.42) T uniformly in m for proving (5.21). In order to achieve this goal, observe that it holds √ (CoJ) E[ n|FbT,n,m |1Ω(nCoJ) |F] T X X ≤ Kn (l) (3−l) (l) (l) |∆i,n X (l) |(∆j,n X (3−l) )2 |σ̂ (l) (ti,n , −) + σ̂ (l) (ti,n , +)| l=1,2 i,j:t(l) ∧t(3−l) ≤T i,n j,n q (l) (l) × |Ii−1,n | + |Ii+1,n |1{|∆(l) X (l) |>β|I (l) |$ } 1{|∆(3−l) X (3−l) |>β|I (3−l) |$ } i,n i,n 15 j,n j,n × 1{I (l) ∩I (3−l) 6=∅} 1Ω(nCoJ) i,n j,n 1 X ≤ Kn ε T (l) X (3−l) (∆i,n X (l) )2 (∆j,n X (3−l) )2 1{|∆(l) X (l) |>β|I (l) |$ } i,n l=1,2 i,j:t(l) ∧t(3−l) ≤T i,n i,n j,n × 1{|∆(3−l) X (3−l) |>β|I (3−l) |$ } 1{I (l) ∩I (3−l) 6=∅} 1Ω(nCoJ) j,n j,n i,n j,n T X X (3−l) (3−l) 2 (l) (l) (l) (∆j,n X ) (σ̂ (ti,n , −) + σ̂ (l) (ti,n , +))2 + Knε (6.43) l=1,2 i,j:t(l) ∧t(3−l) ≤T i,n × (l) (|Ii−1,n | j,n (l) + |Ii+1,n |)1{I (l) ∩I (3−l) 6=∅} 1Ω(nCoJ) i,n j,n (6.44) T where we used the elementary inequality |ab| ≤ a2 /ε + εb2 which holds for any a, b ∈ R, ε > 0. First we consider (6.43). Note that by Proposition A.5 from Martin and Vetter (2017) it suffices on the subset where no common jumps are present to consider (l) (3−l) (6.43) where (∆i,n X (l) )2 (∆j,n X (3−l) )2 is replaced with (l) (l) (3−l) (3−l) ((∆i,n C (l) )2 + (∆i,n N (q)(l) )2 )((∆j,n C (3−l) )2 + (∆j,n N (q)(3−l) )2 ). (nCoJ) (l) (3−l) Because we are on ΩT the sum with (∆i,n N (q)(l) )2 (∆j,n N (q)(3−l) )2 vanishes as n → ∞. Further we obtain using the Cauchy-Schwarz inequality, inequalities (A.31) from Martin and Vetter (2017) and (6.2) X (1) (2) E[n (∆i,n C (1) )2 (∆j,n C (2) )2 1{|∆(1) X (1) |>β|I (1) |$ } (1) i,n (2) i,n i,j:ti,n ∧tj,n ≤T × 1{|∆(2) X (2) |>β|I (2) |$ } 1{I (1) ∩I (2) 6=∅} |S] j,n j,n i,n j,n X (1) (1) 8 (2) (1) (1) E[(∆i,n C ) |S]E[(∆j,n C (2) )8 |S]P(|∆i,n X (1) | > β|Ii,n |$ |S) ≤n (1) (2) i,j:ti,n ∧tj,n ≤T 1/4 (2) (2) × P(|∆j,n X (2) | > β|Ij,n |$ |S) 1{I (1) ∩I (2) 6=∅} i,n (1/2−$)/2 ≤ K(|πn |T ) j,n H1,n (T ) which vanishes as n → ∞ by Condition 3.4. Hence for showing that (6.43) vanishes as n → ∞ it remains to discuss X X (l) (3−l) Y (l) (n, q) = n (∆i,n C (l) )2 (∆j,n N (q)(3−l) )2 1{|∆(l) X (l) |>β|I (l) |$ } i,n l=1,2 i,j:t(l) ∧t(3−l) ≤T i,n i,n j,n × 1{|∆(3−l) X (3−l) |>β|I (3−l) |$ } 1{I (l) ∩I (3−l) 6=∅} 1Ω(nCoJ) , j,n j,n i,n j,n l = 1, 2. T (6.45) (l) Let Sq,p denote an enumeration of the jump times of N (q)(l) . Then Y (l) (n, q) is asymptotically equal to X X (3−l) (∆N (q) (3−l) )2 Sq,p l=1,2 S (3−l) ≤T q,p ×n X (l) i:ti,n ≤T (l) (∆i,n C (l) )2 1{|∆(l) X (l) |>β|I (l) |$ } 1{I (l) ∩I (3−l) i,n i,n 16 i,n (3−l) (3−l) in (Sq,p ),n 1 (nCoJ) 6=∅} ΩT . In the proof of Proposition A.3 in Martin and Vetter (2017) it is shown that n L−s (l) X (∆i,n C (l) )2 1{I (l) ∩I (3−l) i,n (l) (3−l) (3−l) in (Sq,p ),n i:ti,n ≤T 6=∅} (l) (3−l) −→ R1 (Sq,p ) (3−l) for any jump time Sq,p of N (q)(3−l) . Further we obtain from (A.31) in Martin and Vetter (2017) and inequality (6.2) X (l) (∆i,n C (l) )2 1{|∆(l) X (l) |>β|I (l) |$ } 1{I (l) ∩I (3−l) |S] E[n 6=∅} i,n (l) i,n i,n i:ti,n ≤T ≤n (l) X (l) (3−l) (3−l) in (Sq,p ),n (l) (E[(∆i,n C (l) )4 |S]P(|∆i,n X (l) | > β|Ii,n |$ |S))1/2 1{I (l) ∩I (3−l) i,n (l) i:ti,n ≤T (l) X ≤ K(|πn |T )1/2−$ n |Ii,n |1{I (l) ∩I (3−l) i,n (l) i:ti,n ≤T (3−l) (3−l) in (Sq,p ),n 6=∅} (3−l) (3−l) in (Sq,p ),n 6=∅} . Here the expression in the last line vanishes as n → ∞ because of X (l) E[n |Ii,n |1{I (l) ∩I (3−l) ] 6=∅} i,n (l) (3−l) in (s),n i:ti,n ≤T (l) (l) (l) in (s),n (l) = E[η1,n,− (s) + (∆ (l) W (l) 2 (l) ) + η1,n,+ (s)] (l) = E[η1,n,− (s) + (s − t (l) in (s)−1,n (l) ) + (t (l) in (s),n − s) + η1,n,+ (s)] where the right hand side converges as n → ∞ by Condition 3.4. This estimate yields X (l) lim P (|πn |T )−(1/2−$)/2 n |Ii,n |1{|∆(l) X (l) |>β|I (l) |$ } 1{I (l) ∩I (3−l) 6=∅} n→∞ i,n (l) i,n i,n i:ti,n ≤T (3−l) (3−l) in (Sq,p ),n  (l) (3−l) ≤ R1 (Sq,p ) = 1, (3−l) for any jump time Sq,p from which we conclude (CoJ) lim P((|πn |T )−(1/2−$)/2 Y (l) (n, q) ≤ D1,T n→∞ ) = 1, l = 1, 2. Hence we finally get that (6.45) and also (6.43) vanish as n → ∞. Next we consider (6.44). Using Condition 3.2 and the assumption that Γt is bounded we obtain that there exists a q 0 > 0 with X = B(q 0 ) + C + M (q 0 ). Plugging this decomposition into (6.44) and using Lemma A.4 from Martin and Vetter (2017) we obtain that the S-conditional expectation of (6.44) is bounded by Knε X (3−l) X |Ij,n | l=1,2 i,j:t(l) ∧t(3−l) ≤T i,n 1 bn j,n (l) (l) (l) ι:Iι,n ⊂[ti,n −bn ,ti,n +bn ] × 1{I (l) ∩I (3−l) 6=∅} i,n  (l) (l) (l) |Iι,n | (|Ii−1,n | + |Ii+1,n |) X j,n ≤ KεH2,n (T ) + OP (n(|πn |T )2 ). 17 This expression converges to KεH2 (T ) as n → ∞. Hence alltogether we have shown √ (CoJ) lim lim sup P(E[ n|FbT,n,m |1Ω(nCoJ) |F] > δ) = 0 ε→0 n→∞ T for any δ > 0 which by Lemma 6.2 yields (6.42). The following Lemma is needed in the proof of Lemma (5.3). The techniques used in the proof are taken from the proof of Theorem 3.1 in Hayashi and Yoshida (2005). Lemma 6.3. Let Y (1) , Y (2) be martingales which are adapted to the filtration (Ft )t≥0 . Further let (s, t] be a deterministic interval and the observation scheme be exogenous. If the inequalities (1) E[(Yt0 (1) (2) − Ys0 )2 |Fs0 ] ≤ KY (1) (t − s), E[(Yt0 (2) − Ys0 )2 |Fs0 ] ≤ KY (2) (t − s), (6.46)  (1)  (1) (2) (2) E (Yt0 − Ys0 )2 (Yt0 − Ys0 )2 Fs0 ≤ KY (1) ,Y (2) (t0 − s0 )2 , (6.47) hold, then for all s ≤ s0 ≤ t0 ≤ t we have X (1) (2) e (1) (2) (t − s)2 E[( ∆i,n Y (1) ∆j,n Y (2) 1{I (1) ∩I (2) 6=∅} )2 |σ(Fs , S)] ≤ K Y ,Y (1) i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂(s,t] e (1) (2) = 14K (1) K (2) + 3K (1) (2) . where K Y ,Y Y Y Y ,Y Proof. To shorten notation set 1i,j = 1{I (1) ∩I (2) 6=∅} . Note that it holds i,n (1) X (1) (2) ∆i,n Y (1) ∆j,n Y (2) 1i,j j,n 2 (2) i,j:Ii,n ∪Ij,n ⊂(s,t] (1) X = (1) (2) (2) (1) (2) ∆i,n Y (1) ∆j,n Y (2) ∆i0 ,n Y (1) ∆j 0 ,n Y (2) 1i,j 1i0 ,j 0 (1) (2) i,j,i0 ,j 0 :Ii,n ∪Ij,n ∪Ii0 ,n ∪Ij 0 ,n ⊂(s,t] = (1) X (2) (∆i,n Y (1) )2 (∆j,n Y (2) )2 1i,j (6.48) i=i0 ,j=j 0 + (1) X (2) (2) (∆i,n Y (1) )2 ∆j,n Y (2) ∆j 0 ,n Y (2) 1i,j 1i,j 0 (6.49) i=i0 ,j6=j 0 + X (1) (1) (1) (1) (2) ∆i,n Y (1) ∆i0 ,n Y (1) (∆j,n Y (2) )2 1i,j 1i0 ,j (6.50) i6=i0 ,j=j 0 + X (2) (2) ∆i,n Y (1) ∆i0 ,n Y (1) ∆j,n Y (2) ∆j 0 ,n Y (2) 1i,j 1i0 ,j 0 (6.51) i6=i0 ,j6=j 0 and we discuss the terms (6.48)–(6.51) separately. (1,2) For discussing (6.48) we denote by ∆(i,j),n Y (l) the increment of Y (l) , l = 1, 2, (1) (2) (l\3−l) over the interval Ii,n ∩ Ij,n and by ∆(i,{j}),n Y (l) the increment of Y (l) over the (l) (3−l) interval Ii,n \ Ij,n (which might also be the sum of increments over two distinct 18 intervals). Using this notation, iterated expectations and the inequalities (6.46), (6.47) we obtain that the σ(Fs , S)-conditional expectation of (6.48) is equal to X (1,2) (1\2) (1,2) (2\1) E[(∆(i,j),n Y (1) + ∆(i,{j}),n Y (1) )2 (∆(i,j),n Y (2) + ∆(j,{i}),n Y (2) )2 |σ(Fs , S)]1i,j i,j = X (1,2) (1\2) (2\1) E[((∆(i,j),n Y (1) )2 + (∆(i,{j}),n Y (1) )2 )(∆(j,{i}),n Y (2) )2 i,j (1\2) (1,2) + (∆(i,{j}),n Y (1) )2 (∆(i,j),n Y (2) )2 |σ(Fs , S)]1i,j + X (1,2) (1,2) E[(∆(i,j),n Y (1) )2 (∆(i,j),n Y (2) )2 |σ(Fs , S)] i,j ≤ X (1) (2) (1) (1) (2) (1) (2) KY (1) KY (2) (|Ii,n ||Ij,n \ Ii,n | + |Ii,n \ Ij,n ||Ii,n ∩ Ij,n |)1i,j i,j + X (1) (2) KY (1) ,Y (2) |Ii,n ∩ Ij,n |2 i,j (1) X = (2KY (1) KY (2) + KY (1) ,Y (2) ) (1) (2) |Ii,n ||Ij,n |1i,j (2) i,j:Ii,n ∪Ij,n ⊂(s,t] ≤ 3(2KY (1) KY (2) + KY (1) ,Y (2) )(|πn |t ∧ (t − s))(t − s). (1\2) For treating (6.49) we additionally denote by ∆(i,{j,j 0 }),n Y (1) the increment of Y (1) (1) (2) (2) over the interval Ii,n \ (Ij,n ∪ Ij 0 ,n ) (which might also be the sum of increments over up to three distinct intervals). We then obtain by using iterated expectations, the fact that Y (1) , Y (2) are martingales, the Cauchy-Schwarz inequality and (6.46) that the σ(Fs , S)-conditional expectation of (6.49) is bounded by X (1,2) (1,2) (1\2) E[(∆(i,j),n Y (1) + ∆(i,j 0 ),n Y (1) + ∆(i,{j,j 0 }),n Y (1) )2 i,j6=j 0 (1,2) (1,2) × ∆(i,j),n Y (2) ∆(i,j 0 ),n Y (2) |σ(Fs , S)] 1i,j 1i,j 0 =2 (1,2) X (1,2) (1,2) (1,2) E[∆(i,j),n Y (1) ∆(i,j),n Y (2) ∆(i,j 0 ),n Y (1) ∆(i,j 0 ),n Y (2) |σ(Fs , S)] i,j6=j 0 ≤2 X (1) (2) (1) (2) KY (1) KY (2) |Ii,n ∩ Ij,n ||Ii,n ∩ Ij 0 ,n | i,j6=j 0 ≤ KY (1) KY (2) (1) X |Ii,n |2 ≤ 6KY (1) KY (2) (|πn |t ∧ (t − s))(t − s). (1) i:Ii,n ⊂(s,t] By symmetry (6.50) can be treated similarly as (6.49). Hence it remains to discuss (6.51). Using the notation introduced above we obtain that the σ(Fs , S)-conditional expectation of (6.50) is equal to X (1,2) (1,2) (1\2) E[(∆(i,j),n Y (1) + ∆(i,j 0 ),n Y (1) + ∆(i,{j,j 0 }),n Y (1) ) i6=i0 ,j6=j 0 (1,2) (1,2) (1\2) × (∆(i0 ,j),n Y (1) + ∆(i0 ,j 0 ),n Y (1) + ∆(i0 ,{j,j 0 }),n Y (1) ) (1,2) (1,2) (2\1) × (∆(i,j),n Y (2) + ∆(i0 ,j),n Y (2) + ∆(j,{i,i0 }),n Y (2) ) (1,2) (1,2) (2\1) × (∆(i,j 0 ),n Y (2) + ∆(i0 ,j 0 ),n Y (2) + ∆(j 0 ,{i,i0 }),n Y (2) )|σ(Fs , S)]1i,j 1i0 ,j 0 19 = (1,2) X (1,2) (1,2) (1,2) E[(∆(i,j),n Y (1) + ∆(i,j 0 ),n Y (1) )(∆(i0 ,j),n Y (1) + ∆(i0 ,j 0 ),n Y (1) ) i6=i0 ,j6=j 0 (1,2) (1,2) (1,2) (1,2) × (∆(i,j),n Y (2) + ∆(i0 ,j),n Y (2) )(∆(i,j 0 ),n Y (2) + ∆(i0 ,j 0 ),n Y (2) )|σ(Fs , S)]1i,j 1i0 ,j 0 . (1\2) Here terms including factors like ∆(i,{j,j 0 }),n Y (1) vanish using iterated expectations (1) (2) (2) because Y (1) is a martingale and Ii,n \ (Ij,n ∪ Ij 0 ,n ) overlaps with none of the other intervals for i 6= i0 . Expanding the product above and observing that the expectation vanishes for all products where one of i, i0 , j, j 0 occurs exactly once yields that the σ(Fs , S)-conditional expectation of (6.51) is equal to X (1,2) (1,2) (1,2) (1,2) E[∆(i,j),n Y (1) ∆(i0 ,j 0 ),n Y (1) ∆(i,j),n Y (2) ∆(i0 ,j 0 ),n Y (2) i6=i0 ,j6=j 0 (1,2) (1,2) (1,2) (1,2) + ∆(i,j),n Y (1) ∆(i0 ,j 0 ),n Y (1) ∆(i0 ,j),n Y (2) ∆(i,j 0 ),n Y (2) (1,2) (1,2) (1,2) (1,2) (1,2) (1,2) (1,2) (1,2) (1,2) (1,2) (1,2) + ∆(i,j 0 ),n Y (1) ∆(i0 ,j),n Y (1) ∆(i,j),n Y (2) ∆(i0 ,j 0 ),n Y (2) + ∆(i,j 0 ),n Y (1) ∆(i0 ,j),n Y (1) ∆(i0 ,j),n Y (2) ∆(i,j 0 ),n Y (2) |σ(Fs , S)]1i,j 1i0 ,j 0 X (1,2) (1,2) (1,2) (1,2) E[∆(i,j),n Y (1) ∆(i0 ,j 0 ),n Y (1) ∆(i,j),n Y (2) ∆(i0 ,j 0 ),n Y (2) = i6=i0 ,j6=j 0 (1,2) + ∆(i,j 0 ),n Y (1) ∆(i0 ,j),n Y (1) ∆(i0 ,j),n Y (2) ∆(i,j 0 ),n Y (2) |σ(Fs , S)]1i,j 1i0 ,j 0 (1) (2) (1) (2) where the second identity holds as one of the intervals Ii,n ∩ Ij,n , Ii,n ∩ Ij 0 ,n , (1) (2) (1) (2) Ii0 ,n ∩ Ij,n , Ii0 ,n ∩ Ij 0 ,n always has to be empty for i = i0 , j 6= j 0 . We then further obtain using iterated expectations, the conditional Cauchy-Schwarz inequality and (6.47) that the σ(Fs , S)-conditional expectation of (6.51) is bounded by X (1) (2) (1) (2) (1) (2) (1) (2) KY (1) KY (2) (|Ii,n ∩ Ij,n ||Ii0 ,n ∩ Ij 0 ,n | + |Ii,n ∩ Ij 0 ,n ||Ii0 ,n ∩ Ij,n |)1i,j 1i0 ,j 0 i6=i0 ,j6=j 0 ≤ 2KY (1) KY (2) X (1) i:Ii,n ⊂(s,t] (1) |Ii,n | X (1) |Ii0 ,n | ≤ 2KY (1) KY (2) (t − s)2 . (1) i0 :Ii0 ,n ⊂(s,t] We conclude the proof by combining the bounds for (6.48)–(6.51) and using Lemma 6.2. Proof of Lemma 5.3. First note that (5.22) follows from Lemma 5.1. Further (5.23) follows from (5.22) and P (1) P (1) (2) (2) κ̂n (Sp , −)1{Sp <T } −→ ρSp − σSp − σSp − 1{Sp <T } , (6.52) κ̂n (Sp , +)1{Sp <T } −→ ρSp σSp σSp 1{Sp <T } (1) (2) because of σs σs 6= 0 almost surely for almost all s ∈ [0, T ] by Condition 3.7. Hence it remains to prove (6.52). We will only give the proof for κ̂n (s, +) as the proof for κ̂n (s, −) is identical. Further we will give a formal proof only for constant stopping times s ∈ (0, T ) and assume without loss of generality s+bn ≤ T as we did for Lemma 5.1. The extension to jump times Sp then works analogous as for the estimators for σSp − , σSp ; compare the proof of Theorem 9.3.2 in Jacod and Protter (2012). 20 We first prove κ̂(l) (s, +) − 1 bn (1) X (2) ∆i,n C (1) ∆j,n C (2) 1{I (1) ∩I (2) 6=∅} −→ 0. P i,n (1) (2) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] (6.53) j,n Therefore consider κ̂(l) (s, +) − = 1 bn 1 bn (l) X (1) i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] (1) X (1) (l) ∆i,n C (1) ∆j,n C (2) 1{I (1) ∩I (2) 6=∅} (2) (1) (2) ∆i,n B (1) (q)∆j,n X (2) + ∆i,n X (1) ∆j,n B (2) (q)  (2) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ]  (1) (2) − ∆i,n B (1) (q)∆j,n B (2) (q) 1{I (1) ∩I (2) 6=∅} i,n + 1 bn (1) X (1) (6.54) j,n (2) ∆i,n N (1) (q)∆j,n (X (2) − B (2) (q)) (2) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ]  (1) (2) + ∆i,n (C (1) + M (1) (q))∆j,n N (2) (q) 1{I (1) ∩I (2) 6=∅} i,n + X 1 bn l=1,2 + 1 bn (l) X (3−l) ∆i,n M (l) (q)∆j,n C (3−l) 1{I (l) ∩I (3−l) 6=∅} i,n (l) (3−l) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] (1) X (2) i,n (6.56) j,n ∆i,n M (1) (q)∆j,n M (2) (q)1{I (1) ∩I (2) 6=∅} (1) (2) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] (6.55) j,n (6.57) j,n where all terms (6.54)–(6.57) vanish for first n → ∞ and then q → ∞ as we will show in the following. The S-conditional expectation of the absolute value of (6.54) is using (6.1) and (6.5) bounded by X 1 bn l=1,2 (l) X Kq |Ii,n | (l) i:Ii,n ⊂[s,s+bn ]  (3−l) × E (X (3−l) (3−l) t t (3−l) (l) (3−l) in (t )∧(in (s+bn )−1),n i,n + 1 bn (1) X (1) )2 S − X (3−l) 1/2 (3−l) (l) (3−l) in (t )−1)∨in (s),n i−1,n (2) Kq |Ii,n ||Ij,n |1{I (1) ∩I (2) 6=∅} i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] ≤ (3|πn |T )1/2 X 1 bn l=1,2 X (l) Kq |Ii,n | + Kq |πn |T (l) i:Ii,n ⊂[s,s+bn ] ≤ Kq (|πn |T )1/2 which vanishes as n → ∞. (6.55) vanishes as n → ∞ because of lim P(∃t ∈ (s, s + bn ] : ∆N (q)t 6= 0) = 0. n→∞ Using (6.2), (6.3) and (l) (3−l) 2 (3−l) 2 ) |Fs ] ≤ Keq t, E[(M (q)s+t − M (q)(l) s ) (Cs+t − Cs 21 s, t ≥ 0, l = 1, 2, which is obtained from Lemma A.4 in Martin and Vetter (2017) we may apply Lemma 6.3 for (6.56) which yields E[(6.56)2 |S] ≤ Keq . The right hand side vanishes as q → ∞. Next consider the following decomposition of (6.57) using the notation from the proof of Lemma 6.3 (6.57) = 1 bn (1) l=1,2 1 bn + (1,2) ∆(i,j),n M (1) (q)∆(i,j),n M (2) (q) (2) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] X 1 bn + (1,2) X (l\3−l) X (l) (3−l) i,j:Ii,n ∪Ij,n i,n j,n ⊂[s,s+bn ] (1\2) X (1) (l,3−l) ∆(i,{j}),n M (l) (q)∆(i,j),n M (3−l) (q)1{I (l) ∩I (3−l) 6=∅} (2\1) ∆(i,{j}),n M (1) (q)∆(j,{i}),n M (2) (q)1{I (1) ∩I (2) 6=∅} i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] where the S-conditional expectation of the first sum can be bounded by eq using (1,2) (1,2) (1,2) (6.3) because of |∆(i,j),n M (1) (q)∆(i,j),n M (2) (q)| ≤ Kk∆(i,j),n M (q)k2 . Further the expectation of the square of the second and third sum can be bounded as in the proof of Lemma 6.3 by (eq )2 because the condition (6.47) is only needed for the treatment (1,2) (1,2) of the sum involving terms of the form (∆(i,j),n M (1) (q))2 (∆(i,j),n M (2) (q))2 which we here treat separately. Hence we have proven (6.53). (1) (2) P Then (6.53) yields that for proving κ̂n (s, +) −→ ρs σs σs 1 bn (1) X (2) it suffices to verify ∆i,n C (1) ∆j,n C (2) 1{I (1) ∩I (2) 6=∅} −→ ρs σs(1) σs(2) . P i,n (1) (2) i,j:Ii,n ∪Ij,n ⊂[s,s+bn ] (6.58) j,n For proving (6.58) define ζin = 1 bn (1) X (1) (2) ∆j,n C (1) ∆k,n C (2) 1{I (1) ∩I (2) 6=∅} − j,n (2) k,n j,k:Ij,n ∪Ik,n ⊂(s+(i−1)bn /rn ,s+ibn /rn ] 1 ρs σs(1) σs(2) rn and Gin = σ(Fs+ibn /rn , S) for some deterministic sequence (rn )n∈N ⊂ N with rn → ∞ P P and rn |πn |T /bn −→ 0. Such a sequence (rn )n∈N always exists because of |πn |T /bn −→ 0. Denote (1) (2) L(n, i) = {(j, k) : Ij,n ∪ Ik,n ⊂ (s + (i − 1)bn /rn , s + ibn /rn ]}. We then obtain using iterated expectations, the fact that C (1) , C (2) are martingales and a form of the conditional Itô isometry for two different integrals rn X n E[ζin |Gi−1 ] i=1 = = rn 1 X bn 1 bn X (1,2) i=1 (j,k)∈L(n,i) Z s+ibn /rn rn X E[ i=1 (1,2) n E[∆(j,k),n C (1) ∆(j,k),n C (2) |Gi−1 ] − ρs σs(1) σs(2) s+(i−1)bn /rn ) n (ρu σu(1) σu(2) − ρs σs(1) σs(2) )du||Gi−1 ] + OP (rn |πn |T /bn ). 22 As argued for (6.20) this term vanishes in probability as n → ∞ because ρ, σ (1) , σ (2) are right-continuous and bounded and because of the condition on rn . Note that we obtain from the Cauchy-Schwarz inequality and inequality (6.2) that (1) (2) − Cs(1) )2 (Ct E[(Ct (1) ≤ (E[(Ct − Cs(2) )2 |σ(Fs , S)] (2) − Cs(1) )4 |σ(Fs , S)]E[(Ct − Cs(2) )4 |σ(Fs , S)])1/2 ≤ K(t − s)2 holds for any t ≥ s ≥ 0. Hence we obtain using (a − b)2 ≤ 2a2 + 2b2 and Lemma 6.3 rn X n E[(ζin )2 |Gi−1 ] i=1 ≤ rn  2 X E 2 bn i=1 ≤ (l) X (l) ∆j,n C (1) ∆k,n C (2) 1{I (1) ∩I (2) 6=∅} j,n (j,k)∈L(n,i)  K n Gi−1 + rn 2 rn 2 k,n rn 2 X K b2n K = K + 2 2 bn rn rn rn i=1 which vanishes as rn → 0 for n → ∞. Lemma 2.2.12 from Jacod and Protter (2012) then yields rn X i=1 ζin = 1 bn (1) X (1) (2) ∆i,n C (1) ∆i,n C (2) 1{I (1) ∩I (2) 6=∅} − ρs σs(1) σs(2) i,n (2) j,n i,j:Ii,n ∪Ij,n ⊂(s,s+bn ] P + OP (rn |πn |T /bn ) −→ 0 (1) (2) P which is equivalent to (6.53) and therefore implies κ̂n (s, +) −→ ρs σs σs . Proof of Proposition 5.4. Let Sp , p = 1, . . . , P , denote the jump times of the P biggest common jumps of X on [0, T ]. Similarly as in the proof of Proposition 5.2 we define R(P, n, m) = 4 P X (1) (2) ∆ip ,n X (1) ∆jp ,n X (2) 1{|∆(1) ip ,n X p=1 × h (2) ∆jp ,n X (2)  σ̂ (1) (1) |>β|I (1) |$ } ip ,n 1{|∆(2) jp ,n X (2) |>β|I (2) |$ } jp ,n q (1,2),− Lbn,m (τip ,jp ,n )Un,(ip ,jp ),m q (1) b n,m (τip ,jp ,n )U (1,2),+ + σ̂ (1) (tip ,n , +) R n,(ip ,jp ),m (1) (tip ,n , −)  (1) b + (σ̂ (1) (tip ,n , −))2 (Lb(1) n,m − Ln,m )(τip ,jp ,n ) 1/2  (1) (1) b (1) − R b n,m )(τip ,jp ,n ) + (σ̂ (1) (tip ,n , +))2 (R U n,m n,ip ,m q  (1) (2) (1,2),− + ∆ip ,n X (1) σ̂ (2) (tjp ,n , −)ρ̂(τip ,jp ,n , −) Lbn,m (τip ,jp ,n )Un,(ip ,jp ),m q (2) b n,m (τip ,jp ,n )U (1,2),+ + σ̂ (2) (tjp ,n , +)ρ̂(τip ,jp ,n , +) R n,(ip ,jp ),m  (2) + (σ̂ (2) (tjp ,n , −))2 (1 − (ρ̂(τip ,jp ,n , −))2 )Lbn,m (τip ,jp ,n ) 1/2 (2) (2) b n,m (τip ,jp ,n ) + (σ̂ (2) (tjp ,n , +))2 (1 − (ρ̂(τip ,jp ,n , +))2 )R Un,jp ,m 23  (2) b + (σ̂ (2) (tjp ,n , −))2 (Lb(2) n,m − Ln,m )(τip ,jp ,n ) (2) b (2) − R b n,m )(τip ,jp ,n ) + (σ̂ (2) (tjp ,n , +))2 R n,m (1) 1/2 (3) Un,jp ,m i (2) where (ip , jp ) = (in (Sp ), in (Sp )) and further set R(P ) = 4 P X (1) (2) ∆XSp ∆XSp p=1 q q  (2) (1) (1),− (1) (1),+ × ∆XSp σSp − L(Sp )USp + σSp R(Sp )USp q  (2) (1) (1) + (σSp − )2 (L(1) − L)(Sp ) + (σSp )2 (R(1) − R)(Sp )USp q q  (1) (2) (2) (1),− (1),+ + ∆XSp σSp − ρSp − L(Sp )USp + σSp ρSp R(Sp )USp q (2) (2) (3) + (σSp − )2 (1 − (ρSp − )2 )L(Sp ) + (σSp )2 (1 − (ρSp )2 )R(Sp )USp q i (2) (2) (4) + (σSp − )2 (L(2) − L)(Sp ) + (σSp )2 (R(2) − R)(Sp )USp . h Using this notation we obtain e P Mn  1 X (CoJ)  e 1{R(P,n,m)≤Υ} − P(R(P ) ≤ Υ|X ) > ε ∩ ΩT →0 Mn (6.59) m=1 and P e (CoJ) e P(R(P ) ≤ Υ|X )1Ω(CoJ) −→ P(F ≤ Υ|X )1Ω(CoJ) 2,T T T as in Step 1 and Step 3 in the proof of Proposition 5.2. Here, Lemma 5.3 is needed in the proof of (6.59). Hence it remains to prove the equivalent to (6.25) in Step 2 of the proof of Proposition 5.2 which is e lim lim sup P P →∞ n→∞ Mn  1 X (CoJ)  (1{R(P,n,m)≤Υ} − 1{Fb(CoJ) ≤Υ} ) > ε ∩ ΩT = 0, Mn T,n,m m=1 implied by lim lim sup P →∞ n→∞ Mn 1 X (CoJ) e P(|R(P, n, m) − FbT,n,m | > ε) = 0 Mn (6.60) m=1 for all ε > 0. On the set Ω(q, P, n) on which the common jumps of N (q) are among the P largest common jumps and on which two different jumps of N (q) are further apart 24 than 2|πn |T it holds   (CoJ) E |R(P, n, m) − FbT,n,m |1Ω(q,P,n) F X X (l) (3−l) |∆i,n (X (l) − N (l) (q))||∆j,n X (3−l) | ≤K l=1,2 i,j:t(l) ∧t(3−l) ≤T i,n j,n q (l)  (3−l) (l) b (l) × |∆j,n X (3−l) |σ̂ (l) (n, i)E (Lbn,m + R n,m )(ti,n ) S q (3−l)  (l) (3−l) (l) (l) (3−l) b (3−l) + |∆i,n (X − N (q))|σ̂ (n, j)E (Lbn,m + R n,m )(tj,n ) S  × 1{|∆(l) X (l) |>β|I (l) |$ ∧|∆(3−l) X (3−l) |>β|I (3−l) |$ } 1{I (l) ∩I (3−l) 6=∅} i,n i,n j,n j,n i,n (6.61) j,n with (l) (l) σ̂ (l) (n, i) = (σ̂n(l) (ti,n , −))2 + (σ̂n(l) (ti,n , +))2 1/2 . Because of P(Ω(q, P, n)) → ∞ as n, P → ∞ for all q > 0 it suffices to show that (6.61) vanishes as first n → ∞ and then q → ∞ for proving (6.60). Reconsidering the notation introduced in Step 2 in the proof of (5.17) we define the following terms (l) Y(i,j),n = (l) Ye(i,j),n = 1 bn 1 bn (l,3−l) X (∆(k,j),(1,1),n X (l) )2 1/2 , (l) (l) (l) k6=i:Ik,n ⊂[ti,n −bn ,ti,n +bn ] (l\3−l) X (∆(k,j),(1,1),n , X (l) )2 1/2 , (l) (l) (l) k6=i:Ik,n ⊂[ti,n −bn ,ti,n +bn ] l = 1, 2. Then the Minkowski inequality yields (l) (l) σ̂ (l) (n, i) ≤ Y(i,j),n + Ye(i,j),n , (3−l) (3−l) σ̂ (3−l) (n, j) ≤ Y(j,i),n + Ye(j,i),n (6.62) which allows to treat the increments in the estimation of σ (l) , σ (3−l) over intervals (3−l) (l) which do overlap with Ij,n , Ii,n and those which do not, separately. We apply the inequalities (6.62) to get an upper bound for the S-conditional expectation of (6.61). First we get using the Cauchy-Schwarz inequality, iterated expectations and Lemma 6.1 the following bound for (6.61) where we replaced the estimators for σ (l) , σ (3−l) with the first summands from (6.62) X  (l) (3−l) (l) E |∆i,n (X (l) − N (l) (q))||∆j,n X (3−l) |2 Y(i,j),n K (l) (3−l) i,j:ti,n ∧tj,n ≤T p (3−l) (l) (3−l)  + |∆j,n X (3−l) ||∆i,n (X (l) − N (l) (q))|2 Y(j,i),n S 2n|πn |T 1{I (l) ∩I (3−l) 6=∅} j,n i,n X  1/2 (l) 2 (l) (l) (3−l) (3−l) 1/2 ≤K (Kq |Ii,n | + |Ii,n | + eq |Ii,n |)(|Ij,n |/bn ) K|Ij,n | (l) (3−l) i,j:ti,n ∧tj,n ≤T 1/2 (3−l) (l) (l) (l) (l) 1/2  + (K|Ij,n |)(|Ii,n |/bn ) Kq |Ii,n |4 + |Ii,n |2 + eq |Ii,n | p × n|πn |T 1{I (l) ∩I (3−l) 6=∅} i,n j,n 25 ≤ K(Kq (|πn |T + (|πn |T )3 ) + 1 + eq )1/2 (|πn |T /bn )1/2 X  √ (l) (3−l) (l) (3−l) × n |Ii,n |1/2 |Ij,n | + |Ii,n ||Ij,n |1/2 1{I (l) ∩I (3−l) 6=∅} (6.63) (l) (3−l) i,j:ti,n ∧tj,n i,n j,n ≤T for l = 1, 2 which vanishes as n → ∞ for any q > 0 because of Condition 3.11(ii) P and |πn |T /bn −→ 0. (l) (3−l) Further we obtain by treating increments over Ii,n ∩ Ij,n and increments over (l) (3−l) non-overlapping parts of Ii,n , Ij,n differently the following bound for the S-conditional expectation of (6.61) where we replaced the estimators for σ (l) , σ (3−l) with the second summands from (6.62) X  (l\3−l) (3−l) (3−l) 2 e (l) K E[|∆ (X (l) − N (l) (q))||∆ X | Y j,n (i,j),(1,1),n (l) (3−l) i,j:ti,n ∧tj,n (i,j),n ≤T (l,3−l) (3−l\l) (l) + |∆(i,j),(1,1),n (X (l) − N (l) (q))||∆(j,i),(1,1),n X (3−l) |2 Ye(j,i),n |S] (3−l\l) (l) (3−l) + E[|∆(j,i),(1,1),n X (3−l) ||∆i,n (X (l) − N (l) (q))|2 Ye(i,j),n p (3−l,l) (l\3−l) (3−l) + |∆(j,i),(1,1),n X (3−l) ||∆(i,j),(1,1),n (X (l) − N (l) (q))|2 Ye(j,i),n |S] n|πn |T (l,3−l) (l,3−l) (l) + E[|∆(i,j),(1,1),n (X (l) − N (q)(l) )||∆(i,j),(1,1),n X (3−l) |2 Ye(i,j),n |S] q (l) (l) (3−l) b (l) × E[ (Lbn,m + R n,m )(ti,n ∧ tj,n )|S]  (3−l) (l,3−l) (l,3−l) + E[|∆(i,j),(1,1),n X (3−l) ||∆(i,j),(1,1),n (X (l) − N (l) (q)|2 Ye(j,i),n |S] q  (3−l) (l) (3−l) b (3−l) × E[ (Lbn,m + R n,m )(ti,n ∧ tj,n )|S] 1{I (l) ∩I (3−l) 6=∅} i,n j,n X  (l) (l) 1/2 (3−l) ≤K (Kq |Ii,n | + 1 + eq )|Ii,n |) |Ij,n | (l) (3−l) i,j:ti,n ∧tj,n ≤T (3−l) (Kq |Ii,n | p (3−l) (l) + + 1 + eq )|Ii,n |)|Ij,n |1/2 n|πn |T 1{I (1) ∩I (2) 6=∅} i,n j,n  (l) (3−l) 2 (l) (3−l) 1/2 (l) (3−l) 1/3 (l) (3−l) + ((Kq |Ii,n ∩ Ij,n | + |Ii,n ∩ Ij,n | + eq )|Ii,n ∩ Ij,n |) |Ii,n ∩ Ij,n |2/3 q (l) (l) (3−l) b (l) × E[ (Lbn,m + R n,m )(ti,n ∧ tj,n )|S] (l) (3−l) (l) (3−l) (l) (3−l) (l) (3−l) + |Ii,n ∩ Ij,n |1/2 ((Kq |Ii,n ∩ Ij,n |3 + |Ii,n ∩ Ij,n | + eq )|Ii,n ∩ Ij,n |)1/2 q  (3−l) (l) (3−l) b (3−l) × E[ (Lbn,m + R (6.64) n,m )(ti,n ∧ tj,n )|S] . Here, we used iterated expectations and (l) E[Ye(i,j),n |S] ≤ 1 bn (l) X (l) (l) K|Ik,n | 1/2 ≤K (l) k6=i:Ik,n ⊂[ti,n −bn ,ti,n +bn ] (3−l) along with the similar bound for E[Ye(j,i),n |S]. The sum over the expression in the first set of square brackets in (6.64) vanishes similarly as in (6.63) while the sum over the expression in the second set of square 26 brackets is bounded by (l) X K(Kq (|πn |T )1/2 + eq )1/3 (l) (3−l) i,j:ti,n ∧tj,n Ln X × ≤T + (l) n|Ii+l1 +1,n |)1/2 (l) (l) ((n|Ii−1,n | Ln X + (3−l) (3−l) |Ii,n ∩ Ij,n | (3−l) (l) n|Ii+1,n |)1/2 (3−l) + (n|Ij+l2 −1,n | + n|Ij+l2 +1,n |)1/2 ) (l) X i,j:ti,n ∧tj,n × −1 (l) (3−l) |Ii+k1 ,n ∩ Ij+k2 ,n | k1 ,k2 =−Ln (l) ((n|Ii+l1 −1,n | = K(Kq (|πn |T )1/2 + eq ) × Ln X  (l) (3−l) |Ii+l1 ,n ∩ Ij+l2 ,n | l1 ,l2 =−Ln × (3−l) |Ii,n ∩ Ij,n | ≤T (3−l) (3−l) + (n|Ij−1,n | + n|Ij+1,n |)1/2 ) Ln X  (l) (3−l) |Ii+l1 ,n ∩ Ij+l2 ,n | −1 (l) (3−l) |Ii+l1 +k1 ,n ∩ Ij+l2 +k2 ,n | k1 ,k2 =−Ln l1 ,l2 =−Ln √ + OP ( n|πn |T ) (6.65) where the sum in the last line is less or equal than 4 which can be shown similarly to √ (6.29). The OP ( n|πn |T )-term is due to boundary effects. Hence (6.65) is bounded by √  K(Kq (|πn |T )1/2 + eq ) n (l) X (l) (3−l) i≥2:ti,n ≤T  √ (3−l) |Ij,2,n |3/2 + OP ( n|πn |T ) X |Ii,2,n |3/2 + j≥2:tj,n ≤T which vanishes as n, q → ∞ by Condition 3.4(ii); compare also (6.36). 6.3 Proofs regarding the Poisson setting (l) Lemma 6.4. Consider an observation scheme given by t0,n = 0 and (l) (l) (l) ti,n = ti−1,n + α(l) (n)−1 Ei,n , i ≥ 1, (l) where the Ei,n , i, n ∈ N, are i.i.d. random variables with values in R≥0 and α(l) (n) are positive functions with α(l) (n) → ∞ as n → ∞. (l) If E1,1 ∈ L2 (Ω) it holds (l) (l) −1 α (n) X (l) g α (n) (l) Ii,n  P −→ E[g(E1,1 )] (l) E[E1,1 ] (l) i:ti,n ≤T T (l) for all functions g : R≥0 → R with g(E1,1 ) ∈ L1 (Ω). Proof. This proof is based on the proof of Proposition 1 in Hayashi and Yoshida (l) (l) (l) (l) (2008). Set m1 = E[E1,1 ], mg = E[g(E1,1 )] and (l) λ(l) (n) = dα(l) (n)T /m1 e. 27 Further define (l) NT (n) = X i∈N 1{t(l) ≤T } . i,n Step 1. We prove (l) P(|NT (n) − λ(l) (n)| > α(l) (n)1/2+ε ) → 0 (6.66) as n → ∞ for ε ∈ (0, 1/2); compare Lemma 9 in Hayashi and Yoshida (2008). (l) Denote ν (l) = Var(E1,1 ). The case ν (l) = 0 is simple, so let us assume ν (l) > 0 in the following. It holds λ(l) (n)+α(l) (n)1/2+ε (l) NT (n) (l) (l) − λ (n) > α (n) 1/2+ε (l) X ⇔ |Ii,n | < T i=1 Pλ(l) (n)+α(l) (n)1/2+ε ⇔ (l) (Ei,n (l) m1 ) − i=1 p (λ(l) (n) + α(l) (n)1/2+ε )ν (l) < α(l) (n)T − (λ(l) (n) + α(l) (n)1/2+ε )m1 p (λ(l) (n) + α(l) (n)1/2+ε )ν (l) where the left hand side is approximately standard normal distributed by the classical central limit theorem while the right hand side is approximately −α(l) (n)1/2+ε m1 −α(l) (n)1/2+ε  p = O (l) 1/2∨(1/2+ε)/2 ≤ O(−α(l) (n)ε/2 ) α (n) (λ(l) (n) + α(l) (n)1/2+ε )ν (l) which tends to −∞ as n → ∞. Hence (l) P(NT (n) − λ(l) (n) > α(l) (n)1/2+ε ) → 0. (6.67) Analogously it follows (l) P(NT (n) − λ(l) (n) < −α(l) (n)1/2+ε ) → 0 which together with (6.67) yields (6.66). Step 2. By the weak law of large numbers it holds λ(l) (n) (l) α (n) −1 X (l) (l) g α (n) (l) Ii,n i=1  λ (n) X λ(l) (n) 1 T (l)  P = (l) g Ei,n −→ (l) m(l) g . (l) α (n) λ (n) i=1 m1 Hence it remains to prove (l) λ(l) (n)∨NT (n) (l) X α(l) (n)−1 g α(l) (n) Ii,n  P −→ 0 (6.68) (l) i=λ(l) (n)∧NT (n)+1 as n → ∞. Because of (6.66) it suffices to prove the convergence (6.68) restricted to the set (l) Ω(l) (n, ε) = {|NT (n) − λ(l) (n)| ≤ α(l) (n)1/2+ε }. On this set (6.68) is bounded by (l) 2α (n) (1/2+ε)−1  1 (l) 2α (n)1/2+ε λ(l) (n)+dα(l) (n)1/2+ε e X (l)  g Ei,n  1Ω(l) (n,ε) i=λ(l) (n)−bα(l) (n)1/2+ε c+1 which converges in probability to zero for ε < 1/2 as n → ∞ because the term in (l) parantheses converges by the weak law of large numbers in probability to mg . 28 Proof of (2.7). Because of |Ii,n | ∼ Exp(nλ) it holds |Ii,k,n | ∼ Γ(k, nλ). Lemma 6.4 (l) with α(l) (n) = n/k and E1,1 ∼ ζ/k with ζ ∼ Γ(k, λ) yields k−1 1Xn Gn,k (t) = k k l=0 P −→ X k−1 1Xk |Iki+l,k,n | = k n X 2 i:tki+l,n ≤t l=0 i:tki+l,n ≤t 2 n |Iki+l,k,n | k k−1 k(k + 1)/(kλ)2 k+1 1 X E[(ζ/k)2 ] t= t= t. k E[ζ/k] k/(kλ) kλ l=0 Lemma 6.5. Under the Poisson sampling introduced in Example 3.6 we have the pointwise convergences P e e k,n (t) −→ G Gk (t), P Hk,n (t) −→ Hk (t), e k , Hk which are linear in t with positive slope. for functions G e k,n is similar. First Proof. We will proof the claim for Hk,n only, since the proof for G observe X n (1) (2) Hk,n (t) = 3 |Ii,k,n ||Ij,k,n |1{I (1) ∩I (2) 6=∅} k i,k,n j,k,n (1) (2) i,j:ti,n ∧tj,n ≤t L = = 1 1 k3 n (1) X (2) |Ii,k,1 ||Ij,k,1 |1{I (1) (2) i,k,1 ∩Ij,k,1 6=∅} (1) (2) i,j:ti,1 ∧tj,1 ≤nt bntc 1 bntc 1 X Ym + OP (n−1 ) k 3 n bntc m=1 L where = denotes equality in law and X (1) (2) Ym = |Ii,k,1 ||Ij,k,1 |1{I (1) (2) i,k,1 ∩Ij,k,1 6=∅} (1) (2) i,j:ti−k,1 ∧tj−k,1 ∈(m−1,m] , m ∈ N. Because the Poisson process has stationary increments, the sequence Ym , m ∈ N, is a stationary square integrable time series. Because further Ym1 , Ym2 become asymptotically independent as |m1 − m2 | → ∞, compare (A.51) in Martin and Vetter (2017), it is possible to conclude Cov(Ym1 , Ym2 ) → 0 as |m1 − m2 | → ∞. We then obtain by a law of large numbers for stationary processes, compare Theorem 7.1.1 in Brockwell and Davis (1991), P Hk,n (t) −→ This yields the claim as clearly E[Y1 ] > 0. 29 t E[Y1 ]. k3 6.3.1 Proof of Condition 2.7(ii), 3.4(iii) and 3.7 in the Poisson setting The convergences claimed in Condition 2.7(ii), 3.4(iii) and 3.7 are all special cases of the general statement made in the following Lemma and can be easily verified by finding suitable functions f (1) , f (2) , f . (1) (2) (3) Lemma 6.6. Let d ∈ N and Zn (s), Zn (s), Zn (s) be Rd -valued random variables, which can be written as √ (l) (l) (l) Zn(l) (s) = f (l) n(s − t (l) ), n(t (l) − s), ( n∆i(l) (s)+j,n W )j∈[k−1] , n in (s)−1,n in (s),n  √ (l) (l) ( n∆ (l) W 1{I (3−l) ) , l = 1, 2, (l) ∩I 6=∅} i∈{0,...,k−1},j∈Z in (s)+j,k,n (l) Zn(3) (s) = f (n|I (l)  in (s)+i,n (3−l) in (s)+i,k,n |)i∈[k]  (l) in (s)+j,k,n (1) (2) , (n|I (1) l=1,2 in (s)+i,n ∩ I (2) in (s)+j,n |)i,j∈[k]  with [k] = {−k, . . . , k} for measurable functions f (1) , f (2) , f and a fixed k ∈ N. Then under the Poisson sampling introduced in Example 3.6 the integral Z P1 hY 0 0 00 00 (1) g(x1 , . . . , xP1 , x1 , . . . , xP2 , x1 , . . . , xP3 )E h(1) p (Zn (xp )) [0,T ]P1 +P2 +P3 × P2 Y p=1 (2) 0 h(2) p (Zn (xp )) p=1 P3 Y i (3) 00 0 0 00 00 h(3) p (Zn (xp )) dx1 . . . dxP1 dx1 . . . dxP2 dx1 . . . dxP3 p=1 (6.69) converges for n → ∞ to Z P1 Z Y 0 0 00 00 (1) g(x1 , . . . , xP1 , x1 , . . . , xP2 , x1 , . . . , xP3 ) h(1) p (y)Γ (dy) [0,T ]P1 +P2 +P3 × P2 Z Y p=1 p=1 (2) h(2) p (y)Γ (dy) P3 Z Y (3) 0 0 00 00 h(3) p (y)Γ (dy)dx1 . . . dxP1 dx1 . . . dxP2 dx1 . . . dxP3 p=1 (6.70) (l) for all bounded continuous functions g : RP1 +P2 +P3 → R, hp : Rd → R and all P1 , P2 , P3 ∈ N. Note that Γ(1) , Γ(2) , Γ(3) are probability measures on Rd which do not depend on xp , x0p , x00p as in the general case in Conditions 2.7(ii), 3.4(iii) and 3.7 because the Poisson process has stationary increments. (l ) (l ) Proof. First we show that any two random variables Zn 1 (x1 ) and Zn 2 (x2 ) with l1 , l2 ∈ {1, 2, 3}, x1 , x2 ∈ [0, T ], become asymptotically independent for x1 6= x2 which yields that the expectation in (6.69) factorizes in the limit. Let x1 < x2 and define  (l1 ) x1 + x2 , ≤ Ω− n (l1 , x1 ) = t (l1 ) (3−l1 ) )+k,n in (t (3−l ) 2 1 in (x1 )+k  x1 + x2 (l ) Ω+ ≤ t (l22 ) (3−l2 ) n (l2 , x2 ) = in (t (3−l ) 2 2 in )−k−1,n (x2 )−k−1 + Ωn (l1 , l2 , x1 , x2 ) = Ω− n (l1 , x1 ) ∩ Ωn (l2 , x2 ). 30 , Here, Ωn (l1 , l2 , x1 , x2 ) describes the subset of Ω on which the set of intervals used (l ) for the construction of Zn 1 (x1 ) and the set of intervals used for the construction of (l ) Zn 2 (x2 ) are separated by (x1 + x2 )/2. Then there exist measurable functions g1 , g2 such that (l ) = g1 ((Nn (t))t∈[0,(x1 +x2 )/2] , (W t 1 )t∈[0,(x1 +x2 )/2] )1Ω− , Zn(l1 ) (x1 )1Ω− n (l1 ,x1 ) n (l1 ,x1 ) Zn(l2 ) (x2 )1Ω+ = g2 ((Nn (t) − Nn ((x1 + x2 )/2))t∈[(x1 +x2 )/2,∞) , n (l2 ,x2 ) (l ) (l ) (W t 2 − W (x21 +x2 )/2 )t∈[(x1 +x2 )/2,∞) )1Ω+ n (l2 ,x2 ) (1) (2) (l) where Nn (t) = (Nn (t), Nn (t))∗ , Nn (t) = P i∈N 1{t(l) ≤t} , l = 1, 2, denotes the i,n Poisson processes which create the stopping times. These identities yield that the (l ) random variables Zn 1 (x1 )1Ω− , 1Ω− are independent from the random n (l1 ,x1 ) n (l1 ,x1 ) (l ) because the processes W and Nn (t), have , 1Ω+ variables Zn 2 (x2 )1Ω+ n (l2 ,x2 ) n (l2 ,x2 ) independent increments. Hence we get (l1 ) (l2 ) (l2 ) 1) 1 + ] E[h(l p1 (Zn (x1 ))hp2 (Zn (x2 ))1Ω− n (l1 ,x1 ) Ωn (l2 ,x2 ) (l1 ) (l2 ) 1) 2) = E[h(l ]E[h(l ] p1 (Zn (x1 ))1Ω− p2 (Zn (x2 ))1Ω+ n (l1 ,x1 ) n (l2 ,x2 ) (l ) = = (l ) (l ) (l ) E[hp11 (Zn 1 (x1 ))1Ω− 1 − ] 1 + ] E[hp22 (Zn 2 (x2 ))1Ω+ n (l2 ,x2 ) Ωn (l1 ,x1 ) n (l1 ,x1 ) Ωn (l2 ,x2 ) E[1Ω+ E[1Ω− ] ] n (l2 ,x2 ) n (l1 ,x1 ) (l1 ) (l1 ) (l2 ) (l2 ) E[hp1 (Zn (x1 ))1Ωn (l1 ,l2 ,x1 ,x2 ) ]E[hp2 (Zn (x2 ))1Ωn (l1 ,l2 ,x1 ,x2 ) ] E[1Ωn (l1 ,l2 ,x1 ,x2 ) ] which is equivalent to (l2 ) 2) E[hp(l11 ) (Zn(l1 ) (x1 ))h(l p2 (Zn (x2 ))|Ωn (l1 , l2 , x1 , x2 )] (l1 ) (l2 ) (l2 ) 1) = E[h(l p1 (Zn (x1 ))|Ωn (l1 , l2 , x1 , x2 )]E[hp2 (Zn (x2 ))|Ωn (l1 , l2 , x1 , x2 )]. (6.71) Further we obtain as in (A.51) of Martin and Vetter (2017) P(Ωn (l1 , l2 , x1 , x2 )) ≥ 1 − K/n . |x1 − x2 | (6.72) Denote P = P1 + P2 + P3 , AT (ε) = {(x1 , . . . , xk ) ∈ [0, T ]k |∃i, j : |xi − xj | ≤ ε} and BT (σ) = {(x1 , . . . xk ) ∈ [0, T ]k : xσ(1) ≤ xσ(2) ≤ . . . ≤ xσ(k) } where σ denotes a permutation of {1, . . . , k}. Using (6.71), (6.72), the inequality |E[X] − E[X|A]| ≤ 2K 1 − P(A) P(A) which holds for any bounded random variable X ≤ K and any set A with P(A) > 0 31 (l) together with the boundedness of g, hp , l = 1, 2, 3, yields Z P1 Y   (1) E h(1) (6.69) − g(x1 , . . . , xP1 , x01 , . . . , x0P2 , x001 , . . . , x00P3 ) p (Zn (xp )) [0,T ]P1 +P2 +P3 P2 Y × p=1   (2) 0 E h(2) p (Zn (xp )) +   0 0 00 00 (3) 00 E h(3) p (Zn (xp )) dx1 . . . dxP1 dx1 . . . dxP2 dx1 . . . dxP3 p=1 p=1 ≤ Kλλ P3 Y P⊗ (AT (ε)) P1 P1Y +P2 Y (1) (1) (2) g(x1 , . . . , xP )E hp (Zn (xp )) h(2) p (Zn (xp )) Z X (6.69) − BT (σ)\AT (ε) σ∈SP P Y × (3) h(3) p (Zn (xp ))| X Z P m≤l Pm p= P m<l 3 Y l=1 P \  (l) (l)  E hp (Zn (xp ))| Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) ) dx1 . . . dxP Y ×  Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) ) dx1 . . . dxP g(x1 , . . . , xP ) BT (σ)\AT (ε) σ∈SP p=P1 +1 i=2 p=P1 +P2 +1 + p=1 P \ i=2 Pm +1 Z g(x1 , . . . , xP ) − BT (σ)\AT (ε) 3 Y P l=1 p= m≤l Pm Y P m<l   (l) E h(l) p (Zn (xp )) dx1 . . . dxP Pm +1 ≤ KλλP ⊗ (AT (ε)) Z X + K  T 1 − P Pi=2 Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) )  dx1 . . . dxP T P Pi=2 Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) ) BT (σ)\AT (ε) σ∈SP  T Z P X X 1 − P Pi=2 Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) ) + K  dx1 . . . dxP T P Pi=2 Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) ) BT (σ)\AT (ε) j=1 σ∈SP ≤ KλλP ⊗ (AT (ε)) Z X + K P (P + 1) ki=2 P(Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) )c ) dx1 . . . dxP P 1 − Pi=2 P(Ω(lσ(i−1) , lσ(i) , xσ(i−1) , xσ(i) ))c BT (σ)\AT (ε) σ∈SP   K/n −1 P P K/n  P⊗ ≤ Kλλ (AT (ε)) + K(P + 1) 1−P 2 ε ε for all ε > 0 where λλP ⊗ denotes the Lebesgue measure in RP . This term vanishes as n → ∞ and then ε → 0. Next observe that it holds Z P1 Y   (1) (6.70) − E h(1) g(x1 , . . . , xP1 , x01 , . . . , x0P2 , x001 , . . . , x00P3 ) p (Zn (xp )) [0,T ]P1 +P2 +P3 × P2 Y p=1   (2) 0 E h(2) p (Zn (xp )) p=1 ≤ KεP + KT P P3 Y   (3) 00 0 0 00 00 E h(3) p (Zn (xp )) dx1 . . . dxP1 dx1 . . . dxP2 dx1 . . . dxP3 p=1 Pl 3 X X Z sup (l) (l) (l) h(l) p (y)Γ (dy) − E[hp (Zn (x))] . l=1 p=1 x∈[ε,T ] 32 Hence it remains to show that there exists some measure Γ(l) on Rd with Z (l) (l) (l) sup h(l) p (y)Γ (dy) − E[hp (Zn (x))] → 0 (6.73) x∈[ε,T ] as n → ∞ for all ε > 0. L (l) (l) For proving (6.73) observe that it holds Zn (x) = Z1 (nx). Since the Poisson (l) processes generating the observation times in π1 and the Brownian motion W are (l) independent stationary processes the law of Z1 (nx) depends on n only through the fact that all occuring intervals are bounded to the left by 0. Hence if Ω(n, m, x) de(l) notes the set on which all intervals needed for the construction of Z1 (mx) are within (l) [mx − nx, ∞), it follows that all members of the sequence (Z1 (mx)1Ω(n,m,x) ))m≥n have the same law. This yields because of P(Ω(n, m, x)) = P(Ω(n, n, x)), m ≥ n, and lim P(Ω(n, n, x)) = 1, x > 0, n→∞ (l) that the sequence (Z1 (nx))n∈N converges in law for x > 0. Hence the sequence (l) (Zn (x))n∈N converges in law for x > 0. By the stationarity of the processes the law of the limit, which we denote by Γ(l) , does not depend on x. Finally we obtain the uniform convergence in (6.73) because of P(Ω(n, n, x)) ≤ P(Ω(n, n, ε)) for x ≥ ε. In the proof of Lemma 6.6 we obtained L Zn(l) (s)1Ω(l) (s) = Z (l) 1Ωe (l) (s) n n (l) (l) (l) e n (s))n with Ωn (s) ↑ Ω, for a random variable Z (l) ∼ Γ(l) and sequences (Ωn (s))n , (Ω (l) e n (s) ↑ Ω. e Hence if the limit law would admit a part which is singular to the Ω (l) Lebesgue measure, this would also be true for the law of the Zn (s) with n suffi(l) ciently large. In all three cases considered in this paper it can be shown that Zn (s) has no atom. Hence in Condition 2.7(ii) and Condition 3.7 we obtain Γ(x, {0}) = 0. Further by Theorem 6.7 from Billingsley (1999) there exist random variables e en(l) )n∈N ) defined on a probability space (Ω0 , F 0 , P0 ) with (Z (l) , (Z L e(l) )n∈N ) = (Z (l) , (Z (l) )n∈N ) (Ze(l) , (Z n n en(l) → Ze almost surely. An application of Fatou’s lemma then yields and Z e(l) kp ] ≤ lim inf E[kZe(l) (s)kp ] ≤ sup E[kZ e(l) (s)kp ] = sup E[kZ (l) (s)kp ] E[kZ (l) kp ] = E[kZ n n n n∈N n∈N n∈N for p ≥ 0. For the cases in Condition 2.7(ii), 3.4(iii) and 3.7 it can be shown that the supremum is finite for all p ≥ 0 and hence Γ(l) has finite moments. This yields the requirement of uniformly bounded first and second moments in Condition 2.7(ii), 3.4(iii) and 3.7. 33 6.3.2 Proof of Condition 2.10(i) and 3.11(i) in the Poisson setting Note that it holds L (ξˆk,n,1,− (s), ξˆk,n,1,+ (s)) =S (ξk,n,− (Un (s)), ξk,n,+ (Un (s))) with Un (s) ∼ U[tin (s)−Ln −1 , tin (s)+Ln ] where LS denotes equality of the S-conditional distributions. Hence the proof of Condition 2.10(i) in the Poisson setting is identical to the proof of (4.3) in Martin and Vetter (2017). The same holds true for Condition bn,1 (s) L=S Zn (Vn (s)) with 3.11(i) as Z (1) Vn (s) ∼ U[t (1) in (s)−Ln −1,n (1) (2) ∨ t (2) in (s)−Ln −1,n 34 , t (1) in (s)+Ln ,n (2) ∧ t (2) in (s)+Ln ,n ].
10math.ST
1 Hybrid Analog-Digital Beamforming for Massive MIMO Systems arXiv:1712.03485v1 [cs.IT] 10 Dec 2017 Shahar Stein, Student IEEE and Yonina C. Eldar, Fellow IEEE Abstract—In massive MIMO systems, hybrid beamforming is an essential technique for exploiting the potential array gain without using a dedicated RF chain for each antenna. In this work, we consider the data phase in a massive MIMO communication process, where the transmitter and receiver use fewer RF chains than antennas. We examine several different fully- and partially connected schemes and consider the design of hybrid beamformers that minimize the estimation error in the data. For the hybrid precoder, we introduce a framework for approximating the optimal fully-digital precoder with a feasible hybrid one. We exploit the fact that the fully-digital precoder is unique only up to a unitary matrix and optimize over this matrix and the hybrid precoder alternately. Our alternating minimization of approximation gap (Alt-MaG) framework improves the performance over state-of-the-art methods with no substantial increase in complexity. In addition, we present a special case of Alt-MaG, minimal gap iterative quantization (MaGiQ), that results in low complexity and lower mean squared error (MSE) than other common methods, in the case of very few RF chains. MaGiQ is shown to coincide with the optimal fully-digital solution in some scenarios. For combiner design, we exploit the structure of the MSE objective and develop a greedy ratio trace maximization technique, that achieves low MSE under various settings. All of our algorithms can be used with multiple hardware architectures. I. I NTRODUCTION Massive MIMO wireless systems have emerged as a leading candidate for 5G wireless access [1], [2]. Along with mmWave technologies, that were recently recognized as essential for coping with the spectrum crunch [3], it offers higher data rates and capacities than traditional MIMO systems. The use of large-scale antenna arrays at both the transmitter and receiver holds the potential for higher array gain than before. To utilize this gain, precoding and combining techniques are used. Traditionally implemented in the baseband (BB), these methods require a dedicated RF hardware per antenna. Unfortunately, when taking a massive amount of antennas into account, this results in a huge computational load and cost, as the RF components are expensive and have high power consumption, especially for mmWave technologies. Hence, it is desirable to design economical hardware that will utilize the potential gain from a large number of cheap antenna elements using a small number of expensive RF chains. To achieve this goal, several hybrid analog-digital schemes have been suggested [4]–[13]. In hybrid precoding and combining, the operations are split between the digital and analog domains: at the transmitter side a low dimensional digital precoder operates on the transmitted signal at BB. An analog This project has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No. 646804-ERCCOG-BNYQ, and from the Israel Science Foundation under Grant no. 335/14. precoder then maps the small number of digital outputs to a large number of antennas, and the same is performed at the receiver side. Common analog architectures are based on analog phase shifters and switches. The specific schemes can vary according to the power and area budget. Two main families of architectures are the fully- and partially connected structures. Fully connected networks offer a mapping from each antenna to each RF chain and allow to maximize the precoding and combining gain. In the partially connected scheme, a reduced number of analog components is used. This degrades the achieved gain but offers lower power consumption and hardware complexity. In the data phase of each coherence interval of the MIMO communication process, the transmitter sends multiple data streams to the receiver over the constant and known channel, using precoding techniques. At the receiver side, a combiner is used to estimate the data vector from the received signal at the antennas. The precoder and combiner are chosen to optimize some desired performance measure, such as estimation error or spectral efficiency of the system. Unlike the fully-digital case, when considering a hybrid beamformer, the precoder and combiner matrices cannot have arbitrary entries, but are constrained according to the specific hardware choice. For example, when using a phase shifter network at the analog side, only unimodular matrices for the analog beamformer are considered. The goal then is to optimize the performance measure over all pairs of digital and analog precoder matrices, which yields a non-convex difficult optimization problem. The majority of past works considered the fully-connected phase shifter network. It was suggested in [4] to separate the joint precoder and combiner design problem into two subproblems and solve each independently. Although this approach is sub-optimal, it greatly simplifies the difficult joint optimization problem. For the precoder, it was shown that minimizing the gap between the hybrid precoder and the optimal fullydigital one over all hybrid precoders, approximately leads to the maximization of the system’s spectral efficiency. On the combiner side, minimizing the mean squared error (MSE) over all hybrid combiners was shown to be equivalent to minimizing the weighted approximation gap between the fullydigital combiner and the hybrid one, with the weights given by the received signal correlation. Methods for solving these approximation problems were suggested in [4]–[6], [8]–[10], [12]. The works [4]–[6] considered precoder design and aimed at maximizing the system’s spectral efficiency. They exploit the mmWave sparse multipath channel structure and deduce that the optimal fully-digital precoder is composed of a small sum of steering vectors. They then suggest a variant of the or- 2 thogonal matching pursuit (OMP) [14] algorithm to construct a feasible precoder that approximates the optimal one using a dictionary of steering vectors. This solution greatly reduces the problem complexity but results in a large performance gap from the optimal precoder due to restricting the space of possible precoder vectors to steering vectors. Similar approaches were taken in [4], [8], [9] for the combiner design, under a MSE or spectral efficiency objective. In [10], two algorithms were suggested to approximate the optimal fully-digital precoder with a feasible one. The first, MO-AltMin, is based on manifold optimization. In each iteration of the algorithm, it assumes a given digital precoder and develops a conjugate gradient method to find an analog precoder that is a local minimizer of the approximation gap from the fully-digital one. Next, the digital precoder is computed using a least squares solution. This method achieves good performance but suffers from high complexity and run time, and is only suitable for fully-connected phase shifter networks. The second approach is a low complexity algorithm that assumes the digital precoder to be a scaled unitary matrix and uses this assumption to produce an upper bound on the approximation gap, which is then minimized over all analog precoders. However, limiting the combiner to such a structure results in performance loss. For the partially connected architecture, most of the existing works concentrated on fixed sub-arrays where each RF chain is connected to a predetermined sub-array. In [11], the authors suggested a low complexity codebook design producing a small dictionary of feasible precoding vectors, that are chosen based on the transmitted signal strength. They then exhaustively search over all possible combinations from the small dictionary to maximize the mutual information between the receiver and transmitter. This last step can result in heavy computational load when the channel is not sparse. In [10], the authors consider disjoint sub-arrays and suggest an iterative method to approximate the optimal fully-digital precoder, which optimizes over the analog and digital combiners alternately. It is shown that for disjoint arrays, the analog precoder problem is separable in the antennas, and has a closed form solution. For the digital precoder, a semidefinite relaxation is suggested. However, this is both computationally heavy and unnecessary since a closed form solution for the digital combiner is available. This alternating minimization approach still results in large performance gap. Some less restrictive schemes were considered in [12], [13]. In [12], a double phase shifter with dynamic mapping is considered. The use of two phase shifters per antenna relaxes the unit modulo constraint that limits the previous solutions, but costs twice the power consumption, area and complexity. In [13], a dynamic subarray approach is considered for orthogonal frequency-division multiplexing (OFDM), and a greedy algorithm to optimize the array partition based on the long-term channel characteristics is suggested. However, this method is relevant only for OFDM transmissions. Most of the above works concentrated on specific channel models and hardware architectures and suggested a tailored algorithm for the chosen scenario. In this work, we develop a general framework suitable for various channels and hardware: full and partial networks with different amounts of phase shifters and switches. We consider the data estimation problem in a single user massive MIMO system where both the receiver and transmitter are equipped with large antenna arrays and fewer RF chains than antennas. We assume a Bayesian model where the data and interference are both random, and aim at minimizing the MSE of the transmitted data from the low dimensional received digital signal at the receiver, over all hybrid precoders and combiners, assuming a fixed number of RF chains. Like previous works, we relax the difficult joint optimization to two separate problems in the precoder and combiner. To design the precoder, we present a framework for approximating the optimal fully-digital precoder with a feasible hybrid one. Our alternating minimization of approximation gap (AltMaG) method, exploits the fact that there exists an infinite set of optimal precoders, which differ by a unitary matrix. We suggest optimizing over this matrix and the hybrid precoder alternately, to find the fully-digital solution that results in the smallest approximation gap from its hybrid decomposition. For the hybrid precoder optimization step, any of the previously suggested methods [4]–[6], [8]–[10], [12] may be used, according to the hardware constraints. By optimizing over the unitary matrix as well, Alt-Mag achieves additional reduction in MSE compared to state-of-the-art algorithms, with no significant increase in complexity. We then present a simple possible solution for to hybrid precoder optimization step, that results in a low complexity algorithm, termed minimal gap iterative quantization (MaGiQ). In each iteration of MaGiQ, the fully-digital solution is approximated using the analog precoder alone. This results in a closed form solution, given by a simple quantization function which depends on the hardware structure. We demonstrate in simulations that MaGiQ achieves lower MSE than other low complexity algorithms when using very few RF chains. In addition, in some specific cases, it coincides with the optimal fully-digital solution. Next, we show that MaGiQ can also be used for the combiner design with mild adjustments. We then suggest an additional greedy ratio trace maximization (GRTM) algorithm that directly minimizes the estimation error using a suitable dictionary that is chosen according to the hardware scheme. In simulations, we demonstrate that GRTM enjoys good performance and short running time, especially when the number of RF chains increases. The rest of this paper is organized as follows: In Section II we introduce the signal model and problem formulation, and review common combiner hardware schemes. In Section III we introduce the massive MIMO data estimation problem and derive the relevant MSE minimization over the precoder and combiner. Sections IV and V consider the precoder and combiner design problems, respectively. In Section VI we evaluate the proposed algorithms using numerical experiments under different scenarios. The following notations are used throughout the paper: boldface upper-case X and lower-case x letters are used to denote matrices and vectors respectively, and non-bold letters x are scalars. The ith element in the jth column of X is [X]ij . 3 We let I denote the identity matrix of suitable size, X T , X ∗ the transpose and conjugate-transpose of X respectively, E {⋅} the expectation and ∥ ⋅ ∥2p , ∥ ⋅ ∥2F the `p and Forbenius norms. The determinant of X is ∣X∣, and CN (x, X) is the complexGaussian distribution with mean x and covariance matrix X. The real part of a variable is denoted as R {⋅}, R (X) is the range space of X and P X is the orthogonal projection onto R (X). r with W RF ∈ W Nr ×NRF the analog combining matrix and r r W BB the NRF × NRF digital combining matrix. The properties of the set W vary according to the specific analog hardware scheme. We wish to design W RF , W BB , F RF , F BB to minimize the estimation error of s from r, under different hardware constraints, i.e different feasible sets W, F. Thus, we consider the problem min W RF ,W BB F RF ,F BB II. P ROBLEM F ORMULATION s.t. ∶ A. Signal Model r x = F RF F BB s, (1) with s the Ns × 1 symbol vector. We assume, without loss of generality, that s ∼ N (0, I Ns ). The signal is transmitted over a narrowband block-fading propagation channel such that the Nr × 1 received analog signal vector y at the receiver’s antennas is √ (2) y = pr HF RF F BB s + z, where H is the Nr × Nt channel matrix, pr is the received average power and z is an Nr ×1 interference vector with zero mean and autocorrelation matrix E [zz H ] = Rz . Here, z can represent either noise or an interfering signal. We assume full channel state information (CSI), i.e. the matrix H is known. The interference correlation matrix Rz is also known and assumed to be full rank. At the receiver, an analog combiner maps the Nr inputs to r NRF RF chains that are then processed at baseband using a digital combiner. This yields the discrete signal √ t F RF ∈ F Nt ×NRF , ∥F RF F BB ∥2F ≤ Ns , W RF ∈ W Nr ×NRF . Consider a single user massive MIMO system in which a transmitter with Nt antennas communicates Ns independent t t data streams using NRF RF chains, Ns ≤ NRF ≤ Nt , to a r receiver equipped with Nr antennas and NRF RF chains, Ns ≤ r NRF ≤ Nr . At the transmitter, the RF chains are followed by a network of switches and phase shifters that expands the NRF digital outputs to Nt precoded analog signals which feed the transmit antennas. Similarly, at the receiver, the antennas are followed by a network of switches and phase shifters that feed r the NRF RF chains. The specific architecture of the analog hardware at each end can vary according to budget constraints. Some possible choices are presented in the next subsection. t r For simplicity, we assume in this paper that NRF = NRF = Ns , which corresponds to the minimal possible number of chains, and hence the worst-case scenario. The hybrid architecture enables the transmitter to apply a t t BB precoder F BB ∈ CNRF ×NRF , followed by a RF precoder t F RF ∈ F Nt ×NRF . The properties of the set F are determined by the specific hardware scheme in use. For example, in fullyconnected phase shifter networks, F is the set of unimodular matrices. The transmitter obeys a total power constraint such that ∥F RF F BB ∥2F = Ns . The discrete-time Nt × 1 transmitted signal can be written as r= E [∥s − W ∗BB W ∗RF y∥22 ] pr W ∗BB W ∗RF HF RF F BB s + W ∗BB W ∗RF z, (3) (4) B. Precoder and Combiner Hardware Schemes We now present some possible hardware schemes and the feasible sets they dictate. We focus on the combiner architecture and the feasible set W, but all the options presented here can be translated to similar precoder architectures and feasible sets F. The networks below are based on phase shifters and/or switches, with different connectivity levels. In practice, the choice of scheme can be affected by a variety of budget constraints such as power consumption, price, and area against the requested array gain. For example, phase shifters allow for a more flexible design than switches, but their power consumption is higher. (S1) Fully Connected Phase Shifters and Switches Network: Each antenna is connected to an RF chain through an independent on/off switch followed by a phase shifter. All the incoming analog signals from the different antennas are combined before feeding the RF chain. In this architecture, the entries of the analog combining matrix W RF can be either zeros or unit r modulus. Hence the set W is defined by W Nr ×NRF = r {W ∈ CNr ×NRF ∶ ∣wij ∣ ∈ {0, 1}}. The orthogonal projection P W onto the feasible set dictated by this scheme is defined as ⎧ ⎪ej2π∠[A]il , ∣[A]il ∣ ≥ 1 ⎪ 2 . (5) [P W (A)]il = ⎨ ⎪ 0, ∣[A]il ∣ < 12 ⎪ ⎩ The fully-connected phase shifters and switches network is presented in Fig. 1a. (S2) Fully Connected Phase Shifters Network: (architecture A1 in [8]). In this case, each antenna is connected to an RF chain through an independent phase shifter, without a switch. Thus, W RF is a unimodular matrix: r r W Nr ×NRF = {W ∈ CNr ×NRF ∶ ∣wij ∣ = 1}. Here we have [P W (A)]il = ej2π∠[A]il . (6) Figure 1b. demonstrates this setting. (S3) Switching Network: (architecture A5 in [8]). This is an antenna selection scheme, where each RF chain is preceded by a single switch that can togr gle between the Nr antennas. A total of NRF antennas are selected and sampled. It follows that r W RF is partial permutation matrix: W Nr ×NRF = 4 r {W ∈ CNr ×NRF ∶ wij ∈ {0, 1} , ∥wj ∥0 = 1}. The projection P W (A) chooses the entry with largest absolute value in each column of A and sets the corresponding entry of the output to 1, while all other entries are set to 0. This setting is shown in Fig. 1c. Another structure is the switching network with sub-arrays (architecture A6 in [8]), where the switching range is limited to a certain subset of antennas for each RF chain. Note that in switching networks, phase shifters are unnecessary, since each input can be processed independently at BB. (S4) Partially Connected Phase Shifters Network with Fixed Sub Arrays: (architecture A2 in [8]). This scheme divides the array into (possibly overlapping) sub-arrays of size G. Each sub-array operates as a fully-connected phase shifter network with a single r RF chain. This results in the feasible set W Nr ×NRF = r Nr ×NRF {W ∈ C ∶ ∣wij ∣ = 1, ∀i ∈ Sj , ∣wij ∣ = 0, ∀i ∉ Sj }, with Sj the group of indices corresponding to antennas that belong to the jth sub-array. In this case P W (A) applies the mapping in (6) to the G entries of aj that belong to Sj , and sets all others to 0. The fixed subarrays network is presented in Fig. 1d. The parameter G represents the connectivity factor; for G = Nr , we get the fully-connected network S2. (S5) Flexible Partially Connected Phase Shifters Network with Sub Arrays: This architecture is similar to the previous one, except that the contributing antennas to each RF chain are not fixed, but can be optimized, i.e. the subgroups Sj are flexible. This is implemented using a Nr -to-1 switch that precedes the phase r shifters. The corresponding feasible set is W Nr ×NRF = r Nr ×NRF {W ∈ C ∶ ∣wij ∣ ∈ {0, 1} , ∥wj ∥0 = G}. The function P W (A) in this case chooses the G entries with largest absolute value in each column of A and applies (6) to them while setting all others to 0. The networks above constitute some of the main classes of existing analog architectures. They present different levels of challenge in terms of hardware complexity, as in Table I. For a complete power consumption comparison, the reader is referred to the example in [8]. One may define many additional architectures based on the above networks by modifying their components and connectivity factors. (a) (b) (c) (d) (e) Fig. 1: Combiner hardware schemes. (1a) S1: Fully connected phase shifters and switches network, (1b) S2: Fully connected phase shifters network, (1c) S3: Switching network, (1d) S4: Fixed partially connected phase shifters network with subarrays, and (1e) S5: Flexible partially connected phase shifters network with sub-arrays. The resulting estimate ∗ ∗ ŝ = H̄ W RF [W RF ∗ (H̄ H̄ + Rz ) W RF ] and its MSE is equal to ∗ To optimally design the precoder and combiner we consider the MSE minimization problem in (4). Since there are no constraints on W BB in (4), it can be r r matrix, including the linear MMSE chosen as any NRF ×NRF estimator of s from the measurements ỹ = W ∗RF y. This estimator depends on the matrices F RF , F BB ,W RF and is given by ∗ ∗ W ∗RF y, (9) E[∥s − ŝ∥2 ] = Ns − III. MSE M INIMIZATION W ∗BB,opt = H̄ W RF [W RF ∗ (H̄ H̄ + Rz ) W RF ] −1 −1 with −1 W ∗RF H̄) . (10) The digital combiner of (7) is the optimal linear estimator in the MSE sense for any given precoders F RF , F BB and analog combiner W RF . Thus, our remaining goal is to design W RF , F RF , F BB to minimize (10). This is equivalent to the following maximization problem: max , (7) ∗ tr (H̄ W RF [W RF ∗ (H̄ H̄ + Rz ) W RF ] f (W RF , F RF , F BB ) W RF ,F RF ,F BB s.t. ∶ t F RF ∈ F Nt ×NRF , ∥F RF F BB ∥2F ≤ Ns , r √ H̄ = pr HF RF F BB . W RF ∈ W Nr ×NRF , (8) (11) 5 TABLE I: Hardware complexity comparison between different architectures Phase Shifters # Switches # Switch Type S1 r Nr ⋅ NRF r Nr ⋅ NRF on-off S2 r Nr ⋅ NRF 0 - with f (W RF , F RF , F BB ) = (12) ∗ ∗ tr(W ∗RF H̄ H̄ W RF [W RF ∗ (H̄ H̄ −1 + Rz ) W RF ] ). Similar to the case considered in [4], the joint optimization problem (11) is non-convex and has no known solution. Hence, as in previous works [4], [10], [12], we simplify the problem by decoupling it. First, in Section IV, we optimize the precoders F RF , F BB given W RF . Then, in Section V, we assume fixed F RF , F BB and optimize the combiner W RF . IV. P RECODER DESIGN We now consider the problem of designing the precoder F = F RF F BB . For this purpose we assume a fixed analog combiner and rewrite (11) as ∗ ∗ −1 max tr (H̃F F ∗ H̃ [H̃F F ∗ H̃ + R̃] ) (13) F RF ,F BB t Nt ×NRF ∥F RF F BB ∥2F , ≤ Ns , s.t. ∶ F RF ∈ F √ with H̃ = pr W ∗RF H and R̃ = W ∗RF Rz W RF . If we relax t the constraint F RF ∈ F Nt ×NRF to the fully-digital case F RF ∈ t CNt ×NRF , then (13) has a closed form solution [15] F opt = V ΦT (14) t where V contains the first NRF eigenvectors of the matrix ∗ −1 t t H̃ R̃ H̃, Φ is an NRF × NRF diagonal matrix with power t t allocation weights as in [15], and T denotes any NRF × NRF unitary matrix. It follows that there is a set of optimal solutions, one for every unitary T . t Previous works [4]–[6], [8], [10], [12] considered T = I NRF and suggested to approximate F opt with a decomposition F RF F BB , which yields the problem S3 0 r NRF Nr -to-1 S4 r G ⋅ NRF 0 - S5 r G ⋅ NRF r G ⋅ NRF Nr -to-1 the amount of phase shifters per RF chains, and is irrelevant for other schemes. None of the above works take advantage of the flexibility in choosing T , which we show below can boost the performance of existing solutions with only a small computational price. Specifically, we exploit the flexibility in the choice of T , and find the unitary matrix that results in the smallest approximation gap ∥V ΦT − F RF F BB ∥2F . To this end, we develop an alternating minimization framework inspired by [16], that alternately optimizes over F RF , F BB and T . In the simulations in Section VI, we demonstrate that by adding the optimization over T , we are able to improve the performance of state-of-the-art algorithms, with no substantial increase in complexity. A. Alt-MaG - Alternating Minimization of Approximation Gap Plugging (14) into (15) and adding T as an optimization variable, our problem becomes min T ,F RF ,F BB ∥V ΦT − F RF F BB ∥2F t F RF ∈ F Nt ×NRF , ∥F RF F BB ∥2F ≤ Ns , s.t. ∶ T ∈U t NRF , M with U the set of all M ×M unitary matrices. To approximate the solution of (16), we consider an alternating minimization approach. First, we fix F RF , F BB and solve min T s.t. ∶ ∥V ΦT − F RF F BB ∥2F (17) t T ∈ U NRF . Next, using the resulting T , we solve min F RF ,F BB ∥V ΦT − F RF F BB ∥2F (18) t s.t. ∶ F RF ∈ F Nt ×NRF , ∥F RF F BB ∥2F ≤ Ns , min ∥F opt − F RF F BB ∥2F F RF ,F BB s.t. ∶ F RF ∈ F t Nt ×NRF (16) (15) , ∥F RF F BB ∥2F ≤ Ns . This approach is motivated by the approximation in [4] which shows that minimizing (15) approximately maximizes the system’s spectral efficiency. However, these solutions suffer from several disadvantages. In [4]–[6], [8], [9], the suggested design for the analog precoder is based on a dictionary composed of candidate vectors. This constraint leads to a large performance gap from the fully-digital precoder, especially in the case where the number of RF chains is much less than the number of multipath components in the channel H. In [10], the author suggests a manifold optimization algorithm termed MO-AltMin that achieves good performance, but has very high complexity and is suitable only for scheme S2. The solution in [12] is based on the network S4 but with twice and repeat iteratively. If both (17) and (18) can be solved, then the objective value decreases in each step, and since it is bounded from below, the algorithm will converge to a local optimum. However, while (17) has a (closed form) optimal solution, problem (18) is a special case of (15), which has no known solution. To attain a sub optimal solution, any of the mentioned algorithms [4]– [6], [8]–[10], [12] can be used (provided it is compatible with the hardware choice). Each of these solutions will result in a different algorithm from the Alt-MaG family. A closed form solution to problem (17) is given in the next theorem. Theorem 1. Given the svd decomposition F ∗BB F ∗RF V Φ = ∗ Ũ ΛṼ , the optimal solution T opt to (17) is given by T opt = Ṽ Ũ ∗ (19) 6 Proof. First note that ∥V ΦT − F RF F BB ∥2F = ∥V Φ∥2F + ∥F RF F BB ∥2F − 2R(tr (F ∗BB F ∗RF V ΦT )). Thus, it is sufficient to maximize R(tr (F ∗BB F ∗RF V ΦT )). Denote Λ = Ω2 . Then, R(tr(F ∗BB F ∗RF V ΦT )) Note that since the optimal solution V ΦT obeys the transmit power constraint ∥V ΦT ∥2F ≤ NS , it follows that ∥F RF F BB ∥2F = ∥F RF [F ∗RF F RF ] −1 F ∗RF V ΦT ∥2F = ∥P F RF V ΦT ∥2F ≤ ∥V ΦT ∥2F ≤ NS . Therefore, (23) complies with the transmit power constraint. MaGiQ is summarized in Algorithm 2. Since the nonnegative objective function ∥V ΦT −F RF ∥2F does not increase at each iteration of Algorithm 2, the algorithm converges. In the simulations in Section VI, we demonstrate that MaGiQ has lower MSE than other low-complexity methods in the case of very few RF chains. ≤ ∣tr(F ∗BB F ∗RF V ΦT )∣ = ∣tr((Ũ Ω) (T ∗ Ṽ Ω)∗ )∣ √ ∗ ∗ tr(Ũ Ω2 Ũ )tr(T ∗ Ṽ Ω2 Ṽ T ) ≤ (∗) = tr (Ω2 ) , where (∗) stems from the Cauchy-Schwarz inequality. For ∗ T = Ṽ Ũ , the upper bound is achieved with equality. The resulting family of Alt-MaG algorithms is outlined in Algorithm 1. Algorithm 2 MaGiQ - Minimal Gap Iterative Quantization Input: fully-digital precoder V Φ, threshold t Output: digital precoder F BB , analog precoder F RF t t Initialize T = I NRF , F RF = 0, F BB = I NRF 2 While ∥V ΦT − F RF ∥F ≥ t do: 1) Calculate F RF = P F (V ΦT ) ∗ 2) Calculate the svd decomposition F ∗RF V Φ = Ũ ΛṼ ∗ 3) Set T = Ṽ Ũ −1 Calculate F BB = [F ∗RF F RF ] F ∗RF V ΦT Algorithm 1 Alt-MaG - alternating minimization of approximation gap Input: fully-digital precoder V Φ, threshold t Output: digital precoder F BB , analog precoder F RF t Initialize T = I NRF , F RF = 0, F BB = 0 While ∥V ΦT − F RF F BB ∥2F ≥ t do: 1) Calculate (F RF , F BB ) from (18) using one of A version of Algorithm 2, PE-AltMin, has been previously the methods in Section IV suggested in [10] for the scheme S2, as an optimization of ∗ 2) Calculate the svd decomposition F ∗BB F ∗RF V Φ = Ũ ΛṼ an upper bound on (18). There, the authors assumed that the ∗ optimal F BB is an orthogonal matrix of the form F BB = αT , 3) Set T = Ṽ Ũ and considered the problem Next we suggest a possible simplification to (18) that can min ∥V Φ − αF RF T ∥2F T ,α,F RF be used with any of the schemes S1-S5, and results in a simple t (24) closed form solution for F RF , F BB . s.t. ∶ F RF ∈ F Nt ×NRF , ∥αF RF T ∥2F = Ns , t T ∈ U NRF . B. MaGiQ - Minimal Gap Iterative Quantization √ The complexity of Algorithm 1 is determined by the method used in step 1 and could be accordingly high. One approach that yields a low complexity algorithm, is approximating V ΦT by an analog precoder only, setting the digital precoder t to F BB = I NRF . That is, in every iteration, solving min ∥V ΦT − F RF ∥2F F RF t (20) s.t. ∶ F RF ∈ F Nt ×NRF . The benefit of this approach is the existence of a simple, optimal, closed form solution to (20), given by F RF = P F (V ΦT ) , (21) with P F (⋅) defined in Section II. Once F RF and T converge, F BB is calculated as the solution to min ∥V ΦT − F RF F BB ∥2F F BB (22) s.t. ∶ ∥F RF F BB ∥2F ≤ NS . Note that F RF depends on V ΦT which are a function of the random channel and interference. With high probability, it has full column rank. In this case, (22) has a closed form solution −1 F BB = [F ∗RF F RF ] F ∗RF V ΦT . (23) Setting α = ∥F RFNTS ∥F , and using some mathematical manipulations, they bounded the objective in (24) from above with the expression ∥V ΦT ∗ − F RF ∥2F . After optimizing the upper bound √ over T and F RF alternately, they set F BB as F BB = ∥F RFNTS ∥F T , which is sub optimal, as demonstrated in Section VI, where we compare both methods. The main difference between this approach and MaGiQ, is that while we exploit the degree of freedom in the fully-digital precoder, PE-AltMin restricts the set of possible digital precoders to orthogonal matrices alone, which degrades performance. In addition, Algorithm 2 is relevant to any hardware structure. Alt-MaG and MaGiQ where developed for MSE minimization, but can be generalized to other performance criteria, by replacing the optimal unconstrained solution expressions. For example, when considering the system’s spectral efficiency as in [4], [10], the optimal solution is (14) with Φ now being a diagonal matrix with the power allocation weights given by water filling [15]. C. Optimality of MaGiQ One advantage of MaGiQ over other existing techniques is that in some special cases it converges to the optimal fullydigital solution, as we show in the following proposition. 7 Proposition 1. Define F opt as in (14), and assume that Φ has positive values on its diagonal. Denote F̃ the hybrid precoder t produced by MAGiQ for the input F opt . If V Ψ ∈ F Nt ×NRF , for some diagonal, positive definite matrix Ψ, then It follows that F ∗RF V Φ = ΨΦ, which is diagonal. Thus, step 3 yields t T = I NRF with αil the complex gain of the lth ray in the ith cluster, r t ), (φril , θil ) the azimuth and elevation angles and (φtil , θil of departure and arrival for the lth ray in the ith cluster t r ) and ar (φril θil respectively. The vectors at (φtil , θil ) represent the transmit and receive array responses, and depend on the array geometry. For Nray = 1, (26) can be written √ as (25) with Λ a Ncl ×Ncl Nt Nr diagonal matrix with the elements α on its diagNcl Nray il onal, and Ar , At being steering matrices composed of the r t ) respectively. As Nr , Nt vectors ar (φril , θil ) and at (φtil , θil grow to infinity, the matrices Ar , At become orthogonal. Hence, in a noise-limited system, V is asymptotically equal to the steering vectors (up to some normalization constant) and thus unimodular. The above example can be generalized easily to a hybrid fully-connected phase shifter combiner rather than a fullydigital one. To summarize the results of this section, MaGiQ is a low complexity algorithm that yields low MSE with very few RF chains. When the number of chains increases, Alt-Mag can be used with other methods to solve problem (18) in its first step. The choice of algorithm should be made according to the complexity requirement. For example, if the computational load is not a limiting factor, then MO-AltMin from [10] is a good choice. and the algorithm converges after a single iteration. For F BB we get V. A NALOG C OMBINER D ESIGN F̃ = V Φ. Proposition 1 implies that if each column of V lies in the feasible set F Nt ×1 up to some scaling, then MaGiQ will produce a globally optimal solution. This property is not assured in other algorithms, as demonstrated in the simulations in Section VI. Proof. For simplicity, we consider here the fully-connected phase shifter network S2. However, Proposition 1 holds for any t hardware scheme with the same condition V Ψ ∈ F Nt ×NRF , and similar proofs can be constructed accordingly. t For S2, V Ψ ∈ F Nt ×NRF implies that V has columns with constant modulus. In the first iteration of MaGiQ, T is t initialized as T = I NRF . Since Φ is a diagonal matrix with positive diagonal values, step 1 of the algorithm leads to F RF = P F (V Φ) = V Ψ. F BB = [F ∗RF F RF ] −1 We now turn to design the analog combiner F ∗RF V Φ = Ψ−1 Φ, W = W BB W RF . and the received hybrid precoder is F̃ = V Φ. Note that the diagonal matrix Φ represents non-negative weights [15]. If one of the weights is equal to zero, then the corresponding column in the optimal fully-digital precoder is all zeros. In that case, we can drop the relevant column, and use the above proposition with the remaining fully-digital precoder. One example of a scenario in which the conditions of Proposition 1 hold, is a fully-connected phase shifter precoder (for simplicity we assume fully-digital combiner) with Rz = σz2 I Nr , and circulant channel matrix, that is H = Ar ΛA∗t , (25) where Ar , At are DFT matrices and Λ is a diagonal gain matrix. In this case, the singular vectors of H ∗ R−1 z H are columns of the DFT matrix which are unimodular, and thus lie in F Nt ×1 for S2. Therefore, from Proposition 1, MaGiQ produces the optimal solution (14). A common special case of (25), is the well-known narrowband mmWave clustered channel [17]–[19], with a single ray at each cluster and asymptotic number of antennas. This model consists of Ncl clusters with Nray propagating rays, each ray associated with transmit and receive directions, and complex gain. The channel matrix can be written as √ Nt Nr Ncl Nray r r t t ∗ H= ∑ ∑ αil ar (φil , θil ) at (φil , θil ) , (26) Ncl Nray i=1 l=1 (27) For this purpose we assume a fixed precoder, that is, H̄ in (8) is fixed and known. r r In (7), we derived the optimal NRF ×NRF digital combiner, given the hybrid decomposition (27), and reduced the problem to W RF alone. A different approach is to first allow the r entire Nr × NRF hybrid combiner to be digital, and then try and approximate the fully-digital solution with a hybrid decomposition, similar to the precoder optimization in the previous section. This approach was adopted in [4], [9]. There, the authors ber gun by allowing W to be fully-digital, that is W ∈ CNr ×NRF , and solved problem (4) for fixed precoders. This yields the MMSE estimator of s from y, given by W mmse = B −1 H̄, (28) with ∗ B = H̄ H̄ + Rz . (29) It was shown in [4] that minimizing (4) (for fixed precoders) is equivalent to solving the following optimization problem: min W RF ,W BB s.t. ∶ 1 ∥B 2 (W mmse − W RF W BB ) ∥2F r (30) W RF ∈ W Nr ×NRF . That is, finding the hybrid decomposition of W mmse that minimizes the weighted norm in (30). 8 Unlike the precoder case, here (30) and (4) are equivalent. Motivated by this, the authors in [4], [9] assumed a multipath channel structure as in (26), which implies that the optimal combiner is a sparse sum of steering vectors. They then suggested a variation of simultaneous OMP (SOMP) to solve (30), that chooses the analog combiner vectors from a steering dictionary and sets the digital combiner as the corresponding weights. However, this algorithm suffers from degraded performance due to the dictionary constraint. 1 By dropping the weights matrix B 2 , (30) becomes min W RF ,W BB s.t. ∶ combiner, while looking for an S that yields a minimal approximation gap between the two. In (33), S is only required to be invertible, in contrast to r the unitary constraint in (14). Hence, limiting it to U NRF is restrictive. However, this is a convenient way to enforce the invertibility constraint. The optimization problem we consider is therefore 1 min ∥B − 2 U S − W RF ∥2F W RF ,S r W RF ∈ W t Nt ×NRF which constitutes an upper bound on (30) (divided by the 1 constant factor ∥B 2 ∥2F ), and is identical (up to the power constraint) to the precoder problem (15) in the previous section. Hence the methods in [5], [6], [8], [10], [12] can be used to solve it. However, solving (31) is only optimizing an upper bound on the original objective, and is thus sub-optimal. A. MaGiQ for Combiner Design As opposed to previous works, we impose the hybrid decomposition (27), and optimize (4) over both W BB and W RF , as in Section III. This leads to the problem of minimizing (10) over W RF , which is equivalent to −1 max tr (W ∗RF AW RF [W ∗RF BW RF ] ) W RF r (32) s.t. ∶ W RF ∈ W Nr ×NRF , ∗ with A = H̄ H̄ , B as in (29), and W the feasible set of the given hardware scheme as described in Section II. Note that since Rz is positive definite, it follows that B is invertible. Proposition 2. The optimal fully-digital solution to (32), is given by 1 W opt = B − 2 U S, (33) r NRF ∗ where U contains the first eigenvectors of the matrix − 12 − 12 B AB , with A = H̄ H̄ , and H̄, B are as in (8), and r r (29). The matrix S is any NRF × NRF invertible matrix. Proof. The fully-digital problem corresponding to (32) is −1 max tr (W ∗RF AW RF [W ∗RF BW RF ] ) . W RF r S∈U (31) , (34) 1 Let Q ∈ CNr ×NRF be defined by Q = B 2 W RF . Then 1 1 the objective in (34) equals to tr(P Q B − 2 AB − 2 ). Let V r be a Nr × NRF matrix with orthogonal columns that span R(Q). Then, P Q = V V ∗ and Q = V S for some invertr r ible NRF × NRF matrix S. Thus, the objective becomes − 21 ∗ − 21 tr(V B AB V ) which is maximized by choosing V 1 1 r as the first NRF eigenvectors of B − 2 AB − 2 , i.e. V = U . Therefore, Q = U S, (35) s.t. ∶ W RF ∈ W Nr ×NRF , ∥W mmse − W RF W BB ∥2F r NRF , which can be solved using the MaGiQ approach presented in Section IV. Note that here Alt-MaG and MaGiQ coincide, since in step 1 of Alt-MaG the optimization is over W RF alone. Once the algorithm converges, W BB is calculated as in (7). MaGiQ enjoys good performance with very few RF chains, but when the number of chains increases, other methods are preferable. One of the main factors that degrade MaGiQ, is that it only optimizes a bound (31) on (30), rather than directly maximizing (30) (which is equivalent to (32)). Next, we exploit the special structure of the ratio-trace objective in (32), and construct a greedy method for its direct maximization over W RF . The proposed GRTM algorithm solves a ratio-ofscalars problem at each step and achieves low MSE when the number of RF chains increases. Both GRTM and MaGiQ enjoy low complexity and each has its merits. For a small number of RF chains or a noise-limited case with a fully connected phase shifters network and channel that has unimodular singular vectors, one should choose MaGiQ. For a more general channel model and increasing number of RF chains, GRTM is preferable. B. GRTM - Greedy Ratio Trace Maximization We now describe the GRTM approach for directly solving (32). The concept of GRTM is to solve (32) in a greedy manner, where in each iteration we add one RF chain and choose the optimal combiner vector to add to the previously r K selected vectors, 1 ≤ K ≤ NRF . (K) Assume we have a solution W RF ∈ W Nr ×K for the Kr sized problem, i.e. (32) with NRF = K. We now want to add an additional RF chain, and compute the optimal column (K+1) (K) w such that W RF = [W RF w]. First, note that in order (K+1) ∗ (K+1) for W RF BW RF to be invertible and the objective in (32) to be well defined for the (K + 1)-sized case, we require (K) w ∉ R(W RF ). In practice, this condition implies that each RF chain contributes new and independent information with respect to the other chains. Given this condition, the (K + 1)sized optimization problem is (K+1) ∗ which yields the solution (33). max tr (W RF Similarly to the previous section, we can now approximate the optimal unconstrained solution (33) with an analog s.t. ∶ w (K+1) AW RF (K+1) ∗ [W RF (K+1) −1 BW RF ] ) w ∈ W Nr ×1 , w ∉ R (W RF ) . (K) (36) 9 In the next proposition, we show that (36) is equivalent to solving the following vector optimization problem, referred to as the base case: w∗ Cw max w (37) w∗ Dw s.t. ∶ w ∈ W Nr ×1 , w∗ Dw > 0, where D and C are specific Nr × Nr positive semi definite (psd) matrices. Using Proposition 3 leads to the GRTM solution of (32), where in each iteration we choose the best column vector to add to the previously selected combiner vectors by solving (37). The GRTM algorithm is summarized in Algorithm 3. Proposition 3. Problems (37) and (36) are equivalent with 1 1 D = B 2 (I Nr − P B 12 W (K) ) B 2 , RF γ = tr (P B 12 W (K) B − 21 1 1 A 2 B− 2 ) , RF 1 1 1 1 G = B 2 P B 12 W (K) B − 2 A 2 − A 2 , RF C = γD + GG∗ . Proof. To prove the proposition, we rely on the following lemma: −1 Lemma 1. Let W̃ = [W w] and denote Q = (W ∗ W ) . Then [20] −1 ∗ (W̃ W̃ ) with α = =[ Q + αQW ∗ ww∗ W Q∗ −αw∗ W Q∗ −αQW ∗ w ], α 1 . w∗ w−w∗ W QW ∗ w Using Lemma 1 and straightforward algebraic operations we get equality between the two objectives. The first constraint is identical in both problems. It remains to show that the second constraints of (37) and (36) are equivalent. 1 (K) Since B is invertible, w ∈ R(W RF ) if and only if B 2 w ∈ 1 (K) (K) R(B 2 W RF ). Therefore, if w ∈ R(W RF ), then 1 1 w∗ Dw = w∗ Bw − w∗ B 2 P B 12 W (K) B 2 w RF 1 Algorithm 3 GRTM Input effective channel H̄, interference correlation Rz Output analog combiner W RF ∗ ∗ Initialize A = H̄ H̄ , B = H̄ H̄ + Rz , C = A, D = B, (0) W RF = [] r For K = 0 ∶ NRF − 1: 1) Solve the base case (37) to obtain w (K+1) (K) 2) Update W RF = [W RF w] 3) Update 1 = w∗ Bw − w∗ B 2 B 2 w = 0. 1 1 1 D = B 2 (I Nr − P B 21 W (K+1) ) B 2 In the other direction, note that D is a non-negative Hermitian matrix, and can be decomposed as D = QQ∗ for some Q. Thus, w∗ Dw = 0 if and only if Q∗ w = 0. Multiplying 1 both sides of the equation by Q yields Dw = B 2 (I Nr − 1 P B 12 W (K) )B 2 w = 0. Since B is invertible, it follows that RF Problem (37) has several advantages over (36): 1) its objective is a scalar ratio and does not involve matrix inversion. 2) this problem is both a constrained ratio-trace, trace-ratio [21] and a Rayleigh quotient problem, and therefore may be solved using techniques for either of these well known problems. However, these methods require adaptation to fit the additional hardware constraints, which will not be investigated in this work. One simple method for obtaining a (sub-optimal) solution to (37) is searching over a dictionary W̄ with columns in W Nr ×1 , and calculating the objective value in (37) for each vector. The vector that corresponds to the largest objective is then chosen. This solution has low computational load thanks to the simplicity of the scalar ratio objective. The choice of dictionary depends on the channel model and hardware constraints. For a fully-connected scheme and a sparse mmWave channel as in (26), a steering vector dictionary as used in [4] exploits the channel structure and yields good performance. For a more general model, Gaussian randomization can be used to construct a good dictionary. In this case, W̄ = P W (X), with P W (⋅) as defined in Section II, and where X is a random dictionary, commonly chosen as a matrix with columns drawn from CN (0, W opt W ∗opt ), with W opt from (33). In the simulations section, we used the dictionary based solution with Gaussian randomization. To comply with the problem’s second constraint, it is necessary to remove the selected column from the dictionary at each iteration. The other columns are, with (K) high probability, not in R(W RF ) due to the random complexGaussian distribution. 1 1 RF 1 1 1 γ = tr (P B 12 W (K+1) B − 2 A 2 B − 2 ) RF 1 1 1 1 G = B 2 P B 12 W (K+1) B − 2 A 2 − A 2 RF C = γD + GG∗ (K) (I Nr − P B 12 W (K) )B 2 w = 0, or B 2 w ∈ R(B 2 W RF ), RF (K) which is equivalent to w ∈ R(W RF ). Hence, we proved that (K) w ∈ R (W RF ) ⇐⇒ w∗ Dw = 0. Since D is non-negative definite, w∗ Dw ≥ 0, for all w. The previous connection then yields Simulations demonstrate that GRTM has lower MSE than MaGiQ when the number of RF chains increases. In that case, it outperforms other low complexity methods in various SNR, channel models, and hardware scenarios. w ∉ R (W RF ) ⇐⇒ w∗ Dw > 0. C. Combiner Design for Kronecker Model Channel Estimation As there is equality between the objectives and feasible sets of (37) and (36), both problems are equivalent. Problem (32) arises in other communication problems, so that GRTM can be used to address those problems as well. One (K) 10 such scenario is MIMO channel estimation under a Kronecker model in the interference limited case [22], [23]. This problem consists of a single base station (BS) with NBS antennas and NRF < NBS RF chains, and K singleantenna users in a time-synchronized time division duplex (TDD) system. In the uplink channel estimation phase, each user sends τ training symbols to the BS, during which the channel is assumed to be constant. The (discrete time) received signal at the BS can be written as X = W ∗RF HS T + Z, (38) where H ∈ CNBS ×K is the desired channel between the users to the BS, S ∈ Cτ ×K is the pilot sequence matrix, W RF ∈ W NBS ×NRF is the analog combiner and Z is the interference. Both the channel and interference follow the doubly correlated Kronecker model [24] 1 1 1 1 H = Rr2 H̄Rt2 , Z = Rr2 Z̄Q 2 , (39) Nbs ×Nbs with Rr ∈ C the BS’s receive side correlation matrix, Rt ∈ CK×K the transmit side correlation matrix of the users and Q ∈ CKτ ×τ the transmit side correlation matrix N ×K of the interference. The matrices H̄ bs , Z̄ ∈ CNbs ×τ have independent entries with i.i.d complex-normal distribution. All correlation matrices are known. In the channel estimation phase, the goal is to estimate H from X. For this purpose it is desirable to design the analog combiner W RF to minimize the MSE in estimation. As shown in [22], the estimation error of the MMSE estimator for H from X is inversely proportional to µ ≜ tr (W ∗RF R2r W RF (W ∗RF Rr W RF ) ) , −1 (40) which is equal to (32) with A = R2r and B = Rr . Hence, all the previously mentioned methods may be used to design W RF . VI. N UMERICAL E XPERIMENTS We now demonstrate the performance of the suggested beamformers under different hardware constraints. Unless mentioned otherwise, the channel model used is the narrowband mmWave clustered channel defined in (26), with αil ∼ N C (0, 1), Ncl = 6, and Nray = 1. We consider a simple uniform linear array (ULA) with spacing d = λ2 , where λ is the carrier wavelength. It follows that the steering vectors depend only on the azimuth and are given by 1 aq (φ) = √ [1, ej2π sin(φ),⋯, ej2π(Nq −1) sin(φ) ] , N (41) for q ∈ {r, t}. The angles φtil , φril are distributed uniformly over the interval [0, 2π). In all simulations, the number of transmit antennas is Nt = 10 and the number of receive antennas is Nr = 15. Unless mentioned otherwise, the interference is a white complexGaussian noise with zero mean and variance σz2 = 1. For the precoder experiments, we used a fully-digital receiver, i.e. r NRF = Nr , and for the combiner experiments we used a fullyt digital transmitter, i.e. NRF = Nt . A. Hybrid Precoder Performance First, we present the performance of different hybrid precoders. Here, we compare 5 design methods. The first two are the MO AltMin algorithm from [10], which performs manifold optimization to minimize (15) over F RF and F BB , and the PE AltMin method, also from [10], that assumes F BB to be a scaled orthogonal matrix and employs an algorithm similar to Algorithm 2,√ only with F BB calculated in the last step as F BB = ∥F RFNTS ∥F T . The third is the SOMP of [4], that reformulates (15) as a sparsity problem and chooses the columns of F RF from a dictionary of candidate vectors, and simultaneously solves the corresponding weights F BB . Last two are our methods, the MaGiQ in Algorithm 2 and the Alt-MaG method in Algorithm 1 with the MO AltMin solution used to obtain F RF , F BB in step 1. For the AltMin algorithms, we used the Matlab’s functions [25]. The dictionary used for the SOMP algorithm consists of 1000 1000 2π q)}q=1 . As performance measure steering vectors {at ( 1000 we consider the gap between the estimation error of the optimal precoder (14) and the hybrid one, defined by  − opt 2 with  = Ns1⋅Q ∑Q q=1 ∥ŝ − s∥ and Q the number of realizations. We begin by comparing MaGiQ and PE AltMin considering a fully-connected phase shifters network S2. As mentioned before, these two algorithms are very similar except for the last step of calculating F BB which is suboptimal in PE AltMin. This is demonstrated in Fig. 2 where MaGiQ’s lower MSE can be seen. In the next simulations, PE AltMin will not be tested. Figure 3 presents the estimation gap with respect to the number of RF chains for the different design methods. The SOMP algorithm suffers from a large performance gap, especially when the number of RF chains is much less than the t number of multipath components, i.e. NRF ≪ Nc . MaGiQ enjoys both low complexity and good performance due to the alternating minimization over T in (16). The MO-AltMin algorithm achieves low MSE at the cost of long running time. The gap between Alt-MaG and MO AltMin demonstrates the potential gain in optimizing over the matrix T in (16), in comparison to optimizing only F RF , F BB as in (15). This approach substantially improves the performance, with negligible increase in complexity. B. Hybrid Combiner Performance We now investigate the combiner’s performance. Here, an additional algorithm is tested, the GRTM in Algorithm 3, with the dictionary-based method in Section V-B as the base case solution. In contrast to the precoder case, here the SOMP algorithm solves the problem (30) as suggested in [4]. Figure 4 presents the estimation gap from the optimal combiner for all approaches with the fully-connected phase shifters scheme S2. It can be seen that MO-AltMin provides the best results. However, its runtime is an order-of-magnitude larger than the runtime of the other methods. Again, in the case of very few RF chains, MaGiQ is preferred over the greedy methods GRTM and SOMP, which suffer from large performance gap in that scenario. With additional RF chains, 11 10 -3 11 0.06 PE_AltMin MaGiQ 10 SOMP MO_AltMin MaGiQ GRTM 0.05 9 0.04 8 7 0.03 6 0.02 5 0.01 4 3 3 4 5 Fig. 2: Precoder estimation gap  − opt vs. number of RF t chains NRF , for MAGiQ and PE AltMin. 0 3 4 5 6 7 Fig. 4: Combiner estimation gap  − opt vs. number of RF t chains NRF , for a fully-connected phase shifters network S1. 0.035 SOMP MO_AltMin MaGiQ Alt-MaG+MO_AltMin 0.03 0.025 0.02 0.015 0.01 0.005 0 2 3 4 5 Fig. 3: Precoder estimation gap  − opt vs. number of RF t chains NRF , for a fully-connected phase shifters network S1. the greedy algorithms outperform the simple MaGiQ, with a slight advantage to GRTM. Next, we show that in some special cases, MaGiQ coincides with the fully-digital solution and outperforms other methods. We demonstrate this using the scenario from the asymptotic example given in Section IV. Here, H is the mmWave channel r (26), with Nc = 4, NRF = 4, and Nr = 150. In this case, MO-AltMin cannot be used due to its long runtime that grows with the number of antennas. Since Nc ≪ Nr the asymptotic analysis holds, and H has unimodular singular vectors. Figure 5 shows the estimation error  of the different algorithms. It can be seen that MaGiQ’s performance coincides with the fully-digital combiner. We used a steering dictionary for both SOMP and GRTM. Since the singular vectors of the channel also have steering structure, the greedy methods performance gap is small, but still exists. Next, we investigate the partially connected schemes. Figure 6 shows the performance of different combiners for the fixed sub-arrays network S4 and for the flexible one S5. For both scenarios, we set G = 5. We first notice the additional performance gain of the flexible architecture. This result is natural since the additional switches allow each RF chain to choose the antennas that contribute to it, in contrast to the fixed sub-arrays case, that enforces each RF chain to be connected to a small predetermined array. In both scenarios, the greedy methods outperform MaGiQ, but while SOMP yields better performance for the fixed arrays, when adding switches to allow for flexibility, GRTM is preferred. The fully-connected network S1, that involves switches as well as phase shifters is considered next. Here, we use a general channel model such that H has i.i.d complex-Gaussian entries and the interference z is now a complex-Gaussian vector with arbitrary full rank covariance matrix Rz . For the SOMP and GRTM algorithms, we used the dictionary W̄ = P W (X), with W the feasible set for scheme S1, as explained in Section V-B. Figure 7 presents the MSE  as a function of SNR. The complex MO-AltMin algorithm now achieves approximately the same performance as the low complexity GRTM. This is because MO-AltMin cannot use the additional flexibility of the switches and produces only unimodular beamformer vectors. This demonstrates a tradeoff between hardware and computational complexity: one may use the simpler architecture S2 while achieving the same performance as the complex network S1, with the cost of using a heavy computational algorithm such as MO-AltMin. It can be further noticed that SOMP’s performance falls short in comparison to others, possibly due to the different interference structure. C. Hardware Schemes Comparison The last simulation considered the different hardware choices. In Fig. 8, the performance of the MaGiQ combiner under different hardware constraints is demonstrated. As expected, the simple switching network S3 suffers from the largest MSE, as its connectivity is the most limited. The two partially connected cases S4 and S5 with G = 5 offer lower 12 1 Optimal SOMP MaGiQ GRTM 0.7 0.9 0.8 0.65 0.7 0.6 0.6 0.5 0.55 0.4 0.5 0 1 2 3 4 5 Fig. 5: Combiner estimation error vs. SNR for a fullyconnected phase shifters network S2 and mmWave channel with asymptotic number of antennas. 0.3 -5 Optimal S1-Fully Connected S2-Fully Connected Phase Only S3-Switching Network S4-Fixed Sub Arrays S5-Flexible Sub Arrays 0 5 10 Fig. 8: Combiner estimation error vs. SNR for a partially connected phase shifters network S4 with G = 3. MSE, due to the additional hardware complexity, where the second performs better as a result of the additional switches. The two fully-connected architectures S1 and S2 show similar performance, suggesting that the additional switches in the first offer negligible improvement. This surprising result can be explained by the channel model which implies that the optimal combiner consists of sums of steering vectors, that can be efficiently described using unimodular vectors. 0.9 0.85 0.8 0.75 Optimal SOMP_Partiall MaGiQ_Partiall GRTM_Partiall SOMP_Flexible MaGiQ_Flexible GRTM_Flexible 0.7 0.65 0.6 -2 -1 VII. C ONCLUSIONS 0 1 2 Fig. 6: Combiner estimation error vs. SNR for the partially connected phase shifters networks S4 and S5 with G = 5. 0.6 Optimal SOMP MO_AltMin MaGiQ GRTM 0.55 0.5 0.45 0.4 0.35 2 3 4 5 Fig. 7: Combiner estimation error vs. SNR for the fullyconnected phase shifters and switches network S1 with general channel model and interference. We developed a framework for hybrid precoder and combiner design, suitable for various channel models and hardware settings. We considered a data estimation problem and aimed at minimizing the estimation MSE over all possible hybrid precoders/combiners. For the precoder side, we suggested a family of iterative algorithms, Alt-MaG, that approximates the optimal fully-digital precoder while seeking the solution in the optimal set that results in the smallest approximation gap from the hybrid precoder of the previous iteration. The potential gain in exploiting the unitary degree of freedom in the fully-digital precoder was demonstrated in simulations. We further suggested a simple low complexity algorithm, MaGiQ, that achieves good approximation using a simple quantization function, and yields better performance than other methods suggested in the literature. We also showed that in some special cases it coincides with the optimal fully-digital solution. Next, we adjusted the MaGiQ algorithm so that it can be used for combiner design. We then suggested an additional greedy algorithm, GRTM, that directly minimizes the MSE using a simple scalar-ratio objective at each iteration. GRTM achieves lower MSE than MaGiQ when the number of RF chains increases. Experimental results showed that our lowcomplexity algorithms enjoy good performance in various scenarios. Finally, using simulations, we showed that in partially connected schemes and sparse multipath mmWave channels, adding switches to the analog network offers a large increase in 13 performance, while in fully-connected cases the improvement is negligible. R EFERENCES [1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, “What will 5g be?” IEEE Journal on Selected Areas in Communications, vol. 32, no. 6, pp. 1065–1082, Jun. 2014. [2] F. Rusek, D. Persson, Buon Kiong Lau, E. G. Larsson, T. L. Marzetta, and F. Tufvesson, “Scaling up MIMO: opportunities and challenges with very large arrays,” IEEE Signal Processing Magazine, vol. 30, no. 1, pp. 40–60, Jan. 2013. [3] T. S. Rappaport, Shu Sun, R. Mayzus, Hang Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5g cellular: it will work!” IEEE Access, vol. 1, pp. 335–349, 2013. [4] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Transactions on Wireless Communications, vol. 13, no. 3, pp. 1499– 1513, Mar. 2014. [5] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Hybrid precoding for millimeter wave cellular systems with partial channel knowledge,” in Information Theory and Applications Workshop (ITA), 2013. IEEE, 2013, pp. 1–5. [6] Y.-Y. Lee, C.-H. Wang, and Y.-H. Huang, “A hybrid RF/baseband precoding processor based on parallel-index-selection matrix-inversionbypass simultaneous orthogonal matching pursuit for millimeter wave MIMO systems,” IEEE Transactions on Signal Processing, vol. 63, no. 2, pp. 305–317, Jan. 2015. [7] S. Han, I. Chih-Lin, Z. Xu, and C. Rowell, “Large-scale antenna systems with hybrid analog and digital beamforming for millimeter wave 5g,” IEEE Communications Magazine, vol. 53, no. 1, pp. 186–194, 2015. [8] R. Mendez-Rial, C. Rusu, N. Gonzalez-Prelcic, A. Alkhateeb, and R. W. Heath, “Hybrid MIMO architectures for millimeter wave communications: phase shifters or switches?” IEEE Access, vol. 4, pp. 247–267, 2016. [9] M. Kim and Y. H. Lee, “MSE-based hybrid RF/baseband processing for millimeter-wave communication systems in MIMO interference channels,” IEEE Transactions on Vehicular Technology, vol. 64, no. 6, pp. 2714–2720, Jun. 2015. [10] X. Yu, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization algorithms for hybrid precoding in millimeter wave MIMO systems,” IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 485–500, Apr. 2016. [11] J. Singh and S. Ramakrishna, “On the feasibility of codebook-based beamforming in millimeter wave systems with multiple antenna arrays,” IEEE Transactions on Wireless Communications, vol. 14, no. 5, pp. 2670–2683, May 2015. [12] X. Yu, J. Zhang, and K. B. Letaief, “Partially-connected hybrid precoding in mm-Wave systems with dynamic phase shifter networks,” arXiv preprint arXiv:1705.00859, 2017. [13] S. Park, A. Alkhateeb, and R. W. Heath, “Dynamic subarrays for hybrid precoding in wideband mmWave MIMO systems,” IEEE Transactions on Wireless Communications, vol. 16, no. 5, pp. 2907–2920, May 2017. [14] Y. C. Eldar and G. Kutyniok, Compressed sensing: theory and applications. Cambridge University Press, 2012. [15] A. Scaglione, P. Stoica, S. Barbarossa, G. Giannakis, and H. Sampath, “Optimal designs for space-time linear precoders and decoders,” IEEE Transactions on Signal Processing, vol. 50, no. 5, pp. 1051–1064, May 2002. [16] S. X. Yu and J. Shi, “Multiclass spectral clustering,” in Proceedings of the Ninth IEEE International Conference on Computer Vision - Volume 2, ser. ICCV ’03. Washington, DC, USA: IEEE Computer Society, 2003. [17] A. Sayeed, “Deconstructing multiantenna fading channels,” IEEE Transactions on Signal Processing, vol. 50, no. 10, pp. 2563–2579, Oct. 2002. [18] P. F. M. Smulders and L. M. Correia, “Characterisation of propagation in 60 GHz radio channels,” Electronics & communication engineering journal, vol. 9, no. 2, pp. 73–80, 1997. [19] H. Xu, V. Kukshya, and T. S. Rappaport, “Spatial and temporal characteristics of 60-GHz indoor channels,” IEEE Journal on selected areas in communications, vol. 20, no. 3, pp. 620–630, 2002. [20] K. B. Petersen, M. S. Pedersen, J. Larsen, K. Strimmer, L. Christiansen, K. Hansen, L. He, L. Thibaut, M. Baro, S. Hattinger, V. Sima, and W. The, “The matrix cookbook,” Tech. Rep., 2006. [21] H. Wang, S. Yan, D. Xu, X. Tang, and T. Huang, “Trace ratio vs. ratio trace for dimensionality reduction,” in Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE Conference on. IEEE, 2007, pp. 1–8. [22] S. S. Ioushua and Y. C. Eldar, “Pilot contamination mitigation with reduced RF chains,” Signal Processing Advances in Wireless Communications (SPAWC), July 2017. [23] E. Bjornson and B. Ottersten, “A framework for training-based estimation in arbitrarily correlated rician MIMO channels with rician disturbance,” IEEE Transactions on Signal Processing, vol. 58, no. 3, pp. 1807–1820, Mar. 2010. [24] A. Tulino, A. Lozano, and S. Verdu, “Impact of Antenna Correlation on the Capacity of Multiantenna Channels,” IEEE Transactions on Information Theory, vol. 51, no. 7, pp. 2491–2509, Jul. 2005. [Online]. Available: http://ieeexplore.ieee.org/document/1459054/ [25] N. Boumal, B. Mishra, P. Absil, and R. Sepulchre, “Manopt, a matlab toolbox for optimization on manifolds,” J. Mach. Learn. Research, vol. 15, pp. 1455-C1459,, Jan. 2014.
7cs.IT
1 Interference Model Similarity Index and Its Applications to mmWave Networks: arXiv:1710.02659v1 [cs.IT] 7 Oct 2017 Extended version Hossein Shokri-Ghadikolaei, Student Member, IEEE, Carlo Fischione, Member, IEEE, and Eytan Modiano, Fellow, IEEE Abstract In wireless communication networks, interference models are routinely used for tasks such as performance analysis, optimization, and protocol design. These tasks are heavily affected by the accuracy and tractability of the interference models. Yet, quantifying the accuracy of these models remains a major challenge. In this paper, we propose a new index for assessing the accuracy of any interference model under any network scenario. Specifically, it is based on a new index that quantifies the ability of any interference model in correctly predicting harmful interference events, that is, link outages. We consider specific wireless scenario of both conventional sub-6 GHz and millimeter-wave (mmWave) networks and demonstrate how our index yields insights into the possibility of simplifying the set of dominant interferers, replacing a Nakagami or Rayleigh random fading by an equivalent deterministic channel, and ignoring antenna sidelobes. Our analysis reveals that in highly directional antenna settings with obstructions, even simple interference models (such as the classical protocol model) are accurate, while with omnidirectional antennas, more sophisticated and complex interference models (such as the classical physical model) are necessary. We further use the proposed index to develop a simple interference model for mmWave networks that can significantly simplify design principles of the important procedures for wireless communication, such as beamforming, interference management, scheduling, and topology H. Shokri-Ghadikolaei and C. Fischione are with KTH Royal Institute of Technology, Stockholm, Sweden (E-mails: {hshokri, carlofi}@kth.se). E. Modiano is with the Laboratory for Information and Decision Systems, Massachusetts Institute of Technology, Cambridge, USA (E-mail: [email protected]). A preliminary version of this paper [1] has been accepted for presentation at the IEEE International Conference on Communications (ICC), 2016. 2 control. Our new approach makes it possible to adopt the simplest interference model of adequate accuracy for every wireless network. Index Terms Wireless communications, interference model, performance analysis, millimeter wave networks. I. I NTRODUCTION Due to the shared nature of a wireless media, interference plays a critical role in the design and performance analysis of wireless networks, where the intended signal is combined with other undesired wireless signals transmitted at the same (time, frequency, spatial) channel. The receiver typically decodes the received signal by canceling parts of the interference and treating the rest as noise. Successful decoding at the receiver depends on the desired signal strength, the ambient noise level accumulated over the operating bandwidth, and the interference level. Signalto-interference-plus-noise ratio (SINR) is a common metric to evaluate the outage probability (or the probability of successful decoding) of a transmission. However, performance analysis using the SINR expression is complex as it depends on the transmission strategies (transmission power, antenna pattern, and medium access control (MAC) protocol), often unknown or hard to estimate random channel attenuation, receiver design, and the (often partially unknown) network topology. Due to this overwhelming complexity, the design and analysis of wireless networks based on the actual SINR expression, while being accurate, is very challenging. This difficulty is further exacerbated in millimeter-wave (mmWave) networks, where penetration loss, first-order reflection, and antenna pattern introduce further elements of randomness [2]–[4]. This motivates developing different techniques to mathematically model (abstract) various components of the SINR, e.g., the transmission strategy, wireless channel, and network topology. A. Related Works and Motivations Define an interference model as a set of deterministic or stochastic functions that model various components of the SINR expression. There have been many attempts in the literature to design interference models (equivalently, to approximate the SINR expression) that accurately capture the effect of interference, while being tractable for the mathematical analysis. These interference models largely try to answer the following questions under various network settings: 3 Q1. How can we model the set of interferer whose contributions in the aggregated interference term are dominant? Q2. How can we simplify the transmission/reception and propagation models to enhance tractability of the interference model with marginal loss in its accuracy? Answering Q1 demands a careful balance between the accuracy and the simplicity of the interference model. Considering the effects of more interferers in the SINR model generally increases the accuracy but also the complexity. In this regard, the simplest model is the primary interference model [5], wherein an outage event occurs only if two communication links share a common endpoint. In other words, the only interference component in this model is self-interference that leads to a half-duplex operating mode. Interference range model (IRM) is an attempt to improve the accuracy of the primary interference model [6], where an outage event occurs if the closest interferer is located no farther than a certain distance of the receiver, called the interference range. By setting this distance to 0, the IRM can be reduced to the primary interference model. A modified version of IRM is the protocol model (PRM), formalized by the seminal work of Gupta and Kumar [7]. The only modification is that the interference range, instead of being a constant value as in the IRM, depends on the received power from the intended transmitter and a minimum SINR threshold for successful decoding. Although the IRM and PRM are very simple, they fail to capture the effect of interference aggregation (i.e., the sum of the interference power from multiple interferers). It might be that, while there is no interferer inside the interference range, the aggregated interference from several transmitters outside the interference range downs the perceived SINR below the threshold. Thus, these models are generally considered to be overly simplistic. Nonetheless, due to their mathematical tractability, the IRM (including the primary interference model) and PRM are extensively adopted for the performance analysis and for the system design; e.g., transport capacity [7]–[9], delay [10], [11], fairness [12], throughput [13]– [15], topology control [16], [17], routing [18], and backoff design [19]. To alleviate the aforementioned problem of IRM and PRM, the interference ball model (IBM) considers the aggregated impacts of near-field interferers, located no farther than a certain distance. The price is higher complexity of the IBM compared to the IRM and PRM. Nonetheless, the IBM has been extensively adopted in the performance evaluation of wireless networks [4], [20]–[22]. The topological interference model (TIM) [23] is a natural extension of the IBM that 4 considers the aggregated impact of all the transmitters whose individual interference level at the receiver side is not below a certain threshold. In other words, this model neglects weak links based on the “topological” knowledge. The TIM is adopted for capacity and degreeof-freedom analysis [23], [24]. The most accurate and complex answer to Q1 is the physical model (PhyM) [7], which considers the aggregated interference of all transmitters in the entire network.1 The PhyM, also known as the SINR model, is adopted mostly at the physical layer; e.g., beamforming design [26]–[28], capacity evaluation [7], [29], [30], power control [31], [32], coverage analysis [4], energy efficiency characterization [33], and spectrum sharing [34]. The answer to Q2 depends heavily on the transmission and reception strategies and propagation environment. For instance, approximating the random wireless channel gain with its first moment (average) is a common technique to simplify the SINR expression and to design MAC and networking layers [14], [17], [21], [35]–[38]. Reference [39] replaced a Nakagami fading channel by a Rayleigh one for mathematical tractability and numerically concluded from its Fig. 5 that such approximation preserves the main properties of the rate coverage performance. Yet, the impact of these mathematical approximations on the accuracy of the performance analysis is not well understood. Recently, [40] considers the impact of such approximation on the scheduling. In particular, the authors show that, if we design scheduling for n transmitters based on a proper non-fading channel model (deterministic approximation of the random channel gain), the network throughput will be within O(log n) of that of the optimal scheduler, designed based on the actual random channel gains. This result, however, is limited to the Rayleigh fading model. As another example, for mmWave communications with many antenna elements, [14] and [41] assume no emissions from the antenna sidelobe, which affects the SINR distribution. This assumption is relaxed in [4], where the antenna sidelobe is modeled by a small constant value, adding further complexity into the interference model. As a result, the final derivations, while being more accurate, are less tractable and provide less insights. However, without having a mathematical framework that allows assessing the impact of neglecting antenna sidelobes, it is not clear which approach better balances the simplicity-accuracy tradeoff of mathematical analysis. 1 Under very special network settings (e.g., homogenous Poisson field of interferers exhibiting Rayleigh fading channel), the PhyM may be mathematically more tractable than both PRM and IBM [25]; however, the PRM and IBM are yet more desirable models for protocol design and for network optimization [22]. 5 The proper choice of interference model depends on many parameters such as the receiver design, antenna directionality, network topology, channel model, and the choice of medium access protocol [6], [35], [36]. To the best of our knowledge, there has been no systematic method to analyze the accuracy of various interference models, choose the proper interference model, and quantify the amount of error due to adopting other interference models for a given network scenario. The accuracy of different interference models has been mostly evaluated qualitatively, without fully understanding the mutual impacts of different parameters of the physical, medium access, and network layers. This qualitative analysis, however, is often overly simplistic, and may result in the use of interference models that are only marginally more accurate, yet significantly more complex than needed. As we will show throughout this paper, in certain settings of relevant practical interest, even the simplest interference models are sufficiently accurate and can be used to provide significant insights into the network performance and to enable efficient protocol design. B. Contributions In this paper, we substantially extend the preliminary version of this study [1] and propose a new framework to assess the distance of two arbitrary SINR distributions. We use this framework to develop an interference model similarity index that takes on real values between 0 and 1, where higher values correspond to higher similarity. This index builds a universal method to assess the accuracy of any interference model under any network scenario. In other words, instead of introducing a new interference model or a new approach to analyze SINR distribution, we propose a novel framework to investigate the accuracy of the existing interference models. Therefore, our study is complementary to the rich literature of interference analysis. To exemplify the abilities of the proposed index, we mathematically evaluate it for the PRM and IBM under three scenarios: (i) Rayleigh fading channel and omnidirectional communications (a typical sub-6 GHz system); and (ii) Rayleigh fading channel and directional communications; and (iii) deterministic wireless channel, directional communications, and existence of impenetrable obstacles in the environment (a typical mmWave system). Although the applications of the proposed index is general and goes beyond the examples provided in this paper, we use these examples to illustrate fundamental properties of this index and also to provide insights on the 6 mutual effects of various network parameters on the accuracy of the interference model, thus commenting on the proper model for a given network scenario. In the first example scenario, served as a baseline, we derive a closed-form expression for the accuracy index. We show that the accuracy of the IBM monotonically increases with the interference range, at the expense of an increased complexity. In contrast, we show that there is no such monotonic improvement in the accuracy of PRM. Thereby, we find the optimal interference range that maximizes the accuracy of the PRM. In the second example scenario, we show that both the PRM and IBM are significantly more accurate with directional antennas. Further, in the third example scenario, we show significant accuracy improvement of both PRM and IBM due to deterministic channel, directionality, and also blockage. As these conditions hold in mmWave networks, we show that the PRM can be used in the analysis of mmWave networks to significantly improve the mathematical tractability of the problem, with a negligible loss in the analysis accuracy. We further use this index to observe marginal impacts of the first-order reflection and sidelobe transmissions on the accuracy of the interference model, which inspire us to propose a tractable and accurate interference model for mmWave networks. Furthermore, we use the proposed framework to investigate the feasibility of modeling a random fading channel with a deterministic channel. We show that if the spatial distribution of the transmitters follow a Poisson point process on the plane and if the path-loss exponent is 2, then the average of the fading random variable2 is among the best constant approximations of the random fading channel to analyze any ergodic function of the SINR (e.g., transport capacity, throughput, and delay). Throughout the paper, we show how the proposed index can increase our understanding of the mutual interactions among the accuracy of the performance evaluation and various network parameters and modeling techniques. We also signify how we can rigorously develop simple interference models of adequate accuracy to simplify design principles of the main functions of wireless communications such as beamforming, interference management, scheduling, and topology control. 2 Rigorously speaking, the fading should be an absolutely continuous random variable, which holds for almost all wireless channels. 7 The rest of this paper is organized as follows. In Section II, we introduce our interference model similarity index, and investigate it under various network scenarios in Sections III–VI. Future works are presented in Section VII, and the paper is concluded in Section VIII. II. I NTERFERENCE M ODEL S IMILARITY I NDEX A. Interference Model We define a link as the pair of a transmitter and its intended receiver, where transmitter (receiver) i refers to the transmitter (receiver) of link i. Without loss of generality and for brevity, we assume that there is no interference cancellation, so all unintended transmitters act as potential interferers to any receiver. Consider a reference receiver and label its intended transmitter by subscript 0. Denote by I the set of its interferers (all active transmitters excluding the intended transmitter), by pi the transmission power of transmitter i, by σ the power of white Gaussian noise, by di the distance between transmitter i and the reference receiver, and by giCh the channel gain between transmitter i and the reference receiver. We denote by giTx the antenna gain at transmitter i toward the reference receiver, and by giRx the antenna gain at the reference receiver toward transmitter i. Then, the SINR at the reference receiver is p0 g0Txg0Ch g0Rx . γ= P pk gkTxgkCh gkRx + σ k∈I The SINR depends on the transmission powers, antenna patterns, set of active transmitters, channel model, and network topology. Let β > 0 denote the SINR threshold corresponding to a certain target bit error rate. An outage on the reference link occurs when γ < β. Different interference models attempt to approximate the outage probability by ignoring certain components of the interference (see questions Q1 and Q2 in Section I-A). In particular, the IRM, PRM, IBM, TIM, and PhyM characterize the set of interferers I. Neglecting various components of the channel model translates into different distributions for giCh . Power allocation affects pi , and various scheduling protocols further affect I. B. Formal Definition of the Similarity Index Consider reference interference model y under a given set of parameters/functions describing the wireless network. Define γ y as the SINR of a reference receiver under this model. We define a 8 binary hypothesis test, where hypotheses H0 and H1 denote the absence and presence of outage under reference model y, respectively. That is,   H0 , if γ y ≥ β ,  H , 1 (1) y if γ < β . We consider a test interference model x under any set of parameters/functions describing our wireless network, which are not necessarily equal to those of the reference model y. These differences result in possible deviation of the SINR of the reference receiver under x, denoted by γ x , from γ y . From the outage point of view, irrespective of the differences between individual parameters/functions of x and y, we say model x is similar to model y if it gives exactly the same outage result as y. Assume interference model x is a detector of outage events under y. To evaluate the performance of this detector compared to reference model y, we can use the notions of false alarm and miss-detection. A false alarm corresponds to the event that x predicts outage under hypothesis H0 (i.e., y declares no harmful interference); whereas a miss-detection corresponds to the event that x fails to predict outage under hypothesis H1 . Now, the performance of any interference model x can be evaluated using the false alarm and miss-detection probabilities, x|y x|y namely pfa and pmd . Formally, x|y pfa = Pr [γ x < β | γ y ≥ β] , x|y pmd = Pr [γ x ≥ β | γ y < β] . (2) The false alarm and miss-detection probabilities quantify the similarity of any interference model x in detecting outage events compared to any reference model y. Next, we define our index to be a convex combination of these probabilities. Definition 1 (Interference Model Similarity Index). For any constant 0 ≤ ξ ≤ 1, any SINR threshold β, any test interference model x, and any reference interference model y, we define similarity of x to y at β as     x|y x|y x|y x|y Sβ,ξ (xky) = ξ 1 − pfa + (1 − ξ) 1 − pmd = 1 − ξ pfa − (1 − ξ) pmd , x|y (3) x|y where pfa and pmd are given in (2). Notice that random variables γ x and γ y must have a common support. Sβ,ξ (xky) is a unit-less quantity ranging within [0, 1], where higher values represents higher similarity between x and y in capturing outage events at SINR threshold β. Setting ξ = Pr [γ y ≥ β], x|y x|y ξpfa +(1 − ξ) pmd is the average error in detecting the outage events; therefore, Sβ,Pr[γ y ≥β] (xky) 9 shows the probability that interference model x has similar decision as reference interference model y in detecting the outage events. Remark 1 (Accuracy of an Interference Model). Let reference model y perfectly capture the outage events in reality, namely the model y does not make any approximation/simplification. The accuracy of any interference model x is then Sβ,ξ (xky), and we call it the accuracy index throughout the paper. The proposed index is a universal metric that can be used to quantify the accuracy of any interference models, proposed in the literature, as we exemplify in the following sections. C. Comparison to the Existing Statistical Distance Measures Interference model similarity index, formulated in (3), is measuring the distance3 of the PDF of γ x compared to that of γ y . Let fX denote PDF of random variable X. In the following, we highlight three main advantages of using our index with respect to the existing standard distance measures, such as the Bhattacharyya distance and the Kullback-Leibler (KL) divergence [42]. First, the existing standard distance measures mostly map the distance between fγ x and fγ y in their entire support to only one real value. It might be that two distribution are very similar in the meaningful ranges of the SINR values (0–10 dB), but very different outside this range. Still, the classical statistical distance measures may result in a high distance between two distributions, as they compare fγ x to fγ y in the entire SINR range. This is indeed a misleading result that may mistakenly avoid the use of the simplified interference model x in practice. However, our similarity index allows us to investigate whether or not x is accurate at any given SINR threshold. Second, both the Bhattacharyya distance and the KL divergence may fail in a comparative analysis. In particular, fγ y might be more similar to fγ x than fγ z with point-wise comparison, but the Bhattacharyya distance and the KL divergence of fγ x from fγ y become higher than that of fγ z from fγ y , as shown in the following toy example. Example 1. Consider discrete random variables X, Y, and Z with common support of [1, 2, 3] with probability mass functions 3 Rigorously speaking, our similarity index is not a distance measure, as it does not satisfy the subadditivity property. Moreover, we are measuring the similarity, which could be in general a decreasing function of the distance. 10 t fX (t) fY (t) fZ (t) 1 0.05 0.1 0.25 2 0.25 0.45 0.2 3 0.7 0.45 0.55 Then, we have the following metrics (fX is the reference in the KL divergence): Distributions fX ,fY fX ,fZ Euclidean distance 0.324 0.255 Bhattacharyya distance 0.033 0.045 KL divergence 0.059 0.098 In this example, neither the Bhattacharyya distance nor the KL divergence can identify higher point-wise similarity of Z to X than Y to X. Last, but not least, unlike the existing statistical distance metrics that are not necessarily intended for communication systems, our similarity index is developed for these systems so that it has a physical meaning and can provide practical insights. Specifically, setting ξ = Pr [γ y ≥ β], our index Sβ,ξ (xky) evaluates the probability of correct decision of outage events under interference model x. Note that other distance metrics may still be useful to evaluate the accuracy of an interference model, and they may also have some relationship to our proposed index; see the following remark as an example. Remark 2 (Relationship to the Bhattacharyya Coefficient). Let ξ = Pr [γ y ≥ β]. By noting that Sβ,Pr[γ y ≥β] (xky) is the probability of having no hypothesis detection error and following [42, Equation (48)], we get r p 3 1 − ξ − ρ ξ (1 − ξ) ≤ Sβ,ξ (xky) ≤ 1 − ξ + − ξ (1 − ξ) ρ2 , 2 4 R where ρ = fγ x (t)fγ y (t)dt is the Bhattacharyya coefficient. (4) D. Applications of the Interference Model Similarity Index In the following, we provide two class of illustrative examples where our index can be used either to simplify the mathematical analysis or to justify the existing interference models. Use cases of our index, however, goes beyond these examples. 11 1) Simplifying the Set of Interferers: This is one of the first steps in choosing an interference model for performance analysis, protocol design, and network optimization. With omnidirectional transmission/reception and without interference cancelation, an outage occurs under • PRM: if there is an active transmitter no farther than an interference range rPRM = (1+∆)d0 , where ∆ is a constant real positive value [7]; • IBM: if its SINR due to all active transmitters located no farther than an interference range rIBM is less than β [22]; • TIM: if its SINR due to all active transmitters with strong links (with individual channel gains higher than ε) toward receiver i is less than β [23]; and • PhyM: if its SINR due to all active transmitters is less than β [7]. To present a unified view, we associate three random variables aPRM , aIBM , and aTIM to the link k k k between each transmitter k ∈ I and the typical receiver. aPRM is set +∞ if dk ≤ (1 + ∆)d0 , k and otherwise 0. aIBM is set 1 if dk ≤ rIBM , and otherwise 0. Finally, aTIM is set 1 if gkCh > ε, k k and otherwise 0. We define a virtual channel gain for those interference models as gkx = axk gkCh , for interference model x , (5) where x is a label denoting PRM, IBM, TIM, or PhyM, and aPhyM , 1. Despite the virtual i channel gain, all other parameters of interference models x and y are identical. The SINR at the typical receiver under interference model x is given by p0 g0Tx g0Ch g0Rx γx = P . pk gkTxgkx gkRx + σ (6) k∈I The design of many key functions of a wireless network such as scheduling [43] or power allocation [31] need an estimate of (6). To this end, a receiver may need to coordinate with a set of interferers to estimate their individual instantaneous contributions to the SINR expression, namely pk gkTxgkx gkRx for all k ∈ I. The PhyM may imply that every receiver should coordinate with all the interferers in the entire network (global information) whose cost, complexity, and delay may be unaffordable in many networking scenarios. Using IBM implies that each nodes should coordinate with all transmitters within a certain radius (local information), and the PRM necessitates coordination only with the closest unintended transmitter, which are appealing from energy and protocol overhead perspectives. Our proposed index gives a quantitative insight on the accuracy of various interference models, used for protocol development and for network 12 optimization, and allows the use of the right interference model for a given channel model and network scenario. 2) Simplifying the Channel Model: Our accuracy index can be used to adopt tractable channel models (gkCh for every transmitter k) of adequate accuracy. This is specially important for mmWave networks, where LoS and non-LoS conditions have different channel models, nonLoS (blockage) probability follows a rather complicated function, the LoS channel may follow a Nakagami fading in general, and realistic antenna patterns might be a complicated non-linear function. Various researches tried to simplify those complications without rigorous analysis on the validity of such simplifications. For instance, [14] assumed impenetrable obstacles (so communication only in the LoS conditions) and neglected antenna sidelobe, [4] approximated the non-LoS stochastic function by a deterministic LoS ball in which there is no obstacle within a certain range of the receiver and there is no LoS links outside the circle, and [39] replaced the Nakagami fading channel by a Rayleigh fading that facilitates mathematical analysis. Due to lack of a systematic approach to simplify the channel model, the understanding of the cross-layer dynamics between MAC and physical layers of most of the existing standards is a largely open problem, and the existing frameworks such as the one in [44] are not usually mathematically tractable. In the following, we illustrate the utility of our index for four example scenarios. Although our index poses no limitation to these example scenarios, we may simplify some parameters of the system model to avoid unnecessary complications. In the first three examples, we focus on simplifying the set of interferers for various network settings and derive closed-form expressions for the accuracy index to highlight its fundamental properties. In the last example scenario, we use our index to numerically assess the accuracy of various approaches in simplifying the channel model. For the rest of this paper, without loss of generality, we assume ξ = Pr [γ y ≥ β], so Sβ,ξ (xky) evaluates the probability of correct decision under interference model x. 13 III. E XAMPLE S CENARIO 1: R AYLEIGH FADING C HANNEL WITH O MNIDIRECTIONAL C OMMUNICATIONS Consider a wireless network with Rayleigh fading channel and omnidirectional transmission/reception. Assume that the PhyM can perfectly capture the outage events. In this section, we evaluate the accuracy of IBM, PRM, and TIM (see Section II-D where we recalled the definition of these prominent models) for such scenario. We consider a reference receiver (called the typical receiver) at the origin of the Polar coordinate, and its intended transmitter having geometrical/spatial length d0 . We consider a homogeneous Poisson network of interferers (unintended transmitters) on the plane with intensity λt . We assume that all the transmitters are active with transmission power p (no power control), and that there is no interference cancellation, which are natural assumptions in personal and local area networks. With omnidirectional transmission and reception, there is no antenna gains, so gkTx = gkRx = 1, k ∈ I ∪ {0}. Note that, under these assumptions, the PhyM is more tractable for coverage and rate analyses than other models (PRM, IBM, and TIM) [25]; however, we still use this example to derive closed-form expression for the new accuracy index and thereby illustrate its fundamental properties that hold in general. Nonetheless, even in this network setting, the PRM and IBM are more appealing than PhyM for protocol design and for network optimization [22]. We define by B(θ, rin , rout ) a geometrical annulus sector with angle θ, inner radius rin , and outer radius rout , centered at the location of the typical receiver (origin of the Polar coordinate). To model a wireless channel, we consider a constant attenuation c at reference distance 1 m, a distance-dependent attenuation with exponent α, and a Rayleigh fading component h. To avoid the physically unreasonable singularity that arises at the origin under power law attenuation, we change the path loss index to α1B(2π,0,a) , where 1· is the indicator function assuming value 1 over set · and zero otherwise. This modified power law model implies that the signal of all transmitters located outside a disk with radius a will be attenuated by traditional power law method; however, the transmitters inside this disk will observe no channel attenuation. Therefore, the channel gain −α1B(2π,0,a) between transmitter i at radial distance di and the typical receiver is giCh = chi di . To avoid unnecessary complications while illustrating the utility of our index, we eliminate the shadow fading from our channel model. We are now ready to illustrate the utility of our proposed index using the SINR expression (6). 14 A. Accuracy of the Interference Ball Model For mathematical tractability, we assume that rIBM ≥ a and d0 ≥ a, and the extension to the general case is straightforward. The false alarm probability can be reformulated as      IBM  Pr γ IBM < β Pr γ PhyM ≥ β | γ IBM < β IBM|PhyM PhyM . (7) pfa = Pr γ <β|γ ≥β = 1 − Pr [γ PhyM < β] Although the PhyM considers the impacts of all the interferers in the entire network, the IBM considers only the effects of the near-field ones. Consequently, γ PhyM ≤ γ IBM , and thus   IBM|PhyM Pr γ PhyM ≥ β | γ IBM < β = 0 in the nominator of (7). This results in pfa = 0. For the miss-detection probability, we have     IBM|PhyM pmd = Pr γ IBM ≥ β | γ PhyM < β = 1 − Pr γ IBM < β | γ PhyM < β       Pr γ IBM < β Pr γ PhyM < β | γ IBM < β Pr γ IBM < β =1− = 1− , (8) Pr [γ PhyM < β] Pr [γ PhyM < β] where the last equality is  from γ PhyM ≤ γ IBM . In Appendix A, we have derived       α   −α −σβd0 α  α 2 − πλt Eh a2 1 − e−βd0 h + rIBM 1 − e−βd0 hrIBM − Pr γ IBM < β = 1 − exp  pc             2 2 −α 2/α 2/α α α −α α −α α −βdα ha 2  0 , + (βd0 h) Γ 1 − , βd0 hrIBM − (βd0 h) Γ 1 − , βd0 ha a 1−e  α α   (9) and  Pr γ PhyM     σβdα0     −βdα ha−α 2 −βdα h 2 0  0 < β = 1 − exp − − πλt Eh a 1 − e −a 1−e  pc           2 2 α 2/α α −α  α 2/α − (βd0 h) Γ 1 − , βd0 ha , (10) + (βd0 h) Γ 1 −  α α    where Γ (·, ·) is the incomplete Gamma function, Γ (·) is the Gamma function, Eh denotes expectation over random variable h, and the probability density function of h is fh (x) = e−x . Substituting (9) and (10) into (8), the miss-detection probability can be found. Also, from (3), the accuracy of the interference ball model Sβ,ξ (IBMkPhyM) is derived. A simple extension of our analysis gives the accuracy index when d0 is a random variable. Recall that the purpose of this section is to illustrate only the utility of our index, and investigating more practical system models is a subject of our future work; see for instance [45]. 15 Result 1 (Perfect Interference Ball Model). For any constant 0 ≤ ξ ≤ 1 and any β, Sβ,ξ (IBMkPhyM) → 1 as rIBM → ∞. IBM|PhyM Proof: We know that pfa = 0 for any constant 0 ≤ ξ ≤ 1 and any β. Moreover, as rIBM  IBM    IBM|PhyM increases, Pr γ < β tends to Pr γ PhyM < β . Considering (8), Pmd asymptotically goes to zero as rIBM → ∞. With zero false alarm and asymptotically zero miss-detection probabilities, the proof is concluded from (3). Result 1 indicates that the IBM becomes more accurate with higher rIBM , and it can be arbitrary accurate for sufficiently large rIBM . The price, however, is more complicated IBM as its approximations at a receiver demands coordination with more interferers.4 Also, negotiation with other transmitters (e.g., for MAC layer design) within this larger rIBM becomes more challenging in terms of power consumption, signaling overhead, delay, and processing overhead. B. Accuracy of the Protocol Model We now consider the PRM and first note that     1 − Pr γ PRM < β 1 − Pr γ PhyM < β | γ PRM ≥ β PRM|PhyM , (11) pfa =1− 1 − Pr [γ PhyM < β] and that     1 − Pr γ PRM < β Pr γ PhyM < β | γ PRM ≥ β PRM|PhyM pmd = . (12) Pr [γ PhyM < β] In the last two equations, note that Pr[γ PhyM < β] is derived in (10). In the following, we   derive Pr[γ PRM < β] and Pr γ PhyM < β | γ PRM ≥ β . Event γ PRM < β occurs if there is at least one interferer inside B(2π, 0, rPRM ). As I is a homogenous Poisson point process with intensity λt , we have    2 Pr γ PRM < β = 1 − exp −λt πrPRM 4 . (13) Note that for special settings of this section, considering the impact of all interferers (PhyM) simplifies the analysis. However, this does not hold in general, e.g., if we change the spatial distribution of the interferers to a determinantal point process. 16 In Appendix A, we have also derived      −σβdα   −α α 0 2 − πλt Eh  − rPRM 1 − e−βd0 hrPRM Pr[γ PhyM < β | γ PRM ≥ β] = 1 − exp  pc            2 2 α 2/α α 2/α α −α + (βd0 h) Γ 1 − − (βd0 h) Γ 1 − , βd0 hrPRM  . (14)  α α   Substituting (11)–(14) into (3), we can find Sβ,Pr[γ PhyM ≥β ] (PRMkPhyM) for Rayleigh fading channel with omnidirectional transmission/reception. Result 2 (Miss-detection–False Alarm Tradeoff). Consider the protocol model of interference with Rayleigh fading channel. Increasing the interference range rPRM reduces the false alarm probability and increases the miss-detection probability. Decreasing the interference range increases the false alarm probability and reduces the miss-detection probability.   Proof: Pr γ PRM < β is a strictly increasing function of rPRM , see (13). Considering the equations of the false alarm and miss-detection probabilities given in (11) and (12), the proof concludes. Result 3 (Asymptotic Accuracy of the Protocol Model). Consider Equations (3) and (11)–(13). For any 0 ≤ ξ ≤ 1 and any β > 0, we have the following asymptotic results: PRM|PhyM → 0 , pmd PRM|PhyM → 1 , pmd rPRM → a, a → 0 ⇒ pfa rPRM → ∞ ⇒ pfa PRM|PhyM → 1 , Sβ,ξ (PRMkPhyM) → ξ . PRM|PhyM → 0 , Sβ,ξ (PRMkPhyM) → 1 − ξ . Result 3 further confirms the tradeoff between the miss-detection and false alarm probabilities. C. Numerical Illustrations To illustrate the accuracy index in Scenario 1 with Monte Carlo simulation, we consider a spatial Poisson network of interferers and obstacles with density λt and λo per unit area. Length of the typical link is d0 = 20 m. We simulate a traditional outdoor microwave network [4] with average attenuation c = 22.7 dB at the reference distance a = 1 m, path-loss index α = 3.6, and noise power σ = −111 dBm (around 2 MHz bandwidth). We consider p = 20 dBm transmission power and β = 5 dB minimum SINR threshold. For the ease of illustration, we define the notion 17 1 0.9 PRM: Pfa , dt = 30 0.8 PRM: Pfa , dt = 80 0.6 Accuracy index Detection performance 1 PRM: Pmd , dt = 30 PRM: Pmd , dt = 80 0.4 IBM: Pmd , dt = 80 0.2 0 20 PRM: dt = 30 0.8 IBM: dt = 80 0.7 PRM: dt = 80 0.6 0.5 40 60 80 100 Interference range [m] 120 140 0.4 20 40 60 80 100 Interference range [m] (a) Error probabilities 120 140 (b) Accuracy index Fig. 1: Impact of the interference range on the accuracy of interference models under Rayleigh fading channel and omnidirectional communications. √ of the average inter-transmitter distance as dt = 1/ λt . This distance directly relates to the inter-site distance in cellular networks, and also shows the transmitter density in a network. Fig. 1 illustrates the impact of the interference range on the accuracy of both IBM and PRM PRM|PhyM under Scenario 1. From Fig. 1(a), increasing rPRM increases pfa PRM|PhyM and reduces pmd , highlighted as the tradeoff between the miss-detection and false alarm probabilities in Result 2. This tradeoff may lead to increment (see dt = 30) or decrement (see dt = 80) of the accuracy index of the PRM with the interference range. The IBM has zero false alarm probability, not IBM|PhyM depicted in Fig. 1(a) for sake of clarity of the figure. Moreover, as stated in Result 1, pmd decreases with rPRM , leading to a more accurate IBM, as can be confirmed in Fig. 1(b). Note that with the same transmitter density and interference range, the PRM has lower miss-detection probability than the IBM; however, better false alarm performance of the IBM leads to less errors in detecting outage events and therefore higher accuracy index, see Fig. 1(b). The TIM, not depicted in the figure, has a very high accuracy in all simulations. In particular, with ε = −130 dB, its accuracy is about 0.99. However, the corresponding TIM considers many interferers inside an irregular geometrical shape, which substantially decreases the tractability of the resulting interference model. Fig. 2 shows the accuracy of the IBM and PRM under Scenario 1 against the average intertransmitter distance. Again, we can observe an enhancement in the accuracy of the IBM with rIBM , whereas the accuracy index of the PRM shows a complicated behavior as a function of rPRM . By adopting the optimal rPRM that maximizes the accuracy index, as shown in Fig. 2(b), 18 1 1 Accuracy index Accuracy index 0.9 0.8 0.7 rIBM = 20 rIBM = 60 0.6 0.8 0.6 rPRM = 20 rPRM = 60 rPRM = 120 optimal rPRM 0.4 rIBM = 120 0.5 0 50 100 150 Average inter-transmitter distance [m] (a) Interference ball model 200 0.2 0 50 100 150 Average inter-transmitter distance [m] 200 (b) Protocol model of interference Fig. 2: Impact of transmitter density on the accuracy of the interference models under Rayleigh fading channel and omnidirectional communications. The accuracy of the TIM with ε = −130 dB is higher than 0.98. we can maintain a good performance for the PRM. Both interference models are very accurate at extremely dense transmitter deployments. The main reason is the very high interference level   (ξ = Pr γ PhyM ≥ β is almost 0 in this case), implying that the accuracy index is determined only by the miss-detection probability. Increasing the transmitter density through reducing dt decreases the miss-detection probability for both IBM and PRM, see Fig. 1(a), improving their accuracy. For ultra sparse transmitter deployments, again, both interference models work accurately, as ξ goes to 1 in this case and therefore only the false alarm probability determines the accuracy index. This probability is zero for the IBM, and it gets smaller values (asymptotically zero) for the PRM with higher dt , see Fig. 1(a). Finally, the TIM with ε = −130 dB, not shown in Fig.2, has a very high accuracy in modeling the interference. Its accuracy for the same ranges of dt is higher than 0.98. Fig. 3 shows the KL divergence of fγ IBM (x) from fγ PhyM (x) and also their Bhattacharyya distance for the same setting of Fig. 2(a), where lower values translates into higher accuracy of the IBM. From this figure, both the KL divergence and the Bhattacharyya distance can identify higher accuracy of the IBM with rIBM = 60 m. However, they both fail to show that the performance of IBM with rIBM = 20 m converges to that with rIBM = 60 m once the network gets sparser. Moreover, calculating these measures entails almost the same mathematical/numerical complexity as our similarity index. Due to these reasons, we investigate only our accuracy index for the rest of the paper, though one may incorporate those metrics in our proposed interference model similarity analysis framework. 19 0.08 KL: rIBM KL: rIBM BD: rIBM BD: rIBM Distance 0.06 = 20 = 60 = 20 = 60 0.04 0.02 0 0 50 100 150 Average inter-transmitter distance [m] 200 Fig. 3: The KL divergence (labeled by “KL”) of the distribution of γ IBM from that of γ PhyM and their Bhattacharyya distance (labeled by “BD”) corresponding to the accuracy index values of Fig. 2(a). Lower values translate into higher similarity between two distributions. 1 Accuracy index 0.8 0.6 0.4 0.2 0 -10 IBM: IBM: PRM: PRM: -5 dt dt dt dt = 50 = 100 = 50 = 100 0 5 10 SINR threshold [dB] 15 20 Fig. 4: Impact of the SINR threshold on the accuracy of interference models under Rayleigh fading channel and omnidirectional communications. Fig. 4 illustrates the accuracy index against the SINR threshold. Increasing the SINR threshold generally increases the sensitivity of the interference model to any approximation error in x. IV. E XAMPLE S CENARIO 2: R AYLEIGH FADING C HANNEL , D IRECTIONALITY, AND O BSTACLES In this section, we analyze the accuracy of IBM and PRM in modeling a wireless network with Rayleigh fading channels, where all transmitters and receivers use directional communications to boost the link budget and to reduce multiuser interference. We also consider impenetrable 20 obstacles. The application areas of this scenario include modeling and performance evaluation of mmWave networks, where directional communication is inevitable and extreme penetration loss due to most of the solid materials (e.g., 20–35 dB due to the human body [46]) justifies the impenetrable obstacle assumption. In Section VI-B, we will comment on the impact of assuming impenetrable obstacles on the accuracy of the interference model. Note that the interference is not the primary limitation of mmWave networks specially if we take an average over all possible realizations of a random topology [4], [41]. However, even if mmWave networks are noise-limited in a statistical sense (that is, taking an average of the interference over some time or some topologies), there are significant realizations of network topologies at given times where some transmitters can cause strong interference. We cannot use noise-limited arguments, which are valid over some time horizons, when we have to optimize in real-time resource allocations or routing. In the following two sections, we show that special characteristics of mmWave networks, such as blockage and deafness, can be exploited to substantially simplify the interference model, so as to develop efficient scheduling and routing algorithms, which may otherwise be impossible. In fact, our results provide, for the first time, mathematical justifications for the use of simpler interference models in mmWave networks, as extensively done in the literature [14], [15], [17], [47]–[51]. We assume a homogenous Poisson network of interferers as in Section III. If there is no obstacle on the link between transmitter i and the typical receiver located at the origin, we say that transmitter i has line-of-sight (LoS) condition with respect to the typical receiver, otherwise it is in non-LoS condition. We assume that transmitter of every link is spatially aligned with its intended receiver, so there is no beam-searching phase [52]. We model the antenna pattern by an ideal sector model [4], where the antenna gain is a constant in the main lobe and another smaller constant in the side lobe. We assume the same operating beamwidth θ for all devices in both transmission and reception modes. Then, the antenna gain for each transmitter/receiver is [52, Equation (3)]    2π − (2π − θ) z , θ  z, inside the main lobe (15) inside the side lobe , where 0 ≤ z ≪ 1 is the side lobe gain. For mathematical tractability, we assume negligible side lobe gain (i.e., z = 0) throughout this section, and numerically assess the impact of this 21 simplification in Section VI-B. Consider the link between transmitter i and receiver j with distance dij . It is shown that with a random number of obstacles, each having random location and size, this link is in the LoS condition with probability e−ǫλo dij , where λo is the intensity of the obstacles and ǫ is a constant value that depends on the average size of obstacles in the environment [53]. Due to the exponential decrease of the LoS probability with the link length (also see [54, Fig. 4]), very far interferers are most likely blocked. For mathematical simplicity, we assume independent LoS conditions among the typical receiver and all other transmitters, and also impenetrable obstacles. Nonetheless, the following analysis can be extended for more realistic blockage models, introduced in [41]. Notice that we are using this simplified model to investigate the effects of directionality and blockage on the accuracy of the interference models and to characterize fundamental properties of the proposed accuracy index. The exact value of the accuracy index with a more realistic mmWave channel can be readily numerically calculated under any system model, as we highlight in the next sections. To evaluate the accuracy of IBM and PRM, we first notice that an intended transmitter can cause a significant interference contribution to the typical receiver if: (a) the typical receiver is inside its main lobe, (b) it has LoS condition with respect to the typical receiver, and (c) it is inside the main lobe of the typical receiver. Due to random deployment of the transmitters and receivers, the probability that the typical receiver locates inside the main lobe of a transmitter is θ/2π. Moreover, we have independent LoS events among the typical receiver and individual transmitters. Therefore, the interferers for which conditions (a)–(b) hold follow an inhomogeneous Poisson point process I with intensity of λI (r) = λt θe−ǫλo r /2π at radial distance r. Condition (c) reduces the angular region that a potential interferer should be located to contribute in the interference observed by the typical receiver. We note that I ∩ B(θ, 0, rPRM ) is the set of potential interferers inside the vulnerable region of the PRM, shown by red triangles in Fig. 5, and I ∩B(θ, rPRM , ∞) shows the set of potential interferers outside that region, shown by green circles in Fig. 5. Also, I ∩ B(θ, 0, rIBM ) is the set of potential interferers for IBM (near-field interferers). 22 U350 Fig. 5: Illustration of the vulnerable area. A. Impact of Directionality and Blockage Before deriving the accuracy of IBM and PRM, we first evaluate the impact of directionality and blockage on the number of the interferers. We define by ΛB(θ,0,R) the measure of the region B(θ, 0, R), i.e., the average number of interferers inside the region. We have Z R  θ 2 λt  −ǫλo R ΛB(θ,0,R) = θ λI (r)r dr = 1 − (1 + ǫλ R) e . (16) o 2πǫ2 λ2o 0 Then, for any real R > 0, the number of potential interferer inside the region B(θ, 0, R), denoted by NB(θ,0,R) , is a Poisson random variable with probability mass function n Λ B(θ,0,R) Pr[NB(θ,0,R) = n] = e−ΛB(θ,0,R) . (17) n! Result 4 (Impact of Directionality). Consider (16), and let ǫλo → 0. The average number of potential interferers converges to    θ 2 λt 2 θ θ 2 . (18) R = λt R 4π 2π 2 To interpret Result 4, with no obstacle in the environment (ǫλo → 0), we will have a homogenous Poisson network of interferers with density λt θ/2π. Therefore, the average number of interferers over B(θ, 0, R) is the product of the density per unit area and the area of B, which is θR2 /2. It can be concluded that adopting narrower beams reduces the average number of potential interferers within a certain distance R; however, it still tends to infinity almost surely as R → ∞. Result 5 (Impact of Blockage). Consider (16), and let R → ∞. The average number of potential interferers converges to θ 2 λt , 2πǫ2 λ2o (19) which is less than infinity almost surely if ǫλo > 0. Result 5 implies that any receiver observes a finite number of potential interferers almost surely if there is a non-negligible blockage. This unique feature holds for the mmWave bands, as most of 23 the obstacles can severely attenuate the signals.5 Therefore, not only the farther transmitters will contribute less on the aggregated interference (due to higher path-loss), but they will be also thinned by directionality and blockage such that only a finite number of spatially close transmitters can cause non-negligible interference to any receiver. Note that, these fewer interferers may still cause strong interference, if they are located very close to the receiver. The point is that the thinning process due to directionality and blockage makes the SINR distribution under PhyM closer to that of the IBM, which considers only the near-field interferers. To elaborate more, we characterize the average number of far-field interferers in the following. Proposition 1 (Measure of Far-Field Interferers). Let θ be the operating beamwidth, λt be the density of the transmitters, λo be the density of the obstacles, and ǫ > 0 be a constant. Then, the average number of interferers located inside B(θ, R, ∞) is θ 2 λt (1 + ǫλo R) e−ǫλo R , ΛB(θ,R,∞) = 2 2 2πǫ λo and the probability of having no far-field interferer is (20) Pr[NB(θ,R,∞) = 0] = e−ΛB(θ,R,∞) . (21) R∞ Proof: To prove, we only need to compute R θλI (r) r dr, and (20) follows. Moreover, by substituting ΛB(θ,R,∞) into (17) with n = 0, we conclude (21). From Proposition 1, the average number of far-field interferers will be decreased exponentially with distance. Consequently, from (21), the probability of having no far-field interferers increases exponentially with the distance. Fig. 6 shows the probability of having at least one far-field interferer as a function of the distance. By defining any arbitrary minimum threshold κ on this probability, we can find a distance Rκ after which the probability of having far-field interferer(s) is arbitrarily close to 0 (less than κ). This suggests that by setting rIBM = Rκ , IBM can capture, at least, (1 − κ) fraction of the total number of interferers for any arbitrary small κ. Recall that the neglected interferers, if any, are far-field and their contributions to the total interference term are suppressed by the significant distance-dependent path-loss. All these facts result in the following conclusion: 5 In the conventional microwave systems where the transmission is less sensitive to blockage, the number of potential interferers is almost surely infinite, as highlighted in Result 4. 24 Probability of having far−field interferers 0 10 −1 10 −2 10 c = 0.5, λo=0.25, λt=1 c = 0.5, λo=0.25, λt=0.44 −3 10 c = 0.5, λo=0.44, λt=0.44 c = 1, λo=0.44, λt=0.44 −4 10 0 20 40 60 Distance [meters] 80 100 Fig. 6: Probability of having at least one far-field interferer as a function of distance. Simulation parameters are similar to those of Fig. 1. Result 6. Directionality and blockage can substantially increase the accuracy of the interference ball model. We can argue similar accuracy improvement in the PRM, as we numerically illustrate in the next subsections. B. Accuracy of the Interference Ball Model Assume rIBM ≥ a and d0 ≥ a. Using similar claims as in Section III-A, it is straightforward to IBM|PhyM show pfa  Pr γ IBM = 0. To findthe miss-detection probability, in Appendix B, we derive       2 α 2   1 − (ǫλo a + 1) e−ǫλo a σθ βd0 θ −βdα h  0 < β = 1 − exp − − λt E h 1 − e  4pcπ 2 2π ǫ2 λ2o      Z rIBM    −α −ǫλ r −βdα hr o + e r dr  , 1−e 0  a   (22) 25 and   Pr γ PhyM < β = 1 − exp        2 − σθ βdα0 4pcπ 2 2 −   θ λt 1 − e−βdα0 h 1 − (ǫλo a + 1) e−ǫλo a E h 2πǫ2 λ2o    Z ∞  α −α 2 2 −βd0 hr −ǫλo r  − ǫ λo e r dr , (23)  a   and substitute them into (8). Then, Sβ,ξ (IBMkPhyM) can be found using (22), (23), (8), and then (3). Similar to Remark 1, for any 0 ≤ ξ ≤ 1, Sβ,ξ (IBMkPhyM) → 1 as rIBM → ∞. C. Accuracy of the Protocol Model To derive the accuracy of the PRM, we need to derive Pr[γ PhyM < β], Pr[γ PRM ≤ β] and Pr[γ PhyM < β | γ PRM ≥ β], and substitute them into (11) and (12). Pr[γ PhyM < β] is derived in (23). Event γ PRM < β implies that |I ∩ B(θ, 0, rPRM )| ≥ 1, namely there is at least one potential interferer inside B(θ, 0, rPRM ). Considering (17), the probability of this event is Pr[NB(θ,0,rPRM ) ≥ 1], thus n o   Pr γ PRM < β = 1 − exp − ΛB(θ,0,rPRM) , (24) Event γ PRM ≥ β implies that there is no interferer inside B(θ, 0, rPRM ). Assuming rPRM ≥ a, it is easy to find Pr[γ PhyM < β | γ PRM ≥ β]:       Z ∞  σθ2 βdα θ2 λ   α hr −α t 0 PhyM PRM −βd −ǫλ r o   0 . − E Pr[γ <β|γ ≥ β] = 1−exp − 1 − e e r dr h   4pcπ 2 2π rPRM     (25) Substituting (23)–(25) into (11) and (12) gives the accuracy index of the PRM. Note that Results 2 and 3 hold here. D. Numerical Illustrations To numerically illustrate the accuracy index in Scenario 2, we use the same simulation environment of Section III-C. We independently randomly mark some wireless link to be blocked by obstacles, with the exponential blockage probability with ǫλo = 0.008 [53]. We then assume infinite penetration loss for the blocked links, and use the large scale LoS path loss model at 26 Accuracy index 1 0.98 0.96 IBM: IBM: PRM: PRM: 0.94 0.92 20 θ = 20o θ = 40o θ = 20o θ = 40o 40 60 80 Average inter-transmitter distance [m] 100 Fig. 7: Accuracy of IBM and PRM under Rayleigh fading channel and directional communications with obstruction. 28 GHz [2, Table I]. System bandwidth is 1 GHz (noise power σ = −84 dBm). Without loss of generality, we assume rPRM = 40 m and rIBM = 80 m. Fig. 7 illustrates the impact of the operating bandwidth and average inter-transmitter distance on the accuracy index of both IBM and PRM under Scenario 2. As expected, the IBM outperforms PRM. More importantly, directionality and blockage improve the accuracy of both interference models. We show in the following section that changing the underlying channel model from a Rayleigh fading model to a deterministic model further enhances their accuracies. Moreover, the accuracy of the TIM with ε = −130 dB, not depicted for the sake of clarity in Fig. 7, is nearly 1 in our simulations. Notice that a simplified interference model (e.g., PRM, IBM, or TIM) may not be of sufficient accuracy for all range of parameters, still it is substantially improved by directionality and blockage, as highlighted by Results 4–6. V. E XAMPLE S CENARIO 3: D ETERMINISTIC C HANNEL , D IRECTIONALITY, AND O BSTACLES In this section, we investigate how accurately the IBM and PRM can model a wireless network with directional communications, blockage, and deterministic wireless channel. The last assumption holds generally in mmWave networks, where sparse scattering characteristic of mmWave frequencies along with narrow beam operation makes the mmWave channel more deterministic compared to that of microwave systems with rich scattering environment and omnidirectional operation [54]. 27 A. Accuracy of the Interference Ball Model IBM|PhyM Again, it is straightforward to show pfa = 0. However, unlike previous cases, we cannot derive closed-form expression for the miss-detection probability, and consequently for the accuracy index. In Appendix C, we have derived upper bounds on the miss-detection probability using the Chernoff bound. B. Accuracy of the Protocol Model Again, deterministic wireless channel prohibits deriving closed-form expressions for the false alarm and miss-detection probabilities. Nevertheless, we can show that both Remarks 2 and 3 holds here. Moreover, we have the following result: Result 7 (Zero False Alarm Probability). Under the deterministic channel model, the false alarm probability is zero for any rPRM ≤ ζ −1/α , where  2 θ σ d−α 0 − . (26) ζ= β pc 2π Proof: The SINR that the typical receiver experiences due to transmission of the intended transmitter and an unintended receiver located at distance rPRM is 2 −α d0 pc 2π θ .  2 −α pc 2π r + σ PRM θ Comparing the SINR expression to β, we get that any interferer located at distance rPRM less than σ d−α 0 − β pc  θ 2π 2 !−1/α = ζ −1/α can cause packet loss at the typical receiver, namely γ PhyM < β. Now, if we consider the general equation of the false alarm probability, we have      PRM  Pr γ PRM < β Pr γ PhyM ≥ β | γ PRM < β PRM|PhyM PhyM pfa = Pr γ <β|γ ≥β = . 1 − Pr [γ PhyM < β]   For rPRM ≤ ζ −1/α , Pr γ PhyM ≥ β | γ PRM < β = 0; since event γ PRM < β implies that there is PRM|PhyM at least one interferer inside B(θ, 0, rPRM ). This interferer ensures γ PhyM < β, so pfa =0 for rPRM ≤ ζ −1/α . The following proposition characterizes bounds for the accuracy index for the Example Scenario 3 (mmWave networks):   Proposition 2. For ξ = Pr γ PhyM ≥ β and any 0 < rPRM ≤ ζ −1/α , we have   Pr γ PRM < β ≤ Sβ,ξ (PRMkPhyM) ≤ 1 , 28   where Pr γ PRM < β is given in (24). We have provided a proof for this proposition along with other bounds in Appendix C. We have the following scaling law results: Result 8 (Scaling laws for the PRM). The following scaling laws are implied by Proposition 2 and inequality ex ≥ 1 + x for any x ≥ 0: • Scaling with θ: For any constant rPRM no larger than ζ −1/α , limθ→0 Sβ,ξ (PRMkPhyM) ≥ 2 1 − e−θ C , for some constant C ≥ 0. • Scaling with λt : For any constant rPRM no larger than ζ −1/α , limλt →∞ Sβ,ξ (PRMkPhyM) ≥ 1 − e−λt C for some constant C ≥ 0. • Scaling with λo : For any constant rPRM no larger than ζ −1/α , limλo →0 Sβ,ξ (PRMkPhyM) ≥ −2 1 − exp{−C} for some constant C ≥ 0, and limλo →∞ Sβ,ξ (PRMkPhyM) ≥ 1 − e−λo D for some constant D ≥ 0. Due to lack of space and complexity of the analysis, we leave scaling laws of the IBM for a   future publication. In Appendix C, we have used the Chernoff bound to bound Pr γ PRM < β , which is the first step to derive scaling laws for the IBM. C. Numerical Illustrations Using similar setting as in Section IV-D, Fig. 8 shows the accuracy index of both IBM and PRM under Scenario 3 against dt . Comparing this figure to Fig. 7, we observe that directionality and blockage can further boost the accuracy index when we have a deterministic wireless channel. Surprisingly, the PRM is accurate enough to motivate adopting this model to analyze and design of mmWave networks instead of the PhyM, TIM, and even IBM. For relatively pencil-beams (e.g., θ = 10 ∼ 20◦ ), which may be used in wireless backhauling applications, the accuracy of the PRM in detecting outage events is almost 1 in all our simulations. Compared to the PRM, the PhyM and IBM respectively have less than 5% and 2% higher accuracy in modeling the interference and detecting the outage events, but with substantially higher complexities. These complexities often result in limited (mostly intractable) mathematical analysis and little insight. More interestingly, the relative difference between the average rate of the typical link computed by the PRM and that of computed by the PhyM, namely E[log2 (1 + γ x )] and E[log2 (1 + γ y )] is 29 1 Accuracy index 0.9995 0.999 0.9985 IBM: IBM: PRM: PRM: 0.998 0.9975 0.997 20 θ = 20o θ = 40o θ = 20o θ = 40o 25 30 35 40 45 Average inter-transmitter distance [m] 50 Fig. 8: Accuracy of IBM and PRM under deterministic channel and directional communications. rPRM = ζ −1/α where ζ is given in (26), and rIBM = 2rPRM . The relative difference between the average per-user rate computed by the PRM and that of computed by the PhyM is less than 0.002%. less than 0.002%, implying the accuracy of the simple PRM to analyze long-term performance metrics (such as throughput and delay). Fig. 8 together with Results 4-7 support the validity of the previously proposed pseudo-wired model [14], at least for sparse networks like mmWave mesh networks [55]. This highlights the importance of having quantitative (not only qualitative) insight of the accuracy of different interference models we may face in different wireless networks. Thereby, we can adopt a simple yet accurate enough model for link-level and system-level performance analysis. So far, we have observed how we can simplify the set of dominant interferers and how much accuracy loss they entail under three network scenarios. Besides the set of interferers I, computing the SINR expression requires modeling the wireless channel and the antenna patterns. More accurate models generally reduce tractability of the SINR expression and therefore the interference model. In the next section, we analyze the possibility of adopting simple models for the wireless channel and for the antenna pattern. VI. E XAMPLE S CENARIO 4: I MPACT OF OTHER C OMPONENTS OF THE SINR E XPRESSION In this section, we analyze the accuracy loss due to simplifying wireless channel model and antenna pattern of the SINR expression. In particular, we use the proposed accuracy index to investigate the feasibility of modeling a random fading channel with a constant value without affecting the long-term performance of the real system (with random fading). The importance 30 Accuracy index 1 0.9 0.8 Nakagami (m = 9) 0.7 Nakagami (m = 3) Rayleigh 0.6 2 2.5 3 3.5 Path-loss index 4 4.5 5 Fig. 9: Impact of modeling a fading channel by a deterministic one on the accuracy of the resulting interference model (dt = 80 m). of this scenario is due to that numerous studies develop protocols and optimize the network based on deterministic wireless channels, yet no study focuses on the accuracy and validity of this underlying model. In the following, we comment on what this deterministic channel gain should be to maximize its similarity to the actual random wireless channel. We then use the proposed accuracy index to assess the impact of neglecting the reflections, assuming impenetrable obstacles, and neglecting sidelobe gain of the directional antenna on the accuracy of the resulting interference model. We consider the PhyM for both x and y throughout this section. A. Approximating a Fading Channel with a Deterministic One To design many protocols for wireless networks (such as power control, scheduling, and routing), it is often preferable to use deterministic channel gains that depend only on the distance among the transmitters and receivers [14], [17], [36]–[38]. In this subsection, we investigate the accuracy of approximating the fading gain between transmitter i and the reference receiver (hi ) in y by a deterministic value c0 in x. After this approximation, the channel gain in x becomes giCh = ac0 d−α i , and all other parameters of x are identical to those of y. For sake of simplicity, we consider omnidirectional communications without blockage, as in Section III. Using the same simulation setup as of Section III, we numerically find c0 in x that gives the highest similarity between x and y, averaged over all β ∈ [0, 10] dB. Fig. 9 shows the accuracy index, obtained by the optimal c0 , for Rayleigh and Nakagami fading. Moreover, we report in 31 TABLE I: Accuracy of the mathematical analysis when we replace fading channels with a deterministic one (dt = 80 m). “AI” refers to our accuracy index, shown also in Fig. 9. “BC” refers to the Bhattacharyya coefficient of the SINR distributions of x and y, and “TD” refers to the deviation of the throughput obtained by interference model x from that of y. Fading type Rayleigh Nakagami (m = 3) Nakagami (m = 9) α=2 α=3 α=4 α=5 AI 0.68 0.881 0.939 0.956 BC 0.275 0.048 0.014 0.005 TD 13% 9.3% 6.7% 4.5% AI 0.951 0.985 0.995 0.998 BC 0.01 0.004 0.003 0.002 TD 5.8% 4.1% 3.2% 2% AI 0.997 0.9991 0.9996 0.9999 BC 0.001 0.0008 0.0006 0.0003 TD 1.4% 1% 0.7% 0.3% Table I the Bhattacharyya coefficient between SINR distribution of x and that of y, and also the relative difference in corresponding average throughput. From Fig. 9 and Table I, interference model x (with deterministic channel) becomes more similar to y (with fading channel) as the path-loss index grows. This higher similarity manifests itself in higher accuracy indices, in lower Bhattacharyya coefficients, and also in lower errors in the rate analysis. Moreover, approximating a random wireless channel gain with Rayleigh fading and a small path-loss index (outdoor environment) by a constant value6 may lead to a non-negligible inaccuracy in the final throughput analysis (up to 13% error in our example). However, a Nakagami-m fading channel with high m can be well approximated by a deterministic channel gain, substantially simplifying the mathematical analysis and protocol development. The error due to this approximation will be reduced with m. To highlight the importance of this observation, we note that the directional communications will be largely applied in future wireless networks [56]. Therefore, wireless networks with Nakagami-m fading channels will play a major role in future of wireless networks. For mmWave communications, for instance, we are already using narrow beams [3], [55], which result in high m in the corresponding Nakagami-m fading channel. The following conjecture states how we can approximate a Nakagami-m fading channel by a deterministic channel gain. Conjecture 1. Consider a 2D network. Assume that the wireless channel attenuation consists of a constant attenuation at a reference distance, a distance-dependent attenuation with path-loss 6 We observed in our simulations that c0 = Eh [h2/α ] = Γ(1 + 2/α) is roughly the optimal constant that provides the highest similarity index in Rayleigh fading channel. Notice that it is 2/α-th moment of random variable h. 32 index α, and a random fading h. If h has a Nakagami-m distribution with m ≥ 3, the wireless channel can be well approximated by a deterministic LoS channel without significant drop in the accuracy of the resulting interference model or in the analysis of the ergodic performance metrics such as spectral efficiency, energy efficiency, throughput, and delay. If h has Rayleigh fading distribution, replacing h by its 2/α-th moment, namely Eh [h2/α ], results in sufficiently accurate analysis of the ergodic performance metrics. In the following, we further exemplify the proposed index to investigate the accuracy drop due to simplifying other parameters of the SINR expression. B. Other Components of the SINR In this subsection, we focus on mmWave networks and propose a very simple yet accurate interference model. In particular, we consider a PRM wherein we assume that i) obstacles are impenetrable, ii) there is no reflection, and iii) there is no sidelobe transmissions/receptions. Although these assumptions do not generally hold in practice, we show that this simple interference model can be very accurate abstraction of real mmWave networks. Previously, references [14], [38] used this interference model for performance evaluation and protocol development for mmWave networks. Therefore, the discussions of this subsection are a complementary study of those works. We consider a random number of obstacles in the environment each with penetration loss lo . The obstacles are assumed to have rectangular shape whose centers follow a spatial Poisson distribution with density λo on the plane, independent of the Poisson process of the interferers. To each rectangle, we associate a random width that is independently uniformly taken from [0, 4] meters, a random length that is independently uniformly taken from [0, 3] meters, and a random orientation that is independently uniformly taken from [0, 2π]. The obstacles can represent small buildings, human bodies, and cars. We independently randomly mark some obstacles to be reflectors with coefficient r ≤ 1. Without loss of generality, we mark the obstacles as reflectors with probability 0.1. We also assume that the links can be established either by the direct path or by a first-order reflected path. We consider a large scale path loss model at 28 GHz [2], which consists of a constant attenuation, a distance dependent attenuation, and a large scale log-normal fading. Besides these attenuation sources, we consider the penetration 33 TABLE II: Effects of assuming infinite penetration loss, no reflection, and no sidelobe gain on the accuracy of the resulting √ √ interference model. The shown parameters are for reference model y. SINR threshold is β = 5 dB. dt = 1/ λt and do = 1/ λo . Experiment lo r z θ dt do Accuracy ◦ 1 2 3 10 10 20 0.63 0.74 0.9 -10 -10 -10 20 40◦ 40◦ 50 30 50 20 20 50 0.9998 0.9992 0.9993 4 5 6 10 20 20 0.74 0.74 0.74 -10 -10 -10 20◦ 20◦ 20◦ 50 30 30 50 50 20 0.9614 0.9856 0.9588 7 8 9 15 15 15 0.74 0.74 0.74 -5 -5 -10 20◦ 20◦ 40◦ 50 20 30 20 50 20 0.9235 0.7090 0.9311 10 11 12 25 15 15 0.9 0.63 0.74 -10 -15 -10 10◦ 30◦ 20◦ 30 50 100 30 50 50 0.8810 0.9473 0.9718 and reflection losses. Consider path k between transmitter i and the reference receiver. Let dik be the distance of this path (path length), nk be the number of obstacles in this path, lo be the penetration loss due to any obstacle in dB, and lr = −10 log(r) be the reflection loss in dB. Let 1k denote an indicator function that takes 1 if path k contains a reflector, otherwise 0. Then, the channel gain in k-th path between transmitter i and the reference receiver is modelled as Ch gik [dB] = −61.4 − 20 log(dik ) + 1k 10 log(r) − nj lo − X , (27) where X is a zero mean i.i.d. Gaussian random variable with standard deviation 5.8 [2]. Note that the atmospheric absorbtion is almost negligible (0.15 dB/Km) at the 28 GHz [46]. Moreover, changing the carrier frequency will change the parameters of channel model (27), without affecting the generality of the results of this subsection. Again, we consider ideal sector antenna pattern, formulated in (15), at all transmitters and receivers. We consider a realistic reference physical model y with finite penetration loss (lo < ∞), first-order reflections (r > 0), and non-zero antenna side lobe (z > 0). We execute several experiments in which we change the type of the reflectors, type of obstacles, side lobe gain, operating beamwidth, and the average number of interferers and obstacles. We execute four sets of experiments. For each experiment, we compute the average accuracy index over 105 random topologies and report the result in Table II. Effects of no reflection assumption: In the first set of experiments (1–3), we consider three materials for the reflectors: drywall with reflection coefficient 0.63, clear glass with reflection 34 coefficient 0.74, and the tinted glass with reflection coefficient 0.9 [57]. All parameters of interference models x and y are similar (reported in the table), except that r = 0 in x. From Table II, the accuracy index is near 1 for all scenarios. The accuracy marginally decreases with the density of the transmitters, yet it is high enough for typical transmitter densities (dt > 30 m in downlink cellular networks). Increasing the operating beamwidth has similar effect as increasing the transmitter densities. Effects of infinite penetration loss assumption: In the second set of experiments (4–7), we consider different penetration losses for the obstacles. All parameters of x and y are similar (reported in the table), except that lo = ∞ in x. From the table, the accuracy index reduces with the density of the obstacles, as more obstacles correspond to higher source of errors in x. Moreover, the assumption of impenetrable obstacles is more accurate for higher penetration loss values. Moreover, denser mmWave networks (dt = 30) are less sensitive to assuming infinite penetration loss. The main reason is that densifying the network increases the probability of having interferers with LoS condition to the reference receiver. The contribution of those nonblocked interferers in the aggregated interference term dominate that of blocked interferers. Effects of no side lobe gains assumption: In the third set of experiments (7–9), we investigate the impact of neglecting antenna side lobes z in interference model x. All parameters of x and y are similar (reported in the table), except that z = 0 in x. Expectedly, neglecting higher z lowers the accuracy of x, and this error increases also with the number of interferers in the network. Unlike the previous parameters, neglecting side lobe gain may lead to a large deviation of x from y. From the numerical results, not shown in this paper due to the space limitations, if we have either a typical dense network 7 with dt = 30 or enough side lobe suppression (at least 10 dB), we are safe to ignore side lobe gains from the interference model. Increasing the operating beamwidth increases the chance of observing an aligned interferer (which contributes in the link budget with its main lobe gain). As such interferers have dominant role in the aggregated interference term, increasing the operating beamwidth can improve the accuracy of x. Joint effects of all parameters: In the last set of experiments (10–12), we analyze joint effects of 7 Note that using scheduling, we can reduce mutual interference by controlling the number of simultaneous active transmitters. Therefore, the number of transmitters in the environment is not necessarily equal to the number of interferers; rather, it is usually much higher than that. 35 all those parameters by considering infinite penetration loss, zero reflection coefficient, and zero side lobe gain in x. Other parameters of x are similar to those of y, reported in Table II. From the results, our simple interference model x is sufficiently accurate for typical mmWave network scenarios. On the negative side, larger number of interferers magnifies the small error due to neglecting antenna side lobes. This magnified error together with other approximations leads to 12% error in detecting outage event by the simplified interference model in Experiment 10. On the positive side, this higher transmitter density reduces the error due to both neglecting reflection and assuming impenetrable obstacles. VII. F UTURE D IRECTIONS Throughout this paper, we highlighted the tradeoff between the accuracy and mathematical tractability of the interference models, and exemplified the use of our accuracy index to optimize such tradeoff for different wireless network scenarios, with specific reference to mmWave networks. Although we have simplified system models of the examples to avoid unnecessary complications, our index poses no limitation to these example scenarios. We have recently used this index to assess the accuracy of a simple interference model for a mmWave cellular network [45]. Two future directions can be envisioned from this paper. First, one may use our accuracy index to simplify the existing and develop new interference models for various network settings. In particular, illustrative examples of this paper were more suitable for ad hoc networks, and evaluating the generality of the resulting insights is an interesting future research line. Moreover, our proposed index can be used to assess the accuracy of different blockage models like one-ball [53], two-ball [4], cone [41], and queuebased models [58] and even develop novel accurate yet tractable models. Second, we can extend the index itself. In this paper, we have defined the similarity index for any interference model x based on its ability to correctly predict the outage events; see Definition 1. To generalize our approach, one may aim at measuring the similarity based on any other functions of SNIR. For example, given some alternatives for one function inside SINR (e.g., different set of interferers or different antenna models), one may use an extension of our approach to identify which of them better balances the accuracy-complexity tradeoff for a throughput/delay analysis. 36 VIII. C ONCLUSION We developed a new mathematical framework to address very fundamental questions in analysis and design of wireless networks: how accurate different interference models are and how to select the right one. We proposed a new accuracy index that quantifies the ability of any interference model in correctly predicting outage events, under any network setting. We analytically and numerically illustrated the use of our index via many example scenarios. In particular, we evaluated the accuracy of the prominent techniques that model the set of dominant interferers. We then showed that directional antenna and obstructions (basic characteristics of mmWave networks) substantially enhance the accuracy of any interference model, making the simple classical protocol model accurate enough for analysis and optimization of such networks. Furthermore, we measured the accuracy of approximating a random fading wireless channel with a deterministic channel. We conjectured that a Nakagami-m fading channels with m ≥ 3 can be well approximated by a deterministic value without introducing a significant gap in the ergodic performance metrics (e.g., throughput and delay); whereas, such gap is generally nonnegligible under Rayleigh fading channels. Finally, we showed surprisingly high accuracy of a simple interference model that assumes (i) infinite penetration loss, (ii) no reflection, and (iii) no antenna side lobes in modeling a typical mmWave network where none of those assumptions hold. A PPENDIX A: D ERIVING C OMPONENTS OF E XAMPLE S CENARIO 1 A. The Interference Ball Model Let Ex denote expectation over random variable x. From (5) and   (6),    Pr γ IBM < β = Pr   pch0 d−α 0 P k∈I∩B(2π,0,rIBM )  −α1 pchk dk B(2π,0,a)  = 1 − Pr h0 ≥  X X k∈I∩B(2π,0,rIBM )   σ  α βd0 pc    −α1 σ hk dk B(2π,0,a) +  βdα0   pc −α1B(2π,0,a) hk dk k∈I∩B(2π,0,rIBM )    = 1 − EI,h exp −    +σ  < β  + 37       X −α1 EI,h exp −  hk dk B(2π,0,a)  βdα0  = 1 − exp   pc k∈I∩B(2π,0,rIBM )       o n α Y −α1 −σβd0 EI  Eh exp −βdα0 hk dk B(2π,0,a)  . = 1 − exp pc k∈I∩B(2π,0,rIBM ) {z } |  −σβdα0   A (28) From probability generating functionals, we have A = exp            − 2πλt Z ∞  h i −α1 B(2π,0,a) −βdα hr 1B(2π,0,rIBM ) 1 − Eh e 0 r dr            Z ∞    −α1 B(2π,0,a) −βdα hr   0 r dr = exp − 2πλt Eh 1B(2π,0,rIBM ) 1 − e   0            Z Z    rIBM  a− αh α hr −α −βd −βd   0 0 1−e r dr + 1−e r dr = exp − 2πλt Eh   0 a              −α α −α α  α (⋆) 2 1 − e−βd0 hrIBM − a2 1 − e−βd0 ha = exp − πλt Eh a2 1 − e−βd0 h + rIBM    0         2 2 α 2/α α −α α −α  α 2/α + (βd0 h) Γ 1 − , βd0 hrIBM − (βd0 h) Γ 1 − , βd0 ha ,  α α   (29) where Γ (·, ·) is the incomplete Gamma function, (⋆) is derived using integration by parts, and   probability density function of h is fh (x) = e−x . To find Pr γ PhyM < β , we only need to   evaluate Pr γ IBM < β at rIBM → ∞, that is,     Pr γ PhyM < β = lim Pr γ IBM < β rIBM →∞ 38 = 1 − exp        − σβdα0 pc    α −α α  − πλt Eh a2 1 − e−βd0 h − a2 1 − e−βd0 ha         2 2 α 2/α α −α  α 2/α − (βd0 h) Γ 1 − , βd0 ha , + (βd0 h) Γ 1 −  α α   (30) where Γ (·) is the Gamma function. B. The Protocol Model Event γ PRM ≥ β implies that there is no interferer inside B(2π, 0, rPRM). Assuming rPRM ≥ a and d0 ≥ a, and following similar steps as in (28) and (29), we have     Z ∞    −α1 α B(2π,0,a) Pr[γ PhyM < β | γ PRM ≥ β] = 1 − exp − 2πλt Eh  1B(2π,rPRM ,∞) 1 − e−βd0 hr r dr   0      σβdα0  − pc           Z  −σβdα   ∞  α hr −α 0 −βd = 1 − exp − 2πλt Eh  1−e 0 r dr    pc rPRM            −σβdα   −α 2 0 α 2/α −βdα hrPRM 2 0  + (βd0 h) Γ 1 − − πλt Eh − rPRM 1 − e = 1 − exp  pc α         2 −α α 2/α α  − (βd0 h) Γ 1 − , βd0 hrPRM .  α   (31) 39 A PPENDIX B: D ERIVING C OMPONENTS OF E XAMPLE S CENARIO 2 A. The Interference Ball Model We have     Pr γ IBM < β = Pr   p P p  2π 2 θ     = 1 − EI,h exp −    = 1 − exp  2 σθ βdα0 − 4pcπ 2  2π 2 θ k∈I∩B(θ,0,rIBM ) ch0 d−α 0 −α1B(2π,0,a) chk dk X +σ  EI  Y k∈I∩B(θ,0,rIBM ) |  < β  −α1B(2π,0,a) hk dk k∈I∩B(θ,0,rIBM )   Eh   σθ  α  + βd0  4pcπ 2 2    n o −α1 exp −βdα0 hk dk B(2π,0,a)  , {z } B (32) where         Z ∞  h i −α1 α B(2π,0,a) −βd0 hr B = exp − 1B(θ,0,rIBM ) 1 − Eh e θλI (r)r dr   0              Z ∞   θ2 −α1 α B(2π,0,a) −βd0 hr −ǫλo r   λt E h 1B(θ,0,rIBM ) 1 − e e r dr = exp −   2π 0            Z Z  θ2   rIBM  a− α h  −ǫλ r α hr −α −βd −βd −ǫλ r o o = exp − 1−e 0 e r dr + 1−e 0 e r dr  λt E h    2π 0 a             Z rIBM    θ2   −ǫλo a 1 − (ǫλ a + 1) e α hr −α α h o −ǫλ r −βd −βd o 0  . = exp − + e r dr λt E h  1 − e 0 1 − e   2π ǫ2 λ2o a     (33) Also, we have   Pr γ PhyM < β =   Pr γ IBM < β rIBM →∞      σθ2 βdα 2  θ λt 0 1 − e−βdα0 h 1 − (ǫλo a + 1) e−ǫλo a E − = 1 − exp − h  4pcπ 2 2πǫ2 λ2o   lim 40    Z ∞  2 2 −βdα hr −α −ǫλo r 0  − ǫ λo . e r dr  a   (34) B. The Protocol Model We have Pr[γ PhyM < β | γ PRM ≥ β] = 1 − exp         − Eh θ Z ∞  −α1 B(2π,0,a) −βdα 0 hr 1B(θ,rPRM ,∞) 1 − e 0 −        σθ2 βdα  0 2 4pcπ   λI (r)r dr        Z   σθ2 βdα0 θ 2 λt  ∞  −α −βdα hr −ǫλ r o  0 = 1 − exp − − E . 1 − e e r dr h   4pcπ 2 2π rPRM      (35) A PPENDIX C: D ERIVING B OUNDS FOR E XAMPLE S CENARIO 3 In this appendix, we derive bounds on the miss-detection probability in Example Scenario 3. A. The Interference Ball Model To derive an upper bound on the miss-detection probability, we substitute a lower bound of     Pr γ IBM < β and upper bound of Pr γ PhyM < β into (8) of the main manuscript. Recall the definition of ζ in (26). " For any real positive τ , # " # X −α1B(2π,0,a) X −α1B(2π,0,a)  IBM  (⋆) dk dk Pr γ < β = Pr 1B(θ,0,rIBM ) > ζ = 1 − Pr 1B(θ,0,rIBM ) ≤ ζ k∈I k∈I  (⋆⋆) ( ≥ 1 − inf eτ ζ EI  exp −τ τ >0 | X −α1B(2π,0,a) dk k∈I {z C 1B(θ,0,rIBM ) ) . } (36) where (⋆) is due to (6) in the main manuscript, and (⋆⋆) follows from the Chernoff bound and the probability   generating functionals. Z rIBM     −α1 B(2π,0,a) −τ r λI (r)r dr C = exp − θ 1−e   0 41        Z rIBM     −ǫλo r −τ −ǫλo r −τ r −α  re dr + 1−e re dr 1−e  a    Z − θ2  a = exp − λt  2π 0        θ2 λ 1 − e−τ + (1 + ǫλ a) e−ǫλo a−τ − (1 + ǫλ r ) e−ǫλo rIBM o o IBM t = exp − 2 2  2π ǫ λo      Z rIBM  −ǫλo r−τ r −α − re dr  .  a   (37) Using similar technique, we use the Chernoff bound to find exponentially decreasing bound on the tail distribution"of γ PhyM as # X −α1B(2π,0,a)  PhyM  Pr γ < β = Pr dk 1B(θ,0,∞) > ζ k∈I  ( ≤ inf e−τ ζ EI  exp τ τ >0 = inf exp τ >0        − τζ − X k∈I  ) −α1 dk B(2π,0,a) 1B(θ,0,∞)  θ2 λt  1 − eτ + (1 + ǫλo a) e−ǫλo a+τ 2π ǫ2 λ2o    Z ∞  −ǫλo r+τ r −α  − . re dr  a   (38) Note that bounds (36) and (38) are derived using the Chernoff bound. However, easier but looser bounds can be found using Markov and Chebyshev inequalities. These bounds can be readily derived by direct application of the Campbell’s Theorem [25]. B. The Protocol Model We can also find bounds and scaling laws for the accuracy index of the PRM in Example Scenario 3. To this end, we use the following proposition:   Proposition 3. For ξ = Pr γ PhyM ≥ β , we have     max Pr γ IBM < β , Pr γ PhyM ≥ β ≤ Sβ,ξ (IBMkPhyM) ≤ 1 . Also, for any 0 < rPRM ≤ ζ −1/α where ζ is defined in (26) of the manuscript we have     max Pr γ PRM < β , Pr γ PhyM ≥ β ≤ Sβ,ξ (PRMkPhyM) ≤ 1 , (39) (40) 42 where  Pr γ PRM  θ 2 λt  −ǫλo rPRM 1 − (1 + ǫλo rPRM ) e . < β = 1 − exp − 2πǫ2 λ2o   Proof: For the IBM, the upper bound is trivial. To derive the lower bound, from (3) of the manuscript we have IBM|PhyM Sβ,ξ (IBMkPhyM) = 1 − ξpfa (⋆) IBM|PhyM where (⋆) is because pfa IBM|PhyM − (1 − ξ) pmd IBM|PhyM = 1 − (1 − ξ) pmd     (⋆⋆) = Pr γ PhyM ≥ β + Pr γ IBM < β , (41) = 0 for any rIBM ≥ 0, and (⋆⋆) is due to (8) of the manuscript. Then, (39) follows.     To derive the lower bound of the PRM, we first note that Pr γ PhyM < β | γ PRM ≥ β ≤ Pr γ PhyM < β for any rPRM > 0. Therefore, from (12) of the manuscript,    PRM  θ 2 λt  PRM|PhyM −ǫλo rPRM pmd ≤ 1 − Pr γ < β = exp − 1 − (1 + ǫλo rPRM ) e . 2πǫ2 λ2o Then, from (3) we have PRM|PhyM Sβ,ξ (PRMkPhyM) = 1 − ξpfa (⋆) (42) PRM|PhyM − (1 − ξ) pmd PRM|PhyM = 1 − (1 − ξ) pmd PRM|PhyM ≥ 1 − pmd    PRM  θ 2 λt  −ǫλo rPRM , 1 − (1 + ǫλo rPRM ) e ≥ Pr γ < β = 1 − exp − 2πǫ2 λ2o (⋆⋆) (43) PRM|PhyM where (⋆) is because pfa = 0 for any rPRM ≤ ζ −1/α where ζ is defined in Eqn. (27) of the manuscript (see Result 7 of the manuscript), and (⋆⋆) is due to Eqn (42). Moreover, PRM|PhyM Sβ,ξ (PRMkPhyM) = 1 − (1 − ξ) pmd     (⋆) = 1 − Pr γ PRM ≥ β Pr γ PhyM < β | γ PRM ≥ β   ≥ 1 − Pr γ PhyM < β | γ PRM ≥ β   ≥ 1 − Pr γ PhyM < β , where (⋆) is from (12). Combining (43) and (44) results in (40). (44) 43 R EFERENCES [1] H. Shokri-Ghadikolaei, C. Fischione, and E. Modiano, “On the accuracy of interference models in wireless communications,” in Proc. IEEE International Conference on Communications (ICC), 2016. [2] M. Akdeniz, Y. Liu, M. Samimi, S. Sun, S. Rangan, T. Rappaport, and E. Erkip, “Millimeter wave channel modeling and cellular capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1164–1179, Jun. 2014. [3] H. Shokri-Ghadikolaei, C. Fischione, G. Fodor, P. Popovski, and M. Zorzi, “Millimeter wave cellular networks: A MAC layer perspective,” IEEE Trans. Commun., vol. 63, no. 10, pp. 3437–3458, Oct. 2015. [4] M. Di Renzo, “Stochastic geometry modeling and analysis of multi-tier millimeter wave cellular networks,” IEEE Trans. Wireless Commun., vol. 14, no. 9, pp. 5038–5057, Sept. 2015. [5] A. Ephremides, J. E. Wieselthier, and D. J. Baker, “A design concept for reliable mobile radio networks with frequency hopping signaling,” Proc. IEEE, vol. 75, no. 1, pp. 56–73, 1987. [6] A. Iyer, C. Rosenberg, and A. Karnik, “What is the right model for wireless channel interference?” IEEE Trans. Wireless Commun., vol. 8, no. 5, pp. 2662–2671, May 2009. [7] P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEE Trans. Inform. Theory, vol. 46, no. 2, pp. 388–404, Mar. 2000. [8] B. Liu, Z. Liu, and D. Towsley, “On the capacity of hybrid wireless networks,” in Proc. IEEE International Conference on Computer Communications (INFOCOM), vol. 2, 2003, pp. 1543–1552. [9] P. Kyasanur and N. H. Vaidya, “Capacity of multichannel wireless networks under the protocol model,” IEEE/ACM Trans. Netw., vol. 17, no. 2, pp. 515–527, Apr. 2009. [10] A. E. Gamal, J. Mammen, B. Prabhakar, and D. Shah, “Throughput-delay trade-off in wireless networks,” in Proc. IEEE International Conference on Computer Communications (INFOCOM), 2004, pp. 464–475. [11] A. El Gamal, J. Mammen, B. Prabhakar, and D. Shah, “Optimal throughput-delay scaling in wireless networks–part I: The fluid model,” IEEE Trans. Inform. Theory, vol. 52, no. 6, pp. 2568–2592, Jun. 2006. [12] T. Nandagopal, T.-E. Kim, X. Gao, and V. Bharghavan, “Achieving MAC layer fairness in wireless packet networks,” in Proc. ACM International Conference on Mobile Computing and Networking (MobiCom), 2000, pp. 87–98. [13] K. Xu, M. Gerla, and S. Bae, “How effective is the IEEE 802.11 RTS/CTS handshake in ad hoc networks?” in Proc. IEEE Global Communications Conference (GLOBECOM), 2002, pp. 72–76. [14] S. Singh, R. Mudumbai, and U. Madhow, “Interference analysis for highly directional 60-GHz mesh networks: The case for rethinking medium access control,” IEEE/ACM Trans. Netw., vol. 19, no. 5, pp. 1513–1527, Oct. 2011. [15] Y. Xu, H. Shokri-Ghadikolaei, and C. Fischione, “Distributed association and relaying with fairness in millimeterwaves networks,” IEEE Trans. Wireless Commun., vol. 15, no. 12, pp. 7955–7970, Dec. 2016. [16] M. K. Marina, S. R. Das, and A. P. Subramanian, “A topology control approach for utilizing multiple channels in multi-radio wireless mesh networks,” Computer Networks, vol. 54, no. 2, pp. 241–256, Feb. 2010. [17] T. Stahlbuhk, B. Shrader, and E. Modiano, “Topology control for wireless networks with highly-directional antennas,” in Proc. IEEE International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt), 2016. [18] K. Jain, J. Padhye, V. N. Padmanabhan, and L. Qiu, “Impact of interference on multi-hop wireless network performance,” Wireless Networks, vol. 11, no. 4, pp. 471–487, 2005. 44 [19] G. D. Celik, G. Zussman, W. F. Khan, and E. Modiano, “MAC for networks with multipacket reception capability and spatially distributed nodes,” IEEE Trans. Mobile Comput., vol. 9, no. 2, pp. 226–240, Aug. 2010. [20] S. P. Weber, X. Yang, J. G. Andrews, and G. De Veciana, “Transmission capacity of wireless ad hoc networks with outage constraints,” IEEE Trans. Inform. Theory, vol. 51, no. 12, pp. 4091–4102, Dec. 2005. [21] S. P. Weber, J. G. Andrews, X. Yang, and G. De Veciana, “Transmission capacity of wireless ad hoc networks with successive interference cancellation,” IEEE Trans. Inform. Theory, vol. 53, no. 8, pp. 2799–2814, Aug. 2007. [22] L. B. Le, E. Modiano, C. Joo, and N. B. Shroff, “Longest-queue-first scheduling under SINR interference model,” in Proc. ACM International Symposium on Mobile Ad hoc Networking and Computing (MobiHoc), 2010, pp. 41–50. [23] S. Jafar, “Topological interference management through index coding,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp. 529–568, Jan. 2014. [24] X. Yi and D. Gesbert, “Topological interference management with transmitter cooperation,” IEEE Trans. Wireless Commun., vol. 61, no. 11, pp. 6107–6130, Nov. 2015. [25] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge University Press, 2013. [26] M. Schubert and H. Boche, “Solution of the multiuser downlink beamforming problem with individual SINR constraints,” IEEE Trans. Veh. Technol., vol. 53, no. 1, pp. 18–28, Jan. 2004. [27] H. Dahrouj and W. Yu, “Coordinated beamforming for the multicell multi-antenna wireless system,” IEEE Trans. Wireless Commun., vol. 9, no. 5, pp. 1748–1759, May 2010. [28] N. N. Moghadam, H. Shokri-Ghadikolaei, G. Fodor, M. Bengtsson, and C. Fischione, “Pilot precoding and combining in multiuser MIMO networks,” IEEE J. Sel. Areas Commun., vol. 35, no. 7, pp. 1632–1648, Jul. 2017. [29] M. Sharif and B. Hassibi, “On the capacity of MIMO broadcast channels with partial side information,” IEEE Trans. Inform. Theory, vol. 51, no. 2, pp. 506–522, Feb. 2005. [30] F. Baccelli and B. Blaszczyszyn, Stochastic Geometry and Wireless Networks, Volume II–Applications. NOW publishers, 2009. [31] F. Rashid-Farrokhi, L. Tassiulas, and K. Liu, “Joint optimal power control and beamforming in wireless networks using antenna arrays,” IEEE Trans. Commun., vol. 46, no. 10, pp. 1313–1324, Oct. 1998. [32] V. Chandrasekhar, J. G. Andrews, T. Muharemovic, Z. Shen, and A. Gatherer, “Power control in two-tier femtocell networks,” IEEE Trans. Wireless Commun., vol. 8, no. 8, pp. 4316–4328, Aug. 2009. [33] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral efficiency of very large multiuser MIMO systems,” IEEE Trans. Commun., vol. 61, no. 4, pp. 1436–1449, Feb. 2013. [34] H. Shokri-Ghadikolaei, F. Boccardi, C. Fischione, G. Fodor, and M. Zorzi, “Spectrum sharing in mmWave cellular networks via cell association, coordination, and beamforming,” IEEE J. Sel. Areas Commun., vol. 34, no. 11, pp. 2902–2917, Nov. 2016. [35] P. Cardieri, “Modeling interference in wireless ad hoc networks,” IEEE Commun. Surveys Tuts., vol. 12, no. 4, pp. 551–572, Fourth Quarter 2010. [36] L. Badia, A. Erta, L. Lenzini, and M. Zorzi, “A general interference-aware framework for joint routing and link scheduling in wireless mesh networks,” IEEE Netw., vol. 22, no. 1, pp. 32–38, Jan. 2008. [37] L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Cross-layer congestion control, routing and scheduling design in ad hoc wireless networks,” in Proc. IEEE International Conference on Computer Communications (INFOCOM), Apr. 2006. [38] S. Singh, F. Ziliotto, U. Madhow, E. Belding, and M. Rodwell, “Blockage and directivity in 60 GHz wireless personal area 45 networks: From cross-layer model to multihop MAC design,” IEEE J. Sel. Areas Commun., vol. 27, no. 8, pp. 1400–1413, Oct. 2009. [39] A. K. Gupta, J. G. Andrews, and R. W. Heath, “On the feasibility of sharing spectrum licenses in mmwave cellular systems,” IEEE Trans. Commun., vol. 64, no. 9, pp. 3981–3995, Jul. 2016. [40] J. Dams, M. Hoefer, and T. Kesselheim, “Scheduling in wireless networks withrayleigh-fading interference,” IEEE Trans. Mobile Comput., vol. 14, no. 7, pp. 1503–1514, Jul. 2015. [41] H. Shokri-Ghadikolaei and C. Fischione, “The transitional behavior of interference in millimeter wave networks and its impact on medium access control,” IEEE Trans. Commun., vol. 62, no. 2, pp. 723–740, Feb. 2016. [42] T. Kailath, “The divergence and Bhattacharyya distance measures in signal selection,” IEEE Trans. Commun., vol. 15, no. 1, pp. 52–60, Feb. 1967. [43] E. Modiano, D. Shah, and G. Zussman, “Maximizing throughput in wireless networks via gossiping,” in Proc. ACM SIGMETRICS Performance Evaluation Review, Jun. 2006, pp. 27–38. [44] P. Di Marco, C. Fischione, F. Santucci, and K. H. Johansson, “Modeling ieee 802.15.4 networks over fading channels,” IEEE Trans. Wireless Commun., vol. 13, no. 10, pp. 5366–5381, Oct. 2014. [45] X. Jiang, H. Shokri-Ghadikolaei, C. Fischione, and Z. Pang, “A simplified interference model for outdoor millimeter wave networks,” in Proc. EAI International Wireless Internet Conference (EAI WICOM), 2016. [46] S. Rangan, T. Rappaport, and E. Erkip, “Millimeter wave cellular wireless networks: Potentials and challenges,” Proc. IEEE, vol. 102, no. 3, pp. 366–385, Mar. 2014. [47] S. Singh, M. Mudumbai, and U. Madhow, “Distributed coordination with deaf neighbors: Efficient medium access for 60GHz mesh networks,” in Proc. IEEE International Conference on Computer Communications (INFOCOM), 2010. [48] X. An and R. Hekmat, “Directional MAC protocol for millimeter wave based wireless personal area networks,” in Proc. IEEE Vehicular Technology Conference (VTC Spring), 2008, pp. 1636–1640. [49] I. K. Son, S. Mao, M. X. Gong, and Y. Li, “On frame-based scheduling for directional mmWave WPANs,” in Proc. IEEE International Conference on Computer Communications (INFOCOM), 2012, pp. 2149–2157. [50] V. Petrov, M. Komarov, D. Moltchanov, J. M. Jornet, and Y. Koucheryavy, “Interference and SINR in millimeter wave and terahertz communication systems with blocking and directional antennas,” IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 1791–1808, Mar. 2017. [51] J. García-Rois, F. Gómez-Cuba, M. R. Akdeniz, F. J. González-Castaño, J. C. Burguillo, S. Rangan, and B. Lorenzo, “On the analysis of scheduling in dynamic duplex multihop mmwave cellular systems,” IEEE Trans. Wireless Commun., vol. 14, no. 11, pp. 6028–6042, Nov. 2015. [52] H. Shokri-Ghadikolaei, L. Gkatzikis, and C. Fischione, “Beam-searching and transmission scheduling in millimeter wave communications,” in Proc. IEEE International Conference on Communications (ICC), 2015, pp. 1292–1297. [53] T. Bai, R. Vaze, and R. Heath, “Analysis of blockage effects on urban cellular networks,” IEEE Trans. Wireless Commun., vol. 13, no. 9, pp. 5070–5083, Sept. 2014. [54] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements and channel models for future wireless communication system design,” IEEE Trans. Commun., vol. 63, no. 9, pp. 3029–3056, Sept. 2015. [55] F. Yaghoubi, J. Chen, A. Rostami, and L. Wosinska, “Mitigation of rain impact on microwave backhaul networks,” in Proc. IEEE International Conference on Communications (ICC) Workshop, 2016. 46 [56] F. Boccardi, R. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,” IEEE Commun. Mag., vol. 52, no. 2, pp. 74–80, Feb. 2014. [57] H. Zhao, R. Mayzus, S. Sun, M. Samimi, J. K. Schulz, Y. Azar, K. Wang, G. N. Wong, F. Gutierrez, and T. S. Rappaport, “28 GHz millimeter wave cellular communication measurements for reflection and penetration loss in and around buildings in New York city,” in Proc. IEEE International Conference on Communications (ICC), 2013, pp. 5163–5167. [58] R. Congiu, H. Shokri-Ghadikolaei, C. Fischione, and F. Santucci, “On the relay-fallback tradeoff in millimeter wave wireless system,” in Proc. IEEE International Conference on Computer Communications (INFOCOM) Workshop, 2016, pp. 622–627. 1 Accuracy index 0.8 0.6 0.4 0.2 θ = 10o (exact) θ = 30o (exact) θ = 10o (bound) θ = 30o (bound) 0 -10 -5 0 5 10 SINR Threshold [dB] 15 20 0 10 θ = 30o θ = 15o −1 Accuracy index 10 θ = 5o −2 10 −3 10 −4 10 −5 10 −4 10 −3 10 −2 10 Obstacle density −1 10 0 10 0 10 −1 Performance of the detector 10 −2 10 −3 10 −4 Pmd 10 Pfa (upper bound) Pmd (upper bound) −5 10 Accuracy index Accuracy index (bound) −6 10 −6 10 −4 10 −2 10 Transmitter density 0 10 2 10 1 Accuracy index 0.9 0.8 0.7 0.6 θ = 10o (exact) 0.5 θ = 20o (exact) θ = 10o (bound) θ = 20o (bound) 10 -3 10 -2 10 -1 Transmitter density 10 0 10 1 0 10 Performance of the detector Pmd (UB1), λt=0.1 Pmd (UB2), λt=0.1 −1 10 Pmd, λt=0.1 Pmd (UB1), λt=1 Pmd, λt=1 −2 10 Pmd (UB1), λt=10 Pmd, λt=10 −3 10 −4 10 0 10 20 30 40 50 60 70 Operating beamwidth [degrees] 80 90 100 Accuracy index 1 0.8 0.6 rPRM = 20 rPRM = 60 rPRM = 120 optimal rPRM 0.4 0.2 0 50 100 150 Average inter-transmitter distance [m] 200 Accuracy index 1 0.98 0.96 λt = 0.1, λo = 0.02 λt = 0.1, λo = 0.04 0.94 λt = 0.1, λo = 0.07 λt = 1 , λo = 0.04 0.92 5 10 15 20 25 30 35 Penetration loss [dB] 40 45 Accuracy index 1 0.999 0.998 λt = 0.1, θ = 20o 0.997 λt = 1 , θ = 20o 0.996 λt = 1 , θ = 40o 0 0.2 0.4 0.6 Reflection coefficient 0.8 1 Accuracy index 1 0.9 0.8 0.7 0.6 -30 λt = 0.1, θ = 20o , β = 5 λt = 1 , θ = 40o , β = 5 λt = 1, θ = 20o , β = 5 λt = 1, θ = 20o , β = 10 -25 -20 -15 Sidelobe gain [dB] -10 -5 1 Accuracy index 0.99 0.98 0.97 0.96 IBM: θ = 10o IBM: θ = 30o PRM: θ = 10o PRM: θ = 30o 0.95 10 15 20 25 Average inter-transmitter distance [m] 30 1 Accuracy index 0.95 IBM: rIBM =30, λ t = 0.11 IRM: rIRM=30, λ t = 0.11 0.9 IBM: rIBM =10, λ t = 0.0001 IRM: rIRM=10, λ t = 0.0001 IBM: rIBM =10, λ t = 0.11 0.85 IRM: rIRM=10, λ t = 0.11 0.8 -10 -5 0 5 10 SINR threshold [dB] 15 20 1 IRM: β = 1, λ t = 0.11 IBM: β = 1, λ t = 0.11 Accuracy index 0.95 IRM: β = 10, λ t = 0.0001 IRM: β = 10, λ t = 0.11 IBM: β = 10, λ t = 0.0001 0.9 IBM: β = 10, λ t = 0.11 0.85 0.8 0 20 40 60 Interference range [m] 80 100 1 Accuracy index 0.8 0.6 IBM: rIBM = 30, λ t = 0.0001 IRM: rIRM = 30, λ t = 0.0001 IBM: rIBM = 10, λ t = 0.0001 0.4 IRM: rIRM = 10, λ t = 0.0001 IBM: rIBM = 10, λ t = 0.11 IRM: rIRM = 10, λ t = 0.11 0.2 -10 -5 0 5 10 SINR threshold [dB] 15 20 1 IRM: β = 1, λt = 0.0001 Accuracy index 0.8 IBM: β = 1, λt = 0.0001 IRM: β = 10, λt = 0.0001 0.6 IRM: β = 10, λt = 0.11 IBM: β = 10, λt = 0.0001 0.4 IBM: β = 10, λt = 0.11 0.2 0 0 30 60 90 Interference range [m] 120 150
7cs.IT
Approximate Clustering with Same-Cluster Queries Nir Ailon1 , Anup Bhattacharya2 , Ragesh Jaiswal2 , and Amit Kumar2 arXiv:1704.01862v3 [cs.DS] 4 Oct 2017 2 1 Technion, Haifa, Israel.⋆ Department of Computer Science and Engineering, Indian Institute of Technology Delhi.⋆⋆ Abstract. Ashtiani et al. proposed a Semi-Supervised Active Clustering framework (SSAC), where the learner is allowed to make adaptive queries to a domain expert. The queries are of the kind “do two given points belong to the same optimal cluster?”, and the answers to these queries are assumed to be consistent with a unique optimal solution. There are many clustering contexts where such same cluster queries are feasible. Ashtiani et al. exhibited the power of such queries by showing that any instance of the k-means clustering problem, with additional margin assumption, can be solved efficiently if one is allowed O(k2 log k + k log n) same-cluster queries. This is interesting since the k-means problem, even with the margin assumption, is NP-hard. In this paper, we extend the work of Ashtiani et al. to the approximation setting showing that a few of such same-cluster queries enables one to get a polynomial-time (1 + ε)-approximation algorithm for the k-means problem without any margin assumption on the input dataset. Again, this is interesting since the k-means problem is NP-hard to approximate within a factor (1 + c) for a fixed constant 0 < c < 1. The number of same-cluster queries used is poly(k/ε) which is independent of the size n of the dataset. Our algorithm is based on the D2 -sampling technique, also known as the k-means++ seeding algorithm. We also give a conditional lower bound on the number of same-cluster queries showing that if the Exponential Time Hypothesis (ETH) holds, then any such efficient query algorithm  needs to make Ω polyklog k same-cluster queries. Our algorithm can be extended for the case when the oracle is faulty, that is, it gives wrong answers to queries with some bounded probability. Another result we show with respect to the k-means++ seeding algorithm is that a small modification to the kmeans++ seeding algorithm within the SSAC framework converts it to a constant factor approximation algorithm instead of the well known O(log k)-approximation algorithm. 1 Introduction Clustering is extensively used in data mining and is typically the first task performed when trying to understand large data. Clustering basically involves partitioning given data into groups or clusters such that data points within the same cluster are similar as per some similarity measure. Clustering is usually performed in an unsupervised setting where data points do not have any labels associated with them. The partitioning is done using some measure of similarity/dissimilarity between data elements. In this work, we extend the work of Ashtiani et al. [AKBD16] who explored the possibility of performing clustering in a semi-supervised active learning setting for center based clustering problems such as k-median/means. In this setting, which they call Semi-Supervised Active Clustering framework or SSAC in short, the clustering algorithm is allowed to make adaptive queries of the form: do two points from the dataset belong to the same optimal cluster?. where query answers are assumed to be consistent with a unique optimal solution. Ashtiani et al. [AKBD16] started the study of understanding the strength of this model. Do hard clustering problems become easy in this model? They explore such questions in the context of center-based clustering problems. Center based clustering problems such as k-means are extensively used to analyze large data clustering problems. Let us define the k-means problem in the Euclidean setting. ⋆ ⋆⋆ Email address: [email protected] Email addresses: {anupb, rjaiswal, amitk}@cse.iitd.ac.in Definition 1 (k-means problem). Given a dataset X ⊆ Rd containing n points, and a positive integer k, find a set of k points C ⊆ Rd (called centers) such that the following cost function is minimized: X Φ(C, X) = min D(x, c). x∈X c∈C D(x, c) denotes the squared Euclidean distance between c and x. That is, D(x, c) = ||x − c||2 . Note that the k optimal centers c1 , ..., ck of the k-means problem define k clusters of points in a natural manner. All points for which the closest center is ci belong to the ith cluster. This is also known as the Voronoi partitioning and the clusters obtained in this manner using the optimal k centers are called the optimal clusters. Note that the optimal center for the 1-means problem P def. for any dataset X ⊆ Rd is the centroid of the dataset denoted by µ(X) = x∈X |X| x . This means that if X1 , ...., Xk are the optimal clusters for the k-means problem on any dataset X ⊆ Rd and c1 , ..., ck are the corresponding optimal centers, then ∀i, ci = µ(Xi ). The k-means problem has been widely studied and various facts are known about this problem. The problem is tractable when either the number k of clusters or the dimension d equal to 1. However, when k > 1 or d > 1, then the problem is known to be NP-hard [Das08, Vat09, MNV12]. There has been a number of works of beyond the worst-case flavour for k-means problem in which it is typically assumed that the dataset satisfies some separation condition, and then the question is whether this assumption can be exploited to design algorithms providing better guarantees for the problem. Such questions have led to different definitions of separation and also some interesting results for datasets that satisfy these separation conditions (e.g., [ORSS13, BBG09, ABS12]). Ashtiani et al. [AKBD16] explored one such separation notion that they call the γ-margin property. Definition 2 (γ-margin property). Let γ > 1 be a real number. Let X ⊆ Rd be any dataset and k be any positive integer. Let PX = {X1 , ..., Xk } denote k optimal clusters for the k-means problem. Then this optimal partition of the dataset PX is said to satisfy the γ-margin property iff for all i 6= j ∈ {1, ..., k} and x ∈ Xi and y ∈ Xj , we have: γ · ||x − µ(Xi )||< ||y − µ(Xi )||. Qualitatively, what this means is that every point within some cluster is closer to its own cluster center than a point that does not belong to this cluster. This seems to be a very strong separation property. Ashtiani et al. [AKBD16] showed that the k-means clustering problem is NP-hard even √ when restricted to instances that satisfy the γ-margin property for γ = 3.4 ≈ 1.84. Here is the formal statement of their hardness result. Theorem 1 (Theorem 10 in [AKBD16]). Finding an optimal solution to k-means objective function is NP-hard when k = Θ(nε ) for any √ ε ∈ (0, 1), even when there is an optimal clustering that satisfies the γ-margin property for γ = 3.4. In the context of the k-means problem, the same-cluster queries within the SSAC framework are decision questions of the form: Are points x, y such that x 6= y belong to the same optimal cluster? 3 Following is the main question explored by Ashitiani et al. [AKBD16]: √ Under the γ-margin assumption, for a fixed γ ∈ (1, 3.4], how many queries must be made in the SSAC framework for k-means to become tractable? 3 In case where the optimal solution is not unique, the same-cluster query answers are consistent with respect to any fixed optimal clustering. Ashtiani et al. [AKBD16] addressed the above question and gave a query algorithm. Their algorithm, in fact, works for a more general setting where the clusters are not necessarily optimal. In the more general setting, there is a target clustering X̄ = X̄1 , ..., X̄k of the given dataset X ⊆ Rd (not necessarily optimal clusters) such that these clusters satisfy the γ-margin property (i.e., for all i, x ∈ X̄i , and y ∈ / X̄i , γ · ||x − µ(X̄i )||< ||y − µ(X̄i )||) and the goal of the query algorithm is to output the clustering X̄. Here is the main result of Ashtiani et al. Theorem 2 (Theorems 7 and 8 in [AKBD16]). Let δ ∈ (0, 1) and γ > 1. Let X ⊆ Rd be any dataset containing n points, k be a positive integer, and X1 , ..., Xk be any target clustering  of X that satisfies the γ-margin property. Then there is a query algorithm A that makes k+log 1/δ O k log n + k2 log(γ−1) same-cluster queries and with probability at least (1 − δ) outputs the 4   k+log 1/δ clustering X1 , ..., Xk . The running time of algorithm A is O kn log n + k2 log(γ−1) . 4 The above result is a witness to the power of the SSAC framework. We extend this line of work by examining the power of same-cluster queries in the approximation algorithms domain. Our results do not assume any separation condition on the dataset (such as γ-margin as in [AKBD16]) and they hold for any dataset. Since the k-means problem is NP-hard, an important line of research work has been to obtain approximation algorithms for the problem. There are various efficient approximation algorithms for the k-means problem, the current best approximation guarantee being 6.357 by Ahmadian et al. [ANFSW16]. A simple approximation algorithm that gives an O(log k) approximation guarantee is the k-means++ seeding algorithm (also known as D 2 -sampling algorithm) by Arthur and Vassilvitskii [AV07]. This algorithm is commonly used in solving the k-means problem in practice. As far as approximation schemes or in other words (1 + ε)-approximation algorithms (for arbitrary small ε < 1) are concerned, the following is known: It was shown by Awasthi et al. [ACKS15] that there is some fixed constant 0 < c < 1 such that there cannot exist an efficient (1 + c) factor approximation unless P = NP. This result was improved by Lee et al. [LSW17] where it was shown that it is NP-hard to approximate the k-means problem within a factor of 1.0013. However, when either k or d is a fixed constant, then there are Polynomial Time Approximation Schemes (PTAS) for the k-means problem.4 Addad et al. [CAKM16] and Friggstad et al. [FRS16] gave a PTAS for the k-means problem in constant dimension. For fixed constant k, various PTASs are known [KSS10, FMS07, JKS14, JKY15]. Following is the main question that we explore in this work: For arbitrary small ε > 0, how many same-cluster queries must be made in an efficient (1 + ε)-approximation algorithm for k-means in the SACC framework? The running time should be polynomial in all input parameters such as n, k, d and also in 1/ε. Note that this is a natural extension of the main question explored by Ashtiani et al. [AKBD16]. Moreover, we have removed the separation assumption on the data. We provide an algorithm that makes poly(k/ε) same-cluster queries and runs in time O(nd · poly(k/ε)). More specifically, here is the formal statement of our main result: Theorem 3 (Main result: query algorithm). Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd . Then there is a query algorithm A that runs in time O(ndk 9 /ε4 ) and with probability at least 0.99 outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of 4 This basically means an algorithm that runs in time polynomial in the input parameters but may run in time exponential in 1/ε. same-cluster queries used by A is O(k9 /ε4 ). Here ∆k (X) denotes the optimal value of the k-means objective function. Note that unlike Theorem 2, our bound on the number of same-cluster queries is independent of the size of the dataset. We find this interesting and the next natural question we ask is whether this bound on the number of same-cluster queries is tight in some sense. In other words, does there exist a query algorithm in the SSAC setting that gives (1+ε)-approximation in time polynomial in n, k, d and makes significantly fewer queries than the one given in the theorem above? We answer this question in the negative by establishing a conditional lower bound on the number of same-cluster queries under the assumption that ETH (Exponential Time Hypothesis) [IP01, IPZ01] holds. The formal statement of our result is given below. Theorem 4 (Main result: query lower bound). If the Exponential Time Hypothesis (ETH) holds, then there exists a constant c > 1 such that any c-approximation query algorithm for the k queries. k-means problem that runs in time poly(n, d, k) makes at least poly log k Faulty query setting The existence of a same-cluster oracle that answers such queries perfectly may be too strong an assumption. A more reasonable assumption is the existence of a faulty oracle that can answer incorrectly but only with bounded probability. Our query approximation algorithm can be extended to the setting where answers to the same-cluster queries are faulty. More specifically, we can get wrong answers to queries independently but with probability at most some constant q < 1/2. Also note that in our model the answer for a same-cluster query does not change with repetition. This means that one cannot ask the same query multiple times and amplify the probability of correctness. We obtain (1 + ε)-approximation guarantee for the k-means clustering problem in this setting. The main result is given as follows. Theorem 5. Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd . Consider a faulty SSAC setting where the response to every same-cluster query is incorrect with probability at most some constant q < 1/2. In such a setting, there is a query algorithm AE that with probability at least 0.99, outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of same-cluster queries used by AE is O(k15 /ε8 ). The previous theorems summarise the main results of this work which basically explores the power of same-cluster queries in designing fast (1 + ε)-approximation algorithms for the k-means problem. We will give the proofs of the above theorems in Sections 3, 4, and 5. There are some other simple and useful contexts, where the SSAC framework gives extremely nice results. One such context is the popular k-means++ seeding algorithm. This is an extremely simple sampling based algorithm for the k-means problem that samples k centers in a sequence of k iterations. We show that within the SSAC framework, a small modification of this sampling algorithm converts it to one that gives constant factor approximation instead of O(log k)-approximation [AV07] that is known. This is another witness to the power of same-cluster queries. We begin the technical part of this work by discussing this result in Section 2. Some of the basic techniques involved in proving our main results will be introduced while discussing this simpler context. Other related work Clustering problems have been studied in different semi-supervised settings. Basu et al. [BBM04] explored must-link and cannot-link constraints in their semi-supervised clustering formulation. In their framework, must-link and cannot-link constraints were provided explicitly as part of the input along with the cost of violating these constraints. They gave an active learning formulation for clustering in which an oracle answers whether two query points belong to the same cluster or not, and gave a clustering algorithm using these queries. However, they work with a different objective function and there is no discussion on theoretical bounds on the number of queries. In contrast, in our work we consider the k-means objective function and provide bounds on approximation guarantee, required number of adaptive queries, and the running time. Balcan and Blum [BB08] proposed an interactive framework for clustering with split/merge queries. Given a clustering C = {C1 , . . .}, a user provides feedback by specifying that some cluster Cl should be split, or clusters Ci and Cj should be merged. Awasthi et al. [ABV14] studied a local interactive algorithm for clustering with split and merge feedbacks. Voevodski et al. [VBR+ 14] considered one versus all queries where query answer for a point s ∈ X returns distances between s to all points in X. For a k-median instance satisfying (c, ε)-approximation stability property [BBG09], the authors found a clustering close to true clustering using only O(k) one versus all queries. Vikram and Dasgupta [VD16] designed an interactive bayesian hierarchical clustering algorithm. Given dataset X, the algorithm starts with a candidate hierarchy T , and an initially empty set C of constraints. The algorithm queries user with a subtree T |S of hierarchy T restricted to constant sized set S ⊂ X of leaves. User either accepts T |S or provides a counterexample triplet ({a, b}, c) which the algorithm adds to its set of constraints C, and updates T . They consider both random and adaptive ways to select S to query T |S . Our Techniques We now give a brief outline of the new ideas needed for our results. Many algorithms for the k-means problem proceed by iteratively finding approximations to the optimal centers. One such popular algorithm is the k-means++ seeding algorithm [AV07]. In this algorithm, one builds a set of potential centers iteratively. We start with a set C initialized to the empty set. At each step, we sample a point with probability proportional to the square of the distance from C, and add it to C. Arthur and Vassilvitskii [AV07] showed that if we continue this process till |C| reaches k, then the corresponding k-means solution has expected cost O(log k) times the optimal k-means cost. Aggarwal et al. [ADK09] showed that if we continue this process till |C| reaches βk, for some constant β > 1, then the corresponding k-means solution (where we actually open all the centers in C) has cost which is within constant factor of the optimal k-means cost with high probability. Ideally, one would like to stop when size of C reaches k and obtain a constant factor approximation guarantee. We know from previous works [AV07, BR13, BJA16] that this is not possible in the classical (unsupervised) setting. In this work, we show that one can get such a result in the SSAC framework. A high-level way of analysing the k-means++ seeding algorithm is as follows. We first observe that if we randomly sample a point from a cluster, then the expected cost of assigning all points of this cluster to the sampled point is within a constant factor of the cost of assigning all the points to the mean of this cluster. Therefore, it suffices to select a point chosen uniformly at random from each of the clusters. Suppose the set C contains such samples for the first i clusters (of an optimal solution). If the other clusters are far from these i clusters, then it is likely that the next point added to C belongs to a new cluster (and perhaps is close to a uniform sample). However to make this more probable, one needs to add several points to C. Further, the number of samples that needs to be added to C starts getting worse as the value of i increases. Therefore, the algorithm needs to build C till its size becomes O(k log k). In the SSAC framework, we can tell if the next point added in C belongs to a new cluster or not. Therefore, we can always ensure that |C| does not exceed k. To make this idea work, we need to extend the induction argument of Arthur and Vassilvitskii [AV07] – details are given in Section 2. We now explain the ideas for the PTAS for k-means. We consider the special case of k = 2. Let X1 and X2 denote the optimal clusters with X1 being the larger cluster. Inaba et al. [IKI94] showed that if we randomly sample about O(1/ε) points from a cluster, and let µ′ denote the mean of this subset of sampled points, then the cost of assigning all points in the cluster to µ′ is within (1 + ε) of the cost of assigning all these points to their actual mean (whp). Therefore, it is enough to get uniform samples of size about O(1/ε) from each of the clusters. Jaiswal et al. [JKS14] had the following approach for obtaining a (1 + ε)-approximation algorithm for k-means (with running time being nd · f (k, ε), where f is an exponential function of k/ε) – suppose we sample about O(1/ε2 ) points from the input, call this sample S. It is likely to contain at least O(1/ε) from X1 , but we do not know which points in S are from X1 . Jaiswal et al. addressed this problem by cycling over all subsets of S. In the SSAC model, we can directly partition S into S ∩ X1 and S ∩ X2 using |S| same-cluster queries. Having obtained such a sample S, we can get a close approximation to the mean of X1 . So assume for sake of simplicity that we know µ1 , the mean of X1 . Now we are faced with the problem of obtaining a uniform sample from X2 . The next idea of Jaiswal et al. is to sample points with probability proportional to square of distance from µ1 . This is known as D 2 -sampling. Suppose we again sample about O(1/ε2 ) such points, call this sample S ′ . Assuming that the two clusters are far enough (otherwise the problem only gets easier), they show that S ′ will contain about O(1/ε2 ) points from X2 (with good probability). Again, in the SSAC model, we can find this subset by |S ′ | queries – call this set S ′′ . However, the problem is that S ′′ may not represent a uniform sample from X2 . For any point e ∈ X2 , let pe denote the conditional probability of sampling e given that a point from X2 is sampled when sampled using D 2 -sampling. They showed ε , where m denotes the size of X2 . In order for the sampling lemma of Inaba et al. pe is at least m [IKI94] to work, we cannot work with approximately uniform sampling. The final trick of Jaiswal et al. was to show that one can in fact get a uniform sample of size about O(ε|S ′′ |) = O(1/ε) from S ′′ . The idea is as follows – for every element e ∈ S ′′ , we retain it with probability peεm (which is at most 1), otherwise we remove it from S ′′ . It is not difficult to see that this gives a uniform sample from X2 . The issue is that we do not know m. Jaiswal et al. again cycle over all subsets of S ′ – we know that there is a (large enough) subset of S ′ which will behave like a uniform sample from X2 . In the SSAC framework, we first identify the subset of S ′ which belongs to X2 , call this S ′′ (as above). Now we prune some points from S ′′ such that the remaining points behave like a uniform sample. This step is non-trivial because as indicated above, we do not know the value m. Instead, 2 ε and m for most of the points of X2 . Therefore, S ′′ is likely to we first show that pe lies between m εpe contain such a nice point, call it v. Now, for every point e ∈ S ′′ , we retain it with probability 2p v (which we know is at most 1). This gives a uniform sample of sufficiently large size from X2 . For k larger than 2, we generalize the above ideas using a non-trivial induction argument. 2 k-means++ within SSAC framework The k-means++ seeding algorithm, also known as the D 2 -sampling algorithm, is a simple sampling procedure that samples k centers in k iterations. The description of this algorithm is given below. The algorithm picks the first center randomly from the set X of points and after having picked the first (i − 1) centers denoted by Ci−1 , it picks a point x ∈ X to be the ith center with probability proportional to minc∈Ci−1 ||x − c||2 . The running time of k-means++ seeding algorithm is clearly O(nkd). Arthur and Vassilvitskii [AV07] showed that this simple sampling procedure gives an O(log k) approximation in expectation for any dataset. Within the SSAC framework where the algorithm is allowed to make same-cluster queries, we can make a tiny addition to the k-means++ seeding algorithm to obtain a query algorithm that gives constant factor approximation guarantee and makes only O(k2 log k) same-cluster queries. The description of the query algorithm is given in Table 1 (see right). In iteration i > 1, instead of simply accepting the sampled point x as the ith center (as done in k-means++ seeding algorithm), the sampled point x is accepted only if it belongs to a cluster other than those to which centers in Ci−1 belong (if this does not happen, k-means++(X,k) Query-k-means++(X, k) - Randomly sample a point x ∈ X - Randomly sample a point x ∈ X - C ← {x} - C ← {x} - for i = 2 to k - for i = 2 to k - Sample x ∈ X using distribution D - for j = 1 to ⌈log k⌉ defined as D(x) = Φ(C,{x}) - Sample x ∈ X using distribution D Φ(C,X) - C ← C ∪ {x} defined as D(x) = Φ(C,{x}) Φ(C,X) - return(C) - if(NewCluster(C, x)){C ← C ∪ {x}; break} - return(C) NewCluster(C, x) - If(∃c ∈ C s.t. SameCluster(c, x)) return(false) - else return(true) Table 1. k-means++ seeding algorithm (left) and its adaptation in the SSAC setting (right) the sampling is repeated for at most ⌈log k⌉ times). Here is the main result that we show for the query-k-means++ algorithm. Theorem 6. Let X ⊆ Rd be any dataset containing n points and k > 1 be a positive integer. Let C denote the output of the algorithm Query-k-means++(X, k). Then E[Φ(C, X)] ≤ 24 · ∆k (X), where ∆k (X) denotes the optimal cost for this k-means instance. Furthermore, the algorithm makes O(k2 log k) same-cluster queries and the running time of the algorithm is O(nkd + k log k log n + k2 log k). The bound on the number of same-cluster queries is trivial from the algorithm description. For the running time, it takes O(nd) time to update the distribution D which is updated k times. This accounts for the O(nkd) term in the running time. Sampling an element from a distribution D takes O(log n) time (if we maintain the cumulative distribution etc.) and at most O(k log k) points are sampled. Moreover, determining whether a sampled point belongs to an uncovered cluster takes O(k) time. So, the overall running time of the algorithm is O(nkd+k log k log n+k 2 log k). We prove the approximation guarantee in the remaining discussion. We will use the following terminology. Let the optimal k clusters for dataset X are given as X1 , ..., Xk . For any i, ∆i (X) Pdenotes the optimal cost of the i-means problem on dataset X. Given this, note that ∆k (X) = ki=1 ∆1 (Xi ). For any non-empty center set C, we say that a point x is sampled from dataset X using D 2 -sampling w.r.t. center set C if the sampling probability of x ∈ X is given by D(x) = Φ(C,{x}) Φ(C,X) . The proof of Theorem 6 will mostly follow O(log k)-approximation guarantee proof of k-means++ seeding by Arthur and Vassilvitskii [AV07]. The next two lemmas from [AV07] are crucially used in the proof of approximation guarantee. Lemma 1. Let A be any optimal cluster and let c denote a point sampled uniformly at random from A. Then E[Φ({c}, A)] ≤ 2 · ∆1 (A). Lemma 2. Let C be any arbitrary set of centers and let A be any optimal cluster. Let c be a point sampled with D 2 -sampling with respect to the center set C. Then E[Φ(C ∪{c}, A)|c ∈ A] ≤ 8·∆1 (A). The first lemma says that a randomly sampled center from X provides a good approximation (in expectation) to the cost of the cluster to which it belongs. The second lemma says that for any center set C, given that a center c that is D 2 -sampled from X w.r.t. C belong to an optimal cluster A, the conditional expectation of the cost of the cluster A with respect to center set C ∪ {c} is at most 8 times the optimal cost of cluster A. Using the above two lemmas, let us try to qualitatively see why the k-means++ seeding algorithm behave well. The first center belongs to some optimal cluster A and from Lemma 1 we know that this center is good for this cluster. At the time the ith center is D 2 -sampled, there may be some optimal clusters which are still costly with respect to the center set Ci−1 . But then we can argue that it is likely that the ith sampled center c will belong to one of these costly clusters, and conditioned on the center being from one such cluster A, the cost of this cluster after adding c to the current center set is bounded using Lemma 2. The formal proof of O(log k) approximation guarantee in [AV07] involves setting up a clever induction argument. We give a similar induction based argument to prove Theorem 6. We prove the following lemma (similar to Lemma 3.3 in [AV07]). We will need the following definitions: For any center set C, an optimal cluster A is said to be “covered” if at least one point from A is in C, otherwise A is said to be “uncovered”. Let T be a union of a subset of the optimal clusters, then we will use the def. P notation ΦOP T (T ) = Xi ⊆T ∆1 (Xi ). Lemma 3. Let C ⊆ X be any set of centers such that the number of uncovered clusters w.r.t. C is u > 0. Let Xu denote the set of points of the uncovered clusters and Xc denote set of the points of the covered clusters. Let us run t iterations of the outer for-loop in Query-k-means++ algorithm such that t ≤ u ≤ k. Let C ′ denote the resulting set of centers after running t iterations of the outer for-loop. Then the following holds:   u−t t ′ · Φ(C, Xu ). (1) + E[Φ(C , X)] ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + k u Proof. Let us begin by analysing what happens when starting with C, one iteration of the outer forloop in query-k-means++ is executed. The following two observations will be used in the induction argument: Φ(C,Xc ) c) Observation 1: If Φ(C,X Φ(C,X) ≥ 1/2, then we have Φ(C,Xc )+Φ(C,Xu ) ≥ 1/2 which implies that Φ(C, Xu ) ≤ Φ(C, Xc ), and also Φ(C, X) ≤ 2 · Φ(C, Xc ). c) Observation 2: If Φ(C,X Φ(C,X) < 1/2, then the probability that no point will be added after one iteration  ⌈log k⌉ log k c) = k1 . is given by Φ(C,X < 12 Φ(C,X) We will now proceed by induction. We show that if the statement holds for (t − 1, u) and (t − 1, u − 1), then the statement holds for (t, u). In the basis step, we will show that the statement holds for t = 0 and u > 0 and u = t = 1. Basis step: Let us first prove the simple case of t = 0 and u > 0. In this case, C ′ = C. So, we have E[Φ(C ′ , X)] = Φ(C, X) which is at most the RHS of (1). Consider the case when u = t = 1. This means that there is one uncovered cluster and one iteration of the outer for-loop is executed. If a center from the uncovered cluster is added, then E[Φ(C ′ , X)] ≤ Φ(C, Xc ) + 8 · ΦOP T (Xu ) and if no center is picked, then Φ(C ′ , X) = Φ(C, X). The probability of adding a center from the log k  c) . So, we get E[Φ(C ′ , X)] ≤ p · (Φ(C, Xc ) + 8 · uncovered cluster is given by p = 1 − Φ(C,X Φ(C,X) ΦOP T (Xu )) + (1 − p) · Φ(C, X). Note that this is upper bounded by the RHS of (1) by observing c) that 1 − p ≤ Φ(C,X Φ(C,X) . Inductive step: As stated earlier, we will assume that the statement holds for (t − 1, u) and def. c) 1 (t − 1, u − 1) and we will show that the statement holds for (t, u). Suppose p = Φ(C,X Φ(C,X) ≥ 2 , then Φ(C, X) ≤ 2 · Φ(C, Xc ) and so Φ(C ′ , X) ≤ Φ(C, X) ≤ 2 · Φ(C, Xc ) which is upper bounded by the RHS of (1). So, the statement holds for (t, u) (without even using the induction assumption). So, for the rest of the discussion, we will assume that p < 1/2. Let us break the remaining analysis into two cases – (i) no center is added in the next iteration of the outer for-loop, and (ii) a center is added. In case (i), u does not change, t decreases by 1, and the covered and uncovered clusters remain the same after the iteration. So the contribution of this case to E[Φ(C ′ , X)] is at most     t−1 u−t+1 ⌈log k⌉ p · (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) (2) + k u Now, consider case (ii). Let A be any uncovered cluster w.r.t. center set C. For any point a ∈ A, let pa denote the conditional probability of sampling a conditioned on sampling a point from A. Also, let φa denote the cost of A given a is added as a center. That is, φa = Φ(C ∪ {a}, A). The contribution of A to the expectation E[Φ(C ′ , X)] using the induction hypothesis is: (1 − p⌈log k⌉ ) · Φ(C, A) X pa Φ(C, Xu ) a ∈A     t−1 u−t · (Φ(C, Xc ) + φa + 8 · ΦOP T (Xu ) − 8 · ∆1 (A)) · 2 + · (Φ(C, Xu ) − Φ(C, A)) + k u−1 This is at most   t−1 u−t · (Φ(C, Xu ) − Φ(C, A)) (1 − p (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + + k u−1 P The previous inequality follows from the fact that a∈A pa φa ≤ 8 · ∆1 (A) from Lemma 2. Summing over all uncovered clusters, the overall contribution in case (ii) is at most:      Φ(C, Xu ) t−1 u−t · Φ(C, Xu ) − (1 − p⌈log k⌉ ) · (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + + k u−1 u ⌈log k⌉ Φ(C, A) )· Φ(C, Xu )   P The above bound is obtained using the fact that A is uncovered Φ(C, A)2 ≥ u1 Φ(C, Xu )2 . So the contribution is at most     t−1 u−t ⌈log k⌉ (1 − p ) · (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) (3) + k u Combining inequalities (2) and (3), we get the following:     t−1 Φ(C, Xu ) u−t ′ E[Φ(C , X)] ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) + p⌈log k⌉ · + k u u     t−1 u−t = (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) + + k u   Φ(C, Xc ) ⌈log k⌉ Φ(C, Xu ) · Φ(C, X) u     Φ(C, Xc ) u−t t−1 · Φ(C, Xu ) + + ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + k u ku (using the Observation 2, that is p⌈log k⌉ ≤ 1/k)   u−t t · Φ(C, Xu ) + ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + k u This completes the inductive argument and the proof. ⊔ ⊓ Let us now conclude the proof of Theorem 6 using the above lemma. Consider the center set C before entering the outer for-loop. This contains a single center c chosen randomly from the dataset X. Let c belong to some optimal cluster A. Let C ′ denote the center set after the execution of the outer for-loop completes. Applying the above lemma with u = t = k − 1, we get:   k−1 ′ E[Φ(C , X)] ≤ (Φ(C, A) + 8 · ∆k (X) − 8 · ∆1 (A)) · 2 + k ≤ 3 · (2 · ∆1 (A) + 8 · ∆k (X) − 8 · ∆1 (A)) (using Lemma 1) ≤ 24 · ∆k (X) 3 Query Approximation Algorithm (proof of Theorem 3) As mentioned in the introduction, our query algorithm is based on the D 2 -sampling based algorithm of Jaiswal et al. [JKS14, JKY15]. The algorithm in these works give a (1 + ε)-factor approximation for arbitrary small ε > 0. The running time of these algorithms are of the form nd · f (k, ε), where f is an exponential function of k/ε. We now show that it is possible to get a running time which is polynomial in n, k, d, 1/ε in the SSAC model. The main ingredient in the design and analysis of the sampling algorithm is the following lemma by Inaba et al. [IKI94]. Lemma 4 ([IKI94]). Let S be a set of points obtained by independently sampling M points uniformly at random with replacement from a point set X ⊂ Rd . Then for any δ > 0,     1 Pr Φ({µ(S)}, X) ≤ 1 + · ∆1 (X) ≥ (1 − δ). δM Here µ(S) denotes the geometric centroid of the set S. That is µ(S) = P s∈S s |S| Our algorithm Query-k-means is described in Table 2. It maintains a set C of potential centers of the clusters. In each iteration of step (3), it adds one more candidate center to the set C (whp), and so, the algorithm stops when |C| reaches k. For sake of explanation, assume that optimal clusters are X1 , X2 , . . . , Xk with means µ1 , . . . , µk respectively. Consider the ith iteration of step (3). At this time |C|= i − 1, and it has good approximations to means of i − 1 clusters among X1 , . . . .Xk . Let us call these clusters covered. In Step (3.1), it samples N points, each with probability proportional to square of distance from C (D 2 -sampling). Now, it partitions this set, S, into S ∩ X1 , . . . , S ∩ Xk in the procedure UncoveredClusters, and then picks the partition with the largest size such that the corresponding optimal cluster Xj is not one of the (i − 1) covered clusters. Now, we would like to get a uniform sample from Xj – recall that S ∩Xj does not represent a uniform sample. However, as mentioned in the introduction, we need to find an element s of Xj for which the probability of being in sampled is small enough. Therefore, we pick the element in S ∩ Xj for which this probability is smallest (and we will show that it has the desired properties). The procedure UncoveredCluster returns this element s. Finally, we choose a subset T of S ∩ Xj in the procedure UniformSample. This procedure is designed such that each element of Xj has the same probability of being in T . In step (3.4), we check whether the multi-set T is of a desired minimum size. We will argue that the probability of T not containing sufficient number of points is very small. If we have T of the desired size, we take its mean and add it to C in Step (3.6). We now formally prove the approximation guarantee of the Query-k-means algorithm. Theorem 7. Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd . There exists an algorithm that runs in time O(ndk9 /ε4 ) and with probability at least 14 outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of same-cluster queries used by the algorithm is O(k9 /ε4 ). 12 3 23 2 Constants: N = (2 ε2)k , M = 64k , L = (2 ε4)k ε Query-k-means(X, k, ε) UncoveredCluster(C, S, R) (1) R ← ∅ - For all i ∈ {1, ..., k}: Si ← ∅ (2) C ← ∅ -i←1 (3) for i = 1 to k - For all y ∈ R: {Si ← y; i++} (3.1) D2 -sample a multi-set S of N points - For all s ∈ S: from X with respect to center set C - If (∃j, y s.t. y ∈ Sj & SameCluster(s, y)) (3.2) s ← UncoveredCluster(C, S, R) - Sj ← Sj ∪ {s} (3.3) T ← UniformSample(X, C, s) - else (3.4) If (|T |< M ) continue - Let i be any index s.t. Si is empty (3.5) R ← R ∪ {s} - Si ← {s} (3.6) C ← C ∪ µ(T ) - Let Si be the largest set such that i > |R| (4) return(C) - Let s ∈ Si be the element with smallest UniformSample(X, C, s) value of Φ(C, {s}) in Si -T ←∅ - return(s) - For i = 1 to L: - D2 -sample a point x from X with respect to center set C - If (SameCluster(s, x))   - With probability ε 128 · Φ(C,{s}) Φ(C,{x}) add x in multi-set T - return(T ) Table 2. Approximation algorithm for k-means(top-left frame). Note that µ(T ) denotes the centroid of T and D2 sampling w.r.t. empty center set C means just uniform sampling. The algorithm UniformSample(X, C, s) (bottom-left) returns a uniform sample of size Ω(1/ε) (w.h.p.) from the optimal cluster to which point s belongs. Note that the success probability of the algorithm may be boosted by repeating it a constant number of times. This will also prove our main theorem (that is, Theorem 3). We will assume that the dataset X satisfies (k, ε)-irreducibility property defined next. We will later drop this assumption using a simple argument and show that the result holds for all datasets. This property was also used in some earlier works [KSS10, JKS14]. Definition 3 (Irreducibility). Let k be a positive integer and 0 < ε ≤ 1. A given dataset X ⊆ Rd is said to be (k, ε)-irreducible iff ∆k−1 (X) ≥ (1 + ε) · ∆k (X). Qualitatively, what the irreducibility assumption implies is that the optimal solution for the (k − 1)-means problem does not give a (1+ ε)-approximation to the k-means problem. The following useful lemmas are well known facts. Lemma 5. For any dataset X ⊆ Rd and a point c ∈ Rd , we have: Φ({c}, X) = Φ(µ(X), X) + |X|·||c − µ(X)||2 . Lemma 6 (Approximate Triangle Inequality). For any three points p, q, r ∈ Rd , we have ||p − q||2 ≤ 2(||p − r||2 +||r − q||2 ) Let X1 , ..., Xk be optimal clusters of the dataset X for the k-means objective. Let µ1 , ..., µk denote the corresponding optimal k centers. That is, ∀i, µi = µ(Xi ). For all i, let mi = |Xi |. Also, P for all i, let ri = 5 x∈Xi ||x−µi || mi 2 . The following useful lemma holds due to irreducibility.5 This is Lemma 4 from [JKS14]. We give the proof here for self-containment. Lemma 7. For all 1 ≤ i < j ≤ k, ||µi − µj ||2 ≥ ε · (ri + rj ). Proof. For the sake of contradiction, assume that ||µi − µj ||2 < ε · (ri + rj ). WLOG assume that mi > mj . We have: Φ({µi }, Xi ∪ Xj ) = mi ri + mj rj + mj ||µi − µj ||2 (using Lemma 5) ≤ (1 + ε) · mi ri + (1 + ε) · mj rj (since mi > mj ) ≤ mi ri + mj rj + mj · ε · (ri + rj ) ≤ (1 + ε) · Φ({µi , µj }, Xi ∪ Xj ) This implies that the center set {µ1 , ..., µk } \ {µj } gives a (1 + ε)-approximation to the k-means objective which contradicts with the (k, ε)-irreducibility assumption on the data. ⊔ ⊓ Consider the algorithm Query-k-means in Figure 2. Let Ci = {c1 , ..., ci } denote the set of centers at the end of the ith iteration of the for loop. That is, Ci is the same as variable C at the end of iteration i. We will prove Theorem 7 by inductively arguing that for every i, there are i distinct clusters for which centers in Ci are good in some sense. Consider the following invariant: P(i): There exists a set of i distinct clusters Xj1 , Xj2 , ..., Xji such that  ε · ∆1 (Xjr ). ∀r ∈ {1, ..., i}, Φ({cr }, Xjr ) ≤ 1 + 16 P ε Note that a trivial consequence of P (i) is Φ(Ci , Xj1 ∪ ... ∪ Xji ) ≤ (1 + 16 ) · ir=1 ∆1 (Xjr ). We will show that for all i, P (i) holds with probability at least (1 − 1/k)i . Note that the theorem follows if P (k) holds with probability at least (1 − 1/k)k . We will proceed using induction. The base case P (0) holds since C0 is the empty set. For the inductive step, assuming that P (i) holds with probability at least (1 − 1/k)i for some arbitrary i ≥ 0, we will show that P (i + 1) holds with probability at least (1 − 1/k)i+1 . We condition on the event P (i) (that is true with probability at least (1 − 1/k)i ). Let Ci and Xj1 , ..., Xji be as guaranteed by the invariant P (i). For ease of notation and without loss of generality, let us assume that the index jr is r. So, Ci gives a good approximation w.r.t. points in the set X1 ∪ X2 ∪ .... ∪ Xi and these clusters may be thought of as “covered” clusters (in the approximation sense). Suppose we D 2 -sample a point p w.r.t. center set Ci . The probability that p belongs to some “uncovered cluster” Xr where r ∈ [i + 1, k] is given i ,Xr ) 2 as Φ(C Φ(Ci ,X) . If this quantity is small, then the points sampled using D sampling in subsequent iterations may not be good representatives for the uncovered clusters. This may cause the analysis to break down. However, we argue that since our data is (k, ε)-irreducible, this does not occur. 6 Lemma 8. Φ(Ci ,Xi+1 ∪...∪Xk ) Φ(Ci ,X) ≥ 4ε . Proof. For the sake of contradiction, assume that the above statement does not hold. Then we have: Φ(Ci , X) = Φ(Ci , X1 ∪ ... ∪ Xi ) + Φ(Ci , Xi+1 ∪ ... ∪ Xk ) (ε/4) ≤ Φ(Ci , X1 ∪ ... ∪ Xi ) + · Φ(Ci , X1 ∪ ... ∪ Xi ) (using our assumption) 1 − (ε/4) 1 · Φ(Ci , X1 ∪ ... ∪ Xi ) = 1 − ε/4 6 This is Lemma 5 in [JKS14]. We give the proof for self-containment. ≤ i 1 + ε/16 X ∆1 (Xj ) (using invariant) · 1 − ε/4 j=1 ≤ (1 + ε) · k X ∆1 (Xj ) j=1 ⊔ ⊓ This contradicts with the (k, ε)-irreducibility of X. The following simple corollary of the above lemma will be used in the analysis later. Corollary 1. There exists an index j ∈ {i + 1, ..., k} such that Φ(Ci ,Xj ) Φ(Ci ,X) ≥ ε 4k . The above corollary says that there is an uncovered cluster which will have a non-negligible representation in the set S that is sampled in iteration (i + 1) of the algorithm Query-k-means. The next lemma shows that conditioned on sampling from an uncovered cluster l ∈ {i + 1, ..., k}, ε times its sampling probability if it were the probability of sampling a point x from Xl is at least 64 ε · m1l ). 7 sampled uniformly from Xl (i.e., with probability at least 64 Lemma 9. For any l ∈ {i + 1, ..., k} and x ∈ Xl , Φ(Ci ,{x}) Φ(Ci ,Xl ) ≥ ε 64 · 1 ml . Proof. Let t ∈ {1, ..., i} be the index such that x is closest to ct among all centers in Ci . We have: Φ(Ci , Xl ) = ml · rl + ml · ||µl − ct ||2 (using Lemma 5) 2 ≤ ml · rl + 2ml · (||µl − µt || +||µt − ct ||2 ) (using Lemma 6) ε ≤ ml · rl + 2ml · (||µl − µt ||2 + · rt ) (using invariant and Lemma 5) 16 Also, we have: ||x − µt ||2 − ||µt − ct ||2 (using Lemma 6) 2 ||µl − µt ||2 ≥ − ||µt − ct ||2 (since ||x − µt ||≥ ||µl − µt ||/2) 8 ε ||µl − µt ||2 − · rt (using invariant and Lemma 5) ≥ 8 16 ||µl − µt ||2 ≥ (using Lemma 7) 16 Φ(Ci , {x}) = ||x − ct ||2 ≥ Combining the inequalities obtained above, we get the following: ||µl − µt ||2 Φ(Ci , {x})  ≥ Φ(Ci , Xl ) 16 · ml · rl + 2||µl − µt ||2 + 8ε · rt 1 1 ε 1 ≥ · ≥ · 16 · ml (1/ε) + 2 + (1/8) 64 ml This completes the proof of the lemma. ⊔ ⊓ With the above lemmas in place, let us now get back to the inductive step of the proof. Let J ⊆ {i + 1, ..., k} denote the subset of indices (from the uncovered cluster indices) such that 7 This is Lemma 6 from [JKS14]. We give the proof for self-containment. Φ(Ci ,Xj ) ε Φ(Ci ,X) ≥ 8k . For any 2 i ,{y}) Yj , Φ(C Φ(Ci ,Xj ) ≤ mj . That ∀j ∈ J, index j ∈ J, let Yj ⊆ Xj denote the subset of points in Xj such that is, Yj consists of all the points such that the conditional probability ∀y ∈ of sampling any point y in Yj , given that a point is sampled from Xj , is upper bounded by 2/mj . Note that from Lemma 9, the conditional probability of sampling a point x from Xj , given that a ε · m1j . This gives the following simple and useful point is sampled from Xj , is lower bounded by 64 lemma. Lemma 10. For all j ∈ {i + 1, ..., k} the following two inequalities hold: Φ(C ,Y ) Φ(C ,X ) j j ε 1. Φ(Cii ,X) · Φ(Cii ,X) ≥ 128 , and 2. For any y ∈ Yj and any x ∈ Xj , ε 128 · Φ(Ci , {y}) ≤ Φ(Ci , {x}). Φ(Ci ,{y}) ε Φ(Ci ,Xj ) ≥ 64 1 ε i ,{x}) Xj , Φ(C Φ(Ci ,Xj ) ≥ 64 · mj Proof. Inequality (1) follows from the fact that |Yj |≥ mj /2, and Xj . Inequality (2) follows from the fact that for all x ∈ i ,{y}) Yj , Φ(C Φ(Ci ,Xj ) ≤ 2 mj . · 1 mj for all y ∈ and for all y ∈ ⊔ ⊓ Let us see the outline of our plan before continuing with our formal analysis. What we hope to get in line (3.2) of the algorithm is a point s that belongs to one of the uncovered clusters with index in the set J. That is, s belongs to an uncovered cluster that is likely to have a good representation in the D 2 -sampled set S obtained in line (3.1). In addition to s belonging to Xj for some j ∈ J, we would like s to belong to Yj . This is crucial for the uniform sampling in line (3.3) to succeed. We will now show that the probability of s returned in line (3.2) satisfies the above conditions is large. Lemma 11. Let S denote the D 2 -sample obtained w.r.t. center set Ci in line (3.1) of the algorithm. Pr[∃j ∈ J such that S does not contain any point from Yj ] ≤ 1 . 4k Proof. We will first get bound on the probability for a fixed j ∈ J and then use the union bound. Φ(C ,Yj ) ε ε2 ε From property (1) of Lemma 10, we have that for any j ∈ J, Φ(Cii ,X) · 8k = (210 ≥ 128 . Since the )k 12 3 number of sampled points is N = (2 ε2)k , we get that the probability that S has no points from Yj is at most 4k12 . Finally, using the union bound, we get the statement of the lemma. ⊔ ⊓ Lemma 12. Let S denote the D 2 -sample obtained w.r.t. center set Ci in line (3.1) of the algorithm and let Sj denote the representatives of Xj in S. Let max = arg maxj∈{i+1,...,k}|Sj |. Then Pr[max ∈ / 1 J] ≤ 4k . Φ(C ,X ) j ε Proof. From Corollary 1, we know that there is an index j ∈ {i + 1, ..., k} such that Φ(Cii ,X) . ≥ 4k ε Let α = N · 4k . The expected number of representatives from Xj in S is at least α. So, from Chernoff bounds, we have: Pr[|Sj |≤ 3α/4] ≤ e−α/32 On the other hand, for any r ∈ {i + 1, ..., k} \ J, the expected number of points in S from Xr is at ε · N = α/2. So, from Chernoff bounds, we have: most 8k Pr[|Sr |> 3α/4] ≤ e−α/24 So, the probability that there exists such an r is at most k · e−α/24 by union bound. Finally, the probability that max ∈ / J is at most Pr[|Sj |≤ 3α/4] + Pr[∃r ∈ {i + 1, ..., k} \ J||Sr |> 3α/4] which 12 3 1 ⊔ ⊓ is at most 4k due to our choice of N = (2 ε2)k . 1 From the previous two lemmas, we get that with probability at least (1 − 2k ), the s returned in line (3.2) belongs to Yj for some j ∈ J. Finally, we will need the following claim to argue that the set T returned in line (3.3) is a uniform sample from one of the sets Xj for j ∈ {i + 1, ..., k}. Lemma 13. Let S denote the D 2 -sample obtained w.r.t. center set Ci in line (3.1) and s be the point returned in line (3.2) of the algorithm. Let j denote the index of the cluster to which s belongs. 1 ), T returned in line (3.3) is a uniform If j ∈ J and s ∈ Yj , then with probability at least (1 − 4k 64k sample from Xj with size at least ε . Proof. Consider the call to sub-routine UniformSample(X, Ci, s) with s as given in the statement of the lemma. If j is the index of the cluster to which s belongs, then j ∈ J and s ∈ Yj . Let us define L random variables Z1 , ..., ZL one for every iteration of the sub-routine. These random variables are defined as follows: for any r ∈ [1, L], if the sampled point x belongs to the same cluster as s and it is picked to be included in multi-set S, then Zr = x, otherwise Zr = ⊥. We first note that for any r and any x, y ∈ Xj , Pr[Zr = x] = Pr[Zr = y]. This is because for any x ∈ Xj , we have ε ·Φ(C ,{s}) ε ·Φ(C ,{s}) i i ε i ,{s}) i ,{x}) 128 128 · Φ(C Pr[Zr = x] = Φ(C = 128 ≤1 Φ(Ci ,X) · Φ(Ci ,{x}) Φ(Ci ,X) . It is important to note that Φ(Ci ,{x}) from property (2) of Lemma 10 and hence valid in the probability calculations above. Let us now obtain a bound on the size of T . Let Tr = I(Zr ) be the indicator variable that is 1 if Zr 6= ⊥ and 0 otherwise. Using the fact that j ∈ J, we get that for any r: P ε ε ε ε ε3 x∈Xj Φ(Ci , {s}) E[Tr ] = Pr[Tr = 1] = · ≥ · · = 16 . 128 Φ(Ci , X) 128 8k 64 (2 )k Given that L = 223 k 2 , ε4 applying Chernoff bounds, we get the following:      64k 1 64k = 1 − Pr |T |≤ ≥ 1− Pr |T |≥ ε ε 4k  This completes the proof of the lemma. ⊔ ⊓ Since a suitable s (as required by the above lemma) is obtained in line (3.2) with probability at 1 least (1− 2k ), the probability that T obtained in line (3.3) is a uniform sample from some uncovered 1 1 )·(1− 4k ). Finally, the probability that the centroid µ(T ) of the multi-set cluster Xj is at least (1− 2k 1 ) using Inaba’s lemma. Combining T that is obtained is a good center for Xj is at least (1 − 4k 1 everything, we get that with probability at least (1 − k ) an uncovered cluster will be covered in the ith iteration. This completes the inductive step and hence the approximation guarantee of (1 + ε) holds for any dataset that satisfies the (k, ε)-irreducibility assumption. For the number of queries and running time, note that every time sub-routine UncoveredCluster is called, it uses at most kN same cluster queries. For the subroutine UniformSample, the number of same-cluster queries made is L. So, the total number of queries is O(k(kN + L)) = O(k5 /ε4 ). More specifically, we have proved the following theorem. Theorem 8. Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd such that X is (k, ε)irrducible. Then Query-k-means(X, k, ε) runs in time O(ndk5 /ε4 ) and with probability at least (1/4) outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of samecluster queries used by Query-k-means(X, k, ε) is O(k5 /ε4 ). To complete the proof of Theorem 7, we must remove the irreducibility assumption in the above theorem. We do this by considering the following two cases: ε 1. Dataset X is (k, (4+ε/2)k )-irreducible. ε 2. Dataset X is not (k, (4+ε/2)k )-irreducible. In the former case, we can apply Theorem 8 to obtain Theorem 7. Now, consider the latter ε case. Let 1 < i ≤ k denote the largest index such that X is (i, (1+ε/2)k )-irreducible, otherwise i = 1. Then we have: k−i   ε ε · ∆k (X). · ∆k (X) ≤ 1 + ∆i (X) ≤ 1 + (4 + ε/2)k 4 This means that a (1 + ε/4)-approximation for the i-means problem on the dataset X gives the desired approximation for the k-means problem. Note that our approximation analysis works for the i-means problem with respect to the algorithm being run only for i steps in line (3) (instead of k). That is, the centers sampled in the first i iterations of the algorithm give a (1+ε/16)-approximation for the i-means problem for any fixed i. This simple observation is sufficient for Theorem 7. Note since Theorem 8 is used with value of the error parameter as O(ε/k), the bounds on the query and running time get multiplied by a factor of k4 . 4 Query Lower Bound (proof of Theorem 4) In this section, we will obtain a conditional lower bound on the number of same-cluster queries assuming the Exponential Time Hypothesis (ETH). This hypothesis has been used to obtain lower bounds in various different contexts (see [Man16] for reference). We start by stating the Exponential Time Hypothesis (ETH). Hypothesis 1 (Exponential Time Hypothesis (ETH)[IP01, IPZ01]): There does not exist an algorithm that can decide whether any 3-SAT formula with m clauses is satisfiable with running time 2o(m) . Since we would like to obtain lower bounds in the approximation domain, we will need a gap version of the above ETH hypothesis. The following version of the PCP theorem will be very useful in obtaining a gap version of ETH. Theorem 9 (Dinur’s PCP Theorem [Din07]). For some constants ε, d > 0, there exists a polynomial time reduction that takes a 3-SAT formula ψ with m clauses and produces another 3-SAT formula φ with m′ = O(m polylog m) clauses such that: – If ψ is satisfiable, then φ is satisfiable, – if ψ is unsatisfiable, then val(φ) ≤ 1 − ε, and – each variable of φ appears in at most d clauses. Here val(φ) denotes the maximum fraction of clauses of φ that are satisfiable by any assignment. The following new hypothesis follows from ETH and will be useful in our analysis. Hypothesis 2: There exists constants ε, d > 0 such that the following holds: There does not exist an algorithm that, given a 3-SAT formula ψ with m clauses with each variable appearing in atmost d clauses, distinguishes whether ψ is satisfiable or val(ψ) ≤ (1 − ε), Ω runs in time 2 m poly log m . The following simple lemma follows from Dinur’s PCP theorem given above. Lemma 14. If Hypothesis 1 holds, then so does Hypothesis 2. We now see a reduction from the gap version of 3-SAT to the gap version of the Vertex Cover problem that will be used to argue the hardness of the k-means problem. The next result is a standard reduction and can be found in a survey by Luca Trevisan [Tre04]. Lemma 15. Let ε, d > 0 be some constants. There is a polynomial time computable function mapping 3-SAT instances ψ with m variables and where each variable appears in at most d clauses, into graphs Gψ with 3m vertices and maximum degree O(d) such that if ψ is satisfiable, then Gψ has a vertex cover of size at most 2m and if val(ψ) ≤ (1 − ε), then every vertex cover of Gψ has size at least 2m(1 + ε/2). We formulate the following new hypothesis that holds given that hypothesis 2 holds. Eventually, we will chain all these hypothesis together. Hypothesis 3: There exists constants ε, d > 0 such that the following holds: There does not exist an algorithm that, given a graph G with n vertices and maximum degree d, distinguishes between the case when G has a vertex cover of size at most 2n/3 and the case when G has  a vertex cover of size at least 2n 3 Ω · (1 + ε), runs in time 2 n poly log n . The following lemma is a simple implication of Lemma 15 Lemma 16. If Hypothesis 2 holds, then so does Hypothesis 3. We are getting closer to the k-means problem that has a reduction from the vertex cover problem on triangle-free graphs [ACKS15]. So, we will need reductions from vertex cover problem to vertex cover problem on triangle-free graphs and then to the k-means problem. These two reductions are given by Awasthi et al. [ACKS15]. Lemma 17 (Follows from Theorem 21 [ACKS15]). Let ε, d > 0 be some constants. There is a polynomial-time computable function mapping any graph G = (V, E) with maximum degree d to a triangle-free graph Ĝ = (V̂ , Ê) such that the following holds: 3 2 –  |V̂ |= poly(d,1/ε)· |V | and maximum degree   of verticesin Ĝ is O(d /ε ), and Ĝ)| ≤ 1 − |V C( . – 1 − |V C(G)| ≤ (1 + ε) · 1 − |V C(G)| |V | |V | |V̂ | Here V C(G) denote the size of the minimum vertex cover of G. We can formulate the following hypothesis that will follow from Hypothesis 3 using the above lemma. Hypothesis 4: There exists constants ε, d > 0 such that the following holds: There does not exist an algorithm that, given a triangle-free graph G with n vertices and maximum 2n degree d, distinguishes between the case when G has a vertex cover of size at most 3 and  the case when G has a vertex cover of size at least 2n 3 Ω · (1 + ε), runs in time 2 The next claim is a simple application of Lemma 17. Lemma 18. If Hypothesis 3 holds, then so does Hypothesis 4. n poly log n . Finally, we use the reduction from the vertex cover problem in triangle-free graphs to the kmeans problem to obtain the hardness result for the k-means problem. We will use the following reduction from Awasthi et al. [ACKS15]. Lemma 19 (Theorem 3 [ACKS15]). There is an efficient reduction from instances of Vertex Cover (in triangle free graphs) to those of k-means that satisfies the following properties: – if the Vertex Cover instance has value k, then the k-means instance has cost at most (m − k) – if the Vertex Cover instance has value at least k(1 + ε), then the optimal k-means cost is at least m − (1 − Ω(ε))k. Here ε is some fixed constant > 0. Here m denotes the number of edges in the vertex cover instance. The next hypothesis follows from Hypothesis 4 due to the above lemma. Hypothesis 5: There exists constant c > 1 such that the following holds: There does not exist an algorithm that gives guarantee of c for the k-means problem that  an approximation  Ω runs in time poly(n, d) · 2 k poly log k . Claim. If Hypothesis 4 holds, then so does Hypothesis 5. Now using Lemmas 14, 16, 18, and 4, get the following result. Lemma 20. If the Exponential Time Hypothesis (ETH) holds then there exists a constant c > 1 such that any c-approximation algorithm for the k-means problem cannot have running time better   Ω than poly(n, d) · 2 k poly log k . This proves Theorem 4 since if there is a query algorithm that runs in time poly(n, d, k) and makes polyklog k same-cluster queries, then we can convert it to a non-query algorithm that runs in k time poly(n, d) · 2 poly log k in a brute-force manner by trying out all possible answers for the queries and then picking the best k-means solution. 5 Query Approximation Algorithm with Faulty Oracle In this section, we describe how to extend our approximation algorithm for k-means clustering in the SSAC framework when the oracle is faulty. That is, the answer to the same-cluster query may be incorrect. Let us denote the faulty same-cluster oracle as OE . We consider the following error model: for a query with points u and v, the query answer OE (u, v) is wrong independently with probability at most q that is strictly less than 1/2. More specifically, if u and v belong to the same optimal cluster, then OE (u, v) = 1 with probability at least (1 − q) and OE (u, v) = 0 with probability at most q. Similarly, if u and v belong to different optimal clusters, then OE (u, v) = 1 with probability at most q and OE (u, v) = 0 with probability at least (1 − q). The modified algorithm giving (1 + ε)-approximation for k-means with faulty oracle OE is given in Figure 3. Let X1 , . . . , Xk denote the k optimal clusters for the dataset X. Let C = {c1 , . . . , ci } denote the set of i centers chosen by the algorithm at the end of iteration i. Let S denote the sample obtained using D 2 -sampling in the (i + 1)st iteration. The key idea for an efficient (1 + ε)approximation algorithm for k-means in the SSAC framework with a perfect oracle was the following. Given a sample S, we could compute using at most k|S| same-cluster queries the partition S1 , . . . , Sk of S among the k optimal clusters such that Sj = S ∩ Xj for all j. In the following, we discuss how we achieve this partitioning of S among k optimal clusters when the oracle OE is faulty. We reduce the problem of finding the partitions of S among the optimal clusters to the problem of recovering dense (graph) clusters in a stochastic block model (SBM). An instance of an SBM is created as follows. Given any arbitrary partition V1 , . . . , Vk of a set V of vertices, an edge is added between two vertices belonging to the same partition with probability at least (1 − q) and between two vertices in different partitions with probability at most q. We first construct an instance I of an SBM using the sample S. By querying the oracle OE with all pairs of points u, v in S, we obtain a graph I on |S| vertices, where vertices in I correspond to the points in S, and an edge exists in I between vertices u and v if OE (u, v) = 1. Since oracle OE errs with probability at most q, for any u, v ∈ Sj for some j ∈ [k], there is an edge between u and v with probability at least (1 − q). Similarly, there is an edge (u, v) ∈ I for any two points u ∈ Sy and v ∈ Sz , y 6= z belonging to different optimal clusters with probability at most q. Note that the instance I created in this manner would be an instance of an SBM. Since q < 1/2, this procedure, with high probability, creates more edges between vertices belonging to the same partition than the number of edges between vertices in different partitions. Intuitively, the partitions of S would correspond to the dense (graph) clusters in SBM instance I, and if there were no errors, then each partition would correspond to a clique in I. One way to figure out the partitions S1 , . . . , Sk would be to retrieve the dense (graph) clusters from the instance I. Ailon et al. [ACX15] gave a randomized algorithm to retrieve all large clusters of any SBM instance. Their algorithm on a graph of n vertices retrieves √ all clusters of size at least n with high probability. Their main result in our context is given as follows. 13 3 23 2 Constants: N = (2 ε2)k , M = 64k , L = (2 ε4)k ε Faulty-Query-k-means(X, k, ε) UncoveredCluster(C, S, J) (1) J ← ∅ - For all i ∈ {1, ..., k}: Si ← ∅ (2) C ← ∅ -i←1 (3) for i = 1 to k - For all y ∈ J: {Si ← y; i++} (3.1) D2 -sample a multi-set S of N points - T1 , . . . , Tl = PartitionSample(S) for l < k from X with respect to center set C - for j = 1, . . . , l (3.2) s ← UncoveredCluster(C, S, J) - if IsCovered(C, Tj ) is FALSE (3.3) T ← UniformSample(X, C, s) - if ∃t such that St = ∅, then St = Tj (3.4) If (|T |< M ) continue - Let Si be the largest set such that i > |J| (3.5) J ← J ∪ {s} - Let s ∈ Si be the element with smallest (3.6) C ← C ∪ µ(T ) value of Φ(C, {s}) in Si (4) return(C) - return(s) UniformSample(X, C, s) -S←∅ - For i = 1 to L: PartitionSample(S) - D2 -sample point x ∈ X with respect to center set C - Construct SBM instance I by querying OE (s, t) ∀s, t ∈ S - U = U ∪ {x} - Run cluster recovery algorithm of Ailon et al. [ACX15] on I - T1 , . . . , Tl = PartitionSample(U ) for l < k - Return T1 , . . . , Tl for l < k - for j = 1, . . . , l - If (IsCovered(s, Tj ) is TRUE)   - ∀x ∈ Tj , with probability in multi-set S - return (S) ε 128 · Φ(C,{s}) Φ(C,{x}) add x IsCovered(C,U) - for c ∈ C – if for majority of u ∈ U , OE (c, u) = 1, Return TRUE - Return FALSE Table 3. Approximation algorithm for k-means (top-left frame) using faulty oracle. Note that µ(T ) denotes the centroid of T and D2 -sampling w.r.t. empty center set C means just uniform sampling. The algorithm UniformSample(X, C, s) (bottom-left) returns a uniform sample of size Ω(1/ε) (w.h.p.) from the optimal cluster in which point s belongs. Lemma 21 ([ACX15]). There exists a polynomial time algorithm that, given an instance of a √ stochastic block model on n vertices, retrieves all clusters of size at least Ω( n) with high probability, provided q < 1/2. We use Lemma 21 to retrieve the large clusters from our SBM instance I. We also need to make sure that the sample S is such that its overlap with at least one uncovered optimal cluster is large, where an optimal cluster Sj for some j is uncovered if C ∩ Sj = ∅. More formally, we would require p the following: ∃j ∈ [k] such that |Sj |≥ Ω( |S|), and Xj is uncovered by C. From Corollary 1, given a set of centers C with |C|< k, there exists an uncovered cluster such that any point sampled ε using D 2 -sampling would belong to that uncovered cluster with probability at least 4k . Therefore, ε 2 in expectation, D -sample S would contain at least 4k |S| points from one such uncovered optimal p 2 cluster. In order to ensure that this quantity is at least as large as |S|, we need |S|= Ω( 16k ε2 ). Our bounds for N and L, in the algorithm, for the size of D 2 -sample S satisfy this requirement with high probability. This follows from a simple application of Chernoff bounds. 12 2 Lemma 22. For D 2 -sample S of size at least 2 ε2k , there is at least one partition Sj = S ∩ Xj among the partitions returned by the sub-routine PartitionSample corresponding to an uncovered 1 cluster Xj with probability at least (1 − 16k ). Proof. From Corollary 1, for any point p sampled using D 2 -sampling, the probability that point p ε . In expectation, the number of points sampled belongs to some uncovered cluster Xj is at least 4k 10 ε|S| 2 k from uncovered cluster Xj is E[|Sj |] = 4k = ε . Exact recovery using Lemma 21 requires |Sj | to 6 1 be at least 2 εk . Using Chernoff bounds, the probability of this event is at least (1 − 16k ). ⊔ ⊓ Following Lemma 22, we condition on the event that there is at least one partition corresponding to an uncovered cluster among the partitions returned by the sub-routine PartitionSample. Next, we figure out using the sub-routine IsCovered which of the partitions returned by PartitionSample are covered and which are uncovered. Let T1 , . . . , Tl be the partitions returned by PartitionSample where l < k. Sub-routine IsCovered determines whether a cluster is covered or uncovered in the following manner. For each j ∈ [l], we check whether Tj is covered by some c ∈ C. We query oracle OE with pairs (v, c) for v ∈ Tj and c ∈ C. If majority of the query answers for some c ∈ C is 1, we say cluster Tj is covered by C. If for all c ∈ C and some Tj , the majority of the query answers is 0, then we say Tj is uncovered by C. Using Chernoff bounds, we show that with high probability uncovered clusters would be detected. Lemma 23. With probability at least (1 − correctly by the sub-routine IsCovered. 1 16k ), all covered and uncovered clusters are detected Proof. First, we figure out the probability that any partition Tj for j ∈ [l] is detected correctly as covered or uncovered. Then we use union bound to bound the probability that all clusters are detected correctly. Recall that each partition returned by PartitionSample has size at least 6 |Tj | ≥ 2 εk for j ∈ [l]. We first compute for one such partition Tj and some center c ∈ C, the probability that majority of the queries OE (v, c) where v ∈ Tj are wrong. Since each query answer is wrong independently with probability q < 1/2, in expectation the number of wrong query answers would be q|Tj |. Using Chernoff bound, the probability that majority of the queries is wrong is at most e− 1 2 26 k (1− 2q ) 3ε . The probability that the majority of the queries is wrong for at least one center 6 1 2 − 23εk (1− 2q ) c ∈ C is at most ke probability at least (1 − 6 . Again using union bound all clusters are detected correctly with 1 2 2 k k2 e− 3ε (1− 2q ) ) ≥ (1 − 1 16k ). ⊔ ⊓ 1 With probability at least (1 − 8k ), given a D 2 -sample S, we can figure out the largest uncovered optimal cluster using the sub-routines PartitionSample and IsCovered. The analysis of the Algorithm 3 follows the analysis of Algorithm 2. For completeness, we compute the probability of success, and the query complexity of the algorithm. Note that s in line (3.2) of the Algorithm 3 1 1 is chosen correctly with probability (1 − 4k )(1 − 8k ). The uniform sample in line (3.3) is chosen 1 1 properly with probability (1 − 4k )(1 − 8k ). Since, given the uniform sample, success probability 1 ), overall the probability of success becomes (1 − k1 ). For using Inaba’s lemma is at least (1 − 4k 6 query complexity, we observe that PartitionSample makes O( kε8 ) same-cluster queries to oracle 4 OE , query complexity of IsCovered is O( kε4 ). Since PartitionSample is called at most k times, 7 total query complexity would be O( kε8 ). Note that these are bounds for dataset that satisfies (k, ε)irreducibility condition. For general dataset, we will use O(ε/k) as the error parameter. This causes the number of same-cluster queries to be O(k15 /ε8 ). 6 Conclusion and Open Problems This work explored the power of the SSAC framework defined by Ashtiani et al. [AKBD16] in the approximation algorithms domain. We showed how same-cluster queries allowed us to convert the popular k-means++ seeding algorithm into an algorithm that gives constant approximation for the k-means problem instead of the O(log k) approximation guarantee of k-means++ in the absence of such queries. Furthermore, we obtained an efficient (1+ε)-approximation algorithm for the k-means problem within the SSAC framework. This is interesting because it is known that such an efficient algorithm is not possible in the classical model unless P = NP. Our results encourages us to formulate similar query models for other hard problems. If the query model is reasonable (as is the SSAC framework for center-based clustering), then it may be worthwhile exploring its powers and limitations as it may be another way of circumventing the hardness of the problem. For instance, the problem closest to center-based clustering problems such as k-means is the correlation clustering problem. The query model for this problem may be similar to the SSAC framework. It will be interesting to see if same-cluster queries allows us to design efficient algorithms and approximation algorithms for the correlation clustering problem for which hardness results similar to that of k-means is known. References ABS12. ABV14. ACKS15. ACX15. ADK09. AKBD16. Pranjal Awasthi, Avrim Blum, and Or Sheffet. Center-based clustering under perturbation stability. Information Processing Letters, 112(1):49 – 54, 2012. Pranjal Awasthi, Maria-Florina Balcan, and Konstantin Voevodski. Local algorithms for interactive clustering. In ICML, pages 550–558, 2014. Pranjal Awasthi, Moses Charikar, Ravishankar Krishnaswamy, and Ali Kemal Sinop. The Hardness of Approximation of Euclidean k-Means. In Lars Arge and János Pach, editors, 31st International Symposium on Computational Geometry (SoCG 2015), volume 34 of Leibniz International Proceedings in Informatics (LIPIcs), pages 754–767, Dagstuhl, Germany, 2015. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik. Nir Ailon, Yudong Chen, and Huan Xu. Iterative and active graph clustering using trace norm minimization without cluster size constraints. Journal of Machine Learning Research, 16:455–490, 2015. Ankit Aggarwal, Amit Deshpande, and Ravi Kannan. Adaptive sampling for k-means clustering. In Irit Dinur, Klaus Jansen, Joseph Naor, and Jos Rolim, editors, Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, volume 5687 of Lecture Notes in Computer Science, pages 15–28. Springer Berlin Heidelberg, 2009. Hassan Ashtiani, Shrinu Kushagra, and Shai Ben-David. Clustering with same-cluster queries. In NIPS. 2016. ANFSW16. Sara Ahmadian, Ashkan Norouzi-Fard, Ola Svensson, and Justin Ward. Better guarantees for k-means and euclidean k-median by primal-dual algorithms. arXiv preprint arXiv:1612.07925, 2016. AV07. David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, SODA ’07, pages 1027–1035, Philadelphia, PA, USA, 2007. Society for Industrial and Applied Mathematics. BB08. Maria-Florina Balcan and Avrim Blum. Clustering with interactive feedback. In International Conference on Algorithmic Learning Theory, pages 316–328. Springer Berlin Heidelberg, 2008. BBG09. Maria-Florina Balcan, Avrim Blum, and Anupam Gupta. Approximate clustering without the approximation. In Proc. ACM-SIAM Symposium on Discrete Algorithms, pages 1068–1077, 2009. BBM04. Sugato Basu, Arindam Banerjee, and Raymond J. Mooney. Active semi-supervision for pairwise constrained clustering. In Proceedings of the Fourth SIAM International Conference on Data Mining, Lake Buena Vista, Florida, USA, April 22-24, 2004, pages 333–344, 2004. BJA16. Anup Bhattacharya, Ragesh Jaiswal, and Nir Ailon. Tight lower bound instances for k-means++ in two dimensions. Theor. Comput. Sci., 634(C):55–66, June 2016. BR13. Tobias Brunsch and Heiko Rglin. A bad instance for k-means++. Theoretical Computer Science, 505:19 – 26, 2013. CAKM16. Vincent Cohen-Addad, Philip N. Klein, and Claire Mathieu. Local search yields approximation schemes for k-means and k-median in euclidean and minor-free metrics. 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), 00:353–364, 2016. Das08. Sanjoy Dasgupta. The hardness of k-means clustering. Technical Report CS2008-0916, Department of Computer Science and Engineering, University of California San Diego, 2008. Din07. Irit Dinur. The pcp theorem by gap amplification. J. ACM, 54(3), June 2007. FMS07. Dan Feldman, Morteza Monemizadeh, and Christian Sohler. A PTAS for k-means clustering based on weak coresets. In Proceedings of the twenty-third annual symposium on Computational geometry, SCG ’07, pages 11–18, New York, NY, USA, 2007. ACM. FRS16. Zachary Friggstad, Mohsen Rezapour, and Mohammad R. Salavatipour. Local search yields a PTAS for k-means in doubling metrics. 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), 00:365–374, 2016. IKI94. Mary Inaba, Naoki Katoh, and Hiroshi Imai. Applications of weighted Voronoi diagrams and randomization to variance-based k-clustering: (extended abstract). In Proceedings of the tenth annual symposium on Computational geometry, SCG ’94, pages 332–339, New York, NY, USA, 1994. ACM. IP01. Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-sat. Journal of Computer and System Sciences, 62(2):367 – 375, 2001. IPZ01. Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly exponential complexity? Journal of Computer and System Sciences, 63(4):512 – 530, 2001. JKS14. Ragesh Jaiswal, Amit Kumar, and Sandeep Sen. A simple D2 -sampling based PTAS for k-means and other clustering problems. Algorithmica, 70(1):22–46, 2014. JKY15. Ragesh Jaiswal, Mehul Kumar, and Pulkit Yadav. Improved analysis of D2 -sampling based PTAS for k-means and other clustering problems. Information Processing Letters, 115(2):100 – 103, 2015. KSS10. Amit Kumar, Yogish Sabharwal, and Sandeep Sen. Linear-time approximation schemes for clustering problems in any dimensions. J. ACM, 57(2):5:1–5:32, February 2010. LSW17. Euiwoong Lee, Melanie Schmidt, and John Wright. Improved and simplified inapproximability for kmeans. Information Processing Letters, 120:40 – 43, 2017. Man16. Pasin Manurangsi. Almost-polynomial ratio ETH-hardness of approximating densest k-subgraph. CoRR, abs/1611.05991, 2016. MNV12. Meena Mahajan, Prajakta Nimbhorkar, and Kasturi Varadarajan. The planar k-means problem is NPhard. Theor. Comput. Sci., 442:13–21, July 2012. ORSS13. Rafail Ostrovsky, Yuval Rabani, Leonard J. Schulman, and Chaitanya Swamy. The effectiveness of lloyd-type methods for the k-means problem. J. ACM, 59(6):28:1–28:22, January 2013. Tre04. Luca Trevisan. Inapproximability of combinatorial optimization problems. CoRR, cs.CC/0409043, 2004. Vat09. Andrea Vattani. The hardness of k-means clustering in the plane. Technical report, Department of Computer Science and Engineering, University of California San Diego, 2009. VBR+ 14. Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and Yu Xia. Efficient clustering with limited distance information. CoRR, abs/1408.2045, 2014. VD16. Sharad Vikram and Sanjoy Dasgupta. Interactive bayesian hierarchical clustering. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 2081–2090, 2016.
8cs.DS
A Scheme-Driven Approach to Learning Programs from Input/Output Equations Jochen Burghardt jochen.burghardt alumni.tu-berlin.de Feb 2018 arXiv:1802.01177v1 [cs.LO] 4 Feb 2018 Abstract We describe an approach to learn, in a term-rewriting setting, function definitions from input/output equations. By confining ourselves to structurally recursive definitions we obtain a fairly fast learning algorithm that often yields definitions close to intuitive expectations. We provide a Prolog prototype implementation of our approach, and indicate open issues of further investigation. Key words: inductive functional programming Contents 1 Introduction 2 2 The task of learning functions 3 3 Learning functions by anti-unification 4 4 Learning functions by structural recursion 5 5 Possible extensions 10 A Example runs of our prototype implementation 15 References 19 1 1 Introduction This paper describes an approach to learn function definitions from input/output equations. 1 In trivial cases, a definition is obtained by syntactical anti-unification of the given i/o equations. In non-trivial cases, we assume a structurally recursive function definition, and transform the given i/o equations into equations for the employed auxiliary functions. The latter are learned from their i/o equations in turn, until a trivial case is reached. We came up with this approach in 1994 but didn’t publish it until today. In this paper, we explain it mainly along some learning examples, leaving a theoretical elaboration to be done. Also, we indicate several issues of improvement that should be investigated further. However, we provide at least a Prolog prototype implementation of our approach. In the rest of this section, we introduce the term-rewriting setting our approach works in. In Sect. 2, we define the task of function learning. In Sect. 3 and 4, we explain the base case and the inductive case of our approach, that is, how to learn trivial functions, and how to reduce learning sophisticated functions to learning easier functions, respectively. Section 5 sketches some ideas for possible extensions to our approach; it also shows its limitations. Some runs of our Prolog prototype are shown in Appendix A. nat list tree blist 1: 2: 3: 4: ::= ::= ::= ::= 0 nil null nl | s(nat) | nat::list | nd(tree, nat, tree) | o(blist) | i(blist) natural numbers lists of natural numbers binary trees of natural numbers list of binary digits Figure 1. Employed sort definitions We use a term-rewriting setting that is well-known from functional programming: A sort can be defined recursively by giving its constructors. For example, sort definition 1, shown in Fig. 1, defines the sort nat of all natural numbers in 0-s notation. In this example, we use 0 as a nullary, and s as a unary constructor. 5: 6: 7: 8: 9: 10: 11: + ∗ lgth app size dup add : : : : : : : nat × nat nat × nat list list × list tree nat blist × blist −→ −→ −→ −→ −→ −→ −→ nat nat nat list nat nat blist addition of natural numbers multiplication of natural numbers number of elements of a list concatenation of lists number of elements of a binary tree duplicating a natural number addition of binary numbers (lists) Figure 2. Employed function signatures A sort is understood as representing a possibly infinite set of ground constructor terms, 2 e.g. the 1 We will use henceforth “i/o equations” for brevity. We avoid calling them “examples” as this could cause confusion when we explain our approach along example sort definitions, example signatures, and example functions. 2 i.e. terms without variables, built only from constructor symbols 2 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: x+0 x+s(y) x∗0 x∗s(y) lgth(nil) lgth(x::y) app(nil, z) app(x::y, z) size(null) size(nd(x, y, z)) = = = = = = = = = = x s(x+y) 0 x+x∗y 0 s(lgth(y)) z x::app(y, z) 0 s(size(x)+size(z)) Figure 3. Example function definitions sort nat represents the set {0, s(0), s(s(0)), s(s(s(0))), . . .}. A function has a fixed signature; Fig. 2 gives some examples. The signature of a constructor can be inferred from the sort definition it occurs in, e.g. 0 :−→ nat and s : nat −→ nat. We don’t allow non-trivial equations between constructor terms, therefore, we have T1 = T2 iff T1 syntactically equals T2 , for all ground constructor terms T1 , T2 . A non-constructor function can be defined by giving a terminating ([DJ90, Sect.5.1, p.270]) term rewriting system for it such that its left-hand sides are sufficiently complete ([Gut77], [Com86], [DJ90, Sect.3.2, p.264]). Examples for function definitions are shown in Fig. 3. Given some functions f1 , . . . , fm defined by such a term rewriting system, for each i and each ground constructor terms T1 , . . . , Tn we can find a unique ground constructor term T such that fi (T1 , . . . , Tn ) = T . We then say that fi (T1 , . . . , Tn ) evaluates to T . Given a term T , we denote by vars(T ) the set of variables occurring in T . 2 The task of learning functions The problem our approach shall solve is the following. Given a set of sort definitions, a non-constructor function symbol f , its signature, and a set of input/output equations for f , construct a term rewriting system defining f such that it behaves as prescribed by the i/o equations. We say that we want to learn a definition for f , or sloppily, that we want to learn f , from the given i/o equations. For example, given sort definition 1, signature 10, and the following input/output ground equations 22: dup(0) = 0 23: dup(s(0)) = s(s(0)) 24: dup(s(s(0))) = s(s(s(s(0)))) 25: dup(s(s(s(0)))) = s(s(s(s(s(s(0)))))) we are looking for a definition of dup such that equations 3 22, 23, 24, and 25 hold. One such definition is dup(0) = 0 dup(s(x)) = s(s(dup(x))) 26: 27: We say that this definition covers the i/o equations 28: 29: 22, 23, 24, and 25. In contrast, a definition dup(0) = 0 dup(s(x)) = s(s(x)) would cover i/o equations 22 and 23, but neither 24 nor 25. We wouldn’t accept this definition, since we are interested only in function definitions that cover all given i/o equations. It is well-known that there isn’t a unique solution to our problem. In fact, given i/o equations f (L1 ) = R1 , . . . , f (Ln ) = Rn and an arbitrary function g of appropriate domain and range, e.g. the function defined by 3 f (x) = ( if x = L1 then R1 elif . . . elif x = Ln then Rn else g(x) fi ) trivially covers all i/o equations. Usually, the “simplest” function definitions are preferred, with “simplicity” being some user-defined measure loosely corresponding to term size and/or case-distinction count, like e.g. in [Bur05, p.8] and [Kit10, p.77]. However, the notion of simplicity depends on the language of available basic operations. 4 In the end, the notion of a “good” definition can hardly be defined more precisely than being one that meets common human prejudice. From our prototype runs we got the feeling that our approach often yields “good” definition in that sense. 3 Learning functions by anti-unification One of the simplest ways to obtain a function definition is to syntactically anti-unify the given i/o equations. Given i/o equations f (L11 , . . . ,Lm1 ) = R1 .. .. .. . . . f (L1n , . . . ,Lmn ) = Rn , let f ( L1 , . . . , Lm ) = R be their least general generalization (lgg for short, see [Plo70,Plo71,Rey70]). If the variable condition vars(R) ⊆ vars(L1 ) ∪ . . . ∪ vars(Lm ) holds, then the lgg will cover all n given i/o equations. 3 We use common imperative notation here for sake of readability. The “invariance theorem” in Kolmogorov complexity theory (e.g. [LV08, p.105, Thm.2.1.1]) implies that ∀L1 , L2 ∃c ∀x : |CL1 (x) − CL2 (x)| 6 c, where the Li range over Turing-complete algorithm description languages, c is a natural number, x ranges over i/o equation sets, and CL (x) denotes the length of the shortest function definition, written in L, that covers x. This theorem is sometimes misunderstood to enable a language-independent notion of simplicity; however, it does not, at least for small i/o example sets. 4 4 For example, assume we are to generate a definition for a unary function called g2 from the i/o equations 32: g2 ( 0 ) = s(s( 0 )) g2 ( s(s(0)) ) = s(s( s(s(0)) )) g2 (s(s(s(s(0))))) = s(s(s(s(s(s(0)))))) . 33: g2 ( 30: 31: We obtain the lgg x024 ) = s(s( x024 )) As another example, we can generate a definition for a binary function called g4 from the i/o equations 36: g4 (a, 0 ) = s( 0 ) g4 (a, s(0) ) = s( s(0) ) g4 (a,s(s(0))) = s(s(s(0))) . 37: g4 (a, y012 ) = s( y012 ) 34: 35: We obtain the lgg which satisfies the variable condition. 5 equations 36, 35, and 34. Hence when g4 is defined by equation 37, it covers i/o As a counter-example, the lgg of the above dup i/o equations 25 dup( 0 ) dup( s(0) ) dup( s(s(0)) ) dup(s(s(s(0)))) 38: dup( x0123 22 23 24 is computed as = 0 = s(s(0)) = s(s(s(s(0)))) = s(s(s(s(s(s(0)))))) )= x0246 which violates the variable condition, and thus cannot be used to reduce a term dup(T ) to a ground constructor term, i.e. to evaluate dup(T ). The above anti-unification approach can be extended in several ways, they are sketched in Sect. 5.1. However, in all but trivial cases, an lgg will violate the variable condition, and we need another approach to learn a function definition. 4 Learning functions by structural recursion For a function f that can’t be learned by Sect. 3, we assume a defining term rewriting system that follows a structural recursion scheme obtained from f ’s signature and a guessed argument position. For example, for the function dup with the signature given in 10 and the only possible argument position, 1, we obtain the schematic equations 39: 40: dup(0) = g1 dup(s(x)) = g2 (dup(x)) 5 Whenever applied to terms T1 , . . . , Tm that don’t start all with the same function symbol, Plotkin’s lgg algorithm returns a variable that uniquely depends on T1 , . . . , Tm . We indicate the originating terms by an index sequence; e.g. y012 was obtained as lgg(0, s(0), s(s(0))). 5 where g1 and g2 are fresh names of non-constructor functions. If we could learn appropriate definitions for g1 and g2 , we could obtain a definition for dup just by adding equations 39 and 40. The choice of g1 is obvious: 41: 22 39 0 = dup(0) = g1 In order to learn a definition for g2 , we need to obtain appropriate i/o examples for g2 from those for dup. Joining equation 40 with dup’s relevant i/o equations yields three i/o equations for g2 : 23 40 22 s(s(0)) = dup(s(0)) = g2 (dup(0)) = g2 (0) 24 40 23 s(s(s(s(0)))) = dup(s(s(0))) = g2 (dup(s(0))) = g2 (s(s(0))) 25 40 24 s(s(s(s(s(s(0)))))) = dup(s(s(s(0)))) = g2 (dup(s(s(0)))) = g2 (s(s(s(s(0)))) 30 31 32 A definition for g2 covering its i/o examples 30, 31, and 32 has already been derived by antiunification in Sect. 3 as 33 g2 (x024 ) = s(s(x024 )) . Altogether, we obtain the rewriting system dup(0) dup(s(x)) g1 g2 (x024 ) 39 40 41 33 = g1 = g2 (dup(x)) = 0 = s(s(x024 )) as a definition for dup that covers its i/o equations 22, 23, 24, and 25. Subsequently, this system may be simplified, by inlining, to dup(0) = 0 dup(s(x)) = s(s(dup(x))) 39 40 which is the usual definition of the dup function. Returning to the computation of i/o equations for g2 from those for dup, note that g2 ’s derived i/o equations 30, 31, and 32 were necessary in the sense that they must be satisfied by each possible definition of g2 that leads to dup covering its i/o equations (23, 24, and 25). Conversely, g2 ’s i/o equations were also sufficient in the sense that each possible definition of g2 covering them ensures that dup covers 23, 24, and 25, provided it covers 22: Proof of Proof of Proof of 23: 24: 25: 40 22 30 dup(s(0)) = g2 (dup(0)) = g2 (0) = s(s(0)) 40 23 31 dup(s(s(0))) = g2 (dup(s(0))) = g2 (s(s(0))) = s(s(s(s(0)))) 40 24 32 dup(s(s(s(0)))) = g2 (dup(s(s(0)))) = g2 (s(s(s(s(0))))) = s(s(s(s(s(s(0)))))) Observe that the above proofs are based just on permutations of the equation chains from 30, 31, and 32. Moreover, note that the coverage proof for dup(s(T )) relies on the coverage for dup(T ) already being proven. That is, the coverage proofs follow the employed structural recursion scheme. As for the base case, g1 ’s coverage of 41 is of course necessary and sufficient for dup’s coverage of 22. 6 4.1 Non-ground i/o equations As an example that uses i/o equations containing variables, consider the function lgth, with the signature given in 7. Usually, i/o equations for this functions are given in a way that indicates that the particular values of the list elements don’t matter. For example, an i/o equation like lgth(a::b::nil) = s(s(0)) is seen much more often than lgth(s(0)::0::nil) = s(s(0)). Our approach allows for variables in i/o equations, and treats them as universally quantified. That is, a non-ground i/o equation is covered by a function definition iff all its ground instances are. Assume for example we are given the i/o equations lgth(nil) lgth(a::nil) lgth(a::b::nil) lgth(a::b::c::nil) 42: 43: 44: 45: = = = = 0 . s(0) s(s(0)) s(s(s(0))) Given the signature of lgth (see 7) and argument position 1, we obtain a structural recursion scheme 46: lgth(nil) = g3 47: lgth(x::y) = g4 (x, lgth(y)) . Similar to the dup example, we get 48: 42 46 0 = lgth(nil) = g3 , and we can obtain i/o equations for g4 from those for lgth: 6 34 35 36 43 47 42 s(0) = lgth(a::nil) = g4 (a, lgth(nil)) = g4 (a, 0) 44 47 43 s(s(0)) = lgth(a::b::nil) = g4 (a, lgth(b::nil)) = g4 (a, s(0)) 45 47 44 s(s(s(0))) = lgth(a::b::c::nil) = g4 (a, lgth(b::c::nil)) = g4 (a, s(s(0))) Again, a function definition covering these i/o equation happens to have been derived by antiunification in Sect. 3: 37 g4 (a, y012 ) = s(y012 ) Altogether, equations 46, 47, 48, and 37 build a rewriting system for lgth that covers all its given i/o equations. By subsequently inlining g3 ’s and g4 ’s definition, we obtain a simplified definition for lgth: 49: lgth(nil) = 0 50: lgth(x::y) = s(lgth(y)) which agrees with the usual one found in textbooks. Similar to the ground case, g4 ’s derived i/o equations 6 34, 35, and 36 were necessary for lgth In the rightmost equation of each line, we employ a renaming substitution. For example, we apply {a 7→ b, b 7→ c} to i/o equation 44 in line 36. For this reason, our approach wouldn’t work if a, b, c were considered non-constructor constants rather than universally quantified variables. 7 covering its i/o equations. And as in the ground case, they are also sufficient: Proof of Proof of Proof of 43: 44: 45: 47 42 Again, renaming substitutions were used in the application of 4.2 34 lgth(a::nil) = g4 (a, lgth(nil)) = g4 (a, 0) = s(0) 47 43 35 lgth(a::b::nil) = g4 (a, lgth(b::nil)) = g4 (a, s(0)) = s(s(0)) 47 44 36 lgth(a::b::c::nil) = g4 (a, lgth(b::c::nil)) = g4 (a, s(s(0)))) = s(s(s(0))) 43 and 44. Functions of higher arity For functions with more than one argument, we have several choices of the argument on which to do the recursion. In these cases, we currently systematically try all argument positions 7 in succession. This is feasible since • our approach is quite simple, and hence fast to compute, and • we have a sharp and easy to compute criterion (viz. coverage 8 of all i/o examples) to decide whether recursion on a given argument was successful. For the function +, with the signature given in 5, and argument position 2, we obtain the structural recursion scheme 51: 52: x+0 = g5 (x) x+s(y) = g6 (x, x+y) . Appendix A.1 shows a run of our Prolog prototype implementation that obtains a definition for +. In Sect. 5.2, we discuss possible extensions of the structural recursion scheme, like simultaneous recursion. 4.3 Constructors with more than one recursion argument When computing a structural recursion scheme, we may encounter a sort s with a constructor that takes more than one argument of sort s. A common example is the sort of all binary trees (of natural numbers), as given in 3. The function size, with the signature given in 9, computes the size of such a tree, i.e. the total number of nd nodes. A recursion scheme for the size and argument position 1 looks like: 53: 54: size(null) = g9 size(nd(x, y, z)) = g10 (y, size(x), size(z)) 7 In particular, the recursive argument’s sort and the function’s result sort needn’t be related in any way, as the lgth example above demonstrates. 8 Checking if an i/o equation is covered by a definition requires executing the latter on the lhs arguments of the former. Our structural recursion approach ensures the termination of such computations, and establishes an upper bound for the number of rewrite steps. For example, g2 and g4 , defined in 33 and 37, respectively, need one such step, while their callers dup and lgth, defined in 39,40 and 46,47, respectively, need a linear amount of steps. An upper-bound expression for learned functions’ time complexity remains to be defined and proven. 8 In App. A.2, we show a prototype run to obtain a definition for size. 4.4 General approach In the previous sections, we have introduced our approach using particular examples. In this section, we sketch a more abstract and algorithmic description. Given a function and its signature f : s1 × . . . × sn −→ s, and given one of its argument positions 1 6 i 6 n, we can easily obtain a term rewriting system to define f by structural recursion on its ith argument. Assume in the definition of f ’s ith domain sort si we have an alternative si ::= . . . | c(s01 , . . . , s0l ) | . . . , assume {s0ν(1) , . . . , s0ν(m) } 63 si is the set of non-recursive arguments of the constructor c, and s0ρ(1) = . . . = s0ρ(k) = si are the recursive arguments of c. Let g be a new function symbol. We build an equation f (x1 , . . . , xi−1 , c(y1 , . . . , yl ), xi+1 , . . . , xn ) = g(x1 , . . . , xi−1 , xi+1 , . . . , xn , yν(1) , . . . , yν(m) , f (x1 , . . . , xi−1 , yρ(1) , xi+1 , . . . , xn ) ... f (x1 , . . . , xi−1 , yρ(k) , xi+1 , . . . , xn ) ) In a somewhat simplified presentation, we build the equation f (. . . , c(y1 , . . . , yl ), . . .) = g(. . . , f (..., yρ(1) , ...), . . . , f (..., yρ(k) , ...)). From the i/o equations for f , we often 9 can construct i/o equations for g: If we have an i/o equation that matches the above equation’s left-hand side, and we have all i/o equations needed to evaluate the recursive calls to f on its right-hand side, we can build an i/o equation equation for g. This way, we can reduce the problem of synthesizing a definition for f that reproduces the given i/o equations to the problem of synthesizing a definition for g from its i/o equations. As a base case for this process, we may synthesize non-recursive function definitions by anti-unification of the i/o equations. It should be possible to prove that f covers all its i/o equations iff g covers its, under some appropriate conditions. We expect that a sufficient condition is that all recursive calls to f could be evaluated. At least, we could demonstrate this in the above dup and lgth example. 9 Our construction isn’t successful in all cases. We give a counter-example in Sect. 5.3 9 Fct Eqn Lf Rg Fct Eqn Lf Rg 22 23 dup 24 25 2 3 4 5 1 g1 3 5 g2 7 41 30 31 32 1 2 4 6 1 3 5 7 Figure 4. Left- and right-hand term sizes of i/o equations for dup and g2 4.5 Termination In order to establish the termination of our approach, it is necessary to define a criterion by which g is easier to learn from it i/o equations than f is from its. Term size or height cannot be used in a termination ordering; when proceeding from f to g they may remain equal, or may even increase, as shown in Fig. 4 for the dup vs. g2 example. However, the number of i/o equations decreases in this example, and in all other ones we dealt with. A sufficient criterion for this is that f ’s i/o equations don’t all have the same left-hand side top-most constructor. However, the same criterion would have to be ensured in turn for g, and it is not obvious how to achieve this. In any case, by construction of g’s i/o example from f ’s, no new terms can arise. 10 Even more, each term appearing in an i/o example for g originates from a right-hand side of an i/o example for f . Therefore, our approach can’t continue generating new auxiliary functions forever, without eventually repeating the set of i/o equations. Our prototype implementation doesn’t check for such repetitions, however. 5 Possible extensions In this section, we briefly sketch some possible extensions of our approach. Their investigation in detail still remains to be done. 5.1 Extension of anti-unification In Sect. 3 we used syntactical anti-unification to obtain a function definition, as a base case of our approach. Several way to extend this technique can be thought of. Set anti-unification It can be tried to split the set of i/o equations into disjoint subsets such that from each one an lgg satisfying the variable condition is obtained. This results in several defining equations. An additional constraint might be that each subset corresponds to another constructor symbol, observed at some given fixed position in the left-hand side terms. 10 except for the fresh left-hand side top function symbols 10 sq( 0 ) sq( s(0) ) sq( s(s(0)) ) sq(s(s(s(0)))) sq( x0123 ) 56: 57: 58: 59: 60: = 0 = s(0) = s(s(s(s(0)))) = s9 (0) = x0123 ∗x0123 Figure 5. Application of E-anti-unification to learn squaring Anti-unification modulo equational theory Another extension consists in considering an equational background theory E in anti-unification; it wasn’t readily investigated in 1994. See [Hei94b,Hei94a,Hei95] for the earliest publications, and [Bur05,Bur17] for the latest. As of today, the main application of E-anti-unification turned out to be the synthesis of nonrecursive function definitions from input/ output equations [Bur17, p.3]. To sketch an example, let E consist just of definitions 12, 13, 14, and 15. Assume the signature 55: sq : nat −→ nat and the i/o equations 56, 57, 58, and 59 of the squaring function. Applying syntactical antiunification to the left-hand sides yields a variable x0123 , and four corresponding substitutions. Applying constrained E-generalization [Bur05, p.5, Def.2] to the right-hand sides yields a term set that contains x0123 ∗x0123 as a minimal-size member, see Fig. 5. Depth-bounded anti-unification In many cases, defining equations obtained by syntactical anti-unification appear to be too particular. For example, s4 (0) and s9 (0) are generalized to s4 (x05 ), while being by 4 greater than something wouldn’t be the first choice for a common property of both numbers for most humans. As a possible remedy, a maximal depth d may be introduced for the anti-unification algorithm. Beyond this depth, terms are generalized by a variable even if all their root function symbols agree. Denoting by lgg d (t1 , t2 ) the result of an appropriately modified algorithm, it should be easy to prove that lgg d (t1 , t2 ) can be instantiated to both t1 and t2 , and is the most special term with that property among all terms of depth up to d. If d is chosen as ∞, lgg d and lgg coincide. In our prototype implementation, we meanwhile built in such a depth boundary. Figure 6 compares the learned function definitions for size for d = 2, 3, 4 (top to bottom). For example, for d = 2, the —nonsensical— equation size(x) = 0 is learned, while for d > 3 the respective equation reads size(null) = 0. Not surprisingly, for d = 2 only one of the given 9 i/o equations is covered. For d 6 1, the attempt to learn defining equations for size fails. For d = 4, the learned equations agree with those for d = ∞, and hence also with those for all intermediate depths. The prototype run for d = ∞ is shown in App. A.2. Note that the prototype simplifies equations by removing irrelevant function arguments. For this reason, f12 there has only two arguments, while the corresponding function f1 in Fig. 6 has three. 11 x )=0 null null size( nd(x, y, z) ) = f1 ( y, size(x), size(z)) size( y f1 (x, 0 , 0 f1 (x, s(y) , z ) = s( z ) z ) = f2 ( x, z, f1 (x, y, z)) y z z f2 (x, 0 , s(z) ) = s( s(z) ) 0 s(0) s(0) y z z f2 (x, s(y) , s(z) ) = s( s(z) ) s(y) s(s(z)) s(s(z)) Figure 6. Learned tree size definition for anti-unification depth 2, 3, and 4 5.2 Extension of structural recursion Some functions are best defined by simultaneous recursion on several arguments. As an example, consider the sort definition 4 with nl, o, and i denoting an empty list, a 0 digit, and a 1 digit, respectively. For technical reasons, such a list is interpreted in reversed order, e.g. o(i(i(nl))) denotes the number 6. The sum function add, its signature shown in 11, may then be defined by the following rewrite system: 61: 62: 63: 64: 65: 66: add( x , nl ) add( nl , y ) add(o(x),o(y)) add(o(x), i(y) ) add( i(x) ,o(y)) add( i(x) , i(y) ) = = = = = = x y o(add(x, y)) i(add(x, y)) i(add(x, y)) o(inc(add(x, y))) where 67: inc : blist −→ blist is a function to increment a binary digit list. mentation, with inc being used for the carry. This corresponds to the usual hardware imple- It is obvious that this definition cannot be obtained from our simple structural recursion scheme from Sect. 4, neither by recurring over argument position 1 nor over 2. Instead, we would need 12 recursion over both positions simultaneously, i.e. a scheme like 68: 69: 70: 71: 72: 73: 74: 75: 76: add( nl , nl ) add( nl ,o(y)) add( nl , i(y) ) add(o(x), nl ) add(o(x),o(y)) add(o(x), i(y) ) add( i(x) , nl ) add( i(x) ,o(y)) add( i(x) , i(y) ) = = = = = = = = = g15 g16 (y) g17 (y) g18 (x) g19 (add(x, y)) g20 (add(x, y)) g21 (x) g22 (add(x, y)) g23 (add(x, y)) An extension of our approach could provide such a scheme, additionally to the simple structural recursion scheme. If we could prove that each function definition obtainable by the simple recursion scheme can also be obtained by a simultaneous recursion scheme, we needed only to employ the latter. This way, we would no longer need to guess an appropriate argument position to recur over; instead we could always recur simultaneously over all arguments of a given sort. Unfortunately, simultaneous recursion is not stronger than simple structural recursion. For example, the function app to concatenate two given lists can be obtained by simple recursion over the first argument (see 18,19 in Fig. 3), but not by simultaneous recursion: app(w::x, y::z) = g24 (w, y, app(x, z)) doesn’t lead to a sensible definition, for any choice of g24 . One possible remedy is to try simple structural recursion first, on any appropriate argument position, and simultaneous recursion next, on any appropriate set of argument positions. Alternatively, user commands may be required about which recursion to try on which argument position(s). Another possibility might be to employ a fully general structural recursion scheme, like 77: and 78: app(w::x, y::z) = g24 (w, y, app(w::x, z) , app(x, y::z) , app(x, z) ) add(o(x), o(y)) = g25 ( add(o(x), y) , add(x, o(y)) , add(x, y) ) . In this scheme, calls for simple recursion over each position are provided, as well as for simultaneous recursion over each position set. A new symbol Ω, intended to denote an undefined term, could be added to the term language. When e.g. i/o equations are missing to compute add(o(x), y) for some particular instance, the first argument of g25 would be set to Ω in the respective i/o equation. In syntactical anti-unification and coverage test, Ω needed to be handled appropriately. This way, only one recursion scheme would be needed, and no choice of appropriate argument position(s) would be necessary. However, arities of auxiliary functions might grow exponentially. 5.3 Limitations of our approach In this section, we demonstrate an example where our approach fails. Consider again the squaring function, its signature shown in 55, and consider again its i/o equations 56, 57, 58, 13 and 59. Since syntactical anti-unification as in Sect. 3 (i.e. not considering an equational background theory E) doesn’t lead to a valid function definition, we build a structural recursion scheme as in Sect. 4: 79: sq(0) = g11 80: sq(s(x)) = g12 (sq(x)) We get g11 = 0, and the following i/o equations for g12 : 57 80 56 s(0) = sq(s(0)) = g12 (sq(0)) = g12 (0) 58 80 57 s(s(s(s(0)))) = sq(s(s(0))) = g12 (sq(s(0))) = g12 (s(0)) 59 80 58 9 s (0) = sq(s(s(s(0)))) = g12 (sq(s(s(0)))) = g12 (s(s(s(s(0))))) 81: 82: 83: Observe that we are able to obtain i/o equations for g12 only on square numbers. For example, there is no obvious way to determine the value of g12 (s(s(s(0)))). Syntactically anti-unifying g12 ’s i/o equation still doesn’t yield a valid function definition. So we set up a recursion scheme for g12 , in turn: 84: 85: g12 (0) = g13 g12 (s(x)) = g14 (g12 (x)) Again, g13 = s(0) is obvious. Trying to obtain i/o equations for g14 , we get stuck, since we don’t know how g12 should behave on non-square numbers: 86: 87: 88: 89: s(s(s(s(0)))) ?? ?? 9 s (0) 82 = g12 (s(0)) ?? = g12 (s(s(0))) ?? = g12 (s(s(s(0)))) 83 = g12 (s(s(s(s(0))))) As an alternative, by applying 90: 85 85 = g14 (g12 (0)) 85 = g14 (g12 (s(0))) 85 = g14 (g12 (s(s(0)))) 85 = g14 (g12 (s(s(s(0))))) 81 = g14 (s(0)) = g14 (s(s(s(s(0))))) ?? = g14 (??) ?? = g14 (??) 82 sufficiently often rather than just once, we can obtain: 83 s9 (0) = g12 (s(s(s(s(0))))) 85 = g14 (g12 (s(s(s(0))))) 85 = g14 (g14 (g12 (s(s(0))))) 85 82 = g14 (g14 (g14 (g12 (s(0))))) = g14 (g14 (g14 (s(s(s(s(0))))))) However, no approach is known to learn g14 from an extended i/o equation like 90, which determines g14 ◦ g14 ◦ g14 rather than g14 itself. In such cases, we resort to the excuse that the original function, sq isn’t definable by structural recursion. A precise criterion for the class that our approach can handle is still to be found. It is not even clear that such a criterion can be computable. If not, it should still be possible to give computable necessary and sufficient approximations. 14 A A.1 Example runs of our prototype implementation Addition of 0-s numbers ?- SgI = [ + signature [nat,nat] --> nat], | SD = [ nat sortdef 0 ! s(nat)], | ExI = [ 0 + 0 = 0, | s(0) + 0 = s(0), | 0 + s(0) = s(0), | 0 + s(s(0)) = s(s(0)), | s(0) + s(0) = s(s(0)), | s(0) + s(s(0)) = s(s(s(0))), | s(s(0)) + s(0) = s(s(s(0))), | s(s(0)) + 0 = s(s(0))], | run(+,SgI,SD,ExI). +++++ Examples input check: +++++ Example 1: +++++ Example 2: +++++ Example 3: +++++ Example 4: +++++ Example 5: +++++ Example 6: +++++ Example 7: +++++ Example 8: +++++ Examples input check done induce(+) . trying argument position: 1 . inducePos(+,1,0) . . matching examples: [0+0=0,0+s(0)=s(0),0+s(s(0))=s(s(0))] . . anti-unifier: 0+v3 = v3 . inducePos(+,1,0) . inducePos(+,1,s(nat)) . . matching examples: [s(0)+0=s(0),s(0)+s(0)=s(s(0)),s(0)+s(s(0))=s(s(s(0))),s(s(0))+s(0)=s(s(s(0))),s(s(0))+0=s(s(0))] . . new recursion scheme: s(v9)+v8 = f10(v8,v9+v8) . . derive new equation: s(0) = s(0)+0 = f10(0,0) . . derive new equation: s(s(0)) = s(0)+s(0) = f10(s(0),s(0)) . . derive new equation: s(s(s(0))) = s(0)+s(s(0)) = f10(s(s(0)),s(s(0))) . . derive new equation: s(s(s(0))) = s(s(0))+s(0) = f10(s(0),s(s(0))) . . derive new equation: s(s(0)) = s(s(0))+0 = f10(0,s(0)) . . induce(f10) . . . trying argument position: 1 . . . inducePos(f10,1,0) . . . . matching examples: [f10(0,0)=s(0),f10(0,s(0))=s(s(0))] . . . . anti-unifier: f10(0,v13) = s(v13) . . . inducePos(f10,1,0) . . . inducePos(f10,1,s(nat)) . . . . matching examples: [f10(s(0),s(0))=s(s(0)),f10(s(0),s(s(0)))=s(s(s(0))),f10(s(s(0)),s(s(0)))=s(s(s(0)))] . . . . anti-unifier: f10(s(v15),s(v16)) = s(s(v16)) . . . inducePos(f10,1,s(nat)) . . . all examples covered . . induce(f10) . inducePos(+,1,s(nat)) . all examples covered induce(+) +++++ Examples output check: +++++ Examples output check done FUNCTION SIGNATURES: f10 signature [nat,nat]-->nat (+)signature[nat,nat]-->nat FUNCTION EXAMPLES: 0+0=0 s(0)+0=s(0) 0+s(0)=s(0) 0+s(s(0))=s(s(0)) s(0)+s(0)=s(s(0)) s(0)+s(s(0))=s(s(s(0))) s(s(0))+s(0)=s(s(s(0))) s(s(0))+0=s(s(0)) FUNCTION DEFINITIONS: 0+v17=v17 s(v18)+v19=f10(v19,v18+v19) f10(0,v20)=s(v20) f10(s(v21),s(v22))=s(s(v22)) ?- A.2 Size of a tree ?- SgI = [ size signature [tree] --> nat], | SD = [ tree sortdef nl ! nd(tree,nat,tree), | nat sortdef 0 ! s(nat)], | ExI = [ size(nl) = 0, 15 | size(nd(nl,va,nl)) = s(0), | size(nd(nd(nl,va,nl),vb,nl)) = s(s(0)), | size(nd(nl,va,nd(nl,vb,nl))) = s(s(0)), | size(nd(nd(nl,va,nl),vb,nd(nl,vc,nl))) = s(s(s(0))), | size(nd(nl,va,nd(nd(nl,vb,nl),vc,nl))) = s(s(s(0))), | size(nd(nl,va,nd(nl,vb,nd(nl,vc,nl)))) = s(s(s(0))), | size(nd(nd(nl,va,nl),vb,nd(nd(nl,vc,nl),vd,nl))) = s(s(s(s(0)))), | size(nd(nd(nd(nl,va,nl),vb,nl),vc,nd(nl,vd,nl))) = s(s(s(s(0)))) | ], | run(size,SgI,SD,ExI). +++++ Examples input check: +++++ Example 1: +++++ Example 2: +++++ Example 3: +++++ Example 4: +++++ Example 5: +++++ Example 6: +++++ Example 7: +++++ Example 8: +++++ Example 9: Variable sorts: [vd:nat,vc:nat,vb:nat,va:nat] +++++ Examples input check done induce(size) . trying argument position: 1 . inducePos(size,1,nl) . . matching examples: [size(nl)=0] . . anti-unifier: size(nl) = 0 . inducePos(size,1,nl) . inducePos(size,1,nd(tree,nat,tree)) . . matching examples: [size(nd(nl,va,nl))=s(0),size(nd(nd(nl,va,nl),vb,nl))=s(s(0)),size(nd(nl,va,nd(nl,vb,nl)))=s(s(0)),size(nd(nd(nl,va,nl),vb,n... . . new recursion scheme: size(nd(v10,v9,v11)) = f12(v9,size(v10),size(v11)) . . derive new equation: s(0) = size(nd(nl,va,nl)) = f12(va,0,0) . . derive new equation: s(s(0)) = size(nd(nd(nl,va,nl),vb,nl)) = f12(vb,s(0),0) . . derive new equation: s(s(0)) = size(nd(nl,va,nd(nl,vb,nl))) = f12(va,0,s(0)) . . derive new equation: s(s(s(0))) = size(nd(nd(nl,va,nl),vb,nd(nl,vc,nl))) = f12(vb,s(0),s(0)) . . derive new equation: s(s(s(0))) = size(nd(nl,va,nd(nd(nl,vb,nl),vc,nl))) = f12(va,0,s(s(0))) . . derive new equation: s(s(s(0))) = size(nd(nl,va,nd(nl,vb,nd(nl,vc,nl)))) = f12(va,0,s(s(0))) . . derive new equation: s(s(s(s(0)))) = size(nd(nd(nl,va,nl),vb,nd(nd(nl,vc,nl),vd,nl))) = f12(vb,s(0),s(s(0))) . . derive new equation: s(s(s(s(0)))) = size(nd(nd(nd(nl,va,nl),vb,nl),vc,nd(nl,vd,nl))) = f12(vc,s(s(0)),s(0)) . . induce(f12) . . . trying argument position: 1 . . . inducePos(f12,1,0) . . . . matching examples: [] . . . . no examples . . . inducePos(f12,1,0) . . . inducePos(f12,1,s(nat)) . . . . matching examples: [] . . . . no examples . . . inducePos(f12,1,s(nat)) . . . uncovered examples: [f12(va,0,0)=s(0),f12(va,0,s(0))=s(s(0)),f12(va,0,s(s(0)))=s(s(s(0))),f12(vb,s(0),0)=s(s(0)),f12(vb,s(0),s(0))=s(s(s... . . . trying argument position: 2 . . . inducePos(f12,2,0) . . . . matching examples: [f12(va,0,0)=s(0),f12(va,0,s(0))=s(s(0)),f12(va,0,s(s(0)))=s(s(s(0)))] . . . . anti-unifier: f12(va,0,v37) = s(v37) . . . inducePos(f12,2,0) . . . inducePos(f12,2,s(nat)) . . . . matching examples: [f12(vb,s(0),0)=s(s(0)),f12(vb,s(0),s(0))=s(s(s(0))),f12(vb,s(0),s(s(0)))=s(s(s(s(0)))),f12(vc,s(s(0)),s(0))=s(s(s(s... . . . . new recursion scheme: f12(v43,s(v45),v44) = f46(v43,v44,f12(v43,v45,v44)) . . . . derive new equation: s(s(0)) = f12(vb,s(0),0) = f46(vb,0,s(0)) . . . . derive new equation: s(s(s(0))) = f12(vb,s(0),s(0)) = f46(vb,s(0),s(s(0))) . . . . derive new equation: s(s(s(s(0)))) = f12(vb,s(0),s(s(0))) = f46(vb,s(s(0)),s(s(s(0)))) . . . . derive new equation: s(s(s(s(0)))) = f12(vc,s(s(0)),s(0)) = f46(vc,s(0),s(s(s(0)))) . . . . induce(f46) . . . . . trying argument position: 1 . . . . . inducePos(f46,1,0) . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f46,1,0) . . . . . inducePos(f46,1,s(nat)) . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f46,1,s(nat)) . . . . . uncovered examples: [f46(vb,0,s(0))=s(s(0)),f46(vb,s(0),s(s(0)))=s(s(s(0))),f46(vb,s(s(0)),s(s(s(0))))=s(s(s(s(0)))),f46(vc,s(0),s(s(s(0... . . . . . trying argument position: 2 . . . . . inducePos(f46,2,0) . . . . . . matching examples: [f46(vb,0,s(0))=s(s(0))] . . . . . . anti-unifier: f46(vb,0,s(0)) = s(s(0)) . . . . . inducePos(f46,2,0) . . . . . inducePos(f46,2,s(nat)) . . . . . . matching examples: [f46(vb,s(0),s(s(0)))=s(s(s(0))),f46(vb,s(s(0)),s(s(s(0))))=s(s(s(s(0)))),f46(vc,s(0),s(s(s(0))))=s(s(s(s(0))))] . . . . . . anti-unifier: f46(v63,s(v64),s(s(v65))) = s(s(s(v65))) . . . . . inducePos(f46,2,s(nat)) . . . . . all examples covered . . . . induce(f46) . . . inducePos(f12,2,s(nat)) . . . all examples covered . . induce(f12) . inducePos(size,1,nd(tree,nat,tree)) . all examples covered induce(size) +++++ Examples output check: +++++ Examples output check done FUNCTION SIGNATURES: f46 signature [nat,nat,nat]-->nat f12 signature [nat,nat,nat]-->nat size signature [tree]-->nat FUNCTION EXAMPLES: size(nl)=0 size(nd(nl,va,nl))=s(0) 16 size(nd(nd(nl,va,nl),vb,nl))=s(s(0)) size(nd(nl,va,nd(nl,vb,nl)))=s(s(0)) size(nd(nd(nl,va,nl),vb,nd(nl,vc,nl)))=s(s(s(0))) size(nd(nl,va,nd(nd(nl,vb,nl),vc,nl)))=s(s(s(0))) size(nd(nl,va,nd(nl,vb,nd(nl,vc,nl))))=s(s(s(0))) size(nd(nd(nl,va,nl),vb,nd(nd(nl,vc,nl),vd,nl)))=s(s(s(s(0)))) size(nd(nd(nd(nl,va,nl),vb,nl),vc,nd(nl,vd,nl)))=s(s(s(s(0)))) FUNCTION DEFINITIONS: size(nl)=0 size(nd(v66,v67,v68))=f12(size(v66),size(v68)) f12(0,v69)=s(v69) f12(s(v70),v71)=f46(v71,f12(v70,v71)) f46(0,s(0))=s(s(0)) f46(s(v72),s(s(v73)))=s(s(s(v73))) ?- A.3 Reversing a list ?- SgI = [rev signature [list] --> list], | SD = [ list sortdef [] ! [nat|list], | nat sortdef 0 ! s(nat)], | ExI = [ rev([]) = [], | rev([va]) = [va], | rev([vb,va]) = [va,vb], | rev([vc,vb,va]) = [va,vb,vc]], | run(rev,SgI,SD,ExI). +++++ Examples input check: +++++ Example 1: +++++ Example 2: +++++ Example 3: +++++ Example 4: Variable sorts: [vc:nat,vb:nat,va:nat] +++++ Examples input check done induce(rev) . trying argument position: 1 . inducePos(rev,1,[]) . . matching examples: [rev([])=[]] . . anti-unifier: rev([]) = [] . inducePos(rev,1,[]) . inducePos(rev,1,[nat|list]) . . matching examples: [rev([va])=[va],rev([vb,va])=[va,vb],rev([vc,vb,va])=[va,vb,vc]] . . new recursion scheme: rev([v7|v8]) = f9(v7,rev(v8)) . . derive new equation: [va] = rev([va]) = f9(va,[]) . . derive new equation: [va,vb] = rev([vb,va]) = f9(vb,[va]) . . derive new equation: [va,vb,vc] = rev([vc,vb,va]) = f9(vc,[va,vb]) . . induce(f9) . . . trying argument position: 1 . . . inducePos(f9,1,0) . . . . matching examples: [] . . . . no examples . . . inducePos(f9,1,0) . . . inducePos(f9,1,s(nat)) . . . . matching examples: [] . . . . no examples . . . inducePos(f9,1,s(nat)) . . . uncovered examples: [f9(va,[])=[va],f9(vb,[va])=[va,vb],f9(vc,[va,vb])=[va,vb,vc]] . . . trying argument position: 2 . . . inducePos(f9,2,[]) . . . . matching examples: [f9(va,[])=[va]] . . . . anti-unifier: f9(va,[]) = [va] . . . inducePos(f9,2,[]) . . . inducePos(f9,2,[nat|list]) . . . . matching examples: [f9(vb,[va])=[va,vb],f9(vc,[va,vb])=[va,vb,vc]] . . . . new recursion scheme: f9(v22,[v23|v24]) = f25(v22,v23,f9(v22,v24)) . . . . derive new equation: [va,vb] = f9(vb,[va]) = f25(vb,va,[vb]) . . . . derive new equation: [va,vb,vc] = f9(vc,[va,vb]) = f25(vc,va,[vb,vc]) . . . . induce(f25) . . . . . trying argument position: 1 . . . . . inducePos(f25,1,0) . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f25,1,0) . . . . . inducePos(f25,1,s(nat)) . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f25,1,s(nat)) . . . . . uncovered examples: [f25(vb,va,[vb])=[va,vb],f25(vc,va,[vb,vc])=[va,vb,vc]] . . . . . trying argument position: 2 . . . . . inducePos(f25,2,0) . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f25,2,0) . . . . . inducePos(f25,2,s(nat)) . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f25,2,s(nat)) . . . . . uncovered examples: [f25(vb,va,[vb])=[va,vb],f25(vc,va,[vb,vc])=[va,vb,vc]] . . . . . trying argument position: 3 . . . . . inducePos(f25,3,[]) 17 . . . . . . matching examples: [] . . . . . . no examples . . . . . inducePos(f25,3,[]) . . . . . inducePos(f25,3,[nat|list]) . . . . . . matching examples: [f25(vb,va,[vb])=[va,vb],f25(vc,va,[vb,vc])=[va,vb,vc]] . . . . . . anti-unifier: f25(v37,va,[vb|v38]) = [va,vb|v38] . . . . . inducePos(f25,3,[nat|list]) . . . . . all examples covered . . . . induce(f25) . . . inducePos(f9,2,[nat|list]) . . . all examples covered . . induce(f9) . inducePos(rev,1,[nat|list]) . all examples covered induce(rev) +++++ Examples output check: +++++ Examples output check done FUNCTION SIGNATURES: f25 signature [nat,nat,list]-->list f9 signature [nat,list]-->list rev signature [list]-->list FUNCTION EXAMPLES: rev([])=[] rev([va])=[va] rev([vb,va])=[va,vb] rev([vc,vb,va])=[va,vb,vc] FUNCTION DEFINITIONS: rev([])=[] rev([v39|v40])=f9(v39,rev(v40)) f9(v41,[])=[v41] f9(v42,[v43|v44])=f25(v43,f9(v42,v44)) f25(v41,[v45|v46])=[v41,v45|v46] ?- 18 References [Bur05] Jochen Burghardt. 165(1):1–35, 2005. E-generalization using grammars. Artificial Intelligence Journal, [Bur17] Jochen Burghardt. An improved algorithm for E-generalization. Technical report, Berlin, Sep 2017. [Com86] Hubert Comon. Sufficient completeness, term rewriting systems and “anti-unification”. In Proc. 8th International Conference on Automated Deduction, volume 230 of LNCS, pages 128–140. Springer, 1986. [DJ90] N. Dershowitz and J.-P. Jouannaud. Rewrite Systems, volume B of Handbook of Theoretical Computer Science, pages 243–320. Elsevier, 1990. [Gut77] John V. Guttag. Abstract data types and the development of data structures. Communications of the ACM, 20(6):396–404, Jun 1977. [Hei94a] Birgit Heinz. Anti-unification and its application to lemma discovery, 1994. Workshop Talk given in CADE-12. [Hei94b] Birgit Heinz. Lemma discovery by anti-unification of regular sorts. Technical Report 94–21, TU Berlin, 1994. [Hei95] Birgit Heinz. Anti-Unifikation modulo Gleichungstheorie und deren Anwendung zur Lemmagenerierung. PhD thesis, TU Berlin, Dec 1995. [Kit10] Emanuel Kitzelmann. A Combined Analytical and Search-Based Approach to the Inductive Synthesis of Functional Programs. PhD thesis, Univ. Bamberg, May 2010. [LV08] Ming Li and Paul Vitányi. An Introduction to Kolmogorov Complexity and Its Applications. texts in computer science. Springer, New York, 3rd edition, 2008. [Plo70] Gordon D. Plotkin. A note on inductive generalization. Machine Intelligence, 5:153–163, 1970. [Plo71] Gordon D. Plotkin. A further note on inductive generalization. Machine Intelligence, 6:101– 124, 1971. [Rey70] John C. Reynolds. Transformational systems and the algebraic structure of atomic formulas. Machine Intelligence, 5:135–151, 1970. 19
2cs.AI
A Generalized Circuit for the Hamiltonian Dynamics Through the Truncated Series Ammar Daskin Department of Computer Engineering, Istanbul Medeniyet University, Uskudar, Istanbul, Turkey Sabre Kais arXiv:1801.09720v1 [quant-ph] 29 Jan 2018 Department of Chemistry, Department of Physics and Birck Nanotechnology Center, Purdue University, West Lafayette, IN, USA In this paper, we present a fixed-quantum circuit design for the simulation of the Hamiltonian dynamics, H(t), through the truncated Taylor series method described by Berry et al. [1]. In particular, Hamiltonians which are not given as sums of unitary matrices but given as general matrices are considered and a simple divide and conquer method is presented for mapping the Hamiltonians into the circuit. The circuit is general and can be used to simulate any given matrix in the phase estimation algorithm by only changing the angle values of the quantum gates implementing the time variable t in the series. We analyze the circuit complexity and show that it requires O(N 2 ) number of CN OT gates for dense matrices, where N is the matrix size. We finally discuss how it can be used in adaptive processes and eigenvalue related problems along with a slight modified version of the iterative phase estimation algorithm. I. INTRODUCTION Since the standard quantum computing requires any operation to be a unitary transformation in the Hilbert space, quantum algorithms in general deal with the problems represented by unitary matrices. One of these algorithms is the phase estimation[2] used in the eigenvalue related problems. Given an estimate of the eigenvector, the algorithm finds the associated eigenvalue of a unitary matrix. In particular, it estimates the value of φ in the eigenvalue eiφ when given an estimate of the associated eigenvector |ϕi. If U is the time evolution operator of a Hamiltonian, i.e. U = eiHt , the algorithm also yields the eigenvalue of the Hamiltonian H. Therefore, for the simulation of a quantum system on quantum computers; it is necessary to have its time evolution operator as a quantum circuit. A very common approach to estimate eiHt follows the generalized Trotter formula [3, 4] in which a mixed product of the exponential of smaller terms forming the Hamiltonian operator is taken as an approximation. This product accurately yields the evolution if all terms forming the Hamiltonian commute. Otherwise, an error bounded by the order of the approximation is involved [5]. Berry et al.[1] have proposed using the Taylor expansion of eiHt √ directly on quantum circuits by adding an ancillary register to the system. In Ref.[6], we have showed that when I − H2 is available and H is Hermitian, one can use the following unitary matrix in the phase estimation algorithm: ! √ 2 H − I − H √ . (1) I − H2 H This notion of using an extended system to simulate a small one is generalized as quantum signal processing [7], where a unitary matrix similar to the one in Eq.(1) used without the phase estimation algorithm. Recently, the overhead of the truncated Taylor series method is reduced by showing how  to efficiently get the square root of matrix in a different basis [8]. In Ref.[9], successive applications of I + iH/µ with µ ≥ 10||H|| are used to obtain the eigenvalue of H from the sinus value of the phase in the phase estimation algorithm. All of these approaches are assumed that the Hamiltonian is given as a sum of simple unitary matrices. In Ref.[10], the truncated Taylor series method is used for quantum simulations after decomposing the configuration interaction matrix into a sum of sparse matrices. In this paper, we present a general design  for the matrices not given as sums of unitary matrices by  circuit √ employing the Taylor expansion of tH + i I − t2 H2 truncated at the third term. The circuit can be used in the phase estimation algorithm by only changing the part implementing the coefficients related to time step t. The eigenvalue of H can be obtained from the cosine value of the phase. We show that the whole circuit requires O(N 2 ) number of CNOT gates for the dense matrices when it is used in the phase estimation algorithm. This complexity can be further reduced in the case of sparse matrices since the number of gates used in the circuit solely depends on the matrix elements. The remaining part of this paper is organized as follows: In the following subsection, we describe the truncated Taylor series method. Then in the next section, the Taylor expansion used in this paper is described. Then, we II THE SERIES USED FOR THE CIRCUIT explain the divide and conquer method used to obtain the circuit. After analyzing the complexity of the circuit in terms of the number of CNOTs, we discuss how the method can be used with special matrices and in the adaptive processes along with the modified iterative phase estimation algorithm. A. Truncated Taylor Series Method For a matrix U , one can build the following by using an ancilla quantum register:   Ū • U= , • • (2) where “•” represents matrices which have no special meaning. For any arbitrary |ψi on the system register, the above matrix generates the following output state in the standard basis: U |0i |ψi = √ 1 1 |Φi + √ |0i Ū |ψi . 1 − P0 P0 (3) Here, |Φi represents the part of the output in which the first register is not |0i. And obviously when the first register in |0i state, the output yields Ū |ψi with probability ||Ū |ψi ||2 /P0 . Therefore, U can be used to emulate the action of U on any arbitrary state |ψi. This idea is used in various contexts: e.g., in Ref.[11], a programmable circuit design is PL presented for unitary matrices. Given a Hamiltonian H = k αk Hk with Hk representing a unitary matrix; through the Taylor expansion of eiHt truncated at the Kth term, Berry et al.[1] have showed: U (t) = eiHt ≈ Ū (t) = K j X (iHt) j=0 j! = M X βj V j , (4) j=0 where M = LK . And while Vj s are some products of Hk s, βj s are products of αk s. Furthermore, they have showed that the above expansion can be implemented as a circuit by using the following:    Ū • ∗ U = B ⊗ I V (B ⊗ I) = . (5) • • PM −1 p where V = blkdiag (V 1, V 2, . . . , VM ) and B = βj |ji. V can be implemented as a circuit by using an j additional logM qubits and controlling each Vj with the |jith state of these qubits. Here, when Ū is sufficiently close to a unitary matrix, one P can also maximize the probability P0 by using the oblivious amplitude amplification algorithm. In that case, when j βj = 2, the amplification requires only one iteration. II. THE SERIES USED FOR THE CIRCUIT If the Taylor expansion is truncated at the third term; eiHt ≈ I + itH − t2 H 2 . 2 (6) When ||tH|| ≤ 1; instead of using Taylor expansion for eitH , the following can be used to deal with the spectra of the Hamiltonian (similar ones used in prior arts Ref.[10, 12–14]): p U (t) = tH + i I − t2 H2 . (7) Here, U (t) describes a unitary matrix [15] with the eigenvalues whose real parts are equal to those of the Hamiltonian. In addition, U (t) and tH have the same eigenvectors. Eq.(6) also is an approximation to iU (t). Because, if we use the Taylor expansion of the square root; ! t2 H 2 U (t) ≈ Ū (t) = tH + i I − . (8) 2 The most significant part of this approximation is that it does not introduce any error on the real part of the eigenvalues of U (t) given in Eq.(7). Hence, using the phase estimation algorithm, one can obtain the eigenvalue from the cosine value of the phase. In the following section, a circuit design is presented for Ū (t) given in Eq.(8). 2 A Divide and Conquer Method III III. CIRCUIT IMPLEMENTATION OF U (t) CIRCUIT IMPLEMENTATION OF U (t) When H is given as a sum of unitary matrices or matrices which can be easily mapped to quantum gates, then one can design the circuit for U (t) by following Eq.(4) where the summation is converted into a product formula or the standard Trotter-Suzuki decomposition. This is the case for molecular Hamiltonians given in the second quantization[16, 17]. However, if H is not given as a sum of simple1 unitaries, then the following divide an conquer method can be used to rewrite the Hamiltonian as a sum of unitary matrices: i) The matrix is first recursively divided into submatrices. ii) Then, each submatrix is mapped to a quantum gate. iii) Using a coefficient matrix BH , the Hamiltonian is generated as a part of the following matrix:   H • UH = . • • iv) Finally, UH is employed to obtain Ū (t) in Eq.(8) as a part of the following matrix:   Ū (t) • . U (t) = • • (9) (10) The details are given in the following subsections: A. Divide and Conquer Method First, a given Hamiltonian H ∈ RN with (N = 2n ) is divided into four blocks:       A0 A1 A0 A1 H= = + . A2 A3 A3 A2 (11) Using the vectors in the standard basis, this can be rewritten as:  H = |0i h0| A0 + |1i h1| A3 + |0i h1| A1 + |1i h0| A2     = |0i h0| A0 + |1i h1| A3 + X ⊗ IN/2 |0i h0| A1 + |1i h1| A2 , where IN/2 describes an N/2 dimensional identity matrix and   0 1 X= . 1 0 (12) (13) This division into blocks is recursively continued until each block dimension becomes 2 or circuit representations of the blocks become known. For instance, after the second recursion step, we have the following (please see Appendix A 1 and A 2 for the mathematical steps and the third recursion):  H = |00i h00| A00 + |01i h01| A03 + |10i h10| A30 + |11i h11| A33    + I ⊗ X ⊗ IN/4 |00i h00| A01 + |01i h01| A02 + |10i h10| A31 + |11i h11| A32    + X ⊗ I ⊗ IN/4 |00i h00| A10 + |01i h01| A13 + |10i h10| A20 + |11i h11| A23    + X ⊗ X ⊗ IN/4 |00i h00| A11 + |01i h01| A12 + |10i h10| A21 + |11i h11| A22 . 1 simple in the sense that the unitary matrix can be implemented by using a few quantum gates. 3 (14) B Generating Circuit for UH in Eq.(9) III CIRCUIT IMPLEMENTATION OF U (t) In matrix form,  A  00 A02 H=  A20 A22   A01 A10 A03  A12 A21 A30 A23 A32  A11  A13  . A31  A33 (15) Note that following the subscripts of A[... ] s from left to right one can easily find the matrix elements for any A[... ] . After any kth recursive step, the Hamiltonian can be written more concisely in the following form: H= k 2X −1   Vj Pj (X, I) ⊗ IN/2k . (16) j=0 Here, Pj (X, I) is a permutation matrix constructed by using the tensor product of X and identity matrices, i.e.: Pj (X, I) = k−1 O X ji , with j = (j0 . . . jk−1 )2 . (17) i=0 Each Vj describes a multi controlled network: in matrix form, a block diagonal matrix where some A[... ] s are tiled on its diagonal. To find which A[... ] s in step k are involved in any of Vj s, we start from V0 and place all possible words wi with length k from the alphabet {0, 3} as the subscript of Awi , i.e.: V0 = k 2X −1 |ii hi| Awi , (18) i=0 Here, |ii is the ith vector in the standard basis and wi represents the ith word in the set. The subscripts of A[... ] s in each Vj can be determined from V0 and Pj (X, I) as follows: • if there is an X on the ith qubit, then we make the following change in the subscripts of A[... ] s in V0 : – (3 → 2), if the ith letter of the subscript is 3, we make it 2. – (0 → 1), if the ith letter is 0, we make it 1. – no change necessary for the other letters. You can consider this as if we apply X gate to the subscript and switch 3s into 2s and 0s into 1s. B. Generating Circuit for UH in Eq.(9) Each Vj describes a multi controlled network: There are N/2 number of A[... ] s for each Vj . Once the involved A[... ] s are determined, they can be mapped to a circuit by using different control bit schemes for each A[... ] . Note that each A[... ] can be written as a sum of two unitary matrices. For instance,     q q 1 1 2 2 A[... ] + i I − A[... ] + A[... ] − i I − A[... ] (19) A[... ] = 2 2 This will double the number of Vj s. Since the matrix elements of A[... ] are real, it can be also described as a sum of three rotations about the y-axis, which triples the number of terms:   a a A[... ] = 00 01 a10 a11       (20) 1 − (a00 + a11 ) (a10 − a01 ) 1 (a00 + a11 ) − (a10 − a01 ) a a01 = 00 + + . a01 −a00 2 (a10 − a01 ) (a00 + a11 ) 2 (a10 − a01 ) (a00 + a11 ) Then, the circuits for Vj s are combined in select (V ) by using an ancilla register: select (V ) applies Vj when the ancilla is in |ji state. The matrix form of this operation is as follows:   V0   V1   V = (21) . . .   . VN/2 4 B Generating Circuit for UH in Eq.(9) IV TOTAL GATE AND QUBIT COUNT FIG. 1: The circuit for UH for any 8 × 8 Hamiltonian matrix. Here, since there is no coefficient on any A[... ] , BH can be considered as a tensor product of two Hadamard gates. Also note that permutation matrices are simplified into two CNOTs in the end. FIG. 2: The circuit for U (t). In total (2n + 1) qubits are involved  B is a 4 × 4 unitary matrix with √ in the√circuit. t, 1, t/ 2, 0 . the leading row and column Furthermore, using the same control scheme, select (P ) is constructed for the permutation matrices:   P0   P1   (22) V = . . ..   PN/2  ∗ The circuit for UH implementing the Hamiltonian then becomes  the product BH ⊗ I V P (BH ⊗ I), where BH forms a state with the square root of the coefficients of Vj Pj (X, I) s. An example UH for a general 8 × 8 Hamiltonian is presented in Fig.1, where Pj s are simplified into two CNOT gates. C. Combining All for Ū (t) To implement U (t), we will employ the following selection matrix:   UH   iI . V=   UH2 • Then, U (t) is generated by using two more qubits and the matrix B which implements the coefficients: V is generated with the help of a permutation matrix also used in Ref.[9]:   IN . X ⊗ I3N Π= IN (23) √ √  t, 1, t/ 2, 0 . (24) The resulting circuit is shown in Fig.2. IV. TOTAL GATE AND QUBIT COUNT The total complexity of the whole circuit given in Fig.2 is governed by the number of gates, in particular CNOTs used to implement UH . The complexity of UH is determined by the number of Vj s. We can determine the number of 5 B Adaptive Processes V DISCUSSION AND FUTURE WORK qubits and CNOTs by counting the number of A[... ]. In the final step of the recursive division, the matrix elements are put into group of four elements. This yields N 2 /4  number of A[... ] s. If each A[... ] is written as a sum of two unitary 2 matrices,  then the number of gates becomes N /2 . If it is written as a sum of three rotation gates, then it becomes 2 3N /4 . This requires either (2n − 1) or (2n) number of control qubits and 1 target qubit. The decomposition of a multi controlled network into CNOTs and singlerotation gates can be done by following Ref.[18]. In that case, the circuit for UH at this point would require N 2 /2 or N 2 number of CNOTs. The generation of the circuit in Fig.2 necessitates two more qubits. This makes the number of qubits (2n + 1) or (2n + 2) in total depending on the number of gates used for each A[... ] . The circuit has two networks UH controlled by  2 2 (2n) or (2n + 1) qubits. Following Ref.[18] again, the decomposition of each network requires N or 2N CNOTs. Since the  complexity of B and Π are negligible in comparison to controlled UH , the total complexity is bounded by O N2 . V. DISCUSSION AND FUTURE WORK A. Special Matrices The described divide and conquer method groups the neighboring matrix elements into gates. Any sparsity in the considered matrix may potentially reduce the number of A[... ] s. However, when the sparsity of the matrix is structured as in tridiagonal, anti-tridiagonal, and band matrices; the number of terms and so the numbers of qubits and CNOTs used in the circuit are directly affected. For instance, the 16 × 16 Hamiltonian for the hydrogen molecule given in Ref.[17] (It can be easily written as a sum of unitaries through Jordan-Wigner transformation) only have 4 non-diagonal elements, all of which are on the anti-diagonal part of the Hamiltonian. If we consider the diagonal part and the anti-digonal part as a sum of two unitaries, then the Hamiltonian can be written as sum of 4 number of Vj Pj terms. Since each Vj involves multicontrolled 8 quantum gates, in total there are 32 quantum gates. Then, UH requires 5 control- and 1 target-qubits. Since an additional control qubit is necessary for U (t), there are two multi-controlled network, viz. two controlled UH s, with 6 control qubits. The decomposition of these networks will constitute 128 CNOT gates in total. Here, note that using U (t) in the phase estimation introduces an additional control qubit. Then the required number of CNOTs for each iteration of the phase estimation algorithm is doubled to ≈ 256. Also note that in every iteration, only the gates implementing B need to be updated and the rest of the circuit remains the same. B. Adaptive Processes Finding the matrix elements of a Hamiltonian representing the dynamics of a quantum system is a nontrivial task requiring tedious analytical and numerical computations. The circuit can be used to experimentally identify the Hamiltonian dynamic of an unknown large system or estimate parameters of a quantum channels, where a known state is sent through an unknown state and the measurement is used to estimate the parameters associated with the channel [19]. A similar approach is used also in Ref.[20] to compress molecular Hamiltonians. The matrix divided as in Eq.(14) can be also used to represent a layer of a neural network in matrix form. In the learning or adaptive processes based on gradient-descent, when the eigenvalues of an autocorrelation matrix are disparate: i.e. the condition number of the matrix is large, the leaning rate or the performance of an adaptive algorithm are hindered in applications [21]. Therefore, in these processes, using the phase estimation[2] as an ingredient for different algorithms may provide better performances. The phase estimation algorithm requires two registers to store the eigenvalue and the eigenvector respectively. The size of the first register is determined by the desired accuracy: e.g. for a 32-bit precision, it has 32 qubits. When the system size is large, the iterative phase estimation using only a qubit in the first register is more preferred in the experiments and classical simulations. However, the iterative version starts the estimation of the bits from the least significant bit (LSB) toward the most significant bit (MSB). This impedes the employment of the algorithm as a subroutine in various multivariate statistical algorithms in which the eigenvalues above or below some threshold are filtered out. 1. Modified Iterative Phase Estimation Fig.3 describes a slight modified version of the iterative phase estimation algorithm, where the bit values are estimated starting from MSB: The algorithm works as follows: 6 B Adaptive Processes A HAMILTONIAN IN THE RECURSION FIG. 3: The iterative phase estimation for adaptive processes: H represents the Hadamard gate and k ≥ 1. • Assume the circuit U and its eigenvector |ϕi are given. For simplicity also assume t = 1. • In the first iteration k is set to 1: i.e., we have U (2π). After the controlled-U , the following state is obtained (Normalization constants are omitted for simplicity.): |ψ1 i = |0i |ϕi + eiπ(φ1 .φ2 ... )2 |1i |ϕi . (25) Here, (φ1 .φ2 . . . )2 represents the binary form of the phase multiplied by 2: 2φ. • To make the real part of the value eiπ(φ1 .φ2 ... )2 negative when φ1 = 1 and positive when φ1 = 0, we apply a  rotation Rz −π/2 :  Rz π − 2  = 1 0 0 e−iπ/2 ! . (26) Note that in general when rotation gate about the z-axis is defined, the angle is divided by 2, which is neglected here. After this gate, we have: |ψ2 i = |0i |ϕi + eiπ(φ1 .φ2 ... )2 −iπ/2 |1i |ϕi . (27) • The value of τ for the gate Rz (τ ) is computed from the previous iterations and it is zero for the first iteration. • After applying the second Hadamard gate, the final state becomes the following:  |ψ3 i = 1 + cos (α) + isin (α) |0i |ϕi  + 1 − cos (α) − isin (α) |1i) |ϕi , (28) where α = π (φ1 .φ2 . . . )2 − π/2. The probability difference between |0i and |1i is determined by the value of cos(α): h i – If φ1 = 0, then α ∈ − π2 , π2 and cos(α) ≥ 0. Hence, the probability of |0i is higher than |1i. h i – If φ1 = 1, then α ∈ π2 , 3π and cos(α) ≤ 0. And so, the probability of |1i is higher than |0i. 2 • In the second iteration; k = 2 and if φ1 is 1, then τ is set to 1 × π to put the phase in the form (φ2 .φ3 . . . )2 . The probability difference between |0i and |1i in the output may be as high as 1 or in rare cases is equal to 0. For random 5000 matrices, the distribution of the probability differences is drawn Fig.4: The probability differences are obtained by iterating the phase estimation algorithm 20 times for each matrix. As shown in the figure, the probability to see the difference less than 0.1 is around 0.06. Therefore, this iterative version can be used in adaptive processes to filter out some of the eigenvalues or prepare the ground state of the Hamiltonians. MATLAB source codes for the circuits in the figures and for obtaining A[... ] , Vj , and Pj matrices are provided in https://github.com/adaskin/circuitforTaylorseries. Appendix A: Hamiltonian in the Recursion 1. Second step (k = 2) If we plug Eq.(12) in places of A[... ] s: 7 1 Second step (k = 2) A HAMILTONIAN IN THE RECURSION FIG. 4: The histogram of the probability differences in the outputs of the iterative phase estimation for 5000 random matrices. The algorithm is iterated 20 times for each matrix. Here, the probability to see the probability difference less than 0.1 is 0.0644 and greater than 0.9 is 0.2862.      H = |0i h0| |0i h0| A00 + |1i h1| A03 + X ⊗ IN/8 |0i h0| A01 + |1i h1| A02      + |1i h1| |0i h0| A30 + |1i h1| A33 + X ⊗ IN/8 |0i h0| A31 + |1i h1| A32        + X ⊗ IN/4 |0i h0| |0i h0| A10 + |1i h1| A13 + X ⊗ IN/8 |0i h0| A11 + |1i h1| A12        + X ⊗ IN/4 |1i h1| |0i h0| A20 + |1i h1| A23 + X ⊗ IN/8 |0i h0| A21 + |1i h1| A22 . (A1) This gives the following:    H = |00i h00| A00 + |01i h01| A03 + I ⊗ X ⊗ IN/8 |00i h00| A01 + |01i h01| A02    + |10i h10| A30 + |11i h11| A33 + I ⊗ X ⊗ IN/8 |10i h10| A31 + |11i h11| A32       + X ⊗ I ⊗ IN/8 |00i h00| A10 + |01i h01| A13 + X ⊗ X ⊗ IN/8 |00i h00| A11 + |01i h01| A12       + X ⊗ I ⊗ IN/8 |10i h10| A20 + |11i h11| A23 + X ⊗ X ⊗ IN/8 |10i h10| A21 + |11i h11| A22 , which can be rewritten as in Eq.(14). 8 (A2) 1 Second step (k = 2) A 2. HAMILTONIAN IN THE RECURSION Third step (k = 3) Then, in the third recursion, we get the following: H =|0|A000 + |1|A003 + |2|A030 + |3|A033 + |4|A300 + |5|A303 + |6|A330 + |7|A333   + I ⊗ I ⊗ X ⊗ IN/8 |0|A001 + |1|A002 + |2|A031 + |3|A032 + |4|A301 + |5|A302 + |6|A331 + |7|A332   + I ⊗ X ⊗ I ⊗ IN/8  |0|A010 + |1|A013 + |2|A020 + |3|A023 + |4|A310 + |5|A313 + |6|A320 + |7|A323   + I ⊗ X ⊗ X ⊗ IN/8  |0|A011 + |1|A012 + |2|A021 + |3|A022 + |4|A311 + |5|A312 + |6|A321 + |7|A322   + X ⊗ I ⊗ I ⊗ IN/8  |0|A100 + |1|A103 + |2|A030 + |3|A033 + |4|A200 + |5|A203 + |6|A230 + |7|A233   + X ⊗ I ⊗ X ⊗ IN/8  |0|A101 + |1|A102 + |2|A031 + |3|A032 + |4|A201 + |5|A202 + |6|A231 + |7|A232   + X ⊗ X ⊗ I ⊗ IN/8  |0|A110 + |1|A113 + |2|A123 + |3|A123 + |4|A210 + |5|A213 + |6|A220 + |7|A223   + X ⊗ X ⊗ X ⊗ IN/8  (A3)  |0|A111 + |1|A112 + |2|A121 + |3|A122 + |4|A211 + |5|A212 + |6|A221 + |7|A222 , where |i| = |ii hi|. This Hamiltonian in matrix form corresponds to the following matrix:    A000  A002   A020   A022  H =    A200  A202   A220  A222  A001 A003  A021 A023  A201 A203  A221 A223  A010 A012 A030 A032  A210 A212 A230 A232  A011  A013   A031  A033  A211  A213   A231  A233  A  100 A102   A120 A122  A  300 A302   A320 A322  A101 A103  A121 A123  A301 A303  A321 A323  A110 A  112 A130 A132  A310 A  312 A330 A332   A111  A113   A131    A133   .  A311  A323   A331   A333 (A4) [1] Dominic W Berry, Andrew M Childs, Richard Cleve, Robin Kothari, and Rolando D Somma, “Simulating hamiltonian dynamics with a truncated taylor series,” Physical review letters 114, 090502 (2015). [2] Alexei Kitaev, “Quantum measurements and the Abelian Stabilizer Problem,” Electronic Colloquium on Computational Complexity (ECCC) 3 (1996). [3] Hale F Trotter, “On the product of semi-groups of operators,” Proceedings of the American Mathematical Society 10, 545–551 (1959). [4] Masuo Suzuki, “Generalized trotter’s formula and systematic approximants of exponential operators and inner derivations with applications to many-body problems,” Communications in Mathematical Physics 51, 183–190 (1976). [5] David Poulin, Matthew B. Hastings, Dave Wecker, Nathan Wiebe, Andrew C. Doberty, and Matthias Troyer, “The trotter step size required for accurate quantum simulation of quantum chemistry,” Quantum Information and Computation 15, 361–384 (2015). [6] Ammar Daskin and Sabre Kais, “An ancilla-based quantum simulation framework for non-unitary matrices,” Quantum Information Processing 16, 33 (2017). 9 2 Third step (k = 3) A HAMILTONIAN IN THE RECURSION [7] Guang Hao Low and Isaac L Chuang, “Optimal hamiltonian simulation by quantum signal processing,” Physical review letters 118, 010501 (2017). [8] David Poulin, Alexei Kitaev, Damian S Steiger, Matthew B Hastings, and Matthias Troyer, “Fast quantum algorithm for spectral properties,” arXiv preprint arXiv:1711.11025 (2017). [9] Ammar Daskin and Sabre Kais, “Direct application of the phase estimation algorithm to find the eigenvalues of the hamiltonians,” arXiv preprint arXiv:1703.03597 (2017). [10] Ryan Babbush, Dominic W Berry, Yuval R Sanders, Ian D Kivlichan, Artur Scherer, Annie Y Wei, Peter J Love, and Aln Aspuru-Guzik, “Exponentially more precise quantum simulation of fermions in the configuration interaction representation,” Quantum Science and Technology 3, 015006 (2018). [11] Anmer Daskin, Ananth Grama, Giorgos Kollias, and Sabre Kais, “Universal programmable quantum circuit schemes to emulate an operator,” The Journal of chemical physics 137, 234112 (2012). [12] Ryan Babbush, Dominic Berry, Mária Kieferová, Guang Hao Low, Yuval Sanders, Artur Scherer, and Nathan Wiebe, “Improved techniques for preparing eigenstates of fermionic hamiltonians,” arXiv preprint arXiv:1711.10460 (2017). [13] Guang Hao Low and Isaac L Chuang, “Hamiltonian simulation by uniform spectral amplification,” arXiv preprint arXiv:1707.05391 (2017). [14] Guang Hao Low and Isaac L Chuang, “Hamiltonian simulation by qubitization,” arXiv preprint arXiv:1610.06546 (2016). [15] Pei Wu, “Additive combinations of special operators,” Banach Center Publications 30, 337–361 (1994). [16] Benjamin P Lanyon, James D Whitfield, Geoff G Gillett, Michael E Goggin, Marcelo P Almeida, Ivan Kassal, Jacob D Biamonte, Masoud Mohseni, Ben J Powell, Marco Barbieri, et al., “Towards quantum chemistry on a quantum computer,” Nature Chemistry 2, 106–111 (2010). [17] Jacob T. Seeley, Martin J. Richard, and Peter J. Love, “The bravyi-kitaev transformation for quantum computation of electronic structure,” The Journal of Chemical Physics 137, 224109 (2012). [18] Mikko Möttönen, Juha J. Vartiainen, Ville Bergholm, and Martti M. Salomaa, “Quantum circuits for general multiqubit gates,” Physical Review Letters 93, 130502 (2004). [19] BM Escher, RL de Matos Filho, and L Davidovich, “General framework for estimating the ultimate precision limit in noisy quantum-enhanced metrology,” Nature Physics 7, 406–411 (2011). [20] Jonathan Romero, Jonathan P Olson, and Alan Aspuru-Guzik, “Quantum autoencoders for efficient compression of quantum data,” Quantum Science and Technology 2, 045001 (2017). [21] Bernard Widrow, John M McCool, Michael G Larimore, and C Richard Johnson, “Stationary and nonstationary learning characteristics of the lms adaptive filter,” Proceedings of the IEEE 64, 1151–1162 (1976). 10
8cs.DS
Gaussian and exponential lateral connectivity on distributed spiking neural network simulation Elena Pastorelli arXiv:1803.08833v1 [cs.DC] 23 Mar 2018 INFN Sezione di Roma and PhD Program in Behavioural Neuroscience, “Sapienza” University of Rome Rome, Italy [email protected] Pier Stanislao Paolucci, Francesco Simula, Andrea Biagioni, Fabrizio Capuani, Paolo Cretaro, Giulia De Bonis, Francesca Lo Cicero, Alessandro Lonardo, Michele Martinelli, Luca Pontisso, Piero Vicini INFN Sezione di Roma Rome, Italy {name.surname}@roma1.infn.it Roberto Ammendola INFN Sezione di Tor Vergata and Electronic Engineering Dept., University of Roma “Tor Vergata” Rome, Italy [email protected] Abstract—We measured the impact of long-range exponentially decaying intra-areal lateral connectivity on the scaling and memory occupation of a distributed spiking neural network simulator compared to that of short-range Gaussian decays. While previous studies adopted short-range connectivity, recent experimental neurosciences studies are pointing out the role of longer-range intra-areal connectivity with implications on neural simulation platforms. Two-dimensional grids of cortical columns composed by up to 11 M point-like spiking neurons with spike frequency adaption were connected by up to 30 G synapses using short- and long-range connectivity models. The MPI processes composing the distributed simulator were run on up to 1024 hardware cores, hosted on a 64 nodes server platform. The hardware platform was a cluster of IBM NX360 M5 16-core compute nodes, each one containing two Intel Xeon Haswell 8-core E5-2630 v3 processors, with a clock of 2.40 G Hz, interconnected through an InfiniBand network, equipped with 4× QDR switches. Index Terms—cortical simulation; distributed computing; spiking neural network; lateral synaptic connectivity; hardware/software co-design; I. I NTRODUCTION We present the impact of the range of intra-areal lateral connectivity on the scaling of distributed point-like spiking neural network simulations when run on up to 1024 software processes (and hardware cores) for cortical models including tens of billions of synapses. A simulation including a few tens of billions of synapses is what is required to simulate the activity of one cm2 of cortex at biological resolution (e.g. 54K neuron/mm2 and about 5K synapses per neuron in the rat neocortex area [1]). The capability to scale a problem up to such a size allows simulating an entire cortical area. Our study focuses on the computational cost of the implementation of connectivities, pointed out in recent studies reporting about long range intra-areal lateral connectivity in many different cerebral areas, from cat primary visual cortex [2], to rat neocortex [1], [3], just as examples. For instance, in rat neocortex, the impact of lateral connectivity on the pyramidal cells in layer 2/3 and layer 6a, results in ∼75% of incoming remote synapses to neurons of these layers. Longer-range intra-areal connectivity can be modeled by a distance-dependent exponential decay of the probability of synaptic connection between pairs of neurons: i.e. A·exp( −r λ ), where r stands for the distance between neurons, λ is the exponential decay constant and A is a normalization factor that fixes the total number of lateral connections. Decay constants in the range of several hundred microns are required to match experimental results. Previous studies considered intra-areal synaptic connections dominated by local connectivity: e.g. [4] estimated at least 55% the fraction of local synapses, reaching also a ratio of 75%. Such shorter-range lateral connectivity has often been modeled with a distance dependent Gaussian decay [5] 2 B · exp( −r 2σ 2 ), where r stands again for distance between neurons, σ 2 is the variance that determines the lateral range and B fixes the total number of projections. Here we present measures about the scaling of simulations of cortical area patches of different sizes represented by two-dimensional grids of “cortical modules”. Each cortical module is composed of 1240 single-compartment, point-like neurons (no dendritic tree is represented) each one receiving up to ∼2390 recurrent synapses (instantaneous membrane potential charging) plus those bringing external stimuli. The larger simulated cerebral cortex tile includes 11.4 M neurons and 29.6 G total synapses. Exponentially decaying lateral connectivity (longer-range) are compared to a Gaussian connectivity decay (shorter-range), analyzing the scaling and the memory usage of our Distributed and Plastic Spiking Neural Network simulation engine (DPSNN in the following). On DPSNN, the selection of the connectomic model has consequences due to: 1) the mapping strategy (neurons and incoming synapses are placed on MPI processes according to spatial contiguity) and, 2) synaptic messages exchanged between neurons simulated on different MPI processes entail communication tasks among those processes; the higher the number of lateral synaptic connections and the longer the interaction distance is, the more intensive the communication task among processes becomes. The impact of other biologically plausible, or experimentally demonstrated, connectivity patterns is worth of investigation, but is not covered by this work. One of the directions could be the study of the effect of connectivity patterns with local modular/clustered connection and global (inter-areal) non-homogenoeus lateral connectivity. Such a connectivity has been studied theoretically for network dynamical behaviors on a small local neural network [6]. Experimentally, complex connectivity has been seen mostly for across-area studies, however see also the emerging strong evidence of local motifs [7]. The article is structured as follows: Section II describes the main features of the simulation engine and its distributed implementation; network models are summarized in Section III, with a specific description of the different schemes adopted for the lateral intra-area connectivity; Section IV reports the impact of lateral connectivity on the scaling. A discussion section closes the paper. several competing platforms. Coded as a network of C++ processes, it is designed to be easily interfaced to both MPI and other (custom) Software/Hardware Communication Interfaces. In this work, the neural network is described as a twodimensional grid of cortical modules made up of singlecompartment, point-like neurons spatially interconnected by a set of incoming synapses. Cortical modules are composed of several populations of excitatory and inhibitory neurons. Cortical layers can be modeled by a subset of those populations. Each synapse is characterized by a specific synaptic weight and transmission delay, accounting for the axonal arborization. The two-dimensional neural network is mapped on a set of C++ processes interconnected with a message passing interface. Each C++ process simulates the activity of a cluster of neurons. The spikes generated during the neural activity of each neuron are delivered to the target synapses belonging to the same or to other processes. The “axonal spikes”, that carry the information about the spiking neuron identity and the original spike emission time, constitute the payload of the messages. Axonal spikes are sent only toward those C++ processes where a target synapse exists. The knowledge of the original emission time and of the transmission delay introduced by each synapse is necessary for synaptic Spike Timing Dependent Plasticity (STDP) management, supporting Long Term Potentiation/Depression (LTP/LTD) of the synapses. II. D ESCRIPTION OF THE S PIKING N EURAL N ETWORK Simulation undergoes two phases: 1. creation and initialization of the network of neurons, of the axonal arborization and of the synapses; 2. simulation of the neural and synaptic dynamics. A combined event-driven and time-driven approach has been adopted, inspired by [27]: the dynamic of neurons and synapses (STDP) is simulated when the event arises (event driven integration), while the message passing conveying the axonal spikes among processes is performed at regular time driven steps (in the present study set to 1 ms). Simulation (see Fig. 1) can be further decomposed into the following steps: 2.1) spike-producing neurons during the previous time-driven simulation step are identified and the consequent contribution to STDP is calculated; 2.2) spikes are sent through axonal arborizations to the cluster of neurons where target synapses exist; 2.3) inside each process, incoming axonal spikes are queued into lists, for later usage during the time-step corresponding to the synaptic delays; 2.4) synapses inject currents into target neurons and the consequent contribution to STDP is calculated; 2.5) neurons sort input currents coming from recurrent and external synapses; 2.6) neurons integrate their dynamic equation for each input current in the queue, using an event-driven solver. At a slower timescale, which in the current implementation is every second, STDP contributions are integrated in a Long Term Plasticity and applied to each single synapse. SIMULATOR The main focus of several neural network simulation projects is the search for a) biological correctness; b) flexibility in biological modeling; c) scalability using commodity technology — e.g. NEST [8], [9], NEURON [10], GENESIS [11]. A second research line focuses more explicitly on computational challenges when running on commodity systems, with varying degrees of association to specific platform ecosystems [12]–[15]. Another research pathway is the development of specialized hardware, with varying degrees of flexibility allowed — i.e. SpiNNaker [16], BrainScaleS [17]. Instead, the DPSNN simulation engine is meant to address two objectives: (i) quantitative assessment of requirements and benchmarking during the development of embedded [18] and HPC systems [19] — focusing either on network [20] or on power efficiency [21] — and (ii) the acceleration of the simulation of specific models in computational neuroscience — e.g. to study slow waves in large scale cortical fields [22], [23] in the framework of the HBP [24] project. The simulation engine follows a mixed time- and event-driven approach and implements synaptic spike-timing dependent plasticity ( [25], [26]). It has been designed from the ground up to be natively distributed and parallel, and should not pose obstacles against distribution and parallelization on A. Execution flow: a mixed time and event-driven approach from the source to the target neuron. Spikes, whose target neurons belong to the same process, are packed in the axonal spike message. The arborization of this message is deferred to the target process. Deferring as much as possible the arborization of the “axon” reduces the communication load and unnecessary wait barrier. To this purpose, preparatory actions are performed during network initialization (performed once at the beginning of the simulation), to reduce the number of active communication channels during the iterative simulation phase. D. Initial construction of the connectivity infrastructure Fig. 1. Execution flow of DPSNN simulator. Labels on the left identify event- or time-driven tasks. Orange blocks are used for the inter-processes communication tasks. B. Distributed generation of synaptic connections The described simulation engine exploits its parallelism also during the creation and initialization phase of the network, as detailed in a following section. In a given process, a set of local neurons i = 1, .., N projects their set of synapses j = 1, .., M , toward their target neurons K i,j , each synapse characterized by individual delays Di,j and plastic weights W i,j . Synaptic efficacies are randomly chosen from Gaussian distributions, while synaptic delays can be generated according to exponential or uniform distribution. The moments of the distributions depend on the source and target populations that each specific synapse interconnects. In addition to recurrent synapses, the system simulates also a number of external synapses: they represent afferent (thalamo-) cortical currents coming from outside the simulated network. C. Representation of spiking messages Spike messages are defined using an Address Event Representation (AER, [28]) including the spiking neuron identifier and the exact spiking time. During simulation, spikes travel During initialization, each process contributes to create awareness about the subset of processes that should be listened to during next simulation iterations. This knowledge is based on information extracted from the locally constructed matrix of outcoming and incoming synapses. At the end of this construction phase, each “target” process should know about the subset of “source” processes that need to communicate with it, and should have created its database of locally incoming axons and synapses. A simple implementation of the construction phase can be carried out using two steps. During the first step, each source process informs other processes about the existence of incoming axons and the number of incoming synapses. A single word, the synapse counter, is communicated among pairs of processes. Under MPI, this can be achieved by the MPI_Alltoall() routine. This is performed once, and with a single word payload. The second construction step transfers the identities of synapses to be created onto each target process. Under MPI, the payload — a list of synapses specific for each pair in the subset of processes to be connected — can be transferred using a call to the MPI_Alltoallv() library function. The number of messages depends on the lateral connectivity range and on the distribution of cortical modules among processes, while the cumulative load is always proportional to the total number of synapses in the system. The knowledge about the existence of a connection between a pair of processes can be reused to reduce the cost of spiking transmission during the iterations of simulation. E. Delivery of spiking messages during the simulation phase At each iteration, spikes are exchanged between pairs of processes connected by the synaptic matrix. The delivery of spiking messages can be split in two steps, with communications directed toward subsets of decreasing size. During the first step, single word messages (spike counters) are sent to the subset of potentially connected target processes. On each pair of source-target process subset, the individual spike counter informs about the actual payload — i.e. axonal spikes — to be delivered, or about the absence of spikes to be transmitted. The knowledge of the subset was created during the first step of initialization (see Section II-D). The second step uses the spiking counter to establish a communication channel only between pairs of processes that actually need to transfer an axonal payload during the current iteration. On MPI, both steps can be implemented using calls to the MPI_Alltoallv() library function. For two-dimensional grids of neural columns and for their mapping on processes used in this experiment, this implementation proved to be quite efficient, as reported in Section IV, further refined in Section V. III. N EURAL N ETWORK C ONFIGURATION A. Spiking Neuron Model and Synapses The single-compartment, point-like neurons used in this paper are based on the Leaky Integrate and Fire (LIF) neuron model with spike-frequency adaptation (SFA) due to calcium- and sodium-dependent after-hyperpolarization (AHP) currents [29]. Neuronal dynamics is described by the following equations: X V −E c dV = − gc + Ji δ(t − ti ) dt τm Cm (1) c dc =− dt τc (2) where V (t) represents the membrane potential and c(t) the fatigue variable used to model the SFA as an activity-dependent inhibitory current. τm is the membrane characteristic time, Cm the membrane capacitance, E the resting potential and τc the decay time for the fatigue variable c. gc , paired with the membrane capacitance, determines the timescales of the coupling of the fatigue (2) and membrane potential (1) equations. For inhibitory neurons, the SFA term is set to zero. Synaptic spikes, reaching the neuron at times ti , produce instantaneous membrane potential changes of amplitude Ji , the weights of activated synapses. When the membrane potential exceeds a threshold Vθ , a spike occurs. Thereafter, the membrane potential is reset to Vr for a refractory period τarp , whereas the variable c is increased by the constant amount αc . During the construction phase of the network, recurrent synapses are established between pre- and post-synaptic neurons (see Section III-B). Synaptic efficacies and delays are randomly chosen from probabilistic distributions (see Section II-B). In addition to the recurrent synapses, the system simulates also a number of external synapses: they represent afferent (thalamo-)cortical currents coming from outside the simulated network, collectively modeled as a Poisson process with a given average spike frequency. The recurrent synapses plus the external synapses yield the number of total synapses afferent to the neuron, referred to as “total equivalent” synapses in the following. For all the measurements in this work, synaptic plasticity has been disabled, to simplify the comparison between different configurations used in the scaling analysis, ensuring higher stability of the states of the networks. B. Cortical Columns and their connectivity Neurons are organized in cortical modules (mimicking columns), each one composed of 80% excitatory and 20% inhibitory neurons. Modules are assembled in two-dimensional square grids, representing a cortical area slab, with a grid step α ∼ 100 µm (inter-columnar spacing). The size of these grids has been varied as per Table I, to perform the scalability experiments here reported. Each cortical module includes 1240 neurons, while the number of synapses projected by each neuron depends on the implemented connectivity. The neural network connectivity is set by the user defining the probabilistic connection law between neural populations, spatially located in the two-dimensional grid of cortical modules. Connectivity can be varied according to the simulation needs, leading to configurations with different numbers of synapses per neuron. We adopted the following lateral connectivity rules to evaluate the impact of different inter-module connectivity laws: • Gaussian connectivity — shorter range and lower number of remote synapses: considering preeminent local connectivity with respect to lateral, the rule used to calculate remote connectivity has been set proportional to 2 A · exp( −r 2σ 2 ), with A = 0.05 and σ = 100 µm being the lateral spread of the connection probability. The remote connectivity function is similar to that adopted by the [5] model, although with different A and σ parameters. In this case only ∼20% of the synapses (specifically ∼250) are remotely projected and reach modules placed within a short distance, spanning a few steps in the two-dimensional grid of cortical modules. The majority of connections (∼80%) is local to the module. • Exponential decay connectivity — longer range and higher number of remote synapses: the connectivity rule for remote synapses calculation is proportional to A · exp( −r λ ), with A = 0.03 and λ = 290 µm (the exponential decay constant, in the range of experimental biological values, see e.g. [1]). This turns out into an increased number of remote connections (∼59%), i.e. ∼1400 lateral synapses per neuron. It is worth nothing that full biological realism would require to increase the total number of lateral connections above ∼4 K synapses/neuron. For both studied connectivities, a local connection probability of 80% (producing about 990 local synapses) has been adopted. For classical short-range configuration, the dominance of local synapses enables mean-field theory prediction of the dynamical regime of the modules, that perceive the influence of remote modules as small perturbations. In summary, the average number of projected synapses per neuron is ∼2390 for the longer-range exponential connectivity while for the Gaussian connectivity is ∼1240. In both systems, a cut-off has been set in the synapses generation, limiting the projection to the subset of modules with connection probability greater than 1/1000. This turns out into stencils of projected connections centered on the source module. A 7×7 stencil is generated in the first case (Gaussian) and a 21 × 21 in the second case (exponential decay). They are marked in green and orange in Fig. 2. For each connectivity scheme, measurements were taken on different problem sizes obtained varying the dimension of the grid of modules and, once fixed the problem size, distributing it over a span of MPI processes to evaluate the scaling behaviour. We selected three grid dimensions: 24 × 24, 48 × 48 and 96 × 96 (see Table I). For a columnar spacing of few hundreds of microns, they can be considered representative of interesting biological cortical slab dimensions. The number of processes over which each network size is distributed varied from a minimum, bounded by memory, and a maximum, bounded by communication (or HPC platform constraints). Using the Gaussian shorter-range connectivity, an extensive campaign of measures has been conducted, spanning over the three configurations above described. The impact of longer-range exponential decay interconnects has been evaluated on the 24 × 24 and 48 × 48 configurations. Fig. 3. Four snapshot of a slow wave propagating on a 48×48 grid of cortical modules spaced at 400 µm, with a connection probability exponentially decaying with λ = 240 µm. C. Toward biological modeling The network size and execution speed reached in the reported scaling experiments makes this engine a valuable candidate tool for the acceleration of large-scale simulations. Here, we report a preliminary example of usage in a specific case of our interest: the modeling of cortical Slow Wave Activity (SWA). To this purpose, we use a three-dimensional variant of the two-dimensional model [30]. The development of the variant and its biological meaning will be presented in a forthcoming publication (preliminary info in [31]). Snapshots of an exemplary propagating wave are reported in Fig. 3. Simulations express delta rhythms, the main feature of SWA, Fig. 4. Power spectral density of a population of excitatory neurons showing a high quantity of energy in delta band (< 4 Hz). as shown in their power spectral density (Fig. 4). The model includes 2.9 M neurons projecting 3.2 G synapses arranged in a grid of 48 × 48 cortical modules, spaced at 400 µm, with a connection probability exponentially decaying with λ = 240 µm. However, the focus of this paper is on the parallel and distributed computing aspects of the engine development and the cost of longer-range lateral connectivity. Papers targeting biological realism are currently under preparation in cooperation with the partners of the WaveScalES experiment in the Human Brain Project. D. Normalized simulation Cost per Synaptic Event Fig. 2. Example of Gaussian (green) and exponential longer-range (orange) connectivity in a grid composed by 24 × 24 cortical modules: total number of synapses (in thousands) projected by excitatory neurons located in the column marked in grey. Inhibitory neurons project only local connections. Different network sizes and connectivity models have been used in this scaling analysis. This results in heterogeneous measures of the elapsed time due to different numbers of projected synapses and to the different firing rates of resulting models. For example, the observed firing rate is ∼7.5 Hz for the shorter range connectivity, and in the range between 32 and 38 Hz for the longer range one (all other parameters being kept constant). However, a direct comparison is possible converting the execution time into a simulation cost per synaptic event. This normalized cost is computed dividing the elapsed time per simulated second of activity by the number of synapses and by their mean firing rate. In this way, a simple comparison among different simulated configurations is possible: measures from different simulations can be compared on the same plot. Our simulations include two kind of synapses: recurrent — i.e. projected by simulated neurons — and synapses bringing an external stimulus. Summing the number of events generated by recurrent and external synapses, in the following we can normalize the cost to the total number of equivalent synaptic events. E. Hardware Platform B. Impact of longer range exponential decay connectivity Fig. 7 compares the impact of shorter and longer lateral connectivity on the strong scaling behaviour. Circles represent measurements for the Gaussian decay while diamonds involve the longer range exponential one. The introduction of longer range connectivity increases the simulation cost per synaptic event, with a slow-down between 1.9 and 2.3 times, (see Fig. 8). The actual elapsed simulation time increased up to 16.6 times for the exponential longer-range connectivity due to the combined effect produced by: (i) the number of synapses projected by each neuron is higher (by a factor of 1.65), (ii) the firing rates expressed by We run the simulations on a partition of 64 IBM nodes (1024 cores) of the GALILEO server platform, provided at the CINECA [32] supercomputing center. Each 16-core computational node contains two Intel Xeon Haswell 8-core E5-2630 v3 processors, with a clock of 2.40 GHz. All nodes are interconnected through an InfiniBand network, equipped with 4× QDR switches. Due to the specific configuration of the server platform, no hyper-threading is allowed. Therefore, in all following measures, the number of cores corresponds exactly to the number of MPI processes launched on a computational node at each execution. IV. R ESULTS A. Scaling for shorter range Gaussian connectivity We collected wall clock execution times simulating different problem sizes (detailed in Table I), spanning from 1 to 1024 MPI processes (or, equivalently, hardware cores). Fig. 5 is about the strong scaling of the execution time per synaptic event. The black dotted line is the ideal scaling: doubling the resources, the execution time should halve. For the 24 × 24 grid (0.9 G recurrent synapses and 1.2 G total equivalent synapses) the time scales from 275 ns per synaptic event, using a single core, down to 4.09 ns per event using 96 cores. The corresponding speed-up is 67.3 times, losing ∼30% compared to the ideal (96 times). For the 48 × 48 grid (3.5 G recurrent, 5 G equivalent synapses) the speed-up is 54.2 times (ideal 96 times). For the 96 × 96 grid (14.2 G recurrent/ 20.4 G equivalent synapses) the speed-up is 10.8 times (in this case 16 times would be the ideal). Figure 6 reports six curves of weak scaling: constant workload per core, while increasing the number of resources and the problem size by up to 16 times. The weak scaling efficiency ranges from 72% (for a workload of 110.7 M synapses per core) down to 54% (when only 13.8 M synapses per core are allocated). Ideal weak scaling (100% efficiency) would produce horizontal lines. Three points per workload are reported: indeed, data have been extracted from the run times of the three configurations 24 × 24, 48 × 48, 96 × 96 used for strong scaling analysis. In our experience main factors affecting the scaling are collective communications and timing jitter of individual processes due to both operating system interruptions and fluctuations in local workload [18]. Fig. 5. Strong scaling for Gaussian connectivity model: the measures are expressed in elapsed time per equivalent synaptic event. Fig. 6. Weak scaling for Gaussian connectivity model. Fig. 7. Impact of lateral connectivity: the graph compares the execution time per synaptic event for the configurations with Gaussian connectivity (shorter range, lower number of synapses - circles) and the one with exponential connectivity (longer range, higher number of synapses - diamonds). TABLE I P ROBLEM SIZES FOR THE COMPARISON OF SIMULATOR PERFORMANCE APPLIED TO EXPONENTIAL ( LONGER - RANGE ) AND G AUSSIAN ( SHORTER - RANGE ) CONNECTIVITY. Grid Columns Neurons 24 × 24 48 × 48 96 × 96 576 2304 9216 0.7 M 2.9 M 11.4 M Number of Synapses Gaussian Connectivity Exponential Connectivity Recurrent Total Recurrent Total 0.9 G 1.2 G 1.5 G 1.8 G 3.5 G 5.0 G 5.9 G 7.4 G 14.2 G 20.4 G 23.4 G 29.6 G MPI Procs Min Max 1 4 64 64 256 1024 Fig. 9. Memory occupation in byte per synapse for different configurations in the two connectivity systems Fig. 8. Time per simulated synaptic event increased between 1.9 and 2.3 times on changing the decay of connection probability from the shorter range Gaussian scheme to the longer range exponential one. the model is between 4.3 and 5.0 times higher and (iii) the higher cost of longer range communication and demultiplexing neural spiking messages. Point (iii) should be well estimated by the slow-down of the normalized simulation cost per synaptic event. The execution of longer range connectivity on 96 cores reached about 83% for the 48 × 48 (5.9 G recurrent synapses) and 79% of the ideal scaling for the 24 × 24 case (1.5 G recurrent synapses). C. Memory cost per synapse We measured the total amount of memory allocated and divided it by the number of represented synapses. With no plasticity, each synapse should cost 12 Byte. Peak memory usage is observed at the end of initialization, when each synapse is represented at both the source and target process. Afterwards, memory is released on the source process. The forecast of minimum peak cost is therefore 24 Byte/synapse for static synapses. Fig. 9 shows the maximum memory footprint for different networks sizes and projection ranges, distributed over different numbers of MPI processes. The values are in the range between 26 and 34 Byte per synapse. We observed that the growing cost for higher number of MPI processes is mainly due to the memory allocated by the MPI libraries. V. D ISCUSSION Recent experimental results suggest the need of supporting long range lateral connectivity in neural simulation of cortical areas — e.g. modeled by simple exponential decay of the connection probability — with layer to layer specific decay constants, in the order of several hundreds of microns. A distributed spiking neural network simulation engine (DPSNN) has been applied to two-dimensional grids of neural columns spaced at 100 µm connected using two schemes. The longer-range connectivity model corresponds to an exponential connectivity decay (λ = 290 µm) and to the projection of approximately ∼2390 synapses per neuron. The performance of the engine is compared to that obtained with a shorter range Gaussian decay of connectivity, with a decay constant of the order of the columnar spacing and a lower number of synapses per neuron (∼1240). The impact of longer-range intra-areal exponential connectivity is indeed observable: it increases the simulation cost per synaptic event between 1.9 and 2.3 times compared to traditional shorter-range. The trends of the scaling are quite similar for the two studied connectivities. Notwithstanding the slow-down due to longer range connectivity, the engine demonstrates the ability to simulated large grids of neural columns (up to 96 × 96), containing a total of up to 11.4 M LIF neurons with spike-frequency adaptation, and representing up to 30 G equivalent synapses on a 1024 core execution platform, with a memory occupation always below 35 Byte/synapse. This is enough to simulate, on clusters of moderate size, cortical slabs with long-range intra-areal lateral interconnect, enabling the modeling of cortical slow waves, a first objective of our team. A second objective of DPSNN is to support the hardwaresoftware co-design of architectures dedicated to neural simulation. In perspective, we note that more detailed biological simulations of cortical areas could require further extensions of lateral connectivity models and the support of more complex connection motifs at different spatial scales. A further element in future whole brain simulations will be the co-design with white matter inter-areal connectome, which brings sparser connections at system scale. A balance between approaches focalizing on sparse connectivity like [33] and those considering spatial localization (like the one adopted by DPSNN) will have to be carefully addressed for efficient multiscale simulations of the whole brain. The results here presented, combined with previous experiences related to jitter of execution times of individual processes and the impact of collective communications when profiling DPSNN execution on distributed platforms, jointly suggest the utility of designing improved hierarchical communication infrastructures for spiking messages, mechanisms of synchronization of computing nodes and dedicated hardware accelerators. The improvements should consider requirements imposed by biological connectivity, at least for those engines that adopt mapping strategies of neurons and incoming synapses based on spatial contiguity. In such a context, DPSNN can be used to measure the impact of improved designs of execution platforms. ACKNOWLEDGMENT This work has received funding from the European Union Horizon 2020 Research and Innovation Programme under Grant Agreement No. 720270 (HBP SGA1) and under Grant Agreement No. 671553 (ExaNeSt). Simulations have been performed on the Galileo platform, provided by CINECA in the frameworks of HBP SGA1 and of the INFN-CINECA Computational theoretical physics initiative collaboration. We acknowledge G. Erbacci (CINECA) and L. Cosmai (INFN) for the platform setup support. R EFERENCES [1] P. Schnepel, A. Kumar, M. Zohar, A. Aertsen, and C. Boucsein, “Physiology and impact of horizontal connections in rat neocortex,” Cerebral Cortex, vol. 25, no. 10, pp. 3818–3835, 2015. [2] A. Stepanyants, L. M. Martinez, A. S. Ferecsk, and Z. F. Kisvrday, “The fractions of short- and long-range connections in the visual cortex,” Proceedings of the National Academy of Sciences, vol. 106, no. 9, pp. 3555–3560, 2009. [3] C. Boucsein, M. Nawrot, P. Schnepel, and A. Aertsen, “Beyond the cortical column: Abundance and physiology of horizontal connections imply a strong role for inputs from the surround,” Frontiers in Neuroscience, vol. 5, p. 32, 2011. [4] A. Schz, D. Chaimow, D. Liewald, and M. Dortenman, “Quantitative aspects of corticocortical connections: A tracer study in the mouse,” Cerebral Cortex, vol. 16, no. 10, pp. 1474–1486, 2006. [5] T. C. Potjans and M. Diesmann, “The cell-type specific cortical microcircuit: Relating structure and activity in a full-scale spiking network model,” Cerebral Cortex, vol. 24, no. 3, pp. 785–806, 2014. [6] A. Litwin-Kumar and B. Doiron, “Slow dynamics and high variability in balanced cortical networks with clustered connections,” Nature Neuroscience, vol. 15, p. 1498, 2012. [7] E. Gal, M. London, A. Globerson, S. Ramaswamy, M. W. Reimann, E. Muller, H. Markram, and I. Segev, “Rich cell-type-specific network topology in neocortical microcircuitry,” Nature Neuroscience, vol. 20, p. 1004, 2017. [8] M.-O. Gewaltig and M. Diesmann, “Nest (neural simulation tool),” Scholarpedia, vol. 2, no. 4, p. 1430, 2007. [9] S. Kunkel et al., “Nest 2.12.0,” Mar. 2017. [10] R. Brette et al., “Simulation of networks of spiking neurons: A review of tools and strategies,” Journal of Computational Neuroscience, vol. 23, pp. 349–398, Dec 2007. [11] M. A. Wilson, U. S. Bhalla, J. D. Uhley, and J. M. Bower, “Genesis: A system for simulating neural networks,” in Advances in Neural Information Processing Systems 1 (D. S. Touretzky, ed.), pp. 485–492, Morgan-Kaufmann, 1989. [12] M. Mattia and P. D. Giudice, “Efficient event-driven simulation of large networks of spiking neurons and dynamical synapses,” Neural Computation, vol. 12, pp. 2305–2329, Oct 2000. [13] J. M. Nageswaran, N. Dutt, J. L. Krichmar, A. Nicolau, and A. V. Veidenbaum, “A configurable simulation environment for the efficient simulation of large-scale spiking neural networks on graphics processors,” Neural Networks, vol. 22, no. 5, pp. 791 – 800, 2009. Advances in Neural Networks Research: IJCNN2009. [14] E. M. Izhikevich and G. M. Edelman, “Large-scale model of mammalian thalamocortical systems,” Proceedings of the National Academy of Sciences, vol. 105, no. 9, pp. 3593–3598, 2008. [15] D. S. Modha, R. Ananthanarayanan, S. K. Esser, A. Ndirango, A. J. Sherbondy, and R. Singh, “Cognitive computing,” Commun. ACM, vol. 54, pp. 62–71, Aug. 2011. [16] S. B. Furber, D. R. Lester, L. A. Plana, J. D. Garside, E. Painkras, S. Temple, and A. D. Brown, “Overview of the spinnaker system architecture,” IEEE Transactions on Computers, vol. 62, pp. 2454–2467, Dec 2013. [17] S. Schmitt et al., “Neuromorphic hardware in the loop: Training a deep spiking network on the brainscales wafer-scale system,” CoRR, vol. abs/1703.01909, 2017. [18] P. S. Paolucci et al., “Dynamic many-process applications on many-tile embedded systems and HPC clusters: The EURETILE programming environment and execution platforms,” Journal of Systems Architecture, vol. 69, pp. 29–53, 2016. [19] “Euroexa website.” Accessed: 30/Oct/2017. [20] M. Katevenis et al., “The ExaNeSt Project: Interconnects, Storage, and Packaging for Exascale Systems,” in 2016 Euromicro Conference on Digital System Design (DSD), pp. 60–67, Aug 2016. [21] P. S. Paolucci, R. Ammendola, A. Biagioni, O. Frezza, F. Lo Cicero, A. Lonardo, M. Martinelli, E. Pastorelli, F. Simula, and P. Vicini, “Power, Energy and Speed of Embedded and Server Multi-Cores applied to Distributed Simulation of Spiking Neural Networks: ARM in NVIDIA Tegra vs Intel Xeon quad-cores,” arXiv:1505.03015, June 2015. [22] M. Ruiz-Mejias, L. Ciria-Suarez, M. Mattia, and M. V. SanchezVives, “Slow and fast rhythms generated in the cerebral cortex of the anesthetized mouse,” Journal of Neurophysiology, vol. 106, no. 6, pp. 2910–2921, 2011. [23] A. Stroh, H. Adelsberger, A. Groh, C. Rhlmann, S. Fischer, A. Schierloh, K. Deisseroth, and A. Konnerth, “Making waves: Initiation and propagation of corticothalamic ca2+ waves invivo,” Neuron, vol. 77, no. 6, pp. 1136 – 1150, 2013. [24] “Human Brain Project.” [25] R. Gütig, R. Aharonov, S. Rotter, and H. Sompolinsky, “Learning input correlations through nonlinear temporally asymmetric hebbian plasticity,” Journal of Neuroscience, vol. 23, no. 9, pp. 3697–3714, 2003. [26] S. Song, K. D. Miller, and L. F. Abbott, “Competitive hebbian learning through spike-timing-dependent synaptic plasticity,” 2000. [27] A. Morrison, C. Mehring, T. Geisel, A. Aertsen, and M. Diesmann, “Advancing the boundaries of high-connectivity network simulation with distributed computing,” Neural Computation, vol. 17, pp. 1776–1801, Aug 2005. [28] J. Lazzaro, J. Wawrzynek, M. Mahowald, M. Sivilotti, and D. Gillespie, “Silicon auditory processors as computer peripherals,” pp. 820–827, 1993. [29] G. Gigante, M. Mattia, and P. D. Giudice, “Diverse population-bursting modes of adapting spiking neurons,” Phys. Rev. Lett., vol. 98, p. 148101, Apr 2007. [30] C. Capone, B. Rebollo, A. Muoz, X. Illa, P. Del Giudice, M. V. SanchezVives, and M. Mattia, “Slow waves in cortical slices: How spontaneous activity is shaped by laminar structure,” Cerebral Cortex, pp. 1–17, 2017. [31] E. Pastorelli, C. Capone, F. Simula, P. Del Giudice, M. Mattia, and P. S. Paolucci, “Distributed large scale simulation of synchronour slow-wave / asynchronous awake-like cortical activity,” 2017. Poster presented at: 5th Annual Human Brain Project Summit, 17-20 October 2017, Glasgow. http://apegate.roma1.infn.it/mediawiki/images/6/62/HBPSummitDPSNN.pdf. [32] GALILEO at CINECA, The Italian Tier-1 cluster for industrial and public research. [33] S. Kunkel et al., “Spiking network simulation code for petascale computers,” Frontiers in Neuroinformatics, vol. 8, p. 78, 2014.
9cs.NE
arXiv:1602.03479v1 [math.GR] 10 Feb 2016 COMMUTATORS AND CARTAN SUBALGEBRAS IN LIE ALGEBRAS OF COMPACT SEMISIMPLE LIE GROUPS J. MALKOUN AND N. NAHLUS Abstract. First we give a new proof of Goto’s theorem for Lie algebras of compact semisimple Lie groups using Coxeter transformations. Namely, every x in L = Lie(G) can be written as x = [a, b] for some a, b in L. By using the same method, we give a new proof of the following theorem (thus avoiding the classification tables of fundamental weights): in compact semisimple Lie algebras, orthogonal Cartan subalgebras always exist (where one of them can be chosen arbitrarily). Some of the consequences of this theorem are the following. (i) If L = Lie(G) is such a Lie algebra and C is any Cartan subalgebra of L, then the G-orbit of C ⊥ is all of L. (ii) The consequence in part (i) answers a question by L. Florit and W. Ziller on fatness of certain principal bundles. It also shows that in our case, the commutator map L × L → L is open at (0, 0). (iii) given any regular element x of L, there exists a regular element y such that L = [x, L] + [y, L] and x, y are orthogonal. Then we generalize this result about compact semisimple Lie algebras to the class of non-Hermitian real semisimple Lie algebras having full rank. Finally, we survey some recent related results , and construct explicitly orthogonal Cartan subalgebras in su(n), sp(n), so(n). 1. Introduction The Lie algebra version of Goto’s Theorem [10] for compact semisimple Lie groups is not as well known as it ought to be. It says that Theorem 1.1. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group G. Then every element x in L, can be written as x = [a, b] for some a, b in L. The first proof of Theorem 1.1 (i.e. Goto’s Theorem, additive version) was proved by Karl-Hermann Neeb [13, p. 653] (that he has communicated to the authors of [13]). His proof was based on Kostant’s convexity Theorem [16, Thm. 8.1], or more precisely, the version of Date: Received: date / Accepted: date. 1 2 J. MALKOUN AND N. NAHLUS Kostant’s convexity Theorem for compact semisimple Lie groups. Such proof will be presented in Appendix A but slightly simplified using [1, Lemma 2.2]. In section 2, we prove Theorem 1.1 directly by using any Coxeter transformation of the Weyl group W (L, C) where C is a maximal toral subalgebra of L. We remark that certain Coxeter transformations were used to prove Goto’s Theorem (on the group level) as in Bourbaki’s book [4, Corollary, section 4 of chapter 9] or Hofmann and Morris’s book [13, Corollary 6.56]. To the best of our knowledge, this direct proof is new. In section 3, we prove the following theorem. Theorem 1.2. Let Lie(G) be the Lie algebra of a compact semisimple Lie group G. Then, with respect to the negative of its Killing form, Lie(G) has orthogonal Cartan subalgebras (maximal toral subalgebras) where one of them can be chosen arbitrarily. This interesting result was proved recently by d’Andrea and Maffei in [7, Lemma 2.2], via the classification tables of fundamental weights for all types of simple Lie algebras (except type An ). Specifically, their proof uses the tables in [4] to verify that (in any root system of (L, C) where C is a Cartan subalgebra of L), the highest root is either equal or twice some fundamental weight (in all simple Lie algebras except of type An ). However our proof is a simple consequence of our methods in section 2. Theorem 1.2 has the following Corollary. Corollary 1.3. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group G. Let C be a Cartan subalgebra of L and let C ⊥ be the orthogonal complement of C (with respect to negative of the Killing form on L). Then the G-orbit of C ⊥ is all of L. We note that Corollary 1.3 can be easily obtained directly from Kostant’s convexity Theorem (see Appendix A). In fact, Corollary 1.3 was essentially the key step in Karl-Hermann Neeb’s proof of Goto’s Theorem, additive version. In section 4, we present two “applications” of Corollary 1.3. First, Corollary 1.3 answers a question by L. Florit and W. Ziller on fatness of certain principal bundles as follows Corollary 1.4. If G is a compact semisimple Lie group, endowed with minus its Killing form, and G ,→ P → B is a principal G-bundle on B with total space P , endowed with a connection 1-form θ : T P → g, then θ is fat if and only if C ⊥ is fat, where C can be chosen to be any Cartan subalgebra of g. Our second application of Corollary 1.3 is the following. Corollary 1.5. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group G. Then the commutator map L×L → L is open at (0, 0). COMMUTATORS AND CARTAN SUBALGEBRAS 3 This fact was obtained in [7] using their Lemma 2.2 which was proved by using some tables about fundamental weights as explained above. (Recall that their Lemma 2.2 is our Theorem 1.2 above). However, after a careful study of their proof, we found out that one only needs the partial result stated in Corollary 1.3 above. In section 5.1, we make the following remark. Corollary 1.6. Let L be a compact semisimple Lie algebra, and let a be a regular element of L. Then there exists a regular element b in L, such that L = [a, L] + [b, L] and a, b are orthogonal (in fact the centralizers of a and b in L are also orthogonal). More generally, we have Corollary 1.7. Let L = k ⊕ p be a Cartan decomposition of a semisimple Lie algebra L (where k is a maximal compact Lie subalgebra) such that L is non-Hermitian of full rank. That is, k is semisimple and rank(k) = rank(L). Let a be a regular element of k. Then there exists a regular element b in k, such that L = [a, L] + [b, L] and a, b are orthogonal. In section 5.1, we make the following trivial remark. Remark 1.8: The following two conjectures are equivalent. 1) Every element x in a real semisimple Lie algebra L is the commutator of two elements where one element can be chosen to be regular (hence semisimple). 2) Every element x in a real semisimple Lie algebra L is orthogonal to some Cartan subalgebra of L. In section 5.2, we survey most of the results of D. Akhiezer in his recent interesting paper [1] that shows that the above conjecture is valid in many real simple Lie algebras (see Theorem 5.3) In section 5.3, we survey one result by G. Bergman and N. Nahlus in [2] and another result in preparation by the second author in [19] related to 1.5 generators of simple Lie algebras. For example, in any simple Lie algebra L over a field of characteristic 0, and for any x in L \ {0}, there exists a regular (hence semisimple) element y in L such that L = [L, x] + [L, y] ([19]). This last property is related to the concept of 1.5 generators as follows: if x and y generate L as a Lie algebra then L = [L, x] + [L, y] by [3, Lemma 25c]. Finally in section 6, we construct explicit examples of orthogonal Cartan subalgebras in the cases of su(n), sp(n), based on the idea of circulant matrices (cf. [17]), and in the case of so(n). 4 J. MALKOUN AND N. NAHLUS Acknowledgements. The authors are grateful to Anthony Knapp for his help in answering some questions about the general theory of noncompact semisimple Lie algebras. We are also indebted to Karl-Hermann Neeb for his encouragement and interesting discussions about two results in the paper. Finally, we are also indebted to Wolfgang Ziller for kindly pointing out that our Corollary 1.3 answers a question about fatness in principal bundles. 2. Goto’s theorem by Coxeter transformations In this section, we give a proof of Goto’s theorem, additive version, by using Coxeter transformations. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Let C be a maximal toral subalgebra of L. Let Σ ⊆ C ∗ (the dual of C) be the root system of G with respect to C and we fix an ordering of the roots, with ∆ as the set of positive roots, and let {a1 , . . . , an } be the set of simple positive roots. Lemma 2.1. Let {s1 , . . . , sn } be the Weyl reflections corresponding to the simple positive roots {a1 , . . . , an }. Then, for any permutation of {1, . . . , n}, the Coxeter transformation c = s1 s2 · · · sn has no fixed points, other than 0. That is, 1 is not an eigenvalue of c. See the very short proof in [14], p.76. Lemma 2.2. (cf. [12], Lemma 6.53) Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Let C = Lie(T ) be a maximal toral subalgebra. Let c be any Coxeter transformation with respect to (L, C) as in lemma 2.1. c = Ad(n) for some n ∈ N (T ), the normalizer of T in G, since c is in the Weyl group of (G, T ). Then Ad(n)|C − Id : C → C is an isomorphism. Proof. Immediate from lemma 2.1.  Theorem 2.3. In the setting of lemma 2.2, suppose n = exp(N ) for some N in L = Lie(G) (which exists since the exponential map is surjective in connected compact semisimple Lie groups). Then C is contained in [L, N ]. Proof. Let x ∈ C = Lie(T ). Then by lemma 2.2, x = Ad(n)(t) − t for some t in C. But the exponential map is surjective in compact connected semisimple Lie groups, so n = exp(N ), for some N ∈ L. Hence x = Ad(n)(t) − t = Ad(exp(N ))(t) − t = exp(ad(N ))(t) − t COMMUTATORS AND CARTAN SUBALGEBRAS 5 But x = exp(ad(N ))(t) − t can be written as [N, y(x)], for some y(x) ∈ L (by factorization) through writing the exponential series as the limit of its partial sums. Hence x = [N, y(x)].  Next we strengthen Theorem 2.3 to replace N by a regular element of L. Theorem 2.4. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group G, and let C be a Cartan subalgebra of L. Then there exists a regular element a of L such that C ⊆ [a, L]. In particular, every element x of L can be written as x = [a, b] where a can be chosen to be regular. Comment: for comparison, the proof by Kostant’s convexity Theorem in the Appendix shows that C ⊥ ⊆ [x, L] for some regular element x of L. Proof. By Theorem 2.3, C ⊆ [N, L]. Since N is semisimple, N is contained in a Cartan subalgebra C 0 of L, so C 0 ⊆ Z(N ). Now C 0 = Z(a) for some regular element a of L. So Z(a) ⊆ Z(N ). Hence Z(N )⊥ ⊆ Z(a)⊥ . But Z(N )⊥ = [N, L] and Z(a)⊥ = [a, L] by Remark 5.2. Hence [N, L] ⊆ [a, L]. So C ⊆ [a, L] as desired. The last conclusion follows from the fact that every element of L belongs to a Cartan subalgebra, and all Cartan subalgebras of L are conjugate.  Example 2.5: Consider su(2k + 1), which is a compact semisimple Lie algebra. It can be shown by a simple calculation that   0 0 ··· 0 1  1 0 ··· 0 0   . . . .. ..  . . . n= . . . . .     0 0 ··· 0 0  0 0 ··· 1 0 is an element of N (T ) which gives modulo T a Coxeter transformation of SU (2k + 1). Consider the unitary change of basis matrix g = (gab ) for 1 ≤ a, b ≤ 2k + 1, with gab = √ 1 (a−1)(b−1) γk 2k + 1 2πi where γk = exp( 2k+1 ). Let D = diag(1, γ̄, γ̄ 2 , · · · , γ̄ 2k ). Then a simple calculation shows that n = gDg−1 . So we let Λ = diag(0, −ick , −2ick , · · · , −kick , kick , (k − 1)ick , · · · , ick ), 6 J. MALKOUN AND N. NAHLUS 2π where ck = 2k+1 . It is clear that Λ is an element of su(2k + 1) and that exp(Λ) = D. Hence we obtain that N = gΛg−1 is an element of su(2k + 1) which satisfies exp(N) = n. Example 2.6: The case of su(2k) can be dealt with in a similar fashion. We obtain that   0 0 · · · 0 −1  1 0 ··· 0 0   . . . .. ..  . . . n= . . . . .     0 0 ··· 0 0  0 0 ··· 1 0 is an element of N (T ) which gives modulo T a Coxeter transformation of SU (2k). Consider the unitary change of basis given by g = (gab ), where 1 (a−1)(2b−1) gab = √ βk 2k πi where βk = exp( 2k ). Also, let D = (Dab ), where   (2a − 1)πi Dab = exp − δab 2k Then a simple calculation shows that n = gDg−1 . Finally, let Λ = diag(−idk , −3idk , · · · , −(2k−1)idk , (2k−1)idk , (2k−3)idk , · · · , idk ) π . As in example 2.5, N = gΛg−1 is an element in su(2k) where dk = 2k satisfying n = exp(N). 3. Orthogonal Cartan subalgebras In this section, we prove the following non-trivial theorem which was proved in [7] lemma 2.2, via the classification tables of all fundamental weights. Their proof relies on the observation found in the classification tables [4], that the highest root is equal or twice some fundamental weight, in all simple Lie algebras except those of type An . It would be nice to have a more direct argument for this nice result, and this is what we shall present in this section. Theorem 3.1. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Then L has orthogonal Cartan subalgebras (where one of them can be chosen arbitrarily) COMMUTATORS AND CARTAN SUBALGEBRAS 7 Proof. Let C be a Cartan subalgebra of L. Then by Theorem 2.4, C ⊆ [a, L] for some regular element a of L. Since a regular in L, Z(a) is a Cartan subalgebra of L. Moreover, Z(a)⊥ = [a, L] by Remark 5.2. Hence C ⊆ Z(a)⊥ . Thus the Cartan subalgebras C and Z(a) are orthogonal. Finally, C can be chosen arbitrarily since all Cartan subalgebras (in Lie algebras of compact semisimple Lie groups) are conjugate and the Killing form is invariant under automorphisms of L [15, Prop. 1.119]. This proves the theorem.  Theorem 3.2. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Let C be a given Cartan subalgebra, and let C ⊥ be the orthogonal of C (with respect to the negative of the Killing form on L). Then for every element x of L, there exists an element g in G , such that g.x belongs to C ⊥ . That is, C ⊥ intersects every G-orbit of L. Proof. By Theorem 3.1, there exists a Cartan subalgebra C 0 which is orthogonal to C. For every element x in L, we know that there exist a g in G such that g.x ∈ C 0 . Hence g.x is orthogonal to C, so g.x belongs to C ⊥ . Thus, C ⊥ intersects every G-orbit in L. (Or equivalently, L is the union of all g.C ⊥ as g varies over G).  As noted in the introduction, Theorem 3.2 can be obtained almost immediately from Kostant’s Convexity Theorem (See Appendix, Theorem A.3). Corollary 3.3. Let L be a compact semisimple Lie algebra and let a be a given regular element of L. Then there exists a regular element b in L, such that L = [a, L] + [b, L] and b is orthogonal to a (under the negative of the Killing form of L). Proof. By Theorem 3.1, there exist orthogonal Cartan subalgebras C and D in L, so D ⊆ C ⊥ . Since L = D ⊕ D⊥ , it follows that L = C ⊥ + D⊥ . Let x be any regular element of L in C and let y be any regular element of L in D, so C = Z(x) and D = Z(y). Thus L = Z(x)⊥ + Z(y)⊥ . But Z(x)⊥ = [x, L] and Z(y)⊥ = [y, L] by Remark 5.2. Hence L = [x, L] + [y, L]. Note that x and y are regular in L and orthogonal (since they belong to C and D which are orthogonal). The given regular (semisimple) element a belongs to a Cartan subalgebra and all Cartan subalgebras in L are conjugate. Hence there exists an inner automorphism f such that f (a) is in C. Since a and hence f (a) is regular in L, the element x in the previous paragraph can be chosen to be f (a). Hence L = [f (a), L] + [y, L]. Now we apply f −1 on both sides to get L = [a, L] + [f −1 (y), L]. Since x = f (a) and y are orthogonal, then so are a and f −1 (y) since automorphisms preserve 8 J. MALKOUN AND N. NAHLUS orthogonality [15, Prop. 1.119]. Finally, since y is regular, then so is f −1 (y).  Corollary 3.3 can be easily generalized to the case of non-Hermitian real semisimple Lie algebras of full rank. Corollary 3.4. Let L = k ⊕ p be a Cartan decomposition of a semisimple Lie algebra L (where k is the compact Lie algebra) such that L is non-Hermitian of full rank. That is, k is semisimple and rank(k) = rank(L). Let a be a regular element of k. Then there exists a regular element b in k, such that L = [a, L] + [b, L] and b is orthogonal to a. Moreover the centralizers Z(a) and Z(b) of a and b in L are also orthogonal. Proof. The proof is very similar to the proof of Corollary 3.3 with the evident modifications as follows. By Theorem 3.1, there exist orthogonal Cartan subalgebras C and D in k. In fact, C and D are also orthogonal with respect to the Killing form on L (see Remark 3.5 below). Since rank(k) = rank(L), C and D are also Cartan subalgebras of L. Since the restriction of the Killing form of L on every Cartan subalgebra is non-degenerate, we have L = D + D⊥ . But D is a subset of C ⊥ in L. Hence L = C ⊥ + D⊥ . Let x be any regular element of L in C and let y be any regular element of L in D. Then the rest of the proof is verbatim as in the proof of Corollary 3.3, leading to L = [x, L] + [y, L] and finally, L = [a, L] + [f −1 (y), L] where a and f −1 (y) are regular and orthogonal in L.  Apart from the Lie algebras of compact semisimple Lie groups, some other examples of Lie algebras satisfying the hypotheses of Corollary 3.4 are so(2j, 2k) with j and k both > 1, so(2j, 2k + 1) if j > 1 and sp(p, q) where p and q are both positive. Remark 3.5: Over a field of characteristic 0, let S be a semisimple Lie subalgebra of a semisimple Lie algebra L, and let C be a Cartan subalgebra of S. Then the orthogonal space of C in S (with respect to the Killing form of S) is contained in the orthogonal space of C in L (with respect to the Killing form of L). Proof. We know that C = Z(a) for some regular element a of S. Moreover, the restriction of the Killing form of L to S is also non-degenerate by a simple application of Cartan’s solvability criterion by [3, chapter 1, section 6.1, Proposition 1]. Hence, by [1, Lemma 2.2] (stated after Remark 5.2 below), the orthogonal space of C = Z(a) in S coincides COMMUTATORS AND CARTAN SUBALGEBRAS 9 with [a, S] independently of the invariant non-degenerate symmetric bilinear form on S.  4. Applications 4.1. Openness of commutator maps. We start with the following immediate application of Theorem 3.2. Lemma 4.1. Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group, and let C = Lie(T ) be a maximal toral subalgebra of L. Let V be a G-stable neighborhood of 0 ∈ L. Then G.(V ∩ C ⊥ ) = V Proof. This follows immediately from Theorem 3.2 as follows. Since V is G-stable, G.(V ∩ C ⊥ ) ⊆ V . Conversely, if v ∈ V , then by Theorem 3.2, there exists a g ∈ G such that g.v ∈ C ⊥ . Hence g.v ∈ V ∩ C ⊥ ; so v ∈ G.(V ∩ C ⊥ ).  Using lemma 4.1, one can prove the following theorem. Theorem 4.2. ([7], theorem 2.1). Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Then the commutator map Comm : L × L → L given by Comm(x, y) = [x, y] is open at (0, 0). The proof of this theorem as given in [7] relies on their lemma 2.2, that we proved in Theorem 3.1 above (without the classification tables of the fundamental weights versus the highest roots). However, a careful study of their proof of Theorem 4.2 only requires the partial result in our lemma 4.1. Remark 4.3: It is worth noting that the authors in [7] were able to use the fact that the commutator map Comm : L × L → L given by Comm(x, y) = [x, y] is open at (0, 0) to prove their main result that CommG : G × G → G given by CommG (g1 , g2 ) = [g1 , g2 ] = g1 g2 g1−1 g2−1 is open at (Id, Id) in G × G ([7], Corollary 3.1). 4.2. Fatness of certain connections on principal fiber bundles. We first review the basic definition of fatness, in the context of principal fiber bundles with connection. Such material can easily be found in the literature (see the original paper [22] by A. Weinstein, or later works π such as [9] or [18]). Let G ,→ P → B be a principal G-bundle, with base manifold B, and total space P . Let θ : T P → g be a connection 1form on P , where T P denotes the tangent bundle of P . We shall denote the curvature of θ by Ω(−, −) : H × H → g, where T P = H ⊕ V is the decomposition into horizontal and vertical subbundles induced by θ. 10 J. MALKOUN AND N. NAHLUS Fix a bi-invariant (positive-definite) inner product B(−, −) : g×g → R on g. Definition 4.4. In the setting above of a principal G-bundle P with connection θ, an element u ∈ g is said to be fat if Ωu : H × H → R is non-degenerate, where Ωu is defined by Ωu (Y1 , Y2 ) = B(Ω(Y1 , Y2 ), u) We also say that a subspace V ⊆ g is fat if every non-zero element of V is fat. The connection θ is said to be fat if g is fat. We can easily deduce the following theorem, answering a question raised in [9, Remark 3.13] Theorem 4.5. If G is a compact semisimple Lie group, B(−, −) is π minus its Killing form, and G ,→ P → B is a principal G-bundle on B with total space P , endowed with a connection 1-form θ : T P → g, then θ is fat if and only if C ⊥ is fat, where C can be chosen to be any Cartan subalgebra of g. Proof. The “only if” part is trivial. The other direction follows from the fact that G.C ⊥ = g by 1.3 (which follows immediately from Kostant’s convexity theorem), and the known fact that if u ∈ g is fat, then g.u is also fat, for any g ∈ G (see for example [9]).  5. Commutators in semisimple Lie algebras and survey of some related results 5.1. Conjectures about commutators in semisimple Lie algebras. Remark 5.1: The following two conjectures are equivalent. 1) Every element x in a real semisimple Lie algebra L is the commutator of two elements where one element can be chosen to be regular (hence semisimple). 2) Every element x in a real semisimple Lie algebra L is orthogonal to some Cartan subalgebra of L. Proof. Note that every regular element of a semisimple Lie algebra (over a field of characteristic 0) is semisimple [4, Chapter 7, section 2, Corollary 2]. First we prove conjecture 1) implies conjecture 2). Suppose x = [a, b] where a is regular (semisimple). Then x belongs to [a, L]. But [a, L] = Z(a)⊥ by Remark 5.2. Hence x belongs to Z(a)⊥ . Now a is regular (semisimple), Z(a) is a Cartan subalgebra C of L. Consequently, x belongs to C ⊥ . Conversely, if x belongs to C ⊥ where C is a Cartan subalgebra of L. Then C = Z(a) for some regular COMMUTATORS AND CARTAN SUBALGEBRAS 11 element a of L. So x belongs to Z(a)⊥ which coincides with [a, L] by Remark 5.2. Hence x belongs to [a, L].  Remark 5.2: Let a be a semisimple element of a semisimple Lie algebra L over a field of characteristic 0. Then Z(a)⊥ = [a, L] with respect to the negative of the Killing form on L (where Z(a) is the centralizer of a in L. Remark 5.2 is part of Theorem 4.1.6 in [20] which can be checked directly. Roughly, Z(a) is orthogonal to [a, L], and both Z(a)⊥ and [a, L] have the same dimension. However, it is worth noting that Remark 5.2 (with essentiall the same proof) is also valid for any element of L and, for any invariant non-degenerate symmetric bilinear form on L as shown in [1, Lemma 2.1]. 5.2. Survey of some recent results on the commutator conjecture. In the recent paper [1], the commutator conjecture was proved in many cases. Theorem 5.3. [1, Thms 1.1–1.4] Let L = Lie(G) be a simple real Lie algebra. Let L = k ⊕ p be a Cartan decomposition of L where k is the reductive part. Let a be a Cartan subspace of p, and let m be the centralizer of a in k. Then every element x of L can be written as x = [a, b] where a can be chosen to be a regular (semisimple) element of L, in each of the following cases: (i) L is compact or split (i.e. m = 0) (ii) L is non-hermitian (i.e. k is semisimple) and rank(k) = rank(L) (iii) m is a semisimple Lie algebra The paper [1] relies on the Kostant convexity theorem and some interesting arguments by the author. For more information about complex or split semisimple Lie algebras, see [7, Thm. 4.3] and [6, Lemma 2]. 5.3. Survey on generators and 1.5 generators of simple Lie algebras. Theorem 5.4. [2, Thm. 26] Let L be a semisimple Lie algebra over a field of characteristic 0 (or characteristic p > 3). Then there exist x, y in L such that L = [x, L] + [y, L]. Theorem 5.5. [19, Thm. 2] (on 1.5 generators). Let L be a semisimple Lie algebra over any field of characteristic 0. For any fixed x ∈ L \ {0}, there exists a regular element y ∈ L such that L = [x, L] + [y, L] 12 J. MALKOUN AND N. NAHLUS We note that the above two theorems are related respectively to the concepts of generators and 1.5 generators in Lie algebras in view of the following fact: If x and y generate, in the Lie algebra sense, a Lie algebra L, then L = [L, x] + [L, y] by [3, Lemma 25c]. 6. Examples of orthogonal Cartan subalgebras In the examples of orthogonal Cartan subalgebras we will see, corresponding to the classical groups, circulant matrices play a special role. We begin by reviewing circulant matrices, following the presentation in [17]. Given a vector a = (a0 , a1 , . . . , an ) ∈ Cn , we define   a0 a1 · · · an−2 an−1  an−1 a0 · · · an−3 an−2   . .. . . .. ..  A= . . . .    ..  a a ··· a a  2 a1 Let  = e 2πi n 3 0 a2 · · · an−1 1 a0 , which is a primitive n’th root of unity, and let 1 xl = √ (1, l , 2l , . . . , (n−1)l ) ∈ Cn n for l = 0, . . . , n. We let U = (x0 , x1 , . . . , xn−1 ) In other words, U is the square matrix having as columns x0 , x1 ,. . ., xn−1 . It is easy to check that U is both unitary and symmetric. We introduce the numbers n−1 X λl = lj aj j=0 for l = 1, 2, . . . , n. Let Λ = diag(λ0 , λ1 , . . . , λn−1 ). A calculation shows that A = U ΛU −1 so that the λl s are the eigenvalues of A. An important observation is that U does not depend on A, so that the same unitary matrix U diagonalizes all complex circulant n by n matrices. In particular, this also implies that the space Circ(n) of all complex circulant n by n matrices is abelian (which can be seen directly too). COMMUTATORS AND CARTAN SUBALGEBRAS 13 6.1. The Lie algebras su(n). Theorem 6.1. If C is the Cartan subalgebra of su(n) consisting zero trace skew-hermitian diagonal n by n complex matrices, and C 0 = Circ(n) ∩ su(n) then C and C 0 are two orthogonal Cartan subalgebras of su(n), with respect to minus the Killing form of su(n). Proof. We already know that C 0 is abelian, since Circ(n) is abelian, so it only remains to prove that dimR (C 0 ) = n − 1. This is so because C 0 corresponds to vectors a having  a0 = 0 an−l = −āl , for 1 ≤ l ≤ n − 1 from which it follows indeed that the real dimension of C 0 is n − 1.  6.2. The Lie algebras sp(n). It is well known that gl(n, H) ⊆ gl(2n, C). This relies on the observation that a quaternion x can be written in a unique way as x = u + jv, where u and v are complex numbers. It can also be verified that the map   u −v̄ x 7→ v ū is an injective algebra homomorphism from H into gl(2, C). The Lie algebras sp(n) can be written as sp(n) = gl(n, H) ∩ u(2n) where both Lie algebras on the right-hand side are understood as Lie subalgebras of gl(2n, C), using the remark above. We then have the following theorem. Theorem 6.2. If C is the Cartan subalgebra of sp(n) consisting of complex diagonal pure imaginary 2n by 2n matrices, and C 0 = Circ(2n) ∩ sp(n) where both Lie algebras on the right-hand side are thought of as Lie subalgebras of gl(2n, C) in the usual way, then C and C 0 are two orthogonal Cartan subalgebras of sp(n), with respect to minus the Killing form of sp(n). Proof. The only thing to check is that dimR (C 0 ) = n. This essentially follows from the fact that Circ(2n) is a complex 2n-dimensional subspace of gl(2n, C), which admits the decomposition Circ(2n) = C 0 ⊕ iC 0 ⊕ jC 0 ⊕ kC 0 14 J. MALKOUN AND N. NAHLUS the latter being a real vector space decomposition, from which it can be deduced that dimR (C 0 ) = n, as claimed.  6.3. The Lie algebras so(n). Define    0k Λ C2k = : Λ is a real k by k diagonal matrix −Λ 0k    0k Λ 0k,1   C2k+1 =  −Λ 0k 0k,1  : Λ is a real k by k diagonal matrix   0 0 1,k 01,k Here 0k denotes the zero k by k matrix, while for instance 0j,k denotes the j by k zero matrix. It is easy to check that Cn is a Cartan subalgebra of so(n) for all n ≥ 3. Our plan in this section is to construct for each n, a Cartan subalgebra Cn0 which is orthogonal to Cn . Our next lemma will allow us to restrict our attention to so(n), with n even. 0 Lemma 6.3. if C2k is a Cartan subalgebra of so(2k) orthogonal to C2k , and if ι : so(2k) ,→ so(2k + 1) denotes the inclusion of Lie algebras 0 which satisfies ι(C2k ) = C2k+1 , then ι(C2k ) is a Cartan subalgebra of so(2k + 1) which is orthogonal to C2k+1 . Proof. This follows from the fact that so(2k) and so(2k + 1) both have rank k, and that the restriction of the Killing form of so(2k + 1) to so(2k) is a constant multiple of the Killing form of so(2k).  The next lemma will enable us to restrict our attention further to just so(4) and so(6). But first, we introduce the following notation. If X is a k by k matrix, we denote by X̃ the following k + 2 by k + 2 matrix X̃ = X ⊕ 02 Also, if x is 2k by 2k matrix, one can write x as   A B x= C D where A, B, C and D are each k by k matrices. We then denote by x̂ the following 2k + 4 by 2k + 4 matrix   Ã B̃ x̂ = C̃ D̃ Lemma 6.4. If SpanR {xi : 1 ≤ i ≤ k} is a Cartan subalgebra of so(2k) which is orthogonal to C2k , then SpanR {x̂i , y, z; 1 ≤ i ≤ k} is a Cartan COMMUTATORS AND CARTAN SUBALGEBRAS 15 subalgebra of so(2k + 4) which is orthogonal to C2k+4 , where   0k S y= −S 0k   0k T z= T 0k where  0 1 1 0  0 −1 1 0 S = 0k ⊕ T = 0k ⊕   Proof. It suffices to prove that the x̂i s, y and z all mutually commute, which is easy to check.  Hence, using lemmas 6.3 and 6.4, it suffices for our purposes to construct a Cartan subalgebra of so(4), respectively so(6) which is orthogonal to C4 , respectively C6 . Then we can, using the constructions in these two lemmas, construct orthogonal Cartan subalgebras of so(n) for any n ≥ 4 (the case of so(3) is trivial). Let us consider first so(4). Define  0 0 0 1  0 0 1 0 x1 =   0 −1 0 0 −1 0 0 0  0 0 0 −1  0 0 1 0 x2 =   0 −1 0 0 1 0 0 0         It is easy to check that x1 and x2 span a Cartan subalgebra of so(4) which is orthogonal to C4 . 16 J. MALKOUN AND N. NAHLUS We finally consider the case of so(6). Define   0 −1 0 A= 1 0 0  0 0 0   0 0 0 B= 0 0 0  1 0 0   0 0 0 C =  0 0 −1  0 1 0 We then define x1 = A ⊕ 03 , and   03 B x2 = −B T 03 as well as x3 = 03 ⊕ C. Then a straightforward computation shows that x1 , x2 and x3 mutually commute, so that they span a Cartan subalgebra of so(6), which is also orthogonal to C6 (since each of x1 , x2 and x3 is orthogonal to C6 ). This finishes our description of orthogonal Cartan subalgebras for so(n), for n ≥ 4, with the case of n = 3 being trivial. Appendix A. Some consequences of Kostant’s convexity theorem As noted in the introduction, we shall present, in this Appendix, a slight simplification of Karl-Hermann Neeb’s proof of Goto’s Theorem, additive version, using Kostant’s Convexity Theorem [13, p. 653]. His proof essentially passes through the proof of Theorem 3.2. Theorem A.1 (Kostant’s convexity theorem). Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Let C = Lie(T ) be a maximal toral subalgebra of L, corresponding to a maximal torus T ⊆ G. Let p : L → C be the orthogonal projection of L onto C with respect to the negative of the Killing form on L. Let x be an element of C, and let W be the Weyl group of (G, T ) or (L, C). Then p(G.x) = Conv(W.x). This version of Kostant’s Convexity Theorem follows very easily from the original version [16, Theorem 8.2] by taking the complexification L(G)C of L(G), viewed as a real semisimple Lie algebra. Then L(G)C has a Cartan decomposition k ⊕ p where k = L(G) and p = iL(G). So COMMUTATORS AND CARTAN SUBALGEBRAS 17 we apply Theorem 8.2 of [16] to L(G)C viewed as a real semisimple Lie algebra, with k and p as above. Lemma A.2. [1, lemma 2.2] Let V be a real vector space and W ⊂ GL(V ) a finite linear group acting without fixed vectors (other than 0). Then the convex hull of any W -orbit contains 0 ∈ V . For the reader’s convenience, we explain the idea of its short proof in [1] which is that the centroid z of the points of any W -orbit in V is fixed under the action of W . But z = 0 is the only fixed point under W . Hence z = 0 belongs to the convex hull of any W -orbit. Now we give another proof of Theorem 3.2. Theorem A.3. (= Theorem 3.2) In the setting of theorem A.1, let C ⊥ be the orthogonal complement of C (with respect to negative the Killing form on L). Then for every element x of L, there exists an element g ∈ G such that g.x belongs to C ⊥ . Proof. Combine A.1 and A.2.  Theorem A.4 (Goto’s theorem, additive version). In the Lie algebra L of a compact semisimple Lie group, every element x is a bracket. That is, x = [a, b] for some a, b in L. Moreover, a or b can be chosen to be regular. Proof. By A.3, there exists an element g in G such that g.x belongs to C ⊥ . But C = Z(a) for some regular element a of L. Moreover, Z(a)⊥ = [a, L] by 5.2. Hence g.x belongs to [a, L]. Consequently, x belongs to [g −1 .a, L], and g −1 .a is regular.  Appendix B. Alternative proof of Cor. 1.3 by SU (2)-rotations In this section, we give a more direct proof of Cor. 1.3 without using Kostant’s convexity theorem. The proof is based on SU (2) rotations in the orthogonal “root” space decomposition of the Lie algebra L = Lie(G) where G is a compact semisimple Lie group. We equip L with its natural inner product (negative the Cartan Killing form of L). Let C be a maximal toral subalgebra of L. Let Σ ⊆ C ∗ (the dual of C) be the root system of G with respect to C and we fix an ordering of the roots with ∆ as the set of positive roots. Then under the adjoint action of C, L has an orthogonal decomposition X L=C⊕ Lα α∈∆ 18 J. MALKOUN AND N. NAHLUS Let a be a positive root, and consider Ker(a) = {x ∈ C : a(x) = 0}, and its orthogonal space Ker(a)⊥ in C. Then we have X L = Ker(a) ⊕ Ker(a)⊥ ⊕ Lα α∈∆ where Ker(a)⊥ ⊕ La ' su(2). More specifically, each Lα has an orthogonal basis {uα , vα } and Ker(α)⊥ has a basis {hα } such that (i) [h, uα ] = α(h)vα for all h in C, (ii) [h, vα ] = −α(h)uα for all h in C, and {α, uα , vα }, (iii) [uα , vα ] = hα in C, (iv) Ker(α)⊥ ⊕ Lα = SpanR {hα , uα , vα } ' su(2). (cf. [12], 6.48, 6.49], [15], p. 353, [23], p. P 59) Let Sa = Ker(a)⊥ ⊕ La , and let ma = α∈∆\{a} Lα . Then L = Ker(a) ⊕ Sa ⊕ ma where Sa ' su(2). Lemma B.1. (cf. [21], lemma 2.2) In the above setting where L = Ker(a) ⊕ Sa ⊕ ma where a ∈ ∆ is a positive root, let x ∈ C. Then there exists a g in the Lie subgroup Ga ' SU (2) corresponding to Sa ' su(2), such that (i) g.x has no Ker(a)⊥ component (ii) x and g.x have the same Ker(a) components (iii) the ma components of x and g.x have the same norms Proof. Let x0 be the component of x in Ker(a)⊥ . So x0 ∈ Sa = Ker(a)⊥ ⊕ La ' su(2). Hence there exists a g in Ga ' SU (2) such that g.x0 has no Ker(a)⊥ component. We do this via rotations in SU (2) (or the fact that all Cartan subalgebras of su(2) are conjugate). Because we are dealing with Ker(a), the decomposition (B.1) L = Ker(a) ⊕ Sa ⊕ ma is easily checked to be invariant under the action of Sa = Ker(a)⊥ ⊕ La . Hence g.x has no Ker(a)⊥ component, which proves (i). Moreover, [Sa , Ker(a)] = [La , Ker(a)] = 0. Consequently, the group Ga ' SU (2) fixes Ker(a). Hence x and g.x have the same Ker(a) components, thus proving (ii). Finally, since the action of Ga preserves the decomposition B.1, and every element of Ga acts by orthogonal transformations (with respect to minus the Killing form), (iii) follows.  COMMUTATORS AND CARTAN SUBALGEBRAS 19 Theorem B.2 (equiv. to Cor. 1.3). Let L = Lie(G) be the Lie algebra of a compact semisimple Lie group. Let C be a maximal toral subalgebra of L, and let x ∈ L. Then there exists a g ∈ G, such that g.x belongs to C ⊥ (in particular, the G-orbit of any C ⊥ is all of L). Proof. The following is a proof by contradiction. Assume not. Then there is an x ∈ L, necessarily non-zero, such that g.x is not in C ⊥ , for all g ∈ G. But G is compact, so that G.x is also compact (being the continuous image of G), so there is a y ∈ G.x with the property that its C-component has minimal norm, among all points in G.x. By our assumption, the norm of the C-component of y is positive. In order to obtain a contradiction, we shall exhibit a y 0 ∈ G.x whose C-component has norm which is strictly less than that of y. This will follow by an easy application of lemma B.1. Indeed, since y is not in C ⊥ , it follows that there is a positive root a ∈ ∆ such that the C-component of y is not in the kernel of a. Apply lemma B.1 to y and such a positive root a. Thus, there exists an y 0 = g.y ∈ G.y = G.x, for some g ∈ G, such that (i) y 0 has no Ker(a)⊥ component (ii) y and y 0 have the same Ker(a) components (iii) the ma -components of y and y 0 have the same norms Writing y = y1 + y2 + y3 y 0 = y10 + y20 + y30 where y1 , y10 are elements of Ker(a)⊥ , y2 , y20 are in Ker(a), and y3 , y30 are elements of C ⊥ , we have that y10 = 0 and that y20 = y2 . |y10 + y20 |2 = |y2 |2 < |y1 |2 + |y2 |2 = |y1 + y2 |2 The previous inequality is strict since y1 is non-zero, by our choice of positive root a. Since y1 + y2 is the C-component of y (and similarly y10 + y20 is the C-component of y 0 ), this contradicts the property of y having minimal C-component norm among all points in G.x, thus finishing the proof.  References [1] D. Akhiezer, On the commutator map for real semisimple Lie algebras, Moscow Mathematical Journal, Vol. 15, No. 4 (2015), pp. 609 - 613. [2] G. M. Bergman and N. Nahlus, Homomorphisms on infinite direct product algebras, especially Lie algebras. J. Algebra 333 (2011), 67-104. 20 J. MALKOUN AND N. NAHLUS [3] N. Bourbaki, Lie groups and Lie algebras. Chapters 1-3. Translated from the French. Reprint of the 1975 English translation. Elements of Mathematics (Berlin). Springer-Verlag, Berlin, 1989, xviii+450 pp. [4] N. Bourbaki, Lie groups and Lie algebras. Chapters 4-6. Translated from the 1968 French original by Andrew Pressley. Elements of Mathematics (Berlin). Springer-Verlag, Berlin, 2002, xii+300 pp. [5] N. Bourbaki, Lie groups and Lie algebras. Chapters 7-9. Translated from the 1975 and 1982 French originals by Andrew Pressley. Elements of Mathematics (Berlin). Springer-Verlag, Berlin, 2005, xii+434 pp. [6] G. Brown, On commutators in a simple Lie algebra. Proc. Amer. Math. Soc. 14 1963, 763-767. [7] A. D’Andrea and A. Maffei, Small commutators in compact semisimple Lie groups and Lie algebras, J. Lie theory 26 (2016), no. 1, 683-690. [8] D. Z. ¯Doković and T.-Y. Tam, Some questions about semisimple Lie groups originating in matrix theory. Canad. Math. Bull. 46 (2003), no. 3, 332-343. [9] L. A. Florit and W. Ziller, Topological obstructions to fatness. Geom. Topol. 15 (2011), no.2, 891-925. [10] M. Gotô, A theorem on compact semi-simple groups. J. Math. Soc. Japan 1 (1949), 270-272. [11] J. Hilgert and K.-H. Neeb, Structure and geometry of Lie groups. Springer Monographs in Mathematics. Springer, New York, 2012. x+744 pp. [12] K. H. Hofmann and S. A. Morris, The Structure of Compact Groups. A primer for the student–a handbook for the expert, Second revised and augmented edition, De Gruyter Studies in Mathematics, 25, Walter de Gruyter and Co., Berlin, 2006. and , The Lie theory of connected pro-Lie groups, a structure [13] theory for pro-Lie algebras, pro-Lie groups, and connected locally compact groups, EMS tracts in Mathematics, 2, European Mathematical Society (EMS), Zürich, 2007. [14] J. E. Humphreys, Reflection groups and Coxeter groups. Cambridge Studies in Advanced Mathematics, 29. Cambridge University Press, Cambridge, 1990. xii + 204 pp. [15] A. W. Knapp, Lie groups beyond an introduction, Second edition, Progress in Mathematics, 140, Birkhäuser Boston, Inc., Boston, MA, 2002. [16] B. Kostant, On convexity, the Weyl group and the Iwasawa decomposition. Ann. Sci. École Norm. Sup. (4) 6 (1973), 413-455 (1974). [17] I. Kra and S. Simanca, On circulant matrices. Notices Amer. Math. Soc. 59 (2012), no. 3, 368-377. [18] E. Lerman, How fat is a fat bundle? Lett. Math. Phys. 15 (1988), no.4, 335339. [19] N. Nahlus, On 1.5 generators and L = [L, a] + [L, b] in simple Lie algebras, in preparation, tentative title. [20] V. S. Varadarajan, Lie groups, Lie algebras, and their representations. Prentice-Hall Series in Modern Analysis. Prentice-Hall, Inc., Englewood Cliffs, N.J., 1974, xiii+430 pp. [21] N. J. Wildberger, Diagonalization in compact Lie algebras and a new proof of a theorem of Kostant. Proc. Amer. Math. Soc. 119 (1993), no. 2, 649-655. COMMUTATORS AND CARTAN SUBALGEBRAS 21 [22] A. Weinstein, Fat bundles and symplectic manifolds, Adv. Math. 37 (1980), 239–250 [23] W. Ziller, Lie groups, representation theory and symmetric spaces, notes from courses taught Fall 2010 at UPenn, and 2012 at IMPA, http://www.math. upenn.edu/~wziller/ Department of Mathematics and Statistics, Notre Dame UniversityLouaize, Zouk Mikael, Lebanon E-mail address: [email protected] Department of Mathematics, Faculty of Arts and Sciences, American University of Beirut, Beirut, Lebanon, E-mail address: [email protected]
4math.GR
Harmonic Bayesian prediction under α-divergence arXiv:1605.05899v4 [math.ST] 28 Jul 2017 Yuzo Maruyama∗ and Toshio Ohnishi† University of Tokyo∗ and Kyushu University† e-mail: [email protected]; [email protected] Abstract: We investigate Bayesian shrinkage methods for constructing predictive distributions. We consider the multivariate normal model with a known covariance matrix and show that the Bayesian predictive density with respect to Stein’s harmonic prior dominates the best invariant Bayesian predictive density, when the dimension is greater than three. Alpha-divergence from the true distribution to a predictive distribution is adopted as a loss function. AMS 2000 subject classifications: Primary 62C20; secondary 62J07. Keywords and phrases: harmonic prior, minimaxity, Bayesian predictive density. 1. Introduction Let X ∼ Nd (µ, vx I) and Y ∼ Nd (µ, vy I) be independent d-dimensional multivariate normal vectors with common unknown mean µ. We assume that d ≥ 3 and that vx and vy are known. Let φ(·, σ 2 ) be the probability density of Nd (0, σ 2 I). Then the probability density of X and that of Y are φ(x − µ, vx ) and φ(y − µ, vy ), respectively. Based on only observing X = x, we consider the problem of obtaining a predictive density p̂(y | x) for Y that is close to the true density φ(y − µ, vy ). In most earlier papers on such prediction problems, a predictive density p̂(y | x) is often evaluated by Z φ(y − µ, vy ) dy, (1.1) DKL {φ(y − µ, vy ) || p̂(y | x)} = φ(y − µ, vy ) log p̂(y | x) Rd which is called the Kullback-Leibler divergence loss (KL-div loss) from φ(y − µ, vy ) to p̂(y | x). The overall quality of the procedure p̂(y | x) for each µ is then summarized by the Kullback-Leibler divergence risk Z RKL {φ(y − µ, vy ) || p̂(y | x)} = DKL {φ(y − µ, vy ) || p̂(y | x)} φ(x − µ, vx )dx. Rd (1.2) Aitchison (1975) showed that the Bayesian solution with respect to a prior π(µ) under KL-div loss given by (1.1) is the Bayesian predictive density Z p̂π (y | x) = φ(y − µ, vy )π(µ | x)dµ, (1.3) Rd 1 Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 2 where π(µ | x) = φ(x − µ, vx )π(µ)/mπ (x, vx ) is the posterior density corresponding to π(µ) and Z φ(x − µ, v)π(µ)dµ mπ (x, v) = (1.4) Rd is the marginal density of X ∼ Nd (µ, vI) under the prior π(µ). For the prediction problems in general, many studies suggest the use of the Bayesian predictive density rather than plug-in densities of the form φ(y − µ̂(x), vy ), where µ̂(x) is an estimated value of µ. Liang and Barron (2004) showed that the Bayesian predictive density with respect to the uniform prior πU (µ) = 1, (1.5) which is given by Z p̂U (y | x) = φ(y − µ, vy )πU (µ | x)dµ = φ(y − x, vx + vy ) (1.6) Rd is best invariant and minimax. Although the best invariant Bayesian predictive density is generally a good default procedure, it has been shown to be inadmissible in some cases. Specifically, Komaki (2001) showed that the Bayesian predictive density with respect to Stein’s (1974) harmonic prior πH (µ) = kµk−(d−2) (1.7) dominates the best invariant Bayesian predictive density p̂U (y | x). George, Liang and Xu (2006) extended Komaki’s (2001) result to general shrinkage priors including Strawderman’s (1971) prior. From a more general viewpoint, the KL-div loss given by (1.1) is in the class of α-divergence loss (α-div loss) introduced by Csiszár (1967) and defined by   Z p̂(y | x) φ(y − µ, vy )dy, (1.8) Dα {φ(y − µ, vy ) || p̂(y | x)} = fα φ(y − µ, vy ) Rd where   2 (1+α)/2  ,  4/(1 − α ) 1 − z fα (z) = z log z,   − log z, |α| < 1, α = 1, α = −1. When α = −1, we have D−1 {φ(y − µ, vy ) || p̂(y | x)} = DKL {φ(y − µ, vy ) || p̂(y | x)} , where DKL is given by (1.1). When α = 0, we have f0 (z) = 4(1 − z 1/2 ), Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence Z D0 {φ(y − µ, vy ) || p̂(y | x)} = 2 3 n o2 p̂1/2 (y | x) − φ1/2 (y − µ, vy ) dy, Rd p where D0 {φ(y − µ, vy ) || p̂(y | x)}/2 is the Hellinger distance between p̂(y | x) and φ(y − µ, vy ). As in the Kullback-Leibler divergence risk given by (1.2), the overall quality of the procedure p̂(y | x) for each µ is summarized by the α-divergence risk Z Rα {φ(y − µ, vy ) || p̂(y | x)} = Dα {φ(y − µ, vy ) || p̂(y | x)} φ(x − µ, vx )dx. Rd Corcuera and Giummolè (1999) showed that a Bayesian predictive density under α-div loss is Z 2  1−α 1−α    , −1 ≤ α < 1, φ 2 (y − µ, vy )φ(x − µ, vx )π(µ)dµ d RZ  p̂π (y | x; α) ∝   exp {log φ(y − µ, vy )} φ(x − µ, vx )π(µ)dµ , α = 1. Rd (1.9) By (1.9), in the prediction problem under α-div loss with α = 1 from the Bayesian point of view, the Bayesian solution is the normal density p̂π (y | x; 1) = φ(y − µ̂π (x), vy ), where µ̂π (x) is the posterior mean given by Z µ̂π (x) = µπ(µ | x)dµ = x + vx ∇x log m(x, vx ). (1.10) (1.11) Rd In general, the Bayesian prediction problem under α = 1 reduces to the estimation problem under the KL-div loss in the case of the exponential family density. This is because the exponential family density is closed under the calculation in (1.9) with α = 1, as pointed out in Yanagimoto and Ohnishi (2009). As demonstrated in Maruyama and Strawderman (2012), the α-div loss in the case of α = 1 is written as D1 {φ(y − µ, vy ) || φ(y − µ̂π (x), vy )} = kµ̂π (x) − µk2 , 2vy and hence the prediction problem under α = 1 reduces to the estimation problem of µ under the quadratic loss. Stein (1981) showed that " # 1/2   ∆x mπ (X, vx ) 2 2 , (1.12) EX kµ̂π (X) − µk = dvx + 4vx EX 1/2 mπ (X, vx ) which implies that the risk difference under α = 1 is expressed as R1 {φ(y − µ, vy ) || p̂U (y | x; 1)} − R1 {φ(y − µ, vy ) || p̂π (y | x; 1)} " # 1/2 2vx2 ∆x mπ (X, vx ) = EX − . 1/2 vy mπ (X, vx ) (1.13) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 4 Under the KL-div loss or α-div loss with α = −1, George, Liang and Xu (2006) showed that the risk difference is given by R−1 {φ(y − µ, vy ) || p̂U (y | x; −1)} − R−1 {φ(y − µ, vy ) || p̂π (y | x; −1)} " # Z vx 1/2 ∆z mπ (Z, v) EZ − =2 dv, 1/2 v∗ mπ (Z, v) (1.14) where p̂U (y | x; −1) is given by (1.6), Z ∼ Nd (µ, vI) and v∗ = vx vy /(vx + vy ). From this viewpoint, George, Liang and Xu (2006) and Brown, George and Xu (2008) considered the prediction problem under α-div loss with two extreme cases α = ±1 and found a beautiful relationship of risk differences for two cases via ∆z {mπ (z, v)}1/2 for some v. Under both risks R1 and R−1 , any shrinkage prior of the satisfier of the superharmonicity ( ∀v ∈ (v∗ , vx ) for α = −1, 1/2 ∆z mπ (z, v) ≤ 0 for (1.15) v = vx for α = 1, implies the improvement over the best invariant Bayesian procedure. It is wellknown that the superharmonicity of π(µ), ∆µ π(µ) ≤ 0, implies the superharmonicity of mπ (z, v), ∆z mπ (z, v) ≤ 0. Further the superharmonicity of mπ (z, v) implies the superharmonicity of {mπ (z, v)}1/2 . Hence the harmonic prior πH (µ) = kµk−(d−2) is one of the satisfiers of the superharmonicity of {mπ (z, v)}1/2 . Because of the relationship given by (1.13), (1.14) and (1.15), it is of great interest to find the corresponding link via ∆z {mπ (z, v)}1/2 for α-div loss with general α ∈ (−1, 1). To our knowledge, decision-theoretic properties seem to depend on the general structure of the problem (the general type of problem (location, scale), and the dimension of the parameter space) and on the prior in a Bayesian-setup, but not on the loss function, as Brown (1979) pointed out in the estimation problem. In this paper, we investigate the risk difference, diffRα,U,π , in the case of α-div loss, defined by diffRα,U,π = Rα {φ(y − µ, vy ) || p̂U (y | x; α)} − Rα {φ(y − µ, vy ) || p̂π (y | x; α)} . (1.16) In (1.16), p̂π (y | x; α) is given by (1.9) and p̂U (y | x; α) is the Bayesian predictive density under the uniform prior (1.5), the form of which will be derived in (2.5) of Section 2. As a generalization of Liang and Barron’s (2004) result, p̂U (y | x; α) for general α ∈ (−1, 1) is best invariant and minimax, as shown in Appendix A. Further, analyzing diffRα,U,π , we provide some asymptotic results and a nonasymptotic decision-theoretic result. Asymptotic results We show not only somewhat expected relationship lim diffRα,U,π = diffR1,U,π , α→1−0 lim α→−1+0 diffRα,U,π = diffR−1,U,π , (1.17) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 5 where diffR1,U,π and diffR−1,U,π are given in (1.13) and (1.14) respectively, but also the asymptotic relationship for general α ∈ (−1, 1), lim vx /vy →+0 diffRα,U,π = diffR1,U,π . (1.18) Hence, the asymptotic situation vx /vy → 0 corresponds to the case α → 1 and ∆z {mπ (z, v)}1/2 plays an important role for general α ∈ (−1, 1). Non-asymptotic result We particularly investigate a decision-theoretic property of the Bayesian predictive density with respect to πH (µ) = kµk−(d−2) under α-div loss with general α ∈ (−1, 1). We show that, the Bayesian predictive density with respect to πH (µ) = kµk−(d−2) dominates the best invariant Bayesian predictive density with respect to πU (µ) = 1 if  d+2 2   if is a positive integer,  vx d(1 + α) 1 − α 2 ≤  2 d + 2 1 − {κ − 2/(1 − α)}  vy   otherwise, 1−α d 2κ(κ − 1) where κ is the smallest integer larger than 2/(1 − α). The organization of this paper is as follows. In Section 2, we derive the exact form of p̂π (y | x; α), propose a general sufficient condition for diffRα,U,π ≥ 0, where diffRα,U,π is given by (1.16), and demonstrate the asymptotic relationship described in (1.17) and (1.18). In Section 3, we propose the non-asymptotic result under the harmonic prior πH (µ) = kµk−(d−2) described in the above. Some technical proofs are given in Sections A and B of Appendix. 2. Bayesian predictive density under α-divergence loss As in (1.9), the Bayes predictive density under α-div loss is Z p̂π (y | x; α) ∝ 1/β φ (x − µ, vx ) φβ (y − µ, vy )π(µ)dµ , (2.1) Rd where 1−α . 2 Clearly, it follows from α ∈ (−1, 1) that 0 < β < 1. Let β= γ= 1 . 1 + βvx /vy (2.2) (2.3) Since the relation of completing squares with respect to µ, for φ (x − µ, vx ) φβ (y− µ, vy ), is given by 1 β kx − µk2 + ky − µk2 vx vy Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence = = = = 6   1 1−γ 2 2 kx − µk + ky − µk vx γ   kγx + (1 − γ)yk2 1−γ 1 1 2 2 2 kµ − {γx + (1 − γ)y}k − + kxk + kyk vx γ γ γ   1 1 kµ − {γx + (1 − γ)y}k2 + (1 − γ)ky − xk2 vx γ γ 1 kµ − {γx + (1 − γ)y}k2 + β ky − xk2 , vx γ vy we have the identity, φ (x − µ, vx ) φβ (y − µ, vy ) = γ (1−β)d/2 φ(γx + (1 − γ)y − µ, vx γ)φβ (y − x, vy /γ). (2.4) Under the uniform prior πU (µ) = 1, we have, from (2.4), Z φ (x − µ, vx ) φβ (y − µ, vy )πU (µ)dµ = γ (1−β)d/2 φβ (y − x, vy /γ) Rd in (2.1). Therefore the Bayesian predictive density under the uniform prior is p̂U (y | x; α) = φ(y − x, vy /γ) = φ(y − x, vy + βvx ), (2.5) which is the target predictive density so that the risk difference diffRα,U,π = Rα {φ(y − µ, vy ) || p̂U (y | x; α)} − Rα {φ(y − µ, vy ) || p̂π (y | x; α)} is going to be investigated in this paper. As shown in Appendix A, p̂U (y | x; α) for general α ∈ (−1, 1) is best invariant and minimax, which is regarded as a generalization of Liang and Barron’s (2004) minimaxity result. Hence p̂π (y | x; α) with diffRα,U,π ≥ 0 for all µ ∈ Rd is minimax. The exact form of Bayes predictive density p̂π (y | x; α) for (2.1) with normalizing constant, which is regarded as a generalization of Theorem 1 of Komaki (2001) as well as Lemma 2 of George, Liang and Xu (2006), is provided as follows. Theorem 2.1. The Bayes predictive density under π(µ) is 1/β p̂π (y | x; α) = mπ (γx + (1 − γ)y, vx γ) h i p̂U (y | x; α), 1/β EZ1 mπ (x + ξZ1 , vx γ) (2.6) where Z1 ∼ Nd (0, I) and ξ = (1 − γ)(vy /γ)1/2 . (2.7) Proof. By (2.1), (2.4) and (2.5), we have p̂π (y | x; α) ∝ φ(y − x, vy /γ)m1/β π (γx + (1 − γ)y, vx γ). (2.8) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 7 The normalizing constant of (2.8) is Z φ(y − x, vy /γ)m1/β π (γx + (1 − γ)y, vx γ)dy Rd Z   x + (1 − γ)(vy /γ)1/2 z1 , vx γ dz1 = φ(z1 , 1)m1/β π Rd h i = EZ1 m1/β π (x + ξZ1 , vx γ) , where the first equality is from the transformation, z1 = (γ/vy )1/2 (y − x). In the following, as a generalization of the Bayes predictive density, we consider f (γx + (1 − γ)y) p̂f (y | x; α) = p̂U (y | x; α) (2.9) EZ1 [f (x + ξZ1 )] R where f : Rd → R+ is general. As in the proof of Theorem 2.1, p̂f (y | x; α)dy = 1 follows. Also p̂f (y | x; α) is nonnegative for any y ∈ Rd and hence p̂f (y | x; α) is regarded as a predictive density. By the definition of the α-div loss given by (1.8), the risk difference between p̂U and p̂f is written as diffRα,U,f = Rα {φ(y − µ, vy ) || p̂U (y | x; α)} − Rα {φ(y − µ, vy ) || p̂f (y | x; α)} 1−β  1−β ) Z ( p̂f (y | x; α) p̂U (y | x; α) 1 − = β(1 − β) R2d φ(y − µ, vy ) φ(y − µ, vy ) (2.10) × φ(x − µ, vx )φ(y − µ, vy )dxdy. Then we have a following result. Theorem 2.2. 1. The risk difference diffRα,U,f given by (2.10) is written by E[ρ(W, Z)] where W ∼ Nd (µ, vx γ), Z ∼ Nd (0, I) and Z 4γ (1−β)d/2 ξ −∆w %(w + tz; t; f ) ρ(w, z) = 2 β−1 t dt (2.11) β f (w) 0 %2/β−1 (w + tz; t; f ) where β/2 %(u; t; f ) = {EZ1 [f (tZ1 + u)]} , for Z1 ∼ Nd (0, I). (2.12) 2. A sufficient condition for diffRα,U,f ≥ 0 for ∀µ ∈ Rd is ∆u %(u; t; f ) ≤ 0 ∀u ∈ Rd , 0 ≤ ∀t ≤ ξ. (2.13) Proof. Part 2 easily follows from Part 1 and, in the following, we show Part 1. By (2.4), (2.5), and (2.9), the integrand of (2.10) is rewritten as ( β−1  β−1 ) φ(y − µ, vy ) φ(y − µ, vy ) − φ(y − µ, vy )φ(x − µ, vx ) p̂f (y | x; α) p̂U (y | x; α) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence =γ (1−β)d/2 ( EZ1 [f (x + ξZ1 )] f (γx + (1 − γ)y) β−1 8 ) −1 × φ(γx + (1 − γ)y − µ, vx γ)φ(y − x, vy /γ). By the change of variables, w = γx + (1 − γ)y and z = −(γ/vy )1/2 (y − x), where Jacobian of the matrix below is (γ/vy )d/2 ,      γId (1 − γ)Id w x = , (2.14) z y (γ/vy )1/2 Id −(γ/vy )1/2 Id the risk difference is expressed as # " β−1  f (W + ξ(Z1 + Z)) γ (1−β)d/2 −1 EW,Z EZ1 β(1 − β) f (W )   γ (1−β)d/2 EW f (W )1−β {g(ξ; W ) − g(0; W )} β(1 − β) # " Z ξ γ (1−β)d/2 ∂ = EW f (W )1−β g(t; W )dt , β(1 − β) 0 ∂t = (2.15) where ξ = (1 − γ)(vy /γ)1/2 as in (2.7), W ∼ Nd (µ, vx γI), Z1 ∼ Nd (0, I), Z ∼ Nd (0, I) and h i β−1 g(t; w) = EZ EZ1 [f (w + t{Z1 + Z})] . (2.16) In the following, EZ1 [f ] = EZ1 [f (w + t{Z1 + z})] for notational simplicity. Then we have   ∂ ∂ β−1 g(t; w) = EZ {EZ1 [f ]} ∂t ∂t h i β−2 (2.17) = (β − 1)EZ {EZ1 [f ]} EZ1 [(Z1 + Z)T ∇w f ] h i β−2 = (β − 1)EZ {EZ1 [f ]} (EZ1 [Z1T ∇w f ] + Z T EZ1 [∇w f ]) . In (2.17), we have   1 1 EZ1 [Z1T ∇w f ] = EZ1 Z1T ∇z1 f = EZ1 [∆z1 f ] t t (2.18) = tEZ1 [∆w f ] = t∆w EZ1 [f ] where the second equality follows from the Gauss divergence theorem. Similarly Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 9 we have h i β−2 T (β − 1)EZ {EZ1 [f ]} Z EZ1 [∇w f ]   β−2 T 1 = (β − 1)EZ {EZ1 [f ]} Z EZ1 [∇z f ] t h i 1 β−2 T = (β − 1)EZ {EZ1 [f ]} Z ∇z EZ1 [f ] t h i 1 β−1 = EZ Z T ∇z {EZ1 [f ]} t h i 1 β−1 = EZ ∆z {EZ1 [f ]} t h i = tEZ ∆w {EZ1 [f ]} β−1 (2.19) , where the fourth equality follows from the Gauss divergence theorem. By (2.17), (2.18) and (2.19), we have h i ∂ β−1 β−2 g(t; w) = tEZ ∆w {EZ1 [f ]} + (β − 1) {EZ1 [f ]} ∆w EZ1 [f ] . (2.20) ∂t Recall the formula of Laplacian for a function h(u),   ∆u h(u) a a 2 ∆u h (u) = ah (u) + (a − 1)k∇u log h(u)k , h(u) (2.21) for a 6= 0. Then, in (2.20), we have ∆w {EZ1 [f ]} = = = β−1 (β − 1) 1−β {EZ1 [f ]} 2(β − 1) {EZ1 [f ]} 1−β 2(β − 1) {EZ1 [f ]} 1−β β−2 + (β − 1) {EZ1 [f ]} ∆w EZ1 [f ]   ∆w EZ1 [f ] + (β − 2)k∇w log EZ1 [f ] k2 2 EZ1 [f ]   ∆w EZ1 [f ] 2 + (β/2 − 1) k∇w log EZ1 [f ] k EZ1 [f ] ∆w {EZ1 [f ]} (2.22) β/2 β/2 (β/2) {EZ1 [f ]} β/2 = 4(β − 1) ∆w {EZ1 [f ]} . 1−β/2 β {EZ1 [f ]} By (2.15), (2.20) and (2.22), we completes the proof. Remark 2.1. In the previous version of this article as well as George, Liang and Xu (2006), not only the Stein identity but also the heat equation ∂ 1 φ(u, v) = ∆u φ(u, v), ∂v 2 was efficiently applicable for deriving a nice expression of the risk difference, like Part 1 of Theorem 2.2. It seemed to us that the heat equation was an additional Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 10 necessary tool for investigating the Stein phenomenon of predictive density. But it is not true, the heat equation is no longer necessary. As seen in the proof of Theorem 2.2, only the Stein identity or the Gauss divergence theorem is the key, as in Stein “estimation” problem. The superharmonicity of f implies the superharmonicity of EZ1 [f (tZ1 + u)]. Furthermore, using the relationship (2.21), we see that the superharmonicity of EZ1 [f (tZ1 + u)] implies the superharmonicity of %(u; t; f ) = {EZ1 [f (tZ1 + u)]}β/2 for β ∈ (0, 1). Hence, for Part 2 of Theorem 2.2, we have a following corollary. Corollary 2.1. Suppose f : Rd → R+ is superharmonic. Then the predictive density p̂f (y | x; α) given by (2.9) as p̂f (y | x; α) = f (γx + (1 − γ)y) p̂U (y | x; α), EZ1 [f (x + ξZ1 )] dominates p̂U (y | x; α). In Section 3, we will investigate the properties of the Bayesian predictive density p̂π (y | x; α) where f (u) = {mπ (u, vx γ)}1/β is assumed in Theorem 2.2 and Corollary 2.1. Actually in this case, Corollary 2.1 is not useful since the superharmonicity of {mπ (u, vx γ)}1/β for β ∈ (0, 1) is very restrictive. Recall the relationship given by (2.21). For example, the superharmonicity of mπ (u, vx γ) does not imply the superharmonicity of {mπ (u, vx γ)}1/β . Hence, in Section 3, we will seriously consider the superharmonicity of n h ioβ/2 1/β %(u; t; m1/β . π ) = EZ1 {mπ (tZ1 + u, vx γ)}   Further, When 1/β = 2/(1−α) is not an integer, EZ1 {mπ (tZ1 + u, vx γ)}1/β in Part 2 of Theorem 2.2 is not tractable for our current methodology in Section 3. Thus we propose a variant of Theorem 2.2 with f (u) = {mπ (u, vx γ)}1/β , for a non-integer 1/β as follows. Let κ be the smallest integer among integers which is strictly greater than 1/β, κ = min{n ∈ Z | n > 1/β}. (2.23) Then κ − 1 < 1/β < κ. As in (2.15), the risk difference is expressed as Rα {φ(y − µ, vy ) || p̂U (y | x; α)} − Rα {φ(y − µ, vy ) || p̂π (y | x; α)}   " 1/β #β−1 m (W + ξ(Z + Z), v γ) γ (1−β)d/2 π 1 x EW,Z EZ1 − 1 = β(1 − β) mπ (W, vx γ) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 11 where W ∼ Nd (µ, vx γI), Z1 ∼ Nd (0, I) and Z ∼ Nd (0, I). From Jensen’s inequality, we have h i EZ1 m1/β π (w + ξ(Z1 + Z), vx γ) i h (2.24) = EZ1 {mκπ (w + ξ(Z1 + Z), vx γ)}1/(βκ) 1/(βκ) ≤ {EZ1 [mκπ (w + ξ(Z1 + Z), vx γ)]} , since 0 < 1/(βκ) < 1 and hence Rα {φ(y − µ, vy ) || p̂U (y | x; α)} − Rα {φ(y − µ, vy ) || p̂π (y | x; α)} # " (β−1)/(βκ)  κ γ (1−β)d/2 mπ (W + ξ(Z1 + Z), vx γ) −1 . EW,Z EZ1 ≥ β(1 − β) mκπ (W, vx γ) Applying the same technique starting (2.15) through (2.22) to the lower bound above, we have a variant of Part 2 of Theorem 2.2. Theorem 2.3. Assume 1/β is not a positive integer. Let κ be the smallest integer greater than 1/β. A sufficient condition for diffRα,U,π ≥ 0 is c(β)/κ ∆u {EZ1 [mκπ (tZ1 + u, vx γ)]} ≤ 0, ∀u ∈ Rd , 0 ≤ ∀t ≤ ξ (2.25) where Z1 ∼ Nd (0, I) and c(β) = κ − 1/β + 1 ∈ (1/2, 1). 2 (2.26) 2.1. Asymptotics 1/β In this subsection, using Theorem 2.2 with f = mπ , we investigate asymptotics of the risk difference diffRα,U,π = Rα {φ(y − µ, vy ) || p̂U (y | x; α)} − Rα {φ(y − µ, vy ) || p̂π (y | x; α)} where p̂U (y | x; α) and p̂π (y | x; α) are given by (2.5) and (2.6), respectively. 2.1.1. α → −1 Let v∗ = vx vy /(vx + vy ). When α → −1 or equivalently β → 1, we have γ→ and hence 1 v∗ vx2 = and ξ 2 → = vx − v∗ 1 + vx /vy vx vx + vy 2γ (1−β)d/2 {mπ (w, vx γ)}1/β−1 → 2, β2 (2.27) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 12 which are parts of ρ(w, z) given by (2.11). Further, in %(t; u) given by (2.12), we have EZ1 [mπ (tZ1 + u, vx γ)] = mπ (u, vx γ + t2 ) → mπ (u, v∗ + t2 ). (2.28) By (2.27) and (2.28), we have 2 %(t; u) → m1/2 π (u, v∗ + t ), √ Z vx −v∗ Z 1/2 −∆u mπ (u, v∗ + t2 ) t EZ [ρ(w, Z)] → 4 φ(u − w, t2 )dudt 1/2 2 d (2.29) 0 R mπ (u, v∗ + t ) Z vx −v∗ Z 1/2 −∆u mπ (u, v∗ + t) =2 φ(u − w, t)dudt. 1/2 0 Rd mπ (u, v∗ + t) By (2.29), we have Z Z vx −v∗ Z EW,Z [ρ(W, Z)] → 2 Rd Z =2 Rd 0 ! 1/2 −∆u mπ (u, v∗ + t) 1/2 φ(u − w, t)dudt mπ (u, v∗ + t) × φ(w − µ, v∗ )dw Z 1/2 −∆u mπ (u, v∗ + t) ! vx −v∗ 1/2 mπ (u, v∗ + t) " # Z vx 1/2 ∆z mπ (Z, v) =2 EZ − dv 1/2 v∗ mπ (Z, v) Rd 0 φ(u − µ, v∗ + t)du dt = R−1 {φ(y − µ, vy ) || p̂U (y | x; −1)} − R−1 {φ(y − µ, vy ) || p̂π (y | x; −1)}, where Z ∼ Nd (µ, vI) and v∗ = vx vy /(vx + vy ). The last equality follows from George, Liang and Xu’s (2006) result which was already explained in (1.14) of Section 1. Hence we have lim α→−1+0 diffRα,U,π = diffR−1,U,π . 2.1.2. (1 − α)vx /vy → 0 Consider the asymptotic situation where (1 − α)vx /vy → 0 ⇔ β(vx /vy ) → 0 ⇔ γ → 1. Note that EZ [ρ(w, Z)] is rewritten as the product ρ1 (w)ρ2 (w) where 2γ (1−β)d/2 {mπ (w, vx γ)}1/β−1 ξ 2 , β2  Z ξ Z 2 −∆u %(t; u) 2 ρ2 (w) = 2 t φ(u − w, t )du dt 2/β−1 (t; u) ξ 0 Rd % ρ1 (w) = (2.30) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence = Z 1 ξ2 ξ2 Z Rd 0 13 √  −∆u %( t; u) √ φ(u − w, t)du dt. %2/β−1 ( t; u) Since ξ 2 is rewritten as ξ2 = (1 − γ)2 vy = γ  1−γ γ 2 vy γ = vx2 2 β γ, vy (2.31) we have ρ1 (w) = 2 vx2 (1−β)d/2+1 γ {mπ (w, vx γ)}1/β−1 vy and lim ρ1 (w) = 2 γ→1 vx2 {mπ (w, vx )}1/β−1 . vy (2.32) When γ → 1, we have ξ 2 → 0 by (2.31) and hence √ Z  −∆u %( t; u) √ lim ρ2 (w) = lim φ(u − w, t)du 2/β−1 ( t; u) γ→1 t→0 Rd % √   Z −∆u %( t; u) √ = lim δ(u − w)du, 2/β−1 ( t; u) Rd t→0 % (2.33) where δ(·) is the Dirac delta function. By (2.33) and √ Z lim %( t; u) = t→0 γ→1 Rd m1/β π (u1 β/2 + u, vx γ)δ(u1 )du1 = m1/2 π (u, vx ), we have   1/2−1/β lim ρ2 (w) = −∆w m1/2 (w, vx ). π (w, vx ) mπ γ→1 (2.34) By (2.32) and (2.34), we have 1/2 lim EZ [ρ(w, Z)] = lim ρ1 (w)ρ2 (w) = 2 γ→1 γ→1 vx2 −∆w mπ (w, vx ) , vy m1/2 π (w, vx ) which implies that lim diffRα,U,π α→1 lim vx /vy →0 vy diffRα,U,π vx " # 1/2 vx2 −∆w mπ (W, vx ) = diffR1,U,π = 2 E , 1/2 vy mπ (W, vx ) " # 1/2 vy −∆w mπ (W, vx ) = diffR1,U,π = 2vx E . 1/2 vx mπ (W, vx ) Therefore the asymptotic situation vx /vy → 0 corresponds to the case α → 1 and ∆z {mπ (z, v)}1/2 plays an important role for general α ∈ (−1, 1). Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 14 3. Improvement under the harmonic prior Under the harmonic prior πH (µ) = kµk−(d−2) , let Z mH (w, v) = φ(w − µ, v)πH (µ)dµ. (3.1) Rd Let ν be an integer larger than or equal to 2. The superharmonicity related to EZ1 [mνH (tZ1 + u, v)] with Z1 ∼ Nd (0, I) is as follows. Theorem 3.1. Let c ∈ (0, 1) and Z1 ∼ Nd (0, I). Let ν be an integer larger than or equal to 2. Then, we have c/ν ∆u {EZ1 [mνH (tZ1 + u, v)]} ≤ 0, when  0≤t≤ (d + 2)(1 − c)v dν(ν − 1) ∀u ∈ Rd , 1/2 . (3.2) Proof. Section B of Appendix. When 1/β is an integer larger than or equal to 2, namely, α = 0, 1/3, 1/2, 3/5, 2/3, . . . , β = 1/2, 1/3, 1/4, 1/5, 1/6, . . . , (3.3) let ν = 1/β, v = vx γ and c = 1/2 in Theorem 3.1 and compare (3.2) in Theorem 3.1 with 0 ≤ t2 ≤ ξ 2 = β 2 vx2 γ/vy in Theorem 2.2. If β 2 vx (d + 2)(1 − c) vx γ vx γ ≤ vy dν(ν − 1) or equivalently vx d+2 d+2 = , ≤ vy d(1 + α) 2d(1 − β) mH (w, vx γ) satisfies the sufficient condition of Theorem 2.2 and we have a following result of the Bayesian predictive density with respect to Stein’s harmonic prior πH (µ) = kµk−(d−2) , which is given by 1/β p̂H (y | x; α) = mH (γx + (1 − γ)y, vx γ) i p̂U (y | x; α). h 1/β EZ1 mH (x + ξZ1 , vx γ) (3.4) Theorem 3.2. Suppose 2/(1−α) is an positive integer for α ∈ (−1, 1). Suppose vx d+2 ≤ . vy d(1 + α) (3.5) Then, under α-div loss, the Bayesian predictive density p̂H (y | x; α) with respect to the harmonic prior πH (µ) = kµk−(d−2) dominates the best invariant Bayesian predictive density p̂U (y | x; α) = φ(y − x, vy /γ). Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 15 Remark 3.1. For any d ≥ 3 and α ∈ (−1, 1), we have 1 d+2 > . d(1 + α) 2 Note that, in most typical situations, vx 1 ≤ , vy 2 is easily assumed as follows. Suppose that we have a set of observations x1 , . . . , xn from Nd (µ, σ 2 I). An unobserved set xn+1 , . . . , xn+m from the same distribution is predicted by using a predictive density as a function of x1 , . . . , xn . From sufficiency, Xn Xm x = n−1 xi ∼ Nd (µ, σ 2 I/n) and y = m−1 xn+i ∼ Nd (µ, σ 2 I/m) i=1 i=1 and clearly vx /vy = m/n in this case. Since, m is typically 1 or 2 whereas n is relatively large, the condition (3.5) is satisfied. When β = 2/(1 − α) is not an integer, Theorem 2.3 can be applied. Let κ be the smallest integer greater than 1/β. Suppose β2 (d + 2){1 − c(β)}vx γ vx vx γ ≤ , vy dκ(κ − 1) (3.6) where c(β) is given by (2.26) as c(β) = c({1 − α}/2) = {κ − 2/(1 − α) + 1}/2, the left-hand side is the upper bound of t of Theorem 2.3 and the right-hand side is the upper bound of t of Theorem 3.1. When vx ≤ vy  2 1−α 2 d + 2 1 − {κ − 2/(1 − α)} , d 2κ(κ − 1) which is equivalent to (3.6), mH (w, vx γ) satisfies the sufficient condition of Theorem 2.3 and we have a following result. Theorem 3.3. Suppose 2/(1 − α) is not an positive integer for α ∈ (−1, 1). Let κ be the smallest integer greater than 2/(1 − α). Suppose vx ≤ vy  2 1−α 2 d + 2 1 − {κ − 2/(1 − α)} . d 2κ(κ − 1) (3.7) Then the Bayesian predictive density p̂H (y | x; α) with respect to the harmonic prior πH (µ) = kµk−(d−2) dominates the best invariant Bayesian predictive density p̂U (y | x; α) = φ(y − x, vy /γ). By the definition of κ, κ−1< 2 < κ. 1−α 16 0.8 0.6 0.4 0.0 0.2 upper bound of vx /vy 1.0 1.2 Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence −1.0 −0.5 0.0 0.5 α Fig 1. The upper bound of vx /vy in Theorems 3.2 and 3.3 As 2/(1 − α) ↑ κ, the upper bound given by (3.7) approaches (d + 2)/{d(1 + α)} which is exactly the upper bound given by (3.5) of Theorem 3.2. On the other hand, as 2/(1 − α) ↓ κ − 1, the upper bound given by (3.7) approaches 0. Figure 1 gives a graph of behavior of the upper bound of vx /vy for improvement in Theorems 3.2 and 3.3. This undesirable discontinuity with respect to the upper bound of Theorem 3.3 is due to Jensen’s inequality (2.24) which was not used in the proof of Theorem 2.2. However, we would like to emphasize that, for any α ∈ (−1, 1), there exists a positive upper bound of of vx /vy for improvement. We can naturally make a conjecture that the lower bound of vy /vx for improvement, d(1 + α)/(d + 2), of Theorem 3.2 is still valid even if 2/(1 − α) is not an  integer. For that purpose,  the methodology for appropriately treating EZ1 {mH (tZ1 + u, vx γ)}2/(1−α) or more generally   EZ1 {mπ (tZ1 + u, vx γ)}2/(1−α) for non-integer 2/(1 − α) is needed and it remains an open problem. Appendix A: Minimaxity of p̂U (y | x; α) In this section, we show that p̂U (y | x; α) = φ(y − x, vy /γ) = φ(y − x, vy + βvx ) (A.1) is minimax, by following Sections II and III of Liang and Barron (2004). We start with the definition of invariance under location shift. Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 17 Definition A.1. A predictive density p̂(y | x) is invariant under location shift, if for all a ∈ Rd and all x, y, p̂(y + a | x + a) = p̂(y | x). Hence any invariant predictive density should be of the form p̂(y | x) = q(y − x) which satisfies Z q(y)dy = 1. Rd Clearly p̂U (y | x; α) is invariant under location shift. Note that invariant procedures have constant risk since the risk of the invariant predictive density q(y −x) is Rα {φ(y − µ, vy ) || q(y − x)}    Z Z q(y − x) φ(y − µ, vy )dy φ(x − µ, vx )dx = fα φ(y − µ, vy ) Rd Rd    Z Z q(zy − zx ) = fα φ(zy , vy )dzy φ(zx , vx )dzx φ(zy , vy ) Rd Rd (A.2) where zx = x − µ and zy = y − µ, which does not depend on µ. More specifically, the risk of the invariant predictive density q(y − x) is as follows. Lemma A.1. The risk of an invariant predictive density q(y − x) is Rα {φ(y − µ, vy ) || q(y − x)} = 1 − γ (1−β)d/2 + γ (1−β)d/2 Dα {φ(z, vy /γ) || q(z)} . β(1 − β) (A.3) Proof. By (A.2) and the definition of α-div loss, Rα {φ(y − µ, vy ) || q(y − x)}   Z Z 1 1− q 1−β (y − x)φβ (y, vy )φ(x, vx )dxdy . = β(1 − β) Rd Rd By the identity (2.4) with µ = 0, we have φ (x, vx ) φβ (y, vy ) = γ (1−β)d/2 φ(γx + (1 − γ)y, vx γ)φβ (y − x, vy /γ), and hence 1 Rα {φ(y − µ, vy ) || q(y − x)} = β(1 − β)  1−γ (1−β)d/2 Z Rd Z q 1−β (y − x) Rd × φβ (y − x, vy /γ)φ(γx + (1 − γ)y, vx γ)dxdy . By the change of variables,    w γId = z −Id (1 − γ)Id Id   x y (A.4) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 18 where Jacobian of the matrix is 1, we have Rα {φ(y − µ, vy ) || q(y − x)}   Z Z 1 = 1 − γ (1−β)d/2 q 1−β (z)φβ (z, vy /γ)φ(w, vx γ)dzdw β(1 − β) Rd Rd   Z 1 1 − γ (1−β)d/2 q 1−β (z)φβ (z, vy /γ)dz = β(1 − β) Rd = 1 − γ (1−β)d/2 + γ (1−β)d/2 Dα {φ(z, vy /γ) || q(z)} . β(1 − β) In (A.3) of Lemma A.1, Dα {φ(z, vy /γ) || q(z)} is non-negative and takes zero if and only if q(z) = φ(z, vy /γ). Hence the best invariant procedure is p̂U (y | x; α) = φ(y − x, vy /γ), where the constant risk is 1 − γ (1−β)d/2 . β(1 − β) Since the risk is constant for invariant predictive density, the best invariant p̂U (y | x; α) is the minimax procedure among all invariant procedures. If a constant risk procedure is shown to have an extended Bayes property defined in the below, then it is, in fact, minimax over all procedures. See Theorem 5.18 of Berger (1985) and Theorem 5.1.12 of Lehmann and Casella (1998) for the detail. Definition A.2. A predictive procedure p̂∗ (y | x) is called extended Bayes, if there exists a sequence of Bayes procedures p̂πc (y | x; α) with proper prior densities πc (µ) for c = 1, . . . , such chat their Bayes risk differences go to zero, that is, Z lim Rα {φ(y − µ, vy ) || p̂∗ (y | x)}πc (µ)dµ c→∞ Rd  Z − Rα {φ(y − µ, vy ) || p̂πc (y | x; α)}πc (µ)dµ = 0. Rd Recall that Z p̂π (y | x; α) ∝ 1/β φβ (y − µ, vy )φ(x − µ, vx )π(µ)dµ (A.5) Rd for β = (1 − α)/2 and α ∈ (−1, 1). Under the prior µ ∼ Nd (0, {cvx γ}I) with the density πc (µ) = φ(µ, cvx γ), the Bayesian solution is   cγ 1+c p̂πc (y | x; α) = φ y − x, vy 1 + cγ 1 + cγ Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 19 by the identity φβ (y − µ, vy )φ (x − µ, vx ) φ(µ, cvx γ) d(1−β)/2    γx + (1 − γ)y cvx γ 1 + cγ φ µ−c , = 1+c 1+c 1+c   cγx 1 + c × φβ y − , vy φ (x, vx (1 + cγ)) . 1 + cγ 1 + cγ (A.6) Furthermore, by the identity (A.6), the Bayes risk of p̂πc (y | x; α) is given by  1−β Z Z Z  1 p̂πc (y | x; α) 1− β(1 − β) φ(y − µ, vy ) Rd Rd Rd  × φ (x − µ, vx ) φ(y − µ, vy )φ(µ, cvx γ)dxdydµ (  d(1−β)/2 Z Z Z   1 + cγ γx + (1 − γ)y cvx γ 1 1− , φ µ−c = β(1 − β) 1+c 1+c 1+c Rd Rd Rd    cγx 1+c × φ y− , vy φ (x, vx (1 + cγ)) dµdydx 1 + cγ 1 + cγ (  d(1−β)/2 ) 1 1 + cγ = 1− , β(1 − β) 1+c which approaches (1 − γ (1−β)d/2 )/{β(1 − β)} as c goes to infinity, the constant risk of p̂U (y | x; α). Hence p̂U (y | x; α) is extended Bayes and hence minimax. Appendix B: Proof of Theorem 3.1 Recall the identity kµk−(d−2) = b Z 0 ∞   kµk2 g d/2−2 exp −g dg 2v (B.1) for any v > 0, where b = 1/{Γ(d/2 − 1)2d/2−1 v d/2−1 }. Then we have Z mH (w, v) = φ(w − µ, v)kµk−(d−2) dµ d R   Z ∞ Z kµk2 1 kw − µk2 exp − − g dµ =b g d/2−2 dg d/2 v d/2 2v 2v 0 Rd (2π)   Z ∞ g d/2−2 gkwk2 =b exp − dg 2(g + 1)v (1 + g)d/2 0   Z 1 λkwk2 =b λd/2−2 exp − dλ, 2v 0 Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 20 where the third equality is from the relation of completing squares with respect to µ kw − µk2 + gkµk2 = (g + 1)kµ − w/(g + 1)k2 + {g/(g + 1)}kwk2 and the fourth equality is from the transformation λ = g/(g + 1). Note that mνH (w, v) for a positive integer ν is expressed as ν mH (w, v) = b Z ν ν Y Dν i=1 d/2−2 λi Y  Pν 2 i=1 λi kwk dλi , exp − 2v Qν where Dν is ν-dimensional unit hyper-cube. In the following, dλ denotes Q i=1 dλi for notational simplicity. Furthermore the subscript and superscript of and P is omitted for simplicity if they are i = 1 and i = ν respectively. Hence mνH (w, v) in the above is written as  P  Z Y λi kwk2 d/2−2 ν ν λi exp − mH (w, v) = b dλ. 2v Dν For the calculation of EZ1 [mνH (tZ1 + u, v)] = Z mνH (x + u, v)φ(x, t2 )dx (B.2) Rd under Z1 ∼ Nd (0, I), note the relation of completing squares with respect to x, P o kxk2 1 nX ( λi ) kx + uk2 + 2 = λi kx + uk2 + skxk2 t v ) (v P P (B.3) 2  s λi 1 X λi 2 = u +P kuk , λi + s x + P λi + s λi + s v where s = v/t2 . Then, by (B.3), we have EZ1 [mνH (tZ1 + u, v)] = bν v d/2 td Z Dν Q d/2−2 P   λ s λi kuk2 P P i exp − dλ. v( λi + s) 2 ( λi + s)d/2 Re-define u := {s/v}1/2 u and let Z ψ(u; ν, s) = Dν Q d/2−2 P   λi λi kuk2 P P exp − dλ. λi + s 2 ( λi + s)d/2 c/ν By (2.21), the super-harmonicity of {EZ1 [mνH (tZ1 + u, v)]} u ∈ Rd is equivalent to c  − 1 k∇u ψk2 + ψ∆u ψ ≤ 0, ∀u ∈ Rd . ν (B.4) with respect to (B.5) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 21 The integrand of ψ given by (B.4) is denoted by P Q d/2−2   λi λ P exp − z ζ(λ) = ζ(λ1 , . . . , λν ) = P i λi + s ( λi + s)d/2 where z = kuk2 /2. Then we have ∂ ψ = −uj ∂uj Z P λi dλ, ζ(λ) P λi + s for j = 1, . . . , d and ∂2 ψ= ∂u2j Z ( P  P 2 ) λi λ i ζ(λ) − P + u2j P dλ. λi + s λi + s Noting z = kuk2 /2, we have 2 Z k∇u ψk = 2z P 2 Z 2 λ1 λi 2 dλ = 2ν z ζ(λ) P dλ ζ(λ) P λi + s λi + s (B.6) and P  P 2 Z λi λi ∆u ψ = −d ζ(λ) P dλ + 2z ζ(λ) P dλ λi + s λi + s Z Z λ1 λ2 dλ = −dν ζ(λ) P dλ + 2νz ζ(λ) P 1 ( λi + s)2 λi + s Z λ1 λ2 + 2ν(ν − 1)z ζ(λ) P dλ. ( λi + s)2 Z (B.7) In (B.6) and (B.7), the second equalities are from symmetry with respect to λi ’s. Let Z X ρ(j1 , j2 , l) = λj11 λj22 ( λi + s)l ζ(λ)dλ, Dν Z  l X Y η(j2 , l) = λj22 1 + λi + s ζ(1, λ2 , . . . , λν ) dλi , Dν−1 i=2 i=2 where j1 and j2 are nonnegative integers. Then k∇u ψk2 and ∆u ψ given by (B.6) and (B.7) is rewritten as k∇u ψk2 = 2ν 2 zρ(1, 0, −1)2 , ∆u ψ = −dνρ(1, 0, −1) + 2νzρ(2, 0, −2) + 2ν(ν − 1)zρ(1, 1, −2). Here are some useful relationships and inequalities. (B.8) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 22 Lemma B.1. szρ(j1 , j2 , l) = −η(j2 , l + 2) + (j1 + d/2 − 2)ρ(j1 − 1, j2 , l + 2) + (l − d/2 + 2)ρ(j1 , j2 , l + 1), for j1 ≥ 1, ρ(0, 0, l) = νρ(1, 0, l − 1) + sρ(0, 0, l − 1), (B.9) (B.10) ρ(1, 0, l) = ρ(2, 0, l − 1) + (ν − 1)ρ(1, 1, l − 1) + sρ(1, 0, l − 1), (B.11) η(0, 1) = η(0, 0) + (ν − 1)η(1, 0) + sη(0, 0), η(0, 1)ρ(0, 0, −1) ≥ η(0, 0)ρ(0, 0, 0), ρ(1, 0, −1) 1 ≥ . ρ(1, 0, 0) νd/(d + 2) + s (B.12) (B.13) (B.14) Proof. See Sub-section B.1. Applying the identity (B.9) to k∇u ψk2 and ∆u ψ given in (B.8), we have sk∇u ψk2 = 2ν 2 {szρ(1, 0, −1)}ρ(1, 0, −1) = ν 2 {−2η(0, 1) + (d − 2)ρ(0, 0, 1) − (d − 2)ρ(1, 0, 0)} ρ(1, 0, −1), s∆u ψ = −dνsρ(1, 0, −1) + ν{−2η(0, 0) + dρ(1, 0, 0) − dρ(2, 0, −1)} + ν(ν − 1) {−2η(1, 0) + (d − 2)ρ(1, 0, 0) − dρ(1, 1, −1)} = −2ν{η(0, 0) + (ν − 1)η(1, 0)} + ν(ν − 1)(d − 2)ρ(1, 0, 0), where the second equality of s∆u ψ follows from (B.11). Then we have   s c−ν 2 k∇u ψk + ψ∆u ψ ν ν (B.15) = (ν − c) [2η(0, 1) − (d − 2){ρ(0, 0, 1) − ρ(1, 0, 0)}] ρ(1, 0, −1) − 2{η(0, 0) + (ν − 1)η(1, 0)}ρ(0, 0, 0) + (ν − 1)(d − 2)ρ(1, 0, 0)ρ(0, 0, 0). By applying (B.10), (B.12) and (B.13), the terms of (B.15) including η(·, ·), divided by 2, is (ν − c)η(0, 1)ρ(1, 0, −1) − {η(0, 0) + (ν − 1)η(1, 0)}ρ(0, 0, 0) = (ν − c)η(0, 1)ρ(1, 0, −1) − {η(0, 1) − sη(0, 0)}ρ(0, 0, 0) = (ν − c)η(0, 1)ρ(1, 0, −1) − η(0, 1) {νρ(1, 0, −1) + sρ(0, 0, −1)} + sη(0, 0)ρ(0, 0, 0) (B.16) = −cη(0, 1)ρ(1, 0, −1) − s{η(0, 1)ρ(0, 0, −1) − η(0, 0)ρ(0, 0, 0)} ≤ 0, where the first equality follows from (B.12), the second equality follows from (B.10) and the inequality follows from (B.13). Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 23 The terms of (B.15) not including η(·, ·), divided by (d − 2), are rewritten as (ν − c) {−ρ(0, 0, 1) + ρ(1, 0, 0)} ρ(1, 0, −1) + (ν − 1)ρ(1, 0, 0)ρ(0, 0, 0) = −(ν − c)(ν − 1)ρ(1, 0, 0)ρ(1, 0, −1) − (ν − c)sρ(0, 0, 0)ρ(1, 0, −1) + (ν − 1)ρ(1, 0, 0)ρ(0, 0, 0)   (ν − c)s ≤− − (ν − 1) ρ(1, 0, 0)ρ(0, 0, 0) νd/(d + 2) + s (1 − c)s − ν(ν − 1)d/(d + 2) =− ρ(1, 0, 0)ρ(0, 0, 0), νd/(d + 2) + s (B.17) which is nonpositive for s ≥ ν(ν − 1)d/{(1 − c)(d + 2)}, where the first equality follows from (B.10) and the inequality follows from (B.14). By (B.16) and (B.17), we have  c − 1 k∇u ψk2 + ψ∆u ψ ≤ 0, ∀u ∈ Rd ν or equivalently ∆u {EZ1 [mνH (tZ1 + u, v)]} c/ν ≤ 0, ∀u ∈ Rd , when t ≤ {(d + 2)(1 − c)v/{dν(ν − 1)}}1/2 . B.1. Proof of Lemma B.1 [Part of (B.9)] Note P P     ∂ sz z λi z λi P =− P exp − . exp − P λi + s λi + s ∂λ1 ( λi + s)2 (B.18) Then, by an integration by parts, we have Z 1 X sz λj11 λj22 ( λi + s)l ζ(λ)dλ1 0 d/2−2+j2 = −λ2 Y i=3 d/2−2 Z λi 1 d/2−2+j1 λ1 ( X λi + s)l−d/2+2 0 P   z λi ∂ exp − P dλ1 × ∂λ1 λi + s ( P  1 X Y d/2−2 z λi d/2−2+j2 d/2−2+j1 = −λ2 λi λ1 ( λi + s)l−d/2+2 exp − P λi + s 0 i=3 P   Z 1 X z λi d/2−3+j1 l−d/2+2 P −(d/2 − 2 + j1 ) λ1 ( λi + s) exp − dλ1 λi + s 0 P    Z 1 X z λi d/2−2+j1 l−d/2+1 dλ1 . −(l − d/2 + 2) λ1 ( λi + s) exp − P λi + s 0  Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 24 (B.9) follows from integration with respect to λ2 , . . . , λν in the both hand side of the above equality. [Parts of (B.10), (B.11) and (B.12)] The equalities (B.10), (B.11) and (B.12) easily follows from symmetry with respect to λi ’s. [Part of (B.13)] Note that (B.13) is equivalent to η(0, 0)ρ(0, 0, 0) − η(0, 1)ρ(0, 0, −1) = {ρ(0, 0, 0) − ρ(0, 0, −1)}η(0, 1) − {η(0, 1) − η(0, 0)}ρ(0, 0, 0) Z Z Y Y = f1 (λ2 , . . . , λν ) dλi f2 (λ2 , . . . , λν ) dλi Dν−1 − Dν−1 i=2 Z f3 (λ2 , . . . , λν ) Dν−1 Y i=2 i=2 Z dλi f4 (λ2 , . . . , λν ) Dν−1 Y dλi i=2 ≤ 0, where 1   1 ζ(λ1 , . . . , λν )dλ1 1− P λi + s 0 X f2 (λ2 , . . . , λν ) = (1 + λi + s)ζ(1, λ2 , . . . , λν ) i=2 X f3 (λ2 , . . . , λν ) = ( λi + s)ζ(1, λ2 , . . . , λν ) i=2 Z 1 f4 (λ2 , . . . , λν ) = ζ(λ1 , . . . , λν )dλ1 . Z f1 (λ2 , . . . , λν ) = 0 P P Since both 1−1/ ( λi + s) and λi +s are increasing in each of its arguments, we have o n X X λi + s (1 + ξi + s) 1 − 1/ i=2  n o X 1 P (λ1 ∨ 1) + (λi ∨ ξi ) + s ≤ 1− i=2 (λ1 ∨ 1) + i=2 (λi ∨ ξi ) + s X = (λi ∨ ξi ) + s, (B.19) i=2 where ∨ is the maximum operator, i.e. λi ∨ ξi = max(λi , ξi ). In the following, ∧ denotes the minimum operator, i.e. λi ∧ ξi = min(λi , ξi ). Note that a function h: Rν → R is said to be multivariate totally positive of order two (MTP2) if it satisfies h(x1 , . . . , xν )h(y1 , . . . , yν ) ≤ h(x1 ∨ y1 , . . . , xν ∨ yν )h(x1 ∧ y1 , . . . , xν ∧ yν ) for any x, y ∈ Rν . By Lemma B.2 below, ζ(λ1 , . . . , λν ) is MTP2 as a function of ν-variate function and hence the inequality ζ(λ1 , . . . , λν )ζ(1, ξ2 , . . . , ξν ) ≤ ζ(λ1 ∨ 1, λ2 ∨ ξ2 , . . . , λν ∨ ξν )ζ(λ1 ∧ 1, λ2 ∧ ξ2 , . . . , λν ∧ ξν ) = ζ(1, λ2 ∨ ξ2 , . . . , λν ∨ ξν )ζ(λ1 , λ2 ∧ ξ2 , . . . , λν ∧ ξν ) (B.20) Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 25 follows. By (B.19) and (B.20), we have f1 (λ2 , . . . λν )f2 (ξ2 , . . . ξν ) Z 1 h nX o ≤ (λi ∨ ξi ) + s ζ(1, λ2 ∨ ξ2 , . . . , λν ∨ ξν ) i=2 0 i × ζ(λ1 , λ2 ∧ ξ2 , . . . , λν ∧ ξν ) dλ1 (B.21) = f3 (λ2 ∨ ξ2 , . . . , λν ∨ ξν )f4 (λ2 ∧ ξ2 , . . . , λν ∧ ξν ). From Theorem B.1 below, shown by Karlin and Rinott (1980), the theorem follows. [Part of (B.14)] By Jensen’s inequality, we have Z 1 λ1 ζ(λ) ρ(1, 0, −1) Pν = dλ ρ(1, 0, 0) λ1 + i=2 λi + s ρ(1, 0, 0) 1 . ≥ ρ(2, 0, 0) ρ(1, 1, 0) + (ν − 1) +s ρ(1, 0, 0) ρ(1, 0, 0) (B.22) Let f be a probability density given by ν−1  ν Y d d d/2−2 d/2−1 λi , f (λ1 , . . . , λν ) = −1 λ1 2 2 i=2 which is clearly MTP2. Also let g1 (λ1 , . . . , λν ) = λ1 , g2 (λ1 , . . . , λν ) = − P exp (sz/{ λi + s}) P , ( λi + s)d/2 which are both increasing increasing in each of its arguments. Hence, by so-called FKG inequality given in Theorem B.2 below, Z g1 (λ1 , . . . , λν )g2 (λ1 , . . . , λν )f (λ1 , . . . , λν )dλ Dν Z Z ≥ g1 (λ1 , . . . , λν )f (λ1 , . . . , λν )dλ g2 (λ1 , . . . , λν )f (λ1 , . . . , λν )dλ Dν Dν or equivalently R g1 (λ1 , . . . , λν )g2 (λ1 , . . . , λν )f (λ1 , . . . , λν )dλ R g (λ , . . . , λν )f (λ1 , . . . , λν )dλ Dν 2 1 Z ≤ g1 (λ1 , . . . , λν )f (λ1 , . . . , λν )dλ, Dν Dν since g2 < 0. Since ρ(2, 0, 0)/ρ(1, 0, 0) is expressed as R g1 (λ1 , . . . , λν )g2 (λ1 , . . . , λν )f (λ1 , . . . , λν )dλ ρ(2, 0, 0) R = Dν , ρ(1, 0, 0) g (λ , . . . , λν )f (λ1 , . . . , λν )dλ Dν 2 1 Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence we have 26 ρ(2, 0, 0) d ≤ . ρ(1, 0, 0) d+2 (B.23) d−2 d ρ(1, 1, 0) ≤ ≤ . ρ(1, 0, 0) d d+2 (B.24) Similarly we have Hence, by (B.22), (B.23) and (B.24), we have ρ(1, 0, −1) 1 ≥ . ρ(1, 0, 0) νd/(d + 2) + s Lemma B.2. Let Q d/2−2 P   λi λi exp − P z . ζ(λ1 , . . . , λν ) = P λi + s ( λi + s)d/2 Then ζ(λ1 , . . . , λν ) is MTP2. Proof. Note P     sz λi z = exp(−z) exp P . exp − P λi + s λi + s From the form of ζ, we have only to check X X X X ( λi + s)( ξi + s) ≥ ( λi ∨ ξi + s)( λi ∧ ξi + s) or equivalently X X X X ( λi )( ξi ) ≥ ( λi ∨ ξi )( λi ∧ ξi ). We have X X X X ( λi )( ξi ) − ( λi ∨ ξi )( λi ∧ ξi ) X = {λi ξj + λj ξi − (λi ∨ ξi )(λj ∧ ξj ) − (λj ∨ ξj )(λi ∧ ξi )} . i6=j Without the loss of generality, assume λi ≥ ξi . Then we have λi ξj + λj ξi − (λi ∨ ξi )(λj ∧ ξj ) − (λj ∨ ξj )(λi ∧ ξi ) = λi ξj + λj ξi − λi (λj ∧ ξj ) − (λj ∨ ξj )ξi = λi {ξj − (λj ∧ ξj )} − ξi {(λj ∨ ξj ) − λj } = (λi − ξi ){ξj − (λj ∧ ξj )} ≥ 0, which completes the proof. Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 27 Theorem B.1 (Theorem 2.1 of Karlin and Rinott (1980)). Let f1 , f2 , f3 and f4 be nonnegative functions satisfying for all x, y ∈ Rν f1 (x)f2 (y) ≤ f3 (x ∨ y)f4 (x ∧ y). Then Z Z f1 (x)dx Z f2 (x)dx ≤ Z f3 (x)dx f4 (x)dx. Theorem B.2 (FKG Inequality, e.g. Theorem 2.3 of Karlin and Rinott (1980)). Let f (x) for x ∈ Rν be a probability density satisfying MTP2. Then for any pair of increasing functions g1 (x) and g2 (x), we have Z Z Z g1 (x)g2 (x)f (x)dx ≥ g1 (x)f (x)dx g2 (x)f (x)dx. References Aitchison, J. (1975). Goodness of prediction fit. Biometrika 62 547–554. MR0391353 Berger, J. O. (1985). Statistical decision theory and Bayesian analysis, second ed. Springer Series in Statistics. Springer-Verlag, New York. Brown, L. D. (1979). A heuristic method for determining admissibility of estimators—with applications. Ann. Statist. 7 960–994. MR536501 Brown, L. D., George, E. I. and Xu, X. (2008). Admissible predictive density estimation. Ann. Statist. 36 1156–1170. MR2418653 Corcuera, J. M. and Giummolè, F. (1999). A generalized Bayes rule for prediction. Scand. J. Statist. 26 265–279. MR1707658 Csiszár, I. (1967). Information-type measures of difference of probability distributions and indirect observations. Studia Sci. Math. Hungar. 2 299–318. MR0219345 George, E. I., Liang, F. and Xu, X. (2006). Improved minimax predictive densities under Kullback-Leibler loss. Ann. Statist. 34 78–91. MR2275235 Karlin, S. and Rinott, Y. (1980). Classes of orderings of measures and related correlation inequalities. I. Multivariate totally positive distributions. J. Multivariate Anal. 10 467–498. MR599685 Komaki, F. (2001). A shrinkage predictive distribution for multivariate normal observables. Biometrika 88 859–864. MR1859415 Lehmann, E. L. and Casella, G. (1998). Theory of point estimation, second ed. Springer Texts in Statistics. Springer-Verlag, New York. MR1639875 Liang, F. and Barron, A. (2004). Exact minimax strategies for predictive density estimation, data compression, and model selection. IEEE Trans. Inform. Theory 50 2708–2726. MR2096988 Maruyama, Y. and Strawderman, W. E. (2012). Bayesian predictive densities for linear regression models under α-divergence loss: some results and open problems. In Contemporary developments in Bayesian analysis and Y. Maruyama and T. Ohnishi/A Bayesian prediction under α-divergence 28 statistical decision theory: a Festschrift for William E. Strawderman. Inst. Math. Stat. (IMS) Collect. 8 42–56. Inst. Math. Statist., Beachwood, OH. MR3202501 Stein, C. (1974). Estimation of the mean of a multivariate normal distribution. In Proceedings of the Prague Symposium on Asymptotic Statistics (Charles Univ., Prague, 1973), Vol. II 345–381. Charles Univ., Prague. MR0381062 Stein, C. (1981). Estimation of the mean of a multivariate normal distribution. Ann. Statist. 9 1135–1151. MR630098 Strawderman, W. E. (1971). Proper Bayes minimax estimators of the multivariate normal mean. Ann. Math. Statist. 42 385–388. MR0397939 Yanagimoto, T. and Ohnishi, T. (2009). Bayesian prediction of a density function in terms of e-mixture. J. Statist. Plann. Inference 139 3064–3075. MR2535183
10math.ST
Quotients of surface groups and homology of finite covers via quantum representations arXiv:1510.00677v2 [math.GT] 4 Feb 2016 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE Abstract. We prove that for each sufficiently complicated orientable surface S , there exists an infinite image linear representation ρ of π1 (S ) such that if γ ∈ π1 (S ) is freely homotopic to a simple closed curve on S , then ρ(γ) has finite order. Furthermore, we prove that given a sufficiently complicated orientable surface S , there exists a regular finite cover S 0 → S such that H1 (S 0 , Z) is not generated by lifts of simple closed curves on S , and we give a lower bound estimate on the index of the subgroup generated by lifts of simple closed curves. We thus answer two questions posed by Looijenga, and independently by Kent, Kisin, Marché, and McMullen. The construction of these representations and covers relies on quantum SO(3) representations of mapping class groups. 1. Introduction 1.1. Main results. Let S = S g,n be an orientable surface of genus g ≥ 0 and n ≥ 0 boundary components, which we denote by S g,n . A simple closed curve on S is an essential embedding of the circle S 1 into S . We will call an element 1 , g ∈ π1 (S ) simple if there is a simple representative in its conjugacy class. Note that contrary to a common convention, we are declaring curves which are freely homotopic to boundary components to be simple. Our main result is the following: Theorem 1.1. Let S = S g,n be a surface of genus g and with n boundary components, excluding the (g, n) pairs {(0, 0), (0, 1), (0, 2), (1, 0)}. There exists a linear representation ρ : π1 (S ) → GLd (C) such that: (1) The image of ρ is infinite. (2) If g ∈ π1 (S ) is simple then ρ(g) has finite order. Thus, Theorem 1.1 holds for all surface groups except for those which are abelian, in which case the result obviously does not hold. In Subsection 4.4, we will give an estimate on the dimension of the representation ρ. Date: February 5, 2016. 1 2 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE We will show that the representation ρ we produce in fact contains a nonabelian free group in its image (see Corollary 4.4, cf. [10]). If S 0 → S is a finite covering space, we define the subspace H1s (S 0 , Z) ⊆ H1 (S 0 , Z) to be simple loop homology of S 0 . Precisely, let g ∈ π1 (S ) be simple, and let n(g) be the smallest positive integer such that gn(g) lifts to S 0 . Then H1s (S 0 , Z) := h[gn(g) ] | g ∈ π1 (S ) simplei ⊆ H1 (S 0 , Z), where [gn(g) ] denotes the homology class of gn(g) in S 0 . It is easy to check that H1s (S , Z) = H1 (S , Z). Identifying π1 (S 0 ) with a subgroup of π1 (S ), we write π1s (S 0 ) for the simple loop subgroup of π1 (S 0 ), which is generated by elements of the form gn(g) . Here again, g ranges over simple elements of π1 (S ). Observe that H1s (S 0 , Z) is exactly the image of π1s (S 0 ) inside of H1 (S 0 , Z). We obtain the following result as a corollary to Theorem 1.1: Theorem 1.2. Let S = S g,n be a genus g surface with n boundary components, excluding the (g, n) pairs {(0, 0), (0, 1), (0, 2), (1, 0)}. Then there exists a finite cover S 0 → S such that H1s (S 0 , Z) ( H1 (S 0 , Z), i.e. the simple loop homology of S 0 is properly contained in the full homology of S 0 . Again, Theorem 1.2 holds for all nonabelian surface groups, and in the abelian case the result obviously cannot hold. We remark that I. Irmer has proposed a version of Theorem 1.2 in [15], and she proves that H1s (S 0 , Z) = H1 (S 0 , Z) whenever the deck group S 0 → S is abelian. We obtain the following immediate corollary from Theorem 1.2: Corollary 1.3. Let S = S g,n be a genus g surface with n boundary components, excluding the (g, n) pairs {(0, 0), (0, 1), (0, 2), (1, 0)}. Then there exists a finite cover S 0 → S such that π1s (S 0 ) is properly contained in π1 (S 0 ). The representation ρ in Theorem 1.1 is produced using the quantum SO(3) representations of the mapping class group of S . We then use the Birman Exact Sequence to produce representations of the fundamental group of S . We use integral TQFT representations in order to approximate the representation ρ by finite image representations {ρk }k∈N which converge to ρ in some suitable sense. Since the representations {ρk }k∈N each have finite image, each such homomorphism classifies a finite cover S k → S . The cover S 0 → S furnished in Theorem 1.2 is any one of the covers S k → S , where k  0. Thus, the covers coming from integral TQFT representations will produce an infinite sequence of covers for which H1s (S k , Z) is a proper subgroup of H1 (S k , Z). 3 It is unclear whether H1s (S k , Z) has finite or infinite index inside of H1 (S k , Z), or equivalently if Theorem 1.2 holds when integral coefficients are replaced by rational coefficients. However we will show that for a fixed S , the index of H1s (S k , Z) in H1 (S k , Z) can be arbitrarily large. I. Agol has observed that Theorem 1.2 holds for the surface S 0,3 even when integral coefficients are replaced by rational coefficients (see [16]). We will show that if Theorem 1.2 holds over Q, then it in fact implies Theorem 1.1: see Proposition 6.1. 1.2. Notes and references. The question of whether the homology of a regular finite cover S 0 → S is generated by pullbacks of simple closed curves on S appears to be well–known though not well–documented in literature (see for instance [16] and [17], and especially the footnote at the end of the latter). The problem itself is closely related to the congruence subgroup conjecture for mapping class groups ([16], [4], [5]), to the virtually positive first Betti number problem for mapping class groups (see [21], and also [7] for a free group–oriented discussion), and to the study of arithmetic quotients of mapping class groups (see [14]). It appears to have been resistant to various “classical” approaches up to now. The problem of finding an infinite image linear representation of a surface group in which simple closed curves have finite order has applications to certain arithmetic problems, and the question was posed to the authors by M. Kisin and C. McMullen (see also Questions 5 and 6 of [17]). The existence of such a representation for the free group on two generators is an unpublished result of O. Gabber. Our work recovers Gabber’s result, though our representation is somewhat different. In general, the locus of representations X s ⊂ R(π1 (S )), GLd (C)) of the representation variety of π1 (S ) which have infinite image but under which every simple closed curve on π1 (S ) has finite image is invariant under the action of Aut(π1 (S )). Theorem 1.1 shows that this locus is nonempty, and it may have interesting dynamical properties. For instance, Kisin has asked whether any infinite image representation in R(π1 (S )) has a finite orbit under the action of Aut(π1 (S )), up to conjugation by elements of π1 (S ). The locus X s is a good candidate in the search for such representations. At least on a superficial level, Theorem 1.1 is related to the generalized Burnside problem, i.e. whether or not there exist infinite, finitely generated, torsion groups. The classical Selberg’s Lemma (see [22] for instance) implies that any finitely generated linear group has a finite index subgroup which is torsion–free, so that a finitely generated, torsion, linear group is finite. In the context of Theorem 1.1, we produce a finitely generated linear group which is not only generated by torsion elements (i.e. the images of finitely many simple closed curves), but the image of every element in the mapping class group orbit of these generators is torsion. 4 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE Finally, we note that G. Masbaum used explicit computations of TQFT representations to show that, under certain conditions, the image of quantum representations have an infinite order element (see [18]). This idea was generalized in [1] where Andersen, Masbaum, and Ueno conjectured that a mapping class with a pseudoAnosov component will have infinite image under a sufficiently deep level of the TQFT representations. In [1], the authors prove their conjecture for a four–times punctured sphere (cf. Theorem 4.1). It turns out that their computation does not imply Theorem 4.1, for rather technical reasons. Although the papers [18] and [1] do not imply our results, these computations of explicit mapping classes whose images under the TQFT representations are of infinite order is similar in spirit to our work in this paper. The reader may also consult the work of L. Funar (see [9]) who proved independently (using methods different from those of Masbaum) that, under certain conditions, the images of quantum representations are infinite. Again, Funar’s work does not imply our results. 2. Acknowledgements The authors thank B. Farb, L. Funar, M. Kisin, V. Krushkal, E. Looijenga, J. Marché, G. Masbaum, and C. McMullen for helpful conversations. The authors thank the hospitality of the Matematisches Forschungsinstitut Oberwolfach during the workshop “New Perspectives on the Interplay between Discrete Groups in LowDimensional Topology and Arithmetic Lattices”, where this research was initiated. The authors are grateful to an anonymous referee for a report which contained many helpful comments. 3. Background In this section we give a very brief summary of facts we will require from the theory of TQFT representations of mapping class groups. We have included references for the reader to consult, but in the interest of brevity, we have kept the discussion here to a minimum. 3.1. From representations of mapping class groups to representations of surface groups. Let S be an oriented surface with or without boundary and let x0 ∈ S be a marked point, which we will assume lies in the interior of S . Recall that we can consider two mapping class groups, namely Mod(S ) and Mod1 (S ) = Mod1 (S , x0 ), the usual mapping class group of S and the mapping class group of S preserving the marked point x0 , respectively. By convention, we will require that mapping classes preserve ∂S pointwise. 5 When the Euler characteristic of S is strictly negative, these two mapping class groups are related by the Birman Exact Sequence (see [2] or [8], for instance): 1 → π1 (S , x0 ) → Mod1 (S ) → Mod(S ) → 1. Thus, from any representation of Mod1 (S ), we obtain a representation of π1 (S ) by restriction. The subgroup π1 (S )  π1 (S , x0 ) < Mod1 (S ) is called the point–pushing subgroup of Mod1 (S ). When S has a boundary component B, one can consider the boundary–pushing subgroup of Mod(S ). There is a natural operation on S which caps off the boundary component B and replaces it with a marked point b, resulting in a surface Ŝ with one fewer boundary components and one marked point. There is thus a natural map Mod(S ) → Mod1 (Ŝ , b), whose kernel is cyclic and generated by a Dehn twist about B. The boundary–pushing subgroup BP(S ) of Mod(S ) is defined to be the preimage of the point–pushing subgroup of Mod1 (Ŝ , b). In general whenever Ŝ has negative Euler characteristic, we have an exact sequence 1 → Z → BP(S ) → π1 (Ŝ , b) → 1. The left copy of Z is central, and this extension is never split if Ŝ is closed and has negative Euler characteristic. In fact, BP(S ) is isomorphic to the fundamental group of the unit tangent bundle of Ŝ . The reader is again referred to [2] or [8] for more detail. Lemma 3.1. Let ρ : Mod1 (S ) → Q be a quotient such that for each Dehn twist T ∈ Mod1 (S ), we have ρ(T ) has finite order in Q. Then for every simple element g in the point–pushing subgroup of Mod1 (S ), we have that ρ(g) has finite order in Q. Proof. Let γ be an oriented simple loop in S based at x0 . Identifying γ with a simple or boundary parallel element g of the point–pushing subgroup of Mod1 (S ), we can express g as a product of two Dehn twists thus: let γ1 , γ2 ⊂ S be parallel copies of the loop γ, separated by the marked point x0 (i.e. one component of S \ {γ1 , γ2 } is an annulus containing the marked point x0 ). Then the point pushing map about γ is given, up to a sign, by g = T γ1 T γ−12 . Since γ1 and γ2 are disjoint, the corresponding Dehn twists commute with each other. Since ρ(T γi ) has finite order for each i, the element ρ(g) has finite order as well.  We will see in the sequel that if ρ is a TQFT representation of Mod1 (S ), then by Theorem 3.2 below, Lemma 3.1 applies. 3.2. SO(3)–TQFT representations. Let S g,n be a genus g closed, oriented surface with n boundary components. The SO(3) topological quantum field theories (TQFTs) take as an input an odd integer p ≥ 3 and a 2pth primitive root of unity. As an output, they give a projective representation ρ p : Mod1 (S g,n ) → PGLd (C), 6 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE which moreover depends on certain coloring data (which will be specified later on), and where here the dimension d depends the input data. The notion of a TQFT was introduced by Witten (see [25]). His ideas were based on a physical interpretation of the Jones polynomial involving the Feynman path integral, and the geometric quantization of the 3–dimensional Chern–Simons theory. The first rigorous construction of a TQFT was carried out by Reshetikhin and Turaev, using the category of semisimple representations of the universal enveloping algebra for the quantum Lie algebra SL(2)q (see [23] and [24]). We will work in the TQFT constructed by Blanchet, Habegger, Masbaum, and Vogel in [3], wherein an explicit representation associated to a TQFT is constructed using skein theory. Perhaps the most important feature of these representations is the following well–known fact (see [3]): Theorem 3.2. Let T ∈ Mod1 (S g,n ) be a Dehn twist about a simple closed curve. Then ρ p (T ) is a finite order element of PGLd (C). It is verifying that certain mapping classes have infinite order under TQFT representations which is often nontrivial and makes up most of the content of this paper. Here and in Subsection 3.3 we will survey some basic properties and computational methods for TQFT representations which we will require. One can define a certain cobordism category C of closed surfaces with colored banded points, in which the cobordisms are decorated by uni-trivalent colored banded graphs. The details of this category are not essential to our discussion; for details we direct the reader to [3]. The SO(3)–TQFT is a functor Z p from the category C to the category of finite dimensional vector spaces over C. A banded point (or an ribbon point) on a closed oriented surface is an oriented submanifold which is homeomorphic to the unit interval. If a surface has multiple banded points, we will assume that these intervals are disjoint. A banded point provides a good substitute for a boundary component within a closed surface, and a simple loop on S which encloses a single banded point can be thought of a boundary parallel loop. When one wants to study a surface with boundary from the point of view of TQFTs, one customarily attaches a disk to each boundary component and places a single banded point in the interior of each such disk. The banded points are moreover colored, which is to say equipped with an integer. By capping off boundary components, we can start with a surface S g,n and produce a closed surface Ŝ g,n equipped with n colored banded points. We will include a further colored banded point x in the interior of the surface with boundary S g,n , which will play the role of a basepoint. We denote by (Ŝ g,n , x) the closed surface with n + 1 colored banded points thus obtained. Now for p ≥ 3 odd, the SO(3)–TQFT defines a finite dimensional vector space V p (Ŝ g,n , x). 7 For the sake of computations, it is useful to write down an explicit basis for the space V p (Ŝ g,n , x). Denote by y the set of n + 1 colored banded points on (Ŝ g,n , x), and by S g the underlying closed surface without colored banded points. Let H be a handlebody such that ∂H = S g , and let G be a uni-trivalent banded graph such that H retracts to G. We suppose that G meets the boundary of H exactly at the banded points y and this intersection consists exactly of the degree one ends of G. A p–admissible coloring of G is a coloring, i.e. an assignment of an integer, to each edge of G such that at each degree three vertex v of G, the three (non–negative integer) colors {a, b, c} coloring edges meeting at v satisfy the following conditions: (1) |a − c| ≤ b ≤ a + c; (2) a + b + c ≤ 2p − 4; (3) each color lies between 0 and p − 2; (4) the color of an edge terminating at a banded point yi must have the same color as yi . To any p-admissible coloring c of G, there is a canonical way to associate an element of the skein module S A p (H, (Ŝ g,n , x)), where here the notation refers to the usual skein module with the indeterminate evaluated at A p , which in turn is a 2pth primitive root of unity. The skein module element is produced by cabling the edges of G by appropriate Jones-Wenzl idempotents (see [3, Section 4] for more detail). If moreover all the colors are required to be even, it turns out that the vectors associated to p–admissible colorings give a basis for V p (Ŝ g,n , x) (see [3, Theorem 4.14]). We can now coarsely sketch the construction of TQFT representations. If x0 ∈ x, we may contract x down to a point in order to obtain a surface with a marked point (S g,n , x0 ). To construct this representation, one takes a mapping class f ∈ Mod1 (S g,n , x0 ) and one considers the mapping cylinder of f −1 . The TQFT functor gives as an output a linear automorphism ρ p ( f ) of V p (Ŝ g,n , x). This procedure gives us a projective representation ρ p : Mod1 (S g,n , x0 ) → PAut(V p (S g,n , x)), since the composition law is well–defined only up to multiplication by a root of unity. Our primary interest in the representation ρ p lies in its restriction to the point pushing subgroup of Mod1 (S g,n , x0 ). More precisely, we need to compute explicitly the action of a given element of π1 (S g,n , x0 ) on the basis of V p (Ŝ g,n , x) described above. We will explain how to do these calculations in Subsection 4.1, and some illustrative examples will be given in the proof of Theorem 4.1. 8 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE 3.3. Integral TQFT representations. In this Subsection we follow some of the introductory material of [12] and of [13]. The TQFT representations of mapping class groups discussed in Subsection 3.2 are defined over C and may not have good integrality properties. In our proof of Theorem 1.2, we will require an “integral refinement” of the TQFT representations which was constructed by Gilmer [11] and Gilmer–Masbaum [12]. These integral TQFT representations have all the properties of general SO(3)–TQFTs which we will require, and in particular Theorem 1.1 holds for them. To an oriented closed surface S equipped with finitely many colored banded points and an odd prime p (with p ≡ 3 (mod 4)) , one can associate a free, finitely generated module over the ring of integers O p = Z[ζ p ], where ζ p is a primitive pth root of unity. We will denote this module by S p (S ). This module is stable by the action of the mapping class group, and moreover tensoring this action with C gives us the usual SO(3)–TQFT representation. Most of the intricacies of the construction and the properties of the integral TQFT representations are irrelevant for our purposes, and we therefore direct the reader to the references mentioned above for more detail. We will briefly remark that the construction of the integral TQFT representations is performed using the skein module, so that the computations in integral TQFT are identical to those in the TQFTs described in Subsection 3.2. The feature of these representations which we will require is the following filtration by finite image representations. Let h = 1 − ζ p , which is a prime in Z[ζ p ]. We consider the modules S p,k (S ) = S p (S )/hk+1 S p (S ), which are finite abelian groups for each k ≥ 0. The representation ρ p has a natural action on S p,k (S ), and we denote the corresponding representation of the mapping class group by ρ p,k . Observe (see [13]) that the natural map O p → lim O p /hk+1 O p ←−− is injective, where here the right hand side is sometimes called the h–adic completion of O p . We immediately see that \ ker ρ p,k = ker ρ p . k 4. Infinite image TQFT representations of surface groups 4.1. Quantum representations of surface groups. In this Subsection we describe the key idea of this paper, which is the use of the Birman Exact Sequence (see Subsection 3.1) together with TQFT representations in order to produce exotic representations of surface groups. Using the notation of Subsection 3.2, we have a 9 projective representation ρ p : Mod1 (S g,n , x0 ) → PAut(V p (Ŝ g,n , x)). Restriction to the point–pushing subgroup gives a projective representation of the fundamental group of S g,n : ρ p : π1 (S g,n , x0 ) → PAut(V p (Ŝ g,n , x)), which is defined whenever the Euler characteristic of S g,n is strictly negative. In order to make the computations more tractable, let us describe this action more precisely. Let γ : [0, 1] → S g,n be a loop based at x0 . By the Birman Exact Sequence, γ can be seen as a diffeomorphism fγ of (S g,n , x0 ). Pick a lift f˜γ of fγ to the ribbon mapping class group of (S g,n , x) , i.e. the mapping class group preserving the orientation on the banded point x, together with any coloring data that might be present. Note that any two lifts of fγ differ by a twist about the banded point x, which is a central element of the ribbon mapping class group. The preimage of the point pushing subgroup inside of the ribbon mapping class group is easily seen to be isomorphic to the boundary pushing subgroup BP(S g,n+1 ), where the extra boundary component is the boundary of a small neighborhood of the banded point x. By definition, f˜γ−1 is isotopic to the identity via an isotopy which is allowed to move x. Following the trajectory of the colored banded point x in this isotopy gives a colored banded tangle γ̃ inside S g × [0, 1]. Observe that γ̃ is just a thickening of the tangle t ∈ [0, 1] 7→ (γ(1 − t), t) ∈ S g,n × [0, 1]. We can naturally form the decorated cobordism Cγ̃ by considering S g × [0, 1] equipped with the colored banded tangle (γ̃, a1 × [0, 1], . . . , an × [0, 1]), where (a1 , . . . , an ) are the n colored banded points on Ŝ g,n . By applying the TQFT functor, we obtain an automorphism ρ p (γ) := ρ p ( f˜γ ) = Z p (Cγ̃ ). We crucially note that this automorphism generally depends on the choice of the lift f˜γ , but that changing the lift will only result in ρ p (γ) being multiplied by a root of unity, since the Dehn twist about x acts by multiplication by a root of unity. In particular, different choices of lift give rise to the same element PAut(V p (S g,n , x)). Thus, we do indeed obtain a representation of π1 (S g,n ), and not of the boundary pushing subgroup. We now briefly describe how to compute the action of the loop γ on the basis of TQFT described in Subsection 3.2. Let H be a handlebody with ∂H = S g and let G be a uni-trivalent banded graph such that H retracts to G. Let Gc be a coloring of 10 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE this graph as explained in Subsection 3.2. In the TQFT language, applying Z p (Cγ̃ ) to Gc simply means that we glue the cobordism Cγ̃ to the handlebody H equipped with Gc . This gluing operation gives the same handlebody H, but with a different colored banded graph inside. To express this new banded colored graph in terms of the basis mentioned above, we have to use a set of local relations on colored banded graphs, namely a colored version of the Kauffman bracket and the application of Jones– Wenzl idempotents. For the sake of brevity, we refer the reader to the formulas computed by Masbaum and Vogel (see [20]). 4.2. The three-holed sphere. The main technical case needed to establish Theorem 1.1 is the case of the surface S 0,3 , a sphere with three boundary components. Let Mod1 (S 0,3 ) be the mapping class group of S 0,3 , preserving a fourth marked point in the interior of S 0,3 . The correct setup for TQFT representations of Mod1 (S 0,3 ) is that of spheres with banded points. Let (S 2 , 1, 1, 2, 2) be a sphere equipped with two banded points colored by 1 and two banded points colored by 2 (where one of these two serves as the base point). Let p ≥ 5 and A p be a 2pth primitive root of unity. According to [3, Theorem 4.14] and the survey given in Subsection 3.2, the vector space V p (S 2 , 1, 1, 2, 2) has a basis described by the following two colored banded graphs (1) Ga = with a ∈ {0, 2}, where these graphs live in the 3–ball whose boundary is (S 2 , 1, 1, 2, 2). The computations can be done in this basis as described in Subsection 4.1. In order to simplify the computations, we will work in a different basis. We note that these colored banded graphs can be expanded to banded tangles in the 3-ball using a well–studied procedure (see [20] or [3, section 4]). It is easy to check that the following elements form a basis of V p (S 2 , 1, 1, 2, 2): u1 = u2 = More precisely, u1 = A2p G0 + G2 + A−2 p and u2 = G0 . 11 Here again, the ambient 3–ball in which these tangles live is not drawn. The arcs drawn stand for banded arcs with the blackboard framing. The two end points of the arcs in the top left of the picture are attached to the points colored by 1 on the boundary of the 3–ball. The two rectangles labeled by 2 represent the second Jones– Wenzl idempotent, and these are attached to the two points colored by 2 on the boundary of the 3–ball. The construction and most properties of the Jones–Wenzl idempotents are irrelevant for our purposes, and the interested reader is directed to [3, Section 3]. The TQFT representation of Mod1 (S 0,3 ) furnishes a homomorphism ρ p : Mod1 (S 0,3 ) → PAut(V p (S 2 , 1, 1, 2, 2))  PGL2 (C) in whose image all Dehn twists have finite order by Theorem 3.2, which then by restriction gives us a homomorphism ρ p : π1 (S 0,3 ) → PGL2 (C) in whose image all simple loops have finite order, by Lemma 3.1. Theorem 4.1. For all p  0, the image of the representation ρ p : π1 (S 0,3 ) → PGL2 (C) contains an element of infinite order. Proof. We can compute the action of π1 (S 0,3 ) on V p (S 2 , 1, 1, 2, 2) explicitly, and thus find an element of g ∈ π1 (S 0,3 ) such that ρ p (g) has infinite order. Let {γ1 , γ2 , γ3 } be the usual generators of the fundamental group of the three– holed sphere: Following the discussion in Subsection 4.1, we can write down matrices for the action of ρ(γi ) for each i. A graphical representation is as follows: ρ p (γ1 )u2 = ρ p (γ1 )u2 = 12 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE ρ p (γ2 )u1 = ρ p (γ2 )u2 = We can then reduce these diagrams using the skein relations = Ap + A−1 p = −A2p − A−2 p in order to obtain diagrams without crossing and without trivial circles. We then use the Jones–Wenzl idempotents, and in particular the rule = 0, in order to simplify the diagrams further. We thus obtain for each diagram a linear combination of the tangles {u1 , u2 }. One easily checks that we have the following matrices: ! ! −2 1 A−10 A−8 A2p − A−6 p − Ap p p ρ p (γ1 ) = ρ p (γ2 ) = −10 −12 . 0 A−12 A p − A−14 1 − A−8 p p p + Ap Similarly, we can compute ! A−8 0 p ρ p (γ3 ) = . −6 −A−2 1 p + Ap Here we recall that p ≥ 5 is an odd integer and A p is 2pth primitive root of unity. 4 −4 −12 Now, one checks that tr(ρ p (γ1 )ρ p (γ2 )−1 ) = A12 p − A p + 2 − A p + A p . So we have that |tr(ρ p (γ1 γ2−1 ))| −→iπ 5 > 2. A p →e 6 th iπ If we take a sequence of 2p primitive roots of unity {A p } such that A p → e 6 as p → ∞, we see that |tr(ρ p (γ1 γ2−1 ))| > 2 for p  0. Whenever |tr(ρ p (γ1 γ2−1 ))| > 2, an elementary calculation shows that ρ p (γ1 γ2−1 ) has an eigenvalue which lies off the unit circle. It follows that no power of ρ p (γ1 γ2−1 ) is a scalar matrix, since the determinant of this matrix is itself a root of unity. Thus, ρ p (γ1 γ2−1 ) has infinite order for p  0.  13 In the proof above, we note that the based loop γ1 γ2−1 is freely homotopic to a “figure eight” which encircles two punctures, and such a loop is not freely homotopic to a simple loop. In light of the discussion above, Theorem 4.1 could be viewed as giving another proof that γ1 γ2−1 is in fact not represented by a simple loop. Corollary 4.2. Let ρ p be as above. For p  0, the image of ρ p contains a nonabelian free group. Proof. Let g = γ1 γ2−1 as in the proof of Theorem 4.1, and let p  0 be chosen so that ρ p (g) has infinite order. We have shown that ρ p (g) in fact admits an eigenvalue which does not lie on the unit circle, so that λ = ρ p (g) ∈ PGL2 (C) can be viewed as a loxodromic isometry of hyperbolic 3–space. By a standard Ping–Pong Lemma argument, it suffices to show that there exists a loxodromic isometry µ in the image of ρ p whose fixed point set on the Riemann sphere Ĉ is disjoint from that of λ. Indeed, then sufficiently high powers of λ and µ will generate a free subgroup of PGL2 (C), which will in fact be a classical Schottky subgroup of PGL2 (C). See [6], for instance. To produce µ, one can just conjugate λ by an element in the image of ρ p . It is easy to check that for p  0, the element ρ p (γ1 ) ∈ PGL2 (C) has two fixed points, namely ∞ and the point −10 A−2 p − Ap z0 = , 1 − A−12 p which is distinct from infinity if A p is not a twelfth root of unity. An easy computation using ! 4 −4 −8 −12 6 2 −6 A24 −A14 p − Ap + 2 − Ap − Ap + Ap p + Ap − Ap + Ap λ= −14 −A−10 A−8 p + Ap p shows that λ does not fix ∞. Similarly, a direct computation shows that λ does not fix z0 . If p  0, the isometry ρ p (γ1 ) will not preserve the fixed point set of λ setwise, so that we can then conjugate λ by a power of ρ p (γ1 ) in order to get the desired µ.  4.3. General surfaces. In this Subsection, we bootstrap Theorem 4.1 in order to establish the main case of Theorem 1.1. Proof of Theorem 1.1. Retaining the notation of Subsection 4.1, we start with the data of a surface S g,n which is equipped with a marked point x0 in its interior. We then cap off the boundary components to get (Ŝ g,n , x), which is a closed surface with n + 1 colored banded points. Recall that x is a thickening of the marked point x0 . Suppose furthermore that x is colored by 2. 14 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE Restricting the quantum representation of Mod1 (S g,n , x0 ) to the point pushing subgroup gives a representation ρ p : π1 (S g,n , x0 ) → PAut(V p (Ŝ g,n , x)) Proving that ρ p has infinite image can be done by combining Theorem 4.1 with a standard TQFT argument which has already appeared in [18]. For the sake of concreteness, we now give this argument in the case g ≥ 2, and the other cases covered by Theorem 1.1 can be obtained in a similar fashion. We adopt the standing assumption that all banded points on Ŝ g,n are colored by 2. Consider the three-holed sphere inside S g,n whose boundary components are the two curves drawn in the following diagram: Thus, we can map π1 (S 0,3 ) into π1 (S g,n ) and get an action of π1 (S 0,3 ) on V p (Ŝ g,n , x). From Theorem [3, Theorem 1.14], we see that this action of π1 (S 0,3 ) on V p (Ŝ g,n , x) contains as a direct summand a vector space V ⊗ W, where π1 (S 0,3 ) acts on V as on V p (S 2 , 1, 1, 2, 2) as discussed in Subsection 4.2 and where W is another representation. The conclusion of the theorem follows from observation that some element of π1 (S 0,3 ) acts with infinite order on V = V p (S 2 , 1, 1, 2, 2) for p  0 by Theorem 4.1, and thus this element also acts with infinite order on V ⊗ W.  We briefly remark that, as mentioned in Subsection 3.3, Theorem 1.1 also holds for integral TQFT representations, with the same proof carrying over. Thus we have: Corollary 4.3. Let p be an odd prime, let ρ p be the associated integral TQFT representation of Mod1 (S ), and let π1 (S ) < Mod1 (S ) be the point–pushing subgroup. Then for all p  0, we have that: (1) The representation ρ p has infinite image. (2) If g ∈ π1 (S ) is simple or boundary parallel then ρ p (g) has finite order. A direct consequence of Corollary 4.2 and the splitting argument in TQFT used in the proof of Theorem 1.1 is the following fact: Corollary 4.4. The image of the representation ρ p : π1 (S g,n ) → PAut(V p (Ŝ g,n , x)) 15 contains a nonabelian free group for p  0 as soon as π1 (S g,n ) is not abelian. 4.4. Dimensions of the representations. In this subsection, we give a quick and coarse estimate on the dimension of the representation ρ in Theorem 1.1. It suffices to estimate the dimension of the TQFT representation ρ p of Mod1 (S ) which restricts to an infinite image representation of π1 (S ), and then embed the corresponding projective general linear group in a general linear group. The dimension of the space V p (S , x) is given by a Verlinde formula, which simply counts the number of even p-admissible colorings of trivalent ribbon graphs in a handlebody bounded by S , as sketched in Subsection 3.2. In principle, it is possible to compute the dimension of V p (S , x), though a closed formula is often quite complicated. See [3] and [12] for instance. We note that, fixing p, the dimension of V p (S , x) grows exponentially in the genus of S , since in the proof of Theorem 1.1 we have that the color assigned to each boundary component of S is 2. The argument for Theorem 1.1 furnishes one p which works for all genera, since the infinitude of the image of ρ p is proven by considering the restriction of ρ p to a certain three–holed sphere inside of S . It follows that the target dimension of ρ p : π1 (S ) → PGLd (C) is d ∼ C for some constant C > 1. Since PGLd (C) can be embedded in GLd2 (C) using the adjoint action, we obtain the following consequence: g Corollary 4.5. There is a constant C > 1 such that if S = S g is a closed surface of genus g, there is a representation ρ of π1 (S ) satisfying the conclusions of Theorem 1.1, of dimension bounded by C g . 5. Homology and finite covers In this section we use integral TQFT representations to prove Theorem 1.2 and Corollary 1.3. 5.1. From infinite image representations to finite covers. Before using integral TQFT to establish Theorem 1.2 we show how from any projective representation ρ : π1 (S ) → PGLd (C) satisfying the conclusions of Theorem 1.1 we can build a covering S 0 → S satisfying the conclusions of Corollary 1.3. Thus, any representation ρ : π1 (S ) → PGLd (C) satisfying the conclusions of Theorem 1.1, even one not coming from TQFTs, already gives a somewhat counterintuitive result. More precisely: Theorem 5.1. Let ρ : π1 (S ) → PGLd (C) be a projective representation such that (1) The image of ρ is infinite. (2) If g ∈ π1 (S ) is simple then ρ(g) has finite order. 16 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE Then there exists a finite cover S 0 → S such that π1s (S 0 ) is an infinite index subgroup of π1 (S 0 ). Proof. Let ρ : π1 (S ) → PGLd (C) as in the statement of the present theorem. The image of ρ is an infinite, finitely generated, linear group. Using Selberg’s Lemma, we can find a finite index torsion–free subgroup H C ρ(π1 (S )). So H 0 = ρ−1 (H) C π1 (S ) is a finite index subgroup which classifies a finite cover S 0 → S . Now let g ∈ π1 (S ) be simple and let n(g) be the smallest integer such that gn(g) ∈ 0 H . We note that the element ρ(gn(g) ) ∈ ρ(H 0 ) = H is torsion since ρ(g) has finite order, which forces ρ(gn(g) ) = 1, since H is torsion– free. On the one hand, we have that n(g) is precisely the order of ρ(g), and that π1s (S 0 ) < ker ρ. On the other hand H 0 = ρ−1 (H) so ker ρ < H 0 . But ρ has infinite image, so that ker ρ is an infinite index subgroup of π1 (S ), from which we can conclude that π1s (S 0 ) is an infinite index subgroup of H 0 = π1 (S 0 ).  5.2. Homology of finite covers. Let p ≥ 7 be an odd integer and let S = S g,n be compact surface as in the statement of Theorem 1.2. Using the notation of Subsection 4.1 we consider the representation ρ p : π1 (S ) → PAut(V p (Ŝ g,n , x)). th which depends on a 2p primitive root of unity A p . We suppose, as in the proof of Theorem 1.1, that all the banded points on (Ŝ g,n , x) are colored by 2. For compactness of notation, the space V p (Ŝ g,n , x) will be denoted by V p (S ). By Theorem 1.1, we may take p such that ρ p has infinite image. In particular, R := ker ρ p is an infinite index subgroup of π1 (S ). Now we suppose that p is a prime number such that p ≡ 3 (mod 4) and we define ζ p = A2p which is a pth root of unity. The integral TQFT as described in Subsection 3.3 defines a representation ρ p : π1 (S ) → PAut(S p (S )), where S p (S ) is a free Z[ζ p ]–module of finite dimension d p . If k ≥ 0 is an integer, we can consider the representations ρ p,k : π1 (S ) → PAut(S p (S )/hk+1 S p (S )), where here h = 1 − ζ p . Since the abelian groups S p (S )/hk+1 S p (S ) are finite, the groups Rk := ker ρ p,k are finite index subgroups of π1 (S ). Let D be the normal subgroup of π1 (S ) generated by {gn(g) | g ∈ π1 (S ) simple and n(g) the order of ρ p (g)}. 17 Note that since simple elements of π1 (S ) have finite order image under ρ p , the value of n(g) is always finite, so that the definition of D makes sense. Similarly for k ≥ 0 let Dk be the normal subgroup generated by {gn(g,k) | g ∈ π1 (S ) simple , n(g, k) the order of g in π1 (S )/Rk }. Observe that if the subgroup Rk < π1 (S ) classifies a cover S k → S , then Dk is identified with the subgroup π1s (S k ) < π1 (S ). We have the following filtration D ⊂ R ⊂ · · · ⊂ Rk+1 ⊂ Rk ⊂ · · · ⊂ R1 ⊂ R0 , and we have that \ Rk = R k (see Subsection 3.3), so that f < R if and only if f < Rk for k  0. Lemma 5.2. For all k  0, we have that Dk = D. Proof. If g ∈ π1 (S ) is simple then ρ p (g) has finite order n(g). Since \ Rk = R, k we have that the order of the image of g in π1 (S )/Rk is exactly n(g) for all k  0. Finally, we have that ρ p is the restriction of a representation of the whole mapping class group Mod1 (S ), under whose action there are only finitely many orbits of simple closed curves. In particular, for all k  0 and all simple g ∈ π1 (S ), the order of the image of g in π1 (S )/Rk is exactly n(g).  With our choice of p fixed, let us write N0 for the smallest k for which Dk = D, as in Lemma 5.2. Let φ ∈ π1 (S ) such that ρ p (φ) has infinite order. There exists m0 such that φm0 ∈ RN0 , since RN0 has finite index inside the group π1 (S ). For compactness of notation, we will write ψ for φm0 . Observe that ψ < R, since ρ p (φ) has infinite order, and it follows that for k  N0 , we have that ψ < Rk . We set N ≥ N0 to be the integer such that ψ ∈ RN \ RN+1 . Notice that RN is a finite index subgroup of π1 (S ), whereas DN = D is an infinite index subgroup of RN . In particular, RN can be naturally identified with the fundamental group of a finite regular cover S N → S , and D can be naturally identified with a subgroup of π1 (S N ), i.e. the subgroup π1s (S N ). For each N, we may write qN : RN → RN /[RN , RN ] for the abelianization map. 18 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE Theorem 5.3. There is a proper inclusion qN (D) ( RN /[RN , RN ]. In fact, for every δ ∈ [RN , RN ], we have that δ · ψ < D. Theorem 5.3 implies Theorem 1.2 fairly quickly: Proof of Theorem 1.2. Setting S 0 in the statement of Theorem 1.2 to be the cover S N → S classified by the subgroup RN , we have that D can be identified with π1s (S 0 ). The image of D in RN /[RN , RN ] under qN is exactly H1s (S 0 , Z). The conclusion of Theorem 1.2 now follows from Theorem 5.3.  Proof of Theorem 5.3. We will fix ψ and N as in the discussion before the theorem. First, a standard and straightforward computation shows that [RN , RN ] ⊂ R2N+1 . Now, suppose there exists an element δ ∈ [RN , RN ] such that δ · ψ ∈ D ⊂ RN+1 . Then, we would obtain ψ ∈ δ−1 RN+1 . Since [RN , RN ] ⊂ R2N+1 by the claim above, we have that δ ∈ R2N+1 . But then we must have that ψ ∈ RN+1 as well, which violates our choice of N, i.e. ψ ∈ RN \ RN+1 .  5.3. Index estimates. In this subsection, we estimate the index of H1s (S N , Z) inside of H1 (S N , Z) as a function of p and of N. In particular, we will show that the index can be made arbitrarily large by varying both p and N. Proposition 5.4. The index of qN (D) in RN /[RN , RN ] is at least pe , where here $ % N e= + 1. p−1 We need the following number–theoretic fact which can be found in [19, Lemma 3.1]: Lemma 5.5. There exists an invertible element z ∈ Z[ζ p ] such that p = z · h p−1 . The following is an easy number–theoretic fact, whose proof we include for the convenience of the reader: Lemma 5.6. If k and p are relatively prime, then k is invertible modulo hn , for all n ≥ 1. Proof. Since p ≡ 0 (mod h), we have that pn ≡ 0 (mod hn ). Since k and p are relatively prime, we have that for each n ≥ 1, there exist integers a and b such that a · k + b · pn = 1. Thus, a · k ≡ 1 (mod hn ).  Lemma 5.7. Let 1 ≤ k ≤ pe − 1. With the notation of Theorem 5.3, we have that ψk < R2N+1 . 19 Proof. We compute the “h–adic” expansion of ρ p (ψ) in a basis, as in the proof of Theorem 5.3. Up to an invertible element of Z[ζ p ], we have ρ p (ψ) = I + hN+1 ∆, where here ∆ . 0 (mod h), since ψ ∈ RN \ RN+1 . If 1 ≤ k ≤ pe − 1, we obtain the expansion ρ p (ψ)k ≡ I + k · hN+1 ∆ (mod h2N+2 ). Since 1 ≤ k ≤ pe − 1, we have that k can be written k = m · pl , with 0 ≤ l < e and with m relatively prime to p. Lemma 5.5 implies that p = z · h p−1 , so that k · hN+1 ∆ = m · zl · hl(p−1)+N+1 ∆. Now if k · hN+1 ∆ ≡ 0 (mod h2N+2 ), we see that hl(p−1)+N+1 ∆ ≡ 0 (mod h2N+2 ), since m and z are invertible modulo h2N+2 . Note however that l < e, so that l(p − 1) < N + 1. The expression hl(p−1)+N+1 ∆ ≡ 0 (mod h2N+2 ) now implies that ∆ ≡ 0 (mod hN+1−l(p−1) ), which is impossible since N + 1 − l(p − 1) > 0 and since ∆ . 0 (mod h). It follows that k · hN+1 ∆ . 0 (mod h2N+2 ), which in turn implies that ψk < R2N+1 .  Proof of Proposition 5.4. In Theorem 5.3, we established that for each δ ∈ [RN , RN ] ⊂ R2N+1 , we have that ψ · δ < D ⊂ R2N+1 . From Lemma 5.7, it follows that powers of ψ represent at least pe distinct cosets of R2N+1 in RN , and hence of [RN , RN ] in RN , whence the claim of the proposition.  20 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE 6. Representations of surface groups, revisited In this final short section, we illustrate how a rational version of Theorem 1.2 implies Theorem 1.1, thus further underlining the interrelatedness of the two results. We will write S for a surface as above. Proposition 6.1. Let S 0 → S be a finite regular cover of S , and suppose that rk(H1s (S 0 , Q)) < rk(H1 (S 0 , Q)). Then there exists a linear representation ρ : π1 (S ) → GLd (Z) such that: (1) The image of ρ is infinite. (2) The image of every simple element of π1 (S ) has finite order. In fact, the image of ρ can be virtually abelian (i.e. the image of ρ has a finite index subgroup which is abelian). We remark again that our general version of Theorem 1.2 implies a proper inclusion between H1s (S 0 , Z) and H1 (S 0 , Z), which may no longer be proper when tensored with Q. Therefore, we do not get that Theorem 1.1 and Theorem 1.2 are logically equivalent. Indeed, in order to deduce Theorem 1.2, we had to use specific properties of the SO(3)–TQFT representations. Proof of Proposition 6.1. Let S 0 → S be a finite regular cover as furnished by the hypotheses of the proposition, and let G be the deck group of the cover. Write H1 (S 0 , Q)  A ⊕ B, where A  H1s (S 0 , Q) and B , 0. Note that the natural action of G on H1 (S 0 , Q) respects the summand A, since being simple is a conjugacy invariant in π1 (S ). Write AZ and BZ for the intersections of these summands with H1 (S 0 , Z). Note that H1s (S 0 , Z) ⊂ AZ . Notice that for each integer m ≥ 1, the subgroup m · AZ is characteristic in AZ , and is hence stable under the G–action on H1 (S 0 , Z). Let Γ be the group defined by the extension 1 → H1 (S 0 , Z) → Γ → G → 1, which is precisely the group Γ  π1 (S )/[π1 (S 0 ), π1 (S 0 )]. Write Γm = Γ/(m · AZ ). This group is naturally a quotient of π1 (S ). Note that for all m, the group Γm is virtually a finitely generated abelian group, since it contains a quotient of H1 (S 0 , Z) with finite index. Note that every finitely generated abelian group is linear over Z, as is easily checked. Moreover, if a group K contains a finite index subgroup H < K which is linear over Z, then K is also linear over Z, as is seen by taking the induced representation. It follows that for 21 all m, the group Γm linear over Z. Furthermore, there is a natural injective map BZ → Γm , so that Γm is infinite. Finally, if g ∈ π1 (S ) is simple, then for some n = n(g) > 0, we have that [gn ] ∈ H1s (S 0 , Z) ⊂ AZ . It follows that g has finite order in Γm . Thus, the group Γm has the properties claimed by the proposition.  References [1] J.E. Andersen, G. Masbaum, K. Ueno. Topological quantum field theory and the NielsenThurston classification of M(0, 4). Math. Proc. Cam. Phil. Soc. 141(2006) 447–488. [2] J. Birman. Braids, links, and mapping class groups. Annals of Mathematics Studies, No. 82. Princeton University Press, Princeton, N.J.; University of Tokyo Press, Tokyo, 1974. [3] C. Blanchet, N. Habegger, G. Masbaum, P. Vogel. Topological quantum field theories derived from the Kauffman bracket. Topology 34(4) (1992), no. 4, 883–927. [4] M. Boggi. The congruence subgroup property for the hyperelliptic modular group: the open surface case. Hiroshima Mathematical Journal, 39(3):351–362, 2009. [5] M. Boggi. The congruence subgroup property for the hyperelliptic modular group: the closed surface case. Preprint, 2014. [6] P. de la Harpe. Topics in geometric group theory. Chicago Lectures in Mathematics. University of Chicago Press, Chicago, IL, 2000. [7] B. Farb and S. Hensel. Moving homology classes in finite covers of graphs. Preprint, 2015. [8] B. Farb and D. Margalit. A primer on mapping class groups. Princeton Mathematical Series, 49. Princeton University Press, Princeton, NJ, 2012. [9] L. Funar. On the TQFT representations of the mapping class groups. Pacific J. Math. 188(1999), 251–274. [10] L. Funar and T. Kohno. Free subgroups within the images of quantum representations. Forum Math. 26 (2014), no. 2, 337–355. [11] P. Gilmer. Integrality for TQFTs. Duke Math. J. 125 (2004), no. 2, 389–413. [12] P. Gilmer and G. Masbaum. Integral lattices in TQFT. Ann. Sci. École Norm. Sup. (4) 40 (2007), no. 5, 815–844. [13] P. Gilmer and G. Masbaum. Integral TQFT for the one–holed torus. Pacific J. Math. (252) no. 1 (2011), 93–112. [14] F. Grunewald, M. Larsen, A. Lubotzky, and J. Malestein. Arithmetic quotients of the mapping class group. Geom. Funct. Anal., to appear. [15] I. Irmer. Lifts of simple curves in finite regular coverings of closed surfaces. Preprint, 2015. [16] R. Kent and J. Marché, with comments by I. Agol. Math Overflow post, http://mathoverflow.net/questions/86894/homology-generated-by-lifts-of-simple-curves, 2012. [17] E. Looijenga. Some algebraic geometry related to the mapping class group. Oberwolfach Reports, June 2015. [18] G. Masbaum. An element of infinite order in TQFT-representations of mapping class groups. Low-dimensional topology (Funchal, 1998), volume 233 of Contemp. Math., pages 137–139. Amer. Math. Soc., Providence, RI, 1999. [19] G. Masbaum and J. D. Roberts. A simple proof of integrality of quantum invariants at prime roots of unity. Math. Proc. Cambridge Philos. Soc. 121 (1997) no. 3, 443-454 [20] G. Masbaum and P. Vogel. 3-valent graphs and the Kauffman bracket. Pacific J. Math. 162(2) : 361–381, 1994 22 THOMAS KOBERDA AND RAMANUJAN SANTHAROUBANE [21] A. Putman and B. Wieland. Abelian quotients of subgroups of the mapping class group and higher Prym representations. J. London Math. Soc. (2) 88 (2013), no. 1, 79–96. [22] M.S. Raghunathan. Discrete subgroups of Lie groups. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 68. Springer-Verlag, New York–Heidelberg, 1972. [23] N. Reshetikhin and V. G. Turaev. Invariants of 3-manifolds via link polynomials and quantum groups. Invent. Math., 103(3):547–597, 1991. [24] V. G. Turaev. Quantum invariants of knots and 3-manifolds. de Gruyter Studies in Mathematics, vol. 18. Walter de Gruyter and Co., Berlin, 2010. [25] E. Witten. Quantum field theory and the Jones polynomial. Comm. Math. Phys., 121:351–399, 1989. Department of Mathematics, University of Virginia, Charlottesville, VA 22904-4137, USA E-mail address: [email protected] Institut de Mathématiques de Jussieu (UMR 7586 du CNRS), Equipe Topologie et Géométrie Algébriques, Case 247, 4 pl. Jussieu, 75252 Paris Cedex 5, France E-mail address: [email protected]
4math.GR
Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic arXiv:cs/0205034v1 [cs.DS] 18 May 2002 Robert Lupton∗ F. Miller Maley† Neal Young‡ Abstract This paper describes an NP-hard combinatorial optimization problem arising in the Sloan Digital Sky Survey and a practical approximation algorithm that has been implemented and will be used in the Survey. The algorithm is based on network flow theory and Lagrangian relaxation. 1 The Sloan Digital Sky Survey “The Sloan Digital Sky Survey [is] a joint project of the Astrophysical Research Consortium. ... The goal of the project, which is scheduled to begin in 1997 and take five years, is to make a much better map of the universe than is currently available. The volume of the universe to be surveyed will be 100 times larger than the volume of previous surveys. The number of galaxies with known distances is expected to increase by a factor of 100 to 1,000,000 galaxies and the number of quasars to increase to 100,000. “The Sloan Foundation ... has contributed $8 million to the $18 million capital costs of the project. ... “In order to do the survey, ARC is designing and building a special purpose 2.5 meter (100-inch) telescope at its Apache Point Observatory. ... “[The Sky Survey will proceed in two phases. In the first phase, a two-dimensional map of the sky will be made. For the second phase, the] million brightest stars and the one hundred thousand brightest quasars will be selected for spectroscopic analysis from the two-dimensional map... [29, Savani, 1994] To gather the spectroscopic data in the second phase, the telescope will be pointed repeatedly at the sky to take a series of “snapshots”. Each snapshot will capture data for up to 660 galaxies and quasars in the circular portion of the sky visible through the telescope. For each captured galaxy, light from that galaxy will enter the telescope and travel through an optical fiber to a spectral analyzer. The optical fibers (one for each galaxy) will be held in place by a “plug plate” drilled to hold the up to 660 fibers, each aligned to accept the light of its respective galaxy [5]. ∗ Astrophysics Department, Princeton University, Princeton, NJ 08540. E-mail: [email protected]. Mathematics Department, Princeton University, Princeton, NJ 08540. E-mail: [email protected]. ‡ Computer Science Department, Dartmouth College, Hanover, NH 03755. Parts of this research were done at: AT&T Bell Laboratories, Murray Hill, NJ 07974; the School of ORIE, Cornell University, Ithaca NY 14853 on Éva Tardos’ NSF PYI grant DDM-9157199; and the Dept’s of Astrophysics and Computer Science, Princeton University. Corresponding author. E-mail: [email protected]. † 1 Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 2 Figure 1: Sample instance (points are dark); near-uniform cover; better cover. This near-uniform cover is from an earlier implementation not using Hardin et al.’s covers, which are more uniform. 1.1 A Capacitated Covering Problem. The second phase of the survey is expected to cost on the order of $4-5 million. This cost will depend primarily on the number of snapshots taken. This paper concerns the following problem: given the “2-dimensional” locations of the desired galaxies, determine a minimum-size set of snapshots that capture them. Formally: Euclidean Capacitated Covering by Disks (ECCD) Given a collection of points on the unit sphere, a radius r, and a capacity c, find a small set of discs of radius r (located on the sphere) such that each given point can be assigned to a disc containing it, with no disc being assigned more than c points. The sphere corresponds to the view-sphere centered at the telescope. The points correspond to the images of the galaxies projected on the view-sphere. Each disc represents one snapshot to be taken through the telescope; the points assigned to that disc correspond to those galaxies for which data will be collected in that snapshot. The capacity c is the maximum number of galaxies for which spectral data can be gathered in a single snapshot (due to limitations in packing the optical fibers). The ECCD problem is NP-hard [21]. The instances we need to solve will have hundreds of thousands of points. Luckily, as Figure 1 illustrates, the instances we need to solve are nicely structured. In this paper we describe a heuristic algorithm for the problem. The algorithm is effective for instances arising in the Survey and will be used for it. The basic idea behind the algorithm is to start with a near-uniform cover of the sphere by discs [13] and then to iteratively improve the cover. The key observation is that a given cover can be improved by first solving a relaxation of the problem in which the “point-in-disc” constraints are replaced by penalties for assigning points to discs not containing them, and then moving the discs to minimize the cost of the assignment found. The relaxed problem reduces to the minimum-cost flow problem. In our tests, the algorithm runs in nearly linear time and finds covers that are roughly 20% better than comparable near-uniform covers. 2 Related Work The NP-completeness of the variant when the points lie in the plane was proven by Megiddo and Supowit [21]. The proof adapts easily to our problem. The NP-completeness of the planar problem when the discs are required to be centered on the given points was proven by Marchetti-Spaccamela Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 3 [20]. When the covering regions are rings, instead of discs, Maass [19] showed the problem NPcomplete even if the points all lie on a single line. Papadimitriou [26] (improving results by Fisher and Hochbaum [7]), considered the related pmedians problem in the plane, which is that of covering the given points with p discs (of arbitrary radii, but centered at p of the given points) so as to minimize the sum of the disc radii. He showed the problem to be NP-complete and presented average-case analyses of several algorithms. One of the heuristics is a uniform (“honeycomb”) covering of the points by discs, which he shows gives a near-optimal solution with high probability when p is ω(log n) and o(n/ log n) and the points are randomly distributed in the unit square. The problem can be modeled as a capacitated set-cover problem. The well-known greedy algorithm of Johnson [15] and Lovász [17], as modified for the capacitated case by Bar-Ilan, Kortsarz, and Peleg [2], would yield a ln n-approximate solution, where n is the number of galaxies. This algorithm is not good enough in practice. For this particular set-cover problem the dual of the set system has bounded VC-dimension; in this case an improved approximation algorithm is known for the uncapacitated case [3], but, judging from a few small trials, this algorithm does not appear to take sufficient advantage of the structure of our problem instances to perform well in practice. Numerous generalizations of our problem have been considered under various names, including “(un)capacitated facility (or plant) location,” “p-centers”, and “minimax facility location”. These problems have been studied under various metrics and also in general graphs. In general, polynomial-time exact algorithms are known only when the number of covering regions (in our case, discs) is small (e.g., [1]) or when the underlying metric space (or network) is tree-like (e.g., [23, 10, 22, 12, 14, 6]). Generally, these algorithms are for uncapacitated problems. There is a large literature on these problems in Operations Research. Relevant books include [18, 25, 9, 8]. Much of this research has concentrated on adapting integer-programming techniques to fairly general formulations of the problem. For example, recent works on the Capacitated Facility Location Problem (a generalization of our problem to arbitrary networks) include [4, 31]. Quoting from the conclusion of “Approximate Solutions to Large Scale Capacitated Facility Location Problems” (1990) [30]: The problem of locating facilities has inspired a rich body of literature which spans well over two decades. Numerous algorithms have been devised and successfully applied to problems with as many as 200 customers and 100 facilities. The computational experience on larger problems, however, has been virtually non-existent... In the work leading to this paper, the objective was to develop a heuristic algorithm that can be used to generate effective solutions for large scale facility locations problems. The computational results obtained so far seem to indicate that this requirement can be met for problems with as many as 1000 customers and 100 facilities. 3 The Algorithm The instances arising in the Sky Survey exhibit particular structure. Within any given region, the galaxies are distributed densely throughout the region, somewhat uniformly but with clustering tendencies and variation in density. The density of the galaxies means that virtually the entire region must be covered by discs. The variation in density means that more discs must be concentrated within densely populated regions. As a reference point, consider the sparsest possible covering of the area by discs (resembling a “honeycomb”). This cover provides roughly the right total capacity and does well in sparse areas, but in dense areas does not provide sufficient capacity. Any good solution will have to maintain a honeycomb-like structure in sparse areas while bunching discs more densely in dense areas. Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 4 Figure 2: Composite of intermediate covers The outer loop of the algorithm does a binary search for the smallest value of a density parameter δ that leads to success in the inner loop. The inner loop begins with a near-uniform cover of normalized density 1 + δ and iteratively improves it (see Figure 1 for “before” and “after” covers). Each iteration of the loop perturbs the discs, as described below, in an attempt to improve the cover (Figure 2 shows the results of such a series of improvement steps). If the desired coverage is obtained, the inner loop stops (successfully). If the perturbation ceases to improve the cover, the inner loop stops (unsuccessfully). Next we describe how the algorithm perturbs a given cover in order to improve it. We start with the observation that for a given set of discs (with known locations), the problem of finding the maximum number of galaxies that can be assigned reduces to a generalized maximum bipartite matching problem in a graph G = (U, V, E), where the vertices in U correspond to galaxies, the vertices in V correspond to the discs, and edge (u, v) is present if u’s galaxy is in v’s disc. A maximum legal assignment of galaxies to discs then corresponds to a maximum size set S of edges such that each u ∈ U is incident to at most one edge in S while each v ∈ V is incident to at most c edges in S. Since the latter problem reduces in a standard way to the maximum flow problem [27], which can be efficiently solved, it follows that for a given set of discs, one can efficiently find a maximum legal assignment of galaxies to discs. Of course, the maximum legal assignment may still leave many galaxies unassigned, even though many discs are not used to capacity. In this case, how can discs be moved to improve the coverage? Consider the following relaxation of the problem: Relaxed Problem Given a set of discs, a set of galaxies, and a capacity c, find a minimum-penalty assignment of the galaxies to discs such that no disc is assigned more than c galaxies. Here a galaxy can be assigned to a disc not containing it, but there is a penalty for doing so that encourages assignments of galaxies to nearby discs (details of the penalty function are in § 3.2). The relaxed problem can be solved efficiently (even for arbitrary penalties) by reducing it to the assignment problem or to minimum-cost maximum flow. We reduced it to the latter, more general, Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 5 Figure 3: Relaxed assignment problem in anticipation of having to incorporate more complex constraints on the assignment (that no sufficiently close pairs of galaxies should be assigned to the same disc) at a later point. As described below, even the more general problem can be solved quickly enough for our purposes. A solution to the relaxed problem will assign all galaxies to discs, but a given disc may be assigned galaxies outside of it. The advantage of the relaxed problem is that a solution to it can give information about how to improve a given set of discs. The intuition is that if excess demand (i.e. a high density of galaxies relative to discs) exists in one area, and excess capacity exists in another, then a disc between the two areas will tend to be assigned galaxies that are outside of the disc and that lie towards the area of excess demand. Figure 3 illustrates this. Once a minimum-penalty solution to the relaxed problem has been found, the algorithm moves the discs to minimize the cost of the particular assignment of galaxies to discs specified by the minimum-penalty solution. This problem can be solved independently for each disc. For a given disc, for a fixed set of galaxies assigned to it, the sum of the penalties for those assignments is a function f (x, y) of the coordinates (x, y) of the center of the disc. As long as the penalty function is convex and reasonably smooth, f will be also. Starting with the current location (x0 , y0 ) of the disc, a simple gradient-based method (described in § 3.2) is used to find (x, y) maximizing f (x, y). This gives us the essentials of the inner loop of the algorithm. It starts with a near-uniform cover of some specified (normalized) density 1 + δ. It improves the cover by finding a minimum-penalty assignment of the galaxies to the discs and then moving the discs to their optimal locations given that assignment. It continues, alternately improving the assignment and then moving the discs, until the net penalty ceases to decrease appreciably. At the end of the inner loop, the algorithm finds a legal (not relaxed) assignment of galaxies to discs maximizing the number of assigned galaxies. Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 6 Inner Loop (ǫ is fixed, δ is determined by the outer loop) 1. Compute a minimum-size near-uniform cover C of the region by discs so that the total capacity of the discs in C is at least 1 + δ times the number of galaxies. 2. Repeat until convergence (after “polishing”) or 1−ǫ of the discs are legally assigned. (a) Compute a minimum-penalty assignment A of the galaxies to the discs in C. Do this by solving (an approximation of) the corresponding minimum-cost flow problem. (b) Move the discs in C to minimize the penalty associated with the assignment A. Move each disc independently to minimize its associated penalty by a simple gradient-descent method. 3. Find a maximum legal assignment of the galaxies to C. Do this by solving a corresponding maximum-flow problem. 4. If at least 1 − ǫ of the galaxies are assigned, succeed, else fail. Figure 4: Given a desired coverage 1 − ǫ, where ǫ ≥ 0, the outer loop of the algorithm does a binary search for the smallest value of δ ≥ 0 such that the above inner loop succeeds. Further details, including the “polishing” step, the “approximation” of the flow problem, and the criteria for convergence, are described in § 3.2. Figure 2 shows a sequence of covers generated by a single run of the inner loop. The outer loop performs a binary search to find the smallest d that causes the inner loop to successfully cover the galaxies. The presentation here is a slight simplification of the actual algorithm, in that the actual algorithm uses a “polishing” heuristic before terminating the inner loop, and a heuristic is applied to reduce size of the network-flow problem before solving it. These heuristics and other details about convergence of the inner and outer loops, and starting conditions for the outer loop, are described in § 3.2. 3.1 Example Run of Inner Loop. The sample instance in Figure 1 contains 12642 points — a random 10% of the points in a subregion of the sky previously scanned. The size of this subregion is about 10% of that of the region that will be mapped by the Survey. A uniform cover of 218 discs of capacity 60 (total capacity 13080) allows 81% of the galaxies to be assigned. After 16 iterations of the inner loop of the algorithm, the improved cover captures 97.8% of the points. Figure 1 shows the initial near-uniform cover and the final cover; Figure 2 shows a composite of the successive covers. Section 4 describes comprehensive testing of quality of solutions given by the algorithm and its running time. 3.2 Implementation Details. For the initial near-uniform covers, we use Hardin, Sloane, and Smith’s catalogue of packings of points on the sphere [13]. These packings give covers of the entire sphere, but we need a cover of 7 Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 0 20000 rounded penalty -10 18000 rounded penalty 16000 -20 14000 12000 -30 10000 -40 8000 6000 -50 4000 -60 2000 -70 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 1 1.2 1.4 1.6 1.8 2 Figure 5: The assignment penalty as a function of the distance d (in disc radii) between the disc center and the galaxy. The plot on the left is for d ≤ 1; the plot on the right is for d ≥ 1. only a (usually rectangular) subregion of the sky. To prune a “global” cover C the algorithm first finds a maximum legal assignment of galaxies to discs in C, then discards all discs having at most a few assigned galaxies. (The cutoff for discarding a disc is chosen so that the resulting number of discs is as desired.) The inner loop of the algorithm is implemented in C++ using LEDA [24] for basic data structures. We use a scaling algorithm by Andrew Goldberg to solve the minimum-cost flow problems [11]. We use TCL for the outer loop of the algorithm and to collect performance data. Penalty function: is proportional to The penalty for assigning a galaxy to a disc whose center is distance d away p(d) =  d2 − r 2 if d ≤ r 100(d2 − r 2 ) if d ≥ r. Recall that r is the disc radius. When solving the relaxed problem, the algorithm first rounds the penalties. Rounding so that there are few distinct penalties allows a heuristic reduction in the size of the resulting flow problem. (This heuristic is discussed further below.) Figure 5 shows plots of p and the rounded penalties. The rounding is chosen to preserve the distance between the galaxy and the edge of the disc within roughly a factor of 2. The edge of the disc is important because the penalty function is least smooth for points near the edge. The factor of 2 is somewhat arbitrary, it was chosen to balance between the advantages of rounding and the resulting loss of accuracy. After rounding, only 14 or so distinct penalties (each an integer power of 2) arise. Reducing the size of the flow problem: We expected the bottleneck in the algorithm to be solving the minimum-cost flow problems. To minimize this time, the algorithm uses a heuristic to reduce the minimum-cost flow problem to a smaller, approximately equivalent, problem. This is the “approximation” of the flow problem mentioned in the high-level description of the algorithm. First, the algorithm only considers assigning each galaxy to discs whose centers are within a distance of 2 disc radii, and of these at most the three closest discs. Second, it rounds the penalties as described above to reduce the number of distinct penalties. Finally, instead of having vertices for individual galaxies, it has vertices for equivalence classes of galaxies, where two galaxies are equivalent if they have the same assignable discs with the same rounded penalties. With these heuristics, even for Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 8 very dense sets of galaxies, the number of equivalence classes will be proportional to the number of discs as long as each disc intersects O(1) other discs. This is true in our case. The precaution of using equivalence classes turned out to be unnecessary for two reasons. First, the average number of galaxies per equivalence class was typically no more than 3. More fundamentally, solving the flow problems was not in fact a substantial bottleneck (see the data in § 4.2 and the subsequent discussion).1 Constructing the flow problem: The algorithm stores all the discs in a two-dimensional array so that discs near any given point can be found rapidly. To construct the flow problem, the algorithm iterates through the galaxies. For each galaxy, it finds the discs whose centers are within 2 disc radii. It selects the three nearest of these discs and computes the rounded distances to each. These discs and their rounded distances determine the equivalence class of the galaxy. The equivalence class is found (or created if necessary). From the equivalence classes, the flow network is constructed. So that the equivalence classes can be found quickly, each equivalence class is stored in a hash table maintained at its nearest disc. The hash table for a disc D contains those equivalence classes whose nearest disc is D. This method preserves locality of reference. In an earlier implementation, a single large hash table held all the equivalence classes. For large problems, this table was too large to fit in main memory. This slowed the algorithm by a factor of roughly 50. Moving the discs: After the minimum-penalty relaxed assignment is found, recall that each disc is moved individually to minimize the penalty associated with that disc. The “simple gradientdescent method” used to do this is as follows. To minimize f (x, y), starting at a point (x0 , y0 ), compute the gradient (direction of maximum rate of increase), then move (x, y) in steps of α (approximately 16/1000 disc radii, chosen to balance speed and accuracy) in the direction opposite the gradient until such steps ceased to decrease the value of f (x, y). Recompute the gradient at the new location and repeat the process with α halved. Continue in this fashion, halving α each time, until α is decreased to approximately 2/1000 disc radii. Convergence and “Polishing”: The outermost loop of the algorithm does a binary search on the size of the uniform starting cover. Within this loop, the inner loop iteratively improves the given cover. We describe convergence of the inner loop first. Recall that the inner loop starts with a given cover and improves it until the desired number of galaxies are legally covered, or until “convergence” occurs. Convergence is determined as follows: after each iteration, if the gap between the actual number of galaxies covered and the desired number did not decrease by at least 5%, then the algorithm considers the process “stuck”. At this point it changes the basic improvement step (this is the “polishing” heuristic mentioned in the high-level descriptions of the algorithm) as follows: it solves the relaxed problem as if the disc radius were 2% smaller. It continues with this heuristic until it also becomes stuck. Every time the process becomes stuck, the algorithm alternates between the standard improvement step and the modified one. If the process is ever stuck for at least two sequential rounds, it is considered to have converged. The purpose of the polishing heuristic is that in the original relaxed problem, a disc may be assigned galaxies that are just barely outside of it at little penalty. These galaxies cannot be legally assigned, yet may “hold” discs in place 1 It is conceivable that the rounding of the penalties decreased the time used by the minimum-cost flow algorithm, as the latter works by scaling. 9 Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 1.35 name b c d e f g r. ascens. 35 to 55 32 to 57 30 to 59 28 to 62 25 to 65 20 to 70 declination -55 to -35 -57 to -32 -55 to -30 -57 to -28 -60 to -20 -70 to -18 galaxies 29933 45344 52520 70339 109681 157126 1.3 1.25 near-uniform cover 1.2 algorithm’s cover 1.15 1.1 1.05 0 20 40 60 80 100 galaxies/1000 120 140 160 Figure 6: Regions from which sample instances were generated; number of discs needed to achieve a 98% coverage (normalized by capacity lower bound). in the subsequent disc-moving step. “Shrinking” the effective radius of the disc for a few rounds encourages these galaxies to be assigned elsewhere. Next we describe initial conditions and the convergence criterion for the outer loop. The outer loop maintains a lower bound L and an upper bound U on the minimum sufficient cover size. It also maintains covers CL and CU obtained by starting with a uniform cover of size L or U (respectively) and applying the basic algorithm to improve the cover until the desired coverage is obtained or convergence occurs. Initially L and U are taken to be 1.05 and 1.15, respectively, times the number of galaxies divided by the capacity per disc. The binary search maintains the invariant that CL and CU are, respectively, insufficient and sufficient to achieve the desired coverage. If this invariant does not hold initially, L and/or U are adjusted in increments of 5% to achieve the invariant. The algorithm halts the binary search as soon as the following condition ceases to be met: CU has more than one more disc than CL , CU is at least 0.5% bigger than CL , and CU legally covers at least 0.5% more galaxies than CL . Once the search halts, the algorithm returns CU . 4 Performance of the Algorithm We tested the running time and the quality of the solutions found by the algorithm on sample instances. In this section we describe the results. The Survey will map roughly 25% of the sky — the region having right ascension zero through 360 degrees and declination 30 degrees through 90 degrees. Roughly one million galaxies will be mapped. Because the two phases of the Survey will be pipelined (the second will be started before the first is done), the second phase will be done in pieces. We generated the problem instances from data from a region of the sky that had been previously scanned for a different purpose. We selected 6 subregions, and for these subregions we generated 4 problem instances by randomly sampling 30, 50, 70 or 100% of the galaxies. This gave us 24 sample problems. We took the disc radius to be 1.5 arc-seconds and the capacity to be 600 times 0.3, 0.5, 0.7, or 1 corresponding to the sampling percentage above. (The base capacity is 600 instead of 660 because approximately 60 points in each disc will be reserved for quasars not in the sample.) The largest region has an area roughly 4% of the entire sky. For each subregion, the right ascension and declination ranges and the number of galaxies are shown in Figure 6. 10 0.2 1.8 0.18 1.7 0.16 1.6 milliseconds per galaxy seconds per galaxy Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 0.14 0.12 0.1 total time 0.08 moving discs 0.06 0.04 0 0 20 40 60 80 100 galaxies/1000 120 140 1.4 1.3 1.2 1.1 1 building graph solving flow 0.02 1.5 time per iteration 0.9 160 0.8 0 20 40 60 80 100 galaxies/1000 120 140 160 Figure 7: Net time per galaxy and main components; time per galaxy per iteration. Each vertical bar represents a group of points with close x-coordinates: the center of the bar is the average; the endpoints are one standard deviation away. 4.1 Quality of solutions. Figure 6 illustrates the quality of the solutions returned by the algorithm on the 24 problem instances. The figure plots the size of the cover needed to assign 98% of the galaxies in each region, normalized by dividing by the number of discs needed just to provide enough capacity to hold 98% of the galaxies. The plot shows the same information for covering by near-uniform covers. The algorithm (very roughly) requires 5% to 15% extra capacity, whereas using uniform covers requires 25% to 35% extra capacity. 4.2 Running time. Plots of the time per galaxy to solve each problem instance as a function of the number of galaxies appear in Figure 7. This net time includes all of the iterations needed to find the final cover for the given problem instance, including the binary search “outer loop”. The three main components of the running time are the time building the graphs (including finding the equivalence classes of galaxies), the time solving the flow problems, and the time moving the discs. These plots show that the net running time is on the order of 0.1 cpu seconds per galaxy (850000 galaxies per day), with the three main components each taking a substantial fraction of the time. These tests were carried out on a Silicon Graphics machine with 6 150 MHZ processors, a 16 Kbyte data cache, a 1 Mbyte secondary cache, and 256 Mbytes of main memory. Most of the variation in the time per galaxy is due to the number of iterations, which varied from 30 to 100 per problem instance, and which increases (in our implementation) with the problem size. Figure 7 plots the average time per galaxy per iteration versus the problem size. Each iteration represents the solution of one relaxed problem and one perturbation of one set of discs. The time per iteration grows linearly with the number of galaxies. This is as expected, except for the surprising speed of the flow computations. We note that the binary search is fairly naive, given that in principle a fairly precise guess about the correct size of the starting cover could be made. Similarly, we feel that a more careful and less conservative estimate of convergence, possibly interleaving the two loops in some fashion, might be warranted. These improvements might reduce the total number of iterations substantially. 11 Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 2.6 0.22 2.4 0.2 edges per galaxy 2.2 milliseconds per edge normalized flow problem size 2 1.8 1.6 time to solve flow, per edge 0.18 0.16 0.14 0.12 1.4 1.2 0.2 0.3 0.4 0.5 0.6 0.7 0.8 sample density 0.9 1 1.1 0.1 0 50 100 150 edges/1000 200 250 Figure 8: Number of edges per galaxy vs. density; time per edge vs. number of edges Time to solve flow problems. The heuristics for keeping the flow problems small appear to be effective. Figure 8 plots the average number of edges per galaxy in each flow problem as a function of the sampling density of the instance. Figure 8 plots the average time per edge to solve the individual flow problems. The time appears to grow only near-linearly with the number of edges. Better than worst-case behavior on certain classes of problems is not uncommon; further the flow problems arising here are not particularly hard ones. See [16] for computational studies related to this issue. 5 Retrospective The Euclidean capacitated covering problem arising here is very natural. Looking in the Operations Research literature, we found numerous capacitated covering algorithms based on integer linear programming, but these were not fast enough problems of the desired size. The Computer Science literature had a number of efficient approximation algorithms for covering that had provable worstcase performance guarantees, yet these algorithms would not produce good enough solutions in practice. Nonetheless, our final solution rests on theoretical foundations. Our algorithm works in the spirit of Lagrangian relaxation. We decompose the problem into two parts: finding the cover C and finding the assignment A. We relax the constraints on A by replacing the “disc-containment” constraint by a penalty function. Then, for any given cover, finding the minimum-penalty assignment is a tractable problem. Likewise, for any given assignment, finding the minimum-penalty cover is tractable. Thus, the relaxation yields a scheme that iteratively reduces the minimum penalty and so drives the pair (C, A) closer to feasibility. Lagrangian relaxation is a common technique in both the Operations Research [31] and the Computer Science [28] literature. Finding the decomposition required an understanding of network flow theory. The ability to solve large problems hinges on a fast network flow algorithm. Classic augmenting paths algorithms are far too slow for our purpose. Goldberg’s algorithm incorporates both recent research within the worst-case model and heuristics discovered by empirical studies (in the spirit of [16]). Also useful were Hardin, Sloane, and Smith’s sphere covers [13]. These enabled us to start with better uniform covers than we might have otherwise. Finally, in prototyping and testing ideas, it helped to have a pre-existing library of relevant high-level data types and algorithms. For this we Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 12 used LEDA [24]. Worst-case analysis did side-track us slightly. Although worst-case analysis suggested that network flow would be the bottleneck for large problems, it was not at all. Ironically, as described in § 3.2, our first attempt to keep the flow problems small by using equivalence classes backfired: our original implementation used a single hashing data structure to hold all the equivalence classes; although the standard worst-case model suggests hashing is quite fast, its incautious use slowed the solutions of large problems by a factor of 50 due to the lack of locality of reference. In conclusion, our experience suggests that a successful approach rested on theoretical understanding, but required that it be creatively adapted to take advantage of the particular structure of our problem instances. 6 Acknowledgements Thanks to Ken Steiglitz for introducing two of the coauthors and to an anonymous referee for helpful suggestions. References [1] Pankaj K. Agarwal and Micha Sharir. Planar geometric location problems. Algorithmica, 11(2):185–195, February 1994. [2] Judit Bar-Ilan, Guy Kortsarz, and David Peleg. How to allocate network centers. Journal of Algorithms, 15(3):385–415, November 1993. [3] Hérve Brönnimann and Michael T. Goodrich. Almost optimal set covers in bounded VCdimension. Discrete and Computational Geometry, 14(4), 1995. [4] Gérard Cornuéjols, R. Sridharan, and Jean-Michel Thizy. A comparison of heuristics and relaxations for the capacitated plant location problem. European journal of operational research, 50(3):280, 1991. [5] Robert P. Crease. How technique is changing science. Science, 257:344–353, July 7 1992. [6] Ehan Erkut, Richard L. Francis, and Arie Tamir. Distance-constrained multifacility minimax location problems on tree networks. Networks, 22(1):37–55, 1992. [7] Marshall L. Fisher and Dorit S. Hochbaum. Probabilistic analysis of the planar k-median problem. Mathematics of Operations Research, 5:27–34, 1980. [8] Richard L. Francis. Discrete location theory. Wiley-Interscience series in discrete mathematics and optimization. J. Wiley (New York), 1990. [9] Richard L. Francis, Leon F. McGinnis Jr., and John A. White. Facility layout and location: an analytical approach. Prentice-Hall international series in industrial and systems engineering. Prentice-Hall, Englewood Cliffs, N.J., 1991. [10] Greg N. Frederickson and Donald B. Johnson. Finding kth paths and p-centers by generating and searching good data structures. Journal of Algorithms, 4(1):61–80, March 1983. Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 13 [11] Andrew V. Goldberg. An efficient implementation of a scaling minimum-cost flow algorithm. Journal of Algorithms, 22(1):1–29, January 1997. [12] Yuri Gurevich, Larry Stockmeyer, and Uzi Vishkin. Solving NP-hard problems on graphs that are almost trees and an application to facility location problems. Journal of the ACM, 31(3):459–473, July 1984. [13] Ronald H. Hardin, Neil J. A. Sloane, and Warren D. Smith. Tables of spherical codes in dimensions 3, 4, 5. AT&T Bell Laboratories, Murray Hill, NJ. available via http://research.att.com, 1993-1997. [14] Xin He and Yaacov Yesha. Efficient parallel algorithms for r-dominating set and p-center problems on trees. Algorithmica, 5:129–145, 1990. [15] David S. Johnson. Approximation algorithms for combinatorial problems. Journal of Computer and System Sciences, 9(3):256–278, December 1974. [16] David S. Johnson and Catherine C. McGeoch, editors. Network Flows and Matchings — First DIMACS Implementation Challenge, volume 12 of Series in Discrete Mathematics and Computer Science. AMS, 1993. [17] László Lovász. On the ratio of optimal integral and fractional covers. Discrete Mathematics, 13:383–390, 1975. [18] Robert F. Love, James G. Morris, and George O. Wesolowsky. Facilities Location: Models and Methods, volume 7 of Publications in Operations Research. North-Holland, New York, 1988. [19] Wolfgang Maass. On the complexity of nonconvex covering. SIAM Journal on Computing, 15(2):453–467, May 1986. [20] Alberto Marchetti-Spaccamela. The p-center problem in the plane is NP-complete. In Proc. 19th Allerton Conf. Commun. Control Comput., pages 31–40, 1981. [21] Nimrod Megiddo and Kenneth J. Supowit. On the complexity of some common geometric location problems. SIAM Journal on Computing, 13(1):182–196, February 1984. [22] Nimrod Megiddo and Arie Tamir. New results on the complexity of p-center problems. SIAM J. Comput., 12:751–758, 1983. [23] Nimrod Megiddo, Arie Tamir, Eitan Zemel, and R. Chandrasekaran. An O(n log2 n) algorithm for the kth longest path in a tree with applications to location problems. SIAM Journal on Computing, 10(2):328–337, May 1981. [24] Kurt Melhorn and Stefan Näher. LEDA — a platform for combinatorial and geometric computing. available via http://www.mpi-sb.mpg.de/LEDA/leda.html. [25] George L. Nemhauser and Laurence A. Wolsey. Integer and Combinatorial Optimization. John Wiley & Sons, New York, 1988. [26] Christos H. Papadimitriou. Worst-case and probabilistic analysis of a geometric location problem. SIAM Journal on Computing, 10(3):542–557, August 1981. Data Collection for the Sloan Digital Sky Survey — A Network-Flow Heuristic 14 [27] Christos H. Papadimitriou and Kenneth Steiglitz. Combinatorial Optimization: Algorithms and Complexity. Prentice-Hall, Englewood Cliffs, NJ, 1982. [28] Serge A. Plotkin, David B. Shmoys, and Éva Tardos. Fast approximation algorithms for fractional packing and covering problems. In 32nd Annual Symposium on Foundations of Computer Science, pages 495–504, San Juan, Puerto Rico, 1–4 October 1991. IEEE. [29] Jacquelyn Savani. Sky survey receives $5 million from NSF. Princeton Weekly Bulletin, November 1994. [30] Bala Shetty. Approximate solutions to large scale capacitated facility location problems. Applied mathematics and computation, 39(2):159, September 1990. [31] R. Sridharan. A lagrangian heuristic for the capacitated plant location problem with single source constraints. European journal of operational research, 66(3):305, 1993.
5cs.CE
On the Reliability Function of Distributed Hypothesis Testing Under Optimal Detection arXiv:1801.03687v1 [cs.IT] 11 Jan 2018 Nir Weinberger, Member, IEEE and Yuval Kochman, Member, IEEE Abstract The distributed hypothesis-testing problem with full side-information is studied. The trade-off (reliability function) between the type 1 and type 2 error exponents under limited rate is studied in the following way. First, the problem of determining the reliability function of distributed hypothesis-testing is reduced to the problem of determining the reliability function of channel-detection codes (in analogy to a similar result which connects the reliability of distributed compression and ordinary channel codes). Second, a single-letter random-coding bound based on an hierarchical ensemble, as well as a single-letter expurgated bound, are derived for the reliability of channel-detection codes. Both bounds are derived for the optimal detection rule. We believe that the resulting bounds are ensemble-tight, and hence optimal within the class of quantization-and-binning schemes. Index Terms Binning, channel-detection codes, distributed hypothesis-testing, error exponents, expurgated bounds, hierarchical ensembles, multiterminal data compression, random coding, side information, statistical inference, superposition codes. I. I NTRODUCTION In the hypothesis-testing (HT) problem, a detector needs to decide between two hypotheses regarding the underlying distribution of some observed data; the hypotheses are commonly known as the null hypothesis and the alternative hypothesis. Two types of error probability are defined - the type 1 error probability of deciding on the alternative hypothesis when the null hypothesis prevails, and type 2 error probability for the opposite event. The celebrated Neyman-Pearson lemma (e.g., [1, Prop. II.D.1]) states that the detector which achieves the optimal trade-off between the two error probabilities takes the form of comparing the likelihood-ratio to a threshold. In the context of this work, we only consider data which is a sequence Z1 , . . . , Zn of independent and identically distributed observations, and therefore the hypotheses correspond to the distribution of a random variable Z . In information-theoretic literature (e.g., [2], [3, Ch. 11], [4, Ch. 1], [5, Sec. 2]), large-deviations theory, most notably Sanov’s theorem, is usually applied to this problem. In particular, Stein’s theorem provides the largest exponential The material in this paper was presented in part at the Information Theory and Applications (ITA) Workshop, San Diego, California, USA, February 2017. 2 decrease rate of the type 2 error probability when the type 1 error probability is bounded away from one. More generally, the reliability function, i.e., the optimal trade-off between the two types of exponents when both are strictly positive, is also known. As a special case of the HT problem, one may consider a pair of random variables Z = (X, Y ), instances of which are fully observed by the detector. If, however, the detector does not directly observe the data sequence, the characterization of the optimal performance is much more challenging. A common model for such a scenario is known as distributed hypothesis-testing (DHT), and its reliability function is the subject of this paper. The DHT model was introduced by Berger [6] as an example for a problem at the intersection of multi-terminal information theory and statistical inference. In this model, one encoder observes X1 , . . . , Xn , while the other observes the corresponding Y1 , . . . , Yn ; they produce codewords to be sent over limited-rate noiseless links to a common detector. The goal is to characterize the optimal detection performance, under a given pair of encoding rates. As a starting point, usually an asymmetric model (sometimes referred to as the side-information case) is studied, in which the Y -observations are fully available to the decoder, and thus there is only a single rate constraint. The first major breakthrough on this problem was in a notable paper by Ahlswede and Csiszár [7], who addressed a special scenario termed testing against independence. In this case, the null hypothesis states that (X, Y ) have a given joint distribution PXY , whereas the alternative hypothesis states that they are independent, but with the same marginals as in the null hypothesis. This case is special since the Kullback–Leibler divergence, which is usually associated with Stein’s exponent of HT, can be identified as the mutual information between X and Y under PXY , which is naturally related to compression rates. This allowed the authors to use distributed compression, specifically quantization-based encoding techniques from [8], [9], to derive Stein’s exponent for the side-information testingagainst-independence problem [7, Th. 2]. Quantization-based encoding was also used for an achievable Stein’s exponent for a general pair of memoryless hypotheses [7, Th. 5], but without a converse bound. As summarized in [10, Sec. IV], later progress on this problem for a pair of general hypotheses was nonconsecutive, and contributions were made by several groups of researchers. First, in [11], the achievable bound on Stein’s exponent from [7, Th. 5] was improved, and also generalized beyond the side-information case. Then, in [12], achievable bounds on the full trade-off between the two types of exponents were derived. In [13], Stein’s exponent for side-information cases was further significantly improved using binning, as will be described in the sequel. Interestingly, when either the X -marginal or the Y -marginal of the hypotheses is different, positive exponents can be obtained even for zero-rate encoders [10, Sec. V].1 For this case, achievable Stein’s exponents and exponential trade-offs were derived in [10, Th. 5.5] and [11], [12], [15], along with matching converse bounds when some kind of assumptions are imposed. In the last decade, a renewed interest in the problem arose, aimed both at tackling more elaborated models, as well as at improving the results on the basic model. As for the former, Stein’s exponents under positive rates 1 There is also a variant of one-bit encoding, see [10, Sec. V] and [14]. 3 were explored in successive refinement models [16], for multiple encoders [17], for interactive models [18], [19], under privacy constraints [20], combined with lossy compression [21], over noisy channels [22], [23], for multiple decision centers [24], as well as over multi-hop networks [25]. Exponents for the zero-rate problem were studied under restricted detector structure [14] and for multiple encoders [26]. The finite blocklength and second-order regimes were addressed in [27]. As for the basic model, which this work also investigates, the encoding approach used in [13] is currently the best known. It is based on quantization and binning, just as used, e.g., for distributed lossy compression (the Wyner-Ziv problem [28, Ch. 11] [29]). First, the encoding rate is reduced by quantization of the source vector to a reproduction vector chosen from a codebook. Second, the rate is further reduced by binning of the reproduction vectors. As the detector is equipped with side information, it can identify the true reproduction vector with high reliability. In the context of DHT, it can then decide on one of the hypotheses using this reproduction and the side information. In [17] it was shown that a quantization-and-binning scheme achieves the optimal Stein’s exponent in a testing against conditional independence problem (as well as in a model inspired by the Gel’fand-Pinsker problem [30], and also in a Gaussian model). In [31], the quantization-and-binning scheme was shown to be necessary for the case of DHT with degraded hypotheses. In [32], [33], improved exponents were derived (for a doubly-symmetric binary source) by refining the analysis of the effect of binning errors. In addition, a full achievable exponent trade-off was presented, and Körner-Marton coding [34] was used in order to extend the analysis to the symmetric-rate case (for the symmetric source). Finally, [35] suggested an improved detection rule, in which the reproduction vectors in the bin are exhausted one-by-one, and the null hypothesis is declared if a single vector is jointly typical with the side-information vector. It is evident that the detectors above are all sub-optimal, as they are based upon the two-stage process of reproduction and then testing. However, the decoding of the source vector (or its quantized version) is totally superfluous for the DHT system, as the requirement from the latter is only to distinguish between the hypotheses. In fact, this detection procedure is still sub-optimal even if quantization is used without binning, and only the second stage of the detector is employed. While [35] offered an improved detector, it is still sub-optimal.2 In this work, we investigate the performance of the optimal detector. In fact, the optimal detector directly follows from the standard Neyman-Pearson lemma (see Section III). Nonetheless, analyzing its performance is highly non-trivial for DHT. Specifically, we study the reliability function in the side-information case, and will be guided by the following methodology. Recall that for distributed lossless compression systems (the Slepian-Wolf problem [28, Ch. 10] [37]), the side information helps to fully reproduce the source. The concept of binning the source vectors was originally conceived for this problem, and a common wisdom for this problem states that the source vectors which belong to the same bin should constitute a good channel code for the memoryless channel induced by the conditional distribution 2 Recently in the zero-rate regime, [27] considered the use of an optimal Neyman-Pearson-like detector, rather than the possibly sub-optimal Hoeffeding-like detector [36] that was used in [12]. 4 of Y given X . This intuition was made precise in [38, Th. 1][39], [40], which showed that the reliability function3 of distributed compression is directly related to the reliability of channel coding. The idea is to use structured binning,4 using a sequence of channel codes which achieves the channel reliability function. At a given blocklength, such a channel code corresponds to one bin of the distributed compression system.5 All other bins of the system are generated by permuting6 the source vectors of the first bin. Due to the memoryless nature of the problem, all bins generated this way are essentially as good as the original one, and this allows to directly link the reliability function of distributed compression to that of channel coding. While the reliability of channel codes is itself only known above the critical rate [4, Corollary 10.4], this characterization has two advantages nonetheless. First, analyzing channel codes is simpler than analyzing a distributed compression systems. Second, any bound on the reliability of channel codes immediately translates into a bound on the reliability of distributed compression systems. Specifically, the expurgated bound [4, Problem 10.18] and the sphere-packing bound [4, Th. 10.3] can be immediately used, rather than just a random-coding bound [4, Th. 10.2]. Noting the similarity between the distributed compression problem and the DHT problem, it is natural to ask whether structured binning is useful for the DHT problem, and what are the the properties of a “good” bin for the DHT problem? To address these questions, we introduce the concept of channel-detection (CD) codes. Such codes are not required to carry information, but are rather designed for the task of distinguishing between two possible channel distributions.7 Namely, a codeword from the code is chosen with a uniform probability over the codewords, and the detector should decide on the prevailing channel, based only on the output vector (and its knowledge of the codebook). It will be evident that this is the same problem encountered by the detector of a DHT system, given the encoded message. It will be shown that optimal DHT systems (in the exponential sense) can be generated by optimal CD codes, just like distributed compression systems are generated from ordinary channel codes. From this observation, the close relation between the reliability of DHT systems and CD codes will be determined. An illustration of the analogy between the relations distributed compression/channel coding and DHT/CD relations appears in Fig. 1 (with all quantities there will be formally defined in the sequel). . This intimate connection allows us to derive bounds on the reliability function of DHT using bounds on the reliability of CD codes. Concretely, we will derive both random-coding bounds and expurgated bounds on the reliability of CD codes under the optimal Neyman-Pearson detector. The analysis goes beyond that of [42] in two senses: first, it is based on a Chernoff distance characterization of the optimal exponents, which leads to simpler single-letter bounds; and second, the analysis is performed for an hierarchical ensemble 8 , corresponding 3 The reliability function of distributed lossless compression is the optimal exponential decrease of the error probability as a function of the compression rate. 4 Also mentioned in [17] for the DHT problem, though recognized as inessential. 5 More precisely, this is done type-by-type, i.e., per the subsets of source sequences that share the same empirical distribution. 6 This permutation technique, originally developed in [38], [41] will be useful here too, and will be reviewed in more detail in what follows. 7 In [42], a a somewhat different channel-detection setting is considered, where the code is required to simultaneously be a good channel code (in the ordinary sense), as well as a good CD code. In this work, it will only be required that the codewords of the CD code are different from one another. 8 Yielding superposition codes [43], such as the ones used for the broadcast channel, wee, e.g., [28, Ch. 5]) 5 Distributed Source Coding xn encoder decoder 010 : : : fn (xn ) Distributed Hypothesis Testing x ^n 'n (fn (xn ); y n ) xn encoder 010 : : : fn (xn ) detector H=H̄ 'n (fn (xn ); y n ) yn yn PXY =P̄XY PXY Channel Coding Channel Detection PY jX X PY jX Y Y X P̄Y jX Figure 1. An analogy between distributed compression and DHT. to quantization-and-binning schemes. The outline of the rest of the paper is as follows. The system model and preliminaries, such as notation conventions and background on ordinary HT, will be given in Section II. The main result of the paper, namely an achievable bound on the reliability function of DHT under optimal detection, will be stated in Section III, along with some consequences. For the sake of proving these bounds, the reduction of the DHT reliability problem to the CD reliability problem will be considered in Section IV. While only achievability bounds will ultimately be derived, the reduction to CD codes has both an achievability part as well as a converse part. Derivation of single-letter achievable bounds on the reliability of CD codes will be considered in Section V. From this, the achievability bounds on the DHT reliability will immediately follow. Afterwards, a discussion on computational aspects along with a numerical example are given in Section VI. Several directions for further research are highlighted in Section VII. 6 II. S YSTEM M ODEL A. Notation Conventions Throughout the paper, random variables will be denoted by capital letters, specific values they may take will be denoted by the corresponding lower case letters, and their alphabets will be denoted by calligraphic letters. Random vectors and their realizations will be superscripted by their dimension. For example, the random vector X n = (X1 , . . . , Xn ) (n positive integer), may take a specific vector value xn = (x1 , . . . , xn ) ∈ X n , the nth order Cartesian power of X , which is the alphabet of each component of this vector. The Cartesian product of X and Y (finite alphabets) will be denoted by X × Y . We will follow the standard notation conventions for probability distributions, e.g., PX (x) will denote the probability of the letter x ∈ X under the distribution PX . The arguments will be omitted when we address the entire distribution, e.g., PX . Similarly, generic distributions will be denoted by Q, Q, and in other similar forms, subscripted by the relevant random variables/vectors/conditionings, e.g. QXY , QX|Y . The joint distribution induced by QX and QY |X will be denoted by QX × QY |X . In what follows, we will extensively utilize the method of types [4], [44] and use the following notations. The type class of a type QX at blocklength n, i.e., the set of all xn ∈ X n with empirical distribution QX , will be denoted by Tn (QX ). The set of all type classes of vectors of length n from X n will be denoted by Pn (X ), and the def S set of all possible types over X will be denoted by P(X ) = ∞ n=1 Pn (X ). Similar notations will be used for pairs of random variables (and larger collections), e.g., Pn (U × X ), and Tn (QU XY ) ⊆ U n × X n × Y n . The conditional type class of xn for a conditional type QY |X , namely, the subset of Tn (QY ) such that the joint type of (xn , y n ) is QXY , will be denoted by Tn (QY |X , xn ) (sometimes called the Q-shell of xn [4, Definition 2.4]). For a given QX ∈ Pn (X ), the conditional type classes QY |X such that Tn (QY |X , xn ) is not empty when xn ∈ Tn (QX ) will be denoted by Pn (Y, QX ). The probability simplex for an alphabet X will be denoted by S(X ). For two distributions PX , QX over the same finite alphabet X , the variational distance (L1 norm) will be denoted by def kPX − QX k = X x∈X |PX (x) − QX (x)| . (1) When optimizing a function of a distribution QX over a probability simplex S(X ), the explicit display of the constraint will be omitted. For example, for a function f (Q), minQ f (Q) will be used instead of minQ∈S(X ) f (Q). The expectation operator with respect to a given distribution, e.g., Q, will be denoted by EQ [·] where the subscript Q will be omitted if the underlying probability distribution is clear from the context. In general, information-theoretic quantities will be denoted by the standard notation [3], with subscript indicating the distribution of the relevant random variables, e.g. HQ (X|Y ), IQ (X; Y ), IQ (X; Y |U ), under Q = QU XY . As an exception, the entropy of X under Q will be denoted by H(QX ). The binary entropy function will be denoted by hb (q) for 0 ≤ q ≤ 1. The conditional Kullback–Leibler divergence between conditional two distributions, e.g., QX|U and PX|U , when 7 averaged with the distribution QU will be denoted by D(QX|U ||PX|U |QU ), and in case that U is degenerated, the notation will be simplified to D(QX ||PX ). The Hamming distance between two vectors, xn ∈ X n and xn ∈ X n will be denoted by dH (xn , xn ). The complement of a set A will be denoted by Ac . For a finite multiset A, the number of distinct elements will be denoted by |A|. The probability of the event A will be denoted by P(A), and its indicator function will be denoted by I(A). . For two positive sequences, {an } and {bn } the notation an = bn , will mean asymptotic equivalence in the exponential scale, that is, limn→∞ 1 n ˙ bn will mean lim supn→∞ log( abnn ) = 0. Similarly, an ≤ so on. The notation an ∼ = bn will mean that limn→∞ an bn 1 n log( abnn ) ≤ 0, and = 1. The ceiling function will be denoted by ⌈·⌉. The notation |t|+ will stand for max{t, 0}. Logarithms and exponents will be understood to be taken to the natural base. Throughout, for the sake of brevity, we will ignore integer constraints on large numbers. For example, ⌈enR ⌉ will be written as enR . For n ∈ N, the set {1, . . . , n} will be denoted by [n]. B. Ordinary Hypothesis-Testing Before getting into the distributed scenario, we will shortly review in this section ordinary HT between a pair of hypotheses. Consider a random variable Z over a finite alphabet Z , whose distribution according to the hypothesis H (respectively, H ) is given by P (respectively, P ). It is common in the literature to refer to H (respectively, H ) as the null hypothesis (respectively, the alternative hypothesis). However, we will refrain from using such terminology, and the two hypotheses will be considered to have an equal stature. Given n i.i.d. observations Z n , a (possibly randomized) detector  φ : Z n → H, H , (2)   def p1 (φ) = P φ(Z n ) = H , (3) has type 1 and type 2 error probabilities9 given by and def p2 (φ) = P [φ(Z n ) = H] . (4) The family of detectors which optimally trades between the two types of error probabilities is given by the NeymanPearson lemma [1, Prop. II.D.1], [3, Th. 11.7.1] by10 9  def  φ∗n,T (z n ) = I P (z n ) ≤ enT · P (z n ) , (5) Also called the false-alarm probability and misdetection probability in engineering applications. Since the two hypotheses are assumed to be discrete, randomized tie-breaking should be used if a given constraint on one of the error probabilities should be matched exactly. Nonetheless, this randomization has no effect on the exponential behavior, which is the focus of this paper (in the distributed scenario). Thus, we will not dwell on randomized detectors too much in what follows. 10 8 where T ∈ R is a threshold parameter. This parameter controls the trade-off between the probability of the two types of error - the larger T is, the type 1 error probability increases and the type 2 error probability decreases, and vice-versa. To describe bounds on the error probabilities of the optimal detector, let us define the hypothesis-testing reliability function [2, Section II] as def D2 (D1 ; P, P ) = min Q: D(Q||P )≤D1 D(Q||P ). (6) For brevity, we shall omit the dependence on P, P as they remain fixed and can be understood from context. As is well known [2, Th. 3], for a given D1 ∈ (0, D(P ||P )), there exists a proper choice of T such that p1 (φ∗n.T ) ≤ exp(−n · D1 ), (7) p2 (φ∗n.T ) ≤ exp [−n · D2 (D1 )] . (8) Furthermore, it is also known that this exponential behavior is optimal [2, Corollary 2], in the sense that if 1 lim inf − log p1 (φ∗n,T ) ≥ D1 n→∞ n (9) 1 lim sup − log p2 (φ∗n,T ) ≤ D2 (D1 ). n n→∞ (10) then It should be noted, however, that the detector (5) and the bounds on its error probability (7)-(8) are exactly optimal for any given n. In fact, in what follows, we will use these relations for n = 1. The function D2 (D1 ) is known to be a convex function of D1 , continuous on (0, ∞) and strictly decreasing prior to any interval on which it is constant [2, Th. 3]. Furthermore, it is known [2, Th. 7] that it can be represented as D2 (D1 ) = sup {−τ · D1 + (τ + 1) · dτ } , (11) τ ≥0 where def dτ = − log " X z∈Z P τ/τ +1 (z)P 1/τ +1 # (z) , (12) is the Chernoff distance between distributions. The representation (11) will be used in the sequel to derive bounds on the reliability of DHT systems. C. Distributed Hypothesis-Testing Let {(Xi , Yi )}ni=1 be n independent copies of a pair of random variables (X, Y ) ∈ (X , Y), where X and Y are finite alphabets. Under H , the joint distribution of (X, Y ) is given by PXY , whereas under H , this distribution is given by P XY . We assume that both probability measures are absolutely continuous with respect to one another, 9 xn encoder detector 010 : : : fn (xn ) H=H̄ 'n (fn (xn ); y n ) yn PXY =P̄XY Figure 2. A DHT system. i.e., PXY ≫ P XY , as well as PXY ≪ P XY ,11 and thus it can be assumed without loss of generality (w.l.o.g.) that supp(PX ) = supp(P X ) = X and supp(PY ) = supp(P Y ) = Y . For brevity, we denote the probability of an event A under H (respectively, H ) by P (A) [respectively, P (A)]. def A DHT system Hn = (fn , ϕn ), as depicted in Fig. 2, is defined by an encoder fn : X n → [mn ], (13) which maps a source vector into an index i = fn (xn ), and a detector (possibly randomized12) ϕn : [mn ] × Y n → {H, H}. (14) The inverse image of fn for i ∈ [mn ], i.e., def fn−1 (i) = {xn ∈ X n : fn (xn ) = i} , is called the bin associated with index i. The rate of Hn is defined as Hn is defined as 1 n (15) log mn , the type 1 error probability of   def p1 (Hn ) = P ϕn (fn (X n ), Y n ) = H , (16) and the type 2 error probability is defined as def p2 (Hn ) = P [ϕn (fn (X n ), Y n ) = H] . (17) In the sequel, conditional error probabilities given an event A will be abbreviated as, e.g., 11   def p1 (Hn |A) = P ϕn (fn (X n ), Y n ) = H|A . (18) This implies that both types of Stein’s exponent for this problem are finite. Randomized encoding can also be defined. In this case, the encoder takes the form fn : X n → S([mn ]), where fn (xn ) is a probability vector whose ith entry is the probability of mapping xn to the index i ∈ [mn ]. In the sequel, we will also use a rather simple form of randomized encoding, which does not require this general definition. There, the source vector xn will be used to randomly generate a new source vector X̃ n , and the latter will be encoded by a deterministic encoder (see the proof of the achievability part of Theorem 6 in Appendix B-A). 12 10 def A sequence of DHT systems will be denoted by H = {Hn }n≥1 . The sequence H is associated with two different exponents for each of the two probabilities defined above. The infimum type 1 exponent of a sequence of systems H is defined by 1 lim inf − log p1 (Hn ), n→∞ n (19) and the supremum type 1 exponent is defined by 1 lim sup − log p1 (Hn ). n n→∞ (20) Analogous exponents can be defined for the type 2 error probability. The reliability function of a DHT system is the optimal trade-off between the two types of exponents achieved by any encoder-detector pair of a given rate R. Specifically, the infimum DHT reliability function is defined by   1 def nR −n·E1 − , (21) E2 (R, E1 ; PXY , P XY ) = sup lim inf − log p2 (Hn ) : ∀n, mn ≤ e , p1 (Hn ) ≤ e n→∞ n H and the supremum DHT reliability function E2+ (R, E1 ; PXY , P XY ) is analogously defined, albeit with a lim sup. For brevity, the dependence on PY |X , P Y |X will be omitted henceforth whenever it is understood from context. III. M AIN R ESULT: B OUNDS ON T HE R ELIABILITY F UNCTION OF DHT To begin the discussion on the reliability function of DHT systems, we note that for a given encoder fn , the form of the optimal detector is just a comparison of likelihoods, as in ordinary HT. Indeed, it readily follows from (5) that the optimal detector has the form  P P  n n nT · n n H xn : fn (xn )=i P XY (x , y ) xn : fn (xn )=i PXY (x , y ) < e n def ∗ ϕn,T (i, y ) =  H otherwise (22) for some T ∈ R which sets the trade-off between the two error probabilities. Hence, the characterization of the DHT reliability function is reduced to finding optimal encoders, to wit, a partition of the alphabet X n into bins, and then, for a given sequence of optimal encoders, finding singleletter expressions for the resulting error exponents, under optimal detection (22). These problems are much more challenging then the characterization of the optimal detector. Nonetheless, just like in distributed compression and channel coding problems mentioned in the introduction, achievability bounds can be derived using random-coding ensembles. Specifically, the main result of this paper, which we next describe, is a random-coding bound and an expurgated bound, obtained under an optimal Neyman-Pearson detector. Before that, we state the trivial converse bound, obtained when X n is not compressed, or alternatively, when R = log|X | (immediately deduced from the discussion in Section II-B). 11 Proposition 1. The supremum DHT reliability function is bounded as E2+ (R, E1 ; PXY , P XY ) ≤ min QXY :D(QXY ||PXY )≤E1 D(QXY ||P XY ). (23) To state our achievability bound, we will need several additional notations. We denote the Chernoff distance between symbols by def dτ (x, x̃) = − log τ/τ +1 1/τ +1 X PY |X (y|x)P Y |X (y|x̃), def 1X dτ (xi , x̃i ). n (24) y∈Y and between vectors by dτ (xn , x̃n ) = n (25) i=1 Further, when (X, X̃) are distributed according to QX X̃ we define the average Chernoff distance as h i def dτ (QX X̃ ) = EQ dτ (X, X̃) , (26) and when X is distributed according to QX , we denote, for brevity, def dτ (QX ) = EQ [dτ (X, X)] . (27) Next, we denote ′ Brc (R, Rb , QU X , τ ) def = min (QU XY ,QU XY ):QU X =QU X , QY =QY ( τ · D(QY |U X ||PY |X |QU X ) + D(QY |U X ||P Y |X |QU X )  + max |IQ (U ; Y ) − Rb |+ , IQ (U, X; Y ) − H(QX ) + R  ) + τ · max IQ (U ; Y ) − Rb , IQ (U, X; Y ) − H(QX ) + R , + (28) and ′′ Brc (R, Rb , QU X , τ ) def = min (QU XY ,QU XY ):QU X =QU X , QU Y =QU Y , IQ (U ;Y )>Rb + |IQ (X; Y |U ) − H(QX ) + R + Rb |+ + τ · IQ (X; Y |U ) − H(QX ) + R + Rb + ( τ · D(QY |U X ||PY |X |QU X ) + D(QY |U X ||P Y |X |QU X ) ) , (29) as well as  ′ def ′′ Brc (R, Rb , QU X , τ ) = min Brc (R, Rb , QU X , τ ), Brc (R, Rb , QU X , τ ) , (30) 12 which are all related to a random-coding based bound on the reliability function. We also denote def Bex (R, QX , τ ) = (τ + 1) · min QX X̃ : QX =QX̃ , HQ (X|X̃)≥R n o dτ (QX X̃ ) + R − HQ (X|X̃) , which is related to an expurgated based bound on the reliability function. Finally, we also denote ) ( def B(R, QX , τ ) = max sup sup QU |X Rb : Rb ≥|IQ (U ;X)−R|+ Brc (R, Rb , QU X , τ ), Bex (R, QX , τ ) . (31) (32) For brevity, arguments such as (R, Rb , QU X , τ ) will sometimes be omitted henceforth. Theorem 2. The infimum DHT reliability function bounded as E2− (R, E1 ; PXY , P XY ) " # o ≥ min sup − τ · E1 + D(QX ||P X ) + τ · D(QX ||PX ) + min (τ + 1) · dτ (QX ), B(R, QX , τ ) . (33) n QX τ ≥0 As hinted by (32), the best of a random-coding bound and an expurgated bound can be chosen for any given input type QX . In the case of a random-coding bound, the achieving scheme is based on quantization and binning. In this respect, for such QX (with H(QX ) > R), the conditional type QU |X is the test channel for quantizing . |Tn (QX )|= enH(QX ) source vectors into one of enRq possible reproduction vectors, where the quantization rate Rq satisfies Rq > R. Then, these reproduction vectors are grouped to bins of size (at most) enRb each, the binning rate Rb satisfies Rb = Rq − R. Both QU |X and Rb may be optimized, separately for any given QX , to obtain the best type 2 exponent. In case the expurgated bound is better than the random-coding bound for QX , the scheme which achieves it is based on binning at rate Rb = |H(QX ) − R|+ , without quantization. We next discuss several implications of Theorem 2. First, simpler bounds, perhaps at the cost of worse exponents, can be obtained by considering two extermal choices. To obtain a binning-based scheme, without quantization, we choose U to be a degenerated random variable (deterministic, i.e., |U |= 1) and Rb = H(QX ) − R. We then get ′ dominates the minimization in (30), and that Brc Brc (R, H(QX ) − R, QU X , τ ) = Brc,b (R, QX , τ ) def = min (QXY ,QXY ):QX =QX , QY =QY (34) ( τ · D(QY |X ||PY |X |QX ) + D(QY |X ||P Y |X |QX ) + |R − HQ (X|Y )|+ + τ · R − HQ (X|Y ) + ) . (35) To obtain a quantization-based scheme, without binning, we choose Rb = 0, and limit QU |X to satisfy R ≥ IQ (U ; X). Second, if the rate is large enough then no loss is expected in the reliability function of DHT. We can deduce from Theorem 2 an upper bound on this no-loss rate, as follows. 13 Corollary 3. Suppose that R is sufficiently large such that B(R, QX , τ ) ≥ dτ (QX ) (36) E2− (R, E1 ) = E2− (∞, E1 ) (37) for all QX ∈ S(X ) and τ ≥ 0. Then, = D2 (E1 ). (38) The proof of this corollary appears in Appendix A. Third, by setting E1 = 0, Theorem 2 yields an achievable bound on Stein’s exponent, as follows. Corollary 4. Stein’s exponent is bounded as E2− (R, 0) ≥ D(PX ||P X ) + sup min {(τ + 1) · dτ (PX ), B(R, PX , τ )} τ ≥0    ≥ min D(PXY ||PX × P Y |X ), D(PX ||P X ) + sup sup  QU |X R : R ≥ I  b b PX ×QU |X (U ;X)−R + (39)    lim Brc (R, Rb , PX × QU |X , τ ) . τ →∞   (40) The first term in (40) can be identified as Stein’s exponent when the rate is not constrained at all. The proof of this corollary also appears in Appendix A, and it seems that no further significant simplifications are possible. It is worth to note, however, that the resulting bound is quite different from the best known bound by Shimokawa, Han and Amari [10, Th. 4.3], [13] (and its refinement in [32], [33]). Fourth, it is interesting to examine the case R = 0. Using analysis similar to the proof of Corollary 4, it is easy to verify that using a binning-based scheme [i.e., substituting (35) in (33) for B(R, QX , τ )] achieves the lower bound E2− (R = 0, E1 ) ≥ min (QXY ,QXY ):QX =QX , QY =QY , D(QXY ||PXY )≤E1 D(QXY ||P XY ). (41) As expected, this is the same type 2 error exponent obtained when y n is also encoded at zero rate, as obtained in [10, Th. 5.4], [11, Th. 6]. For this bound, a matching converse is known [10, Th. 5.5]. When E1 = 0 then QXY = PXY , and then Stein’s exponent is given by E2− (R = 0, E1 = 0) ≥ min QXY :QX =PX , QY =PY D(QXY ||P XY ). (42) In [15, Th. 2] it was determined that this exponent is optimal (even when y n is not encoded and given as sideinformation to the detector). The rest of the paper is mainly devoted to the proof of Theorem 2, and is based on the following methodology, comprised of two steps. We will introduce CD codes, which, in a nutshell, correspond to a single bin of a DHT 14 system. The first step of the proof is the reduction of the DHT reliability problem to the CD reliability problem which will be considered in Section IV. The second step is derivation of single-letter achievable bounds on the reliability of CD codes, and this will be considered in Section V. The bound of Theorem 2 on the DHT reliability function then follow as easy corollary to these results. IV. F ROM D ISTRIBUTED H YPOTHESIS -T ESTING TO C HANNEL -D ETECTION C ODES In this section, we formulate CD codes, and then use them to characterize the reliability of DHT systems. CD codes were considered in [42] for the problem of joint detection and decoding. For this purpose, the code has to be chosen to allow the receiver to detect the channel conditional probability, as well as for transmitting messages, just like an ordinary channel code. In this paper, each quantization cell of a DHT system will be considered and analyzed as a CD code. Since a DHT system is only required to decide on the hypothesis but not on the actual source vector, the error probability of CD codes (for transmitting messages) is irrelevant in this paper. However, this does not imply that all the codewords of CD code can be chosen to be identical (which is optimal in terms of detection performance), since by definition, the members of a quantization cell are different from one another. Hence, in what follows we will define CD codes of a given cardinality, and enforce their codewords to be different from one another. Here too, for brevity, we will denote the probability of an event A under H and H by P (A) and P (A), respectively. The required definitions for CD codes are quite similar to the ones required for DHT systems, but as some differences do exist, we explicitly outline them in what follows. A CD code for a type class QX ∈ Pn (X ) is given by Cn ⊆ Tn (QX ) (where all codewords must be different). An input X n ∈ Cn to the channel is chosen with a uniform distribution over Cn , and sent over n uses of a DMC which may be either PY |X when H is active or P Y |X when H is. The random channel output is given by Y n ∈ Y n . The detector has to decide based on y n whether the DMC conditional probability distribution is PY |X or P Y |X . As for the DHT problem, we assume that PY |X ≪ P Y |X and PY |X ≫ P Y |X . A detector (possibly randomized) for Cn is given by φn : Y n → {H, H}. (43) In accordance, two error probabilities can be defined, to wit, the type 1 error probability and the type 2 error probability   def p1 (Cn , φn ) = P φn (Y n ) = H , def p2 (Cn , φn ) = P [φn (Y n ) = H] . As for the DHT problem, the Neyman-Pearson lemma implies that the optimal detector is given by ) ( X X n n n n nT ∗ n def P Y |X (y |x ) , PY |X (y |x ) < e · φn,T (y ) = I xn ∈Cn xn ∈Cn (44) (45) (46) 15 for some threshold T ∈ R. Let QX ∈ P(X ) be a given type, and let {nl }∞ l=1 be the subsequence of blocklengths such that Pn (QX ) is def not empty. As for a DHT sequence of systems H, a sequence of CD codes C = {Cnl }∞ l=1 is associated with two exponents. The infimum type 1 exponent of a sequence of codes C and detector {φnl }∞ l=1 is defined as lim inf − l→∞ 1 log p1 (Cnl , φnl ), nl (47) and the supremum type 1 exponent is similarly defined, albeit with a lim sup. Analogous exponents are defined for the type 2 error probability. In the sequel, we will construct DHT systems whose bins are good CD codes, for each QX ∈ P(X ). Since to obtain an achievability bound for a DHT system, good performance of CD codes of all types of the source vectors will be simultaneously required, the blocklengths of the components CD codes must match. Thus, the limit inferior definition of exponents must be used, as it assures convergence for all sufficiently large blocklength. For the converse bound, we will use the limit superior definition. For a given type QX ∈ P(X ), rate ρ ∈ [0, H(QX )), and type 1 constraint F1 > 0, we define the infimum CD reliability function as F2− (ρ, QX , F1 ; PY |X , P Y |X )   1 def nl ρ −nl ·F1 lim inf − log p2 (Cnl , φnl ) : ∀l, Cnl ⊆ Tnl (QX ), |Cnl |≥ e , p1 (Cnl , φnl ) ≤ e = sup , (48) l→∞ nl C,{φnl }∞ l=1 and the supremum CD reliability function F2+ (ρ, QX , F1 ; PY |X , P Y |X ) is analogously defined, albeit with a lim sup. For brevity, the dependence on PY |X , P Y |X will be omitted whenever it is understood from context. Thus, the only difference in the reliability function of CD codes from ordinary HT, is that in CD codes the distributions are to be optimally designed under the rate constraint |Cn |≥ enρ . Indeed, for |Cn |= 1 symmetry implies that any xn ∈ Tn (QX ) is an optimal CD code. The detector in this case has no ambiguity regarding the transmitted symbol at any given time point. This, however, does not hold when |Cn |> 1, and ambiguity exists at at least a single time point. This additional uncertainty complicates the operation of the detector, and reduces the reliability function. Basic properties of F2± (ρ, QX , F1 ) are given as follows. Proposition 5. As a function of F1 , F2± (ρ, QX , F1 ) are non-increasing and have both limit from the right and from the left at every point. They have no discontinuities of the second kind and the set of first kind discontinuities (i.e., jump discontinuity points) is at most countable. Similar properties hold as a function of ρ ∈ [0, H(QX )) Proof: It follows from their definition that F2± (ρ, QX , F1 ) are non-increasing in F1 . The continuity statements follow from properties of monotonic functions [45, Th. 4.29 and its Corollary, Th. 4.30] (Darboux-Froda’s theorem). With the above, we can state the main result of this section, which is a characterization of the reliability of DHT systems using the reliability of CD codes. 16 Theorem 6. The DHT reliability functions E2± (R, E1 ) satisfy: • Achievability part: E2− (R, E1 ) ≥ lim  D(QX ||P X ) + F2− (H(QX ) − R, QX , E1 − D(QX ||PX ) + δ) . (49)  D(QX ||P X ) + F2+ (H(QX ) − R + δ, QX , E1 − D(QX ||PX ) − δ) . (50) inf δ↓0 QX ∈P(X ) • Converse part: E2+ (R, E1 ) ≤ lim inf δ↓0 QX ∈P(X ) The achievability and converse part match up to two discrepancies. First, in the achievability (respectively, converse) part the infimum (supremum) reliability function appears. This seems unavoidable, as it is not known if the infimum and supremum reliability functions are equal even for ordinary channel codes [4, Problem 10.7]. Second, the bounds include left and right limits of E2+ (R, E1 ) at rate R and exponent E1 . Nonetheless, due to monotonicity, E2+ (R, E1 ) is continuous function of R and E1 for all rates and exponents, perhaps excluding a countable set (Proposition 5). Thus, for any given (R, E1 ) there exists an arbitrarily close (R̃, Ẽ1 ) such that Theorem 6 holds with δ = 0. The proof of Theorem 6 appears in Appendix B. The achievability part (Appendix B-A) is proved by first constructing a DHT system for source vectors from a a single type class of the source. The bins are generated by permutations of a “good” CD code. The fact that the two hypotheses are memoryless implies that bins generated this way have approximately the same exponents as the original CD code. Furthermore, since type classes are closed under permutations, they can be covered using permutations of a CD code. A covering lemma by Ahlswede [41, Section 6, Covering Lemma 2] shows that in fact such covering method can be effective, in the sense that the required number of permutations is close to minimal number possible (up to the first order in the exponent). This allows to prove that the encoding rate is as required. Ideas from this spirit were used for the DHT problem in [7], as well as for distributed compression [39], [40], and secure lossy compression [46]. Afterwards, all types of the source are considered simultaneously to generate a DHT system for any possible type of the source vector. This requires proving the uniform convergence of the error probabilities to their asymptotic exponential behavior, uniformly over all possible types. The proof of the converse part (Appendix B-B) is based upon identifying a sequence of bins whose size and conditional exponents are close to typical values of the entire DHT system. Such a bin corresponds to a CD code, and thus clearly cannot have better exponents than the ones dictated by the reliability function of CD codes. This restriction is then translated back to bound the reliability of DHT systems. This theorem parallels a similar result of [39], [40] for the reliability functions of distributed compression. By analogy, the reliability function of distributed compression can be characterized by the reliability function of ordinary channel codes (see Fig. 1). The latter can be bounded using well-known classic bounds, such as the random-coding and expurgated achievability bounds, and the sphere-packing, zero-rate, and straight-line converse 17 bounds [4, Ch. 10]. The characterization of Theorem 6 articulates the fact that the reliability of DHT systems is directly connected to the problem of determining the reliability of CD codes. However, even this apparently simpler problem is challenging to solve exactly, and just as in the case of ordinary channel coding, perhaps only lower and upper bounds are at reach. Hence, even though DHT is a source coding problem in nature, for the such reliability functions are known (e.g., Marton’s exponent [47]), Theorem 6 reveals that the reliability of DHT systems depends on the reliability of a channel coding problem, for the such optimal exponents are typically not entirely known, even for the most basic settings. This might manifest the intrinsic difficulty of the DHT problem. Nonetheless, in the next section, we derive concrete bounds on the reliability of CD codes, which, using Theorem 6, lead directly to bounds on the reliability of the DHT problem of Theorem 2. V. B OUNDS ON THE R ELIABILITY OF CD C ODES In the previous section, we have shown that the reliability function of DHT can be directly obtained by the reliability of CD codes. In this section, we derive bounds on the latter, using random-coding arguments. A naive approach would be to randomly draw the codewords independently from some distribution. However, as discussed in the introduction, the best known achievable bounds for DHT systems are obtained using quantization-and-binningbased schemes. As CD codes correspond to a single bin of a DHT system,13 it follows that CD codes can also benefit from adding dependence between the randomly drawn codewords. In more detail, a bin of a DHT system which was constructed by the quantization-and-binning method corresponds to a superposition code. To wit, it will contain a set of reproduction vectors which are sufficiently different from one another, so that the side-information vector y n enables to decide on the true reproduction vector with high reliability, and each reproduction vector is surrounded by a quantization cell, which correspond to all source vectors which are mapped to it in the quantization process. The quantization cell should be sufficiently “small”, such that as long as the detector correctly decodes the reproduction vector, the detection reliability given the reproduction vector is close to the detection reliability given the true vector. A CD code which corresponds to a binning-based scheme will contain source vectors which correspond to a good channel code for the channels PY |X and P Y |X , so that the side-information vector y n can be used to decode the true vector with high reliability. A CD code which corresponds to a quantization-based scheme will look like a quantization cell of a single reproduction vector. The various types of CD codes are depicted in Fig. 3. Since a single bin of a quantization-and-binning DHT system corresponds to a superposition code, it should be drawn from an hierarchical ensemble, which we next define: Definition 7. A fixed-composition hierarchical ensemble for a type QX ∈ Pn (X ) and rate ρ is defined by a conditional type QU |X ∈ Pn (U , QX ), where U is an auxiliary random variable from a finite alphabet U , a cloudcenter rate ρc , and a satellite rate ρs such that ρ = ρc + ρs . A codebook Cn from this ensemble is drawn in 13 Strictly speaking, this is true only when the type of the source vectors is constant within each bin. Since sending the type class of the source vector requires negligible rate, it can be assumed w.l.o.g. that this is indeed the case, as otherwise, one can further partition the bin into type-classes. 18 Quantization and Binning Quantization Binning Figure 3. An illustration of various types of CD codes which pertain to a bin of a DHT system. The grey dots within the large circle represent the members of Tn (QX ). In a quantization based scheme, a bin corresponds to a single reproduction cell of the quantization scheme, and thus all the codewords of the CD code share a common “center” (reproduction vector). In a binning-based scheme, the codewords of the CD code are scattered over the type class with no particular structure. In a quantization-and-binning scheme, the codewords are partitioned to “distant” clouds, where the black dots within one of the small circles represent the satellite codebook pertaining to one of the cloud centers. two stages. First, enρc cloud centers Cc,n are drawn, independently and uniformly over Tn (QU ). Second, for each of the cloud centers un ∈ Cc,n , enρs satellites are drawn independently and uniformly over Tn (QX|U , un ). When considered a random entity, the codebook will be denoted by Cn . Evidently, codewords which pertain to the same cloud are dependent, whereas codewords from different clouds are independent. Further, the ordinary ensemble, in which each codeword is drawn uniformly over Tn (QX ), independently of all other codewords, can be obtained as a special case by choosing U = X and ρc = ρ. On the other hand, when ρs = ρ, then a single14 cloud center is drawn, and all codewords are satellites of this center. This corresponds to the bin of a DHT system which is based only on quantization, without binning. More generally, for source vectors of type QX , a quantization-and-binning scheme of rate R, binning rate Rb , and quantization rate Rq = R + Rb , leads to CD codes of rate ρ = H(QX ) − R, cloud-center rate ρc = Rb and satellite rate ρs = H(QX ) − Rq . When used for a DHT system, the cloud centers Cc,n are the reproduction vectors. The joint distribution of any source vector and its reproduction vector is exactly QU X , and the choice of the test channel QU |X is used to control the distortion of the quantization.15 def In this section, it will be more convenient to use the parameter λ = 1 τ +1 (so that τ = 1−λ λ ). Using this convention, we will use, e.g., dλ instead of dτ for the Chernoff distance. Note also that λ ∈ [0, 1]. To state a random-coding bound on the reliability of CD codes, we denote A′rc (ρ, ρc , QU X , λ) def = min (QU XY ,QU XY ):QU |X =QU |X , QY =QY 14  (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X ), Or a sub-exponential number of cloud centers. Usually in rate-distortion theory, the test channel is used to control the average distortion EQ [d(X, U )] for some additive distortion measure d : X × U → R+ . Here, xn is always chosen from Tn (QX|U , un ) and therefore the distortion between un and xn is constant depending on QU X . 15 19  + λ · max |IQ (U ; Y ) − ρc |+ , IQ (U, X; Y ) − ρ + (1 − λ) · max  IQ (U ; Y ) − ρc + , IQ (U, X; Y ) − ρ  , (51) and A′′rc (ρ, ρc , QU X , λ) def = min (QU XY ,QU XY ):QU |X =QU |X , QU Y =QU Y , IQ (U ;Y )>ρc  (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X ), + λ · |IQ (X; Y |U ) − ρ + ρc |+ + (1 − λ) · IQ (X; Y |U ) − ρ + ρc +  , (52) as well as  def Arc (ρ, ρc , QU X , λ) = min A′rc (ρ, ρc , QU X , λ), A′′rc (ρ, ρc , QU X , λ) . (53) For brevity, when can be understood from context, the dependency on (ρ, ρc , QU X , λ) will be omitted. Our randomcoding bound is as follows. Theorem 8. The infimum CD reliability function is bounded as ( " #) 1−λ 1 − F2 (ρ, QX , F1 ) ≥ max − · F1 + · min dλ (QX ), sup Arc (ρ, ρc , QU X , λ) . sup 0≤λ≤1 λ λ QU |X ρc : ρc ≥|ρ−HQ (X|U )|+ (54) The proof of Theorem 8 appears entirely in Appendix C, and here we provide a brief outline. The main idea of the proof is that when a CD code is used, the output distribution under H is given by the induced distribution (C ) def PY nn (y n ) = 1 X PY |X (y n |xn ). |Cn | n (55) x ∈Cn (C ) (Cn ) (C ) (Cn ) Based on the output vector y n , the detector should distinguish between two distributions PY nn and P Y n [the latter defined similarly to (55)]. This corresponds to an ordinary HT problem between the PY nn and P Y n , and therefore the exponents of this HT problem are simply given by (7) and (8) (when letting Z = Y n ). In turn, using (C ) the characterization (11), the reliability function can be expressed using the Chernoff distance between PY nn and (Cn ) P Y n . As the CD code which maximizes this Chernoff distance is difficult to find, we turn to a random-coding approach, and evaluate its ensemble average, namely    X h i1−λ h iλ  (Cn ) (C ) PY nn (y n ) · P Y n (y n ) E .  n n  (56) y ∈Y In the same spirit, we will evaluate the number of unique codewords of a randomly chosen CD code. In fact, we will show that under proper conditions, it is asymptotically equal to the number of drawn codewords, i.e., enρ . Then, a sequence of codebooks, such that each codebook has effectively enρ unique codewords, and the Chernoff distance is typical to the ensemble average will be considered. For this sequence, it will then be shown that removing 20 duplicates, i.e., keeping only a single instance of codewords which were drawn more than once (thus making it a valid CD code), may only cause a negligible loss in the achieved exponent. This will prove a single-letter random-coding bound on the reliability function. We make the following comments: 1) Loosely speaking, in (53), the exponent A′rc corresponds to the contribution to the error probability from codewords which belong to different cloud centers, whereas the exponent A′′rc corresponds to the contribution to the error probability from codewords which belong to the same cloud center as the transmitted codeword. Thus, for a given rate ρ, A′rc is monotonically non-increasing with ρc , while A′′rc is monotonically nondecreasing with ρc (or, monotonically non-increasing with ρs ). The cloud-center rate ρc and the test channel QU |X therefore should be chosen to optimally balance between these two contributions to the error probability. 2) For a given QU |X , the random-coding analysis of the ensemble average of the Chernoff distance is based on the type-enumeration method [48, Sec. 6.3], which leads to ensemble-tight bounds in many problems. While we cannot rigorously claim ensemble-tightness here, our analysis still avoids any use of bounds such as Jensen’s inequality. Thus, we conjecture that our bounds are ensemble-tight, and cannot be improved. 3) In comparison to [42], we have generalized the random-coding analysis of the detection error exponents to hierarchical ensembles, and also obtained simpler expressions using the ensemble average of the exponent of the Chernoff distance. 4) In fact, a stronger claim than the one appears in Theorem 8 can be made. It can be shown that there exists a single sequence of CD codes {Cn∗l }∞ l=1 such that lim inf − l→∞ 1 log min p2 (Cn∗ l , φ∗nl ,T ) nl T : p1 (Cn∗ l ,φ∗nl ,T )≤e−nF1 ( #) " 1−λ 1 ≥ max − , (57) Arc · F1 + · min dλ (QX ), sup sup 0≤λ≤1 λ λ QU |X ρc : ρc ≥[ρ−HQ (X|U )]+ simultaneously for all F1 . Thus, when using such a CD code, the operating point along the trade-off curve between the two exponents can be determined solely by the detector, and can be arbitrarily changed from block to block. For details regarding the proof of this claim, see Remark 18. Next, we state our expurgated exponent, and to this end we denote def Aex (ρ, QX , λ) = min QX X̃ : QX =QX̃ , IQ (X;X̃)≤ρ n o dλ (QX X̃ ) + IQ (X; X̃) − ρ . Theorem 9. The infimum CD reliability function is bounded as   1 1−λ − · F1 + · min {dλ (QX ), Aex (ρ, QX , λ)} . F2 (ρ, QX , F1 ) ≤ max − 0≤λ≤1 λ λ (58) (59) Proof of Theorem 9: As was shown in [42, Appendix E], for any δ > 0 and all n sufficiently large, there 21 exists a CD code Cn∗ (of rate ρ) such that i1−λ h iλ X h (C ∗ ) (C ∗ ) n n PY n (y ) P Y nn (y n ) ≤ exp [−n · min {dλ (QX ), Aex (ρ, QX , λ)}] . (60) y n ∈Y n Substituting this bound to the characterization (11), taking n → ∞ and δ ↓ 0 completes the proof of the theorem. We conclude with two comments: 1) A similar expurgated bound can be derived for hierarchical ensembles. However, when optimizing the rates (ρs , ρc ) for this expurgated bound, it turns out that choosing ρs = 0 is optimal. Thus, the resulting bound exactly equals the bound of Theorem 9, which corresponds to an ordinary ensemble. 2) Since the expurgated exponent only improves the random-coding exponent of the ordinary ensemble (which is inferior in performance to the hierarchical ensemble), it is anticipated that expurgation does not play a significant role in this problem, compared to the channel coding problem. This might be due to the fact that the aim of expurgation is to remove codewords which have “close” neighbors, while this is not actually required in the DHT problem. This can also be attributed to the bounding technique of the expurgated bound, which is based on pairwise Chernoff distances. After deriving the bounds on the reliability of CD codes, we return to the DHT problem, and conclude the section with a short proof of Theorem 2. Proof: Up to the arbitrariness of δ > 0, in Theorem 6 we have obtained E2− (R, E1 ) = inf QX ∈P(X )  D(QX ||P X ) + F2− (H(QX ) − R, QX , E1 − D(QX ||PX )) . (61) Further, the random-coding bound of Theorem 8 and the expurgated bound of Theorem 9 both imply that F2− (H(QX ) − R, QX , E1 − D(QX ||PX )) ( " 1−λ 1−λ · E1 + · D(QX ||PX ) ≥ max max − 0≤λ≤1 λ λ )# ( 1 + · min dλ (QX ), sup Arc (H(QX ) − R, Rb , QU X , λ) , sup λ QU |X Rb : Rb ≥|IQ (U ;X)−R|+ " 1−λ 1−λ max − · E1 + · D(QX ||PX )+ 0≤λ≤1 λ λ #) 1 + · min {dλ (QX ), Aex (H(QX ) − R, QX , λ)} λ ( 1−λ 1−λ max − = sup sup · E1 + · D(QX ||PX ) λ λ QU |X Rb : Rb ≥|IQ (U ;X)−R|+ 0≤λ≤1  ) 1 . + · min dλ (QX ), max {Arc (H(QX ) − R, Rb , QU X , λ) , Aex (H(QX ) − R, QX , λ)} λ (62) (63) 22 def 1−λ λ The bound of Theorem 2 is obtained by substituting in (63) in (61), while changing variables from λ to τ = and from ρc to Rb , as well as using the definitions of Brc (30) and Bex (31) and B (32). VI. C OMPUTATIONAL A SPECTS AND A N UMERICAL E XAMPLE The bound of Theorem 2 is rather involved, and therefore it is of interest to discuss how to compute it efficiently. ′ and B ′′ for a given (R, R , Q Evidently, the main computational task is the computation of Brc U X , τ ). To this end, b rc ′ and B ′′ are strictly convex functions of (Q it can be seen that the objective functions of both Brc Y |U X , QY |U X ) (as rc ′ is a convex set (only has linear constraints) PY |X ≪≫ P Y |X was assumed).16 Furthermore, the feasible set of Brc ′ is a convex optimization problem, which can be solved efficiently [49]. However, and thus the computation of Brc ′′ is not convex, due to the additional constraint I (U ; Y ) > R beyond the linear constraints. the feasible set of Brc Q b Nevertheless, the value of Brc can be computed efficiently, by solving only convex optimization problems, in the following algorithm: ′ , and let the optimal value be v ′ . 1) Solve the optimization problem (28) defining Brc ′′ , but without the constraint I (U ; Y ) > R (this is a convex 2) Solve the optimization problem (29) defining Brc Q b ∗ optimization problem). Let the solution be (Q∗U XY , QU XY ) and the optimal value be v ′′ . If IQ∗ (U ; Y ) > Rb ′′ = v ′′ , and otherwise, set B ′′ = ∞. then set Brc rc 3) The result is Brc = min{v ′ , v ′′ }. This algorithm is correct due to the following argument. It is easily verified that if IQ∗ (U ; Y ) > Rb then the ′′ = v ′′ . However, if this constraint IQ (U ; Y ) > Rb is inactive, and therefore can be omitted. Thus, in this case Brc is not the case, then the solution must be on the boundary, i.e., must satisfy IQ (U ; Y ) = Rb . This is because the ′′ is a strictly convex function. In the latter case, it can be easily seen that B ′ ≤ B ′′ , and as B is objective in Brc rc rc rc ′′ = ∞ can be set. Computing B (R, Q , τ ) of (31) is a convex optimization the minimum between the two, Brc ex X problem, over QX̃|X . Given the value of Brc , the next step is to optimize over QU |X and ρc . While this should be done exhaustively,17 any specific choice of QU |X and ρc (or a restricted optimization set for them) leads to an achievable bound on E2− (R, E1 ). It should be mentioned, however, that it is not clear to us how to apply standard cardinality-bounding techniques (based on the support lemma [4, p. 310]) to bound |U | in this problem. Thus, in principle, the cardinality of U is unrestricted, improved bounds are possible. Finally, both τ and QX should be optimized, which is feasible when X is not very large and exhausting the simplex S(X ) in search of the minimizer QX is possible. Furthermore, as we have seen in Corollary 4, when only Stein’s exponent is of interest, i.e., E1 = 0, the minimal value in (33) must be attained for QX = PX . Thus, there is no need to minimize over QX ∈ S(X ), but rather only on QX = PX . We can also set τ → ∞ if the weak version 16 Indeed, the divergence terms and I(U, X; Y ) are convex in QY |U X . The term I(U ; Y ) is also convex in QY |U X , as a composition of a linear function which maps QY |U X to QY |U and the mutual information I(U ; Y ). The pointwise maximum of two convex functions is also a convex function (note that |f (t)|+ = max{0, f (t)}). 17 Or by any other general-purpose global optimization algorithm. 23 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 Figure 4. A binary example. (40) of Corollary 4 is used as a bound. More generally, the minimizer of QX must satisfy D(QX ||PX ) ≤ E1 , and this can decrease the size of the feasible set of QX whenever the required E1 is not very large. A simple example for using the above methods to compute bounds on the DHT reliability function is given as follows. Example 10. Consider the case X = Y = {0, 1}, and PX = P X = (1/2, 1/2), where PY |X and P Y |X are binary symmetric channels with crossover probabilities 10−1 and 10−2 , respectively. We have used an auxiliary alphabet of size |U |= |X |+1 = 3 since this cardinality suffices for the best known bounds [13], and due to the symmetry in the problem, we have only optimized over symmetric QU |X . Achievable bounds on the reliability of the DHT are shown in Fig. 4 for two different rates. VII. C ONCLUSION AND F URTHER R ESEARCH In this paper, we have considered the trade-off between the two types of error exponents of a DHT system, with full side-information. We have shown that its reliability is intimately related to the reliability of CD codes, and thus the latter simpler problem should be considered. Achievable bounds on the reliability of CD codes were derived, under the optimal Neyman-Pearson detector. There are multiple directions in which our understanding of the problem can be broadened: 1) Variable-rate coding: As was noted in the past, the reliability of compression systems can be increased when variable-rate is allowed, either with an average rate constraint [39], or under excess-rate exponent constraint [40]. It is straightforward to use the techniques developed there to the DHT problem studied here (see also the informal discussion in [46, Appendix]). 24 2) Computation of the bounds: As discussed in the paper, the main challenge in computing the random-coding bound computation is the optimization over the test channel QU |X . First, deriving cardinality bounds on the auxiliary random variable alphabet U is of interest. Second, finding an efficient algorithm to optimize the test channel, perhaps an alternating-maximization algorithm in the spirit of the Csiszár-Tusnády [50] and the Blahut-Arimoto algorithms [51], [52], [53], could be developed for this problem. As was noted in [16], [21], the Stein exponent of the DHT problem of testing against independence is identical to the information bottleneck problem [54], for which such alternating-maximization algorithm was developed. 3) Converse bounds: While we have shown that to obtain converse bounds on the reliability of DHT systems, it suffices to obtain converse bounds on the reliability of CD codes, no concrete bounds were derived. To obtain converse bounds which explicitly depend on the rate (in contrast to Proposition 1), two challenges are visible. First, it is tempting to conjecture that the Chernoff characterization (11) characterizes the reliability of CD codes, in the sense that18 F2+ (ρ, QX , F1 ; PY |X , P Y |X ) = lim sup sup max n ρ l→∞ Cnl ⊆Tnl (QX ): |Cnl |≥e l τ ≥0      i h i h 1/1+τ  τ/1+τ X 1 (Cn ) (Cn ) , (64) −τ · F1 − (τ + 1) · log PY nll (y nl ) · P Y nll (y nl )    nl nl nl y ∈Y just as a similar quantity was used to derive the random-coding and expurgated bounds. Second, even if this conjecture holds, the value of   X h  y nl ∈Y nl  iτ/1+τ h i1/1+τ  (Cn ) (Cn ) PY nll (y nl ) · P Y nll (y nl )  (65) should be lower bounded for all CD codes whose size is larger than enρ . As this term can be identified as a Rényi divergence [55], [56], the problem of bounding its value is a Rényi divergence characterization. This problem seems formidable, as the methods developed in [8] for the entropy characterization problem rely heavily on the chain rule of mutual information; a property which is not naturally satisfied by Rényi entropies and divergences. Hence, the problem of obtaining a non-trivial converse bound for the reliability of DHT systems with general hypotheses and a positive encoding rate remains an elusive open problem. 4) Rate constraint on the side-information: The reliability of a DHT systems in which the side-information vector y n is also encoded at a limited rate should be studied under optimal detection. Such systems will naturally lead to multiple-access CD codes, as studied for ordinary channel coding (see [57], [58] and references therein). However, for such a scenario, it was shown in [33] that the use of linear codes (a-la Körner-Marton coding) dramatically improves performance. Thus, it is of interest to analyze DHT systems with both linear codes and optimal detection. However, it is not yet known how to apply the type-enumeration method, used here for analysis of optimal detection, to linear codes. Hence, either the type-enumeration method should be refined, 18 As usual {nl }∞ l=1 is the subsequence of blocklength such that Tn (QX ) is not empty. 25 or an alternative approach should be sought after. 5) Generalized hypotheses: Hypotheses regarding the distributions of continuous random variables, or regarding the distributions of sources with memory can be considered. Furthermore, the case of composite hypotheses, in which the distribution under each hypotheses is not exactly known (e.g., belongs to a subset of a given parametric family), and finding universal detectors which operate as well as the for simple hypotheses can also be considered. For preliminary results along this line see [15], [42]. ACKNOWLEDGMENT Discussions with N. Merhav are acknowledged with thanks. A PPENDIX A P ROOFS OF C OROLLARIES TO T HEOREM 2 Proof of Corollary 3: Suppose that in the bound of Theorem 2, the inner minimization is dominated by (τ + 1) · dτ (QX ) for all QX . Then, the bound (33) reads E2− (R, E1 ; PXY , P XY )  ≥ min sup −τ · E1 + D(QX ||P X ) + τ · D(QX ||PX ) + (τ + 1) · dτ (QX ) QX τ ≥0 (A.1) (a) = min sup min { − τ · E1 + D(QX ||P X ) + τ · D(QX ||PX ) QX τ ≥0 QY |X + τ · D(QY |X ||PY |X |QU X ) + D(QY |U X ||P Y |X |QU X )}  = min sup min −τ · E1 + D(QX ||P X ) + τ · D(QXY ||PXY ) + D(QXY ||P XY ) QX τ ≥0 QY |X (b)  = min sup τ · [D(QXY ||PXY ) − E1 ] + D(QXY ||P XY ) QXY τ ≥0 = min QXY :D(QXY ||PXY )≤E1 D(QXY ||P XY ), (A.2) (A.3) (A.4) (A.5) where (a) follows since as can be easily shown that (see (C.31) in the proof of Lemma 16 and the discussion surrounding it)   (τ + 1) · dτ (QX ) = min τ · D(QY |U X ||PY |X |QU X ) + D(QY |U X ||P Y |X |QU X ) QY |U X   = min τ · D(QY |X ||PY |X |QX ) + D(QY |X ||P Y |X |QX ) , QY |X (A.6) (A.7) and (b) follows since the objective function is linear in τ (and hence concave) and convex in QY |X , and therefore the minimization and maximization order can be interchanged [59]. Thus, the achievability bound of Theorem 2 coincides with the converse bound of Proposition 1, the latter obtained when the rate of the DHT system is not constrained at all and given by the reliability function of ordinary HT problem between PXY and P XY . Proof of Corollary 4: It can be seen that the outermost minimum in (33) must be attained for QX = PX . Intuitively, since we are only interested in negligible type 1 exponent, any event with QX 6= PX has exponentially 26 decaying probability exp[−nD(QX ||PX )], and does not affect the exponent. More rigorously, if QX 6= PX then by taking τ → ∞ the objective function becomes unbounded. Hence (39) immediately follows. Further simplifications are possible if the bound is weakened by ignoring the expurgated term, i.e., setting Bex (R, QX , τ ) = 0 in (32). In this case, since   (τ + 1) · dτ (PX ) = min τ · D(QY |X ||PY |X |PX ) + D(QY |X ||P Y |X |PX ) , QY |X (A.8) ′ , B ′′ [see (28) and (29)], it [see (A.7)], and since τ only multiplies positive terms in the objective functions of Brc rc is evident that the supremum in (39) is obtained as τ → ∞. Hence, the supremum and minimum in (39) can be interchanged to yield the bound E2− (R, 0) ≥ D(PX ||P X ) + min    lim (τ + 1) · dτ (PX ), sup sup  QU |X R : R ≥ I τ →∞ PX ×QU |X (U ;X)−R b b    (a) = min D(PXY ||PX × P Y |X , D(PX ||P X ) + sup sup  QU |X R : R ≥ I  b b PX ×QU |X (U ;X)−R + +    lim Brc (R, Rb , PX × QU |X , τ ) τ →∞   (A.9)    lim Brc (R, Rb , PX × QU |X , τ ) , τ →∞   (A.10) where (a) follows since   sup(τ + 1) · dτ (PX ) = sup min τ · D(QY |X ||PY |X |PX ) + D(QY |X ||P Y |X |PX ) τ ≥0 τ ≥0 QY |X   = sup min sup τ · D(QY |X ||PY |X |PX ) + D(QY |X ||P Y |X |PX ) τ ≥0 QY |X τ ≥0 = D(PY |X ||P Y |X |PX ). (A.11) (A.12) (A.13) A PPENDIX B P ROOF OF T HEOREM 6 A. Proof of the Achievability Part We begin by showing that an asymptotically good CD code must have a large subcode, such that the error probabilities of any of the subcodes of the latter are not much worse than the corresponding probabilities of the entire code. Lemma 11. Let Cn be a CD code, and φn be a detector. Then, there exists a CD code C˜n with |C˜n |≥ |Cn |/3 which 27 satisfies the following: For any subcode C n ⊆ C˜n , there exists a of detector φn such that pi (C n , φn ) ≤ 3 · pi (Cn , φn ) (B.1) holds for both i = 1, 2. Proof: Note that the error probabilities in (44) and (45) are averaged over the transmitted codeword X n ∈ Cn . We first prove that by expurgating enough codewords from a codebook with good average error probabilities, a codebook with maximal (over the codewords) error probabilities can be obtained (for both types of error). The proof follows the standard expurgation argument from average error probability to maximal error probability (which in turn follows from Markov’s inequality). Denoting the conditional type 1 error probability by19 we may write   def p1 (Cn , φn |X n = xn ) = P φn (Y n ) = H|X n = xn , p1 (Cn , φn ) = X xn ∈Cn = (B.2) P (X n = xn ) · p1 (Cn , φn |X n = xn ) 1 X p1 (Cn , φn |X n = xn ) . |Cn | n (B.3) (B.4) x ∈Cn Thus, at least 2/3 of the codewords in xn ∈ Cn satisfy p1 (Cn , φn |X n = xn ) ≤ 3 · p1 (Cn , φn ). (B.5) Using a similar notation for the conditional type 2 error probability, and repeating the same argument, we deduce that there exists C˜n ⊂ Cn such that |C˜n |≥ |Cn |/3 and both (B.5) as well as p2 (Cn , φn |X n = xn ) ≤ 3 · p2 (Cn , φn ), (B.6) hold for any xn ∈ C˜n . Let us now consider any C n ⊆ C˜n . For the code C n , the detector φn is possibly sub-optimal, and thus might be improved. Using the standard Neyman-Pearson lemma [1, Prop. II.D.1], one can find a detector φn (perhaps randomized) to match any prescribed type 1 error probability value, which is optimal in the sense that if any other detector φ̂n satisfies p1 (C n , φ̂n ) ≤ p1 (C n , φn ) then p2 (C n , φ̂n ) ≥ p2 (C n , φn ). Specifically, let us require that p1 (C n , φn ) = 3 · p1 (Cn , φn ), (B.7) and choose φ̂n = φn . Then, as (B.5) holds for any xn ∈ C n , p1 (C n , φn ) = X xn ∈C n P (X n = xn ) · p1 C n , φn |X n = xn  (B.8) 19 Note that conditioned on X n = xn , p1 (Cn , φn |X n = xn ) depends on the code Cn only if the detector φn depends on the code. In this lemma, the detector φn is arbitrary, and the use of this notation is therefore just for the sake of consistency. 28 ≤ 3 · p1 (Cn , φn ) (B.9) = p1 (C n , φn ), (B.10) and as φn is optimal in the Neyman-Pearson sense, (B.6) implies that p2 (C n , φn ) ≤ p2 (C n , φn ) X P (X n = xn ) · p2 (Cn , φn |X n = xn ) = xn ∈C (B.11) (B.12) n ≤ 3 · p2 (Cn , φn ). (B.13) The result follows from (B.7) and (B.13). Note that C n = C˜n is a valid choice. Next, we construct optimal DHT systems for a single type class of the source. Lemma 12. Let δ > 0 and QX ∈ P(X ) be given, and let {nl } the subsequence of blocklengths such that Tn (QX ) is not empty. Further, let C be a sequence of CD codes of type QX and rate ρ, and {φnl }∞ l=1 be a sequence of detectors. Then, there exist a sequence of DHT systems H of rate H(QX ) − ρ such that lim inf − l→∞ 1 1 log pi [Hnl |X nl ∈ Tnl (QX )] ≥ lim inf − log pi (Cnl , φnl ) − δ, l→∞ nl nl (B.14) holds for both i = 1, 2. Proof: We only need to focus on of xn ∈ Tn (QX ). For notational simplicity, let us assume that n is always such that Tn (QX ) is not empty. Let us first extract from C the sequence of CD codes C˜ whose existence is assured by Lemma 11. The rate of C˜ is chosen to be larger than ρ − δ (for all sufficiently large n), and for any given codeword, the error probability of each type is assured to be up to a factor of 3 of its average error probability. Now, for a given permutation π of [n], we define the permutation of xn = (x1 , x2 , . . . , xn ) as def  def π(xn ) = xπ(1) , xπ(2) , . . . , xπ(n) , (B.15) and for a set Dn = {xn (0), . . . , xn (|Dn |−1)}, we define def π(Dn ) = {π(xn (0)), . . . , π(xn (|Dn |−1))} . (B.16) n As C˜n ∈ Tn (QX ), clearly so is π(C˜n ) ∈ Tn (QX ). Thus, there exists a set of permutations {πn,i }κi=i such that κn [ i=1 πn,i (C˜n ) = Tn (QX ). By a simple counting argument, the minimal number of permutations required κn is at least (B.17) |Tn (QX )|/|C˜n |. This is achieved when the permuted sets are pairwise disjoint, i.e., πn,i (C˜n ) ∩ πn,i′ (C˜n ) = φ, for all i 6= i′ . While this property is difficult to assure, Ahlswede’s covering lemma [41, Section 6, Covering Lemma 2] implies that up to the first order in the exponent, this minimal number can be achieved. To wit, there exists a set of permutations 29 κ∗ ∗ } n such that {πn,i i=1 κ∗n ≤ |Tn (QX )| nδ en[H(QX )+δ] nδ ·e ≤ · e = en[H(QX )−ρ+3δ] , en(ρ−δ) |C˜n | (B.18) ∗ is the identity permutation, and thus C ∗ = C˜ . Further, for all n sufficiently large. W.l.o.g., we assume that πn,1 n n,1 for 2 ≤ i ≤ κ∗n , let ∗ Cn,i     i−1 [ def ∗ ∗ πn,j (C˜n ) . = πn,i (C˜n )\   (B.19) j=1 ∗ is the permutation π ∗ of the code C˜ , excluding codewords which belong to a permutation In words, the code Cn,i n n,i ∗ ∗ }κn forms a disjoint partition of T (Q ). Moreover, recall that from Lemma of C˜n with smaller index. Thus, {Cn,i n X i=1 11, for any given def ∗ −1 ∗ C n,i = [πn,i ] (Cn,i ) ⊆ C˜n (B.20) (where π −1 is the inverse permutation of π ), one can find a detector φn,i such that p1 (C n,i , φn,i ) = 3 · p1 (Cn , φn ), (B.21) p2 (C n,i , φn,i ) ≤ 3 · p2 (Cn , φn ). (B.22)  ∗ −1 n  def φ∗n,i (y n ) = φn,i [πn,i ] (y ) . (B.23) and Now, let Due to the memoryless nature of the hypotheses, the permutation does not change the probability distributions. Indeed, for an arbitrary CD code Cn′ , a detector φ′n and a permutation π , p1 (Cn′ , φ′n ) = X P (X n = xn ) · p1 (xn , φ′n ) X P (X n = xn ) · xn ∈Cn′ = xn ∈Cn′ 1 X = ′ |Cn | n ′ x ∈Cn 1 X = ′ |Cn | n ′ x ∈Cn 1 = ′ |Cn | yn : yn : X X (B.24) y n : φ′n (y n )=H P (Y n = y n |X n = xn ) (B.25) X P (Y n = y n |X n = xn ) (B.26) X P [Y n = π(y n )|X n = π(xn )] (B.27) φ′n (y n )=H φ′n (y n )=H xn ∈π −1 (Cn′ ) yn : X φ′n [π −1 (y n )]=H  = p1 π −1 (Cn′ ), φ′n,π . P (Y n = y n |X n = xn ) (B.28) (B.29) 30 def where φ′n,π (y n ) = φ′n [π −1 (y n )]. Hence, ∗ p1 (C n,i , φn,i ) = p1 (Cn,i , φ∗n,i ), (B.30) ∗ , φ∗n,i ). p2 (C n,i , φn,i ) = p2 (Cn,i (B.31) and κ∗ ∗ } n will serve as the bins of f , and detectors We thus construct Hn = (fn , ϕn ) as follows. The codes {Cn,i n i=1 κ∗ n as the decision function, given that the bin index is i. As said above, only xn ∈ Tn (QX ) will be encoded. {φ∗n,i }i=1 ∗ , and by f (xn ) = 0 More rigorously, the encoding of xn ∈ Tn (QX ) is given by fn (xn ) = i whenever xn ∈ Cn,i n whenever xn 6∈ Tn (QX ). Clearly, the rate of the code is less than 1 log κ∗n ≤ H(QX ) − ρ + 3δ, n (B.32) for all n sufficiently large. The detector ϕn is given by ϕn (i, y n ) = φ∗n,i (y n ). The conditional type 1 error probability of this DHT system is given by κ∗ n  X   P ϕn (fn (X ), Y ) = H|X ∈ Tn (QX ) = P [fn (X n ) = i] · P ϕn (fn (X n ), Y n ) = H|fn (X n ) = i (B.33)  n n n i=1 ∗ = (a) = (b) = κn X i=1 κ∗n (B.34) X ∗ P [fn (X n ) = i] · p1 (Cn,i , φ∗n,i ) (B.35) X P [fn (X n ) = i] · p1 (C n,i , φn,i ) (B.36) i=1 κ∗n i=1 (c)   P [fn (X n ) = i] · P φ∗n,i (Y n ) = H|fn (X n ) = i ≤ 3 · p1 (Cn , φn ), (B.37) ∗ , (b) follows where (a) follows because given fn (X n ) = i the source vector X n is distributed uniformly over Cn,i from (B.30), and (c) follows from (B.21). Similarly, the conditional type 2 error probability is upper bounded as P [ϕn (fn (X n ), Y n ) = H|X n ∈ Tn (QX )] ≤ 3 · p2 (Cn , φn ). (B.38) The factor 3 in the error probabilities is negligible asymptotically. The prove the achievability part of Theorem 6, the DHT system of Lemma 12 can be used to encode a single type class. It still remains, however, to assure convergence of the error probabilities to their exponential bounds, uniformly over all types. In other words, the DHT system of Lemma 12 is designed for a given QX , and as such, the minimal blocklength for which the error probabilities are close to their exponential bounds may depend on QX . However, as the blocklength n gets larger, source vectors types which did not exist at smaller blocklengths are also emitted by the source. For such type class, the blocklength required to achieve good error exponents may be much 31 larger. It may be that these types dominate the type 1 error probability, and thus prevent convergence. The method to solve this problem (see also [40], [46]) is to define a finite grid of types. CD codes are constructed only for types in the grid, and as the grid is finite, uniform convergence is assured (by taking the maximum required blocklength over all types within the grid). Now, if the type of xn belongs to the grid, it can be encoded using the CD which pertains to its type. Otherwise, xn is slightly modified to a different vector X̃ n , where the type of the latter does belong to the grid. Then, X̃ n is encoded using the CD code which pertain to its type. Since the CD codes are designed for the pair (X n , Y n ), rather than for (X̃ n , Y n ), the side-information vector Y n is also modified to a vector Ỹ n , using additional information sent from the encoder. To analyze the effect of this modification on the error probabilities, we will need the following partial mismatch lemma. def n ) is Lemma 13. Let Cn be a CD code, and φn a detector. For d = δn ∈ [n], assume that Ỹ n = (Ỹ1d , Ỹd+1 drawn as follows: Given xn ∈ Cn , under the hypothesis H , Ỹ n = Y n ∼ PY |X (·|xn ), and under the hypothesis H , n ∼P n Ỹ1d ∼ PY |X (·|xd1 ) and Ỹd+1 Y |X (·|xd+1 ). Then, i h P φn (Ỹ n ) = H = p1 (Cn , φn ), (B.39) h i P φn (Ỹ n ) = H ≤ enΩδ · p2 (Cn , φn ) (B.40) and where def Ω = max x∈X , y∈Y log PY |X (y|x) P Y |X (y|x) < ∞. (B.41) Proof: Fix some blocklength n. The statement regarding the type 1 error probability is trivial. As for the type 2 error probability, we will show that the “wrong” distribution of Y n in the first d coordinates does not change likelihoods and probabilities significantly. Indeed, conditioning on Y1d = y1d i i h h P φn (Ỹ n ) = H|Ỹ1d = y1d = P φn (Y n ) = H|Y1d = y1d . (B.42) Then, since,   1 X P Ỹ1d = y1d = P Y |X (y1d |xd1 ) |Cn | n x ∈Cn 1 X PY |X (y1d |xd1 ) ≤ edΩ · |Cn | n x ∈Cn   dΩ = e · P Y1d = y1d , (B.43) (B.44) (B.45) we obtain i  h i  h X P φn (Ỹ n ) = H = P Ỹ1d = y1d · P φn (Ỹ n ) = H|Ỹ1d = y1d y1d ∈Y d (B.46) 32 ≤ X y1d ∈Y d  i  h edΩ · P Y1d = y1d · P φn (Y n ) = H|Y1d = y1d = edΩ · p2 (Cn , φn ). (B.47) (B.48) We will also use the following simple lemma: Lemma 14. Let QX , Q̃X ∈ Pn (X ) and assume that ||QX − Q̃X ||= min x̃n ∈Tn (Q̃X ) 2d n where d > 0. If xn ∈ Tn (QX ) then dH (x̃n , xn ) ≤ d. (B.49) We are now ready to prove the achievability part. Since a similar method was applied in [40, Appendix A, proof of Th. 5][46, Sec. VI.C], we will shorten the details of some of the arguments. Proof of the achievability part of the Theorem 6: The encoded message begins with a description of the type QX of X n . Then, to encode a given X n ∈ Tn (QX ), a DHT system which is based on good CD codes, as in Lemma 12, will be used. Given ǫ > 0 (to be specified later), we choose the grid as Pn0 (X ), and denote20 def Φǫ (QX ) = arg min ||QX − Q̃X ||. (B.50) Q̃X ∈Pn0 (X ) The blocklength n0 is chosen to be sufficiently large, such that Φǫ (QX ) ≤ ǫ 2 for any QX ∈ P(X ). Note that Φǫ (QX ) = QX whenever QX ∈ Pn0 (X ). We shall only consider blocklength of the form n = kn0 for some k ∈ N+ . For n = kn0 + k0 with 1 ≤ k0 < n0 , the last k0 symbols of (X n , Y n ) can be simply ignored by both the encoder and detector. The loss in exponents is a factor n kn0 , which is negligible for sufficiently large k.21 For n = kn0 , Pn0 (X ) ⊂ Pn (X ), and thus if QX is the type of the source vector xn , then either QX ∈ Pn0 (X ) or QX ∈ Pn (X )\Pn0 (X ). In the former case, (X n , Y n ) remains unchanged, and in the latter case, it is modified as follows. First, the encoder will randomly generate X̃ n ∈ Tn (Φǫ (QX )) with a uniform distribution over the set n x̃n ∈ Tn (Φǫ (QX )) : dH (xn , x̃n ) = nǫ o . 4 (B.51) Lemma 14 assures that this set is not empty. Further, due to the permutation symmetry of type classes, if X n is distributed uniformly over Tn (QX ) then X̃ n is distributed uniformly over Tn (Φǫ (QX )). Now, as xn is altered to X̃ n = x̃n , a similar change has to be made to the side-information vector. Indeed, given xn , the actual distribution of Y n is either PY |X (·|xn ) or P Y |X (·|xn ) rather than either PY |X (·|x̃n ) or P Y |X (·|x̃n ). To account for this discrepancy, the encoder will send the location of the nǫ 4 coordinates in which xn and x̃n differ, and the values of x̃n at those coordinates. Given this knowledge, the detector draws a new side-information vector Ỹ n in the following way. If xi = x̃i then Ỹi = Yi , and if xi 6= x̃i then Ỹi ∼ PY |X (·|x̃i ). Note that if the hypothesis PY |X is active then 20 21 If the minimizer is not unique, one of the minimizers can be arbitrarily and consistently chosen. As PXY ≪≫ P XY the exponents achieved by the DHT system are always finite. 33 n are Ỹ n ∼ PY |X (·|x̃n ) as required. If, however, the hypothesis P Y |X is active, then n − ⌈ nǫ 4 ⌉ coordinates of Ỹ drawn “correctly” according to Yi ∼ P Y |X (·|xi ), whereas nǫ 4 coordinates are drawn according to the “wrong” distribution Ỹi ∼ PY |X (·|x̃i ). As we will see, Lemma 13 assures that this mismatch has small effect on the type 1 error probability. At this point, the pair (X̃ n , Ỹ n ) can be encoded and detected using a DHT for the type Φǫ (QX ). For the sake of the brevity of next derivations, let us assume that (X̃ n , Ỹ n ) are fictitiously generated even when QX ∈ Pn0 (X ). Thus, the encoded vector is always x̃n , and the detector always uses the side-information vector ỹ n . We next turn to discuss the encoding of the modified source vector. We construct DHT subsystems Hn,Q̃X = (fn,Q̃X , ϕn,Q̃X ) for all Q̃X ∈ Pn0 (X ), according to Lemma 12. Each DHT subsystem is constructed from a constituent CD code whose type 1 error exponent is tuned to meet a constraint of E1 − D(Q̃X ||PX ). Now, as |Pn0 (X )| is finite, uniform convergence over Pn0 (X ) is assured, to wit, there exists subsystems Hn,Q̃X and k sufficiently large such that h  i p1 (fn,Q̃X , ϕn,Q̃X ) ≤ exp −n · E1 − D(Q̃X ||PX ) − δ , (B.52) n h io p2 (fn,Q̃X , ϕn,Q̃X ) ≤ exp −n · E2− (R, Q̃X , E1 − D(Q̃X ||PX )) − δ , (B.53) and for all Q̃X ∈ Pn0 (X ). A vector x̃n of type Q̃X ∈ Pn0 (X ) is encoded using a DHT subsystem Hn,Q̃X . Let us denote the constructed system by Hn = (fn , ϕn ) which includes both the randomly generated (X̃ n , Ỹ n ) and the subsystems Hn,Q̃X . It is required to analyze its asymptotic rate and both types of error exponent. To analyze the rate, let us now summarize the different parts of the encoded message: • An index of the type class within [|Pn (X )|]. As |Pn (X )|≤ (n + 1)|X | , it can be encoded in no more than ⌈|X |· log(n + 1)⌉ nats. • The location of the nǫ 4 coordinates in which x̃n (might) differ from xn . The number of sets of possible nǫ 4 coordinates satisfies  n nǫ 4  ǫ ≤ enhb ( 4 ) , (B.54) and thus can be encoded in no more than nhb ( 4ǫ ) nats. • • The value of xn in the possibly changed d coordinates. Each letter can be encoded using ⌈log|X |⌉ nats. The nR encoded nats of the DHT system for x̃n . Hence, the required rate is upper bounded by ǫ ǫ 1 + ⌈log|X |⌉ + R. ⌈|X |· log(n + 1)⌉ + hb n 4 4 (B.55) By choosing ǫ > 0 sufficiently small, the required rate can be made less than R + δ for all n sufficiently large. 34 For the type 1 error exponent, note that for all n = kn0 sufficiently large X   P [X n ∈ Tn (QX )] · P ϕn (Y n ) = H|X n ∈ Tn (QX ) (B.56)   e−nD(QX ||PX ) · P ϕn (Y n ) = H|X n ∈ Tn (QX ) QX ∈Pn (X ) h i (b) nδ = e · max e−nD(QX ||PX ) · P ϕn,Φǫ (QX ) (Ỹ n ) = H|X̃ n ∈ Tn (Φǫ (QX )) (B.57) p1 (Hn ) = QX ∈Pn (X ) (a) ≤ enδ · max QX ∈Pn (X ) (c)   e−nD(QX ||PX ) · P ϕn,Φǫ (QX ) (Y n ) = H|X n ∈ Tn (Φǫ (QX )) QX ∈Pn (X )   max exp −n · min [D(QX ||PX ) + E1 − D (Φǫ (QX )||PX ) − δ] = enδ · (d) ≤ max QX ∈Pn (X ) (e) QX ∈Pn (X ) ≤ exp [−n · (E1 − 2δ)] , (B.58) (B.59) (B.60) (B.61) where (a) follows since |Pn (X )|≤ (n + 1)|X | ≤ enδ and P[X n ∈ Tn (QX )] ≤ exp[−n · D(QX ||PX )], (b) follows from the definition of the system Hn , (c) follows from Lemma 13 (the trivial part regarding the type 1 error exponent), (d) follows from (B.52), and (e) follows from the fact that D(QX ||PX ) is a continuous function of QX in S(X ), and thus uniformly continuous. To analyze the type 2 error exponent, first note that, as for the type 1 error probability, p2 (Hn ) X = QX ∈Pn (X ) P [X n ∈ Tn (QX )] · P [ϕn (Y n ) = H|X n ∈ Tn (QX )]  ≤ exp −n · (B.62)   h i 1 n n min D(QX ||PX ) − log P ϕn,Φǫ (QX ) (Ỹ ) = H|X̃ ∈ Tn (Φǫ (QX )) − δ . n QX ∈Pn (X ) (B.63) Now, recalling that n = kn0 lim inf − k→∞ 1 log p2 (Hn ) kn0  h i 1 n n = lim inf min D(QX ||PX ) − log P ϕn,Φǫ (QX ) (Ỹ ) = H|X̃ ∈ Tn (Φǫ (QX )) − δ k→∞ QX ∈Pn (X ) n   (a)   ǫΩ 1 n ≥ lim inf min D(QX ||PX ) − log p2 Hn,Φǫ(QX ) |X ∈ Tn (Φǫ (QX )) − δ − k→∞ QX ∈Pn (X ) n 4 (b)  D(QX ||PX ) + F2− [H(Φǫ (QX )) − R, Φǫ (QX ), E1 − D(QX ||PX )] ≥ lim inf min (B.64) (B.65) k→∞ QX ∈Pn (X ) − ǫΩ − 2δ 4 (c) ≥ lim inf min k→∞ QX ∈Pn (X ) − 2δ − = min QX ∈Pn0 ǫΩ − δ1 4  (B.66)  D (Φǫ (QX )||PX ) + F2− [H(Φǫ (QX )) − R, Φǫ (QX ), E1 − D (Φǫ (QX )||PX ) + δ1 ] D(QX ||PX ) + F2− [H(QX ) − R, QX , E1 − D(QX ||PX ) + δ1 ] − 2δ − (B.67) ǫΩ − δ1 4  (B.68) 35 ≥ inf ( QX ∈Pn (X ) D (QX ||PX ) + F2− ) ǫΩ − δ1 , [H(QX ) − R, QX , E1 − D(QX ||PX ) + δ1 ] − 2δ − 4 (B.69) where (a) follows from Lemma 13 (and the way Ỹ n was defined), (b) follows from the uniform convergence of the type 1 error exponents of the DHT systems for Φ(QX ) ∈ Pn0 (X ) [see (B.53)]. Passage (c) holds for some δ1 > 0 that satisfies δ1 ↓ 0 as ǫ ↓ 0, and follows from the fact that D(QX ||PX ) is a continuous function of QX in S(X ), and thus uniformly continuous. Finally, by choosing ǫ > 0 sufficiently small, and then δ > 0 sufficiently small, the loss in exponents can be made arbitrarily small. B. Proof of the Converse Part Proof of the converse part of Theorem 6: For a given DHT system Hn , let us denote the (random) bin index def by In = fn (X n ). We will show that the converse hold even if the detector of the DHT systems Hn is aware of the type of xn . Consequently, as an optimal Neyman-Pearson detector will only average the likelihoods of source vectors from the true type class, it can be assumed w.l.o.g. that each bin contains only sequences from a unique type class. Recall that mn is the number of possible bins, and let mn,QX be the number of bins associated with a specific def type class QX , i.e., mn,QX = |Mn,QX | where def  i ∈ [mn ] : fn−1 (i) ∩ Tn (QX ) = fn−1 (i) . (B.70) h i −1 def µQX = E fn−1 (In ) |X n ∈ Tn (QX ) X −1 = fn−1 (i) · P [In = i|X n ∈ Tn (QX )] (B.71) Mn,QX = Further, conditioned on the type class QX (note that In is a function of X n ), (B.72) i∈Mn,QX = X fn−1 (i) i∈Mn,QX −1 · fn−1 (i) |Tn (QX )| (B.73) mn,QX |Tn (QX )| en(R+δ) ≤ n[H(Q )−δ] , X e = (B.74) (B.75) as for all n sufficiently large, mn ≤ en(R+δ) , and thus clearly mn,QX ≤ en(R+δ) . Hence, for any γ > 1, Markov’s inequality implies " P In : fn−1 (In ) # h 1 n X ∈ Tn (QX ) = P In : fn−1 (In ) ≥ γ · µQ X ≥1− 1 . γ −1 i ≤ γ · µQX X n ∈ Tn (QX ) (B.76) (B.77) 36 Thus, using (B.75), conditioned on X n ∈ Tn (QX ) fn−1 (In ) ≥ with probability larger than 1 − 1 γ 1 1 ≥ · en·[H(QX )−R−2δ] , γ · µQ X γ (B.78) > 0. Now, assume by contradiction that the statement of the theorem does not hold. This implies that there exists an increasing subsequence of blocklengths {nk }∞ k=1 and δ > 0 such that p1 (Hnk ) ≤ exp {−nk · [E1 − δ]} , (B.79)    p2 (Hk ) ≤ exp −nk · E2+ (R − 3δ, E1 − 3δ) + 5δ , (B.80) and for all k sufficiently large. For brevity of notation, we assume w.l.o.g. that these bounds hold for all n sufficiently large, and thus omit the subscript k. Let δ > 0 be given. Then, for all n sufficiently large [which only depends on (δ, |X |)], p1 [Hn |X n ∈ Tn (QX )] ≤ exp {−n · [E1 − D(QX ||PX ) − 2δ]} , (B.81) for all QX such that Tn (QX ) is not empty. Indeed, for all n sufficiently large, it holds that exp [−n · (E1 − δ)] ≥ p1 (Hn ) X = P [X n ∈ Tn (QX )] · p1 [Hn |X n ∈ Tn (QX )] (B.83) X e−n·[D(QX ||PX )+δ] · p1 [Hn |X n ∈ Tn (QX )] (B.84) max e−n·[D(QX ||PX )+δ] · p1 [Hn |X n ∈ Tn (QX )] , (B.85) QX ∈Pn (X ) ≥ ≥ (B.82) QX ∈Pn (X ) QX ∈Pn (X ) and (B.81) is obtained by rearranging. Writing p1 [Hn |X n ∈ Tn (QX )] = X i∈Mn,QX P [In = i|X n ∈ Tn (QX )] · p1 (Hn |In = i) , (B.86) Markov’s inequality implies n o P In : p1 (Hn |In ) ≥ enδ · p1 [Hn |X n ∈ Tn (QX )] ≤ = E [p1 (Hn |In )] · p1 [Hn |X n ∈ Tn (QX )] P n i∈Mn,Q P [In = i|X ∈ Tn (QX )] · p1 (Hn |In = i) enδ = e−nδ . X enδ · p1 [Hn |X n ∈ Tn (QX )] (B.87) (B.88) (B.89) The same arguments can be applied for the type 2 exponent. Thus, from the above and (B.78), with probability 37 larger than 1 − γ −1 − 2 · e−nδ , which is strictly positive for all sufficiently large n, the bin index satisfies (B.78), p1 (Hn |In ) ≤ enδ · p1 [Hn |X n ∈ Tn (QX )] ≤ exp {−n · [E1 − D(QX ||PX ) − 3δ]} , (B.90) (B.91) as well as    p2 (Hn |In ) ≤ exp −n · E2+ (R + 3δ, E1 − 3δ) − D(QX ||P X ) + 3δ . (B.92) Now, let Q∗X ∈ P(X ) be chosen to achieve E2+ (R + 3δ, E1 − 3δ) up to δ, i.e., to be chosen such that E2+ (R + 3δ, E1 − 3δ) − D(Q∗X ||P X ) ≥ F2+ (H(Q∗X ) − R − 3δ, Q∗X , E1 − D(Q∗X ||PX ) − 3δ) − δ, (B.93) ∗ and let {nl }∞ l=1 be the subsequence of blocklengths such that Tn (QX ) is not empty. From the above discussion, ∗ there a sequence of bin indices {i∗nl }∞ l=1 such that (B.78), (B.91) and (B.92) hold for QX . Consider the sequence (i∗nl ) to be a sequence of CD codes, whose rate is larger than H(QX ) − R − 3δ, its detectors of bins Cn∗l = fn−1 l are induced by the DHT system detector as φ∗nl (y nl ) = ϕnl (i∗nl , y nl ), and such that  p1 Cn∗ l , φ∗nl ≤ exp {−nl · [E1 − D(Q∗X ||PX ) − 3δ]} , and     p2 Cn∗ l , φ∗nl ≤ exp −nl · E2+ (R + 3δ, E1 − 3δ) − D(Q∗X ||P X ) + 3δ    ≤ exp −nl · F2+ (H(Q∗X ) − R − 3δ, Q∗X , E1 − D(Q∗X ||PX ) − 3δ) + 2δ , (B.94) (B.95) (B.96) where the last inequality follows from (B.93). However, this is a contradiction, since whenever (B.94) holds the definition of CD reliability function implies that     p2 Cn∗ l , φ∗nl ≥ exp −nl · F2+ (H(Q∗X ) − R − 3δ, Q∗X , E1 − D(Q∗X ||PX ) − 3δ) + δ (B.97) for all l sufficiently large. A PPENDIX C P ROOF OF T HEOREM 8 We will follow the outline which appears right after the statement of the theorem. We will consider CD codes drawn from the fixed-composition hierarchical ensemble defined in Definition 7, with conditional distribution QU |X , cloud-center rate ρc , and satellite rate ρs (which satisfy ρ = ρc + ρs ). In the course of the proof, we shall consider various types of the form QU XY and QU XY . All of them are assume to have (U, X) marginal QU X = QU X even if it is not explicitly stated. Furthermore, we shall assume that the blocklength n is such that Tn (QU X ) is not empty. 38 In this case, the notation for exponential equality (or inequality) needs to be clarified as follows. We will say that . an = bn if 1 an lim (C.1) log l = 1, l→∞ nl bn l where {nl }∞ l=1 is the subsequence of blocklengths such that Tn (QU X ) is not empty. The proof of Theorem 8 relies on the following result, which is stated and proved by means of the typeenumeration method (see [48, Sec. 6.3]). Specifically, for a given y n , we define type-class enumerators for a random CD code Cn by def Myn (QU XY ) = |{xn ∈ Cn : ∃un such that xn ∈ Cn,s (un ), (un , xn , y n ) ∈ Tn (QU XY )}| . (C.2) To wit, Myn (QU XY ) counts the random number of codewords which have joint type QU XY ∈ Pn (U × X × Y) with their cloud center un and y n . To derive a random-coding bound on the achievable CD exponents, we will (QU XY )Myλn (QU XY )] for an arbitrary sequence of {y n } taken need to evaluate the exponential order of E[My1−λ n from Tn (QY ) = Tn (QY ). The result is summarized in the following proposition, which is interesting on its own right. To this end, it will be convenient to denote  def ∆λ (QU XY , QU XY ) = (1 − λ) · [ρ − IQ (U, X; Y )] − λ · max |IQ (U ; Y ) − ρc |+ , IQ (U, X; Y ) − ρ   i h + λ ρ − IQ (U, X; Y ) − (1 − λ) · max IQ (U ; Y ) − ρc , IQ (U, X; Y ) − ρ . + (C.3) Proposition 15. Let QU XY , QU XY ∈ Pn0 (U × X × Y) be given for some n0 , with QU X = QU X and QY = QY . Also let {nl } be the subsequence of blocklengths such that Tnl (QU XY ) and Tnl (QU XY ) are both non-empty, and nl let {y nl }∞ l=1 satisfy y ∈ Tnl (QY ) for all l. Then, for any λ ∈ (0, 1) h i 1 λ log E My1−λ ) nl (QU XY )M nl (Q y U XY nl →∞ nl     ρ − IQ (U, X; Y ),    = ∆λ (QU XY , QU XY ),      ∆λ (QU XY , QU XY ) − |IQ (U ; Y ) − ρc |+ , lim QU XY = QU XY QU Y 6= QU Y and QU XY 6= QU XY . (C.4) QU Y = QU Y (and QU XY 6= QU XY ) It is interesting to note that the expression (C.4) is not continuous, as, say, QU XY → QU XY . The proof of Proposition 15 is of technical nature, and thus relegated to Appendix D. For the rest of the proof, no knowledge of the type-enumeration method is required. Equipped with this result we proceed to analyze typical properties of randomly generated codes. As the CD (Cn ) (C ) problem corresponds to an HT problem between P Y n and PY nn , and the performance of the latter is determined by the Chernoff distance [see (11)], we next analyze the average exponent of the Chernoff distance over a random choice of CD codes. 39 Lemma 16. Let Cn be drawn randomly from the hierarchical ensemble of Definition 7, with conditional distribution QU |X , cloud-center rate ρc , and satellite rate ρs (which satisfy ρ = ρc + ρs ). Then,    X h i1−λ h iλ  (Cn ) (C ) ˙ exp [−n · min {dλ (QX ), Arc }] . E P nn (y n ) · P Y n (y n ) ≤  n n Y  (C.5) y ∈Y where dλ (QX ) is defined in (27) when setting τ = 1−λ λ , and Arc is defined in (53). Proof: For (xn , y n ) ∈ Tn (QXY ), let us denote the log-likelihood of PY |X by 1 def L(QXY ) = − log PY |X (y n |xn ) n X =− QXY (x, y) log P (y|x), (C.6) (C.7) x∈X , y∈Y and the log-likelihood of P Y |X by L(QXY ) (with P Y |X replacing PY |X ). For any given n,    X h i1−λ h iλ  (Cn ) (C ) E PY nn (y n ) · P Y n (y n )  n n  y ∈Y " #λ  #1−λ "   X X X 1 P Y |X (y n |xn ) (C.8) = nρ PY |X (y n |xn ) E ·   n e xn ∈Cn x ∈Cn y n ∈Y n " #1−λ " #λ   X  X X X 1 PY |X (y n |xn ) E · = nρ P Y |X (y n |xn ) (C.9)  n  e n n x ∈Cn x ∈Cn QY ∈Pn (Y) y ∈Tn (QY )  1−λ  λ     X  X X (a) 1 −nL(Q ) −nL(Q XY )  XY    |Tn (QY )| · E Myn (QU XY )e · Myn (QU XY )e = nρ   e  QU XY  QY ∈Pn (Y) Q U XY (C.10)  1−λ  λ     X  X −nL(Q n[H(QY )−ρ] −nL(Q ) ) XY XY   Myn (QU XY )e ≤ |Pn (Y)| · max e ·E  Myn (QU XY )e ·   QY ∈Pn (Y)  QU XY  Q U XY (C.11) ≤ |Pn (Y)| |Pn (U × X × Y)|2 " × max QY ∈Pn (Y) (QU XY )e−n(1−λ)·L(QXY ) · max Myλn (QU XY )e−nλL(QXY ) en[H(QY )−ρ] · E max My1−λ n QU XY ≤ |Pn (Y)| |Pn (U × X × Y)|2  × max QY ∈Pn (Y) en[H(QY )−ρ] · E  X QU XY QU XY My1−λ (QU XY )e−n(1−λ)·L(QXY ) · n X QU XY # (C.12)  (C.13) Myλn (QU XY )e−nλL(QXY )  = |Pn (Y)| |Pn (U × X × Y)|2 i X X h E My1−λ (QU XY ) · Myλn (QU XY ) · e−n[(1−λ)·L(QXY )+λL(QXY )] × max en[H(QY )−ρ] n QY ∈Pn (Y) QU XY QU XY (C.14) 40 ≤ |Pn (Y)| |Pn (U × X × Y)|4 i h × max max en[H(QY )−ρ] · E My1−λ (QU XY ) · Myλn (QU XY ) · e−n[(1−λ)·L(QXY )+λL(QXY )] n QU XY QU XY def = cn · max (QU XY ,QU XY )∈Qn ζn (QU XY , QU XY ), (C.15) (C.16) where (a) follows since by symmetry, the expectation only depends on the type of y n , and by using the definitions of the enumerators in (C.2), and the log-likelihood in (C.6). After passage (a) and onward, y n is an arbitrary member of Tn (QY ), and the sums and maximization operators are over (QU XY , QU XY ) restricted to the set def Qn =  (QU XY , QU XY ) ∈ Pn2 (U × X × Y) : QY = QY , QXU = QXU . (C.17)  (QU XY , QU XY ) ∈ S 2 (U × X × Y) : QY = QY , QXU = QXU , (C.18) In the last equality we have implicitly defined cn and ζn (QU XY , QU XY ). By defining def Q = and using standard arguments (e.g., as in the proof of Sanov’s theorem [3, Theorem 11.4.1]) we get    i h h i X 1−λ λ 1 1 (Cn ) (C ) lim inf − log ζn (QU XY , QU XY ). PY nn (y n ) · P Y n (y n ) lim inf − log E ≥ min n→∞  n n  (QU XY ,QU XY )∈Q n→∞ n n y ∈Y (C.19) Hence, to obtain the exponential upper bound required (C.5), we may minimize over (QU XY , QU XY ) the exponential order of ζn (·), denoted as 1 def Λλ (QU XY , QU XY ) = lim inf − log ζn (QU XY , QU XY ) n→∞ n h i 1 = −H(QY ) + ρ − lim inf − log E My1−λ (QU XY )Myλn (QU XY ) n n→∞ n   + (1 − λ) · L(QXY ) + λL(QXY ) . (C.20) (C.21) We now evaluate this expression in three cases, which correspond to the three cases of Proposition 15. In each one, we substitute for lim inf n→∞ − n1 log E[My1−λ (QU XY )Myλn (QU XY )] the appropriate result,22 as follows: n Case 1. For QU XY = QU XY , Λλ (QU XY , QU XY ) = −H(QY ) + ρ − [ρ − IQ (U, X; Y )] + (1 − λ) · L(QXY ) + λ · L(QXY )   = (1 − λ) · [−HQ (Y |X, U ) + L(QXY )] + λ · −HQ (Y |X, U ) + L(QXY ) (a) = (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X ), 22 In fact, Proposition 15 implies that the limit inferior of this sequence is a proper limit. (C.22) (C.23) (C.24) 41 where (a) follows from the identity − HQ (Y |X, U ) + L(QXY ) = D(QY |U X ||PY |X |QU X ). (C.25) By defining the distribution λ def (λ,x) PY (y) = PY1−λ |X (y|x)P Y |X (y|x) P y ′ ∈Y and observing that λ ′ ′ PY1−λ |X (y |x)P Y |X (y |x)   min (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X ) QY |U X      X X λ (λ,x)  (y|x) (y|x)P PY1−λ QX (x) log  = min D(QY |X ||PY |QX ) − Y |X |X  QY |X  y∈Y x∈X   X X λ  =− QX (x) log  PY1−λ |X (y|x)P Y |X (y|x) , x∈X (C.26) (C.27) (C.28) y∈Y it is evident that   Λλ,1 (QU XY , QU XY ) = min (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X ) (C.29) QY |X   X X λ (C.30) (y|x)P Y |X (y|x) PY1−λ =− QX (x) log  |X x∈X y∈Y = dλ (QX ). Case 2. (C.31) For QU XY 6= QU XY and QU Y 6= QU Y Λλ (QU XY , QU XY ) = −H(QY ) + ρ + (1 − λ) · L(QXY ) + λ · L(QXY )  − (1 − λ) · [ρ − IQ (U, X; Y )] + λ · max |IQ (U ; Y ) − ρc |+ , IQ (U, X; Y ) − ρ   i h − λ ρ − IQ (U, X; Y ) + (1 − λ) · max IQ (U ; Y ) − ρc , IQ (U, X; Y ) − ρ + (C.32) (C.33) (a) = (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X )  + λ · max |IQ (U ; Y ) − ρc |+ , IQ (U, X; Y ) − ρ   + (1 − λ) · max IQ (U ; Y ) − ρc , IQ (U, X; Y ) − ρ + def = Λλ,2 (QU XY , QU XY ), where (a) follows from (C.25) again and rearranging. (C.34) (C.35) 42 Case 3. For QU XY 6= QU XY and QU Y = QU Y Λλ (QU XY , QU XY ) = Λλ,2 (QU XY , QU XY ) − |IQ (U ; Y ) − ρc |+ (C.36) def = Λλ,3 (QU XY , QU XY ). (C.37) Hence, the required bound on the Chernoff distance is given by min ( (QU XY ,QU XY )∈Q Λλ (QU XY , QU XY ) = min dλ (QX ), min (QU XY ,QU XY )∈Q:QU Y 6=QU Y Λλ,2 (QU XY , QU XY ), min (QU XY ,QU XY )∈Q:QU Y =QU Y ) Λλ,3 (QU XY , QU XY ) . (C.38) Observing (C.37), we note that the third term in (C.38) satisfies min (QU XY ,QU XY )∈Q:QU Y =QU Y = min ( Λλ,3 (QU XY , QU XY ) min (QU XY ,QU XY )∈Q:QU Y =QU Y , IQ (U ;Y )≤ρc min (QU XY ,QU XY )∈Q:QU Y =QU Y , IQ (U ;Y )>ρc (a) = min ( min (QU XY ,QU XY )∈Q:QU Y =QU Y ,  Λλ,2 (QU XY , QU XY ), Λλ,2 (QU XY , QU XY ) − IQ (U ; Y ) + ρc ) (C.39) Λλ,2 (QU XY , QU XY ), min (QU XY ,QU XY )∈Q:QU Y =QU Y , IQ (U ;Y )>ρc  Λλ,2 (QU XY , QU XY ) − IQ (U ; Y ) + ρc ) , (C.40) where in (a) we have removed the constraint IQ (U ; Y ) ≤ ρc in the first term of the outer minimization, since for QU Y with IQ (U ; Y ) > ρc the second term will dominate the minimization. Thus, the first term in (C.40) may be unified with the second term of (C.38). Doing so, the constraint QU Y 6= QU Y may be removed in the second term of (C.38). Consequently, the required bound on the Chernoff distance is given by ( min dλ (QX ), min (QU XY ,QU XY )∈Q Λ2 (QU XY , QU XY ), min (QU XY ,QU XY )∈Q:QU Y =QU Y , IQ (U ;Y )>ρc  Λ2 (QU XY , QU XY ) − IQ (U ; Y ) + ρc ) . (C.41) The second term in (C.41) corresponds to A′rc defined in (51). The third term corresponds to A′′rc defined in (52), when using ρs = ρ − ρc , and noting that when IQ (U ; Y ) > ρc , it is equal to Λλ,2 (QU XY , QU XY ) − IQ (U ; Y ) + ρc 43 = (1 − λ) · D(QY |U X ||PY |X |QU X ) + λ · D(QY |U X ||P Y |X |QU X ) + λ · |IQ (X; Y |U ) − ρs |+ + (1 − λ) · IQ (X; Y |U ) − ρs + . (C.42) Using the definition (53) of Arc as the minimum of the last two cases, (C.5) is obtained. Next, recall that by its definition, a valid CD code of rate ρ is comprised of at least enρ distinct codewords. However, when the codewords are independently drawn, some of them might be identical. Nonetheless, the next lemma shows that the average number of distinct codewords is asymptotically close to enρ . Lemma 17. Let Cn be drawn randomly from the hierarchical ensemble of Definition 7, with conditional distribution QU |X , cloud-center rate ρc , and satellite rate ρs (which satisfy ρ = ρc + ρs ). If ρc + HQ (X|U ) ≥ ρ then ˙ enρ . E [|Cn |] ≥ (C.43) nρc Proof: Let us enumerate the random cloud centers as {U n (i)}ei=1 and the random satellite codebooks as nρc {Cn,s (U n (i))}ei=1 . When a random satellite codebook Cn,s (un ) is drawn, then for any given (un , xn ) ∈ Tn (QU X ) n n P [x ∈ Cn,s (u )] = 1 − using 1 − (1 − t)K ≥ 1 2 1− 1 Tn (QU |X , un ) !enρs ( ) enρs 1 ≥ · min 1, 2 Tn (QX|U , un )   ≥ exp −n · |HQ (X|U ) − ρs |+ + δ , (C.44) (C.45) (C.46) · min {1, tK} [60, Lemma 1]. Further, for a random cloud center U n and xn ∈ Tn (QX ) P [xn ∈ Cn,s (U n )] = P [(xn , U n ) ∈ Tn (QU X )] · P [xn ∈ Cn,s (U n )|(xn , U n ) ∈ Tn (QU X )] (a) (C.47) = P [(xn , U n ) ∈ Tn (QU X )] · P [xn ∈ Cn,s (un )] (C.48) ≥ (C.49)   Tn (QU |X , xn ) · exp −n · |HQ (X|U ) − ρs |+ + δ |Tn (QU )|   ≥ exp −n · IQ (U ; X) + |HQ (X|U ) − ρs |+ + 2δ def = e−n(ξ+2δ) , (C.50) (C.51) where (a) is due to symmetry, and the definition def ξ = IQ (U ; X) + |HQ (U |X) − ρs |+ = max {IQ (U ; X), H(QX ) − ρs } . Therefore, the average number of distinct codewords in the random CD code Cn is lower bounded as   X I (xn ∈ Cn ) E [|Cn |] = E  xn ∈T n (QX ) (C.52) (C.53) (C.54) 44 = X P (xn ∈ Cn ) X P xn ∈Tn (QX ) = xn ∈Tn (QX ) (a) ≥ X xn ∈T n (QX ) (enρc [ i=1 (C.55) [xn ∈ Cn,s (U n (i))] ) 1 · min {enρc · P [xn ∈ Cn,s (U n (1))] , 1} 2 (C.56) (C.57) ≥ exp {n · [min {H(QX ) + ρc − ξ, H(QX )} − 3δ]} (C.58) = exp {n · [min {H(QX ) + ρc − IQ (U ; X), ρc + ρs , H(QX )} − 3δ]} (C.59) = en(ρ−3δ) , (C.60) (b) where (a) holds since for a given set of K pairwise independent events {Ak }K k=1 [61, Lemma A.2] "K ( K ) # X [ 1 P P (Ak ) . Ak ≥ min 1, 2 k=1 (C.61) k=1 The passage (b) follows from the assumptions ρ < H(QX ) and ρc + HQ (X|U ) ≥ ρ. Thus, on the average, a randomly chosen Cn has more than en(ρ−3δ) distinct codewords. We are now ready to prove Theorem 8. Proof of Theorem 8: As noted in Section IV, given a CD code Cn , the detector faces an ordinary HT problem (C ) (Cn ) between the distributions PY nn and P Y n , and thus the bounds of Section II-B can be used. Specifically, using (11) (with τ = 1−λ λ ), the type 2 error exponent under a type 1 error exponent constraint of F1 , can be lower bounded for any given n by     1−λ   i i h h X λ 1−λ 1 1 (Cn ) n (Cn ) n max max − (y ) P n (y ) · PYn . · F1 − · log   n n Y λ λ n Cn ⊆Tn (QX ): |Cn |≥enρ 0≤λ≤1  (C.62) y ∈Y Instead of maximizing over Cn , we use ensemble averages. To this end, note that up to this point, we have obtained (Lemma 16)    X h iλ  i1−λ h (Cn ) . (C ) E = exp [−n · min {dλ (QX ), Arc }] , P nn (y n ) · P Y n (y n )   n n Y (C.63) y ∈Y . and showed that the average number of distinct codewords in a randomly chosen codebook is |Cn |= enρ (Lemma 17). It remains to prove the existence of a CD code, whose codewords are all distinct, and its Chernoff distance exponent is close to the ensemble average. To this end, let δ > 0 be given, and consider the events   1 def A1 = |Cn | ≥ E [|Cn |] , 2 (C.64) and    X h i1−λ h iλ oi h n (Cn ) def (C ) P nn (y n ) · P Y n (y n ) ≤ e4nδ · exp −n · min dλ (QX ), Ãrc A2 (λ) = .  n n Y  y ∈Y (C.65) 45 Note that since P(|Cn |≤ enδ · E[|Cn |]) = P(|Cn |≤ enρ ) = 1, for all n sufficiently large, the reverse Markov inequality23 [62, Section 9.3, p. 159] implies that P [A1 ] ≥ 1 − 1/2 ≥ e−2nδ . enδ − 1/2 (C.66) Further, for any given λ ∈ [0, 1], Markov’s inequality implies that for all n sufficiently large P [A2 (λ)] ≥ 1 − e−n3δ . (C.67) P [A1 ∩ A2 (λ)] ≥ 1 − P[Ac1 ] − P [Ac2 (λ)]   ≥ 1 − e−3nδ − 1 − e−2nδ (C.68) Then, we note that (C.69) = e−2nδ − e−3nδ (C.70) > 0, (C.71) ad thus deduce that there exists a CD code Cn∗ (λ) such that |Cn∗ (λ)|≥ 41 en(ρ−δ) ≥ en(ρ−2δ) and A2 (λ) holds for all n sufficiently large. Let the CD code obtained after keeping only the unique codewords of Cn∗ (λ) be denoted as Cn∗∗ (λ). It remains to show that the exponent of the Chernoff distance of Cn∗∗ (λ) is asymptotically equal to that of Cn∗ (λ). Indeed, i1−λ h iλ X h (C ∗ (λ)) (C ∗ (λ)) PY nn (y n ) · P Y nn (y n ) y n ∈Y n  1−λ  λ 1 1 P (y n |xn ) PY |X (y n |xn ) · nρ nρ Y |X e e y n ∈Y n xn ∈Cn∗ (λ) xn ∈Cn∗ (λ)  λ 1−λ  X X (a) X 1 1  ≥ P (y n |xn ) P (y n |xn ) · nρ Y |X nρ Y |X e e y n ∈Y n xn ∈Cn∗∗ (λ) xn ∈Cn∗∗ (λ) 1−λ   λ X X X (b) 1 1  P (y n |xn ) · ≥ e−2nδ · P (y n |xn ) , ∗∗ (λ)| Y |X ∗∗ (λ)| Y |X |C |C n n n n n n ∗∗ ∗∗ = X  X y ∈Y x ∈Cn (λ) X (C.72) (C.73) (C.74) x ∈Cn (λ) where (a) follows since Cn∗∗ (λ) ⊆ Cn∗ (λ), and (b) follows since |Cn∗∗ (λ)|≥ en(ρ−2δ) , and therefore X h y n ∈Y n (C ∗∗ (λ)) PY nn i1−λ h i h n oi (C ∗∗ (λ)) n λ (y n ) · P Y nn (y ) ≤ e6nδ · exp −n · min dλ (QX ), Ãrc (ρ, ρc , QU X , λ) . As all the above holds for any λ ∈ [0, 1], (C.62) implies       1−λ i i h h X λ 1−λ 1 1 (Cn ) (C ) · F1 − · log PY nn (y n ) · P Y n (y n ) − max max   n n 0≤λ≤1 Cn ⊆Tn (QX ): |Cn |≥enρ  λ λ n (C.75) y ∈Y 23 The reverse Markov inequality states that if P(0 ≤ X ≤ αE[X]) = 1 for some α > 1. Then, for any β < 1,P (X > βE[X]) ≥ 1−β . α−β 46     i h h i X 1−λ 1 1 (C ∗∗ (λ)) n λ (C ∗∗ (λ)) n 1−λ ≥ max − (y ) PY nn · P Y nn · F1 − · log (y )  n n  0≤λ≤1 λ λ n y ∈Y   1 1−λ · F1 + · [min {dλ (QX ), Arc (ρ, ρc , QU X , λ)} − 6δ] . ≥ max − 0≤λ≤1 λ λ (C.76) (C.77) Now, due to the continuity of Arc (ρ, ρc , QU X , λ) in QU X , for any given QU X ∈ S(U × X ), one can find a blocklength n0 such that for all n ≥ n0 with Tn (QX ) not empty, there exists Q̃U X ∈ Pn (U × X ) with QX = Q̃X such that Arc (ρ, ρc , QU X , λ) − Arc (ρ, ρc , Q̃U X , λ) ≤ δ. With this, we have proved the existence of a CD code such that     1 i h h i X 1−λ λ 1 1 (Cn ) n (Cn ) n − · log P n (y ) · P Y n (y ) ≥ · [min {dλ (QX ), Arc } − 3δ]  n n Y  λ λ n (C.78) (C.79) y ∈Y can be established for any blocklength such that Tn (QX ) is not empty. The proof is completed by taking n → ∞ and δ ↓ 0, and allowing to maximize Arc (ρ, ρc , QU X , λ) over QU |X and ρc (with a constraint set dictated by Lemma 17). Remark 18. As mentioned after Theorem 8, the random-coding bound can be achieved by a single sequence of CD codes, simultaneously for all type 1 error exponent constraint F1 . This can be proved by showing that there exists a CD code such that the event A2 (λ) defined in (C.67) holds for all λ ∈ [0, 1]. To show the latter, we uniformly quantize the interval [0, 1] to {λi }K i=0 with λi = i K and a fixed K . Then, using the union bound, for all n sufficiently large P "K \ i=0 # A2 (λi ) ≥ 1 − K X P [Ac2 (λi )] (C.80) i=0 ≥ 1 − (K + 1) · e−3nδ (C.81) ≥ 1 − e−2nδ , (C.82) and this bound can be used in lieu of (C.67) in the proof. This will prove the simultaneous achievability of A2 (λi ) for all 0 ≤ i ≤ K . Then, utilizing the continuity of Arc (ρ, ρc , QU X , λ), by taking K to increase sub-exponentially in n the same result can be established to the entire [0, 1] interval. A PPENDIX D T HE T YPE -E NUMERATION M ETHOD AND THE P ROOF OF P ROPOSITION 15 We begin with a short review of the type-enumeration method [48, Sec. 6.3]. To begin, let us define type-class enumerators for the cloud centers by def Nyn (QU Y ) = |{un ∈ Cc,n : (un , y n ) ∈ Tn (QU Y )}| . (D.1) 47 To wit, Nyn (QU Y ) counts the random number of cloud centers which have joint type QU Y ∈ Pn (U × Y) with y n . While Myn (QU XY ) defined in (C.2) is an enumerator of an hierarchical ensemble, Nyn (QU Y ) is an enumerator of an ordinary ensemble, and thus simpler to analyze. Furthermore, the analysis of Myn (QU XY ) depends on the properties of Nyn (QU Y ), and thus we begin by analyzing the latter. As the cloud centers in the ensemble are drawn independently, Nyn (QU Y ) is a binomial random variable. It pertains to enρc trials and probability of success of the exponential order of exp[−n · IQ (U ; Y )], and consequently, . E[Nyn (QU Y )] = exp{n · [ρc − IQ (U ; Y )]}. In the sequel, we will need refined properties of the enumerator, and specifically, its large-deviations behavior and the moments E[Nyλn (QU Y )]. To this end, we note that as Nyn (QU Y ) is just an enumerator for a code drawn from an ordinary ensemble, the analysis of [48, Sec. 6.3] [63, Appendix A.2] holds. As was shown there, when IQ (U ; Y ) ≤ ρc , E[Nyn (QU Y )] increases exponentially with n as exp{n · [ρc − IQ (U ; Y )]}, and Nyn (QU Y ) concentrates double-exponentially rapidly around this average. Specifically, letting def Bn (QU Y , δ) = n o e−nδ · E [Nyn (QU Y )] ≤ Nyn (QU Y ) ≤ enδ · E [Nyn (QU Y )] , (D.2) then for any δ > 0 sufficiently small i h P [Bnc (QU Y , δ)] ≤ exp −enδ . (D.3) When IQ (U ; Y ) > ρc holds, E[Nyn (QU Y )] decreases exponentially with n as exp{−n · [IQ (U ; Y ) − ρc ]}, and Nyn (QU Y ) = 0 almost surely. Furthermore, the probability that even a single codeword has joint type QU Y with y n is exponentially small, and the probability that there is an exponential number of such codewords is double- exponentially small. Specifically, for all sufficiently large n P {Nyn (QU Y ) ≥ 1} ≤ exp {−n · [IQ (U ; Y ) − ρc ]} , (D.4) i o h n P Nyn (QU Y ) ≥ e2nδ ≤ exp −enδ . (D.5) and for any given δ > 0 Using the properties above, it can be easily deduced that   i  h . exp {nλ · [ρc − IQ (U ; Y )]} , E Nyλn (QU Y ) =  exp {−n · [IQ (U ; Y ) − ρc ]} , IQ (U ; Y ) ≤ ρc , (D.6) IQ (U ; Y ) > ρc or, in an equivalent and more compact form, h i   . E Nyλn (QU Y ) = exp n · λ [ρc − IQ (U ; Y )] − (1 − λ) · |IQ (U ; Y ) − ρc |+ . (D.7) We can now turn to analyze the behavior of the more complicated enumerator Myn (QU XY ). To this end, note that conditioned on the event Nyn (QU Y ) = enν , Myn (QU XY ) is a binomial random variable pertaining to exp[n(ν +ρs )] 48 trials and probability of success of the exponential order of exp[−n · IQ (X; Y |U )]. Thus,   i  h . exp {nλ · [ν + ρs − IQ (X; Y |U )]} , IQ (X; Y |U ) ≤ ν + ρs E Myλn (QU XY ) Nyn (QU Y ) = enν = ,  exp {−n · [IQ (X; Y |U ) − ν − ρs ]} , IQ (X; Y |U ) > ν + ρs (D.8) and the conditional large-deviations behavior of Myn (QU XY ) is identical to the large-deviations behavior of Nyn (QU Y ), with ν + ρs and IQ (X; Y |U ) replacing ρc and IQ (U ; Y ), respectively. The next lemma provides an asymptotic expression for the unconditional moments of Myn (QU XY ). It can be easily seen that (D.7) is obtained as a special case, when setting U = X and ρ = ρc . Lemma 19. For λ > 0 h i    . E Myλn (QU XY ) = exp n · λ [ρ − IQ (U, X; Y )] − (1 − λ) · max |IQ (U ; Y ) − ρc |+ , IQ (U, X; Y ) − ρ . (D.9) Proof: Let   1 δ ∈ 0, lim inf log E[Nyn (QU Y )] , n→∞ n (D.10) define the events def A=0 n (QU Y ) = {Ny n (QU Y ) = 0} , def A=1 n (QU Y , δ) = n def A≥1 n (QU Y , δ) = o 1 ≤ Nyn (QU Y ) ≤ e2nδ , n o Nyn (QU Y ) ≥ e2nδ , (D.11) (D.12) (D.13) and recall the definition of the event Bn (QU Y , δ) in (D.2). We will consider four cases depending on the relations between the rates and mutual information values. For the sake of brevity, only the first case will be analyzed with a strictly positive δ > 0 and then the limit δ ↓ 0 will be taken. In all other three cases, we shall derive the expressions for the moments assuming δ = 0, with the understanding that upper and lower bounds can be derived in a similar manner to the first case. For notational convenience, when the expressions are derived assuming δ = 0 we will omit δ from the notation of the events defined above [e.g., A=1 n (QU Y )]. We will use the moments (D.6), the strong concentration relation (D.3) and the large-deviations bound (D.4), for both Nyn (QU Y ) and Myn (QU XY ) [when the latter is conditioned on the value of Nyn (QU Y )]. Case 1. If IQ (U ; Y ) > ρc and IQ (X; Y |U ) > ρs , then for all δ > 0 sufficiently small, IQ (U ; Y ) > ρc + δ and IQ (X; Y |U ) > ρs + δ and thus, i h E Myλn (QU XY ) i h    =1  =1 λ ≤ P A=0 n (QU Y ) · 0 + P An (QU Y , δ) · E My n (QU XY )|An (QU Y , δ)   nλρ + P A≥1 n (QU Y , δ) · e (D.14) 49 i h ˙ 0 + exp {−n · [IQ (U ; Y ) − ρc − δ]} · exp {−n · [IQ (X; Y |U ) − ρs − 2δ]} + exp −enδ · enλρ ≤ (D.15) . = exp {n · [ρ − IQ (U, X; Y ) + 3δ]} . (D.16) Similarly, it can be shown that i h ˙ exp {n · [ρ − IQ (U, X; Y ) − 3δ]} . E Myλn (QU XY ) ≥ (D.17) As δ ≥ 0 is arbitrary, we obtain h i . E Myλn (QU XY ) = exp {n · [ρ − IQ (U, X; Y )]} . Case 2. If IQ (U ; Y ) > ρc and IQ (X; Y |U ) < ρs then i h E Myλn (QU XY ) i h    =1  =1 λ (Q )|A (Q ) ≤ P A=0 (Q ) · 0 + P A (Q ) · E M n U XY U Y U Y U Y n n n y  ≥1  nλρ + P An (QU Y ) · e Case 3. (D.18) (D.19) . = 0 + exp {−n · [IQ (U ; Y ) − ρc ]} · exp {n · λ · [ρs − IQ (X; Y |U )]} + 0 (D.20) = exp {n · [ρc − IQ (U ; Y ) + λ · (ρs − IQ (X; Y |U ))]} (D.21) = exp {n · [ρc − IQ (U ; Y ) + ρs − IQ (X; Y |U ) − (1 − λ) · (ρs − IQ (X; Y |U ))]} (D.22) = exp {n · [ρ − IQ (U, X; Y ) − (1 − λ) (ρs − IQ (X; Y |U ))]} . (D.23) If IQ (U ; Y ) < ρc and IQ (X; Y |U ) > ρc − IQ (U ; Y ) + ρs then i h E Myλn (QU XY ) h i = P {Bn (QU Y )} · E Myλn (QU XY )|Bn (QU Y ) h i + P {Bnc (QU Y )} · E Myλn (QU XY )|Bnc (QU Y ) Case 4. (D.24) . = 1 · exp {n · [ρc − IQ (U ; Y ) + ρs − IQ (X; Y |U )]} + 0 (D.25) = exp {n · [ρ − IQ (U, X; Y )]} . (D.26) If IQ (U ; Y ) < ρc and IQ (X; Y |U ) < ρc − IQ (U ; Y ) + ρs then h i E Myλn (QU XY ) i h = P {Bn (QU Y )} · E Myλn (QU XY )|Bn (QU Y ) i h + P {Bnc (QU Y )} · E Myλn (QU XY )|Bnc (QU Y ) . = 1 · exp {n · λ · [ρc − IQ (U ; Y ) + ρs − IQ (X; Y |U )]} + 0 (D.27) (D.28) 50 = exp {n · λ · [ρ − IQ (U, X; Y )]} (D.29) = exp {n · [ρ − IQ (U, X; Y ) − (1 − λ) [ρ − IQ (U, X; Y )]]} . (D.30) Noting that IQ (X; Y |U ) > ρc − IQ (U ; Y ) + ρs is equivalent to IQ (U, X; Y ) > ρ, it is easy to verify that i h  n . E Myλn (QU XY ) = exp n · ρ − IQ (U, X; Y ) − (1 − λ) ρs − IQ (X; Y |U ) + |ρc − IQ (U ; Y )|+ + o (D.31) matches all four cases. The final expression is obtained from the identities ρ − IQ (U, X; Y ) − (1 − λ) · ρs − IQ (X; Y |U ) + |ρc − IQ (U ; Y )|+ + (a) = ρ − IQ (U, X; Y ) − (1 − λ) ρ − IQ (U, X; Y ) + |IQ (U ; Y ) − ρc |+ (D.32) + (b) = λ [ρ − IQ (U, X; Y )] − (1 − λ) |IQ (U ; Y ) − ρc |+ − (1 − λ) IQ (U, X; Y ) − ρ − |IQ (U ; Y ) − ρc |+  (c) = λ [ρ − IQ (U, X; Y )] − (1 − λ) · max |IQ (U ; Y ) − ρc |+ , IQ (U, X; Y ) − ρ , + (D.33) (D.34) where (a) follows from the identity |t|+ = t+|−t|+ with t = ρc −IQ (U ; Y ), (b) follows from the same identity with t = ρ−IQ (U, X; Y )+|IQ (U ; Y )−ρc |+ , and (c) follows from t+|s − t|+ = max {t, s} with t = |IQ (U ; Y )−ρc |+ and s = IQ (U, X; Y ) − ρ. To continue, we will need to separate the analysis according to whether QU Y 6= QU Y (Lemma 22) or QU Y = QU Y (Lemma 25). In the former case Myn (QU XY ) and Myn (QU XY ) count codewords which pertain to different cloud (QU XY ) and Myλn (QU XY ) are asymptotically uncorrelated. In the later centers, and as we shall next see, My1−λ n case, these enumerators may count codewords which pertain to the same cloud center, and correlation between (QU XY ) and Myλn (QU XY ) is possible. We will need two auxiliary lemmas. My1−λ n Lemma 20. Let f (t) be a monotonically non-decreasing function, let B ∼ Binomial(N, p) and B̃ ∼ Binomial(Ñ , p) with Ñ > N . Then, def h i E [f (B)] ≤ E f (B̃) . (D.35) Proof: Let L = N − Ñ and A ∼ Binomial(L, p), independent of B . As the sum of independent binomial random variables with the same success probability is also binomially distributed, we have B + A is equal in distribution to B̃ ∼ Binomial(Ñ , p). Thus, h i E f (B̃) = E [f (B + A)] ≥ E [f (B)] , (D.36) (D.37) where the inequality holds pointwise for any given A = a, and thus also under expectation. . As is well known from the method of types, P[(U n , y n ) ∈ Tn (QU Y )] = exp[−nIQ (U ; Y )]. The next lemma shows that this probability can be upper bounded asymptotically even when IQ (U ; Y ) = 0, and the large-deviations behavior does not hold. 51 Lemma 21. Let QU Y ∈ P(U × Y) be given such that supp(QU ) ≥ 2 and supp(QY ) ≥ 2. Also let y n ∈ Tn (QY ) and assume that U n is distributed uniformly over Tn (QU ). Then, for any given ǫ > 0 there exists n0 (QU , QY ) such that for all n ≥ n0 (QU , QY ) P [(U n , y n ) ∈ Tn (QU Y )] ≤ ǫ. (D.38) Proof: Using Robbins’ sharpening of Stirling’s formula (e.g., [4, Problem 2.2]), it can be shown that the size of a type class satisfies      1 supp(Q ) − 1 log n X |Tn (QX )| ∼ n− 2 [supp(QX )−1] . = exp n · H(QX ) − 2 n (D.39) Hence, P [(U n , y n ) ∈ Tn (QU Y )] (a) = P [(U n , Y n ) ∈ Tn (QU Y )] (D.40) = (D.41) |Tn (QU Y )| |Tn (QU )| |Tn (QY )| 1 ∼ = e−n·IQ (U ;Y ) · n− 2 ·[supp(QU Y )−supp(QU )−supp(QY )+1] , (D.42) where (a) holds by symmetry [assuming that Y n is drawn uniformly over Tn (QY )]. Now, by the Kullback-Csiszár-Kemperman-Pinsker inequality [3, Lemma 11.6.1][4, Exercise 3.18] IQ (U ; Y ) = D(QU Y ||QU × QY ) 1 kQU Y − QU × QY k2 , ≥ 2 log 2 (D.43) (D.44) and thus for any given δ > 0 n o Q̃U Y : Q̃U = QU , Q̃Y = QY , IQ̃ (U ; Y ) ≤ δ n ⊆ Q̃U Y : Q̃U = QU , Q̃Y = QY , Q̃U Y − QU × QY def = J (η, QU , QY ), def where η = ≤η o (D.45) (D.46) √ 2δ log 2. Choose η(QU, QY ) such that supp(Q̃U Y ) = supp(QU × QY ) = supp(QU ) · supp(QY ) for all Q̃U Y ∈ J (δ0 , QU , QY ). We consider two cases: Case 1. If QU Y ∈ J (δ, QU , QY ) then it is elementary to verify that in this event, as supp(QU ) ≥ 2 and supp(QY ) ≥ 2 was assumed, supp(QU Y ) − supp(QU ) − supp(QY ) = supp(QU ) · supp(QY ) − supp(QU ) − supp(QY ) (D.47) ≥ 0, (D.48) 52 and thus (D.42) implies that 1 P [(U n , y n ) ∈ Tn (QU Y )] ≤ √ . n Case 2. If QU Y 6∈ J (δ, QU , QY ) then IQ (U ; Y ) ≥ δ = 1 2 2 log 2 η (D.49) and 1 2 P [(U n , y n ) ∈ Tn (QU Y )] ≤ e−n 2 log 2 η . (D.50) This completes the proof. We are now ready to state and prove the asymptotic uncorrelation lemma for QU Y 6= QU Y . Lemma 22. Let (QU XY , QU XY ) be given such that QU Y 6= QU Y . Then, i i i h h h . λ 1−λ λ 1−λ E Myn (QU XY )Myn (QU XY ) = E Myn (QU XY ) · E Myn (QU XY ) . (D.51) Proof: We begin by lower bounding the correlation. To this end, let us decompose Myn (QU XY ) = Myn (QU XY , 1) + Myn (QU XY , 2), (D.52) where Myn (QU XY , 1) is the enumerator of to the subcode of Cn of codewords which pertain to half of the cloud centers (say, for cloud centers with odd indices) and Myn (QU XY , 2) is the enumerator pertaining to the rest of the codewords. Note that Myn (QU XY , 1) and Myn (QU XY , 2) are independent for any given (QU XY , QU XY ). Hence, i h (QU XY )Myλn (QU XY ) E My1−λ n n λ o  = E [Myn (QU XY , 1) + Myn (QU XY , 2)]1−λ · Myn (QU XY , 1) + Myn (QU XY , 2) h i (a) ≥ E My1−λ (QU XY , 1) · Myλn (QU XY , 2) n i h h i (b) (QU XY , 1) · E Myλn (QU XY , 2) = E My1−λ n (D.53) (D.54) (D.55) (c) i i h h . (QU XY ) · E Myλn (QU XY ) , = E My1−λ n (D.56) where (a) follows since enumerators are positive, and (b) follows since Myn (QU XY , 1) and Myn (QU XY , 2) are independent for any given (QU XY , QU XY ). (c) holds true since Myn (QU XY , i) (i = 1, 2) pertain to codebooks of . . cloud-center size 21 enρc = enρc and satellite rate ρs , and thus clearly E[Myλn (QU XY , i)] = E[Myλn (QU XY )]. To derive an upper bound on the correlation, we first note two properties. First, as the codewords enumerated by Myn (QU XY ) necessarily correspond to different cloud centers from the codewords enumerated by Myn (QU XY ), the following Markov relation holds: Myn (QU XY ) − Nyn (QU Y ) − Nyn (QU Y ) − Myn (QU XY ). (D.57) Second, conditioned on Nyn (QU Y ), Nyn (QU Y ) is a binomial random variable pertaining to enρc −Nyn (QU Y ) ≤ enρc 53 trials with a probability of success given by h i P (U n , y n ) ∈ Tn (QU Y ) (U n , y n ) 6∈ Tn (QU Y ) . (D.58) However, this probability is not significantly different than the unconditional probability. More rigorously, for any ǫ ∈ (0, 1), and all n > n0 (QU , QY ) sufficiently large h i P [(U n , y n ) ∈ Tn (QU Y )] ≤ P (U n , y n ) ∈ Tn (QU Y ) (U n , y n ) 6∈ Tn (QU Y ) ≤ 1 · P [(U n , y n ) ∈ Tn (QU Y )] . 1−ǫ (D.59) To see this, note that since QU Y 6= QU Y we must have supp(QY ) ≥ 2, and thus Lemma 21 implies that for any ǫ ∈ (0, 1), and all n sufficiently large h i P (U n , y n ) ∈ Tn (QU Y ) (U n , y n ) 6∈ Tn (QU Y )   P (U n , y n ) ∈ Tn (QU Y ) ∩ (U n , y n ) 6∈ Tn (QU Y )   = P (U n , y n ) 6∈ Tn (QU Y ) P [(U n , y n ) ∈ Tn (QU Y )] ≤ 1−ǫ 1 · P [(U n , y n ) ∈ Tn (QU Y )] , ≤ 1−ǫ (D.60) (D.61) (D.62) and, similarly, h i P (U n , y n ) ∈ Tn (QU XY ) (U n , X n , y n ) 6∈ Tn (QU XY ) ≥ P [(U n , X n , y n ) ∈ Tn (QU XY )] . (D.63) Equipped with the Markov relation (D.57) and the bound on the conditional probability (D.59), we can derive an upper bound on the correlation that asymptotically matches the lower bound (D.56), and thus prove the lemma. Indeed, i h (QU XY ) · Myλn (QU XY ) E My1−λ n    (a) 1−λ λ = E E Myn (QU XY ) · Myn (QU XY ) Nyn (QU Y ), Nyn (QU Y )      (b) = E E My1−λ (QU XY ) Nyn (QU Y ) · E Myλn (QU XY ) Nyn (QU Y ) n "     (D.64) (D.65) !# (D.66) "  !#    (QU XY ) Nyn (QU Y ) · E E Myλn (QU XY ) Nyn (QU Y ) Nyn (QU Y ) = E E My1−λ n (D.67) (c) = E E E My1−λ (QU XY ) Nyn (QU Y ) · E Myλn (QU XY ) Nyn (QU Y ) n Nyn (QU Y ) (d)       λ n n n ≤ E E My1−λ (Q ) ) ), N (Q ) = 0 N (Q ) · E E M (Q N (Q n U XY y UY y UY y U XY UY yn !   h i 1−λ λ , = E Myn (QU XY ) · E E Myn (QU XY ) Nyn (QU Y ), Nyn (QU Y ) = 0 (D.68) (D.69) 54 where (a) and (c) follows from the law of total expectation, and (b) follows from the Markov relation (D.57). To see (d) note that E[Myλn (QU XY )|Nyn (QU Y ) = s] is a non-decreasing function of s [see (D.8)]. In addition, conditioned on Nyn (QU Y ) = s, Nyn (QU Y ) is a binomial random variable pertaining to less enρc − Nyn (QU Y ) ≤ enρc trials. Thus, (d) follows from Lemma 20. We now note that    λ n n E E Myn (QU XY ) Ny (QU Y ), Ny (QU Y ) = 0 (D.70) and    h i λ = E Myλn (QU XY ) E E Myn (QU XY ) Nyn (QU Y ) (D.71) are both moments of binomial random variables with the same number of trials, but the former has a success probability . P [(U n , y n ) ∈ Tn (QU Y )] = e−nIQ (U ;Y ) , (D.72) and the latter has a success probability   P (U n , y n ) ∈ Tn (QU Y )| (U n , y n ) 6∈ Tn (QU Y ) . (D.73) However, (D.59) shows that the latter success probability has the same exponential order. In turn, the proof of Lemma 19 shows that the exponential order of this expectation only depends on the exponential order of the success probability. Consequently,    λ E E Myn (QU XY ) Nyn (QU Y ), Nyn (QU Y ) = 0  !  . λ = E E Myn (QU XY ) Nyn (QU Y ) h i = E Myλn (QU XY ) . (D.74) (D.75) Using this in (D.69) completes the proof. Next, we move to the case where the cloud centers may be the same, i.e., QU Y = QU Y . In this case, correlation (QU XY ) and Myλn (QU XY ) is possible even asymptotically. Apparently, this is due to the fact between My1−λ n that Nyn (QU Y ) = 0 with high probability whenever IQ (U ; Y ) > ρc , and thus, in this case, Myn (QU XY ) = (QU XY ) and Myλn (QU XY ) are Myn (QU XY ) = 0 with high probability. However, as we will next show, My1−λ n asymptotically uncorrelated when conditioned on Nyn (QU Y ). To show this, we first need a result analogous to Lemma 21. To this end, we first need to exclude possible QU Y from the discussion. Let us say that (QU X , QU Y ) is a joint-distribution-dictator (JDD) pair if it determines QU XY unambiguously. For example, when |U |= |X |= |Y|= 2, QX|U corresponds to a Z-channel and QY |U corresponds to an S-channel,24 (QU X , QU Y ) is a JDD pair. Clearly, in this case no QU XY 6= QU XY exists with the same (U, X) 24 That is QY |U (0|1) = QX|U (1|0) = 0 and all other transition probabilities are non-zero. 55 and (U, Y ) marginals, and thus such QU Y are of no interest to the current discussion. By carefully observing the Z-channel/S-channel example above, it is easy to verify if for all u ∈ U either supp(QX|U =u∗ ) < 2 or supp(QY |U =u∗ ) < 2 then (QU X , QU Y ) is a JDD pair. Therefore, if (QU X , QU Y ) is not a JDD pair then there must exist u∗ ∈ U such that both supp(QX|U =u∗ ) ≥ 2 and supp(QY |U =u∗ ) ≥ 2. This property will be used in the proof of the following lemma. Lemma 23. Let QU XY ∈ P(U × X × Y) be given such that supp(QX ) ≥ 2, supp(QY ) ≥ 2, and (QU X , QU Y ) is not a JDD pair. Also, let (un , y n ) ∈ Tn (QU Y ) and assume that X n is distributed uniformly over Tn (QX|U , un ). Then, for any given ǫ > 0 there exists n0 (QU X , QU Y ) such that for all n ≥ n0 (QU X , QU Y ) P [(un , X n , y n ) ∈ Tn (QU XY )] ≤ ǫ. (D.76) Proof: Had X n been distributed uniformly over Tn (QX ), the claim would follow directly from Lemma 21, where y n and U n there are replaced by (un , y n ) and X n , respectively. However, since X n is distributed uniformly over Tn (QX|U , un ) the proof is not immediate. Nonetheless, it follows the same lines, and thus we will only highlight the required modifications. Just as in (D.39), the size of a conditional type class can be shown to satisfy Tn (QX|U , un ) ∼ = 1 Y u∈supp(QU ) exp {nQU (u) · HQ (X|U = u)} · [nQU (u)]− 2 [supp(QX|U =u )−1] (D.77) and thus P [(un , X n , y n ) ∈ Tn (QU XY )] = (a) ∼ = Tn (QX|U Y , un , y n ) Q (D.78) Tn (QX|U , un ) (u,y)∈supp(QU Y ) exp {nQU Y (u, y) 1 · HQ (X|U = u, Y = Y )} · [nQU Y (u, y)]− 2 [supp(QX|U =u,Y =y )−1] 1 − 2 [supp(QX|U =u )−1] u∈supp(QU ) exp {nQU (u) · HQ (X|U = u)} · [nQU (u)] Q 1 = e−n·IQ (X;Y |U ) · n 2 P 1 u∈supp(QU ) [supp(QX|U =u )−1]− 2 where def c(QU Y ) = Q X u∈supp(QU ) (a) = 1 2 X  Q u∈supp(QU ) 1 X (u,y)∈supp(QU Y )  1 supp(QX|U =u ) − 1 − 2 · c(QU Y ), (D.80) . (D.81) 1 (u, y)− 2 [supp(QX|U =u,Y =y )−1] − 2 [supp(QX|U =u )−1] u∈supp(QU ) QU (u)  1 supp(QX|U =u ) − 1 − 2  (u,y)∈supp(QU Y ) [supp(QX|U =u,Y =y )−1] (u,y)∈supp(QU Y ) QU Y Now, suppose that IQ (X; Y |U ) = 0. Then, 1 2 P (D.79) X   supp(QX|U =u,Y =y ) − 1 (u,y)∈supp(QU Y )   supp(QX|U =u ) − 1 (D.82) 56 = 1 2 1 = 2 X   X  u∈supp(QU ) (b) u∈supp(QU ) 1 ≤− , 2   supp(QX|U =u ) − 1 − X y∈supp(QY |U =u )    supp(QX|U =u ) − 1    1 − supp(QY |U =u ) supp(QX|U =u ) − 1 (D.83) (D.84) (D.85) where (a) follows since QX|U =u = QX|U =u,Y =y for all u ∈ supp(QU ), and (b) follows since (QU X , QU Y ) is not a JDD pair, and thus there must exist u∗ ∈ supp(QU ) such that both supp(QX|U =u∗ ) ≥ 2 and supp(QX|U =u∗ ) ≥ 2 (as noted before the statement of the lemma). Thus, when IQ (X; Y |U ) = 0 we get P [(un , X n , y n ) ∈ Tn (QU XY )] ≤ c(QU Y ) √ . n (D.86) The proof then may continue as the proof of Lemma 21. One can find η > 0 sufficiently small such that any n o Q̃U XY ∈ Q̃U XY : kQ̃U XY − QU × QX|U × QY |U k1 ≤ η has the same support as QU × QX|U × QY |U (at least conditioned on u∗ ). Further, one can find δ(η, QU X , QU Y ) > 0 such that n o n o Q̃U XY : IQ̃ (X; Y |U ) ≤ δ ⊆ Q̃U XY : kQ̃U XY − QU × QX|U × QY |U k1 ≤ η , (D.87) and the two cases considered in Lemma 21 can be considered here as well. In the first case, IQ (X; Y |U ) may vanish, but supp(QU XY ) = supp(QU × QX|U × QY |U ) and thus (D.86) holds. In the second case IQ (X; Y |U ) ≥ δ, and thus for any given ǫ > 0 there exists n0 (QU X , QU Y ) such that P [(un , X n , y n ) ∈ Tn (QU XY )] ≤ e−nIQ (X;Y |U ) · n|U ||X | · c(QU Y ) ≤ ǫ (D.88) for all n ≥ n0 (QU X , QU Y ). Lemma 24. Let (QU XY , QU XY ) be given such that QU Y = QU Y . Then,       . 1−λ λ 1−λ λ E Myn (QU XY )Myn (QU XY ) Nyn (QU Y ) = E Myn (QU XY ) Nyn (QU Y ) · E Myn (QU XY ) Nyn (QU Y ) . (D.89) Proof: The proof follows the same lines of the proof of Lemma 22, and so we only provide a brief outline. For a lower bound on the conditional correlation, one can decompose Myn (QU XY ) = Myn (QU XY , 1) + Myn (QU XY , 2) (D.90) where Myn (QU XY , 1) [respectively, Myn (QU XY , 2)] corresponds to codewords of odd (even) satellite indices (say). For a asymptotically matching upper bound on the correlation, we note that similarly to (D.59), when X n is drawn uniformly over Tn (QX|U , un ), h i P (un , X n , y n ) ∈ Tn (QU XY ) (un , X n , y n ) 6∈ Tn (QU XY ) (D.91) 57 is close to the unconditional probability, in the sense that for any given ǫ > 0, h i P [(un , X n , y n ) ∈ Tn (QU XY )] ≤ P (U n , y n ) ∈ Tn (QU Y ) (U n , y n ) 6∈ Tn (QU Y ) ≤ 1 · P [(un , X n , y n ) ∈ Tn (QU XY )] . (D.92) 1−ǫ To prove this, a derivation similar to (D.62) can be used, while noting that (QU X , QU Y ) is not a JDD pair, and so according to Lemma 23   P (un , X n , y n ) ∈ Tn (QU XY ) ≤ ǫ for all n sufficiently large. Equipped with these results, we get   λ 1−λ E Myn (QU XY )Myn (QU XY ) Nyn (QU Y )     (a) 1−λ λ = E Myn (QU XY ) · E Myn (QU XY ) Nyn (QU Y ), Myn (QU XY ) Nyn (QU Y )     (b) λ 1−λ ≤ E Myn (QU XY ) · E Myn (QU XY ) Nyn (QU Y ), Myn (QU XY ) = 0 Nyn (QU Y ) (D.93) (D.94) (D.95) (c)     . λ 1−λ = E Myn (QU XY ) Nyn (QU Y ) · E Myn (QU XY ) Myn (QU XY ) , (D.96) where (a) follows from the law of total expectation. For (b) note that conditioned on both Nyn (QU Y ), Myn (QU XY ), Myn (QU XY ) is a binomial random variable pertaining to Nyn (QU Y )enρs − Myn (QU XY ) ≤ Nyn (QU Y )enρs trials. (QU XY )|Nyn (QU Y ), Myn (QU XY ) = s] is a non-increasing function of s [see (D.8)], and (b) follows Thus, E[My1−λ n from Lemma 20. For (c), we note that from (D.92), the conditioning on Myn (QU XY ) = 0 does not change the exponential order of the success probability of Myn (QU XY ). As evident from (D.8), this conditioning can be removed without changing the exponential order of the expression. Proceeding with the case of QU Y = QU Y , we next evaluate the expectation over Nyn (QU Y ). We show that the asymptotic uncorrelation result of Lemma 22 holds, albeit with a correction term required when IQ (U ; Y ) > ρc . Lemma 25. Let δ > 0, and (QU XY , QU XY ) be given such that QU Y = QU Y . Then, i i h i h h . 1−λ λ λ ) ) · en|IQ (U ;Y )−ρc |+ . = E M (Q ) · E M (Q E My1−λ (Q )M (Q n U XY U XY U XY U XY yn yn yn (D.97) Proof: We consider two cases separately. First suppose that IQ (U ; Y ) ≤ ρc . In this case, Nyn (QU Y ) concentrates double-exponentially fast around its expected value, where the latter equals exp[n(ρc − IQ (U ; Y ))] up to the first order in the exponent. Thus, the conditional expectation and the unconditional expectation are equal up to the first order in the exponent. More rigorously, let δ > 0 be given and recall the definition of the event Bn (QU Y , δ) in (D.2). Then,   h i 1−λ 1−λ E Myn (QU XY ) = P [Nyn (QU Y ) ∈ Bn (QU Y , δ)] · E Myn (QU XY ) Nyn (QU Y ) ∈ Bn (QU Y , δ) 58   c n N (Q ) ∈ B (Q , δ) (Q ) + P [Nyn (QU Y ) ∈ Bnc (QU Y , δ)] · E My1−λ n y UY UY U XY n (D.98) (a)   . 1−λ n = E Myn (QU XY ) Ny (QU Y ) ∈ Bn (QU Y , δ)   (b) −nδ 1−λ ≥e · E Myn (QU XY ) Nyn (QU Y ) = E [Nyn (QU Y )] , (D.99) (D.100) where (a) follows from the fact P[Bn (QU Y , δ)] decays double-exponentially [see (D.3)], and (b) follows from (D.8). Similarly   h i λ −nδ λ ˙ E Myn (QU XY ) ≥ e · E Myn (QU XY ) Nyn (QU Y ) = E [Nyn (QU Y )] . (D.101) Thus, h i (QU XY )Myλn (QU XY ) E My1−λ n   1−λ λ = P [Nyn (QU Y ) ∈ Bn (QU Y , δ)] · E Myn (QU XY )Myn (QU XY ) Nyn (QU Y ) ∈ Bn (QU Y , δ)   + P [Nyn (QU Y ) ∈ Bnc (QU Y , δ)] · E My1−λ (QU XY )Myλn (QU XY ) Nyn (QU Y ) ∈ Bnc (QU Y , δ) n (D.102) (a)   . λ n = E My1−λ (Q )M (Q ) N (Q ) ∈ B (Q , δ) n U XY y UY n UY U XY yn   X λ 1−λ = P [Nyn (QU Y ) = s] · E Myn (QU XY )Myn (QU XY ) Nyn (QU Y ) = s (D.103) (D.104) s∈Bn (QU Y ,δ) (b) . = X s∈Bn (QU Y ,δ) P [Nyn (QU Y ) = s] · E  (QU XY My1−λ n  ) Nyn (QU Y ) = s · E  Myλn (QU XY ) Nyn (QU Y ) = s  (D.105) (c) ≤ e2nδ · P [Nyn (QU Y ) ∈ Bn (QU Y , δ)]     1−λ λ × E Myn (QU XY ) Nyn (QU Y ) = E [Nyn (QU Y )] · E Myn (QU XY ) Nyn (QU Y ) = E [Nyn (QU Y )] (D.106)     (d) . λ n n n n = e2nδ · E My1−λ (Q ) N (Q ) = E [N (Q )] · E M (Q ) N (Q ) = E [N (Q )] n U XY y UY y UY y UY y UY U XY yn (D.107) (e) o n n o ˙ e4nδ · E My1−λ (QU XY ) · E Myλn (QU XY ) , ≤ n (D.108) where (a) and (d) follow from (D.3), (b) follows from Lemma 24, (c) from (D.8), and (e) from (D.100) and (D.101). We next address the case IQ (U ; Y ) > ρc . In this case, Nyn (QU Y ) = 0 with high probability, 1 ≤ Nyn (QU Y ) ≤ e2nδ with probability exp{−n[IQ (U ; Y ) − ρc ]}, and Nyn (QU Y ) ≥ e2nδ with probability double-exponentially small =1 ≥1 [see (D.4) and (D.5)]. For brevity, we will use the definitions of A=0 n (QU Y ), An (QU Y ) and An (QU Y , δ) in 59 (D.11)-(D.13). Then,   i h  =0  =0 1−λ 1−λ E Myn (QU XY ) = P An (QU Y ) · E Myn (QU XY ) An (QU Y )    =1  =1 1−λ + P An (QU Y , δ) · E Myn (QU XY ) An (QU Y , δ)     ≥1 1−λ A (Q , δ) (Q ) + P A≥1 (Q , δ) · E M n UY U XY UY n n y (D.109) (a)     . 1−λ =1 = 0 + P A=1 (Q , δ) · E M (Q ) A (Q , δ) UY U XY UY yn n n   (b)  =1  1−λ nδ ≤ e · P An (QU Y , δ) · E Myn (QU XY ) Nyn (QU Y ) = 1 , where for (a) and (b) we apply (D.5) and (D.8), respectively. Similarly,   h i  =1  λ nδ λ ˙ E Myn (QU XY ) ≤ e · P An (QU Y , δ) · E Myn (QU XY ) Nyn (QU Y ) = 1 . (D.110) (D.111) (D.112) Then, i h E My1−λ (QU XY )Myλn (QU XY ) n     (a)  =1  2nδ 1−λ λ ˙ ≤e · P An (QU Y , δ) · E Myn (QU XY ) Nyn (QU Y ) = 1 · E Myn (QU XY ) Nyn (QU Y ) = 1 i h i h 1 1−λ λ ˙ e4nδ · ) · E M (Q ) · E M (Q ≤ n n U XY U XY y y P [A=1 n (QU Y , δ)] i h h i (b) . (QU XY ) · E Myλn (QU XY ) , = e4nδ · en[IQ (U ;Y )−ρc ] · E My1−λ n (D.113) (D.114) (D.115) where (a) follows from a derivation similar to (D.108), and (b) follows from (D.4). (QU XY )Myλn (QU XY )] for both cases can be obtained in the Analogous asymptotic lower bounds on E[My1−λ n same manner, when δ is replaced by (−δ). The proof is then completed by taking δ ↓ 0. Using all the above, we are now ready to prove Proposition 15. Proof of Prop. 15: For the first case of (C.4) h i (QU XY )Myλn (QU XY ) = E [Myn (QU XY )] , E My1−λ n (D.116) and the result follows from Lemma 19 with λ = 1. For the second case, using Lemma 22 i i h i h h . (QU XY ) · E Myλn (QU XY ) , E My1−λ (QU XY )Myλn (QU XY ) = E My1−λ n n (D.117) and the result follows from Lemma 19. Similarly, the third case follows from Lemmas 25 and 19. Specifically, the result is just as in the second case, except for the correction term |IQ (U ; Y ) − ρc |+ to the exponent. Standard manipulations lead to the expression shown in the third case. 60 R EFERENCES [1] H. V. Poor, An Introduction to Signal Detection and Estimation, 2nd ed. Springer-Verlag, 1994. [2] R. E. Blahut, “Hypothesis testing and information theory,” IEEE Transactions on Information Theory, vol. 20, no. 4, pp. 405–417, July 1974. [3] T. M. Cover and J. A. Thomas, Elements of Information Theory (Wiley Series in Telecommunications and Signal Processing). WileyInterscience, 2006. [4] I. Csiszár and J. Körner, Information Theory: Coding Theorems for Discrete Memoryless Systems. Cambridge University Press, 2011. [5] I. Csiszár and P. C. Shields, Information Theory and Statistics: A Tutorial, ser. Foundations and Trends in Communications and Information Theory. Now Publishers Inc, 2004. [6] T. Berger, “Decentralized estimation and decision theory,” in IEEE Seven Springs Workshop on Information Theory, Mt. Kisco, NY, 1979. [7] R. Ahlswede and I. Csiszár, “Hypothesis testing with communication constraints,” IEEE Transactions on Information Theory, vol. 32, no. 4, pp. 533–542, July 1986. [8] R. Ahlswede, P. Gács, and J. Körner, “Bounds on conditional probabilities with applications in multi-user communication,” Probability Theory and Related Fields, vol. 34, no. 2, pp. 157–177, 1976. [9] R. Ahlswede and J. Körner, “Source coding with side information and a converse for degraded broadcast channels,” IEEE Transactions on Information Theory, vol. 21, no. 6, pp. 629–637, November 1975. [10] T. S. Han and S. Amari, “Statistical inference under multiterminal data compression,” IEEE Transactions on Information Theory, vol. 44, no. 6, pp. 2300–2324, October 1998. [11] T. S. Han, “Hypothesis testing with multiterminal data compression,” IEEE Transactions on Information Theory, vol. 33, no. 6, pp. 759–772, November 1987. [12] T. S. Han and K. Kobayashi, “Exponential-type error probabilities for multiterminal hypothesis testing,” IEEE Transactions on Information Theory, vol. 35, no. 1, pp. 2–14, January 1989. [13] H. Shimokawa, T. S. Han, and S. Amari, “Error bound of hypothesis testing with data compression,” in Proc. IEEE International Symposium Information Theory (ISIT), June 1994, pp. 114–. [14] Y. Polyanskiy, “Hypothesis testing via a comparator,” in Proc. IEEE International Symposium Information Theory (ISIT), July 2012, pp. 2206–2210, extended version available online: http://people.lids.mit.edu/yp/homepage/data/htstruct_journal.pdf. [15] H. M. H. Shalaby and A. Papamarcou, “Multiterminal detection with zero-rate data compression,” IEEE Transactions on Information Theory, vol. 38, no. 2, pp. 254–267, March 1992. [16] C. Tian and J. J. Chen, “Successive refinement for hypothesis testing and lossless one-helper problem,” IEEE Transactions on Information Theory, vol. 54, no. 10, pp. 4666–4681, October 2008. [17] M. S. Rahman and A. B. Wagner, “On the optimality of binning for distributed hypothesis testing,” IEEE Transactions on Information Theory, vol. 58, no. 10, pp. 6282–6303, October 2012. [18] G. Katz, P. Piantanida, and M. Debbah, “Collaborative distributed hypothesis testing,” arXiv preprint, April 2016, available online: http://arxiv.org/pdf/1604.01292.pdf. [19] Y. Xiang and Y. H. Kim, “Interactive hypothesis testing against independence,” in Proc. IEEE International Symposium Information Theory (ISIT), July 2013, pp. 2840–2844. [20] M. Mhanna and P. Piantanida, “On secure distributed hypothesis testing,” in Proc. IEEE International Symposium Information Theory (ISIT), June 2015, pp. 1605–1609. [21] G. Katz, P. Piantanida, and M. Debbah, “Distributed binary detection with lossy data compression,” IEEE Transactions on Information Theory, 2017. [22] S. Sreekumar and D. Gündüz, “Distributed hypothesis testing over noisy channels,” in Proc. IEEE International Symposium Information Theory (ISIT), June 2017, pp. 983–987. 61 [23] E. Tuncel, “On error exponents in hypothesis testing,” IEEE Transactions on Information Theory, vol. 51, no. 8, pp. 2945–2950, August 2005. [24] S. Salehkalaibar, M. Wigger, and R. Timo, “On hypothesis testing against independence with multiple decision centers,” arXiv preprint, August 2017, available online: http://arxiv.org/pdf/1708.03941.pdf. [25] S. Salehkalaibar, M. Wigger, and L. Wang, “Hypothesis testing in multi-hop networks,” arXiv preprint, August 2017, available online: http://arxiv.org/pdf/1708.05198.pdf. [26] W. Zhao and L. Lai, “Distributed testing with zero-rate compression,” in Proc. IEEE International Symposium Information Theory (ISIT), June 2015, pp. 2792–2796. [27] S. Watanabe, “Neyman-Pearson test for zero-rate multiterminal hypothesis testing,” arXiv preprint, November 2016, available online: http://arxiv.org/pdf/1611.08175.pdf. [28] A. El Gamal and Y. Kim, Network information theory. Cambridge university press, 2011. [29] A. Wyner and J. Ziv, “The rate-distortion function for source coding with side information at the decoder,” IEEE Transactions on Information Theory, vol. 22, no. 1, pp. 1–10, January 1976. [30] S. I. Gel’fand and M. S. Pinsker, “Coding of sources on the basis of observations with incomplete information,” Problemy Peredachi Informatsii, vol. 15, no. 2, pp. 45–57, 1979. [31] G. Katz, R. Couillet, P. Piantanida, and M. Debbah, “On the necessity of binning for the distributed hypothesis testing problem,” in Proc. IEEE International Symposium Information Theory (ISIT), June 2015, pp. 2797–2801. [32] E. Haim and Y. Kochman, “Binary distributed hypothesis testing via Körner-Marton coding,” in Proc. IEEE Information Theory Workshop (ITW), Septmber 2016, pp. 146–150. [33] ——, “On binary distributed hypothesis testing,” Submitted to IEEE Transactions on Information Theory, December 2017, available online: http://arxiv.org/pdf/1801.00310.pdf. [34] J. Körner and K. Marton, “How to encode the modulo-two sum of binary sources (corresp.),” IEEE Transactions on Information Theory, vol. 25, no. 2, pp. 219–221, March 1979. [35] G. Katz, P. Piantanida, and M. Debbah, “A new approach to distributed hypothesis testing,” in Proc. Asilomar Conference on Signals, Systems and Computers, November 2016, pp. 1365–1369. [36] W. Hoeffding, “Asymptotically optimal tests for multinomial distributions,” The Annals of Mathematical Statistics, pp. 369–401, 1965. [37] D. Slepian and J. K. Wolf, “Noiseless coding of correlated information sources,” IEEE Transactions on Information Theory, vol. 19, no. 4, pp. 471–480, July 1973. [38] R. Ahlswede and G. Dueck, “Good codes can be produced by a few permutations,” IEEE Transactions on Information Theory, vol. 28, no. 3, pp. 430–443, May 1982. [39] J. Chen, D. He, A. Jagmohan, and L. A. Lastras-Montaño, “On the reliability function of variable-rate Slepian-Wolf coding,” in Proc. of 45th Annual Allerton Conference Communication, Control, and Computing, September 2007, extended version available online: http://www.ece.mcmaster.ca/~junchen/IT_variableSW.pdf. [40] N. Weinberger and N. Merhav, “Optimum trade-offs between the error exponent and the excess-rate exponent of variable-rate SlepianWolf coding,” IEEE Transactions on Information Theory, vol. 61, no. 4, pp. 2165–2190, April 2015, extended version available online: http://arxiv.org/pdf/1401.0892v3.pdf. [41] R. Ahlswede, “Coloring hypergraphs: A new approach to multi-user source coding, part II,” Journal of Combinatorics, vol. 5, pp. 220–268, 1980. [42] N. Weinberger and N. Merhav, “Channel detection in coded communication,” IEEE Transactions on Information Theory, vol. 63, no. 10, pp. 6364–6392, October 2017. [43] P. Bergmans, “Random coding theorem for broadcast channels with degraded components,” IEEE Transactions on Information Theory, vol. 19, no. 2, pp. 197–207, March 1973. [44] I. Csiszár, “The method of types,” IEEE Transactions on Information Theory, vol. 44, no. 6, pp. 2505–2523, October 1998. [45] W. Rudin, Principles of mathematical analysis, 3rd ed. McGraw-Hill New York, 1976. 62 [46] N. Weinberger and N. Merhav, “A large deviations approach to secure lossy compression,” IEEE Transactions on Information Theory, vol. 63, no. 4, pp. 2533–2559, April 2017. [47] K. Marton, “Error exponent for source coding with a fidelity criterion,” IEEE Transactions on Information Theory, vol. 20, no. 2, pp. 197–199, March 1974. [48] N. Merhav, “Statistical physics and information theory,” Foundations and Trends in Communications and Information Theory, vol. 6, no. 1-2, pp. 1–212, 2009. [49] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge university press, 2004. [50] I. Csiszár and G. Tusnády, “Information Geometry and Alternating Minimization Procedures,” Statistics and Decisions, vol. Supplement Issue 1, 1984. [51] S. Arimoto, “An algorithm for computing the capacity of arbitrary discrete memoryless channels,” IEEE Transactions on Information Theory, vol. 18, no. 1, pp. 14–20, 1972. [52] ——, “Computation of random coding exponent functions,” IEEE Transactions on Information Theory, vol. 22, no. 6, pp. 665–671, November 1976. [53] R. E. Blahut, “Computation of channel capacity and rate-distortion functions,” IEEE Transactions on Information Theory, vol. 18, no. 4, pp. 460–473, 1972. [54] N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,” in Proc. of 37th Annual Allerton Conference Communication, Control, and Computing, September 1999. [55] A. Rényi, “On measures of entropy and information,” in Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics. The Regents of the University of California, 1961. [56] T. Van Erven and P. Harremos, “Rényi divergence and Kullback-Leibler divergence,” IEEE Transactions on Information Theory, vol. 60, no. 7, pp. 3797–3820, 2014. [57] A. Nazari, A. Anastasopoulos, and S. S. Pradhan, “Error exponent for multiple-access channels: Lower bounds,” IEEE Transactions on Information Theory, vol. 60, no. 9, pp. 5095–5115, September 2014. [58] A. Nazari, S. S. Pradhan, and A. Anastasopoulos, “Error exponent for multiple-access channels: Upper bounds,” IEEE Transactions on Information Theory, vol. 61, no. 7, pp. 3605–3621, July 2015. [59] M. Sion, “On general minimax theorems,” Pacific Journal of Mathematics, vol. 8, no. 1, pp. 171–176, 1958. [60] A. Somekh-Baruch and N. Merhav, “Achievable error exponents for the private fingerprinting game,” IEEE Transactions on Information Theory, vol. 53, no. 5, pp. 1827–1838, May 2007. [61] N. Shulman, “Communication over an unknown channel via common broadcasting,” Ph.D. dissertation, Dept. Electrical Engineering, Tel Aviv University, Tel Aviv, Israel, 2003, http://www.eng.tau.ac.il/~shulman/papers/Nadav_PhD.pdf. [62] M. Loève, Probability Theory I. Springer, 1977. [63] N. Merhav, “Relations between random coding exponents and the statistical physics of random codes,” IEEE Transactions on Information Theory, vol. 55, no. 1, pp. 83–92, January 2009.
7cs.IT
arXiv:1802.07740v2 [cs.AI] 12 Mar 2018 Machine Theory of Mind Neil C. Rabinowitz∗ DeepMind [email protected] Frank Perbet DeepMind [email protected] H. Francis Song DeepMind [email protected] Chiyuan Zhang Google Brain [email protected] S. M. Ali Eslami DeepMind [email protected] Matthew Botvinick DeepMind [email protected] Abstract Theory of mind (ToM; Premack & Woodruff, 1978) broadly refers to humans’ ability to represent the mental states of others, including their desires, beliefs, and intentions. We propose to train a machine to build such models too. We design a Theory of Mind neural network – a ToMnet – which uses meta-learning to build models of the agents it encounters, from observations of their behaviour alone. Through this process, it acquires a strong prior model for agents’ behaviour, as well as the ability to bootstrap to richer predictions about agents’ characteristics and mental states using only a small number of behavioural observations. We apply the ToMnet to agents behaving in simple gridworld environments, showing that it learns to model random, algorithmic, and deep reinforcement learning agents from varied populations, and that it passes classic ToM tasks such as the “SallyAnne” test (Wimmer & Perner, 1983; BaronCohen et al., 1985) of recognising that others can hold false beliefs about the world. We argue that this system – which autonomously learns how to model other agents in its world – is an important step forward for developing multi-agent AI systems, for building intermediating technology for machine-human interaction, and for advancing the progress on interpretable AI. * Corresponding author: [email protected]. 1. Introduction For all the excitement surrounding deep learning and deep reinforcement learning at present, there is a concern from some quarters that our understanding of these systems is lagging behind. Neural networks are regularly described as opaque, uninterpretable black-boxes. Even if we have a complete description of their weights, it’s hard to get a handle on what patterns they’re exploiting, and where they might go wrong. As artificial agents enter the human world, the demand that we be able to understand them is growing louder. Let us stop and ask: what does it actually mean to “understand” another agent? As humans, we face this challenge every day, as we engage with other humans whose latent characteristics, latent states, and computational processes are almost entirely inaccessible. Yet we function with remarkable adeptness. We can make predictions about strangers’ future behaviour, and infer what information they have about the world; we plan our interactions with others, and establish efficient and effective communication. A salient feature of these “understandings” of other agents is that they make little to no reference to the agents’ true underlying structure. We do not typically attempt to estimate the activity of others’ neurons, infer the connectivity of their prefrontal cortices, or plan interactions with a detailed approximation of the dynamics of others’ hippocampal maps. A prominent argument from cognitive psychology is that our social reasoning instead relies on highlevel models of other agents (Gopnik & Wellman, 1992). These models engage abstractions which do not describe the detailed physical mechanisms underlying observed be- Machine Theory of Mind haviour; instead, we represent the mental states of others, such as their desires, beliefs, and intentions. This ability is typically described as our Theory of Mind (Premack & Woodruff, 1978). While we may also, in some cases, leverage our own minds to simulate others’ (e.g. Gordon, 1986; Gallese & Goldman, 1998), our ultimate human understanding of other agents is not measured by a 1-1 correspondence between our models and the mechanistic ground truth, but instead by how much these models afford for tasks such as prediction and planning (Dennett, 1991). In this paper, we take inspiration from human Theory of Mind, and seek to build a system which learns to model other agents. We describe this as a Machine Theory of Mind. Our goal is not to assert a generative model of agents’ behaviour and an algorithm to invert it. Rather, we focus on the problem of how an observer could learn autonomously how to model other agents using limited data (Botvinick et al., 2017). This distinguishes our work from previous literature, which has relied on hand-crafted models of agents as noisy-rational planners – e.g. using inverse RL (Ng et al., 2000; Abbeel & Ng, 2004), Bayesian inference (Lucas et al., 2014; Evans et al., 2016), Bayesian Theory of Mind (Baker et al., 2011; Jara-Ettinger et al., 2016; Baker et al., 2017) or game theory (Camerer et al., 2004; Yoshida et al., 2008; Camerer, 2010; Lanctot et al., 2017). In contrast, we learn the agent models, and how to do inference on them, from scratch, via meta-learning. Building a rich, flexible, and performant Machine Theory of Mind may well be a grand challenge for AI. We are not trying to solve all of this here. A main message of this paper is that many of the initial challenges of building a ToM can be cast as simple learning problems when they are formulated in the right way. Our work here is an exercise in figuring out these simple formulations. There are many potential applications for this work. Learning rich models of others will improve decision-making in complex multi-agent tasks, especially where model-based planning and imagination are required (Hassabis et al., 2013; Hula et al., 2015; Oliehoek & Amato, 2016). Our work thus ties in to a rich history of opponent modelling (Brown, 1951; Albrecht & Stone, 2017); within this context, we show how meta-learning could be used to furnish an agent with the ability to build flexible and sampleefficient models of others on the fly. Such models will be important for value alignment (Hadfield-Menell et al., 2016) and flexible cooperation (Nowak, 2006; KleimanWeiner et al., 2016; Barrett et al., 2017; Kris Cao), and will likely be an ingredient in future machines’ ethical decision making (Churchland, 1996). They will also be highly useful for communication and pedagogy (Dragan et al., 2013; Fisac et al., 2017; Milli et al., 2017), and will thus likely play a key role in human-machine interaction. Ex- ploring the conditions under which such abilities arise can also shed light on the origin of our human abilities (Carey, 2009). Finally, such models will likely be crucial mediators of our human understanding of artificial agents. Lastly, we are strongly motivated by the goals of making artificial agents human-interpretable. We attempt a novel approach here: rather than modifying agents architecturally to expose their internal states in a human-interpretable form, we seek to build intermediating systems which learn to reduce the dimensionality of the space of behaviour and represent it in more digestible forms. In this respect, the pursuit of a Machine ToM is about building the missing interface between machines and human expectations (Cohen et al., 1981). 1.1. Our approach We consider the challenge of building a Theory of Mind as essentially a meta-learning problem (Schmidhuber et al., 1996; Thrun & Pratt, 1998; Hochreiter et al., 2001; Vilalta & Drissi, 2002). At test time, we want to be able to encounter a novel agent whom we have never met before, and already have a strong and rich prior about how they are going to behave. Moreover, as we see this agent act in the world, we wish to be able to collect data (i.e. form a posterior) about their latent characteristics and mental states that will enable us to improve our predictions about their future behaviour. To do this, we formulate a meta-learning task. We construct an observer, who in each episode gets access to a set of behavioural traces of a novel agent. The observer’s goal is to make predictions of the agent’s future behaviour. Over the course of training, the observer should get better at rapidly forming predictions about new agents from limited data. This “learning to learn” about new agents is what we mean by meta-learning. Through this process, the observer should also learn an effective prior over the agents’ behaviour that implicitly captures the commonalities between agents within the training population. We introduce two concepts to describe components of this observer network and their functional role. We distinguish between a general theory of mind – the learned weights of the network, which encapsulate predictions about the common behaviour of all agents in the training set – and an agent-specific theory of mind – the “agent embedding” formed from observations about a single agent at test time, which encapsulates what makes this agent’s character and mental state distinct from others’. These correspond to a prior and posterior over agent behaviour. This paper is structured as a sequence of experiments of increasing complexity on this Machine Theory of Mind network, which we call a ToMnet. These experiments Machine Theory of Mind showcase the idea of the ToMnet, exhibit its capabilities, and demonstrate its capacity to learn rich models of other agents incorporating canonical features of humans’ Theory of Mind, such as the recognition of false beliefs. Some of the experiments in this paper are directly inspired by the seminal work of Baker and colleagues in Bayesian Theory of Mind, such as the classic food-truck experiments (Baker et al., 2011; 2017). We have not sought to directly replicate these experiments as the goals of this work differ. In particular, we do not immediately seek to explain human judgements in computational terms, but instead we emphasise machine learning, scalability, and autonomy. We leave the alignment to human judgements as future work. Our experiments should nevertheless generalise many of the constructions of these previous experiments. formalism, we associate the reward functions, discount factors, and conditional observation functions with the agents rather than with the POMDPs. For example, a POMDP could be a gridworld with a particular arrangement of walls and objects; different agents, when placed in the same POMDP, might receive different rewards for reaching these objects, and be able to see different amounts of their local surroundings. The POMDPs are thus tuples of state spaces Sj , action spaces Aj , and transition probabilities Tj only, i.e. Mj = (Sj , Aj , Tj ). In this work, we only consider single-agent POMDPs, though the extension to the multiagent case is simple. When agents have full observability, we use the terms MDP and POMDP interchangeably. S We write the joint state space over all POMDPs as S = j Sj . • In Section 3.2, we show that the ToMnet learns to infer the goals of algorithmic agents (effectively performing few-shot inverse reinforcement learning), as well as how they balance costs and rewards. Separately, we assume we have a family of agents A = S i Ai , with corresponding observation spaces Ωi , conditional observation functions ωi (·) : S → Ωi , reward functions Ri , discount factors γi , and resulting policies πi , i.e. Ai = (Ωi , ωi , Ri , γi , πi ). These policies might be stochastic (as in Section 3.1), algorithmic (as in Section 3.2), or learned (as in Sections 3.3–3.5). We do not assume that the agents’ policies πi are optimal for their respective tasks. The agents may be stateful – i.e. with policies parameterised as πi (·|ωi (st ), ht ) where ht is the agent’s (Markov) hidden state – though we assume agents’ hidden states do not carry over between episodes. • In Section 3.3, we show that the ToMnet learns to characterise different species of deep reinforcement learning agents, capturing the essential factors of variations across the population, and forming abstract embeddings of these agents. We also show that the ToMnet can discover new abstractions about the space of behaviour. In turn, we consider an observer who makes potentially partial and/or noisy observations of agents’ trajectories, via a state-observation function ω (obs) (·) : S → Ω(obs) , and an action-observation function α(obs) (·) : A → A(obs) . Thus, if agent Ai follows its policy πi on POMDP Mj and produces trajectory τij = {(st , at )}Tt=0 , the observer (obs) (obs) (obs) (obs) would see τij = {(xt , at )}Tt=0 , where xt = • In Section 3.4, we show that when the ToMnet is trained on deep RL agents acting in POMDPs, it implicitly learns that these agents can hold false beliefs about the world, a core component of humans’ Theory of Mind. ω (obs) (st ) and at = α(obs) (at ). For all experiments we pursue here, we set ω (obs) (·) and α(obs) (·) as identity functions, so that the observer has unrestricted access to the MDP state and overt actions taken by the agents; the observer does not, however, have access to the agents’ parameters, reward functions, policies, or identifiers. Our contributions are as follows: • In Section 3.1, we show that for simple, random agents, the ToMnet learns to approximate Bayesoptimal hierarchical inference over agents’ characteristics. • In Section 3.5, we show that the ToMnet can be trained to predict agents’ belief states as well, revealing agents’ false beliefs explicitly. We also show that the ToMnet can infer what different agents are able to see, and what they therefore will tend to believe, from their behaviour alone. 2. Model 2.1. The tasks Here we describe the formalisation of the task. We assume we have a family of partiallySobservable Markov decision processes (POMDPs) M = j Mj . Unlike the standard (obs) We set up the meta-learning problem as follows. ToMnet training involves a series of encounters with individual agents, together with a query for which the ToMnet has to make a set of predictions. More precisely, the observer sees a set of full or partial “past episodes”, wherein a single, unNpast labelled agent, Ai , produces trajectories, {τij }j=1 , as it executes its policy within the respective POMDPs, Mj . Generally, we allow Npast to vary, sometimes even setting it to zero. The task for the observer is to predict the agent’s behaviour (e.g. atomic actions) and potentially its latent states (e.g. beliefs) on a “current episode” as it acts within POMDP Mk . The observer may be seeded with a partial trajectory in Mk up to time t. Machine Theory of Mind The observer must learn to predict the behaviour of many agents, whose rewards, parameterisations, and policies may vary considerably; in this respect, the problem resembles the one-shot imitation learning setup recently introduced in Duan et al. (2017) and Wang et al. (2017). However, the problem statement differs from imitation learning in several crucial ways. First, the observer need not be able to execute the behaviours itself: the behavioural predictions may take the form of atomic actions, options, trajectory statistics, or goals or subgoals. The objective here is not to imitate, but instead to form predictions and abstractions that will be useful for a range of other tasks. Second, there is an informational asymmetry, where the “teacher” (i.e. the agent Ai ) may conceivably know less about the environment state st than the “student” (i.e. the observer), and it may carry systematic biases; its policy, πi , may therefore be far from optimal. As a result, the observer may need to factor in the likely knowledge state of the agent and its cognitive limitations when making behavioural predictions. Finally, as a ToM needs to operate online while observing a new agent, we place a high premium on the speed of inference. Rather than using the computationally costly algorithms of classical inverse reinforcement learning (e.g. Ng et al., 2000; Ramachandran & Amir, 2007; Ziebart et al., 2008; Boularias et al., 2011), or Bayesian ToM (e.g. Baker et al., 2011; Nakahashi et al., 2016; Baker et al., 2017), we drive the ToMnet to amortise its inference through neural networks (as in Kingma & Welling, 2013; Rezende et al., 2014; Ho & Ermon, 2016; Duan et al., 2017; Wang et al., 2017). 2.2. The architecture To solve these tasks, we designed the ToMnet architecture shown in Fig 1. The ToMnet is composed of three modules: a character net, a mental state net, and a prediction net. The goal of the character net is to characterise the presented agent, by parsing observed past episode trajectories, (obs) Npast {τij }j=1 , into a character embedding, echar,i . Here we choose to parse each past episode independently   using a (obs) , and sum learned neural net, fθ , as echar,ij = fθ τij PNpast these to form the embedding echar,i = j=1 echar,ij . The goal of the mental state net is to mentalise about the presented agent during the current episode (i.e. infer its mental state; Dennett, 1973; Frith & Frith, 2006), by (obs) parsing the current episode trajectory, τik , up to time t − 1 into a mental state embedding, emental,i , using a learned neural net, gφ . This takes the form emental,i =  (obs) gφ [τij ]0:t−1 , echar,i . For brevity, we drop the agent subscript, i. Lastly, the goal of the prediction net is to leverage the char- Figure 1. ToMnet architecture. The character net parses an agent’s past trajectories from a set of POMDPs to form a character embedding, echar . The mental state net parses the agent’s trajectory on the current episode, to form an embedding of its mental state, emental . These embeddings are fed into the prediction net, which is then queried with a current state. This outputs predictions about future behaviour, such as next-step action probabilities (π̂), probabilities of whether certain objects will be consumed c Dayan, 1993). (ĉ), and predicted successor representations (SR; acter and mental state embeddings to predict subsequent behaviour of the agent. For example, next-step action prediction takes the form of estimating the given agent’s pol(obs) icy with π̂(·|xt , echar , emental ). We also predict other behavioural quantities, described below. We use a shared torso and separate heads for the different prediction targets. Precise details of the architecture, loss, and hyperparameters for each experiment are given in Appendix A. We train the whole ToMnet end-to-end. 2.3. Agents and environments We deploy the ToMnet to model agents belonging to a number of different “species” of agent. In Section 3.1, we consider species of agents with random policies. In Section 3.2, we consider species of agents with full observability over MDPs, which plan using value iteration. In Sections 3.3 – 3.5, we consider species of agents with different kinds of partial observability (i.e. different functions ωi (·)), with policies parameterised by feed-forward nets or LSTMs. We trained these agents using a version of the UNREAL deep RL framework (Jaderberg et al., 2017), modified to include an auxiliary belief task of estimating the locations of objects within the MDP. Crucially, we did not change the core architecture or algorithm of the ToMnet observer to match the structure of the species, only the ToMnet’s capacity. The POMDPs we consider here are all gridworlds with a Machine Theory of Mind current state π̂ partial past traj. 1 predicted action trained on α = 0. 01 0 π̂ 1 trained on α=3 (a) 0 0 ↓ → ↑ ← Figure 2. Example gridworld in which a random agent acts. (a) Example past episode. Coloured squares indicate objects. Red arrows indicate the positions and actions taken by the agent. (b) Example query: a state from a new MDP. Black dot indicates agent position. (c) Predictions for the next action taken by the agent shown in (a) in query state (b). Top: prediction from ToMnet trained on agents with near-deterministic policies. Bottom: prediction from ToMnet trained on agents with more stochastic policies. common action space (up/down/left/right/stay), deterministic dynamics, and a set of consumable objects, as described in the respective sections and in Appendix C. We experimented with these POMDPs due to their simplicity and ease of control; our constructions should generalise to richer domains too. We parameterically generate individual Mj by randomly sampling wall, object, and initial agent locations. 3. Experiments 3.1. Random agents To demonstrate its essential workings, we tested the ToMnet observer on a simple but illustrative toy problem. We created a number of different species of random agents, sampled agents from them, and generated behavioural traces on a distribution of random 11 × 11 gridworlds (e.g. Fig 2a). Each agent had a stochastic policy defined by a fixed vector of action probabilities πi (·) = πi . We defined different species based on how sparse its agents’ policies were: within a species S(α), each πi was drawn from a Dirichlet distribution with concentration parameter α. At one extreme, we created a species of agents with neardeterministic policies by drawing πi ∼ Dir(α = 0.01); here a single agent might overwhelmingly prefer to always move left, and another to always move up. At the other extreme, we created a species of agent with far more stochastic policies, by drawing πi ∼ Dir(α = 3). Next, we trained different ToMnet observers each on a single species of agent. For each α, we formed a training set by sampling 1000 agents from S(α), and for each agent, generating behavioural traces on randomlygenerated POMDPs. We then trained a ToMnet to observe (c) 2.5 N past 5 1 0 0.01 0 0.1 Trained α Trained α 0.01 0.03 0.1 D KL (π||π̂) · (b) 1 e 1 vs. e 2 (c) (b) prob. (a) 0.01 0.3 1 3 0.1 Test α 1 (d) 1 obs. 1 1.5 Trained α 0.01 3 0.01, 3 0 0.01 0.1 Test α 5 obs. 1 Figure 3. ToMnet trained on random agents. (a) Likelihood of agent’s true actions under the ToMnet’s predictions, given that the ToMnet has been trained on species S(α). Priors are shown in light blue, and posteriors after observing that agent perform just that same action in Npast = 1 or 5 past episodes in darker blue. Dots are data from the ToMnet; solid lines are from the analytic Bayes-optimal posteriors specialised to the respective S(α). (b) Character embeddings echar ∈ R2 of different agents. Dots are coloured by which action was observed to occur most during Npast = 10 past episodes, and are darker the higher that count. (c) Average KL-divergence between agents’ true and predicted policies when the ToMnet is trained on agents from one species, S(α), but tested on agents from a different species S(α0 ). Dots show values from the ToMnet; lines show analytic expected KLs when using analytic Bayes-optimal inference as in (a). Values calculated for Npast = 1. The ToMnet thus learns an effective prior for the species it is trained on. (d) Same, but including a ToMnet trained on a mixture of species (with Npast = 5). The ToMnet here implicitly learns to perform hierarchical inference. how randomly-sampled agents Ai ∼ S(α) behave on a variable number of past episodes (Npast ∼ U {0, 10}; for simplicity, limiting the length of each past trajectory to a single observation/action pair) and to use this information to predict the initial action that each agent Ai would take in a new POMDP, Mk (e.g. Fig 2b-c). We omitted the mental net for this task. When the ToMnet observer is trained on a species S(α), it learns how to approximate Bayes-optimal, online inference about agents’ policies πi (·) = πi ∼ Dir(α). Fig 3a shows how the ToMnet’s estimates of action probability increase with the number of past observations of that action, and how training the ToMnet on species with lower α makes it apply priors that the policies are indeed sparser. We can Machine Theory of Mind also see how the ToMnet specialises to a given species by testing it on agents from different species (Fig 3c): the ToMnet makes better predictions about novel agents drawn from the species which it was trained on. Moreover, the ToMnet easily learns how to predict behaviour from mixtures of species (Fig 3d): when trained jointly on species with highly deterministic (α = 0.01) and stochastic (α = 3) policies, it implicitly learns to expect this bimodality in the policy distribution, and specialises its inference accordingly. We note that it is not learning about two agents, but rather two species of agents, which each span a spectrum of individual parameters. There should be nothing surprising about seeing the ToMnet learn to approximate Bayes-optimal online inference; this should be expected given more general results about inference and meta-learning with neural networks (MacKay, 1995; Finn & Levine, 2017). Our point here is that a very first step in reasoning about other agents is an inference problem. The ToMnet is just an engine for learning to do inference and prediction on other agents. The ToMnet does expose an agent embedding space which we can explore. In Fig 3b, we show the values of echar produced by a ToMnet with a 2D embedding space. We note that the Bayes-optimal estimate of an agent’s policy is a Dirichlet posterior, which depends only on α (which is fixed for the species) and on the observed action count (a 5-dim vector). We see a similar solution reflected in the ToMnet’s echar embedding space, wherein agents are segregated along canonical directions by their empirical action counts. In summary, without any changes to its architecture, a ToMnet learns a general theory of mind that is specialised for the distribution of agents it encounters in the world, and estimates an agent-specific theory of mind online for each individual agent that captures the sufficient statistics of its behaviour. 3.2. Inferring goal-directed behaviour An elementary component of humans’ theory of other agents is an assumption that agents’ behaviour is goaldirected. There is a wealth of evidence showing that this is a core component of our model from early infancy (Gergely et al., 1995; Woodward, 1998; 1999; Buresh & Woodward, 2007), and intelligent animals such as apes and corvids have been shown to have similar expectations about their conspecifics (Call & Tomasello, 2008; Ostojić et al., 2013). Inferring the desires of others also takes a central role in machine learning in imitation learning, most notably in inverse RL (Ng et al., 2000; Abbeel & Ng, 2004). We demonstrate here how the ToMnet observer learns how to infer the goals of reward-seeking agents. We defined species of agents who acted within gridworlds with full observability (Fig 2a). Each gridworld was 11 × 11 in size, had randomly-sampled walls, and contained four different objects placed in random locations. Consuming an object yielded a reward for the agent and caused the episode to terminate. Each agent, Ai , had a unique, fixed reward function, such that it received reward ri,a ∈ (0, 1) when it consumed object a; the vectors ri were sampled from a Dirichlet distribution with concentration parameter α = 0.01. Agents also received a negative reward of −0.01 for every move taken, and a penalty of 0.05 for walking into walls. In turn, the agents planned their behaviour through value iteration, and hence had optimal policies πi∗ with respect to their own reward functions. We trained the ToMnet to observe behaviour of these agents in randomly-sampled “past” MDPs, and to use this to predict the agents’ behaviour in a “current” MDP. We detail three experiments below; these explore the range of capabilities of the ToMnet in this domain. First, we provided the ToMnet with a full trajectory of an agent on a single past MDP (Fig 4a). In turn, we queried the ToMnet with the initial state of a current MDP (Fig 4b) and asked for a set of predictions: the next action the agent would take (Fig 4c top), what object the agent would consume by the end of the episode (Fig 4c bottom), and a set of statistics about the agent’s trajectory in the current MDP, the successor representation (SR; the expected discounted state occupancy; Dayan, 1993, Fig 4). The ToMnet’s predictions qualitatively matched the agents’ true behaviours. Second, as a more challenging task, we trained a ToMnet to observe only partial trajectories of the agent’s past behaviour. We conditioned the ToMnet on single observationaction pairs from a small number of past MDPs (Npast ∼ U{0, 10}; e.g. Fig 5a). As expected, increasing the number of past observations of an agent improved the ToMnet’s ability to predict its behaviour on a new MDP (Fig 5b), but even in the absence of any past observations, the ToMnet had a strong prior for the reasonable behaviour that would be expected of any agent within the species, such as movement away from the corners, or consumption of the only accessible object (Fig 5c). We note that unlike the approach of inverse RL, the ToMnet is not constrained to explicitly infer the agents’ reward functions in service of its predictions. Nevertheless, in this simple task, using a 2-dimensional character embedding space renders this information immediately legible (Fig 5d). This is also true when the only behavioural prediction is next-step action. Finally, we added more diversity to the agent species by applying a very high move cost (0.5) to 20% of the agents; these agents therefore generally sought the closest object. Machine Theory of Mind (a) (b) current state (a) (b) partial past traj. 0.65 π̂(at | · ) past traj. 0.5 0 (d) action 1 successor (c) Npast = 0 Npast = 1 10 Npast = 5 π̂ (c) Npast 0 · → ↑ consumption ← prob. 1 ↓ Figure 4. ToMnet on goal-driven agents. (a) Past trajectory of an example agent. Coloured squares indicate the four objects. Red arrows indicate the position and action taken by the agent. (b) Example query: a state from a new MDP. Black dot indicates agent position. (c) ToMnet’s prediction for the agent’s next action (top) and object consumed at the end of the episode (bottom) for the query MDP in (b), given the past observation in (a). (d) ToMnet’s prediction of the successor representation (SR) for query (b), using discount γ = 0.9. Darker shading indicates higher expected discounted state occupancy. We trained a ToMnet to observe a small number of full trajectories (Npast ∼ U{0, 5}) of randomly-selected agents before making its behavioural prediction. The ToMnet learned to infer from even a single trajectory which subspecies of agent it was observing: if the agent went out of its way to consume a distant object on a past episode, then the ToMnet inferred a strong posterior that it would do so in a new episode from any starting position (Fig 6a); if the agent sought the closest object in a past episode, then the ToMnet was more cautious about whether it would seek the same object again on a new episode, deferring instead to a prediction that the agent would act greedily again (Fig 6b). This inference resembles the ability of children to jointly reason about agents’ costs and rewards when observing short traces of past behaviour (Jara-Ettinger et al., 2016; Liu et al., 2017). 3.3. Learning to model deep RL agents The previous experiments demonstrate the ToMnet’s ability to learn models of simple, algorithmic agents which have full observability. We next considered the ToMnet’s ability to learn models for a richer population of agents: those (d) e1 vs. e2 0 Figure 5. ToMnet on goal-driven agents, continued. (a) This ToMnet sees only snapshots of single observation/action pairs (red arrow) from a variable number of past episodes (one shown here). (b) Increasing Npast leads to better predictions; here we show the average posterior probability assigned to the true action. Even when Npast = 0, the action probability is greater than chance, since all agents in the species have similar policies in some regions of the state space. (c) Predicted policy for different initial agent locations in a query MDP, for different numbers of past observations.PArrows show resultant vectors for the predicted policies, i.e. k ak · π̂(ak |x, echar ). When Npast = 0, the ToMnet has no information about the agent’s preferred object, so the predicted policy exhibits no net object preference. When Npast > 0, the ToMnet infers a preference for the pink object. When the agent is stuck in the top right chamber, the ToMnet predicts that it will always consume the blue object, as this terminates the episode as soon as possible, avoiding a costly penalty. (d) 2D embedding space of the ToMnet, showing values of echar from 100 different agents. Agents are colour-coded by their groundtruth preferred objects; saturation increases with Npast , with the grey dots in the centre denoting agents with Npast = 0. with partial observability and neural network-based policies, trained using deep reinforcement learning. In this section we show how the ToMnet learns how to do inference over the kind of deep RL agent it is observing, and show the specialised predictions it makes as a consequence. This domain begins to capture the complexity of reasoning about real-world agents. So long as the deep RL agents Machine Theory of Mind past episode trajectory predicted consumption (b) past episode trajectory (a) predicted consumption 0 ∆ loss (a) -0.6 (b) Figure 6. ToMnet on greedy agents. Left: a single past trajectory. Right: ToMnet predictions on a query MDP. Light shaded regions indicate ToMnet’s prediction for the most probable object the agent will consume by the end of the episode, given that the agent is currently in that location. (a) After seeing the agent take a long path to the orange object, the ToMnet predicts it will try to consume the orange object on the query MDP, no matter its current location. (b) After seeing the agent take the shortest path to the green object, the ToMnet predicts it will generally consume a nearby object on the query MDP. share some overlap in their tasks, structure, and learning algorithms, we expect that they should exhibit at least some shared behavioural patterns. These patterns should also diverge systematically from each other as the aforementioned factors vary, and individual agents may also exhibit idiosyncratic behaviour as they settle in local minima while optimising their respective policies. There are thus opportunities to learn rich general and agent-specific theories of mind for such populations. Moreover, as the tasks and networks become more complex, hand-crafting a Machine Theory of Mind to parse behaviour based on our human knowledge (e.g. Baker et al., 2011; Nakahashi et al., 2016; Baker et al., 2017; Lake et al., 2017) becomes increasingly intractable; instead we seek here a path towards machines which learn how to model others’ minds autonomously (Botvinick et al., 2017). We trained three different species of agents on a modified version of the gridworlds, described below in Section 3.4. In brief, agents received maximum reward for reaching a subgoal location first, then consuming a preferred object that differed from agent to agent. Consuming any of the non-subgoal objects terminated the episode. All agents were based on the UNREAL architecture (Jaderberg et al., 2017), with details given in Appendix D. One species of agent (“blind”) was unable to observe the maze state at all, and could only observe its previous action (at−1 ) and reward (rt−1 ), which it could integrate over time through its LSTM state. The second species had partial observability (“sighted”), but was stateless: these agents could observe the gridworld within a 5 × 5 window centred at their current location, with the rest of the maze shrouded by a binary mask; the agents’ policies however were purely reactive, implemented via feedforward networks without any memory. The third species shared the benefits of the other two, being both sighted (with partial observability) and stateful current state actions consumptions successors with char net with mental net with both SR: blind SR: sighted, stateless SR: sighted, stateful Figure 7. Using the ToMnet to characterise trained neural-net agents. (a) Usefulness of ToMnet components for the three behavioural prediction targets, compared with a simple ToMnet with no character nor mental net. Longer bars are better; including both character and mental nets is best. More details are given in Table A1. (b) A ToMnet’s prediction of agents’ future state occupancy given a query POMDP state at time t = 0 (left), as per Fig 4d. Star denotes the subgoal. The maps on the right are produced after observing behaviour on Npast = 5 past POMDPs from a sampled agent of each subspecies (always preferring the pink object). The ToMnet does not know a priori which subspecies each agent belongs to, but infers it from past behaviour. (with an LSTM-based policy). The ToMnet, on the other hand, observed the behaviour of these agents with full observability of the POMDP state. We constructed a training set for the ToMnet by generating behaviour from 120 trained agents (3 species × 4 preferred objects × 10 neural nets trained with different random seeds). We held out a test set of a further 120 trained agents (i.e. trained using 10 additional random seeds) for evaluating the trained ToMnet. All results below are shown for the test set, i.e. predicting behaviour for novel agents which the ToMnet had never seen before. Unlike previous experiments, these agents’ behaviour depended on both their individual characteristics and their state; the ToMnet thus needed both a character net and a mental net to make the best predictions (Fig 7a). Qualitative evaluations of the ToMnet’s predictions show how it learned the expected behaviour of the three species of agents. Fig 7b shows the ToMnet’s predictions of future state occupancy for the same query state, but given different past observations of how the agent behaves. Without being given the species label, the ToMnet implicitly infers it, and maps out where the agent will go next: blind agents continue until they hit a wall, then turn; sighted but stateless agents consume objects opportunistically; sighted, stateful agents explore the interior and seek out the subgoal. Thus the ToMnet develops general models for the three different species of agents in its world. Machine Theory of Mind blind (a) Figure 8. Variational character embeddings produced from observations of held-out agents. This shows how the ToMnet learns a disentangled conceptual space of agent personalities. Left panel shows the first two of four non-degenerate dimensions of echar ∈ R8 ; right panels show the second two. Ellipses denote the Gaussian covariance (one stdev) of the posteriors q(echar |·). Left: posteriors coloured by agents’ ground-truth species. Right: posteriors coloured by agents’ ground-truth preferred objects. The ToMnet uses the first two dimensions of echar (left panel) to represent which of the three species the agent belongs to, and the next two dimensions (right panels) to represent its preferred object. When the agent is blind, the ToMnet represents the agent’s preferred object by the prior, a unit Gaussian. All posteriors collapsed to the prior in the remaining four dimensions. While we wished to visualise the agent embeddings as in previous experiments, constraining echar to a 2D space produced poor training performance. With the higher dimensionality required to train the ToMnet on this task (e.g. using R8 ), we found it difficult to discern any structure in the embedding space. This was likely due to the relatively deep prediction network, and the lack of explicit pressure to compress or disentangle the embeddings. However, the results were dramatically different when we added an explicit bottleneck to this layer, using the Deep Variational Information Bottleneck technique recently developed in Alemi et al. (2016). By replacing the character embedding vectors echar with simple Gaussian posteriors, q(echar |·), limiting their information content by regularising them towards a unit Gaussian prior, p(echar ), and annealing the respective penalty, Lq = β DKL (q||p) from β = 0 to β = 0.01 over training, the ToMnet was driven to disentangle the factors of variation in agent personality space (Fig 8). Moreover, the ToMnet even discovered substructure amongst the sighted/stateless subspecies that we were not aware of, as it clustered sighted/stateless test agents into two subcategories (Fig 9a-b). By contrasting the ToMnet’s predictions for these two clusters, the structure it discovers becomes obvious: each sighted/stateless agent explores its world using one of two classic memoryless wall-following algorithms, the right-hand rule or the left-hand rule (Fig 9c). 3.4. Acting based on false beliefs It has long been argued that a core part of human Theory of Mind is that we recognise that other agents do not base their decisions directly on the state of the world, but rather on an internal representation of the state of the world (Leslie, 1987; Gopnik & Astington, 1988; Wellman, 1992; e1 vs. e2 (b) (c) ∗1 π̂( · | · , echar ) ∗1 echar ∗2 π̂( · | · , echar ) ∗2 echar difference POMDP 1 stateless sighted, stateful POMDP 2 stateful sighted, stateless e3 vs. e4 e1 vs. e2 blind Figure 9. The ToMnet discovers two subspecies of sighted/ stateless agents. (a) Variational character posteriors, q(echar ), for sighted/stateless agents. Axes show the first two nondegenerate dimensions of echar (as in Fig 8a). Each colour shows the posteriors inferred from a single deep RL agent from the test set, using different behavioural traces. (b) Marginal posteriors for the individual agents shown in (a). These are shown as iso-density contours, enclosing 80% of the total density. Dots show the cluster means. (c) Predicted policy differences between agents in the two clusters. Each row shows a different query POMDP. Each panel shows predicted policy for different agent locations, as in Fig 5c. Left: ToMnet’s prediction for an agent with echar at the one cluster mean. Middle: at the other cluster mean. Arrows are darker where the two policies differ (higher DJS ). Right: vector difference between left and middle. Agents in the first cluster explore in an anti-clockwise direction, while agents in the second cluster explore in a clockwise direction. Baillargeon et al., 2016). This is usually framed as an understanding that other agents hold beliefs about the world: they may have knowledge that we do not; they may be ignorant of something that we know; and, most dramatically, they may believe the world to be one way, when we in fact know this to be mistaken. An understanding of this last possibility – that others can have false beliefs – has become the most celebrated indicator of a rich Theory of Mind, and there has been considerable research into how much children, infants, apes, and other species carry this capability (Baron-Cohen et al., 1985; Southgate et al., 2007; Clayton et al., 2007; Call & Tomasello, 2008; Krupenye et al., 2016; Baillargeon et al., 2016). Machine Theory of Mind recent traj. (a) p = 0. 9 p = 0. 1 p = 0. 9 p = 0. 1 current state (b) 0 · ↓ → ↑ ← · ↓ → ↑ ← · ↓ → ↑ ← · ↓ → ↑ ← empirical SR (d) 1 π̂ actions (c) Figure 10. Subgoal task, where agents can have false beliefs. (a) Trajectory of an agent (red arrows) as it seeks the subgoal (star). Agent has partial observability: dark grey areas have not been observed; light grey areas have been seen previously, but are not observable at the time of subgoal consumption. (b) When the agent consumes the subgoal object, there is a small probability that the other objects will instantaneously swap locations. Left: swap event within the agent’s current field of view. Right: outside it. (c) Effect of swap on agent’s immediate policy. (d) Effect of swap on agent’s empirical successor representation (average discounted state occupancy over 200 stochastic rollouts). Agent prefers the blue object. Here, we sought to explore whether the ToMnet would also learn that agents may hold false beliefs about the world. To do so, we first needed to generate a set of POMDPs in which agents could indeed hold incorrect information about the world (and act upon this). To create these conditions, we allowed the state of the environment to undergo random changes, sometimes where the agents couldn’t see them. In the subgoal maze described above in Section 3.3, we included a low probability (p = 0.1) state transition when the agent stepped on the subgoal, such that the four other objects would randomly permute their locations instantaneously (Fig 10a-b). These swap events were only visible to the agent insofar as the objects’ positions were within the agent’s current field of view; when the swaps occurred entirely outside its field of view, the agent’s internal state and policy at the next time step remained unaffected (policy changes shown in Fig 10c, right side), a signature of a false belief. As agents were trained to expect these lowprobability swap events, they learned to produce corrective behaviour as their policy was rolled out over time (Fig 10d, right side). While the trained agents were competent at the task, they were not optimal. In turn, we trained the ToMnet to predict the behaviour of these agents. We initially focused on agents with 5 × 5 fields of view, as in Section 3.3. We trained the ToMnet on rollouts from 40 sighted/stateful agents, each having a preference for one of the four different objects; we tested it on a set of 40 held-out agents. We used the ToMnet model described above in Section 3.3, with Npast = 4 past episodes for character inference. Our goal was to determine whether the ToMnet would learn a general theory of mind that included an element of false beliefs. However, the ToMnet, as described, does not have the capacity to explicitly report agents’ (latent) belief states, only the ability to report predictions about the agents’ overt behaviour. To proceed, we took inspiration from the literature on human infant and ape Theory of Mind (Call & Tomasello, 2008; Baillargeon et al., 2016). Here, experimenters have often utilised variants of the classic “Sally-Anne test” (Wimmer & Perner, 1983; Baron-Cohen et al., 1985) to probe subjects’ models of others. In the classic test, the observer watches an agent leave a desired object in one location, only for it to be moved, unseen by the agent. The subject, who sees all, is asked where the agent now believes the object lies. While infants and apes have limited ability to explicitly report such inferences about others’ mental states, experimenters have nevertheless been able to measure these subjects’ predictions of where the agents will actually go, e.g. by measuring anticipatory eye movements, or surprise when agents behave in violation of subjects’ expectations (Call & Tomasello, 2008; Krupenye et al., 2016; Baillargeon et al., 2016). These experiments have demonstrated that human infants and apes can implicitly model others as holding false beliefs. We used the swap events to construct a gridworld SallyAnne test. We hand-crafted scenarios where an agent would see its preferred blue object in one location, but would have to move away from it to reach a subgoal before returning to consume it (Fig 11a). During this time, the preferred object might be moved by a swap event, and the agent may or may not see this occur, depending on how far away the subgoal was. We forced the agents along this trajectory (off-policy), and measured how a swap event affected the agent’s probability of moving back to the preferred object. As expected, when the swap occurred within the agent’s field of view, the agent’s likelihood of turning back dropped dramatically; when the swap occurred outside its field of view, the policy was unchanged (Fig 11b, left). In turn, we presented these demonstration trajectories to the ToMnet (which had seen past behaviour indicating the agent’s preferred object). Crucially, the ToMnet was able to observe the entire POMDP state, and thus was aware of swaps when the agent was not. To perform this task Machine Theory of Mind swap dist = 2 swap dist = 3 swap dist = 4 1 2 3 4 5 6 7 DJS (s, sswap ) acting based on false beliefs (b) 0.4 properly, the ToMnet needs to have implicitly learned to separate out what it itself knows, and what the agent can plausibly know, without relying on a hand-engineered, explicit observation model for the agent. Indeed, the ToMnet predicted the correct behavioural patterns (Fig 11b, right): specifically, the ToMnet predicts that when the world changes far away from an agent, that agent will persist with a policy that is founded on false beliefs about the world. This test was a hand-crafted scenario. We validated its results by looking at the ToMnet’s predictions for how the agents responded to all swap events in the distribution of POMDPs. We sampled a set of test mazes, and rolled out the agents’ policies until they consumed the subgoal, selecting only episodes where the agents had seen their preferred object along the way. At this point, we created a set of counterfactuals: either a swap event occurred, or it didn’t. We measured the ground truth for how the swaps would affect the agent’s policy, via the average Jensen-Shannon divergence (DJS ) between the agent’s true action probabilities in the no-swap and swap conditions1 . As before, the agent’s policy often changed when a swap was in view (for 1 For a discussion of why we used the DJS measure, see Appendix F.2. γ = 0. 9 acting based on false beliefs γ = 0. 5 0.0 swap distance (c) 0.3 acting based on changed beliefs 3×3 actions: Figure 11. Sally-Anne test. (a) We force agents to initially move along a hand-constructed trajectory. Agents have 5 × 5 observability and prefer the blue square object, but must seek the subgoal (star) first. When an agent reaches the subgoal, a swap event may or may not occur. If there is no swap, the optimal action is to go left. By extending the length of the path, the swap event will no longer be visible to the agent. (b) Left: effect of a swap event on the agents’ true policies, measured as the relative reduction in their probability of moving back towards the original location where they saw the blue object (∆πL = (π(aL |no swap) − π(aL |swap))/π(aL |no swap) × 100%). If the agent can see that the object has moved from this location (swap dist ≤ 2), it will not return left. If it cannot see this location, its policy will not change. Right: ToMnet’s prediction. acting based on changed beliefs γ = 0. 99 DJS (π, πswap ) swap distance acting based on changed beliefs acting based on false beliefs 9×9 0 (d) 0.4 acting based on changed beliefs DJS (s, sswap ) 1 2 3 4 5 6 7 acting based on false beliefs successors: 0% acting based on changed beliefs 0.0 successors: tomnet tomnet 5×5 actions: recent traj. 100 % true agent ∆πL (b) true agents (a) 0.25 DJS (π, πswap ) (a) swap dist = 1 0 1 2 3 4 swap distance 5 1 2 3 4 swap distance 5 acting based on false beliefs Figure 12. Natural Sally-Anne test, using swap events within the distribution of POMDPs. (a) Left: effect of swap events on 5 × 5 agents’ next-step policies. Right: ToMnet predictions. (b) For SRs of different discount factors (γ). DJS measured between normalised SRs. (c)-(d) As for (a)-(b), but for a ToMnet trained on a range of agents with different fields of view. Showing only 3 × 3 and 9 × 9 results for clarity. For a discussion of why 3 × 3 agents’ next-step actions are particularly sensitive to adjacent swap events, see Appendix F.1. these agents, within a 2 block radius), but wouldn’t change when the swap was not observable (Fig 12a, left). The ToMnet learned that the agents’ policies were indeed more sensitive to local changes in the POMDP state, but were relatively invariant to changes that occurred out of sight (Fig 12a, right). The ToMnet did not, however, learn a hard observability boundary, and was more liberal in predicting that far-off changes could affect agent policy. The ToMnet also correctly predicted that the swaps would induce corrective behaviour over longer time periods, even when they were not initially visible (Fig 12b). These patterns were even more pronounced when we trained the ToMnet on mixed populations of agents with different fields of view. In this task, the ToMnet had to infer what each agent could see (from past behaviour alone) in order to predict each agent’s behaviour in the future. The ToMnet’s predictions reveal an implicit grasp of how dif- Machine Theory of Mind (a) Most surprising of all, we found that the ToMnet learned these statistics even if the ToMnet had never seen swap events during training: the curves in Fig 12 were qualitatively identical for the ToMnet under such conditions (Fig A1). beliefs: true agent no swap ferent agents’ sensory abilities render them differentially vulnerable to acquire false beliefs (Fig 12c-d). We have demonstrated that the ToMnet learns that agents can act based on false beliefs. This is limited, though, in that the ToMnet cannot explicitly report what these agents know and don’t know about the world. Moreover, it is difficult to extract any beliefs that will not manifest immediately in overt behaviour. swap predicted beliefs: tomnet true agents (c) 0.55 tomnet holding changed beliefs 9×9 DJS (b, bswap ) 3.5. Explicitly inferring belief states no swap (b) beliefs: On the one hand, we were impressed that the ToMnet learns a general theory of mind that incorporates an implicit understanding that agents act based on their own persistent representations of the world, even if they are mistaken. On the other hand, we should not attribute this cognitive ability to a special feature of the ToMnet architecture itself, which is indeed very straightforward. Rather, this work demonstrates that representational Theory of Mind can arise simply by observing competent agents acting in POMDPs. swap 0 1 2 3×3 3 4 swap distance 5 1 2 3 4 swap distance 5 holding false beliefs We therefore extended the ToMnet to be able to make declarative statements about agents’ beliefs. We achieved this by constructing a supervised dataset of belief states in the sample gridworld. We trained the UNREAL agents to report their beliefs about the locations of the four objects and the subgoal at every time step, alongside their policy. To do this, we added a head to the LSTM that output a posterior over each object’s current location on the grid (or whether it was absent). During training, the agents learned to report their best estimate of each object’s current location, based on its observations so far during the episode. Example belief states for the query MDP states in Fig 10ab are shown in Fig 13a. Note that these reported beliefs are not causal to the agents’ policy; they are just a readout from the same LSTM hidden state that serves the policy. Figure 13. Supervised prediction of beliefs. (a) Belief states (b = p(obj location|x0:tswap )) reported by the agent in the POMDPs shown in Fig 10 at the time of subgoal consumptions. Here we show b for only two objects (rows). In the left two columns, the swap event occurred within the agent’s field of view, so the agent’s beliefs about the object locations changed given a swap. In the right two columns, the swap event was not within the agent’s field of view, so its beliefs about the object locations did not change. (b) Predictions b̂ made by the ToMnet, given only the trajectory of states and actions. The ToMnet predicts that the observable swap event (left) leads to a change in belief state, whereas the unobservable swap event (right) does not. (c) Statistics, as in Fig 12, but for beliefs. The ToMnet captures the longer range over which the 9 × 9 agents can update their beliefs (though again, it infers a soft observability boundary). In turn, we trained a single ToMnet on agents with a range of different fields of view. This ToMnet had to observe only agents’ overt behaviour, and use it to predict, for a given query state, an agent’s policy, consumptions, successor representations, and reported beliefs (i.e., what the agent would say about the world if it were asked; e.g. Fig 13b). ing to them too (as in Fig 13c). As shown in Fig 13c, the ToMnet learns agent-specific theories of mind for the different subspecies that grasp the essential differences between their belief-forming tendencies: agents with less visibility of changes in their world are more likely to report false beliefs; and behave accord- Last of all, we included an additional variational information bottleneck penalty, to encourage low-dimensional abstract embeddings of agent types. As with the agent characterisation in Fig 7, the character embeddings of these agents separated along the factors of variation (field of view and preferred object; Fig 14). Moreover, these embeddings show the ToMnet’s ability to distinguish different agents’ visibility: blind and 3 × 3 agents are easily distinguishable, whereas there is little in past behaviour to separate 7 × 7 agents from 9 × 9 agents (or little benefit in making this distinction). Machine Theory of Mind (b) e1 vs. e2 blind 9×9 3×3 7×7 5×5 blind sighted e3 vs. e4 vs. e5 (a) Figure 14. Variational character embeddings of agents with different fields of view. (a) First two dimensions of echar represent field of view. Contours are shown of the marginal posteriors for each agent species. (b) Next three dimensions represent preferred objects. Volumes show the approximate marginal posteriors for agents preferring each of the four objects (colours). Blind agents (left) cannot express their preference through their overt behaviour; the ToMnet therefore reverts to the prior. Sighted agents (right) produce embeddings arranged in a roughly tetrahedral arrangement. This same arrangement arises independently of the sighted agents’ field of view. We note that this particular construction of explicit belief inference will likely not scale in its current form. Our method depends on two assumptions that break down in the real world. First, it requires access to others’ latent belief states for supervision. We assume here that the ToMnet gets access to these via a rich communication channel; as humans, this channel is likely much sparser. It is an empirical question as to whether the real-world information stream is sufficient to train such an inference network. We do, however, have privileged access to some of our own mental states through meta-cognition; though this data may be biased and noisy, it might be sufficiently rich to learn this task. Second, it is intractable to predict others’ belief states about every aspect of the world. As humans, we nevertheless have the capacity to make such predictions about arbitrary variables as the need arises. This may require creative solutions in future work, such as forming abstract embeddings of others’ belief states that can be queried. such a system. There is much work to do to scale the ToMnet to richer domains. First, we have worked entirely within gridworlds, due to the control such environments afford. We look forward to extending these systems to operate within complex 3D visual environments, and within other POMDPs with rich state spaces. Second, we did not experiment here with limiting the observability of the observer itself. This is clearly an important challenge within real-world social interaction, e.g. when we try to determine what someone else knows that we do not. This is, at its heart, an inference problem (Baker et al., 2017); learning to do this robustly is a future challenge for the ToMnet. Third, there are many other dimensions over which we may wish to characterise agents, such as whether they are animate or inanimate (Scholl & Tremoulet, 2000), prosocial or adversarial (Ullman et al., 2009), reactive or able to plan (Sutton & Barto, 1998). Potentially more interesting is the possibility of using the ToMnet to discover new structure in the behaviour of either natural or artificial populations, i.e. as a kind of machine anthropology. Fourth, a Theory of Mind is important for social beings as it informs our social decision-making. An important step forward for this research is to situate the ToMnet inside artificial agents, who must learn to perform multi-agent tasks. In pursuing all these directions we anticipate many future needs: to enrich the set of predictions a ToMnet must make; to introduce gentle inductive biases to the ToMnet’s generative models of agents’ behaviour; and to consider how agents might draw from their own experience and cognition in order to inform their models of others. Addressing these will be necessary for advancing a Machine Theory of Mind that learns the rich capabilities of responsible social beings. 4. Discussion Acknowledgements In this paper, we used meta-learning to build a system that learns how to model other agents. We have shown, through a sequence of experiments, how this ToMnet learns a general model for agents in the training distribution, as well as how to construct an agent-specific model online while observing a new agent’s behaviour. The ToMnet can flexibly learn such models over a range of different species of agents, whilst making few assumptions about the generative processes driving these agents’ decision making. The ToMnet can also discover abstractions within the space of behaviours. We’d like to thank the many people who provided feedback on the research and the manuscript, including Marc Lanctot, Jessica Hamrick, Ari Morcos, Agnieszka GrabskaBarwinska, Avraham Ruderman, Christopher Summerfield, Pedro Ortega, Josh Merel, Doug Fritz, Nando de Freitas, Heather Roff, Kevin McKee, and Tina Zhu. We note that the experiments we pursued here were simple, and designed to illustrate the core ideas and capabilities of Albrecht, Stefano V and Stone, Peter. Autonomous agents modelling other agents: A comprehensive survey and References Abbeel, Pieter and Ng, Andrew Y. Apprenticeship learning via inverse reinforcement learning. In ICML, 2004. Machine Theory of Mind open problems. arXiv preprint arXiv:1709.08071, 2017. Alemi, Alexander A, Fischer, Ian, Dillon, Joshua V, and Murphy, Kevin. Deep variational information bottleneck. arXiv:1612.00410, 2016. Baillargeon, Renée, Scott, Rose M, and Bian, Lin. Psychological reasoning in infancy. Annual Review of Psychology, 67:159–186, 2016. Baker, Chris, Saxe, Rebecca, and Tenenbaum, Joshua. Bayesian theory of mind: Modeling joint belief-desire attribution. In Proceedings of the Cognitive Science Society, volume 33, 2011. Baker, Chris L, Jara-Ettinger, Julian, Saxe, Rebecca, and Tenenbaum, Joshua B. Rational quantitative attribution of beliefs, desires and percepts in human mentalizing. Nature Human Behaviour, 1(4):0064, 2017. Baron-Cohen, Simon, Leslie, Alan M, and Frith, Uta. Does the autistic child have a theory of mind? Cognition, 21 (1):37–46, 1985. Barrett, Samuel, Rosenfeld, Avi, Kraus, Sarit, and Stone, Peter. Making friends on the fly: Cooperating with new teammates. Artificial Intelligence, 242:132–171, 2017. Botvinick, M., Barrett, D. G. T., Battaglia, P., de Freitas, N., Kumaran, D., Leibo, J. Z, Lillicrap, T., Modayil, J., Mohamed, S., Rabinowitz, N. C., Rezende, D. J., Santoro, A., Schaul, T., Summerfield, C., Wayne, G., Weber, T., Wierstra, D., Legg, S., and Hassabis, D. Building Machines that Learn and Think for Themselves: Commentary on Lake et al., Behavioral and Brain Sciences, 2017. arXiv:1711.08378, November 2017. Camerer, Colin F, Ho, Teck-Hua, and Chong, Juin-Kuan. A cognitive hierarchy model of games. The Quarterly Journal of Economics, 119(3):861–898, 2004. Carey, Susan. The origin of concepts. Oxford University Press, 2009. Churchland, Paul. The neural representation of the social world. In May, Larry, Friedman, Marilyn, and Clark, Andy (eds.), Mind and morals: essays on cognitive science and ethics. MIT Press, 1996. Clayton, Nicola S, Dally, Joanna M, and Emery, Nathan J. Social cognition by food-caching corvids. the western scrub-jay as a natural psychologist. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 362(1480):507–522, 2007. Cohen, Philip R, Perrault, C Raymond, and Allen, James F. Beyond question answering. Strategies for natural language processing, 245274, 1981. Dayan, Peter. Improving generalization for temporal difference learning: The successor representation. Neural Computation, 5(4):613–624, 1993. Dennett, Daniel C. Two contrasts: folk craft versus folk science, and belief versus opinion. The future of folk psychology: Intentionality and cognitive science, pp. 135– 148, 1991. Dennett, Daniel Clement. The intentional stance. MIT press, 1973. Dragan, Anca D, Lee, Kenton CT, and Srinivasa, Siddhartha S. Legibility and predictability of robot motion. In 8th ACM/IEEE International Conference on HumanRobot Interaction, pp. 301–308. IEEE, 2013. Boularias, Abdeslam, Kober, Jens, and Peters, Jan. Relative entropy inverse reinforcement learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp. 182–189, 2011. Duan, Yan, Andrychowicz, Marcin, Stadie, Bradly, Ho, Jonathan, Schneider, Jonas, Sutskever, Ilya, Abbeel, Pieter, and Zaremba, Wojciech. One-shot imitation learning. arXiv:1703.07326, 2017. Brown, George W. Iterative solution of games by fictitious play. Activity analysis of production and allocation, 13 (1):374–376, 1951. Evans, Owain, Stuhlmüller, Andreas, and Goodman, Noah D. Learning the preferences of ignorant, inconsistent agents. In AAAI, pp. 323–329, 2016. Buresh, Jennifer Sootsman and Woodward, Amanda L. Infants track action goals within and across agents. Cognition, 104(2):287–314, 2007. Finn, Chelsea and Levine, Sergey. Meta-learning and universality: Deep representations and gradient descent can approximate any learning algorithm. arXiv:1710.11622, 2017. Call, Josep and Tomasello, Michael. Does the chimpanzee have a theory of mind? 30 years later. Trends in cognitive sciences, 12(5):187–192, 2008. Camerer, Colin. Behavioral game theory. New Age International, 2010. Fisac, Jaime F, Gates, Monica A, Hamrick, Jessica B, Liu, Chang, Hadfield-Menell, Dylan, Palaniappan, Malayandi, Malik, Dhruv, Sastry, S Shankar, Griffiths, Thomas L, and Dragan, Anca D. Pragmatic-pedagogic value alignment. arXiv:1707.06354, 2017. Machine Theory of Mind Frith, Chris D and Frith, Uta. The neural basis of mentalizing. Neuron, 50(4):531–534, 2006. Gallese, Vittorio and Goldman, Alvin. Mirror neurons and the simulation theory of mind-reading. Trends in cognitive sciences, 2(12):493–501, 1998. Gergely, György, Nádasdy, Zoltán, Csibra, Gergely, and Bı́ró, Szilvia. Taking the intentional stance at 12 months of age. Cognition, 56(2):165–193, 1995. Gopnik, Alison and Astington, Janet W. Children’s understanding of representational change and its relation to the understanding of false belief and the appearance-reality distinction. Child development, pp. 26–37, 1988. Gopnik, Alison and Wellman, Henry M. Why the child’s theory of mind really is a theory. Mind & Language, 7 (1-2):145–171, 1992. Gordon, Robert M. Folk psychology as simulation. Mind & Language, 1(2):158–171, 1986. Hadfield-Menell, Dylan, Russell, Stuart J, Abbeel, Pieter, and Dragan, Anca. Cooperative inverse reinforcement learning. In NIPS, pp. 3909–3917, 2016. Hassabis, Demis, Spreng, R Nathan, Rusu, Andrei A, Robbins, Clifford A, Mar, Raymond A, and Schacter, Daniel L. Imagine all the people: how the brain creates and uses personality models to predict behavior. Cerebral Cortex, 24(8):1979–1987, 2013. Ho, Jonathan and Ermon, Stefano. Generative adversarial imitation learning. In NIPS, pp. 4565–4573, 2016. Hochreiter, Sepp, Younger, A Steven, and Conwell, Peter R. Learning to learn using gradient descent. In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001. Hula, Andreas, Montague, P Read, and Dayan, Peter. Monte carlo planning method estimates planning horizons during interactive social exchange. PLoS computational biology, 11(6):e1004254, 2015. Jaderberg, Max, Mnih, Volodymyr, Czarnecki, Wojciech Marian, Schaul, Tom, Leibo, Joel Z, Silver, David, and Kavukcuoglu, Koray. Reinforcement learning with unsupervised auxiliary tasks. In ICLR, 2017. Jara-Ettinger, Julian, Gweon, Hyowon, Schulz, Laura E, and Tenenbaum, Joshua B. The naı̈ve utility calculus: computational principles underlying commonsense psychology. Trends in cognitive sciences, 20(8):589–604, 2016. Kingma, Diederik P and Welling, Max. Auto-encoding variational bayes. arXiv:1312.6114, 2013. Kleiman-Weiner, Max, Ho, Mark K, Austerweil, Joseph L, Littman, Michael L, and Tenenbaum, Joshua B. Coordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In COGSCI, 2016. Kris Cao, Angeliki Lazaridou, Marc Lanctot Joel Z Leibo Karl Tuyls Stephen Clark. Emergent communication through negotiation. ICLR submission. Krupenye, Christopher, Kano, Fumihiro, Hirata, Satoshi, Call, Josep, and Tomasello, Michael. Great apes anticipate that other individuals will act according to false beliefs. Science, 354(6308):110–114, 2016. Lake, Brenden M, Ullman, Tomer D, Tenenbaum, Joshua B, and Gershman, Samuel J. Building machines that learn and think like people. Behavioral and Brain Sciences, 40, 2017. Lanctot, Marc, Zambaldi, Vinicius, Gruslys, Audrunas, Lazaridou, Angeliki, Perolat, Julien, Silver, David, Graepel, Thore, et al. A unified game-theoretic approach to multiagent reinforcement learning. In NIPS, pp. 4193–4206, 2017. Leslie, Alan M. Pretense and representation: The origins of” theory of mind.”. Psychological review, 94(4):412, 1987. Liu, Shari, Ullman, Tomer D, Tenenbaum, Joshua B, and Spelke, Elizabeth S. Ten-month-old infants infer the value of goals from the costs of actions. Science, 358 (6366):1038–1041, 2017. Lucas, Christopher G, Griffiths, Thomas L, Xu, Fei, Fawcett, Christine, Gopnik, Alison, Kushnir, Tamar, Markson, Lori, and Hu, Jane. The child as econometrician: A rational model of preference understanding in children. PloS one, 9(3):e92160, 2014. MacKay, David JC. Probable networks and plausible predictionsa review of practical bayesian methods for supervised neural networks. Network: Computation in Neural Systems, 6(3):469–505, 1995. Milli, Smitha, Hadfield-Menell, Dylan, Dragan, Anca, and Russell, Stuart. Should robots be obedient? arXiv:1705.09990, 2017. Nakahashi, Ryo, Baker, Chris L., and Tenenbaum, Joshua B. Modeling human understanding of complex intentional action with a bayesian nonparametric subgoal model. In AAAI, pp. 3754–3760, 2016. Ng, Andrew Y, Russell, Stuart J, et al. Algorithms for inverse reinforcement learning. In ICML, pp. 663–670, 2000. Machine Theory of Mind Nowak, Martin A. Five rules for the evolution of cooperation. Science, 314(5805):1560–1563, 2006. Wellman, Henry M. The child’s theory of mind. The MIT Press, 1992. Oliehoek, Frans A and Amato, Christopher. A concise introduction to decentralized POMDPs. Springer, 2016. Wimmer, Heinz and Perner, Josef. Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception. Cognition, 13(1):103–128, 1983. Ostojić, Ljerka, Shaw, Rachael C, Cheke, Lucy G, and Clayton, Nicola S. Evidence suggesting that desire-state attribution may govern food sharing in eurasian jays. PNAS, 110(10):4123–4128, 2013. Woodward, Amanda L. Infants selectively encode the goal object of an actor’s reach. Cognition, 69(1):1–34, 1998. Premack, David and Woodruff, Guy. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4):515–526, 1978. Woodward, Amanda L. Infants ability to distinguish between purposeful and non-purposeful behaviors. Infant Behavior and Development, 22(2):145–160, 1999. Ramachandran, Deepak and Amir, Eyal. Bayesian inverse reinforcement learning. Urbana, 51(61801):1–4, 2007. Yoshida, Wako, Dolan, Ray J, and Friston, Karl J. Game theory of mind. PLoS computational biology, 4(12): e1000254, 2008. Rezende, Danilo Jimenez, Mohamed, Shakir, and Wierstra, Daan. Stochastic backpropagation and approximate inference in deep generative models. arXiv:1401.4082, 2014. Ziebart, Brian D, Maas, Andrew L, Bagnell, J Andrew, and Dey, Anind K. Maximum entropy inverse reinforcement learning. In AAAI, volume 8, pp. 1433–1438, 2008. Schmidhuber, Juergen, Zhao, Jieyu, and Wiering, MA. Simple principles of metalearning. Technical report IDSIA, 69:1–23, 1996. Scholl, Brian J and Tremoulet, Patrice D. Perceptual causality and animacy. Trends in cognitive sciences, 4 (8):299–309, 2000. Southgate, Victoria, Senju, Atsushi, and Csibra, Gergely. Action anticipation through attribution of false belief by 2-year-olds. Psychological Science, 18(7):587–592, 2007. Sutton, Richard S and Barto, Andrew G. Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998. Thrun, Sebastian and Pratt, Lorien. Learning to learn: Introduction and overview. In Learning to learn, pp. 3–17. Springer, 1998. Ullman, Tomer, Baker, Chris, Macindoe, Owen, Evans, Owain, Goodman, Noah, and Tenenbaum, Joshua B. Help or hinder: Bayesian models of social goal inference. In NIPS, pp. 1874–1882, 2009. Vilalta, Ricardo and Drissi, Youssef. A perspective view and survey of meta-learning. Artificial Intelligence Review, 18(2):77–95, 2002. Wang, Ziyu, Merel, Josh S, Reed, Scott E, de Freitas, Nando, Wayne, Gregory, and Heess, Nicolas. Robust imitation of diverse behaviors. In NIPS, pp. 5326–5335, 2017. Machine Theory of Mind Appendices A. Model description: architectures Here we describe the precise details of the architectures used in the main text. We note that we did not optimise our results by tweaking architectures or hyperparameters in any systematic or substantial way. Rather, we simply picked sensible-looking values. We anticipate that better performance could be obtained by improving these decisions, but this is beyond the scope of this work. A.1. Common elements. Pre-processing. Both the character net and the mental state net consume trajectories, which are sequences of observed (obs) (obs) (obs) state/action pairs, τij = {(xt , at )}Tt=0 , where i is the agent index, and j is the episode index. The ob(obs) served states in our experiments, xt , are always tensors of shape (11 × 11 × K), where K is the number of feature planes (comprising one feature plane for the walls, one for each object, and one for the agent). The observed (obs) actions, at , are always vectors of length 5. We combine these data through a spatialisation-concatenation operation, whereby the actions are tiled over space into a (11 × 11 × 5) tensor, and concatenated with the states to form a single tensor of shape (11 × 11 × (K + 5)). Training. All ToMnets were trained with the Adam optimiser, with learning rate 10−4 , using batches of size 16. We trained the ToMnet for 40k minibatches for random agents (Section 3.1), and for 2M minibatches otherwise. A.2. ToMnet for random agents (Section 3.1) Data. For each species, S(α), we trained a single ToMnet. For each agent, the ToMnet was provided with a variable number of past episodes (Npast ∼ U {0, 10}), each of length 1 (i.e. each trajectory consisting of a single stateaction pair). When no past episodes were sampled for a given agent, the character embedding was set to echar = 0. Character net. Each trajectory τij comprises a single state/action pair. We spatialise the action, and concatenate this with the state. This is passed into a 1-layer convnet, with 8 feature planes and ReLU nonlinearity. We then passed the sequence of these (indexed by j) into a convolutional LSTM, with the output passed through an average pooling layer, and a fully-connected layer to a 2D embedding space, to produce echar,i . We obtained similar results with a wide range of different architectures. Mental net. None. Prediction net. In this experiment, we predict only nextstep action (i.e. policy, π̂) We spatialise echar,i , and concatenate with the query state. This is passed to a 2-layer convnet, with 32 feature planes and ReLUs. This is followed by average pooling, then a fully-connected layer to logits in R5 , followed by a softmax. A.3. ToMnet for inferring goals (Section 3.2) A.3.1. E XPERIMENT 1: SINGLE PAST MDP Data. Character embedding formed from a single past episode, comprising a full trajectory on a single MDP. Query state is the initial state of a new MDP, so no mental state embedding required. Character net. For the single trajectory τi in the past episode, the ToMnet forms the character embedding echar,i as follows. We pre-process the data from each time-step by (obs) spatialising the actions, at , concatenating these with the (obs) respective states, xt , passing through a 5-layer resnet, with 32 channels, ReLU nonlinearities, and batch-norm, followed by average pooling. We pass the results through an LSTM with 64 channels, with a linear output to either a 2-dim or 8-dim echar,i (no substantial difference in results). Mental net. None. Prediction net. In this and subsequent experiments, we make three predictions: next-step action, which objects are consumed by the end of the episode, and successor representations. We use a shared torso for these predictions, from which separate heads branch off. For the prediction torso, we spatialise echar,i , and concatenate with the query state; this is passed into a 5-layer resnet, with 32 channels, ReLU nonlinearities, and batch-norm. Action prediction head. From the torso output: a 1-layer convnet with 32 channels and ReLUs, followed by average pooling, and a fully-connected layer to 5-dim logits, followed by a softmax. This gives the predicted policy, π̂. Consumption prediction head. From the torso output: a 1-layer convnet with 32 channels and ReLUs, followed by average pooling, and a fully-connected layer to 4-dims, followed by a sigmoid. This gives the respective Bernoulli probabilities that each of the four objects will be consumed by the end of the episode, ĉ. Successor representation prediction head. From the torso output: a 1-layer convnet with 32 channels and ReLUs, then a 1-layer convnet with 3 channels, followed by a softmax over each channel independently. This gives the predicted normalised SRs for the three discount factors, γ = 0.5, 0.9, 0.99. Machine Theory of Mind A.3.2. E XPERIMENT 2: MANY PAST MDP S , ONLY A SINGLE SNAPSHOT EACH Data. Character embedding is formed from many past episodes (Npast ∼ U {0, 10}); however, we only use a snapshot of a single time point (i.e. a single state/action pair) from each past episode. Query state as for Experiment 1. Character net. For each trajectory τij , the character net adds a contribution echar,ij to the character embedding echar,i as follows. We pre-process the single state/action pair in τij as in Experiment 1, i.e. spatialising, concatenating, resnet, average pooling. There is no LSTM as there is only a single state/action pair; instead we use a fullyconnected layer to form echar,ij ∈ R2 . These are summed across the Npast past trajectories to form echar,i . Mental net. None. Prediction net. As for Experiment 1. A.3.3. E XPERIMENT 3: GREEDY AGENTS Same as Experiment 1, but with Npast ∈ U {0, 5}. A.4. ToMnet for modelling deep RL agents (Section 3.3) Data. Character embedding is formed from observing full trajectories of agents on Npast = 4 POMDPs. While the agents have partial observability, potentially with significant parts of the state masked, the observer sees the entire gridworld (albeit without any indication of the field of view of the agent). The current episode is split at a random time (drawn uniformly from U {0, T − 1} where T is the length of the trajectory). The trajectory prior to the split forms the “recent trajectory”, and is passed to the mental net. The state at the time of the split is used as the query state. Character net. For each past trajectory τij , the character net adds a contribution echar,ij to the character embedding echar,i via the same architecture as in Experiment 1 described in Appendix A.3 above, with an 8-dim echar,ij . These are summed to form echar,i . Mental net. We pre-process each time step’s state/action pair in the recent trajectory as follows: we spatialise the action, concatenate with the state, pass through a 5-layer resnet, with 32 channels, ReLU nonlinearities, and batchnorm. The results are fed into a convolutional LSTM with 32 channels. The LSTM output is also a 1-layer convnet with 32 channels, yielding a mental state embedding emental,i ∈ R11×11×32 . When the recent trajectory is empty (i.e. the query state is the initial state of the POMDP), emental,i is the zero vector. Prediction net. As in Experiment 1 described in Appendix A.3. However, the prediction torso begins by spatialising echar,i and concatenating it with both emental,i and the query state. Also, as these agents act in gridworlds that include the subgoal object, the consumption prediction head outputs a 5-dim vector. DVIB. For the Deep Variational Information Bottleneck experiments, we altered the architecture by making the character net output a posterior density, q(echar,i ), rather than a single latent echar,i ; likewise, for the mental net to produce q(emental,i ), rather than emental,i . We parameterised both densities as Gaussians, with the respective nets outputting the mean and log diagonal of the covariance matrices, as in Kingma & Welling (2013). For the character net, we achieved this by doubling the dimensionality of the final fully-connected layer; for the mental net, we doubled the number of channels in the final convolutional layer. In both cases, we used fixed, isotropic Gaussian priors. For evaluating predictive performance after the bottleneck, we sampled both echar and emental , propagating gradients back using the reparameterisation trick. For evaluating the bottleneck cost, we used the analytic KL for q(echar,i ), and the analytic KL for q(emental,i ) conditioned on the sampled value of echar,i . We scaled the bottleneck costs by βchar = βmental = β, annealing β quadratically from 0 to 0.01 over 500k steps. A.5. ToMnet for false beliefs (Sections 3.4–3.5) The ToMnet architecture was the same as described above in Appendix A.4. The experiments in Section 3.5 also included an additional belief prediction head to the prediction net. Belief prediction head. For each object, this head outputs a 122-dim discrete distribution (the predicted belief that the object is in each of the 11 × 11 locations on the map, or whether the agent believes the object is absent altogether). From the torso output: a 1-layer convnet with 32 channels and ReLU, branching to (a) another 1-layer convnet with 5 channels for the logits for the predicted beliefs that each object is at the 11×11 locations on the map, as well as to (b) a fully-connected layer to 5-dims for the predicted beliefs that each object is absent. We unspatialise and concatenate the outputs of (a) and (b) in each of the 5 channels, and apply a softmax to each channel. Machine Theory of Mind B. Loss function C. Gridworld details Here we describe the components of the loss function used for training the ToMnet. The POMDPs Mj were all 11 × 11 gridworld mazes. Mazes in Sections 3.1–3.2 were sampled with between 0 and 4 random walls; mazes in Sections 3.3–3.5 were sampled with between 0 and 6 random walls. Walls were defined between two randomly-sampled endpoints, and could be diagonal. For each agent, Ai , we sample past and current trajectories, and form predictions for the query POMDP at time t. Each prediction provides a contribution to the loss, described below. We average the respective losses across each of the agents in the minibatch, and give equal weighting to each loss component. Action prediction. The negative log-likelihood of the true action taken by the agent under the predicted policy: (obs) Laction,i = − log π̂(at (obs) |xt , echar,i , emental,i ) Consumption prediction. For each object, k, the negative log-likelihood that the object is/isn’t consumed: X (obs) Lconsumption,i = − log pck (ck |xt , echar,i , emental,i ) k Successor representation prediction. For each discount factor, γ, we define the agent’s empirical successor representation as the normalised, discounted rollout from time t onwards, i.e.: SRγ (s) = T −t 1 X ∆t γ I(st+∆t = s) Z ∆t=0 where Z is the normalisation constant such that P s SRγ (s) = 1. The loss here is then the cross-entropy between the predicted successor representation and the empirical one: XX c γ (s) LSR,i = −SRγ (s) log SR γ s Belief prediction. The agent’s belief states for each object k is a discrete distribution over 122 dims (the 11 × 11 locations on the map, plus an additional dimension for an absent object), denoted bk (s). For each object, k, the loss is the cross-entropy between the ToMnet’s predicted belief state and the agent’s true belief state: XX Lbelief,i = −bk (s) log b̂k (s) k s Deep Varational Information Bottleneck. In addition to these loss components, where DVIB was used, we included an additional term for the β-weighted KLs between posteriors and the priors LDV IB = βDKL (q(echar,i )||p(echar )) + βDKL (q(emental,i )||p(emental )) Each Mj contained four terminal objects. These objects could be consumed by the agent walking on top of them. Consuming these objects ended an episode. If no terminal object was consumed after 31 steps (random and algorithmic agents; Sections 3.1–3.2) or 51 steps (deep RL agents; Sections 3.3–3.5), the episodes terminated automatically as a time-out. The sampled walls may trap the agent, and make it impossible for the agent to terminate the episode without timing out. Deep RL agents (Sections 3.3–3.5) acted in gridworlds that contained an additional subgoal object. Consuming the subgoal did not terminate the episode. Reward functions for the agents were as follows: Random agents (Section 3.1.) No reward function. Algorithmic agents (Section 3.2). For a given agent, the reward function over the four terminal objects was drawn randomly from a Dirichlet with concentration parameter 0.01. Each agent thus has a sparse preference for one object. Penalty for each move: 0.01. Penalty for walking into a wall: 0.05. Greedy agents’ penalty for each move: 0.5. These agents planned their trajectories using value iteration, with a discount factor of 1. When multiple moves of equal value were available, these agents sampled from their best moves stochastically. Deep RL agents (Sections 3.3–3.5). Penalty for each move: 0.005. Penalty for walking into a wall: 0.05. Penalty for ending an episode without consuming a terminal object: 1. For each deep RL agent species (e.g. blind, stateless, 5 × 5, ...), we trained a number of canonical agents which received a reward of 1 for consuming the subgoal, and a reward of 1 for consuming a single preferred terminal object (e.g. the blue one). Consuming any other object yielded zero reward (though did terminate the episode). We artifically enlarged this population of trained agents by a factor of four, by inserting permutations into their observation functions, ωi , that effectively permuted the object channels. For example, when we took a trained blueobject-preferring agent, and inserted a transformation that swapped the third object channel with the first object channel, this agent behaved as a pink-object-preferring agent. Machine Theory of Mind D. Deep RL agent training and architecture Deep RL agents were based on the UNREAL architecture (Jaderberg et al., 2017). These were trained with over 100M episode steps, using 16 CPU workers. We used the Adam optimiser with a learning rate of 10−5 , and BPTT, unrolling over the whole episode (50 steps). Policies were regularised with an entropy cost of 0.005 to encourage exploration. We trained a total of 660 agents, spanning 33 random seeds × 5 fields of view × 2 architectures (feedforward/convolutional LSTM) × 2 depths (4 layer convnet or 2 layer convnet, both with 64 channels). We selected the top 20 agents per condition (out of 33 random seeds), by their average return. We randomly partitioned these sets into 10 training and 10 test agents per condition. With the reward permutations described above in Appendix C, this produced 40 training and 40 test agents per condition. Observations. Agents received an observation at each time step of nine 11 × 11 feature planes – indicating, at each location, whether a square was empty, a wall, one of the five total objects, the agent, or currently unobservable. Beliefs. We also trained agents with the auxiliary task of predicting the current locations of all objects in the map. To do this, we included an additional head to the Convolutional LSTMs, in addition to the policy (πt ) and baseline (Vt ) heads. This head output a posterior for each object’s location in the world, bk (i.e. a set of five 122-dim discrete distributions, over the 11 × 11 maze size, including an additional dimension for a prediction that that the object is absent). For the belief head, we used a 3-layer convnet with 32 channels and ReLU nonlinearities, followed by a softmax. This added a term to the training loss: the cross entropy between the current belief state and the true current world state. The loss for the belief prediction was scaled by an additional hyperparameter, swept over the values 0.5, 2, and 5. E. Additional results Model Train agents Test agents Action loss none char net + shuffled echar mental net + shuffled emental both + shuffled echar + shuffled emental + shuffled both 1.14 0.84 1.61 0.83 1.61 0.72 1.57 1.16 1.99 1.12 0.86 1.62 0.98 1.65 0.73 1.69 1.20 2.02 Consumption loss none char net + shuffled echar mental net + shuffled emental both + shuffled echar + shuffled emental + shuffled both 0.34 0.19 0.83 0.32 0.43 0.16 0.82 0.23 0.83 0.36 0.16 0.77 0.30 0.43 0.14 0.78 0.23 0.77 Successor loss none char net + shuffled echar mental net + shuffled emental both + shuffled echar + shuffled emental + shuffled both 2.48 2.23 3.17 2.36 2.92 2.16 3.27 2.45 3.53 2.53 2.21 3.13 2.29 2.80 2.04 3.19 2.33 3.31 Table A1. Full table of losses for the three predictions in Fig 7. For each prediction, we report the loss obtained by a trained ToMnet that had no character or mental net, had just a character net, just a mental net, or both. For each model, we quantify the importance of the embeddings echar and emental by measuring the loss when echar,i and emental,i are shuffled within a minibatch. The middle column shows the loss for the ToMnet’s predictions on new samples of behaviour from the agents used in the trained set. The right column shows this for agents in the test set. true agents 0.3 DJS (s, sswap ) successors: tomnet trained with swaps tomnet trained without swaps 3×3 actions: DJS (π, πswap ) Machine Theory of Mind F.2. Use of Jensen-Shannon Divergence 9×9 0 0.4 0 1 2 3 events. 4 swap distance 5 1 2 3 4 swap distance 5 1 2 3 4 swap distance 5 Figure A1. ToMnet performance on the Natural Sally-Anne test does not depend on the ToMnet observing swap events during training. The left two columns show the data presented in Fig 12 and Fig 13. The rightmost column shows the predictions of the ToMnet when it is trained on data from the same agents, but rolled out on POMDPs where the probability of swap events was p = 0 instead of p = 0.1. F. Additional notes F.1. Hypersensitivity of 3 × 3 agents to swap events with swap distance 1 In Fig 12c, the policies of agents with 3 × 3 fields of view are seen to be considerably more sensitive to swap events that occur adjacent to the agent than the agents with 9 × 9 fields of view. Agents with 5×5 and 7×7 had intermediate sensitivities. We did not perform a systematic analysis of the policy differences between these agents, but we speculate here as to the origin of this phenomenon. As we note in the main text, the agents were competent at their respective tasks, but not optimal. In particular, we noted that agents with larger fields of view were often sluggish to respond behaviourally to swap events. This is evident in the example shown on the left hand side of Fig 10. Here an agent with a 5 × 5 field of view does not respond to the sudden appearance of its preferred blue object above it by immediately moving upwards to consume it; its next-step policy does shift some probability mass to moving upwards, but only a small amount (Fig 10c). It strongly adjusts its policy on the following step though, producing rollouts that almost always return directly to the object (Fig 10d). We note that when a swap event occurs immediately next to an agent with a relatively large field of view (5 × 5 and greater), such an agent has the luxury of integrating information about the swap events over multiple timesteps, even if it navigates away from this location. In contrast, agents with 3 × 3 fields of view might take a single action that results in the swapped object disappearing altogether from their view. There thus might be greater pressure on these agents during learning to adjust their next-step actions in response to neighbouring swap In Sections 3.4–3.5, we used the Jensen-Shannon Divergence (DJS ) to measure the effect of swap events on agents’ (and the ToMnet’s predicted) behaviour (Figs 1213). We wanted to use a standard metric for changes to all the predictions (policy, successors, and beliefs), and we found that the symmetry and stability of DJS was most suited to this. We generally got similar results when using the KL-divergence, but we typically found more variance in these estimates: DKL is highly sensitive the one of the distributions assigning little probability mass to one of the outcomes. This was particularly problematic when measuring changes in the successor representations and belief states, which were often very sparse. While it’s possible to tame the the KL by adding a little uniform probability mass, this involves an arbitrary hyperparameter which we preferred to just avoid. G. Version history V 1: 21 F EB 2018 • Initial submission to arxiv. V 2: 12 M AR 2018 • Added missing references to opponent modelling in introduction. • Typographical error in citation (Dayan, 1993). • Figure 3 caption fixes: estimated probabilities → action likelihoods; missing descriptor of Npast for panels (c) and (d).
2cs.AI
Multi-label Image Recognition by Recurrently Discovering Attentional Regions Zhouxia Wang1,2 Tianshui Chen1 1 3 Guanbin Li1,3 2 Sun Yat-sen University, China Ruijia Xu1 Liang Lin1,2,3 ∗ SenseTime Group Limited Engineering Research Center for Advanced Computing Engineering Software of Ministry of Education, China arXiv:1711.02816v1 [cs.CV] 8 Nov 2017 {wzhoux, chtiansh, xurj3}@mail2.sysu.edu.cn, [email protected], [email protected] Abstract This paper proposes a novel deep architecture to address multi-label image recognition, a fundamental and practical task towards general visual understanding. Current solutions for this task usually rely on an extra step of extracting hypothesis regions (i.e., region proposals), resulting in redundant computation and sub-optimal performance. In this work, we achieve the interpretable and contextualized multi-label image classification by developing a recurrent memorized-attention module. This module consists of two alternately performed components: i) a spatial transformer layer to locate attentional regions from the convolutional feature maps in a region-proposal-free way and ii) an LSTM (Long-Short Term Memory) sub-network to sequentially predict semantic labeling scores on the located regions while capturing the global dependencies of these regions. The LSTM also output the parameters for computing the spatial transformer. On large-scale benchmarks of multi-label image classification (e.g., MS-COCO and PASCAL VOC 07), our approach demonstrates superior performances over other existing state-of-the-arts in both accuracy and efficiency. ... multi-label prediction ... Figure 1. Multi-label image recognition with discovered attentional regions by our approach. These regions (highlighted by different colors) corresponding to the semantic labels (visualized below the images) are contextualized and discriminative in terms of classification, although they may not preserve object boundaries well. ally needs understanding the image in depth (e.g., associating semantic labels with regions and capturing their dependencies). Recently, convolutional neural networks (CNNs) [18, 24] achieve great success in visual recognition/classification tasks by learning powerful feature representations from raw images, and they have been also applied to the problem of multi-label image classification by combining with some object localization techniques [30, 27]. The resulting common pipeline usually involves two steps. A batch of hypothesis regions are first produced by either exploiting bottom-up image cues [25] or casting extra detectors [4], and these regions are assumed to contain all possible foreground objects in the image. A classifier or neural network is then trained to predict the label score on these hypothesis regions, and these predictions are aggregated to achieve the multi-label classification results. Despite acknowledged successes, these methods take the redundant computational cost of extracting region proposals and usually over-simplify the contextual dependencies among foreground objects, leading to a sub-optimal performance in complex scenarios. Recently, Wang et al. [26] proposed to jointly characterize the semantic label dependency and the image-label relevance by combining recurrent neural networks (RNNs) with CNNs. However, their model disregards the explicit associations between semantic labels and 1. Introduction Recognizing multiple labels of images is a fundamental yet practical problem in computer vision, as real-world images always contain rich and diverse semantic information. Besides the challenges shared with single-label image classification (e.g., large intra-class variation caused by viewpoint, scale, occlusion, illumination), multi-label image classification is much more difficult since accurately predicting the presence of multiple object categories usu∗ Zhouxia Wang and Tianshui Chen contribute equally to this work and share first-authorship. Corresponding author is Liang Lin (Email: [email protected]). This work was supported by the State Key Development Program under Grant 2016YFB1001004, Special Program for Applied Research on Super Computation of the NSFC-Guangdong Joint Fund (second phase), and CCF-Tencent Open Research Fund (NO.AGR20160115). 1 image contents, and lacks fully exploiting the spatial context in images. In contrast to all these mentioned methods, we introduce an end-to-end trainable framework that explicitly discovers attentional regions over image scales corresponding to multiple semantic labels and captures the contextual dependencies of these regions from a global perspective. No extra step of extracting hypothesis regions is needed in our approach. Two examples generated by our approach are illustrated in Figure 1. To search for meaningful and discriminative regions in terms of multi-label classification, we propose a novel recurrent memorized-attention module, which is combined with convolutional neural networks in our framework. Specifically, this module consists of two components: i) a spatial transformer layer to locate attentional regions on the convolutional maps and ii) an LSTM (Long-Short Term Memory) sub-network to sequentially predict the labeling scores over the attentional regions and output the parameters of the spatial transformer layer. Notably, the global contextual dependencies among the attentional regions are naturally captured (i.e., memorized) together with the LSTM sequential encoding. And the two components are alternately performed during the recurrent learning. In this way, our approach enables to learn a contextualized and interpretable region-label relevance while improving the discriminability for multi-label classification. The main contributions of this work are three-fold. • We develop a proposal-free pipeline for multi-label image recognition, which is capable of automatically discovering semantic-aware regions over image scales and simultaneously capturing their long-range contextual dependencies. • We further propose three novel constraints on the spatial transformer, which help to learn more meaningful and interpretable regions, and in turn, facilitate multi-label classification. • We conduct extensive experiments and evaluations on large-scale benchmarks such as PASCAL VOC [6] and Microsoft COCO [20], and demonstrate the superiority of our proposed model in both recognition accuracy and efficiency over other leading multi-label image classification methods. 2. Related Works The performance of image classification has recently witnessed a rapid progress due to the establishment of largescale labeled datasets (i.e., PASCAL VOC [6], COCO [20]) and the fast development of deep CNNs [24, 12]. In recent years, many researchers have attempted to adapt the deep CNNs to multi-label image recognition problem and have achieved great success. 2.1. Multi-label image recognition Traditional multi-label image recognition methods apply the bag-of-words (BOW) model to solve this problem [3, 5]. Although performing well on the simple benchmarks, these methods may fail in classifying images with complex scenes since BOW based models depend largely on the hand-crafted low-level features. In contrast, features learned by deep models have been confirmed to be highly versatile and far more effective than the handcrafted features. Since this paper focuses on deep learning based multi-label image recognition, we discuss the relevant works in the following context. Recently, there have been attempts to apply deep learning to multi-label image recognition task [23, 24, 30, 27, 26]. Razavian et al. [23] applies off-the-shelf features extracted from deep network pretrained on ImageNet [22] for multi-label image classification. Gong et al. [10] propose to combine convolutional architectures with an approximate top-k ranking objective function for annotating multi-label images. Instead of extracting off-the-shelf deep features, Chatfield et al. [2] fine tune the network with the target multi-label datasets, which can learn task-specific features and thus boost the classification performance. To better consider the correlations between labels instead of treating each label independently, traditional graphical models are widely incorporated, such as Conditional Random Field [8], Dependency Network [11], and co-occurrence matrix [29]. Recently, Wang et al. [26] utilize the RNNs to learn a joint image-label embedding to characterize the semantic label dependency as well as the image-label relevance. All of the aforementioned methods consider extracting the features of the whole image with no spatial information, which on one hand were unable to explicitly perceive the corresponding image regions to the detected classification labels, and on the other hand, were extremely vulnerable to the complex background. To overcome this issue, some researchers propose to exploit object proposals to only focus on the informative regions, which effectively eliminate the influences of the non-object areas and thus demonstrate significant improvement in multi-label image recognition task [30, 27]. More specifically, Wei et al. [27] propose a Hypotheses-CNN-Pooling framework to aggregate the label scores of each specific object hypotheses to achieve the final multi-label predictions. Yang et al. [30] formulate the multi-label image recognition problem as a multi-class multi-instance learning problem to incorporate local information and enhance the discriminative ability of the features by encoding the label view information. However, these object proposals based methods are generally not efficient with the preprocessing step of object proposal generation being the bottleneck. Moreover, the training stage is not perfect and can hardly be modeled as an end-to-end scheme in both training and testing. In this paper, we pro- fI M0 cinit , hinit ST LSTM f0 M1 c0 , h0 M1 , c0 , h0 CNN fI M1 c0 , h0 ST LSTM f1 c1 , h1 label distribution M2 s1 category-wise max-pooling … fI cK-1 , hK-1 ST LSTM … fI MK fK sK Figure 2. Overview of our proposed framework for multi-label image recognition. Our model iteratively locates the attentional regions corresponding to semantic labels and predicts the score for the current region. pose to incorporate a recurrent memorized-attention module in the neural network to simultaneously locate the attentional regions and predict the labels on various located regions. Our proposed method does not resort to the extraction of object proposals and is thus very efficient and can be trained in an end-to-end mode. 2.2. Visual attention model Attention model has been recently applied to various computer vision tasks, including image classification [21, 1], saliency detection [19], and image captioning [28]. Most of these works use the recurrent neural network for sequential attentions, and optimized their models with reinforcement learning technique. Works [21, 1] formulate a recurrent attention model and apply it to the digital classification tasks for which the images are low-resolution with a clean background, using the small attention network. The model is non-differential and addressed with reinforcement learning to learn task-specific policies. Jaderberg et al. [15] propose a differential spatial transformer module which could be used to extract attentional regions with any spatial transformation, including scaling, rotation, transition, and cropping. Moreover, it could be easily integrated into the neural network and optimized using the standard backpropagation algorithm without reinforcement learning. 3. Model Figure 2 illustrates the architecture of the proposed model. The input image I is first fed into a VGG-16 ConvNet without additional object proposals. The network first processes the whole image with several convolutional (conv) and max pooling layers to produce the conv feature maps, denoted as fI . Here, we use the conv feature maps from the last conv layer (i.e., conv5 3). The recurrent memorized-attention module, comprising a spatial transformer (ST) [15] and an LSTM network [14] that work collaboratively in an iterative manner, predicts the label distributions directly from the input image features. Specifically, in one iterative procedure, the ST locates an attentional region for the LSTM, and the LSTM predicts the scores regarding this region for multi-label classification and simultaneously updates the parameters of ST. Finally, the scores from several attentional regions are fused to achieve the final label distribution. 3.1. ST for Attentional Region Localization We briefly introduce the spatial transformer (ST) [15] for completeness before diving deep into the recurrent memorized-attention module. ST is a sample-based differential module that spatially transforms its input maps to the output maps with a given size which correspond to a subregion of the input maps. It is convenient to embed an ST layer in the neural network and train it with the standard back-propagation algorithm. In our model, the ST is incorporated in the recurrent memorized-attention module for the localization of attentional regions. Formally, the ST layer extracts features of an attentional region, denoted as fk , from the feature maps fI of the whole input image. The computational procedure is as follows. A transformation matrix M is first estimated by a localization network (explained later). After that, the corresponding coordinate grid in fI is obtained, based on the coordinates of fk . Then the sampled feature maps fk that correspond to the attentional region are generated by bilinear interpolation. Fig. 3 shows an example of coordinate mapping. As we aim to locate the attentional regions, we constrain the transformation matrix M to involve only cropping, translation and scaling, expressed as M xk = relu(Wf x fk + bx ) M ik = σ(Wxi xk + Whi hk−1 + bi ) gk = σ(Wxg xk + Whg hk−1 + bg ) U U ok = σ(Wxo xk + Who hk−1 + bo ) (4) mk = tanh(Wxm xk + Whm hk−1 + bm ) V V (a) ck = gk ck−1 + ik hk = ok ck mk (b) Figure 3. Illustration of coordinate grid mapping on (a) the feature maps and (b) the corresponding input image.  s M= x 0 0 sy  tx , ty (1) where sx , sy , tx , ty are the scaling and translation parameters. In our model, we apply a standard neural network to estimate these parameters to facilitate an end-to-end learning scheme. 3.2. Recurrent Memorized-Attention Module The core of our proposed model is the recurrent memorized-attention module, which combines the recurrent computation process of an LSTM network and a spatial transformer. It iteratively searches the most discriminative regions, and predicts the scores of label distribution for them. In this subsection, we introduce this module in detail. In the k-th iteration, our model searches an attentional region, and extracts the corresponding features by applying the following spatial transformer, expressed as  k s fk = st(fI , Mk ), Mk = x 0 0 sky  tkx , tky where relu(·) is the rectified linear function, σ(·) is the sigmoid function, tanh(·) is the hyperbolic tangent function; hk−1 and ck−1 are the hidden state and memory cell of previous iteration; ik , gk , ok and mk are the outputs of the input gate, forget gate, output gate, and input modulation gate, respectively. These multiplicative gates can ensure the robust training of LSTMs as they work well in exploding and vanishing gradients [14]. The memory cell ck encodes the useful information of previous (k − 1) regions, and it is possible to benefit our task in the following two aspects. First, previous works [29, 26] have shown that different categories of objects exhibit strong co-occurrence dependencies. Therefore, it helps to recognize objects within the current attentional region aided by “remembering” information of previous ones. Second, it is expected that our model can find out all relevant and useful regions for classification. Simultaneously considering the information of previous regions is a feasible approach that implicitly enhances the diversity and complementarity among the attentional regions. Update rule of M. Given the hidden state hk , the classifier and localization network can be expressed as zk = relu(Whz hk + bz ) (2) where st(·) is the spatial transformation function, and Mk is the transformation matrix estimated in the previous round by the localization network. We initialize the attentional region with the whole image at the first iteration, i.e., the initial transformation matrix is set to be   1 0 0 M0 = . (3) 0 1 0 Note that we apply the spatial transformer operation on the feature maps fI instead of the input image to avoid repeating the computational intensive convolutional processes. The LSTM takes the sampled feature map fk as input to compute the memory cell and hidden state. The computation process can be expressed as sk = Wzs zk + bs , k 6= 0 (5) Mk+1 = Wzm zk + bm where sk is the predicted score distribution of the k-th region, and Mk+1 is the transformation matrix for the next iteration. Note that at the first iteration (k = 0), we make no prediction of s and just estimate the matrix M because no attentional region is obtained initially. Category-wise max-pooling. The iterations are repeated for K + 1 times, resulting in K score vectors {s1 , s2 , . . . , sK }, where sk = {s1k , s2k , . . . , sC k } denotes the scores over C class labels. Following [27], we employ the category-wise max-pooling to fuse the scores into the final result s = {s1 , s2 , . . . , sC }. It simply maximizes out the scores over regions for each category sc = max(sc1 , sc2 , . . . , scK ), c = 1, 2, . . . , C. (6) 4. Learning 4.1. Loss for Classification We employ the Euclidean loss as the objective function following [27, 30]. Suppose there are N training samples, and each sample xi has its label vector yi = {yi1 , yi2 , . . . , yiC }. yic (c = 1, 2, . . . , C) is assigned as 1 if the sample is annotated with the class label c, and 0 otherwise. The ground-truth probability vector of the i-th sample is defined as p̂i = yi /||yi ||1 . Given the predicted probability vector pi exp(sci ) pci = PC 0 c0 =1 exp(sci ) c = 1, 2, . . . , C, (7) Figure 4. Anchor selection for left: K=5 and right: K=9. the circle uniformly, as depicted in Figure 4. We use K = 5 in the experiments, so four anchor points are generated at (0.5, 0.5), (0.5, −0.5), (−0.5, 0.5), and (−0.5, −0.5), respectively1 . The anchor constraint is formulated as and the classification loss function is expressed as Lcls = 1 N N X C X (pci − p̂ci )2 . `A = (8) i=1 c=1 4.2. Loss for Attentional Region Constraints As discussed above, we obtain the final result by aggregating the scores of the attentional regions. Thus, we hope that the attentional regions selected by our model contain all of the objects in the input image. If one object is left out unexpectedly, an inevitable error occurs because the LSTM network has never seen this object during the prediction procedure. We experimentally found that the proposed model can be trained with the defined classification loss, however, has notable drawbacks: • Redundancy. The ST layer usually picks up the same region that corresponds to the most salient objects. As a result, it would be difficult to retrieve all of the objects appearing in the input image, since the set of attentional regions are redundant. • Neglect of tiny objects. The ST layer tends to locate regions in a relatively large size and ignores the tiny objects, which hampers the classification performance. • Spatial flipping. The selected attentional region may be mirrored vertically or horizontally. To address these issues, we further define a loss function that consists of three constraints on the parameters of the transformation matrix M. Anchor constraint. It would be better if the attentional regions scatter over different semantic regions in the image. For the first iteration, adding no constraint helps to find the most discriminative region. After that, we push the other (K − 1) attentional regions away from the image center by √ an anchor constraint. We draw a circle of radius 22 centered on the image center, and pick up the anchor points on 1 k {(t − ckx )2 + (tky − cky )2 }, 2 x (9) where (ckx , cky ) is the location of the k-th anchor point. Scale constraint. This constraint attempts to push the scale parameters in a certain range, so that the located attentional region will not be too large in size. It can be formulated as `S = `sx + `sy , (10) in which `sx = (max(|sx | − α, 0))2 `sy = (max(|sy | − α, 0))2 (11) where α is a threshold value, and it is set as 0.5 in our experiments. Positive constraint. The last one also constrains the scale parameters. Positive constraint prefers a transformation matrix with positive scale parameters, leading to attentional regions that are not be mirrored: `P = max(0, β − sx ) + max(0, β − sy ), (12) where β is a threshold value, set as 0.1 in our experiments. Finally, we combine the aforementioned three types of constraints on the parameters of the transformation matrix to define a loss of localization of attentional regions. It is formulated as the weighted sum of the three components: Lloc = `S + λ1 `A + λ2 `P , (13) where λ1 and λ2 are the weighted parameters, and they are set as 0.01 and 0.1, respectively. Our model is jointly trained with the classification loss and the localization loss, so the overall loss function can be expressed as L = Lcls + γLloc . (14) We set the balance parameter γ as 0.1 since the classification task is dominated in our model. Optimization is performed using the recently proposed Adam algorithm [17] and standard back-propagation. 1 The range of coordinate is rescaled to [-1, 1] 5. Experiments 5.1. Settings Implementation details. We implemented our method on the basis of Caffe [16] for deep network training and testing. In the training stage, we employed a two-step training mechanism to initialize the convolutional neural network following [27]. The CNN is first pre-trained on the ImageNet, a large scale single label classification dataset, and further fine-tuned on the target multi-label classification dataset. The learned parameters are used to initialize the parameters of the corresponding layers in our proposed model, while the parameters of other newly added layers in our network are initialized with Xavier algorithm [9, 13] rather than manual tuning. All the training images are first resized to N × N , and randomly cropped with a size of (N − 64) × (N − 64). The training samples are also augmented by horizontal flipping. In our experiments, we trained two models with N = 512 and N = 640, respectively. Both of the models are optimized using Adam with a batch size of 16, momentum of 0.9 and 0.999. The learning rate is set to 0.00001 initially and divided by 10 after 30 epochs. We trained the models for about 45 epochs for each scale, and selected the model with the lowest validation loss as the best model for testing. In the testing phase, we follow [18] to perform ten-view evaluation across different scales. Specifically, we first resized the input image to N × N (N = 512, 640), and extracted five patches (i.e., the four corner patches and the center patch) with a size of (N − 64) × (N − 64), as well as their horizontally flipped versions. Instead of repeatedly extracting features for each patch, the model feeds the N×N image to the VGG-16 ConvNet, and crops the features on the conv5 3 features maps accordingly to achieve the features of all patches. Then, for each patch, the model extracts the features for each located attentional region, and eventually aggregates the features of this patch by max-pooling. The image representation is obtained via averaging the features of all the patches. At last, we trained a one-vs-rest SVM classifier for each category using the LIBLINEAR library [7]. We test our model on a single NVIDIA GeForce GTX TITAN-X, and it takes about 150ms for ten-view evaluation for scale 512, and about 200 ms for scale 640. It reduces the execution time by more than an order of magnitude, compared with previous proposal-based methods, e.g., HCP [27], which costs about 10s per image. Evaluation metrics. We use the same evaluation metrics as [26]. For each image, we assign top k highest-ranked labels to the image, and compare with the ground-truth labels. We compute the overall precision, recall, F1 (OP, OR, OF1) and per-class precision, recall, F1 (CP, CR, CF1) in Eq. 15. Following [27, 5], we also apply average precision (AP) for each category, and the mean average precision (mAP) over all categories as well. Generally, overall F1, per-class F1, and mAP are relatively important. P c N CP = OP = Pi ip N i Pi c N CR = OR = Pi ig i Ni 2 × OP × OR OF 1 = OP + OR CF 1 = 1 X Nic C i Nip 1 X Nc i C i Nig (15) 2 × CP × CR , CP + CR where C is the number of labels, Nic is the number of images that are correctly predicted for the i-th label, Nip is the number of predicted images for the i-th label, Nig is the number of ground truth images for the i-th label. 5.2. Comparison with State-of-the-art Methods To validate the effectiveness of our model, we conduct the experiments on two benchmarks, PASCAL VOC 2007 [6] and Microsoft COCO [20]. VOC 2007 is the most widely used benchmark, and most works have reported the results on this dataset. We compare the performance of our proposed method against the following state-of-the-art approaches: FeV+LV-20-VD [30], HCP [27], RLSD [31], CNN-RNN [26], VeryDeep [24] and CNN-SVM [23] on the VOC 2007 dataset. MS-COCO is released later than VOC and more challenging. Recent works have also used this benchmark for evaluation. We compare with CNNRNN [26], RLSD [31] and WARP [10] on the COCO dataset as well. 5.2.1 Performance on the VOC 2007 dataset The PASCAL VOC 2007 dataset contains 9,963 images from 20 object categories, which is divided into train, val and test sets. We train our model on the trainval set, and evaluate the performance on the test set, following other competitors. Table 1 presents the experimental results. The previous best-performing methods are HCP and FeV+LV, which achieve a mAP of 90.9% and 90.6%, respectively. Both of them share a similar two-step pipeline: they first extract the object proposals of the image, and then aggregate the features of them for multi-label classification. Different from them, our method is proposal-free since the attentional regions are selected by the ST layer that works collaboratively with the LSTM network. In this way, the interaction between attentional region localization and classification is well explored, leading to improvement in performance. Our proposed method achieves a mAP of 91.9%, that outperforms previous state-of-the-art algorithms. Note that our model learned with a single scale of 512 or 640 also surpasses previous works. This better demonstrates the effectiveness of the proposed method. Methods CNN-SVM [23] CNN-RNN [26] VeryDeep [24] RLSD [31] HCP [27] FeV+LV [30] Ours (512) Ours (640) Ours aero 88.5 96.7 98.9 96.4 98.6 97.9 98.5 97.7 98.6 bike 81.0 83.1 95.0 92.7 97.1 97.0 96.7 97.3 97.4 bird 83.5 94.2 96.8 93.8 98.0 96.6 95.6 96.4 96.3 boat 82.0 92.8 95.4 94.1 95.6 94.6 95.7 95.8 96.2 bottle 42.0 61.2 69.7 71.2 75.3 73.6 73.7 74.6 75.2 bus 72.5 82.1 90.4 92.5 94.7 93.9 92.1 91.9 92.4 car 85.3 89.1 93.5 94.2 95.8 96.5 95.8 96.5 96.5 cat 81.6 94.2 96.0 95.7 97.3 95.5 96.8 96.7 97.1 chair 59.9 64.2 74.2 74.3 73.1 73.7 76.5 75.2 76.5 cow 58.5 83.6 86.6 90.0 90.2 90.3 92.9 89.9 92.0 table 66.5 70.0 87.8 74.2 80.0 82.8 87.2 87.1 87.7 dog 77.8 92.4 96.0 95.4 97.3 95.4 96.6 96.0 96.8 horse 81.8 91.7 96.3 96.2 96.1 97.7 97.5 96.9 97.5 mbike 78.8 84.2 93.1 92.1 94.9 95.9 92.8 93.2 93.8 person 90.2 93.7 97.2 97.9 96.3 98.6 98.3 98.4 98.5 plant 54.8 59.8 70.0 66.9 78.3 77.6 76.9 81.3 81.6 sheep 71.1 93.2 92.1 93.5 94.7 88.7 91.3 93.4 93.7 sofa 62.6 75.3 80.3 73.7 76.2 78.0 83.6 81.3 82.8 train 87.2 99.7 98.1 97.5 97.9 98.3 98.6 98.3 98.6 tv 71.8 78.6 87.0 87.6 91.5 89.0 88.1 88.5 89.3 mAP 73.9 84.0 89.7 88.5 90.9 90.6 91.3 91.3 91.9 Table 1. Comparison of AP and mAP in % of our model and state-of-the-art methods on the PASCAL VOC 2007 dataset. The best results and second best results are highlighted in red and blue, respectively. Best viewed in color. Methods WARP [10] CNN-RNN [26] RLSD [31] Ours (512) Ours (640) Ours C-P 59.3 66.0 67.6 77.7 78.0 79.1 C-R 52.5 55.6 57.2 58.1 57.7 58.7 C-F1 55.7 60.4 62.0 66.5 66.3 67.4 O-P 59.8 69.2 70.1 83.4 83.8 84.0 O-R 61.4 66.4 63.4 62.3 62.3 63.0 O-F1 60.7 67.8 66.5 71.3 71.4 72.0 Table 2. Comparison of our model and state-of-the-art methods on the MS-COCO dataset. The best results and second best results are highlighted in red and blue, respectively. Best viewed in color. 5.2.2 Performance on the MS-COCO dataset The MS-COCO dataset is primarily built for object detection, and it is also widely used for multi-label recognition recently. It comprises a training set of 82,081 images, and a validation set of 40,137 images. The dataset covers 80 common object categories, with about 3.5 object labels per image. The label number for each image also varies considerably, rendering MS-COCO even more challenging. As the ground truth labels of the test set are not available, we evaluate the performance of all the methods on the validation set instead. We follow [26] to select the top k = 3 labels for each image, and filter out the labels with probabilities lower than a threshold 0.5, so the label number of some images would be less than 3. We compare the overall precision, recall, F1, and perclass precision, recall, F1 in Table 2. Our model outperforms the existing methods by a sizable margin. Specifically, it achieves a per-class F1 score of 67.4% and an overall F1 score of 72.0, improving those of the previously best method by 5.4% and 4.2%, respectively. Similar to the results on VOC, the model learned with a single scale also beats the state-of-the-art approaches. 5.3. Ablation Study In this subsection, we perform ablative studies to carefully analyze the contribution of the critical components of our proposed model. 5.3.1 Attentional regions v.s. object proposals One of the main contributions of this work is that our model is capable of discovering the discriminative regions, which Figure 5. Comparison of visualization of the attentional regions (indicated by green boxes) located by our method, and the object proposals (indicated by blue boxes) generated by EdgeBox. facilitates the task of multi-label image classification compared with proposal-based methods. In this subsection, we present a comparison to reveal the fact that attentional regions have significant advantages against object proposals. Proposal-based methods are proved to be powerful for objectness detection. However, satisfactory recall rates are difficult to achieve until thousands of proposals are provided. In addition, it is extremely time-consuming to examine all of the provided proposals with a deep network. As an example, although HCP selects some representative proposals, it still needs 500 proposals to obtain desired performance. Besides, computing the object proposals also introduces additional computational overhead. In contrast, our model utilizes an efficient spatial transformation layer to find out a small number of discriminative regions, making the model runs much faster. Here we also present the visualization results of attentional regions discovered by our model, and those generated by EdgeBox [32], a representative proposal method. For our method, K is set as 5, so five attentional regions are found. For EdgeBox, we directly use the codes provided by [32] to extract the proposals, and adopt non-maximum suppression (NMS) with a threshold of 0.7 on them based on their objectness scores to exclude the seriously overlapped proposals. We also visualize the top five ones for a fair comparison. Figure 5 shows that the regions generated by our model better capture the discriminative regions (e.g., the head part of dogs), and most of them concentrate on the area of semantic objects. For EdgeBox, although its top-5 proposals cover most objects in the given image, most of them contain non-object areas that carry less discriminative information for classification. In order to clearly show the advantages of attentional region localization, we conduct experiments to compare the classification performance when using attentional regions or object proposals in the same framework. To this end, we first remove the spatial transformer and replace the attentional region with the selected five object proposals, with the other components left unchanged. Table 3 gives the results on the VOC 2007 dataset. It is shown that attentional regions lead to better performance. In fact, proposal-based methods need hundreds of regions or even more proposals to cover most objects. Our model also achieves better performance than those using hundreds of proposals, such as HCP and FeV+LV, which use 500 and 400 proposals, respectively (see Table 1). type object proposals attentional regions mAP 88.6 90.4 Table 3. Comparison of the mAPs of our model using attentional regions and object proposals, respectively, on the PASCAL VOC 2007 dataset. The results are all evaluated using single-crop at scale of 512×512 5.3.2 Analysis of the attentional region constraints We propose three types of novel constraints for attentional region localization, facilitating the task of multi-label image classification. To validate their contributions, we remove all three constraints, and retrain the model on the VOC 2007 dataset. The results, depicted in Table 4, show a significant drop in mAP, well demonstrating the effectiveness of the constraints as a whole. We further remove one of three constraints, and retrain the model to evaluate the effectiveness of each constraint individually. The performance also declines when any constraint is excluded (see Table 4). Therefore, it suggests that all of the three constraints are of importance for our model, they work cooperatively to facilitate the improvement of classification. We also conduct similar experiments on the MS-COCO dataset. As Table 5 shown, although MS-COCO is far different from VOC, similar results have been observed, again demonstrating their contributions on various scenes. 5.3.3 Multi-scale multi-view evaluation We assess the impact of fusion of multi-scale and multi-crop at the test stage. Two scales (512 × 512 and 640 × 640) are used in our experiments. For each scale, we extract ten crop features. Hence, we reported the performance of singlescale + single-crop, single-scale + multi-crop and multiscale + multi-crop, in Table 6. The results show that aggregating information from multi-crop on a single-scale can constraints null S+P A+S A+P A+S+P mAP 89.9 90.2 90.4 90.3 91.3 Table 4. Comparison of mAP of our model learned using different constraints of attentional region localization on the PASCAL VOC 2007 dataset. The results are all evaluated using multi-crop at the scale of 512×512. We abbreviate anchor, scale and positive constraints as A, S, P for simple illustration. constraints null A+S+P C-F1 65.8 66.5 O-F1 70.9 71.3 mAP 71.5 72.2 Table 5. Comparison of C-F1, O-F1 and mAP of our model learned with and without constraints of attentional region localization on the MS-COCO dataset. The results are all evaluated using multicrop at the scale of 512×512. We abbreviate anchor, scale and positive constraints as A, S, P for simple illustration. s=512 + single-crop s=640 + single-crop s=512 + ten-crop s=640 + ten-crop two scales + ten-crop VOC 2007 90.4 90.4 91.3 91.3 91.9 MS-COCO 70.4 70.5 72.2 72.3 73.4 Table 6. Comparison of mAP with multi-scale and multi-crop on the PASCAL VOC 2007 and MS-COCO datasets. boost the performance, and fusing the results of both of two scale shows a further improvement. 6. Conclusion In this paper, we have introduced a recurrent memorizedattention module into the deep neural network architecture to solve the problem of multi-label image recognition. Specifically, our proposed recurrent memorized-attention module is composed of a spatial transformer layer for localizing attentional regions from the image and an LSTM unit to predict the labeling score based on the feature of a localized region and preserve the past information for the located regions. Experimental results on large-scale benchmarks (e.g., PASCAL VOC, COCO) demonstrate that our proposed deep model can significantly improve the state of the art in both accuracy and efficiency. References [1] J. Ba, V. Mnih, and K. Kavukcuoglu. Multiple object recognition with visual attention. arXiv preprint arXiv:1412.7755, 2014. [2] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman. Return of the devil in the details: Delving deep into convolutional nets. arXiv preprint arXiv:1405.3531, 2014. [3] Q. Chen, Z. Song, Y. Hua, Z. Huang, and S. Yan. Hierarchical matching with side information for image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3426–3433, 2012. [4] M.-M. Cheng, Z. Zhang, W.-Y. Lin, and P. Torr. Bing: Binarized normed gradients for objectness estimation at 300fps. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3286–3293, 2014. [5] J. Dong, W. Xia, Q. Chen, J. Feng, Z. Huang, and S. Yan. Subcategory-aware object classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 827–834, 2013. [6] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303– 338, 2010. [7] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.J. Lin. Liblinear: A library for large linear classification. Journal of machine learning research, 9(Aug):1871–1874, 2008. [8] N. Ghamrawi and A. McCallum. Collective multi-label classification. In Proceedings of the 14th ACM international conference on Information and knowledge management, pages 195–200. ACM, 2005. [9] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In Aistats, volume 9, pages 249–256, 2010. [10] Y. Gong, Y. Jia, T. Leung, A. Toshev, and S. Ioffe. Deep convolutional ranking for multilabel image annotation. arXiv preprint arXiv:1312.4894, 2013. [11] Y. Guo and S. Gu. Multi-label classification using conditional dependency networks. [12] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. [13] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision, pages 1026–1034, 2015. [14] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997. [15] M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial transformer networks. In Advances in Neural Information Processing Systems, pages 2017–2025, 2015. [16] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of ACM international conference on Multimedia, pages 675– 678, 2014. [17] D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. [18] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012. [19] J. Kuen, Z. Wang, and G. Wang. Recurrent attentional networks for saliency detection. arXiv preprint arXiv:1604.03227, 2016. [20] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer, 2014. [21] V. Mnih, N. Heess, A. Graves, et al. Recurrent models of visual attention. In Advances in Neural Information Processing Systems, pages 2204–2212, 2014. [22] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. [23] A. Sharif Razavian, H. Azizpour, J. Sullivan, and S. Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 806– 813, 2014. [24] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. [25] J. R. Uijlings, K. E. Van De Sande, T. Gevers, and A. W. Smeulders. Selective search for object recognition. International journal of computer vision, 104(2):154–171, 2013. [26] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu. Cnn-rnn: A unified framework for multi-label image classification. arXiv preprint arXiv:1604.04573, 2016. [27] Y. Wei, W. Xia, M. Lin, J. Huang, B. Ni, J. Dong, Y. Zhao, and S. Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE transactions on pattern analysis and machine intelligence, 2015. [28] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhutdinov, R. S. Zemel, and Y. Bengio. Show, attend and tell: Neural image caption generation with visual attention. arXiv preprint arXiv:1502.03044, 2015. [29] X. Xue, W. Zhang, J. Zhang, B. Wu, J. Fan, and Y. Lu. Correlative multi-label multi-instance image annotation. In 2011 International Conference on Computer Vision, pages 651– 658. IEEE, 2011. [30] H. Yang, J. Tianyi Zhou, Y. Zhang, B.-B. Gao, J. Wu, and J. Cai. Exploit bounding box annotations for multi-label object recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 280–288, 2016. [31] J. Zhang, Q. Wu, C. Shen, J. Zhang, and J. Lu. Multi-label image classification with regional latent semantic dependencies. arXiv preprint arXiv:1612.01082, 2016. [32] C. L. Zitnick and P. Dollár. Edge boxes: Locating object proposals from edges. In European Conference on Computer Vision, pages 391–405. Springer, 2014.
1cs.CV
arXiv:1506.00019v4 [cs.LG] 17 Oct 2015 A Critical Review of Recurrent Neural Networks for Sequence Learning Zachary C. Lipton [email protected] John Berkowitz [email protected] Charles Elkan [email protected] June 5th, 2015 Abstract Countless learning tasks require dealing with sequential data. Image captioning, speech synthesis, and music generation all require that a model produce outputs that are sequences. In other domains, such as time series prediction, video analysis, and musical information retrieval, a model must learn from inputs that are sequences. Interactive tasks, such as translating natural language, engaging in dialogue, and controlling a robot, often demand both capabilities. Recurrent neural networks (RNNs) are connectionist models that capture the dynamics of sequences via cycles in the network of nodes. Unlike standard feedforward neural networks, recurrent networks retain a state that can represent information from an arbitrarily long context window. Although recurrent neural networks have traditionally been difficult to train, and often contain millions of parameters, recent advances in network architectures, optimization techniques, and parallel computation have enabled successful large-scale learning with them. In recent years, systems based on long short-term memory (LSTM) and bidirectional (BRNN) architectures have demonstrated ground-breaking performance on tasks as varied as image captioning, language translation, and handwriting recognition. In this survey, we review and synthesize the research that over the past three decades first yielded and then made practical these powerful learning models. When appropriate, we reconcile conflicting notation and nomenclature. Our goal is to provide a selfcontained explication of the state of the art together with a historical perspective and references to primary research. 1 Introduction Neural networks are powerful learning models that achieve state-of-the-art results in a wide range of supervised and unsupervised machine learning tasks. 1 They are suited especially well for machine perception tasks, where the raw underlying features are not individually interpretable. This success is attributed to their ability to learn hierarchical representations, unlike traditional methods that rely upon hand-engineered features [Farabet et al., 2013]. Over the past several years, storage has become more affordable, datasets have grown far larger, and the field of parallel computing has advanced considerably. In the setting of large datasets, simple linear models tend to under-fit, and often under-utilize computing resources. Deep learning methods, in particular those based on deep belief networks (DNNs), which are greedily built by stacking restricted Boltzmann machines, and convolutional neural networks, which exploit the local dependency of visual information, have demonstrated record-setting results on many important applications. However, despite their power, standard neural networks have limitations. Most notably, they rely on the assumption of independence among the training and test examples. After each example (data point) is processed, the entire state of the network is lost. If each example is generated independently, this presents no problem. But if data points are related in time or space, this is unacceptable. Frames from video, snippets of audio, and words pulled from sentences, represent settings where the independence assumption fails. Additionally, standard networks generally rely on examples being vectors of fixed length. Thus it is desirable to extend these powerful learning tools to model data with temporal or sequential structure and varying length inputs and outputs, especially in the many domains where neural networks are already the state of the art. Recurrent neural networks (RNNs) are connectionist models with the ability to selectively pass information across sequence steps, while processing sequential data one element at a time. Thus they can model input and/or output consisting of sequences of elements that are not independent. Further, recurrent neural networks can simultaneously model sequential and time dependencies on multiple scales. In the following subsections, we explain the fundamental reasons why recurrent neural networks are worth investigating. To be clear, we are motivated by a desire to achieve empirical results. This motivation warrants clarification because recurrent networks have roots in both cognitive modeling and supervised machine learning. Owing to this difference of perspectives, many published papers have different aims and priorities. In many foundational papers, generally published in cognitive science and computational neuroscience journals, such as [Hopfield, 1982, Jordan, 1986, Elman, 1990], biologically plausible mechanisms are emphasized. In other papers [Schuster and Paliwal, 1997, Socher et al., 2014, Karpathy and Fei-Fei, 2014], biological inspiration is downplayed in favor of achieving empirical results on important tasks and datasets. This review is motivated by practical results rather than biological plausibility, but where appropriate, we draw connections to relevant concepts in neuroscience. Given the empirical aim, we now address three significant questions that one might reasonably want answered before reading further. 2 1.1 Why model sequentiality explicitly? In light of the practical success and economic value of sequence-agnostic models, this is a fair question. Support vector machines, logistic regression, and feedforward networks have proved immensely useful without explicitly modeling time. Arguably, it is precisely the assumption of independence that has led to much recent progress in machine learning. Further, many models implicitly capture time by concatenating each input with some number of its immediate predecessors and successors, presenting the machine learning model with a sliding window of context about each point of interest. This approach has been used with deep belief nets for speech modeling by Maas et al. [2012]. Unfortunately, despite the usefulness of the independence assumption, it precludes modeling long-range dependencies. For example, a model trained using a finite-length context window of length 5 could never be trained to answer the simple question, “what was the data point seen six time steps ago?” For a practical application such as call center automation, such a limited system might learn to route calls, but could never participate with complete success in an extended dialogue. Since the earliest conception of artificial intelligence, researchers have sought to build systems that interact with humans in time. In Alan Turing’s groundbreaking paper Computing Machinery and Intelligence, he proposes an “imitation game” which judges a machine’s intelligence by its ability to convincingly engage in dialogue [Turing, 1950]. Besides dialogue systems, modern interactive systems of economic importance include self-driving cars and robotic surgery, among others. Without an explicit model of sequentiality or time, it seems unlikely that any combination of classifiers or regressors can be cobbled together to provide this functionality. 1.2 Why not use Markov models? Recurrent neural networks are not the only models capable of representing time dependencies. Markov chains, which model transitions between states in an observed sequence, were first described by the mathematician Andrey Markov in 1906. Hidden Markov models (HMMs), which model an observed sequence as probabilistically dependent upon a sequence of unobserved states, were described in the 1950s and have been widely studied since the 1960s [Stratonovich, 1960]. However, traditional Markov model approaches are limited because their states must be drawn from a modestly sized discrete state space S. The dynamic programming algorithm that is used to perform efficient inference with hidden Markov models scales in time O(|S|2 ) [Viterbi, 1967]. Further, the transition table capturing the probability of moving between any two time-adjacent states is of size |S|2 . Thus, standard operations become infeasible with an HMM when the set of possible hidden states grows large. Further, each hidden state can depend only on the immediately previous state. While it is possible to extend a Markov model to account for a larger context window by creating a new state space equal to the cross product of the possible states at each time in the window, this procedure grows the state space exponentially with the size of the 3 window, rendering Markov models computationally impractical for modeling long-range dependencies [Graves et al., 2014]. Given the limitations of Markov models, we ought to explain why it is reasonable that connectionist models, i.e., artificial neural networks, should fare better. First, recurrent neural networks can capture long-range time dependencies, overcoming the chief limitation of Markov models. This point requires a careful explanation. As in Markov models, any state in a traditional RNN depends only on the current input as well as on the state of the network at the previous time step.1 However, the hidden state at any time step can contain information from a nearly arbitrarily long context window. This is possible because the number of distinct states that can be represented in a hidden layer of nodes grows exponentially with the number of nodes in the layer. Even if each node took only binary values, the network could represent 2N states where N is the number of nodes in the hidden layer. When the value of each node is a real number, a network can represent even more distinct states. While the potential expressive power of a network grows exponentially with the number of nodes, the complexity of both inference and training grows at most quadratically. 1.3 Are RNNs too expressive? Finite-sized RNNs with nonlinear activations are a rich family of models, capable of nearly arbitrary computation. A well-known result is that a finite-sized recurrent neural network with sigmoidal activation functions can simulate a universal Turing machine [Siegelmann and Sontag, 1991]. The capability of RNNs to perform arbitrary computation demonstrates their expressive power, but one could argue that the C programming language is equally capable of expressing arbitrary programs. And yet there are no papers claiming that the invention of C represents a panacea for machine learning. A fundamental reason is there is no simple way of efficiently exploring the space of C programs. In particular, there is no general way to calculate the gradient of an arbitrary C program to minimize a chosen loss function. Moreover, given any finite dataset, there exist countless programs which overfit the dataset, generating desired training output but failing to generalize to test examples. Why then should RNNs suffer less from similar problems? First, given any fixed architecture (set of nodes, edges, and activation functions), the recurrent neural networks with this architecture are differentiable end to end. The derivative of the loss function can be calculated with respect to each of the parameters (weights) in the model. Thus, RNNs are amenable to gradient-based training. Second, while the Turing-completeness of RNNs is an impressive property, given a fixed-size RNN with a specific architecture, it is not actually possible to reproduce any arbitrary program. Further, unlike a program composed in C, a recurrent neural network can be regularized via standard techniques that help 1 While traditional RNNs only model the dependence of the current state on the previous state, bidirectional recurrent neural networks (BRNNs) [Schuster and Paliwal, 1997] extend RNNs to model dependence on both past states and future states. 4 prevent overfitting, such as weight decay, dropout, and limiting the degrees of freedom. 1.4 Comparison to prior literature The literature on recurrent neural networks can seem impenetrable to the uninitiated. Shorter papers assume familiarity with a large body of background literature, while diagrams are frequently underspecified, failing to indicate which edges span time steps and which do not. Jargon abounds, and notation is inconsistent across papers or overloaded within one paper. Readers are frequently in the unenviable position of having to synthesize conflicting information across many papers in order to understand just one. For example, in many papers subscripts index both nodes and time steps. In others, h simultaneously stands for a link function and a layer of hidden nodes. The variable t simultaneously stands for both time indices and targets, sometimes in the same equation. Many excellent research papers have appeared recently, but clear reviews of the recurrent neural network literature are rare. Among the most useful resources are a recent book on supervised sequence labeling with recurrent neural networks [Graves, 2012] and an earlier doctoral thesis [Gers, 2001]. A recent survey covers recurrent neural nets for language modeling [De Mulder et al., 2015]. Various authors focus on specific technical aspects; for example Pearlmutter [1995] surveys gradient calculations in continuous time recurrent neural networks. In the present review paper, we aim to provide a readable, intuitive, consistently notated, and reasonably comprehensive but selective survey of research on recurrent neural networks for learning with sequences. We emphasize architectures, algorithms, and results, but we aim also to distill the intuitions that have guided this largely heuristic and empirical field. In addition to concrete modeling details, we offer qualitative arguments, a historical perspective, and comparisons to alternative methodologies where appropriate. 2 Background This section introduces formal notation and provides a brief background on neural networks in general. 2.1 Sequences The input to an RNN is a sequence, and/or its target is a sequence. An input sequence can be denoted (x(1) , x(2) , ..., x(T ) ) where each data point x(t) is a realvalued vector. Similarly, a target sequence can be denoted (y (1) , y (2) , ..., y (T ) ). A training set typically is a set of examples where each example is an (input sequence, target sequence) pair, although commonly either the input or the output may be a single data point. Sequences may be of finite or countably infinite length. When they are finite, the maximum time index of the sequence 5 is called T . RNNs are not limited to time-based sequences. They have been used successfully on non-temporal sequence data, including genetic data [Baldi and Pollastri, 2003]. However, in many important applications of RNNs, the sequences have an explicit or implicit temporal aspect. While we often refer to time in this survey, the methods described here are applicable to non-temporal as well as to temporal tasks. Using temporal terminology, an input sequence consists of data points x(t) that arrive in a discrete sequence of time steps indexed by t. A target sequence consists of data points y (t) . We use superscripts with parentheses for time, and not subscripts, to prevent confusion between sequence steps and indices of nodes in a network. When a model produces predicted data points, these are labeled ŷ (t) . The time-indexed data points may be equally spaced samples from a continuous real-world process. Examples include the still images that comprise the frames of videos or the discrete amplitudes sampled at fixed intervals that comprise audio recordings. The time steps may also be ordinal, with no exact correspondence to durations. In fact, RNNs are frequently applied to domains where sequences have a defined order but no explicit notion of time. This is the case with natural language. In the word sequence “John Coltrane plays the saxophone”, x(1) = John, x(2) = Coltrane, etc. 2.2 Neural networks Neural networks are biologically inspired models of computation. Generally, a neural network consists of a set of artificial neurons, commonly referred to as nodes or units, and a set of directed edges between them, which intuitively represent the synapses in a biological neural network. Associated with each neuron j is an activation function lj (·), which is sometimes called a link function. We use the notation lj and not hj , unlike some other papers, to distinguish the activation function from the values of the hidden nodes in a network, which, as a vector, is commonly notated h in the literature. Associated with each edge from node j 0 to j is a weight wjj 0 . Following the convention adopted in several foundational papers [Hochreiter and Schmidhuber, 1997, Gers et al., 2000, Gers, 2001, Sutskever et al., 2011], we index neurons with j and j 0 , and wjj 0 denotes the “to-from” weight corresponding to the directed edge to node j from node j 0 . It is important to note that in many references the indices are flipped and wj 0 j 6= wjj 0 denotes the “from-to” weight on the directed edge from the node j 0 to the node j, as in lecture notes by Elkan [2015] and in Wikipedia [2015]. The value vj of each neuron j is calculated by applying its activation function to a weighted sum of the values of its input nodes (Figure 1):   X vj = lj  wjj 0 · vj 0  . j0 For convenience, we term the weighted sum inside the parentheses the incoming 6 Figure 1: An artificial neuron computes a nonlinear function of a weighted sum of its inputs. activation and notate it as aj . We represent this computation in diagrams by depicting neurons as circles and edges as arrows connecting them. When appropriate, we indicate the exact activation function with a symbol, e.g., σ for sigmoid. Common choices for the activation function include the sigmoid σ(z) = 1/(1 + e−z ) and the tanh function φ(z) = (ez − e−z )/(ez + e−z ). The latter has become common in feedforward neural nets and was applied to recurrent nets by Sutskever et al. [2011]. Another activation function which has become prominent in deep learning research is the rectified linear unit (ReLU) whose formula is lj (z) = max(0, z). This type of unit has been demonstrated to improve the performance of many deep neural networks [Nair and Hinton, 2010, Maas et al., 2012, Zeiler et al., 2013] on tasks as varied as speech processing and object recognition, and has been used in recurrent neural networks by Bengio et al. [2013]. The activation function at the output nodes depends upon the task. For multiclass classification with K alternative classes, we apply a softmax nonlinearity in an output layer of K nodes. The softmax function calculates eak ŷk = PK k0 =1 eak0 for k = 1 to k = K. The denominator is a normalizing term consisting of the sum of the numerators, ensuring that the outputs of all nodes sum to one. For multilabel classification the activation function is simply a point-wise sigmoid, and for regression we typically have linear output. 7 Figure 2: A feedforward neural network. An example is presented to the network by setting the values of the blue (bottom) nodes. The values of the nodes in each layer are computed successively as a function of the prior layers until output is produced at the topmost layer. 2.3 Feedforward networks and backpropagation With a neural model of computation, one must determine the order in which computation should proceed. Should nodes be sampled one at a time and updated, or should the value of all nodes be calculated at once and then all updates applied simultaneously? Feedforward networks (Figure 2) are a restricted class of networks which deal with this problem by forbidding cycles in the directed graph of nodes. Given the absence of cycles, all nodes can be arranged into layers, and the outputs in each layer can be calculated given the outputs from the lower layers. The input x to a feedforward network is provided by setting the values of the lowest layer. Each higher layer is then successively computed until output is generated at the topmost layer ŷ. Feedforward networks are frequently used for supervised learning tasks such as classification and regression. Learning is accomplished by iteratively updating each of the weights to minimize a loss function, L(ŷ, y), which penalizes the distance between the output ŷ and the target y. The most successful algorithm for training neural networks is backpropagation, introduced for this purpose by Rumelhart et al. [1985]. Backpropagation uses the chain rule to calculate the derivative of the loss function L with respect to each parameter in the network. The weights are then adjusted by gradient descent. Because the loss surface is non-convex, there is no assurance that backpropagation will reach a global minimum. Moreover, exact optimization is known to be an NP-hard problem. However, a large body of work on heuristic 8 pre-training and optimization techniques has led to impressive empirical success on many supervised learning tasks. In particular, convolutional neural networks, popularized by Le Cun et al. [1990], are a variant of feedforward neural network that holds records since 2012 in many computer vision tasks such as object detection [Krizhevsky et al., 2012]. Nowadays, neural networks are usually trained with stochastic gradient descent (SGD) using mini-batches. With batch size equal to one, the stochastic gradient update equation is w ← w − η∇w Fi where η is the learning rate and ∇w Fi is the gradient of the objective function with respect to the parameters w as calculated on a single example (xi , yi ). Many variants of SGD are used to accelerate learning. Some popular heuristics, such as AdaGrad [Duchi et al., 2011], AdaDelta [Zeiler, 2012], and RMSprop [Tieleman and Hinton, 2012], tune the learning rate adaptively for each feature. AdaGrad, arguably the most popular, adapts the learning rate by caching the sum of squared gradients with respect to each parameter at each time step. The step size for each feature is multiplied by the inverse of the square root of this cached value. AdaGrad leads to fast convergence on convex error surfaces, but because the cached sum is monotonically increasing, the method has a monotonically decreasing learning rate, which may be undesirable on highly nonconvex loss surfaces. RMSprop modifies AdaGrad by introducing a decay factor in the cache, changing the monotonically growing value into a moving average. Momentum methods are another common SGD variant used to train neural networks. These methods add to each update a decaying sum of the previous updates. When the momentum parameter is tuned well and the network is initialized well, momentum methods can train deep nets and recurrent nets competitively with more computationally expensive methods like the Hessianfree optimizer of Sutskever et al. [2013]. To calculate the gradient in a feedforward neural network, backpropagation proceeds as follows. First, an example is propagated forward through the network to produce a value vj at each node and outputs ŷ at the topmost layer. Then, a loss function value L(ŷk , yk ) is computed at each output node k. Subsequently, for each output node k, we calculate δk = ∂L(ŷk , yk ) 0 · lk (ak ). ∂ ŷk Given these values δk , for each node in the immediately prior layer we calculate X δj = l0 (aj ) δk · wkj . k This calculation is performed successively for each lower layer to yield δj for every node j given the δ values for each node connected to j by an outgoing edge. Each value δj represents the derivative ∂L/∂aj of the total loss function with respect to that node’s incoming activation. Given the values vj calculated 9 during the forward pass, and the values δj calculated during the backward pass, the derivative of the loss L with respect a given parameter wjj 0 is ∂L = δ j vj 0 . ∂wjj 0 Other methods have been explored for learning the weights in a neural network. A number of papers from the 1990s [Belew et al., 1990, Gruau et al., 1994] championed the idea of learning neural networks with genetic algorithms, with some even claiming that achieving success on real-world problems only by applying many small changes to the weights of a network was impossible. Despite the subsequent success of backpropagation, interest in genetic algorithms continues. Several recent papers explore genetic algorithms for neural networks, especially as a means of learning the architecture of neural networks, a problem not addressed by backpropagation [Bayer et al., 2009, Harp and Samad, 2013]. By architecture we mean the number of layers, the number of nodes in each, the connectivity pattern among the layers, the choice of activation functions, etc. One open question in neural network research is how to exploit sparsity in training. In a neural network with sigmoidal or tanh activation functions, the nodes in each layer never take value exactly zero. Thus, even if the inputs are sparse, the nodes at each hidden layer are not. However, rectified linear units (ReLUs) introduce sparsity to hidden layers [Glorot et al., 2011]. In this setting, a promising path may be to store the sparsity pattern when computing each layer’s values and use it to speed up computation of the next layer in the network. Some recent work shows that given sparse inputs to a linear model with a standard regularizer, sparsity can be fully exploited even if regularization makes the gradient be not sparse [Carpenter, 2008, Langford et al., 2009, Singer and Duchi, 2009, Lipton and Elkan, 2015]. 3 Recurrent neural networks Recurrent neural networks are feedforward neural networks augmented by the inclusion of edges that span adjacent time steps, introducing a notion of time to the model. Like feedforward networks, RNNs may not have cycles among conventional edges. However, edges that connect adjacent time steps, called recurrent edges, may form cycles, including cycles of length one that are selfconnections from a node to itself across time. At time t, nodes with recurrent edges receive input from the current data point x(t) and also from hidden node values h(t−1) in the network’s previous state. The output ŷ (t) at each time t is calculated given the hidden node values h(t) at time t. Input x(t−1) at time t − 1 can influence the output ŷ (t) at time t and later by way of the recurrent connections. Two equations specify all calculations necessary for computation at each time step on the forward pass in a simple recurrent neural network as in Figure 3: h(t) = σ(W hx x(t) + W hh h(t−1) + bh ) 10 Figure 3: A simple recurrent network. At each time step t, activation is passed along solid edges as in a feedforward network. Dashed edges connect a source node at each time t to a target node at each following time t + 1. ŷ (t) = softmax(W yh h(t) + by ). Here Whx is the matrix of conventional weights between the input and the hidden layer and Whh is the matrix of recurrent weights between the hidden layer and itself at adjacent time steps. The vectors bh and by are bias parameters which allow each node to learn an offset. The dynamics of the network depicted in Figure 3 across time steps can be visualized by unfolding it as in Figure 4. Given this picture, the network can be interpreted not as cyclic, but rather as a deep network with one layer per time step and shared weights across time steps. It is then clear that the unfolded network can be trained across many time steps using backpropagation. This algorithm, called backpropagation through time (BPTT), was introduced by Werbos [1990]. All recurrent networks in common current use apply it. 3.1 Early recurrent network designs The foundational research on recurrent networks took place in the 1980s. In 1982, Hopfield introduced a family of recurrent neural networks that have pattern recognition capabilities [Hopfield, 1982]. They are defined by the values of the weights between nodes and the link functions are simple thresholding at zero. In these nets, a pattern is placed in the network by setting the values of the nodes. The network then runs for some time according to its update rules, and eventually another pattern is read out. Hopfield networks are useful for recovering a stored pattern from a corrupted version and are the forerunners of Boltzmann machines and auto-encoders. 11 Figure 4: The recurrent network of Figure 3 unfolded across time steps. An early architecture for supervised learning on sequences was introduced by Jordan [1986]. Such a network (Figure 5) is a feedforward network with a single hidden layer that is extended with special units.2 Output node values are fed to the special units, which then feed these values to the hidden nodes at the following time step. If the output values are actions, the special units allow the network to remember actions taken at previous time steps. Several modern architectures use a related form of direct transfer from output nodes; Sutskever et al. [2014] translates sentences between natural languages, and when generating a text sequence, the word chosen at each time step is fed into the network as input at the following time step. Additionally, the special units in a Jordan network are self-connected. Intuitively, these edges allow sending information across multiple time steps without perturbing the output at each intermediate time step. The architecture introduced by Elman [1990] is simpler than the earlier Jordan architecture. Associated with each unit in the hidden layer is a context unit. Each such unit j 0 takes as input the state of the corresponding hidden node j at the previous time step, along an edge of fixed weight wj 0 j = 1. This value then feeds back into the same hidden node j along a standard edge. This architecture is equivalent to a simple RNN in which each hidden node has a single self-connected recurrent edge. The idea of fixed-weight recurrent edges that make hidden nodes self-connected is fundamental in subsequent work on LSTM networks [Hochreiter and Schmidhuber, 1997]. Elman [1990] trains the network using backpropagation and demonstrates that the network can learn time dependencies. The paper features two sets of experiments. The first extends the logical operation exclusive or (XOR) to 2 Jordan [1986] calls the special units “state units” while Elman [1990] calls a corresponding structure “context units.” In this paper we simplify terminology by using only “context units”. 12 Figure 5: A recurrent neural network as proposed by Jordan [1986]. Output units are connected to special units that at the next time step feed into themselves and into hidden units. the time domain by concatenating sequences of three tokens. For each threetoken segment, e.g. “011”, the first two tokens (“01”) are chosen randomly and the third (“1”) is set by performing xor on the first two. Random guessing should achieve accuracy of 50%. A perfect system should perform the same as random for the first two tokens, but guess the third token perfectly, achieving accuracy of 66.7%. The simple network of Elman [1990] does in fact approach this maximum achievable score. 3.2 Training recurrent networks Learning with recurrent networks has long been considered to be difficult. Even for standard feedforward networks, the optimization task is NP-complete Blum and Rivest [1993]. But learning with recurrent networks can be especially challenging due to the difficulty of learning long-range dependencies, as described by Bengio et al. [1994] and expanded upon by Hochreiter et al. [2001]. The problems of vanishing and exploding gradients occur when backpropagating errors across many time steps. As a toy example, consider a network with a single input node, a single output node, and a single recurrent hidden node (Figure 7). Now consider an input passed to the network at time τ and an error calculated at time t, assuming input of zero in the intervening time steps. The tying of weights across time steps means that the recurrent edge at the hidden node j always has the same weight. Therefore, the contribution of the input at time τ to the output at time t will either explode or approach zero, exponentially fast, as t − τ grows large. Hence the derivative of the error with respect to the input will either explode or vanish. 13 Figure 6: A recurrent neural network as described by Elman [1990]. Hidden units are connected to context units, which feed back into the hidden units at the next time step. Which of the two phenomena occurs depends on whether the weight of the recurrent edge |wjj | > 1 or |wjj | < 1 and on the activation function in the hidden node (Figure 8). Given a sigmoid activation function, the vanishing gradient problem is more pressing, but with a rectified linear unit max(0, x), it is easier to imagine the exploding gradient. Pascanu et al. [2012] give a thorough mathematical treatment of the vanishing and exploding gradient problems, characterizing exact conditions under which these problems may occur. Given these conditions, they suggest an approach to training via a regularization term that forces the weights to values where the gradient neither vanishes nor explodes. Truncated backpropagation through time (TBPTT) is one solution to the exploding gradient problem for continuously running networks [Williams and Zipser, 1989]. With TBPTT, some maximum number of time steps is set along which error can be propagated. While TBPTT with a small cutoff can be used to alleviate the exploding gradient problem, it requires that one sacrifice the ability to learn long-range dependencies. The LSTM architecture described below uses carefully designed nodes with recurrent edges with fixed unit weight as a solution to the vanishing gradient problem. The issue of local optima is an obstacle to effective training that cannot be dealt with simply by modifying the network architecture. Optimizing even a single hidden-layer feedforward network is an NP-complete problem [Blum and Rivest, 1993]. However, recent empirical and theoretical studies suggest that in practice, the issue may not be as important as once thought. Dauphin et al. [2014] show that while many critical points exist on the error surfaces of large neural networks, the ratio of saddle points to true local minima increases exponentially with the size of the network, and algorithms can be designed to 14 Figure 7: A simple recurrent net with one input unit, one output unit, and one recurrent hidden unit. Figure 8: A visualization of the vanishing gradient problem, using the network depicted in Figure 7, adapted from Graves [2012]. If the weight along the recurrent edge is less than one, the contribution of the input at the first time step to the output at the final time step will decrease exponentially fast as a function of the length of the time interval in between. 15 escape from saddle points. Overall, along with the improved architectures explained below, fast implementations and better gradient-following heuristics have rendered RNN training feasible. Implementations of forward and backward propagation using GPUs, such as the Theano [Bergstra et al., 2010] and Torch [Collobert et al., 2011] packages, have made it straightforward to implement fast training algorithms. In 1996, prior to the introduction of the LSTM, attempts to train recurrent nets to bridge long time gaps were shown to perform no better than random guessing [Hochreiter and Schmidhuber, 1996]. However, RNNs are now frequently trained successfully. For some tasks, freely available software can be run on a single GPU and produce compelling results in hours [Karpathy, 2015]. Martens and Sutskever [2011] reported success training recurrent neural networks with a Hessian-free truncated Newton approach, and applied the method to a network which learns to generate text one character at a time in [Sutskever et al., 2011]. In the paper that describes the abundance of saddle points on the error surfaces of neural networks [Dauphin et al., 2014], the authors present a saddle-free version of Newton’s method. Unlike Newton’s method, which is attracted to critical points, including saddle points, this variant is specially designed to escape from them. Experimental results include a demonstration of improved performance on recurrent networks. Newton’s method requires computing the Hessian, which is prohibitively expensive for large networks, scaling quadratically with the number of parameters. While their algorithm only approximates the Hessian, it is still computationally expensive compared to SGD. Thus the authors describe a hybrid approach in which the saddle-free Newton method is applied only in places where SGD appears to be stuck. 4 Modern RNN architectures The most successful RNN architectures for sequence learning stem from two papers published in 1997. The first paper, Long Short-Term Memory by Hochreiter and Schmidhuber [1997], introduces the memory cell, a unit of computation that replaces traditional nodes in the hidden layer of a network. With these memory cells, networks are able to overcome difficulties with training encountered by earlier recurrent networks. The second paper, Bidirectional Recurrent Neural Networks by Schuster and Paliwal [1997], introduces an architecture in which information from both the future and the past are used to determine the output at any point in the sequence. This is in contrast to previous networks, in which only past input can affect the output, and has been used successfully for sequence labeling tasks in natural language processing, among others. Fortunately, the two innovations are not mutually exclusive, and have been successfully combined for phoneme classification [Graves and Schmidhuber, 2005] and handwriting recognition [Graves et al., 2009]. In this section we explain the LSTM and BRNN and we describe the neural Turing machine (NTM), which extends RNNs with an addressable external memory [Graves et al., 2014]. 16 Figure 9: One LSTM memory cell as proposed by Hochreiter and Schmidhuber [1997]. The self-connected node is the internal state s. The diagonal line indicates that it is linear, i.e. the identity link function is applied. The blue dashed line is the recurrent edge, which has fixed unit weight. Nodes marked Π output the product of their inputs. All edges into and from Π nodes also have fixed unit weight. 4.1 Long short-term memory (LSTM) Hochreiter and Schmidhuber [1997] introduced the LSTM model primarily in order to overcome the problem of vanishing gradients. This model resembles a standard recurrent neural network with a hidden layer, but each ordinary node (Figure 1) in the hidden layer is replaced by a memory cell (Figure 9). Each memory cell contains a node with a self-connected recurrent edge of fixed weight one, ensuring that the gradient can pass across many time steps without vanishing or exploding. To distinguish references to a memory cell and not an ordinary node, we use the subscript c. The term “long short-term memory” comes from the following intuition. Simple recurrent neural networks have long-term memory in the form of weights. The weights change slowly during training, encoding general knowledge about the data. They also have short-term memory in the form of ephemeral activations, which pass from each node to successive nodes. The LSTM model introduces an intermediate type of storage via the memory cell. A memory cell is a composite unit, built from simpler nodes in a specific connectivity pattern, with the novel inclusion of multiplicative nodes, represented in diagrams by the letter Π. All elements of the LSTM cell are enumerated and described below. Note that when we use vector notation, we are referring to the values of the 17 nodes in an entire layer of cells. For example, s is a vector containing the value of sc at each memory cell c in a layer. When the subscript c is used, it is to index an individual memory cell. • Input node: This unit, labeled gc , is a node that takes activation in the standard way from the input layer x(t) at the current time step and (along recurrent edges) from the hidden layer at the previous time step h(t−1) . Typically, the summed weighted input is run through a tanh activation function, although in the original LSTM paper, the activation function is a sigmoid. • Input gate: Gates are a distinctive feature of the LSTM approach. A gate is a sigmoidal unit that, like the input node, takes activation from the current data point x(t) as well as from the hidden layer at the previous time step. A gate is so-called because its value is used to multiply the value of another node. It is a gate in the sense that if its value is zero, then flow from the other node is cut off. If the value of the gate is one, all flow is passed through. The value of the input gate ic multiplies the value of the input node. • Internal state: At the heart of each memory cell is a node sc with linear activation, which is referred to in the original paper as the “internal state” of the cell. The internal state sc has a self-connected recurrent edge with fixed unit weight. Because this edge spans adjacent time steps with constant weight, error can flow across time steps without vanishing or exploding. This edge is often called the constant error carousel. In vector notation, the update for the internal state is s(t) = g (t) i(t) + s(t−1) where is pointwise multiplication. • Forget gate: These gates fc were introduced by Gers et al. [2000]. They provide a method by which the network can learn to flush the contents of the internal state. This is especially useful in continuously running networks. With forget gates, the equation to calculate the internal state on the forward pass is s(t) = g (t) i(t) + f (t) s(t−1) . • Output gate: The value vc ultimately produced by a memory cell is the value of the internal state sc multiplied by the value of the output gate oc . It is customary that the internal state first be run through a tanh activation function, as this gives the output of each cell the same dynamic range as an ordinary tanh hidden unit. However, in other neural network research, rectified linear units, which have a greater dynamic range, are easier to train. Thus it seems plausible that the nonlinear function on the internal state might be omitted. In the original paper and in most subsequent work, the input node is labeled g. We adhere to this convention but note that it may be confusing as g does 18 Figure 10: LSTM memory cell with a forget gate as described by Gers et al. [2000]. not stand for gate. In the original paper, the gates are called yin and yout but this is confusing because y generally stands for output in the machine learning literature. Seeking comprehensibility, we break with this convention and use i, f , and o to refer to input, forget and output gates respectively, as in Sutskever et al. [2014]. Since the original LSTM was introduced, several variations have been proposed. Forget gates, described above, were proposed in 2000 and were not part of the original LSTM design. However, they have proven effective and are standard in most modern implementations. That same year, Gers and Schmidhuber [2000] proposed peephole connections that pass from the internal state directly to the input and output gates of that same node without first having to be modulated by the output gate. They report that these connections improve performance on timing tasks where the network must learn to measure precise intervals between events. The intuition of the peephole connection can be captured by the following example. Consider a network which must learn to count objects and emit some desired output when n objects have been seen. The network might learn to let some fixed amount of activation into the internal state after each object is seen. This activation is trapped in the internal state sc by the constant error carousel, and is incremented iteratively each time another object is seen. When the nth object is seen, the network needs to know to let out content from the internal state so that it can affect the output. To accomplish this, the output gate oc must know the content of the internal state sc . Thus sc should be an input to oc . Put formally, computation in the LSTM model proceeds according to the 19 Figure 11: A recurrent neural network with a hidden layer consisting of two memory cells. The network is shown unfolded across two time steps. following calculations, which are performed at each time step. These equations give the full algorithm for a modern LSTM with forget gates: g (t) = φ(W gx x(t) + W gh h(t−1) + bg ) i(t) = σ(W ix x(t) + W ih h(t−1) + bi ) f (t) = σ(W fx x(t) + W fh h(t−1) + bf ) o(t) = σ(W ox x(t) + W oh h(t−1) + bo ) s(t) = g (t) i(i) + s(t−1) h(t) = φ(s(t) ) f (t) o(t) . The value of the hidden layer of the LSTM at time t is the vector h(t) , while h(t−1) is the values output by each memory cell in the hidden layer at the previous time. Note that these equations include the forget gate, but not peephole connections. The calculations for the simpler LSTM without forget gates are obtained by setting f (t) = 1 for all t. We use the tanh function φ for the input node g following the state-of-the-art design of Zaremba and Sutskever [2014]. However, in the original LSTM paper, the activation function for g is the sigmoid σ. Intuitively, in terms of the forward pass, the LSTM can learn when to let activation into the internal state. As long as the input gate takes value zero, no activation can get in. Similarly, the output gate learns when to let the 20 Figure 12: A bidirectional recurrent neural network as described by Schuster and Paliwal [1997], unfolded in time. value out. When both gates are closed, the activation is trapped in the memory cell, neither growing nor shrinking, nor affecting the output at intermediate time steps. In terms of the backwards pass, the constant error carousel enables the gradient to propagate back across many time steps, neither exploding nor vanishing. In this sense, the gates are learning when to let error in, and when to let it out. In practice, the LSTM has shown a superior ability to learn longrange dependencies as compared to simple RNNs. Consequently, the majority of state-of-the-art application papers covered in this review use the LSTM model. One frequent point of confusion is the manner in which multiple memory cells are used together to comprise the hidden layer of a working neural network. To alleviate this confusion, we depict in Figure 11 a simple network with two memory cells, analogous to Figure 4. The output from each memory cell flows in the subsequent time step to the input node and all gates of each memory cell. It is common to include multiple layers of memory cells [Sutskever et al., 2014]. Typically, in these architectures each layer takes input from the layer below at the same time step and from the same layer in the previous time step. 4.2 Bidirectional recurrent neural networks (BRNNs) Along with the LSTM, one of the most used RNN architectures is the bidirectional recurrent neural network (BRNN) (Figure 12) first described by Schuster and Paliwal [1997]. In this architecture, there are two layers of hidden nodes. Both hidden layers are connected to input and output. The two hidden layers are differentiated in that the first has recurrent connections from the past time 21 steps while in the second the direction of recurrent of connections is flipped, passing activation backwards along the sequence. Given an input sequence and a target sequence, the BRNN can be trained by ordinary backpropagation after unfolding across time. The following three equations describe a BRNN: h(t) = σ(W hx x(t) + W hh h(t−1) + bh ) z (t) = σ(W zx x(t) + W zz z (t+1) + bz ) ŷ (t) = softmax(W yh h(t) + W yz z (t) + by ) where h(t) and z (t) are the values of the hidden layers in the forwards and backwards directions respectively. One limitation of the BRNN is that cannot run continuously, as it requires a fixed endpoint in both the future and in the past. Further, it is not an appropriate machine learning algorithm for the online setting, as it is implausible to receive information from the future, i.e., to know sequence elements that have not been observed. But for prediction over a sequence of fixed length, it is often sensible to take into account both past and future sequence elements. Consider the natural language task of part-of-speech tagging. Given any word in a sentence, information about both the words which precede and those which follow it is useful for predicting that word’s part-of-speech. The LSTM and BRNN are in fact compatible ideas. The former introduces a new basic unit from which to compose a hidden layer, while the latter concerns the wiring of the hidden layers, regardless of what nodes they contain. Such an approach, termed a BLSTM has been used to achieve state of the art results on handwriting recognition and phoneme classification [Graves and Schmidhuber, 2005, Graves et al., 2009]. 4.3 Neural Turing machines The neural Turing machine (NTM) extends recurrent neural networks with an addressable external memory [Graves et al., 2014]. This work improves upon the ability of RNNs to perform complex algorithmic tasks such as sorting. The authors take inspiration from theories in cognitive science, which suggest humans possess a “central executive” that interacts with a memory buffer [Baddeley et al., 1996]. By analogy to a Turing machine, in which a program directs read heads and write heads to interact with external memory in the form of a tape, the model is named a Neural Turing Machine. While technical details of the read/write heads are beyond the scope of this review, we aim to convey a high-level sense of the model and its applications. The two primary components of an NTM are a controller and memory matrix. The controller, which may be a recurrent or feedforward neural network, takes input and returns output to the outside world, as well as passing instructions to and reading from the memory. The memory is represented by a large matrix of N memory locations, each of which is a vector of dimension M . Additionally, a number of read and write heads facilitate the interaction between 22 the controller and the memory matrix. Despite these additional capabilities, the NTM is differentiable end-to-end and can be trained by variants of stochastic gradient descent using BPTT. Graves et al. [2014] select five algorithmic tasks to test the performance of the NTM model. By algorithmic we mean that for each task, the target output for a given input can be calculated by following a simple program, as might be easily implemented in any universal programming language. One example is the copy task, where the input is a sequence of fixed length binary vectors followed by a delimiter symbol. The target output is a copy of the input sequence. In another task, priority sort, an input consists of a sequence of binary vectors together with a distinct scalar priority value for each vector. The target output is the sequence of vectors sorted by priority. The experiments test whether an NTM can be trained via supervised learning to implement these common algorithms correctly and efficiently. Interestingly, solutions found in this way generalize reasonably well to inputs longer than those presented in the training set. In contrast, the LSTM without external memory does not generalize well to longer inputs. The authors compare three different architectures, namely an LSTM RNN, an NTM with a feedforward controller, and an NTM with an LSTM controller. On each task, both NTM architectures significantly outperform the LSTM RNN both in training set performance and in generalization to test data. 5 Applications of LSTMs and BRNNs The previous sections introduced the building blocks from which nearly all stateof-the-art recurrent neural networks are composed. This section looks at several application areas where recurrent networks have been employed successfully. Before describing state of the art results in detail, it is appropriate to convey a concrete sense of the precise architectures with which many important tasks can be expressed clearly as sequence learning problems with recurrent neural networks. Figure 13 demonstrates several common RNN architectures and associates each with corresponding well-documented tasks. In the following subsections, we first introduce the representations of natural language used for input and output to recurrent neural networks and the commonly used performance metrics for sequence prediction tasks. Then we survey state-of-the-art results in machine translation, image captioning, video captioning, and handwriting recognition. Many applications of RNNs involve processing written language. Some applications, such as image captioning, involve generating strings of text. Others, such as machine translation and dialogue systems, require both inputting and outputting text. Systems which output text are more difficult to evaluate empirically than those which produce binary predictions or numerical output. As a result several methods have been developed to assess the quality of translations and captions. In the next subsection, we provide the background necessary to understand how text is represented in most modern recurrent net applications. We then explain the commonly reported evaluation metrics. 23 Figure 13: Recurrent neural networks have been used successfully to model both sequential inputs and sequential outputs as well as mappings between single data points and sequences (in both directions). This figure, based on a similar figure in Karpathy [2015] shows how numerous tasks can be modeled with RNNs with sequential inputs and/or sequential outputs. In each subfigure, blue rectangles correspond to inputs, red rectangles to outputs and green rectangles to the entire hidden state of the neural network. (a) This is the conventional independent case, as assumed by standard feedforward networks. (b) Text and video classification are tasks in which a sequence is mapped to one fixed length vector. (c) Image captioning presents the converse case, where the input image is a single non-sequential data point. (d) This architecture has been used for natural language translation, a sequence-to-sequence task in which the two sequences may have varying and different lengths. (e) This architecture has been used to learn a generative model for text, predicting at each step the following character. 24 5.1 Representations of natural language inputs and outputs When words are output at each time step, generally the output consists of a softmax vector y (t) ∈ K where K is the size of the vocabulary. A softmax layer is an element-wise logistic function that is normalized so that all of its components sum to one. Intuitively, these outputs correspond to the probabilities that each word is the correct output at that time step. For application where an input consists of a sequence of words, typically the words are fed to the network one at a time in consecutive time steps. In these cases, the simplest way to represent words is a one-hot encoding, using binary vectors with a length equal to the size of the vocabulary, so “1000” and “0100” would represent the first and second words in the vocabulary respectively. Such an encoding is discussed by Elman [1990] among others. However, this encoding is inefficient, requiring as many bits as the vocabulary is large. Further, it offers no direct way to capture different aspects of similarity between words in the encoding itself. Thus it is common now to model words with a distributed representation using a meaning vector. In some cases, these meanings for words are learned given a large corpus of supervised data, but it is more usual to initialize the meaning vectors using an embedding based on word co-occurrence statistics. Freely available code to produce word vectors from these statistics include GloVe [Pennington et al., 2014], and word2vec [Goldberg and Levy, 2014], which implements a word embedding algorithm from Mikolov et al. [2013]. Distributed representations for textual data were described by Hinton [1986], used extensively for natural language by Bengio et al. [2003], and more recently brought to wider attention in the deep learning community in a number of papers describing recursive auto-encoder (RAE) networks [Socher et al., 2010, 2011a,b,c]. For clarity we point out that these recursive networks are not recurrent neural networks, and in the present survey the abbreviation RNN always means recurrent neural network. While they are distinct approaches, recurrent and recursive neural networks have important features in common, namely that they both involve extensive weight tying and are both trained end-to-end via backpropagation. In many experiments with recurrent neural networks [Elman, 1990, Sutskever et al., 2011, Zaremba and Sutskever, 2014], input is fed in one character at a time, and output generated one character at a time, as opposed to one word at a time. While the output is nearly always a softmax layer, many papers omit details of how they represent single-character inputs. It seems reasonable to infer that characters are encoded with a one-hot encoding. We know of no cases of paper using a distributed representation at the single-character level. R 5.2 Evaluation methodology A serious obstacle to training systems well to output variable length sequences of words is the flaws of the available performance metrics. In the case of captioning or translation, there maybe be multiple correct translations. Further, 25 a labeled dataset may contain multiple reference translations for each example. Comparing against such a gold standard is more fraught than applying standard performance measure to binary classification problems. One commonly used metric for structured natural language output with multiple references is BLEU score. Developed in 2002, BLEU score is related to modified unigram precision [Papineni et al., 2002]. It is the geometric mean of the n-gram precisions for all values of n between 1 and some upper limit N . In practice, 4 is a typical value for N , shown to maximize agreement with human raters. Because precision can be made high by offering excessively short translations, the BLEU score includes a brevity penalty B. Where c is average the length of the candidate translations and r the average length of the reference translations, the brevity penalty is ( 1 if c > r B= . (1−r/c) e if c ≤ r Then the BLEU score is BLEU = B · exp N 1 X log pn N n=1 ! where pn is the modified n-gram precision, which is the number of n-grams in the candidate translation that occur in any of the reference translations, divided by the total number of n-grams in the candidate translation. This is called modified precision because it is an adaptation of precision to the case of multiple references. BLEU scores are commonly used in recent papers to evaluate both translation and captioning systems. While BLEU score does appear highly correlated with human judgments, there is no guarantee that any given translation with a higher BLEU score is superior to another which receives a lower BLEU score. In fact, while BLEU scores tend to be correlated with human judgement across large sets of translations, they are not accurate predictors of human judgement at the single sentence level. METEOR is an alternative metric intended to overcome the weaknesses of the BLEU score [Banerjee and Lavie, 2005]. METEOR is based on explicit word to word matches between candidates and reference sentences. When multiple references exist, the best score is used. Unlike BLEU, METEOR exploits known synonyms and stemming. The first step is to compute an F-score Fα = P ·R α · P + (1 − α) · R based on single word matches where P is the precision and R is the recall. The next step is to calculate a fragmentation penalty M ∝ c/m where c is the smallest number of chunks of consecutive words such that the words are adjacent in both the candidate and the reference, and m is the total number of matched unigrams yielding the score. Finally, the score is METEOR = (1 − M ) · Fα . 26 Empirically, this metric has been found to agree with human raters more than BLEU score. However, METEOR is less straightforward to calculate than BLEU. To replicate the METEOR score reported by another party, one must exactly replicate their stemming and synonym matching, as well as the calculations. Both metrics rely upon having the exact same set of reference translations. Even in the straightforward case of binary classification, without sequential dependencies, commonly used performance metrics like F1 give rise to optimal thresholding strategies which may not accord with intuition about what should constitute good performance [Lipton et al., 2014]. Along the same lines, given that performance metrics such as the ones above are weak proxies for true objectives, it may be difficult to distinguish between systems which are truly stronger and those which most overfit the performance metrics in use. 5.3 Natural language translation Translation of text is a fundamental problem in machine learning that resists solutions with shallow methods. Some tasks, like document classification, can be performed successfully with a bag-of-words representation that ignores word order. But word order is essential in translation. The sentences “Scientist killed by raging virus” and “Virus killed by raging scientist” have identical bag-ofwords representations. Sutskever et al. [2014] present a translation model using two multilayered LSTMs that demonstrates impressive performance translating from English to French. The first LSTM is used for encoding an input phrase from the source language and the second LSTM for decoding the output phrase in the target language. The model works according to the following procedure (Figure 14): • The source phrase is fed to the encoding LSTM one word at a time, which does not output anything. The authors found that significantly better results are achieved when the input sentence is fed into the network in reverse order. • When the end of the phrase is reached, a special symbol that indicates the beginning of the output sentence is sent to the decoding LSTM. Additionally, the decoding LSTM receives as input the final state of the first LSTM. The second LSTM outputs softmax probabilities over the vocabulary at each time step. • At inference time, beam search is used to choose the most likely words from the distribution at each time step, running the second LSTM until the end-of-sentence (EOS ) token is reached. For training, the true inputs are fed to the encoder, the true translation is fed to the decoder, and loss is propagated back from the outputs of the decoder across the entire sequence to sequence model. The network is trained to maximize the likelihood of the correct translation of each sentence in the training set. At inference time, a left to right beam search is used to determine 27 Figure 14: Sequence to sequence LSTM model of Sutskever et al. [2014]. The network consists of an encoding model (first LSTM) and a decoding model (second LSTM). The input blocks (blue and purple) correspond to word vectors, which are fully connected to the corresponding hidden state. Red nodes are softmax outputs. Weights are tied among all encoding steps and among all decoding time steps. 28 which words to output. A few among the most likely next words are chosen for expansion after each time step. The beam search ends when the network outputs an end-of-sentence (EOS ) token. Sutskever et al. [2014] train the model using stochastic gradient descent without momentum, halving the learning rate twice per epoch, after the first five epochs. The approach achieves a BLEU score of 34.81, outperforming the best previous neural network NLP systems, and matching the best published results for non-neural network approaches, including systems that have explicitly programmed domain expertise. When their system is used to rerank candidate translations from another system, it achieves a BLEU score of 36.5. The implementation which achieved these results involved eight GPUS. Nevertheless, training took 10 days to complete. One GPU was assigned to each layer of the LSTM, and an additional four GPUs were used simply to calculate softmax. The implementation was coded in C++, and each hidden layer of the LSTM contained 1000 nodes. The input vocabulary contained 160,000 words and the output vocabulary contained 80,000 words. Weights were initialized uniformly randomly in the range between −0.08 and 0.08. Another RNN approach to language translation is presented by Auli et al. [2013]. Their RNN model uses the word embeddings of Mikolov and a lattice representation of the decoder output to facilitate search over the space of possible translations. In the lattice, each node corresponds to a sequence of words. They report a BLEU score of 28.5 on French-English translation tasks. Both papers provide results on similar datasets but Sutskever et al. [2014] only report on English to French translation while Auli et al. [2013] only report on French to English translation, so it is not possible to compare the performance of the two models. 5.4 Image captioning Recently, recurrent neural networks have been used successfully for generating sentences that describe photographs [Vinyals et al., 2015, Karpathy and Fei-Fei, 2014, Mao et al., 2014]. In this task, a training set consists of input images x and target captions y. Given a large set of image-caption pairs, a model is trained to predict the appropriate caption for an image. Vinyals et al. [2015] follow up on the success in language to language translation by considering captioning as a case of image to language translation. Instead of both encoding and decoding with LSTMs, they introduce the idea of encoding an image with a convolutional neural network, and then decoding it with an LSTM. Mao et al. [2014] independently developed a similar RNN image captioning network, and achieved then state-of-the-art results on the Pascal, Flickr30K, and COCO datasets. Karpathy and Fei-Fei [2014] follows on this work, using a convolutional network to encode images together with a bidirectional network attention mechanism and standard RNN to decode captions, using word2vec embeddings as word representations. They consider both full-image captioning and a model that captures correspondences between image regions and text snippets. At 29 inference time, their procedure resembles the one described by Sutskever et al. [2014], where sentences are decoded one word at a time. The most probable word is chosen and fed to the network at the next time step. This process is repeated until an EOS token is produced. To convey a sense of the scale of these problems, Karpathy and Fei-Fei [2014] focus on three datasets of captioned images: Flickr8K, Flickr30K, and COCO, of size 50MB (8000 images), 200MB (30,000 images), and 750MB (328,000 images) respectively. The implementation uses the Caffe library [Jia et al., 2014], and the convolutional network is pretrained on ImageNet data. In a revised version, the authors report that LSTMs outperform simpler RNNs and that learning word representations from random initializations is often preferable to word2vec embeddings. As an explanation, they say that word2vec embeddings may cluster words like colors together in the embedding space, which can be not suitable for visual descriptions of images. 5.5 Further applications Handwriting recognition is an application area where bidirectional LSTMs have been used to achieve state of the art results. In work by Liwicki et al. [2007] and Graves et al. [2009], data is collected from an interactive whiteboard. Sensors record the (x, y) coordinates of the pen at regularly sampled time steps. In the more recent paper, they use a bidirectional LSTM model, outperforming an HMM model by achieving 81.5% word-level accuracy, compared to 70.1% for the HMM. In the last year, a number of papers have emerged that extend the success of recurrent networks for translation and image captioning to new domains. Among the most interesting of these applications are unsupervised video encoding [Srivastava et al., 2015], video captioning [Venugopalan et al., 2015], and program execution [Zaremba and Sutskever, 2014]. Venugopalan et al. [2015] demonstrate a sequence to sequence architecture that encodes frames from a video and decode words. At each time step the input to the encoding LSTM is the topmost hidden layer of a convolutional neural network. At decoding time, the network outputs probabilities over the vocabulary at each time step. Zaremba and Sutskever [2014] experiment with networks which read computer programs one character at a time and predict their output. They focus on programs which output integers and find that for simple programs, including adding two nine-digit numbers, their network, which uses LSTM cells in several stacked hidden layers and makes a single left to right pass through the program, can predict the output with 99% accuracy. 6 Discussion Over the past thirty years, recurrent neural networks have gone from models primarily of interest for cognitive modeling and computational neuroscience, to 30 powerful and practical tools for large-scale supervised learning from sequences. This progress owes to advances in model architectures, training algorithms, and parallel computing. Recurrent networks are especially interesting because they overcome many of the restrictions placed on input and output data by traditional machine learning approaches. With recurrent networks, the assumption of independence between consecutive examples is broken, and hence also the assumption of fixed-dimension inputs and outputs. While LSTMs and BRNNs have set records in accuracy on many tasks in recent years, it is noteworthy that advances come from novel architectures rather than from fundamentally novel algorithms. Therefore, automating exploration of the space of possible models, for example via genetic algorithms or a Markov chain Monte Carlo approach, could be promising. Neural networks offer a wide range of transferable and combinable techniques. New activation functions, training procedures, initializations procedures, etc. are generally transferable across networks and tasks, often conferring similar benefits. As the number of such techniques grows, the practicality of testing all combinations diminishes. It seems reasonable to infer that as a community, neural network researchers are exploring the space of model architectures and configurations much as a genetic algorithm might, mixing and matching techniques, with a fitness function in the form of evaluation metrics on major datasets of interest. This inference suggests two corollaries. First, as just stated, this body of research could benefit from automated procedures to explore the space of models. Second, as we build systems designed to perform more complex tasks, we would benefit from improved fitness functions. A BLEU score inspires less confidence than the accuracy reported on a binary classification task. To this end, when possible, it seems prudent to individually test techniques first with classic feedforward networks on datasets with established benchmarks before applying them to recurrent networks in settings with less reliable evaluation criteria. Lastly, the rapid success of recurrent neural networks on natural language tasks leads us to believe that extensions of this work to longer texts would be fruitful. Additionally, we imagine that dialogue systems could be built along similar principles to the architectures used for translation, encoding prompts and generating responses, while retaining the entirety of conversation history as contextual information. 7 Acknowledgements The first author’s research is funded by generous support from the Division of Biomedical Informatics at UCSD, via a training grant from the National Library of Medicine. This review has benefited from insightful comments from Vineet Bafna, Julian McAuley, Balakrishnan Narayanaswamy, Stefanos Poulis, Lawrence Saul, Zhuowen Tu, and Sharad Vikram. 31 References Michael Auli, Michel Galley, Chris Quirk, and Geoffrey Zweig. Joint language and translation modeling with recurrent neural networks. In EMNLP, pages 1044–1054, 2013. Alan Baddeley, Sergio Della Sala, and T.W. Robbins. Working memory and executive control [and discussion]. Philosophical Transactions of the Royal Society B: Biological Sciences, 351(1346):1397–1404, 1996. Pierre Baldi and Gianluca Pollastri. The principled design of large-scale recursive neural network architectures–DAG-RNNs and the protein structure prediction problem. The Journal of Machine Learning Research, 4:575–602, 2003. Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, 2005. Justin Bayer, Daan Wierstra, Julian Togelius, and Jürgen Schmidhuber. Evolving memory cell structures for sequence learning. In Artificial Neural Networks–ICANN 2009, pages 755–764. Springer, 2009. Richard K. Belew, John McInerney, and Nicol N. Schraudolph. Evolving networks: Using the genetic algorithm with connectionist learning. In In. Citeseer, 1990. Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994. Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Janvin. A neural probabilistic language model. The Journal of Machine Learning Research, 3:1137–1155, 2003. Yoshua Bengio, Nicolas Boulanger-Lewandowski, and Razvan Pascanu. Advances in optimizing recurrent networks. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on, pages 8624– 8628. IEEE, 2013. James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), volume 4, page 3. Austin, TX, 2010. Avrim L. Blum and Ronald L. Rivest. Training a 3-node neural network is NPcomplete. In Machine Learning: From Theory to Applications, pages 9–28. Springer, 1993. 32 Bob Carpenter. Lazy sparse stochastic gradient descent for regularized multinomial logistic regression. Alias-i, Inc., Tech. Rep, pages 1–20, 2008. Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A matlablike environment for machine learning. In BigLearn, NIPS Workshop, 2011. Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In Advances in Neural Information Processing Systems, pages 2933–2941, 2014. Wim De Mulder, Steven Bethard, and Marie-Francine Moens. A survey on the application of recurrent neural networks to statistical language modeling. Computer Speech & Language, 30(1):61–98, 2015. John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011. Charles Elkan. Learning meanings for sentences. http://cseweb.ucsd.edu/ ~elkan/250B/learningmeaning.pdf, 2015. Accessed: 2015-05-18. Jeffrey L. Elman. Finding structure in time. Cognitive science, 14(2):179–211, 1990. Clement Farabet, Camille Couprie, Laurent Najman, and Yann LeCun. Learning hierarchical features for scene labeling. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 35(8):1915–1929, 2013. Felix A. Gers. Long short-term memory in recurrent neural networks. Unpublished PhD dissertation, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, 2001. Felix A. Gers and Jürgen Schmidhuber. Recurrent nets that time and count. In Neural Networks, 2000. IJCNN 2000, Proceedings of the IEEE-INNS-ENNS International Joint Conference on, volume 3, pages 189–194. IEEE, 2000. Felix A. Gers, Jürgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with LSTM. Neural computation, 12(10):2451–2471, 2000. Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier networks. In Proceedings of the 14th International Conference on Artificial Intelligence and Statistics. JMLR W&CP Volume, volume 15, pages 315–323, 2011. Yoav Goldberg and Omer Levy. word2vec explained: deriving Mikolov et al.’s negative-sampling word-embedding method. arXiv preprint arXiv:1402.3722, 2014. 33 Alex Graves. Supervised sequence labelling with recurrent neural networks, volume 385. Springer, 2012. Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 18(5):602–610, 2005. Alex Graves, Marcus Liwicki, Santiago Fernández, Roman Bertolami, Horst Bunke, and Jürgen Schmidhuber. A novel connectionist system for unconstrained handwriting recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 31(5):855–868, 2009. Alex Graves, Greg Wayne, and Ivo Danihelka. Neural Turing machines. arXiv preprint arXiv:1410.5401, 2014. Frdric Gruau, L’universite Claude Bernard lyon I, Of A Diplome De Doctorat, M. Jacques Demongeot, Examinators M. Michel Cosnard, M. Jacques Mazoyer, M. Pierre Peretto, and M. Darell Whitley. Neural network synthesis using cellular encoding and the genetic algorithm., 1994. Steven A. Harp and Tariq Samad. Optimizing neural networks with genetic algorithms. In Proceedings of the 54th American Power Conference, Chicago, volume 2, 2013. Geoffrey E. Hinton. Learning distributed representations of concepts, 1986. Sepp Hochreiter and Jurgen Schmidhuber. Bridging long time lags by weight guessing and “long short-term memory”. Spatiotemporal Models in Biological and Artificial Systems, 37:65–72, 1996. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies. A field guide to dynamical recurrent neural networks, 2001. John J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79 (8):2554–2558, 1982. Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014. Michael I. Jordan. Serial order: A parallel distributed processing approach. Technical Report 8604, Institute for Cognitive Science, University of California, San Diego, 1986. 34 Andrej Karpathy. The unreasonable effectiveness of recurrent neural networks. http://karpathy.github.io/2015/05/21/rnn-effectiveness/, 2015. Accessed: 2015-08-13. Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. arXiv preprint arXiv:1412.2306, 2014. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 1097–1105, 2012. John Langford, Lihong Li, and Tong Zhang. Sparse online learning via truncated gradient. In Advances in Neural Information Processing Systems, pages 905– 912, 2009. Yann Le Cun, B. Boser, John S. Denker, D. Henderson, Richard E. Howard, W. Hubbard, and Lawrence D. Jackel. Handwritten digit recognition with a back-propagation network. In Advances in Neural Information Processing Systems. Citeseer, 1990. Zachary C. Lipton and Charles Elkan. Efficient elastic net regularization for sparse linear models. CoRR, abs/1505.06449, 2015. URL http://arxiv. org/abs/1505.06449. Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy. Optimal thresholding of classifiers to maximize F1 measure. In Machine Learning and Knowledge Discovery in Databases, pages 225–239. Springer, 2014. Marcus Liwicki, Alex Graves, Horst Bunke, and Jürgen Schmidhuber. A novel approach to on-line handwriting recognition based on bidirectional long shortterm memory networks. In Proc. 9th Int. Conf. on Document Analysis and Recognition, volume 1, pages 367–371, 2007. Andrew L. Maas, Quoc V. Le, Tyler M. O’Neil, Oriol Vinyals, Patrick Nguyen, and Andrew Y. Ng. Recurrent neural networks for noise reduction in robust ASR. In INTERSPEECH. Citeseer, 2012. Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan Yuille. Deep captioning with multimodal recurrent neural networks (m-RNN). arXiv preprint arXiv:1412.6632, 2014. James Martens and Ilya Sutskever. Learning recurrent neural networks with Hessian-free optimization. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 1033–1040, 2011. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. Vinod Nair and Geoffrey E. Hinton. Rectified linear units improve restricted Boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages 807–814, 2010. 35 Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pages 311–318. Association for Computational Linguistics, 2002. Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. arXiv preprint arXiv:1211.5063, 2012. Barak A. Pearlmutter. Gradient calculations for dynamic recurrent neural networks: A survey. Neural Networks, IEEE Transactions on, 6(5):1212–1228, 1995. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. Proceedings of the Empirical Methods in Natural Language Processing (EMNLP 2014), 12, 2014. David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning internal representations by error propagation. Technical report, DTIC Document, 1985. Mike Schuster and Kuldip K. Paliwal. Bidirectional recurrent neural networks. Signal Processing, IEEE Transactions on, 45(11):2673–2681, 1997. Hava T. Siegelmann and Eduardo D. Sontag. Turing computability with neural nets. Applied Mathematics Letters, 4(6):77–80, 1991. Yoram Singer and John C. Duchi. Efficient learning using forward-backward splitting. In Advances in Neural Information Processing Systems, pages 495– 503, 2009. Richard Socher, Christopher D. Manning, and Andrew Y. Ng. Learning continuous phrase representations and syntactic parsing with recursive neural networks. In Proceedings of the NIPS-2010 Deep Learning and Unsupervised Feature Learning Workshop, pages 1–9, 2010. Richard Socher, Eric H. Huang, Jeffrey Pennin, Christopher D. Manning, and Andrew Y. Ng. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In Advances in Neural Information Processing Systems, pages 801–809, 2011a. Richard Socher, Cliff C. Lin, Chris Manning, and Andrew Y. Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 129–136, 2011b. Richard Socher, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 151–161. Association for Computational Linguistics, 2011c. 36 Richard Socher, Andrej Karpathy, Quoc V. Le, Christopher D. Manning, and Andrew Y. Ng. Grounded compositional semantics for finding and describing images with sentences. Transactions of the Association for Computational Linguistics, 2:207–218, 2014. Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised learning of video representations using LSTMs. arXiv preprint arXiv:1502.04681, 2015. Ruslan L. Stratonovich. Conditional markov processes. Theory of Probability & Its Applications, 5(2):156–178, 1960. Ilya Sutskever, James Martens, and Geoffrey E. Hinton. Generating text with recurrent neural networks. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 1017–1024, 2011. Ilya Sutskever, James Martens, George Dahl, and Geoffrey E. Hinton. On the importance of initialization and momentum in deep learning. In Proceedings of the 30th International Conference on Machine Learning (ICML-13), pages 1139–1147, 2013. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104–3112, 2014. Tijmen Tieleman and Geoffrey E. Hinton. Lecture 6.5- RMSprop: Divide the gradient by a running average of its recent magnitude. https://www. youtube.com/watch?v=LGA-gRkLEsI, 2012. Alan M. Turing. Computing machinery and intelligence. Mind, pages 433–460, 1950. Subhashini Venugopalan, Marcus Rohrbach, Jeff Donahue, Raymond Mooney, Trevor Darrell, and Kate Saenko. Sequence to sequence–video to text. arXiv preprint arXiv:1505.00487, 2015. Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3156–3164, 2015. Andrew J. Viterbi. Error bounds for convolutional codes and an asymptotically optimum decoding algorithm. Information Theory, IEEE Transactions on, 13(2):260–269, 1967. Paul J. Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78(10):1550–1560, 1990. Wikipedia. Backpropagation — Wikipedia, the free encyclopedia, 2015. URL http://en.wikipedia.org/wiki/Backpropagation. [Online; accessed 18May-2015]. 37 Ronald J. Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Computation, 1(2):270–280, 1989. Wojciech Zaremba and Ilya Sutskever. Learning to execute. arXiv preprint arXiv:1410.4615, 2014. Matthew D. Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012. Matthew D. Zeiler, M. Ranzato, Rajat Monga, M. Mao, K. Yang, Quoc V. Le, Patrick Nguyen, A. Senior, Vincent Vanhoucke, Jeffrey Dean, et al. On rectified linear units for speech processing. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on, pages 3517– 3521. IEEE, 2013. 38
9cs.NE
1 On the Analysis of Puncturing for Finite-Length Polar Codes: arXiv:1801.05095v1 [cs.IT] 16 Jan 2018 Boolean Function Approach Song-Nam Hong∗ and Dennis Hui† ∗ Ajou University, Suwon, Korea, Email: [email protected] † Ericsson Research, San Jose, CA, Email: [email protected] Abstract This paper investigates the impact of puncturing on finite-length polar codes in which a puncturing pattern pN = (p0 , ..., pN ) is applied to a length-N polar code.. We first introduce two virtual channels to stochastically model the punctured (untransmitted) bits, which are respectively called useless channel model (UCM) and deterministic channel model (DCM). Under each model, we derive boolean functions in variables p0 , ..., pN −1 that can indicate which polarized channels should carry frozen bits. Based on this, we present an efficient method to jointly optimize a puncturing pattern and an information set. Focusing on a fixed information set, we show that there exist the so-called catastrophic puncturing patterns that will surely lead to a block error and derive their weight distributions recursively. We then propose the two construction methods of a rate-compatible (RC) polar code which ensures that each puncturing pattern in the family is non-catastrophic. Simulation results demonstrate that the proposed RC polar code outperform the RC Turbo code adopted in LTE. January 17, 2018 DRAFT 2 Index Terms Polar codes, puncturing, shortening, rate-compatible code, HARQ. I. I NTRODUCTION Polar codes, proposed by Arikan [1], achieve the symmetric capacity of binary-input discrete memoryless channels (BI-DMCs) under a low-complexity successive cancellation (SC) decoder. The finite-length performance of polar codes can be enhanced by using list decoder that enables polar codes to approach the performance of optimal maximum-likelihood (ML) decoder [2]. It was further shown in [2] that a polar code concatenated with a simple CRC outperforms welloptimized LDPC and Turbo codes even for short lengths. Due to their good performance and low-complexity, polar codes are currently considered for possible deployment in future wireless communication systems (i.e., 5G). Puncturing is widely used to support various lengths and to construct rate-compatible (RC) codes. In [5]–[7], a puncturing pattern is produced by a heuristic algorithm and then the information set of the mother polar code is optimized by taking into account the puncturing pattern. Also, an efficient algorithm to jointly optimize a puncturing pattern and the corresponding information set was recently proposed [10], where an exhaustive search over all possible puncturing patterns is performed, significantly reducing the search-space by using certain symmetry of the polar encoder. It can achieve an optimized frame-error-rate (FER) performance but still require a higher optimization complexity. A low-complexity method for joint optimization was further developed in [13]. It is remarkable that the above methods cannot be directly used to construct a RC polar code since in this case, all punctured polar codes in the family and the mother polar code should use the identical information set, i.e., the information set cannot be optimized according to puncturing patterns. In [8], [9], puncturing patterns are optimized for the given information set of the mother polar code. In general, this approach considerably reduces the optimization complexity, at the cost of an increased FER. DRAFT January 17, 2018 3 In this paper, we investigate the impact of puncturing on polarized channels of finite-length polar coded and, based on this, we present an efficient method to construct good (rate-compatible) puncturing patterns. First of all, we introduce two virtual channels that stochastically model punctured bits, which are referred to as useless channel model (UCM) and deterministic channel model (DCM), respectively. On the one hand, in UCM, it is assumed that the output of the virtual channel is independent of its input and hence, the polar decoder is performed by assigning zero log-likelihood ratios (LLRs) for the punctured bits. On the other hand, in DCM, it is assumed that the channel output is equal to its input as a pre-agreed bit (e.g., zero) with probability 1 and hence, the polar decoder is performed by assigning infinite LLRs for the punctured bits. Suppose that a puncturing pattern pN = (p0 , ..., pN −1 ) is applied to a length-N polar code. Our main contributions are summarized as follows. • We derive boolean functions in variables p0 , ..., pN −1 indicating which polarized channels should be assigned by frozen bits. The index sets of such polarized channels in UCM and DCM are denoted by DpN and EpN , respectively. • We define a reciprocal puncturing pattern which ensures that BpN = DpN (or BpN = EpN ), where BpN denote the index set of the locations of punctured bits. Then, we derive a necessary and sufficient condition such that a puncturing pattern is reciprocal. This condition is used to efficiently construct a good puncturing pattern and the corresponding information set. • Next, focus on a fixed information set, we show that there exist the so-called catastrophic puncturing patterns which will surely lead to a block error. Also, we develop an efficient recursive algorithm to characterize those puncturing patterns and their weight distributions. • Based on the above analysis, we present two simple methods to construct a RC polar code which ensures that each puncturing pattern in the family is non-catastrophic. Via simulation results, we demonstrate that the proposed RC polar code can outperform the RC Turbo code adopted in LTE. January 17, 2018 DRAFT 4 This paper is organized as follows. In Section II, we provide some useful notations and definition to be used throughout the paper. In Section III, using boolean functions, we analyze the impact of puncturing on the polarized channels of a polar code, and provide an efficient method to jointly optimize a puncturing pattern and an information set. In Section IV, we define the catastrophic puncturing patterns (which should be avoided) for a given information set and derive their weight distributions. In Section , we propose two simple methods to construct a RC polar code having a family of non-catastrophic puncturing patterns. Simulation results are provided in Section VI. Section VII concludes the paper. II. P RELIMINARIES In this section we provide some useful notations and definitions that will be used in the sequel. A. Notation ∆ Let [a : b] = {a, a + 1, . . . , b} for any integers a and b ≥ a. A polar code of length N = 2n is considered, in which the polarized channels are indexed by 0, 1, . . . , N −1. We let A ⊆ [0 : N −1] denote the information set that contains all the indices of unfrozen-bit channels. Accordingly, ⊗ log(N ) Ac contains all the indices of frozen bit channels. For any N = 2n , let GN = G2 be the rate-one generator matrix of all polar codes with blocklength N , where G2 is the 2-by-2 Arikan Kernel [1]. Also, for any M ⊆ [1 : N ] and N ⊆ [1 : N ], let GN (M, N ) denote the submatrix of GN obtained by selecting rows and columns whose indices belong to M and N , respectively. We define a function g(`) : [0 : N − 1] → {0, 1}n which maps ` onto a binary expansion as g(`) = (bn , . . . , b1 ), such that ` = (1) Pn n−i . i=1 bi 2 B. Punctured Polar Codes In this section, we formally define the polarized channels of a punctured polar code. Let s denote a number of punctured bits. Then, a polar code of blocklength N is punctured by DRAFT January 17, 2018 5 removing a set of s columns from its generator matrix, which has the effect of reducing the codeword length from N to Np = N − s. Formally, a punctured polar code of post-puncturing blocklength Np is characterized by its “mother” (unpunctured) polar code of blocklength N and a puncturing pattern pN = (p0 , . . . , pN −1 ) ∈ {0, 1}N with pi = 0 indicating that the i-th coded bit is punctured and thus not transmitted. Let BpN = {i ∈ [0 : N − 1] : pi = 0} be the index set which contains the zero locations in pN . Also, we let Bpc N = [0 : N − 1] \ BpN . Then, we have Np = wh (pN ) = |Bpc N |. For a given puncturing pattern pN , let πpN : Y N → Y Np be a projection operator that copies Np = wh (pN ) coordinates of its input as its output based on the puncturing pattern specified by pN , where wh (pN ) denotes the number of ones in pN , i.e. yNP = πpN (yN ) containing the coordinates of yN corresponding to the locations of ones in pN . The notion of polarized channels in conventional polar codes can be extended to punctured polar codes in a straightforward manner as follows. For a given (unpunctured) polar code of blocklength N and puncturing pattern pN , we define the transition probability of the i-th polarized channel of the corresponding punctured polar code as W (i) (yNp , ui−1 , pN |ui ) = 1 2N −1 X X uN i+1 Np }) yN ∈π −1 N ({y W N (yN |uN GN ), (2) p where πp−1N (S) , {yN ∈ Y N : πpN (yN ) ∈ S} represents the inverse image of πpN (·) and the channel transition probabilities are W N (yN |xN ) = Y W (yj |xj ), (3) j∈[1:N ] where W (·|·) denotes the channel transition probability of the underlying BI-DMC. For the (i) simplicity of notation, we let WpN denote the ith polarized channel with the transition probability (i) in (2), and let I(WpN ) denote the corresponding symmetric capacity. January 17, 2018 DRAFT 6 III. I MPACT OF P UNCTURING ON P OLARIZED C HANNELS As explained in Section II-B, punctured coded bits are not transmitted through an underlying channel. Conventionally, a length-N polar decoder is used regardless of the number of punctured bits and thus, the proper LLR values for those bits should be assigned. In this section, we introduce the two virtual channels that model those punctured bits stochastically, which are respectively called useless and deterministic channels. Under the transmitter-receiver agreement, one of the modelings is applied to each punctured bit and according to a chosen model, the proper LLR values are assigned. Then, we analyze their impacts on the polarized channels using boolean functions. A. Stochastic Modelings of Punctured Bits We define the two virtual channels to stochastically model punctured bits, which are referred to as useless channel model (UCM) and deterministic channel model (DCM), respectively. To be specific, they are defined as follows. • In UCM, it is assumed that the punctured coded bits are transmitted over a useless BIDMC. In this case, a channel output is independent of its input and thus, the corresponding channel can be defined as W (yj |xj ) = PY (yj ), (4) for some probability distribution PY (·). Under this model, the LLRs for the punctured coded bits are assigned by zeros since they are assumed to be equally likely of being one or zero. • In DCM, it is assumed that the punctured coded bits are transmitted over a deterministic (or noiseless) BI-DMC. In this case, a channel output is identical to its input with probability 1. In fact, this can be established by assigning the punctured coded bits with some fixed values that have been pre-agreed with the transmitter, so that the encoder restricts the transmitted codewords to be in the offset subspace specified by the fixed values. In this paper, without DRAFT January 17, 2018 7 loss of generality, the pre-agreed values are assumed to be zeros, i.e., xi = 0 for all i ∈ BpN . Accordingly, for an index set EpN , uEpN = (ui : i ∈ EpN ) should be chosen such that uGN ([1 : N ], BpN ) = 0, (5) uEpN = uE c N GN (Epc N , BpN )GN (EpN , BpN )−1 . (6) which is equivalent to p This condition shows that uEpN = (ui : i ∈ EpN ) should not carry information and be used as frozen bits. Also, the condition (6) is satisfied as long as GN (EpN , BpN ) is a full-rank. Under this model, the LLRs for the punctured coded bits are assigned by +∞. In the following Sections III-B and III-C, we investigate the impact of puncturing on polarized channels under each virtual channel model. In UCM, we show that some polarized channel, determined by a puncturing pattern, should have zero capacity and then identify the index set, denoted by DpN , of such polarized channels. Namely, we obtain that (i) I(WN ) = 0 for i ∈ DpN . (7) Next, in DCM, we identify the index set EpN defined above. Definitely, those index sets have an impact on the design of an information set A as in Remark 1 Remark 1: The reliabilities of polarized channels of a polar code can be computed by several techniques such as density evolution under a Gaussian approximation (DE/GA), tracking the mean value, Batthacharyya parameter or mutual information of the Gaussian L-densities [12]. Let I = {i1 , i2 , ..., iN } denote the orderd index set such that the reliability of polarized channel ij is not lower than that of polarized channel ik for i ≥ k. From this, we can construct the information set A of the size K, by taking the first K indices from the I. Whereas, for a punctured polar code, we need to more carefully construct an information set as follows: • In UCM, the information set A is constructed by taking the first K indices from the ordered set I \ DpN where |I \ DpN | = Np . January 17, 2018 DRAFT 8 𝑊) 𝑥" 𝑢" 𝑥# 𝑢# 𝑊" 𝑦" 𝑊# 𝑦# (a) The polar encoder of 𝑁 = 2 𝑊) Puncturing patterns Polarized channel capacities 𝑍 " 𝒑) 𝑝" AND 𝑍 " (𝒑) ) 0 0 0 1 𝑝" 0 0 1 1 𝑍 # (𝒑) ) 0 1 1 1 𝑝# 0 1 0 1 𝑍 # 𝒑) 𝑝# OR (b) The corresponding AND-OR representation Fig. 1. The capacity of polarized channel with a puncturing pattern p2 = (p0 , p1 ) for N = 2. Stage 2 Stage 3 Fig. 2. • 𝑍 " 𝒑+ = 0 AND 𝑍 % 𝒑+ = 0 OR 𝑍 & 𝒑+ = 1 AND 𝑍 ( 𝒑+ = 1 OR 𝑍 , 𝒑+ = 0 AND 𝑍 - 𝒑+ = 0 OR 𝑍 . 𝒑+ = 1 AND 𝑍 / 𝒑+ = 1 OR 0 0 AND AND 0 0 1 OR 1 1 OR 1 0 0 1 1 AND AND OR OR Binary representation Stage 1 0 0 1 1 AND 𝑝" = 0 (0,0,0) AND 𝑝% = 0 (0,0,1) AND 𝑝& = 1 (0,1,0) AND 𝑝( = 1 (0,1,1) OR 𝑝, = 0 (1,0,0) OR 𝑝- = 0 (1,0,1) OR 𝑝. = 1 (1,1,0) OR 𝑝/ = 1 (1,1,1) The capacity of polarized channel with a puncturing pattern p8 = (p0 , p1 , ..., p7 ) for N = 8. Likewise, in DCM, the information set A is constructed by by taking the first K indices from the ordered set I \ EpN where |I \ EpN | = Np . DRAFT January 17, 2018 9 B. Useless Channel Model (UCM) Focusing on the perfect underlying channel (i.e., I(W ) = 1), we first identify the DpN and then show its university with respect to the underlying B-DMC channel. By incorporating puncturing into the channel, the resulting channel in this case can be either a perfect channel or a completely noisy (i.e. zero-capacity) channel because a punctured bit is not transmitted. Without loss of generality, it is assumed that N copies of the channel, denoted by Wi , i ∈ [1 : N ], are used in (i) which I(Wi ) = 0 for i ∈ BpN and I(Wi ) = 1 for i ∈ / BpN . We let W̄pN denote the i-th polarized channel particularly when the underlying B-DMC is a perfect channel. We start with the simplest case of N = 2. From [1, Proposition 4], it is easily verified that (0) (8) (1) (9) I(W̄p2 ) ≤ min{I(W0 ), I(W1 )} I(W̄p2 ) ≥ max{I(W0 ), I(W1 )}. When p2 = (1, 0) is used, the capacities of resulting polarized channels are given by (0) (1) I(W̄p2 ) = 0 and I(W̄p2 ) = 1, (10) because of I(W0 ) = 1 and I(W1 ) = 0. From (8) and (9), we can easily obtain that    0, p2 ∈ {(0, 0), (0, 1), (1, 0)} (0) I(W̄p2 ) =   1, p2 = (1, 1)    0, p2 = (0, 0) (1) I(W̄p2 ) =   1, p2 ∈ {(0, 1), (1, 0), (1, 1)}. (11) (12) (i) Also, we can simply express I(W̄p2 ) as a boolean function in variables p0 and p1 such as (0) I(W̄p2 ) = p0 ∧ p1 and (1) I(W̄p2 ) = p0 ∨ p1 , (13) where “∧” and “∨” denote the AND and OR operations, respectively, and to simplify notation, we will omit the “∨” in the following. To simplify the expression and emphasize the boolean January 17, 2018 DRAFT 10 function, we let Z (i) (pN ) denote the capacity of the ith polarized channel, i.e., Z (i) (pN ) = (i) I(W̄pN ). Note that the boolean functions in (13) can be immediately obtained from Fig. 1 (b). The next level of recursion for N = 4 is shown in Fig. 2. We first define a binary-output function h : [0, 1] × [0, 1] → [0, 1] as hb (p0 , p1 ) =    p0 ∧ p1 , if b = 0   p0 p1 , if b = 1. (14) Letting g(i) = (b0 , b1 ), we can obtain Z (i) (p4 ) = hb1 (hb0 (p0 , p2 ), hb0 (p1 , p3 )). (15) From this, we can obtain the boolean functions for N = 4 as follows: Z (0) (p4 ) = p0 ∧ p1 ∧ p2 ∧ p3 (16) Z (1) (p4 ) = p0 p1 ∧ p0 p3 ∧ p1 p2 ∧ p2 p3 (17) Z (2) (p4 ) = p0 p2 ∧ p1 p3 (18) Z (3) (p4 ) = p0 p1 p2 p3 . (19) Repeatedly applying the above recursions as in (15), we can derive the Z (i) (pN ), i = 0, . . . , N −1 as a function of a puncturing pattern pN . When the perfect underlying channel (i.e., I(W ) = 1) is assumed, we obtain the index set of zero-capacity polarized channels as DpN = {i ∈ [0 : N − 1] : Z (i) (pN ) = 0}. (20) Also, the above set is universal with respect to the underlying B-DMC channel W because (i) (i) (i) from [11, Lemma 4.7], I(WpN ) ≤ I(W̄pN ) for any W , and thus I(WpN ) = 0 for any W if Z (i) (pN ) = 0. DRAFT January 17, 2018 11 C. Deterministic Channel Model (DCM) Focusing on the simplest case of N = 2, we can see that the condition (6) is satisfied by choosing u0 and u1 according to a puncturing pattern    u0 = 0, u1 = 0 p2    u0 = u1 , p2      u =0 p2 1 as = (0, 0) = (0, 1) (21) = (1, 0). From the above, we can see that in the third case, u1 should be frozen bits, and in the second case, either u0 or u1 should be frozen bits. Since the second polarized channel has a higher capacity than the first polarized channel, we choose u0 as frozen bit in the second case. Let p̄N denote the 1’s complement of pN . Likewise the case of UCM, we can express the (21) using boolean function as Z (0) (p̄2 ) = p̄0 ∧ p̄1 and Z (1) (p̄2 ) = p̄0 p̄1 , (22) and the condition (6) is satisfied if ui is chosen as frozen bit when Z (i) (p̄2 ) = 1. From this, we obtain the equivalent result with the (21) as follows: Z (0) (p2 ) = 1 if p2 ∈ {(0, 0)} (23) Z (1) (p2 ) = 1 if p2 ∈ {(0, 1), (1, 0), (1, 1)}. (24) In this simplest case, thus, we can identify the Ep2 = {i ∈ [0 : 1] : Z (i) (p̄2 ) = 1}. (25) Following the same recursion procedures in UCM, we can obtain the EpN = {i ∈ [0 : N − 1] : Z (i) (p̄N ) = 1}. (26) Remark 2: In UCM, the frozen bits ui for i ∈ DpN can be set by zero, without loss of performance. In DCM, however, the frozen bits ui for i ∈ DpN should be determined as a binary addition of other bits, and thus, we require another computations (i.e., encoding) to determine January 17, 2018 DRAFT 12 the frozen bits. In the next section, it will be shown that for some class of puncturing patterns, those frozen bits can be simply chosen as zeros. Example 1: Consider the case of N = 4. From (16)-(19), we obtain the corresponding boolean functions as Z (0) (p̄4 ) = p̄0 ∧ p̄1 ∧ p̄2 ∧ p̄3 (27) Z (1) (p̄4 ) = p̄0 p̄1 ∧ p̄0 p̄3 ∧ p̄1 p̄2 ∧ p̄2 p̄3 (28) Z (2) (p̄4 ) = p̄0 p̄2 ∧ p̄1 p̄3 (29) Z (3) (p̄4 ) = p̄0 p̄1 p̄2 p̄3 . (30) For p4 = (1, 0, 1, 0), we have Ep4 =(1,0,1,0) = {1, 3} from (26), which implies that the condition (6) is satisfied by assigning u2 and u3 as frozen bits. This is verified as    1 0  G4 ({1, 3}, {1, 3}) =   1 1 (31) is a full-rank. D. Reciprocal Puncturing Patterns In Sections III-B and III-C above, we showed that for each pN , there exists the corresponding index sets DpN ⊆ Ac and EpN ⊆ Ac for UCM and DCM, respectively. In order to construct a good information set, they should be identified. In this section, we show that for some puncturing patterns, we can ensure that BpN = DpN (or BpN = EpN ). Formally, we define: Definition 1: A puncturing pattern pN is referred to as reciprocal if BpN = DpN under UCM (or BpN = EpN under DCM). Using this, we provide the main theorems of this section below: Theorem 1: In UCM, a puncturing pattern pN is reciprocal if and only if the following DRAFT January 17, 2018 13 properties are satisfied: zero-inclusion property: 0 ∈ BpN , one-covering property: if i ∈ BpN and i 1 j, then j ∈ BpN , (32) (33) where i 1 j means that for every digit of ’1’ in the binary representation of index j, the corresponding digit in the index i must also be ’1’ and i 1 0 for every i > 0. Proof: We first prove sufficiency and then prove necessity. a) Proof of sufficiency: It suffices to show that the input and output sequences remain the same after each polarization stage in the modified encoder with AND-OR operators in Fig. 1 (b). Let the binary representation of an index i be denoted by g(i) = (bi1 , bi2 , ..., bin ) where n = log2 N . At each polarization stage k for k ∈ {1, ..., n}, there are 2n−1 pairs of basic AND-OR operators in the form of that in Fig. 1 (b), each of which has the input pairs whose indices correspond to `0 = (bi1 , ..., bik−1 , 0, bik+1 , .., bin ) and `1 = (bi1 , ..., bik−1 , 1, bik+1 , .., bin ). The input and the output of each pair are identical if (p`0 , p`1 ) ∈ {(0, 0), (0, 1), (1, 1)} as in Fig. 1 (b). If the zero-inclusion and one-covering properties are satisfied, then the input and output of each pair of these ANDOR operators are identical. Since this argument holds for every polarization stage, it follows that those properties imply reciprocity. b) Proof of necessity: Note that if the input and the output of the modified polar encoder are the same (i.e., BpN = DpN ), then the input and output after each polarization stage must be also the same because the largest bit index with value ’0’ is monotonically non-increasing after each polarization stage as the output of the AND operator should be smaller than that of the OR operator for the same input. Now suppose there exist i and j such that i ∈ BpN and i 1 j but j ∈ / BpN . There must exist `1 , `2 ∈ [1 : N ] such that i 1 `1 1 `2 1 j, `1 ∈ BpN , `2 ∈ / BpN , and that `1 and `2 differ only in one bit (say, k-th bit), where k ∈ {1, 2, ..., n}. It follows that b`k1 , b`k2 = 0, and b`m1 = b`m2 for all m 6= k. Since `1 ∈ BpN and `2 ∈ / BpN , the input to the pair of AND-OR operators at stage k corresponding to `1 and `2 should be different from its output as January 17, 2018 DRAFT 14 in the 3rd column of the table in Fig. 1 (b). It follows that the input and output after (at least) the k-th polarization stage must be different and thus the input and the output of the whole modified polar encoder must be different. Thus, BpN 6= DpN . Theorem 2: In DCM, a puncturing pattern pN is reciprocal if and only if the following properties are satisfied: N -inclusion property: N ∈ BpN , (34) zero-covering property: if i ∈ BpN and i 0 j, then j ∈ BpN , (35) where i 0 j means that for every digit of ’0’ in the binary representation of index j, the corresponding digit in the index i must also be ’0’. Proof: The proof follows the same procedures in the proof of Theorem 1. Corollary 1: Suppose DCM is assumed. If a puncturing pattern pN is reciprocal, then we have ui = 0 for i ∈ EpN . (36) Proof: Using zero-covering property in Theorem 2, we define ΦN (i) = {j ∈ [0 : N − 1] : i 0 j} ∪ {N − 1}, (37) ⊗ log(N ) for any i ∈ [0 : N − 1]. Due to the particular construction of GN = G2 , we can see that xi , i-th coded bit, is determined as an addition of ui ’s with i ∈ ΦN (i). Since the pN is reciprocal by Hypothesis assumption, we know that ΦN (i) ⊆ EpN for any i ∈ EpN . Thus, if we choose ui = 0 for all i ∈ EpN , xi = 0 for all i ∈ BpN . This completes the proof. From now on, we will prove a class of reciprocal puncturing patterns. Definition 2: Let Πn denote the set of all possible permutation of (1, 2, ..., n). For any π ∈ Π, π(i) denote the i-th element of π. For any length-n binary vector b = (b1 , ..., bn ), we define a bit-permutation Ψπ with seed vector π ∈ Πn as Ψπ (b) = (bπ(1) , bπ(2) , ..., bπ(n) ). DRAFT (38) January 17, 2018 15 Proposition 1: Under UCM, suppose that pN is reciprocal. Then, for any π ∈ Πn , the is also reciprocal, where puncturing pattern pN π with zero-location set BpN π  BpNπ = g −1 (Ψπ (g(i))) : i ∈ BpN . (39) The above statement is also hold under DCM. Proof: The proof follows the fact that one-covering property (or zero-covering property) definitely holds for any bit-permutation π. Example 2: In this example, we provide some reciprocal puncturing patterns. First of all, we have: • In UCM, the puncturing pattern pN with BpN = {0, 1, ..., s − 1} is obviously reciprocal. • In DCM, the puncturing pattern pN with BpN = {N − s + 1, N − s + 2, ..., N } is obviously reciprocal. Using the above puncturing patterns and from Proposition 1, we are able to generate several reciprocal puncturing patterns. In particular, when π = (n, n − 1, ..., 1) (called bit-reverse permutation), the corresponding puncturing patterns in UCM and DCM are called quasi-uniform puncturing (QUP) and reverse QUP (RQUP), respectively. Fig. 5 shows the frame-error-rate (FER) performances of punctured polar codes where QUP and RQUP are used for UCM and DCM, respectively. In this example, we observe that UCM approach is slightly better than DCM approach at lower rate and vice versa at higher rate. IV. C ATASTROPHIC P UNCTURING PATTERNS From the analysis in Section III, we learned that an information set A should be carefully designed for a punctured polar code, by taking into account a puncturing pattern and a chosen stochastic modelling. However, this approach cannot be used on constructing a rate-compatible code since an information set should not be changed according to the puncturing patterns in the family, i.e., a common information set, usually optimized for a mother code, should be January 17, 2018 DRAFT 16 100 10-1 R=3/4 FER R=1/3 10-2 10-3 10-4 -2 UCM, (32,96) polar code DCM, (32,96) polar code UCM, (72,96) polar code DCM, (72,96) polar code -1 0 1 2 3 4 5 6 SNR [dB] Fig. 3. Performance comparisons of UCM and DCM approaches for various code rates. Here, the punctured polar codes of Np = 96 are obtained by puncturing 32 bits from mother polar code of N = 128. kept. Motivated by this, we in this section analyze the impact of puncturing patterns for a fixed (pre-designed) information set A. Definition 3: A puncturing pattern pN is said to be catastrophic for a polarized channel i (obtained from any underlying channel W ) if Z (i) (pN ) = 0. (40) More generally, a puncturing pattern pN is called catastrophic for a set of polarized channels i ∈ A ⊆ [0 : N − 1] if Z (i) (pN ) = 0 for any i ∈ A. (41) Obviously, a catastrophic puncturing pattern is universal with respect to the underlying BI-DMC DRAFT January 17, 2018 17 channel W . If a catastrophic puncturing pattern is used, then the block-error-rate (BLER) of the corresponding punctured code is always 1 and thus, it should be avoid. From the boolean functions in Section III, we obtain the two interesting facts: • Any puncturing pattern pN with wh (pN ) > N − 2wt(i) is non-catastrophic for the polarized channel i. • The boolean functions Z (i) (pN ) can be always be expressed in the corresponding maximal forms, in which each term represents a catastrophic puncturing pattern. In the example of N = 4, the first fact is verified from (16)-(19). Also, the maximal form of Z (2) (p4 ) is obtained using the fact pi ∧ p̄i = 0 as Z (2) (p4 ) =p0 p2 ∧ p1 p3 =p0 p2 (p1 ∧ p̄1 )(p̄3 ∧ p3 ) ∧ p1 p3 (p0 ∧ p̄0 )(p̄2 ∧ p2 ) =p0 p1 p2 p3 ∧ p0 p1 p2 p̄3 ∧ p0 p̄1 p2 p3 ∧ p0 p̄1 p2 p̄3 ∧ p0 p1 p̄2 p3 ∧ p̄0 p1 p2 p3 ∧ p̄0 p1 p̄2 p3 , and the corresponding catastrophic puncturing patterns are {(0, 0, 0, 0), (0, 0, 0, 1), (0, 1, 0, 0), (0, 1, 0, 1), (0, 0, 1, 0), (1, 0, 0, 0), (1, 0, 1, 0)}. (42) Using the boolean function Z (i) (pN ), we can check if pN is a catastrophic or not for a given polarized channel i. Also, we can efficiently check it using the following lemma. Lemma 1: A puncturing pattern pN is non-catastrophic for a polarized channel i if the following rank-increment condition is satisfied: Rank GN [i : N ], Bpc N January 17, 2018  − Rank GN [i + 1 : N ], Bpc N  = 1. DRAFT 18 Proof: When pN is applied, the capacity of the polarized channel i is given by Z (i) (pN ) = I(ui ; {yj : j ∈ Bpc N }|u1i−1 )  = H(ui ) − H(ui |uN G [1 : N ], Bpc N , u1i−1 ) = H(ui ) − H(ui |{yj : j ∈ Bpc N }, ui−1 1 )    c N c = 1 − H ui |ui G i, BpN + ui+1 G [i + 1 : N ], BpN , where the interference term caused by the known values u1i−1 , i.e., u1i−1 G([1 : i − 1], Bpc N ), is eliminated. Then, we have that  H ui |{yj : j ∈ Bpc N }, ui−1 = 0, 1 if G(i, Bpc N ) is linearly independent from the rows of G([i + 1 : N ], Bpc N ). This completes the proof. In Section IV-A, we analyze the catastrophic puncturing patterns and then in Section ??, we present an efficient algorithm to construct a non-catastrophic puncturing pattern. A. Analysis of Catastrophic Puncturing Patterns For a given polarized channel i with g(i) = (bn , . . . , b1 ), we characterize all catastrophic puncturing patterns and their weight distributions. 1) Characterization of catastrophic puncturing patterns: Define CaN (bn , . . . , b1 ) by the set of catastrophic puncturing patterns of the polarized channel i, i.e., ∆ CaN (bn , . . . , b1 ) = {pN ∈ {0, 1}N : Z (i) (pN ) = 0}. (43) Now we develop an efficient recursive method to characterize CaN (bn , . . . , b1 ). We first construct a binary tree consisting of N = 2n leaf nodes and n levels. When n = 4, there are four types of binary trees, two of which are illustrated in Fig. 4. Note that the binary tree has two types DRAFT January 17, 2018 19 {0000,0001,0100,0101,0010,1000,1010} {0000,0001,0100,0010,0011,0110,1000,1001,1100} Stage 2 Stage 2 AND OR {00} {00} {00,01,10} {00,01,10} Stage 1 Stage 1 OR 𝑝" Fig. 4. OR 𝑝# 𝑝) AND 𝑝* 𝑝" AND 𝑝# 𝑝) 𝑝* The binary trees to characterize the catastrophic puncturing patterns for n = 2. of function nodes as “AND” node and “OR” node, and leaf nodes. Also, each level m contains the same types of function nodes in which the type is completely determined by the bm as    AND node, bm = 0   OR node (44) bm = 1. For a given node v, we define the subtree that contains all child nodes of v. As before, we can derive the boolean function associated with the subtree. Then, the output of the node v (i.e., the message value of its outgoing edge) represents the set of all binary sequences associated with the leaf nodes in the subtree such that each binary sequence in the set is the root of the boolean function. Due to the symmetry of the binary tree, all edges in the same level have the same message value (see Fig. 4). We let I and O represent the incoming and outgoing message values at a node, respectively. From Fig. 4, the message value of the outgoing edge from v is computed using the following update rules: • Update rule at OR node: O = {(p1 , p2 ) : p1 , p2 ∈ I}. January 17, 2018 (45) DRAFT 20 • Update rule at AND node: O = {(p1 5 p2 ) : p1 , p2 ∈ I}, (46) where, for the ease of expression, we used the notation of ∆ {(p1 5 p2 ) : p1 , p2 ∈ I} = {(p1 , p2 ) : p1 ∈ I, p2 ∈ {0, 1}lI } ∪ {(p1 , p2 ) : p1 ∈ {0, 1}lI , p2 ∈ I}, where lI denote the length of binary sequences in I. Using the binary tree and the above update rules, we can find all catastrophic puncturing patterns recursively. Starting with n = 1, we have: Ca2 (0) = {00, 01, 10} and Ca2 (1) = {00}. (47) Using them, we obtain Ca4 (b2 , b1 ) as    {(p1 , p2 ) : p1 5 p2 ∈ C 2 (b2 )}, b1 = 0 a 4 Ca (b2 , b1 ) =   {(p1 , p2 ) : p1 , p2 ∈ Ca2 (b2 )}, b1 = 1. Repeatedly applying the above procedures, we have: Proposition 2: For a polarized channel i with g(i) = (bn , . . . , b1 ), the set of catastrophic puncturing patterns are obtained recursively as n−m Ca2 (bn , . . . , bm+1 ) =    {(p1 5 p2 ) : p1 , p2 ∈ C 2n−m−1 (bn , . . . , bm+2 ), bm+1 = 0 a   {(p1 , p2 ) : p1 , p2 ∈ Ca2n−m−1 (bn , . . . , bm+2 ), bm+1 = 1 for m = n − 2, . . . , 1, 0, where the recursion starts with Ca2 (0) = {00, 01, 10} and Ca2 (1) = {00}. DRAFT January 17, 2018 21 2) Weight distributions: We derive the weight distributions of catastrophic puncturing patterns recursively, in which the recursion starts with simple polynomial of z at each coded bit location. P Pn s s First of all, for any two input polynomial Dα (z) = m s=0 dα (s)z and Dβ (z) = s=0 dβ (s)z , where dα (s) and dβ (s) denote the corresponding numbers of catastrophic puncturing patterns with zero weight s (i.e. s zeros), we obtain the output polynomials after OR and AND operations as: • OR operation: ∆ f1 (Dα (z), Dβ (z)) = Dα (z)Dβ (z). (48) • AND operation: ∆ f0 (α(z), β(z)) = Dα (z)(1 + z)m + Dβ (z)(1 + z)n − Dα (z)Dβ (z), (49) where m and n denote the highest degrees of Dα (z) and Dβ (z), respectively. Note that (48) and (49) are derived from (45) and (46), respectively. As performed in Section IV-A1, using the binary tree and the above update rules, we get: Proposition 3: The weight distribution of the set of catastrophic puncturing patterns for a polarized channel i of a length-N polar code can be computed recursively by (N ) Di N X ∆ (z) = (N ) di (s)z s s=2wh (i)   (N/2) (N/2) = f[i] mod 2 Dbi/2c (z), Dbi/2c (z) , (1) (50) ∆ for i ∈ {0, 1, 2..., N − 1}, where D0 (z) = z. For example, from Theorem 3, we can derive the (4) D2 (z) = 2z 2 + 4z 3 + z 4 , (51) which is well-matched to (42). January 17, 2018 DRAFT 22 V. C ONSTRUCTION OF R ATE -C OMPATIBLE P OLAR C ODE In this section, we will construct a RC polar code to send k information bits with M various code rates R1 > R2 > · · · > RM , where a puncturing pattern pN i is applied to a mother polar code of length N = k/RM to achieve a target rate Ri . For all code rates, a fixed information set A is used. Also, for the simplicity, it is assumed that N has the form of power of 2. Then, we present the two efficient methods, named greedy and reciprocal constructions, to construct RC polar codes where each puncturing pattern pN i in the family is non-catastrophic. Algorithm 1 Greedy algorithm to find pN 0 1: Initialization: 2: Ao = {i1 , . . . , ik } denotes an ordered information set of A with respect to Hamming weight. 3: = φ, and t = 1. pN 0 = (0, . . . , 0), BpN 0 4: while t ≤ k do for j = 1, . . . , N : j ∈ / BpN0 do 5: if Z (it ) (pN 0 + ej ) = 1 then 6: N pN 0 = p0 + ej 7: . ej denotes an unit vector with 1 in the jth location. t = t + 1 and go to 4. 8: else if j=N then 9: N pN 0 = p0 + er 10: . r is randomly and uniformly chosen from the zero locations in pN 0 t=t+1 11: end if 12: end for 13: 14: end while DRAFT January 17, 2018 23 A. Greedy Construction The main idea of the greedy construction is as follows: 1) Find a possibly minimum weight non-catastrophic pN 0 such that Z (i) (pN 0 ) = 1 for all i ∈ A. (52) N Note that if wH (pN 0 ) > wH (p1 ), then it is impossible to construct the desired RC polar N code. In this section, it is assumed that wH (pN 0 ) ≤ wH (p1 ). N N 2) To obtain a non-catastrophic puncturing pattern pN 1 , we add wH (p1 ) − wH (p0 ) ones to some zero locations in pN 0 . N N 3) Next, to obtain a non-catastrophic puncturing pattern pN 2 , we add wH (p2 ) − wH (p1 ) ones to some zero locations in pN 1 . N N 4) In general, to obtain a non-catastrophic puncturing pattern pN i , we add wH (pi )−wH (pi−1 ) ones to some zero locations in pN 1 . 5) Repeat the above procedures until obtaining all the puncturing patterns in the family. Note that in the above, the locations to add ones are chosen arbitrary (i.e., randomly and uniformly, or via a clever algorithm). Then, the proposed algorithm is established by developing a greedy algorithm to find the pN 0 (see Algorithm 1). In this algorithm, we used an ordered information set Ao = {i1 , . . . , ik } (with respect to a Hamming weight) from A, where wH (g(ij )) ≥ wH (g(ik )) for any j, k ∈ Ao with j ≥ k. This is motivated by the fact that a polarized channel i tends to have more non-catastrophic puncturing patterns as wH (g(i)) becomes larger. Thus, the use of the Ao makes it easier to satisfy the condition in line 6 in Algorithm N 1, by adding fewer ones’ in pN 0 . Eventually, we are able to obtain a lower weight p0 to satisfy (52). Note that Algorithm 1 is a randomized algorithm since, when the condition in line 6 in Algorithm is not satisfied, i.e., we need to add more ones to pN 0 to satisfy the condition, one is added to an arbitrary zero location of pN 0 . Note that although the use of Ao is helpful to find a N lower weight pN 0 , it does not ensure that Algorithm 1 finds the minimum weight p0 . January 17, 2018 DRAFT 24 B. Reciprocal Construction N In this section, we construct a family of rate-compatible puncturing patterns pN 1 , ..., pM −1 where they are reciprocal. We first focus on the design of pN M −1 = (pM −1,0 , ..., pM −1,N −1 ). In the proposed approach, the |BpNM −1 | number of zero locations will be chosen so that the one-covering N N property in Theorem 1 is satisfied. Then, we design the pN M −2 by adding the |pM −2 | − |pM −1 | number of additional zero locations to the pN M −1 . This process will be continued until constructing all the puncturing patterns in the family. The specific construction methods will be explained as follows. Define the useful set operation below: Definition 4: For any i, j ∈ [0 : N − 1], we say that i ⊥ j if wH (g(i) ⊕ g(j)) = wH (g(i)) + wH (g(j)), where ⊕ denotes a binary addition. Then, for any two index sets M and N , define a set operation as  M  N = g −1 (g(i) ⊕ g(j)) : i ⊥ j, i ∈ N , j ∈ M . (i) Using the above definitions and a given information set A, we derive the ordered sets LA , i = 0, ..., n, in a recursive manner as ∆ (0) (0) LA = {0} = {i1 } ∆ (1) (1) LA = {i ∈ [0 : N − 1] : wH (g(i)) = 1} \ A = {i1 , ..., i (j) LA =  (j−1) LA  (1) LA  ∆ (j) \ A = {i1 , ..., i (j) (j) |LA | (1) (1) |LA | } }, P (j) for j = 2, ..., n. Then, we define a sequence of length nj=1 |LA as   (0) (1) (1) (n) (n) Iseed (A) = i1 , i1 , ..., i (1) , ..., i1 , ..., i (n) , |LA | |LA | (53) where this sequence depends on the choice of an information set. Leveraging this sequence, we construct the rate-compatible puncturing patterns (equivalently, BpNi ) as follows: 1) The BpNM −1 is constructed by taking the first |BpNM −1 | elements of Iseed (A). By construction (i) of LA ’s, the corresponding puncturing pattern pN M −1 is reciprocal and non-catastrophic with respect to A. DRAFT January 17, 2018 25 2) Next, the BpNM −2 is constructed by taking the first |BpNM −2 | elements of Iseed (A). Clearly, BpNM −1 ⊂ BpNM −2 , namely, rate-compatibility is satisfied. 3) In general, the BpNi is constructed by taking the first |BpNi | elements of Iseed (A). Note that if |Iseed (A)| < |BpN1 |, then it is impossible to construct the desired RC polar code. In this case, we need to change the highest code rate. In the following example, we show that the supportable highest rate is determined as a function of an information set. Example 3: Suppose that N = 8 (e.g., n = 3) and A = {4, 6}. Then we have: (0) (1) LA = {0} and LA = {1, 2} (2) (1) (1) (3) (2) (1) LA = (LA  LA ) \ A = {g −1 ((0, 0, 1) ⊕ (0, 1, 0))} = {3} LA = (LA  LA ) \ A = φ. Then, we have Iseed ({4, 6}) = (0, 1, 2, 3). Also, in other example of A = {5, 7}, we have: (0) (1) LA = {0} and LA = {1, 2, 4} (2) (1) (1) (3) (2) (1) LA = (LA  LA ) \ A = {3, 6} LA = (LA  LA ) \ A = φ. Then, we have Iseed ({5, 7}) = (0, 1, 2, 4, 3, 6). This example shows that the number of possible punctured bits varies according to the choice of information set A. VI. N UMERICAL R ESULTS In this section, we evaluate the performances of the proposed puncturing methods. For the simulation, AWGN channel and BPSK modulation are assumed. Length-flexibility: In this case, a puncturing is performed to construct a polar code whose code length is not the power of 2. We employed a polar code with N = 128. To construct the punctured polar code of Np = 96, we use the QUP and RQUP in Section III-D for UCM and DCM, respectively. Also, CRC-aided list decoder is used with 8-bit CRC and list-size 8. Fig. 3 January 17, 2018 DRAFT 26 BLACK: LTE Turbo Code RED: RC Polar Code (Greedy) BLUE: RC Polar Code (Reciprocal) FER 10-1 R=3/4 R=1/2 R=1/3 -2 10 10-3 -2 -1 0 1 2 3 4 5 6 7 8 9 SNR [dB] Fig. 5. Performance comparisons of RC polar codes and RC Turbo codes. Black lines denotes the performances of the RC Turbo code. Also, red and blue lines denote the performances of the RC polar codes constructed by greedy and reciprocal constructions, respectively. shows that UCM performs better than DCM at rate 1/3, while vice versa, at rate 3/4. Hence, we can learn that either UCM or DCM should be carefully chosen according to the code rate of a target punctured polar code. Rate-compatible polar code: Fig. 5 shows the performance comparisons of the proposed RC polar codes and RC Turbo code. For the simulation, we constructed the two RC polar codes where where one is developed by greedy construction in Section V-A and the other is by reciprocal construction in Section V-B. As mother codes, both polar and Turbo codes have the length of 256 (e.g., N = 256). Also, for polar code, CRC-aided list decoder is used with 5-bit CRC and list-size 32. Hence, the number of information bits for polar code is 93 while that of Turbo DRAFT January 17, 2018 27 code is 88. From this simulation, we observe that the proposed RC polar codes outperform the RC Turbo code adopted in LTE. Also, we can see that a low-complexity reciprocal construction performs very well. VII. C ONCLUSION We derived the boolean expressions of the capacities of polarized channels of finite-length polar codes. Based on this, we provided a guideline to jointly optimize a puncturing and the corresponding information set, Also, it was shown that for each fixed information set, there exist the catastrophic puncturing patterns that should be avoided to yield a good performance. Furthermore, we presented two efficient methods to construct non-catastrophic puncturing patterns for any fixed information set. Leveraging them, we designed the RC polar codes using noncatastrophic puncturing patterns. Via simulation results, it was demonstrated that the proposed RC polar codes can outperform the RC Turbo code adopted in LTE. Therefore, polar codes can be a good candidate for 5G channel coding. R EFERENCES [1] E. Arikan, “Channel polarization: A method for constructing capacity-achieving codes for symmetric binary-input memoryless channels,” IEEE Trans. Inf. Theory, vol. 55, pp. 3051-3073, Jul. 2009. [2] I. Tal and A. Vardy, “List decoding of polar codes,” IEEE Trans. Inf. Theory, vol. 61, pp. 2213-2226, May 2015. [3] A. Eslami and H. Pishro-Nik, “Practical approach to polar codes,” in Int. Symp. Inf. Th., pp. 16-20, 2011. [4] K. Chen, K. Niu, and J. Lin, “A hybrid ARQ scheme based on polar codes,” IEEE Commun. Letters, vol. 17, pp. 1996-1999, 2013. [5] R. Wang and R. Liu, “A novel puncturing scheme for polar codes,” IEEE Commun. Lett., vol. 18, no. 12, pp. 2081-2084, Dec. 2014. [6] D.-M. Shin, S.-C. Lim, and K. Yang, “Design of length-compatible polar codes based on the reduction of polarizing matrices,” IEEE Trans. Commun., vol. 61, no. 7, pp. 2593-2599, Jul. 2013. [7] K. Niu, K. Chen and J.-R. Lin, “Beyond turbo codes: rate-compatible punctured polar codes,” in Proc. IEEE Int. Conf. Commun. (ICC), Budapest, Hungary, Jun. 2013. [8] L. Zhang, Z. Zhang, X. Wang, Q. Yu, and Y. Chen, “On the Puncturing Patterns for Punctured Polar Codes,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT) Huawaii, USA, Jul. 2014. January 17, 2018 DRAFT 28 [9] H. Saber and I. Marsland, “An incremental redundancy hybrid ARQ scheme via puncturing and extending of polar codes,” IEEE Trans. Commun., vol. 63, no. 11, pp. 3964-3973, Nov. 2015. [10] V. Miloslavskaya, “Shortened Polar Codes,” IEEE Trans. Inf. Theory, vol. 61, pp. 4852-4865, Sep. 2015. [11] S. Korada, “Polar codes for channel and source coding,” Ph.D. thesis, E’cole Polytechnique Fe’de’rale de Lausanne (EPFL), 2009. [12] G. Sarkis, I. Tal, P. Giard, A. Vardy, C. Thibeault, and W. J. Gross, “Flexible and low-complexity encoding and decoding of systematic polar codes,” IEEE trans. on commun., vol. 64, no. 7, pp. 2732-2745, Jul. 2016. [13] V. Bioglio, F. Gabry and I. Land, “Low-Complexity Puncturing and Shortening of Polar Codes,” in Proc. IEEE Wire. Comm. Net. Conf. (WCNC) Works., San Francisco, CA, Mar. 2017. DRAFT January 17, 2018
7cs.IT
1 Low RF-Complexity Technologies to Enable Millimeter-Wave MIMO with Large Antenna Array for 5G Wireless Communications arXiv:1607.04559v3 [cs.IT] 1 Oct 2017 Xinyu Gao, Student Member, IEEE, Linglong Dai, Senior Member, IEEE, and Akbar M. Sayeed, Fellow, IEEE Abstract—Millimeter-wave (mmWave) MIMO with large antenna array has attracted considerable interests from academic and industry communities, as it can provide larger bandwidth and higher spectrum efficiency. However, with hundreds of antennas, the number of radio frequency (RF) chains required by mmWave MIMO is also huge, leading to unaffordable hardware cost and power consumption in practice. In this paper, we investigate low RF-complexity technologies to solve this bottleneck. We first review the evolution of low RF-complexity technologies from microwave frequencies to mmWave frequencies. Then, we discuss two promising low RF-complexity technologies for mmWave MIMO systems in detail, i.e., phased array based hybrid precoding (PAHP) and lens array based hybrid precoding (LAHP), including their principles, advantages, challenges, and recent results. We compare the performance of these two technologies to draw some insights about how they can be deployed in practice. Finally, we conclude this paper and point out some future research directions in this area. I. I NTRODUCTION M Illimeter-wave (mmWave) (30-300 GHz) multipleinput multiple-output (MIMO) with large antenna array has been considered as a promising solution to meet the one thousand times increase in data traffic predicted for further 5G wireless communications [1]. On one hand, mmWave can provide nearly 2 GHz bandwidth [2], which is much larger than the 20 MHz bandwidth in current 4G wireless communications without carrier aggregation. On the other hand, the short wavelengths associated with mmWave frequencies enable a large antenna array to be packed in a small physical size, which means that MIMO with a large antenna array is possible at mmWave frequencies to effectively compensate the high path loss induced by high frequencies and considerably improve the spectrum efficiency [3]. This work has been posted on arXiv (arXiv:1607.04559) on 15 July, 2016 and revised on 16 Sep, 2017. It has also been posted on the personal website at: http://oa.ee.tsinghua.edu.cn/dailinglong/publications/publications.html. X. Gao and L. Dai (corresponding author) are with the Tsinghua National Laboratory for Information Science and Technology (TNList), Department of Electronic Engineering, Beijing 100084, China (e-mails: [email protected], [email protected]). A. Sayeed is with the Department of Electrical and Computer Engineering, University of Wisconsin, Madison, WI 53706, USA (email: [email protected]). This work was supported by the National Natural Science Foundation of China for Outstanding Young Scholars (Grant No. 61722109), the National Natural Science Foundation of China (Grant No. 61571270), the Royal Academy of Engineering through the UK-China Industry Academia Partnership Programme Scheme (Grant No. UK-CIAPP\49), and the US National Science Foundation (Grant Nos. 1548996 and 1444962). However, realizing mmWave MIMO in practice is not a trivial task. One challenging problem is that each antenna in MIMO systems usually requires one dedicated radio-frequency (RF) chain, including digital-to-analog converters (DACs), mixers, and so on [3]. This will result in unaffordable hardware cost and power consumption in mmWave MIMO systems, as the number of antennas is huge (e.g., 256 compared with 8) and the power consumption of RF chain is high (e.g., 250 mW at mmWave frequencies compared with 30 mW at microwave frequencies) [3]. Therefore, the large number of RF chains with prohibitively high power consumption is a bottleneck for mmWave MIMO with large antenna array in practice [1], [3]. In this paper, we investigate low RF-complexity technologies for mmWave MIMO systems. Our contributions can be summarized as follows: i) We provide a review for the evolution of low RF-complexity technologies from microwave frequencies to mmWave frequencies, and highlight two promising technologies proposed recently, i.e., phased array based hybrid precoding (PAHP) [4] and lens array based hybrid precoding (LAHP) [5]. We give detailed overview for these two technologies including their principles, advantages, challenges, and recent results; ii) We propose a novel adaptive selecting network for LAHP with low hardware cost and power consumption. For data transmission, it can select beams like the traditional one, while for channel estimation, it can formulate the beamspace channel estimation as a sparse signal recovery problem and estimate the beamspace channel with considerably reduced pilot overhead; iii) We provide the sumrate and power efficiency comparisons between PAHP and LAHP in a practical outdoor mmWave MIMO system, where the channel estimation error and inter-cell interference are also included. Then, we draw some insights about how these two technologies can be deployed in practice. II. T RADITIONAL L OW RF-C OMPLEXITY T ECHNOLOGIES We first review two typical low RF-complexity technologies, i.e., antenna selection and analog beamforming. Antenna selection may be considered as the most classical low RFcomplexity technology for microwave MIMO systems. By contrast, analog beamforming is the most widely used low RFcomplexity technology for indoor mmWave communications. These two technologies can be regarded as the basics of PAHP and LAHP. 2 Amplifier loss [3]. More importantly, analog beamforming can only support single-stream transmission, which cannot be used in multi-stream or multi-user scenarios [3]. Data Streams RF Chain Baseband Signal Processing RF Chain N s £ N RF N RF < N Switches N (a) Amplifier Power Splitter Data Stream Baseband Signal Processing RF Chain Ns = 1 N RF = 1 Phase Shifters N (b) Fig. 1. Architectures of traditional low RF-complexity technologies: (a) antenna selection; (b) analog beamforming. A. Antenna selection As shown in Fig. 1 (a), the key feature of antenna selection [6] is that there is one selecting network between NRF RF chains and N antennas. Based on the channel state information (CSI), the target of antenna selection is to select NRF best antennas out of total N antennas for data transmission to maximize the achievable sum-rate [6]. An exciting result of antenna selection is that when the number of RF chains NRF is larger than the number of transmitted data streams Ns , the performance loss induced by antenna selection is negligible under independent identically distributed (IID) Rayleigh fading channels [6]. However, when channels are highly correlated, the achievable sum-rate of antenna selection will decrease drastically [6], as antenna selection incurs more channel information loss in this case. B. Analog beamforming As shown in Fig. 1 (b), the key idea of analog beamforming [7] is to use only one RF chain to transmit single data stream, and employ the phase shifter network to control the phases of original signals to maximize the array gain and effective signal-to-noise ratio (SNR). Building the beamforming vectors (include the relative amplitudes and phases applied to the different antenna elements to shape the signal strength at a specific direction in the far field) requires beam training, which involves an iterative and joint design between the transmitter and receiver. For example, in IEEE 802.11ad, a multi-resolution beamforming codebook (consists of several pre-defined beamforming vectors) is adopted to progressively refine the selected beamforming vectors [7]. The advantage of analog beamforming is that it only requires one RF chain, leading to quite low hardware cost and power consumption [7]. However, the analog beamforming can only adjust the phases of the signals, which means that all the elements of beamforming vector have the same amplitude. Such design constraint will incur some performance C. Can traditional low RF-complexity technologies be used in mmWave MIMO systems? Note that the outdoor mmWave MIMO channel is significantly different from the one at microwave frequencies. One of the most important differences is that the scattering of outdoor mmWave communications is usually limited. This is because that the wavelengths at mmWave frequencies are quite small compared to the obstacle size, leading to poor diffraction [3]. Moreover, though scattering occurs, it incurs significant 5-20 dB attenuation at mmWave frequencies [2], [5]. A study in New York has shown that the average number of paths in 28 GHz outdoor mmWave communications is only 2.4 [2]. This means that the outdoor mmWave MIMO channel is usually low-rank with high correlation in the spatial domain and sparse in the angular domain [5]. Based on these facts, we know that antenna selection is not appropriate for mmWave MIMO systems [2], since it suffers from serious performance loss with highly correlated channels. Moreover, although analog beamforming is developed for mmWave communications, it can only support singlestream transmission without multiplexing gains. This means that analog beamforming cannot fully exploit the potential of mmWave MIMO in spectrum efficiency [3], [5]. Next, we will investigate two promising low RF-complexity technologies proposed recently for mmWave MIMO systems, i.e., PAHP and LAHP. III. P HASED A RRAY BASED H YBRID P RECODING A. Principle Precoding is used to adjust the weights of transmitted signals to maximize the achievable sum-rate [3]. As shown in Fig. 2 (a), the conventional fully digital precoding can arbitrarily adjust the amplitudes and phases of the original signals. It can achieve multiplexing gains, and enjoys higher design freedom than analog beamforming. However, it requires one dedicated RF chain for each antenna, which brings unaffordable hardware cost and power consumption when the number of antennas is large. Hybrid precoding can be considered as a promising compromise between the optimal fully digital precoding and the low-cost analog beamforming [4], [5]. Its key idea is to divide the large-size digital precoder into a largesize analog beamformer (realized by the analog circuit) and a small-size digital precoder (requiring a small number of RF chains). PAHP is one of the realization of hybrid precoding, where the analog beamformer is realized by phase shifters. Assume there are Ns single-antenna users to be served. As shown in Fig. 2, the received signal vector y for Ns users in the downlink can be presented as y = HADs + n, (1) where H of size Ns × N , s of size Ns × 1, and n of size Ns × 1 denote the mmWave MIMO channel matrix, transmitted signal vector, and noise vector, respectively, A of size 3 N × NRF is the analog beamformer, and D of size NRF × Ns is the digital precoder. Note that PAHP has two architectures, i.e., full-PAHP [4] and sub-PAHP [8], as illustrated in Fig. 2 (b) and Fig. 2 (c), respectively. In full-PAHP, each RF chain is connected to all N antennas via phase shifters, and the analog beamformer A is a full matrix. It can achieve satisfying performance, but usually requires a large number of N NRF phase shifters (e.g., N NRF = 256 × 16 = 4096), together with the complicated power splitters/cominbers and signal/control lines. By contrast, in sub-PAHP, each RF chain is only connected to a subset of antennas, leading A to be a block diagonal matrix. Obviously, sub-PAHP can reduce the number of phase shifters from N NRF to N and avoid using power combiners [8]. Therefore, although sub-PAHP suffers from a loss in array gains by a factor of 1/NRF , it may be preferred in practice [8]. B. Advantages PAHP can achieve a better tradeoff between the hardware cost/power consumption and the sum-rate performance. It can significantly reduce the number of required RF chains from N (e.g., N = 256) to Ns (e.g., NRF = Ns = 16), leading to lower power consumption. Besides, as explained above, the outdoor mmWave MIMO channel matrix is usually lowrank [2]. This indicates that the maximum number of data streams that can be simultaneously transmitted by such channel is limited. Therefore, as long as the number of RF chains is larger than the rank of channel matrix, the small-size digital precoder is still able to fully achieve the multiplexing gains and obtain the near-optimal performance compared to the fully digital precoding [4], [8]. C. Challenges and recent results Optimal design of hybrid precoder: Maximizing the achievable sum-rate by designing the hybrid precoder P = AD is the main target of PAHP. However, this optimization problem imposes new challenges, since there are several non-convex hardware constraints on the analog beamformer A. For example, all the nonzero elements of A should share the same amplitude due to the constant modulus constraint on phase shifter. To this end, one feasible way is to approximate the original optimization problem as a convex one to obtain a near-optimal hybrid precoder with low complexity. Following this idea, some advanced schemes have been proposed recently. In [4], a spatially sparse scheme is proposed for single-user full-PAHP. It approximates the sumrate optimization problem as the one minimizing the distance between the optimal fully digital precoder and the hybrid precoder. Then, a variant of the orthogonal matching pursuit (OMP) algorithm [9] is developed to obtain the near-optimal hybrid precoder. In [10], full-PAHP is extended to multi-user scenario, where a two-stage multi-user scheme is proposed. In the first stage, the optimal analog beamformer is searched from a pre-defined codebook to maximize the desired signal power of each user. In the second stage, the classical zero forcing (ZF) precoder is used to cancel multi-user interference. In [8], a successive interference cancelation (SIC)-based scheme is proposed for sub-PAHP. It first decomposes the sum-rate optimization problem into a series of simple and convex subproblems, each of which only considers one sub-phased array. Then, inspired by the classical SIC multi-user signal detector, the near-optimal hybrid precoder for each sub-phased array is obtained in an one-by-one fashion. Channel estimation: The maximum gain of PAHP can be only achieved with perfect CSI, which is difficult to obtain in mmWave MIMO systems. Firstly, due to the lack of array gains before the establishment of the transmission link, the SNR for channel estimation in PAHP is quite low [3]. Secondly, the number of RF chains in PAHP is usually much smaller than the number of antennas. Therefore, we cannot directly observe the channel matrix like that in fully digital precoding. Two typical solutions have been proposed to solve this problem. The first divides the channel estimation problem into two steps. In the first step, the BS and users will perform beam training like analog beamforming [7] to determine A. In the second step, the effective channel matrix HA with smaller size Ns × NRF (NRF ≪ N ) is estimated by classical schemes, such as least squares (LS). The second solution is to exploit the low-rank characteristic of mmWave MIMO channel. Instead of estimating the effective channel matrix HA, it can directly obtain the complete channel matrix H with low pilot overhead. For example, an adaptive compressive sensing (CS) [9] based channel estimation scheme is proposed in [3]. It divides the total channel estimation problem into several sub-problems, each of which only considers one channel path. For each channel path, it first starts with coarse direction grids, and determines the direction of this path belonging to which grid by employing OMP algorithm. Then, the narrowed direction grids are used, and the direction of this path is further refined. Note that the first solution usually involves low complexity and is easy to implement. By contrast, the second solution can significantly reduce the pilot overhead, but usually involves higher complexity. IV. L ENS A RRAY BASED H YBRID P RECODING A. Principle Although full-PAHP can achieve the near-optimal performance with reduced number of RF chains, it usually requires a large number of high-resolution phase shifters, together with the complicated power splitters/combiners. Sub-PAHP can partly solve these problems, but it suffers from some performance loss due to the reduced array gains. These problems above can be solved by LAHP [5], another realization of hybrid precoding, where the analog beamformer is realized by lens array and selecting network as shown in Fig. 3 (a). By employing a lens array (an electromagnetic lens with directional energy focusing capability and a matching antenna array with elements located in the focal surface of the lens [5]), the signals from different directions (beams) can be concentrated on different antennas, and the spatial channel can be transformed to the beamspace channel. Mathematically, the lens array plays the role of a spatial discrete fourier transform (DFT) matrix U of size N × N , whose N columns 4 correspond to the orthogonal beamforming vectors of N predefined directions (beams) that cover the whole angular space. The system model of LAHP can be presented by ỹ = HUDs + n = H̃Ds + n, (2) where ỹ, s, and n of size Ns × 1 denote the received signal vector in the beamspace, transmitted signal vector, and noise vector, respectively, D of size N × Ns is the digital precoder, and the beamspace channel H̃ of size Ns × N is defined as H̃ = HU, whose N columns correspond to N orthogonal beams. B. Advantages LAHP can also achieve the near-optimal performance with low hardware cost/power consumption. This advantage comes from the fact that the beamspace channel H̃ at mmWave frequencies is sparse due to the limited scattering [2]. Therefore, similar to antenna selection, we can select only a small number of dominant beams to reduce the MIMO dimension as ỹ ≈ H̃r Dr s + n, where H̃r = H̃(:, l)l∈B is the dimension-reduced beamspace channel, B denotes the set of selected beams, and Dr of size |B| × Ns is the corresponding dimension-reduced digital precoder. As the dimension of Dr is much smaller than that of D in (2), LAHP can significantly reduce the number of required RF chains without obvious performance loss. Another advantage of LAHP is that the array gains can be always preserved by the low-cost lens array. Therefore, even though the simple selecting network is used, the satisfying performance can be still guaranteed [5]. different hardware architectures; ii) we need to estimate the sparse beamspace channel instead of the conventional lowrank spatial channel [5]. This problem can be solved by following two recently proposed solutions. The first one is to reduce the dimension of beamspace channel estimation problem. For example, in [13], a two-step channel estimation scheme is proposed. In the first step, the BS utilizes the selecting network to scan all the beams, and selects several strong beams. In the second step, LS is used to estimate the dimension-reduced beamspace channel. The second solution is to completely estimate the sparse beamspace channel by utilizing CS algorithms. However, if we use the traditional selecting network as shown in Fig. 3 (a) (each RF chain can only select one beam via one switch) to sample the beamspace channel with reduced number of pilots, the sensing matrix will be rank-deficient and CS algorithms cannot work [9]. To this end, we propose an adaptive selecting network for LAHP as shown in Fig. 3 (b), where each RF chain is connected to all antennas via switches. For data transmission, it can select beams like the traditional one, while for channel estimation, it can perform as a sensing matrix with randomly selected 0/1 elements, which has full rank [9]. Then, we can estimate the sparse beamspace channel via efficient CS algorithms such as OMP, and considerably reduce the pilot overhead1. Note that the extra power consumption and hardware cost incurred by the adaptive selecting network is limited, since the switches are easy to implement with low power consumption [15]. Finally, similar to PAHP, the first solution usually involves low complexity while the second one can considerably reduce the pilot overhead. C. Challenges and recent results V. P ERFORMANCE C OMPARISON Optimal design of beam selection: The performance of LAHP depends on beam selection, which aims to select |B| beams out of the total N beams to maximize the achievable sum-rate. The most intuitive beam selection scheme is exhaustive search. It has the optimal performance but prohibitively high complexity, which exponentially increases with the number of selected beams. This means that more efficient beam selection schemes should be designed. In [5], a magnitude maximization (MM) beam selection scheme is proposed, where several beams with large power are selected for data transmission. MM beam selection scheme is simple, but it only aims to preserve the power as much as possible without considering interference, leading to some performance loss. In [11], the authors propose a more efficient beam selection scheme by using the incremental algorithm developed from antenna selection. It selects |B| beams one by one. In each step, the beam with the greatest contribution to the achievable sum-rate is selected. In [12], an interferenceaware (IA) beam selection scheme with better performance is proposed. The key idea is to classify all users into two user groups according to the potential interference. For users with small interference, it directly selects the beams with large power, while for users with severe interference, the incremental algorithm is employed to search the optimal beams. In this section, we compare the performance of full-PAHP, LAHP with the proposed adaptively selecting network, and fully digital precoding. We first consider an outdoor multiuser mmWave MIMO in single-cell scenario, where the BS employs N = 256 antennas to simultaneously serve Ns = 16 single-antenna users. The widely used Saleh-Valenzuela multipath model is adopted to capture the characteristics of mmWave MIMO channels [2]. Each user has one LoS path and two NLoS paths. The gain of LoS path is normalized to 1, while the gain of each NLoS path is assumed to follow CN (0, 0.1). The directions of all paths of users are assumed to follow the IID uniform distribution within [−π/2, π/2]. Fig. 4 shows the comparison of achievable sum-rate against the SNR for data transmission. For full-PAHP with NRF = Ns = 16 RF chains, we use the adaptive CS based channel estimation scheme proposed in [3] (the number of initial direction grids is set as 1024) to estimate the spatial channel, and employ the two-stage multi-user scheme proposed in [10] (4-bit phase shifters are used) to transmit data. For LAHP also with NRF = Ns = 16 RF chains, we utilize the proposed adaptive selecting network to estimate the beamspace channel as illustrated above, and adopt the IA beam selection scheme [12] to transmit data. Finally, for Channel estimation: The channel estimation for LAHP is different from that for PAHP since: i) PAHP and LAHP have 1 In practice, the beamspace channel may not be sparse enough due to the power leakage effect, but this can be relieved following the idea in [14]. 5 1.8 90 Fully digital precoding, perfect channel Fully digital precoding, estimated channel Phased array based hybrid precoding, perfect channel Phased array based hybrid precoding, estimated channel Lens array based hybrid precoding, perfect channel Lens array based hybrid precoding, estimated channel Achievable sum−rate (bits/s/Hz) 70 Fully digital precoding Phased array based hybrid precoding Lens array based hybrid precoding 1.6 1.4 Power efficiency (bps/Hz/W) 80 60 50 40 30 1.2 1 0.8 0.6 0.4 20 0.2 10 0 0 0 5 10 15 20 25 SNR (dB) for data transmission Fig. 2. 5 10 15 20 25 30 Number of users Ns 30 Fig. 3. Comparison of sum-rate against the SNR for data transmission. Comparison of power efficiency against the number of users Ns . 80 Fully digital precoding Phased array based hybrid precoding Lens array based hybrid precoding fully digital precoding with NRF = N = 256 RF chains, the classical LS channel estimation scheme and ZF precoding scheme are adopted. For fair comparison, we set SNR as 20 dB and use 96 pilots for all of the three adopted channel estimation schemes. From Fig. 4, we observe that, although the number of RF chains is significantly reduced, both PAHP and LAHP can achieve the sum-rate performance close to that of fully digital precoding. This is due to the fact that PAHP can fully exploit the low-rank characteristic of mmWave channels in the spatial domain, while LAHP can benefit from the sparse characteristic of mmWave channels in the beamspace (angular domain). Since the lens array plays the role of spatial DFT, the beamspace channel and the spatial channel are essentially equivalent but expressed in different forms, just like the same signal can be expressed in the time domain and equivalently in the frequency domain. Moreover, Fig. 4 also shows that with the perfect channel, PAHP outperforms LAHP by about 2 dB, since the phase shifter has higher design freedom than switch. Finally, Fig. 4 shows that LAHP is more robust to the channel estimation error, since only the dimension-reduced beamspace channel is effective for data transmission and a little inaccurate channel may still achieve satisfying performance. Fig. 5 shows the comparison of power efficiency. We define the power efficiency η as η = R/ (PT + PH ) [11], where R is the achievable sum-rate, PT is the transmission power, which can be set as PT = 2.5W (34 dBm) for outdoor mmWave MIMO in a small cell [2], PH is the power consumed by hardware architecture. For full-PAHP as shown in Fig. 2 (b), we have PT = N PA + NRF N PPS + NRF PSP + N PCO + NRF PRF , where PA , PPS , PSP , PCO , and PRF are the power consumed by amplifier, phase shifter, power splitter, power combiner, and RF chain, respectively. For LAHP with the proposed adaptive selecting network as shown in Fig. 3 (b), we have PT = N PA + NRF N PSW + NRF PSP + N PCO + NRF PRF , where PSW is the power consumed by switch. Finally, for fully digital precoding as shown in Fig. 2 (a), we have Achievable sum−rate (bits/s/Hz) 70 60 50 40 30 20 10 0 0 5 10 15 20 25 30 35 40 45 50 SNR (dB) for data transmission Fig. 4. Comparison of sum-rate in multi-cell scenario. PT = N PA + N PRF . In this paper, some referenced values2 are adopted as PA = 20mW [15], PSP = PCO = 10mW [15], PRF = 250mW [3], PPS = 30mW for 4-bit phase shifter [15], and PSW = 5mW [15]. Fig. 5 shows that both PAHP and LAHP can achieve much higher power efficiency than fully digital precoding when the number of users is not large (e.g., Ns ≤ 8). However, when Ns > 12, LAHP still enjoys high power efficiency, but PAHP performs even worse than fully digital precoding. This conclusion is contrary to the widely accepted misconception that PAHP is more power efficient than fully digital precoding, but consistent with a very recent analysis in [11]–[13]. It can be explained by the fact that as Ns grows, the number of phase shifters required by PAHP increases rapidly. As a result, the power consumption of phase shifters will be huge, even higher than that of RF chains. Next, we extend the performance comparison to multi-cell 2 Note that the power consumption of RF modules usually presents high variability, which depends on the specific implementation type and performance requirement [15]. In this paper, we just adopt the conservative (high) values, and the results in Fig. 5 can be considered as the lower bounds of the power efficiencies in practice. 6 scenario as shown in Fig. 6, where the number of cells is set as 2 for simplicity but without loss of generality. Each cell has one BS with N = 256 antennas to serve K = 16 users with the same transmission power. The channels between the BS and the users in its own cell, and the channels between the BS and the users in the neighbor cell, are generated following the same model in Fig. 4. We assume that no cooperation between BSs exists, and each BS only knows the channels of the users in its own cell. From Fig. 6, we observe that the achievable sum-rates of all technologies will not grow without bound when the SNR for data transmission increases. This is due to the fact that high SNR also incurs high inter-cell interference. However, Fig. 6 still shows that both PAHP and LAHP can achieve the near-optimal performance, even in the multi-cell scenario with inter-cell interference. VI. C ONCLUSIONS In this paper, we have introduced two promising low RFcomplexity technologies for mmWave MIMO with large antenna array, i.e., PAHP and LAHP, in detail. We have also proposed an adaptive selecting network for LAHP with low hardware cost and power consumption, which can formulate the beamspace channel estimation as a sparse signal recovery problem, and considerably reduce the pilot overhead. Finally, we have provided the complete and systematic performance comparison between PAHP and LAHP. It shows that PAHP achieves higher achievable sum-rate than LAHP when the channel is perfectly known, but LAHP is more robust to the channel estimation error. It also shows that LAHP enjoys higher power efficiency than PAHP, since the phase shifter network is replaced by the low-cost lens array and switches. Besides the discussions above, there are still some open issues on low RF-complexity technologies for outdoor mmWave MIMO systems. For example, most of the existing low RFcomplexity technologies are designed for the narrowband and time-invariant channels. However, due to the large bandwidth and the high frequency, the mmWave MIMO channels are more likely to be broadband and time-varying, which incurs new challenges. Take the PAHP for example, “broadband” means that the analog beamformer cannot be adaptively adjusted according to the frequency, leading to more difficulties in signal processing design, while “time-varying” means that we need to re-estimate the channel and re-compute the hybrid precoder frequently, leading to high pilot overhead and computational complexity. Therefore, designing low RF-complexity technologies for broadband time-varying channels will be an urging problem to solve. [5] J. Brady, N. Behdad, and A. Sayeed, “Beamspace MIMO for millimeterwave communications: System architecture, modeling, analysis, and measurements,” IEEE Trans. Ant. and Propag., vol. 61, no. 7, pp. 3814– 3827, July 2013. [6] A. F. Molisch and M. Z. Win, “MIMO systems with antenna selection,” IEEE Microwave Mag., vol. 5, no. 1, pp. 46–56, Mar. 2004. [7] S. Hur et al., “Millimeter wave beamforming for wireless backhaul and access in small cell networks,” IEEE Trans. Commun., vol. 61, no. 10, pp. 4391–4403, Oct. 2013. [8] X. Gao et al., “Energy-efficient hybrid analog and digital precoding for mmWave MIMO systems with large antenna arrays,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 998–1009, Apr. 2016. [9] W. U. Bajwa et al., “Compressed channel sensing: A new approach to estimating sparse multipath channels,” Proc. IEEE, vol. 98, no. 6, pp. 1058–1076, June 2010. [10] A. Alkhateeb, G. Leus, and R. W. Heath, “Limited feedback hybrid precoding for multi-user millimeter wave systems,” IEEE Trans. Wireless Commun., vol. 14, no. 11, pp. 6481–6494, Nov. 2015. [11] P. Amadori and C. Masouros, “Low RF-complexity millimeter-wave beamspace-MIMO systems by beam selection,” IEEE Trans. Commun., vol. 63, no. 6, pp. 2212–2222, June 2015. [12] X. Gao et al., “Near-optimal beam selection for beamspace mmWave massive MIMO systems,” IEEE Commun. Lett., vol. 20, no. 5, pp. 1054– 1057, May 2016. [13] J. Hogan and A. Sayeed, “Beam selection for performance-complexity optimization in high-dimension MIMO systems,” in Proc. CISS, Mar. 2016, pp. 337–342. [14] G. Taubock et al., “Compressive estimation of doubly selective channels in multicarrier systems: Leakage effects and sparsity-enhancing processing,” IEEE J. Sel. Top. Signal Process., vol. 4, no. 2, pp. 255–271, Apr. 2010. [15] R. Méndez-Rial et al., “Hybrid MIMO architectures for millimeter wave communications: Phase shifters or switches?” IEEE Access, vol. 4, pp. 247–267, Jan. 2016. PLACE PHOTO HERE PLACE PHOTO HERE R EFERENCES [1] S. Mumtaz, J. Rodriguez, and L. Dai, MmWave Massive MIMO: A Paradigm for 5G. Academic Press, Elsevier, 2016. [2] T. S. Rappaport et al., “Millimeter wave mobile communications for 5G cellular: It will work!” IEEE Access, vol. 1, pp. 335–349, May 2013. [3] R. W. Heath et al., “An overview of signal processing techniques for millimeter wave MIMO systems,” IEEE J. Sel. Top. Signal Process., vol. 10, no. 3, pp. 436–453, Apr. 2016. [4] O. El Ayach et al., “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014. PLACE PHOTO HERE Xinyu Gao (S’14) received the B.E. degree of Communication Engineering from Harbin Institute of Technology, Heilongjiang, China in 2014. He is currently working towards Ph.D. degree in Electronic Engineering from Tsinghua University, Beijing, China. His research interests include massive MIMO and mmWave communications. He has received the national scholarship in 2015 and the IEEE WCSP Best Paper Award in 2016. Linglong Dai (M’11-SM’14) received the Ph.D. degree from Tsinghua University, Beijing, China, in 2011. He is currently an Associate Professor of Tsinghua University. His current research interests include massive MIMO, millimeter-wave communications, multiple access, and sparse signal processing. He has received five conference Best Paper Awards (IEEE ICC 2013, IEEE ICC 2014, WCSP 2016, IEEE ICC 2017, and IEEE VTC 2017 Fall). Akbar M. Sayeed (F’12) received the Ph.D. degree from the University of Illinois at UrbanaChampaign, USA, in 1996. He is a Professor of Electrical and Computer Engineering at the University of Wisconsin-Madison. His research interests include wireless communications, statistical signal processing, communication and information theory, wireless channel modeling, time-frequency analysis, and development of basic theory, system architectures, and prototypes for new wireless technologies.
7cs.IT
Better s-t-Tours by Gao Trees arXiv:1511.05514v1 [cs.DM] 17 Nov 2015 Corinna Gottschalk RWTH Aachen Jens Vygen University of Bonn November 17, 20151 Abstract We consider the s-t-path TSP: given a finite metric space with two elements s and t, we look for a path from s to t that contains all the elements and has minimum total distance. We improve the approximation ratio for this problem from 1.599 to 1.566. Like previous algorithms, we solve the natural LP relaxation and represent an optimum solution x∗ as a convex combination of spanning trees. Gao showed that there exists a spanning tree in the support of x∗ that has only one edge in each narrow cut (i.e., each cut C with x∗ (C) < 2). Our main theorem says that the spanning trees in the convex combination can be chosen such that many of them are such “Gao trees”. keywords: traveling salesman problem, s-t-path TSP, approximation algorithm, spanning tree 1 Introduction The traveling salesman problem (TSP) is one of the best-known NP-hard problems in combinatorial optimization. In this paper, we consider the s-t-path variant: Given a finite set V (of cities), two elements s, t ∈ V and c : V × V → R≥0 , the goal is to find a sequence v1 , . . . , vN containing PN −1 all cities and with v1 = s and vN = t, minimizing i=1 c(vi , vi+1 ). An equivalent formulation requires that every city is visited exactly once, but then we need to assume that c satisfies the triangle inequality. We work with the latter formulation (to obtain such an instance from a general function c, compute the metric closure). The special case where s = t is the well-known metric TSP; but in this paper we assume s 6= t. The classical algorithm by Christofides [1976] computes a minimum-cost spanning tree (V, S) and then does parity correction by adding a minimum-cost matching on the vertices whose degree in S has the wrong parity. While Christofides’ algorithm is still the best known approximation algorithm for metric TSP (with ratio 32 ), there have recently been improvements for special cases and variants (see e.g. Vygen [2012] for a survey), including the s-t-path TSP. 1 This work was done during the trimester program on combinatorial optimization at the Hausdorff Institute for Mathematics in Bonn. 1 1.1 Previous Work For the s-t-path TSP, Christofides’ algorithm has only an approximation ratio of 35 as shown by Hoogeveen [1991]. An, Kleinberg and Shmoys [2015] were the first to improve on this and obtained √ 1+ 5 an approximation ratio of 2 ≈ 1.618. They first solve the natural LP relaxation (essentially proposed by Dantzig, Fulkerson and Johnson [1954]) and represent an optimum solution x∗ as a convex combination of spanning trees. This idea, first proposed by Held and Karp [1970], was exploited earlier for different TSP variants by Asadpour et al. [2010] and Oveis Gharan, Saberi and Singh [2011]. Given this convex combination, An, Kleinberg and Shmoys [2015] do parity correction for each of the contributing trees and output the best of these solutions. Sebő [2013] improved the analysis of this best-of-many Christofides algorithm and obtained the approximation ratio 58 . Gao [2015] gave a unified analysis. Vygen [2015] suggested to “reassemble” the trees: starting with an arbitrary convex combination of spanning trees, he computed a different one, still representing x∗ , that avoids certain bad local configurations. This led to the slightly better approximation ratio of 1.599. In this paper, we will reassemble the trees more systematically to obtain a convex combination with strong global properties. This will enable us to control the cost of parity correction much better, leading to an approximation ratio of 1.566. This also proves an upper bound of 1.566 on the integrality ratio of the natural LP. The only known lower bound is 1.5. Sebő and Vygen [2014] proved that the integrality ratio is indeed 1.5 for the graph s-t-path TSP, i.e., the special case of graph metrics (where c(v, w) is the distance from v to w in a given unweighted graph on vertex set V ). Gao [2013] gave a simpler proof of this result, which inspired our work: see Section 1.4. 1.2 Notation and Preliminaries Throughout this paper, V denotes the given set of cities, n := |V |, and E denotes the set of edges of the complete graph on V . Moreover, c : V × V → R≥0 is the given metric. For any U ⊆ V we write E[U ] for the set of edges with both endpoints in U and δ(U ) for the set of edges with exactly one endpoint in U ; moreover, δ(v) := δ({v}) for v ∈ V . If F ⊆ E and U ⊆ V , we denote by (V, F )[U ] the subgraph (U, F ∩ E[U ]) of (V, F ) induced by U . For x ∈ RE ≥0 we P P F write c(x) := e={v,w}∈E c(v, w)xe and x(F ) := e∈F xe for F ⊆ E. Furthermore, χ ∈ {0, 1}E denotes the characteristic vector of a subset F ⊆ E, and c(F ) := c(χF ). For F ⊂ E and f ∈ E, we write F + f and F − f for F ∪ {f } and F \ {f }, respectively. For T ⊆ V with |T | even, a T -join is a set J ⊆ E for which |δ(v) ∩ J| is odd if and only if v ∈ T . Edmonds [1965] proved that a minimum weight T -join can be computed in polynomial time. Moreover, the minimum cost of a T -join is the minimum over c(y) for y in the T -join polyhedron {y ∈ RE ≥0 : y(δ(U )) ≥ 1 ∀U ⊂ V with |U ∩ T | odd} as proved by Edmonds and Johnson [1973]. We are going to use that the cost of any vector in the T -join polyhedron is an upper bound for the minimum cost of a T -join. 2 To obtain a solution for the s-t-path TSP, it is sufficient to compute a connected multi-graph with vertex set V in which exactly s and t have odd degree. We call such a graph an {s, t}-tour. As an {s, t}-tour contains an Eulerian walk from s to t, we can obtain a Hamiltonian s-t-path (i.e. an s-t-path with vertex set V ) by traversing the Eulerian walk and shortcutting when the walk encounters a vertex that has been visited already. Since c is a metric, it obeys the triangle inequality. Thus, the resulting path is not more expensive than the multigraph. By S we denote the set of edge sets of spanning trees in (V, E). For S ∈ S, TS denotes the set of vertices whose degree has the wrong parity, i.e., even for s or t and odd for v ∈ V \ {s, t}. Christofides’ algorithm computes an S ∈ S with minimum c(S) and adds a TS -join J with minimum c(J). 1.3 Best-of-Many Christofides Like An, Kleinberg and Shmoys [2015], we begin by solving the natural LP relaxation: min c(x) subject to x(δ(U )) x(δ(U )) x(δ(v)) x(δ(v)) xe ≥ ≥ = = ≥ (∅ = 6 U ⊂ V, |U ∩ {s, t}| even) (∅ = 6 U ⊂ V, |U ∩ {s, t}| odd) (v ∈ V \ {s, t}) (v ∈ {s, t}) (e ∈ E) 2 1 2 1 0 (1) whose integral solutions are precisely the incidence vectors of the edge sets of the Hamiltonian s-t-paths in (V, E). This LP can be solved in polynomial time (either by the ellipsoid method (Grötschel, Lovász and Schrijver [1981]) or an extended formulation). Let x∗ be an optimum solution. Then, x∗ (in factPevery feasible solution) can be written as convex combination of S ∗ spanning P trees, i.e. as x = S∈S pS χ , where p is a distribution on S, i.e., pS ≥ 0 for all S ∈ S and S∈S pS = 1. By Carathéodory’s theorem, we can assume that pS > 0 for less than n2 spanning trees (V, S). Such spanning trees and numbers pS can be computed in polynomial time, using either the ellipsoid method or the splitting-off technique (cf. Genova and Williamson [2015]). The best-of-many Christofides algorithm does the following: Compute an optimum solution x∗ for P (1) and obtain a distribution p with x∗ = S∈S pS χS as above. For each tree (V, S) with pS > 0, ˙ S ) is an {s, t}-tour. We compute a minimum weight TS -join JS . Then, the multigraph (V, S ∪J output the best of these. We will fix x∗ henceforth. An important concept in An, Kleinberg and Shmoys [2015] and the subsequent works are the so-called narrow cuts, i.e., the cuts C = δ(U ) with x∗ (C) < 2. We denote by C the set of all narrow cuts. We are going to exploit their structure as well. 3 Lemma 1 (An, Kleinberg and Shmoys [2015]) The narrow cuts form a chain: there are sets {s} = U0 ⊂ U1 ⊂ · · · ⊂ Uℓ−1 ⊂ Uℓ = V \ {t} so that C = {δ(Ui ) : i = 0, . . . , ℓ}. These sets can be computed in polynomial time. We number the narrow cuts C = {C0 , C1 . . . , Cℓ } with Ci = δ(Ui ) (i = 0, . . . , ℓ). 1.4 Gao trees Our work was inspired by the following idea of Gao [2013]: Theorem 2 (Gao [2013]) There exists a spanning tree S ∈ S with x∗e > 0 for all e ∈ S and |C ∩ S| = 1 for all C ∈ C. In fact, Gao [2013] showed this for any vector x ∈ RE 6 U ⊂ V and ≥0 with x(δ(U )) ≥ 1 for all ∅ = x(δ(U )) ≥ 2 for all ∅ = 6 U ⊂ V with |U ∩ {s, t}| even. For graph s-t-path TSP, one uses only variables corresponding to edges of the given graph. Then every spanning tree has cost n − 1. The approximation guarantee of 32 then follows from the fact that for a tree (V, S) with |S ∩ C| = 1 for all C ∈ C, the vector 21 x∗ is in the TS -join polyhedron. But, as shown by Gao [2015], for the general s-t-path TSP there may be no tree as in Theorem 2 whose cost is bounded by the LP value. Let us call a tree S ∈ S a local Gao tree at C if |C ∩ S| = 1. We call S a global Gao tree if it is a local Gao tree at every narrow cut. An, Kleinberg and Shmoys [2015] and Sebő [2013] observed that for every distribution p with P x∗ = S∈S pS χS and every narrow cut C ∈ C, at least a 2 − x∗ (C) fraction of the trees will be local Gao trees at C. However, in general none of these trees will be a global Gao tree. 1.5 Our contribution Our main contribution is a new structural result: Starting from an arbitrary distribution of trees representing x∗ , we can compute a new distribution in which a sufficient number of trees are Gao trees simultaneously for all sufficiently narrow cuts. For example, if x∗ (C) = 32 for all C ∈ C, at least half of our new distribution will be made of global Gao trees. Here is our main structure theorem: Theorem solution x∗ of (1), there are S1 , . . . , Sr ∈ S and p1 , . . . , pr > 0 Pr 3 For every feasible Pr ∗ with j=1 pj = 1 such that x = j=1 pj χSj and for every C ∈ C there exists a k ∈ {1, . . . , r} P with kj=1 pj ≥ 2 − x∗ (C) and |C ∩ Sj | = 1 for all j = 1, . . . , k. Note that this result immediately implies Theorem 2, simply by taking S1 . However, we do not know whether such a distribution can be computed in polynomial time (or even whether r can be polynomially bounded). Therefore we will start with an arbitrary distribution p 4 and round the coefficients down to integral multiples of nǫ3 for a sufficiently small constant ǫ > 0. P 3 This way we will get a vector x close to x∗ that we can write as x = rj=1 1r χSj , where r ≤ nǫ and Sj ∈ S for j = 1, . . . , r. We will work with x in the following and consider the rounding error in Section 3. We will show that x ∈ RE ≥0 satisfies the properties x(δ(s)) = x(δ(t)) = 1 and x∗ (F ) − ǫ ≤ x(F ) ≤ x∗ (F ) + ǫ for all F ⊆ E. (2) Theorem 4 Given S1 , . . . , Sr ∈ S, a feasible solution x∗ of (1) and ǫ ≥ 0 such that x = Pr 1 Pr Sj satisfies (2), we can find Ŝ , . . . , Ŝ ∈ S in polynomial time such that x = 1 Ŝj 1 r j=1 χ j=1 χ , r r and for every C ∈ C there exists an h ∈ {1, . . . , r} with hr ≥ 2 − x∗ (C) − ǫ and |C ∩ Ŝj | = 1 for all j = 1, . . . , h. P Sj = Note this theorem implies Theorem 3, by letting ǫ = 0 and x = x∗ = S∈S pS χ Pr that 1 1 Sj where all pS are integer multiples of r for some (possibly exponentially large) natural j=1 r χ number r and taking rpS copies of each S ∈ S. This paper is organized as follows: In Section 2, we will prove Theorem 4. In Section 3, we will analyze the best-of-many Christofides algorithm on the resulting distribution. Finally, in Section 4 we provide the computations needed to obtain the approximation guarantee of 1.566. 2 Proof of the Structure Theorem P We will prove Theorem 4 by starting with arbitrary trees S1 , . . . , Sr ∈ S with x = 1r rj=1 χSj satisfying (2), and successively exchange a pair of edges in two of the trees. We will first satisfy the properties for the first tree S1 , then for S2 , and so on. For each Sj , we will work on the narrow cuts C1 , . . . , Cℓ−1 in this order; note that |C0 ∩ Sj | = |Cℓ ∩ Sj | = 1 always holds for all j = 1, . . . , r due to the first property in (2). In the following we write θi := ⌈r(2 − x∗ (Ci ) − ǫ)⌉ for i = 0, . . . , ℓ. Note that θ0 ≥ θi for all i = 1, . . . , ℓ because x∗ (Ci ) ≥ 1 = x∗ (C0 ) = x∗ (δ(s)). Our goal is to obtain |Sj ∩ Ci | = 1 whenever j ≤ θi . Note that |Sj ∩ Ci | = 1 implies that (V, Sj )[Ui ] is connected, and we will first obtain this weaker property by Lemma 9, before obtaining |Sj ∩ Ci | = 1 by Lemma 10. We need a few preparations. P Lemma 5 Let S1 , . . . , Sr ∈ S and ǫ ≥ 0 such that x = 1r rj′ =1 χSj′ satisfies (2). Let 0 ≤ h < i ≤ ℓ and 1 ≤ j ≤ r with j ≤ θh and j ≤ θi . Let M = Ui \ Uh or M = Ui . Then there exists an index k ≥ j such that (V, Sk )[M ] is connected. 5 Proof: Assume the above is not true. Then |E[M ] ∩ Sj ′ | ≤ |M | − 1 for all j ′ < j, and |E[M ] ∩ Sj ′ | ≤ |M | − 2 for j ′ ≥ j. Therefore, P x(E[M ]) = 1r rj′ =1 |E[M ] ∩ Sj ′ |  ≤ 1r (j − 1)(|M | − 1) + (r − j + 1)(|M | − 2) = |M | − 2 + j−1 r .  P ∗ ∗ On the other hand, using (2) we have x(E[M ]) + ǫ ≥ x∗ (E[M ]) = 21 v∈M x (δ(v)) − x (δ(M )) . If M = Ui \Uh , this is equal to |M |− 21 (x∗ (Ch )+x∗ (Ci ))+x∗ (Ch ∩Ci ) ≥ |M |− 21 (x∗ (Ch )+x∗ (Ci )). If M = Ui , this is equal to |M | − 21 − 12 x∗ (Ci ) = |M | − 12 (1 + x∗ (Ci )). j−1 ∗ Now, j ≤ θi and j ≤ θh implies 1 ≤ x∗ (Ci ) < 2 − j−1 r − ǫ and x (Ch ) < 2 − r − ǫ. Thus, x(E[M ]) > −ǫ + |M | − 12 (2 − j−1 r −ǫ+2− j−1 r a contradiction. − ǫ) = |M | − 2 + j−1 r , ✷ P Lemma 6 Let S1 , . . . , Sr ∈ S and ǫ ≥ 0 such that x = 1r rj′ =1 χSj′ satisfies (2). Let 0 ≤ h < i ≤ ℓ and 1 ≤ j ≤ r with j ≤ θh and j ≤ θi . Then there exists an index k ≥ j such that Sk ∩ Ch ∩ Ci = ∅. ′ ∗ Proof: Using first (2) and then x∗ (Ci′ ) < 2 − j−1 r − ǫ for i ∈ {h, i} and x (U ) ≥ 2 for |U ∩ {s, t}| even, we obtain  x(Ch ∩ Ci ) − ǫ ≤ x∗ (Ch ∩ Ci ) = 21 x∗ (Ch ) + x∗ (Ci ) − x∗ (δ(Ui \ Uh )) <2− j−1 r −ǫ−1= r−j+1 r P Therefore, 1r rj′ =1 |Sj ′ ∩ Ch ∩ Ci | = x(Ch ∩ Ci ) < edge in Ch ∩ Ci . − ǫ. r−j+1 , r i.e. at most r − j trees can contain an ✷ P Lemma 7 Let S1 , . . . , Sr ∈ S and ǫ ≥ 0 such that x = 1r rj′ =1 χSj′ satisfies (2). Let 1 ≤ i ≤ ℓ−1 and j ≤ θi with |Ci ∩ Sj | ≥ 2. Then there exists an index k > θi with |Ci ∩ Sk | = 1. P Proof: Suppose there exists no such k. Then we get rx(Ci ) = rj′ =1 |Ci ∩ Sj ′ | ≥ (r − θi + 1)2 + θi − 1 = 2r − ⌈r(2 − x∗ (Ci ) − ǫ)⌉ + 1 > 2r − r(2 − x∗ (Ci ) − ǫ) = r(x∗ (Ci ) + ǫ), which is a contradiction to (2). ✷ Lemma 8 Let 1 ≤ i ≤ ℓ − 1 and M ⊆ Ui and Sj , Sk ∈ S such that (V, Sj )[M ] is disconnected and (V, Sk )[M ] is connected and |Sj ∩ δ(Ui \ M )| ≤ 1. Then there exist edges e ∈ Sj and f ∈ Sk such that Sj − e + f ∈ S and Sk + e − f ∈ S and e ∈ / E[Ui ] and f ∈ E[M ]. 6 Ui M A1 B1 e F B2 A2 Y Bq Aq Ci Figure 1: Tree Sj in green, Sk in red/bold. In this example one can choose p = 1. Note that e could belong to Ci (but not to δ(Ui \ M )). Proof: An illustration of the following proof can be found in Figure 1. Let A1 , . . . , Aq be the vertex sets of the Sqconnected components of (V, Sj )[M ]; note that q ≥ 2. Let F := Sk ∩ p=1 (δ(Ap ) \ δ(M )) be the set of edges of Sk between the sets A1 , . . . , Aq . Note that F ⊆ E[M ]. For p = 1, . . . , q let Bp be the set of vertices reachable from Ap in (V, Sk \ F ). Trivially, Ap ⊆ Bp for all p, and {B1 , . . . , Bq } is a partition of V because (V, Sk )[M ] is connected and (V, Sk ) is a tree. Let Y be the union of the edge sets of the unique v-w-paths in Sj for all v, w ∈ M . Note that Y ⊆ E[V \ (Ui \ M )] because |Sj ∩ δ(Ui \ M )| ≤ 1. Claim: There exists an index p ∈ {1, . . . , q} and an edge e ∈ Y ∩ δ(Bp ) such that for every p′ ∈ {1, . . . , q} \ {p} the path in Sj from Ap to Ap′ contains e. To prove the Claim, note that (V, Y ) consists of a treeSand possibly isolated vertices. Choose an arbitrary root z in this tree and take an edge e ∈ Y ∩ qp′ =1 δ(Bp′ ) with maximum distance from z. Let D := {v ∈ V : e is on the z-v-path in (V, Y )}. We will show that D ∩ M = Ap for some p ∈ {1, . . . , q}. This will immediately imply that p and e satisfy the properties of the Claim. Observe that D ∩ M 6= ∅. Let v ∈ D ∩ M and p such that v ∈ Ap . Since (V, Sj )[Ap ] is connected, this implies Ap ⊆ D. (D, Y ∩ E[D]) is a tree and by the choice of e, it contains no edge from S q p′ =1 δ(Bp′ ). Therefore, D ⊆ Bp and hence, D ∩ M = Ap . The Claim is proved. Now, take an index p and an edge e as in the Claim. Consider the path P in Sk that connects the endpoints of e. Since e ∈ δ(Bp ), P has an edge f ∈ δ(Bp ) ∩ Sk = δ(Ap ) ∩ F . Thus, (V, Sk + e − f ) 7 is a tree. The path in Sj that connects the endpoints of f contains e by the Claim. Thus, (V, Sj − e + f ) is a tree. We have f ∈ E[M ] since f ∈ F , and e 6∈ E[Ui ] as e ∈ Y ∩ δ(Bp ). ✷ Now we are ready to prove our main lemmas: P Lemma 9 Let S1 , . . . , Sr ∈ S and ǫ ≥ 0 such that x = 1r rj′ =1 χSj′ satisfies (2). Let 1 ≤ i ≤ ℓ−1 such that j ≤ θi and |Sj ∩ Ch | = 1 for all h < i with j ≤ θh . Then we can find Ŝ1 , . . . , Ŝr ∈ S in P polynomial time such that x = 1r rj′ =1 χŜj′ and Ŝj ′ = Sj ′ for all j ′ < j and |Ŝj ∩ Ch | = 1 for all h < i with j ≤ θh , and (V, Ŝj )[Ui ] is connected. Proof: Assume that (V, Sj )[Ui ] is disconnected, i.e., |Sj ∩ E[Ui ]| < |Ui | − 1. Let h be the largest index smaller than i with j ≤ θh . Such an index must exist because θ0 ≥ θi ≥ j. Case 1: |Sj ∩ E[Ui \ Uh ]| < |Ui \ Uh | − 1. Let M := Ui \ Uh . Note that |Sj ∩ δ(Ui \ M )| = |Sj ∩ Ch | = 1. Since (V, Sj )[M ] is not connected, by Lemma 5 there exists an index k > j such that (V, Sk )[M ] is connected. Now we apply Lemma 8 and obtain two trees Ŝj := Sj − e + f and Ŝk := Sk + e − f with e ∈ / E[Ui ] and f ∈ E[M ]. We have |Ŝj ∩ E[Ui ]| = |Sj ∩ E[Ui ]| + 1 and |Ŝj ∩ Ch′ | ≤ |Sj ∩ Ch′ | for all h′ ≤ h and hence |Ŝj ∩ Ch′ | = 1 for h′ ≤ h with j ≤ θh′ . Note that j > θi′ for all h < i′ < i, so a new edge f in such cuts Ci′ does no harm. We replace Sj and Sk by Ŝj and Ŝk and leave the other trees unchanged. If (V, Ŝj )[Ui ] is still not connected, we iterate. Case 2: |Sj ∩ E[Ui \ Uh ]| = |Ui \ Uh | − 1. (V, Sj )[Uh ] is connected since |Sj ∩ Ch | = 1. Moreover, (V, Sj )[Ui \Uh ] is connected, but (V, Sj )[Ui ] is disconnected. Therefore, Sj must contain an edge in Ci ∩Ch and Sj ∩Ch ⊂ Sj ∩Ci and (V, Sj )[Ui ] has exactly two connected components: Uh and Ui \ Uh . We will illustrate the following in Figure 2. Case 2a: h > 0. Let g be the largest index smaller than h with j ≤ θg . Set M := Ui \ Ug . Note that |Sj ∩ δ(Ui \ M )| = |Sj ∩ Cg | = 1. By Lemma 5 there exists an index k ≥ j with (V, Sk )[M ] connected. As (V, Sj )[M ] is not connected, we have k > j. Case 2b: h = 0. Set M := Ui . Note that |Sj ∩ δ(Ui \ M )| = |Sj ∩ δ(∅)| = 0. By Lemma 5, there exists an index k ≥ j such that (V, Sk )[M ] is connected. Since (V, Sj )[M ] is not connected, k > j. Note that in both cases 2a and 2b, (V, Sj )[M ] is disconnected. Now we apply Lemma 8 and obtain two trees Ŝj := Sj − e + f and Ŝk := Sk + e − f with e ∈ / E[Ui ] and f ∈ E[M ]. We replace Sj and Sk by Ŝj and Ŝk and leave the other trees unchanged. Then (V, Sˆj )[Ui ] is connected. We have |Ŝj ∩ Ch′ | = 1 for all h′ < h with j ≤ θh′ , but we may have |Ŝj ∩ Ch | = 2. Assume |Ŝj ∩ Ch | = 2 (otherwise we are done). Then Ŝj ∩ Ch ∩ Ci = Sj ∩ Ch ∩ Ci = Sj ∩ Ch , and this set contains precisely one edge ê = {v, w} (where v ∈ Uh and w ∈ V \ Ui ). By Lemma 6 there exists an index k′ > j with Ŝk′ ∩ Ch ∩ Ci = ∅. 8 M connected in Sj v w ê W f Ch Ci Figure 2: Tree Sj in green, Ŝk′ in red/bold. The dashed edge f is added to Sj by applying Lemma 8. Let W be the set of vertices reachable from w in (V, Ŝj \ Ci ). Since (V, Ŝj )[Ui ] is connected, Ŝj ∩ δ(W ) = {ê}. The unique path in (V, Ŝk′ ) from v to w contains at least one edge fˆ ∈ δ(W ). ˆ ˆ Note that fˆ ∈ / Ch by the choice of Ŝk′ . We replace Ŝj and Ŝk′ by Ŝj := Ŝj − ê + fˆ and Ŝk′ := ˆ ˆ Ŝk′ + ê − fˆ. Then (V, Ŝj )[Ui ] is still connected and |Ŝj ∩ Ch′ | = 1 for all h′ < i with j ≤ θh′ . ✷ P Lemma 10 Let S1 , . . . , Sr ∈ S and ǫ ≥ 0 such that x = 1r rj′ =1 χSj′ satisfies (2). Let 1 ≤ i ≤ ℓ − 1 and j ≤ θi such that (V, Sj )[Ui ] is connected and |Sj ∩ Ch | = 1 for all h < i with j ≤ θh . P Then we can find Ŝ1 , . . . , Ŝr ∈ S in polynomial time such that x = 1r rj′ =1 χŜj′ and Ŝj ′ = Sj ′ for all j ′ < j and |Ŝj ∩ Ch | = 1 for all h ≤ i with j ≤ θh . Proof: Suppose |Sj ∩ Ci | ≥ 2. Then by Lemma 7 there exists an index k > θi with |Sk ∩ Ci | = 1. We will swap a pair of edges, reducing |Sj ∩ Ci | and increasing |Sk ∩ Ci | while maintaining the other properties. An illustration of the following construction is given in Figure 3. Let Sk ∩ Ci = {{x, y}} with x ∈ Ui and y ∈ V \ Ui . Let A be the set of vertices reachable from y in (V, Sj \ Ci ). Note that A ∩ Ui = ∅. We have |δ(A) ∩ Sj ∩ Ci | = 1 because (V, Sj )[Ui ] is connected. So let e = {v, w} ∈ (Sj ∩ Ci ) \ δ(A), with v ∈ Ui and w ∈ V \ Ui . Let B be the set of vertices reachable from w in (V, Sj \ Ci ). We have w ∈ B, y ∈ A, and A ∩ B = ∅ by the choice of e. Consider the path P in Sk from w to y. Note that P does not contain any vertex in Ui because |Ci ∩ Sk | = 1. But P contains at least one edge f ∈ δ(B). We will swap e and f . Since Sk ∩ Ci = {{x, y}}, the w-v-path in Sk contains P . Therefore, Ŝk := Sk + e − f is a tree. On the other hand, the path in Sj connecting the endpoints of f must use an edge in δ(B). Since Sj ∩ E[Ui ] is connected and Sj ∩ (δ(B) \ Ci ) = ∅, e is the only edge in δ(B) ∩ Sj and thus, Ŝj := Sj + f − e is a tree. 9 B v e w f P y x A Ci Figure 3: Tree Sj in green, Sk in red/bold Since f ∈ E[V \ Ui ] and e ∈ Ci , we have |Ŝj ∩ Ci | = |Sj ∩ Ci | − 1 and |Ŝj ∩ Ch | = |Sj ∩ Ch | for all h < i with j ≤ θh . Moreover, (V, Ŝj )[Ui ] is still connected. As before, we replace Sj and Sk by Ŝj and Ŝk and leave the other trees unchanged. If |Ŝj ∩ Ci | > 1, we iterate. ✷ Now the proof of Theorem 4 is a simple induction. We scan the indices of the trees j = 1, . . . , r in this order. For each j, we consider all narrow cuts Ci with j ≤ θi . Since x satisfies (2), we always have |Sj ∩ C0 | = 1 and |Sj ∩ Cℓ | = 1 for all j = 1, . . . , r. Now let i ∈ {1, . . . , ℓ − 1} with j ≤ θi . Assuming |Sj ∩ Ch | = 1 for all h < i with j ≤ θh , we first apply Lemma 9 and then Lemma 10. The new tree then satisfies |Sj ∩ Ch | = 1 for all h ≤ i with j ≤ θh , and S1 , . . . Sj−1 remain unchanged. 3 Analysis of the Approximation Ratio In this section, we will explain and analyze the algorithm that was sketched in Section 1. As mentioned before, we use the best-of-many Christofides algorithm on a special distribution of trees. To obtain we start with an optimal solution x∗ of the LP (1) and a distribution P these trees, ∗ S p with x = S∈S pS χ and round down the coefficients to integer multiples of nǫ3 . By taking multiple copies of trees, we will obtain trees S1 , . . . , Sr on which we can apply Theorem 4. Now, there is a small contribution of trees that is lost due to the rounding error. We will deal with these separately in the analysis and use that they can only represent a small fraction of x∗ . Bestof-many Christofides will then be applied to the union of the trees constructed by Theorem 4 (we will call the set of these trees Ŝ) and the "leftover" trees. How exactly we deal with rounding will be explained in the proof of Theorem 14. Our goal is to choose for each tree S a cheap vector y S in the TS -join polyhedron to bound the cost of parity correction for S. Clearly, y S = x∗ is a possible choice as x∗ (δ(U )) ≥ 1 for all ∅ = 6 U ⊂V. But for the trees in Ŝ, we will construct better vectors. 10 We follow the analysis of Vygen [2015] (based on An, Kleinberg and Shmoys [2015] and Sebő [2013]). Every tree S ∈ S can be partitioned into the edge set IS of the s-t-path in S and JS , which is a TS -join. For any narrow cut C ∈ C let eC ∈ C be a minimum cost edge in C. We will choose numbers 0 ≤ β < 21 and 0 ≤ γS ≤ 1 for all S ∈ S later. Then, for S ∈ S, we set z S := X (1 − 2β)γS χ{e} + e∈IS X  max 0, (β(2 − x∗ (C) − ǫ) − (1 − 2β)γS ) χ{eC } . (3) C∈C : |S∩C| even and y S := β(1 + ǫ)x∗ + (1 − 2β)χJS + z S . (4) Observe that z S (C) ≥ β(2 − x∗ (C) − ǫ) for all C ∈ C with |S ∩ C| even. This implies that each y S is in the TS -join polyhedron, because |δ(U ) ∩ S| even ⇔ |U ∩ TS | odd holds for all U ⊂ V with |U ∩ {s, t}| = 1. We adapt Definition 6 and Lemma 7 of Vygen [2015] as follows. Definition 11 Given numbers n 0 ≤ γS ≤ 1 foroS ∈ S and β < 12 , we define the benefit of ∗ (C)−ǫ) , γS if |S ∩ C| is even, bS,C := 1− γS if |S ∩ C| = 1, (S, C) ∈ S × C to be bS,C := min β(2−x 1−2β and bS,C = 0 otherwise. 1 2 Lemma P 12 Let 0 ≤ β < and x = S∈S pS χS . If X S∈S for all C ∈ C, then P S∈S and 0 ≤ γS ≤ 1 for S ∈ S. Let p be an arbitrary distribution over S pS bS,C ≥ P β(2 − x∗ (C) − ǫ)( S∈S:|S∩C|even pS ) 1 − 2β (5) pS c(y S ) ≤ (β + ǫβ)c(x∗ ) + (1 − 2β)c(x). P P Proof: First we note that S∈S pS c(z S ) ≤ (1 − 2β) S∈S pS c(IS ) is proved exactly as in the proof of Lemma 7 in Vygen [2015]. The claim then follows from X pS c(y S ) = β(1 + ǫ)c(x∗ ) + (1 − 2β) S∈S X pS c(JS ) + S∈S ∗ ≤ β(1 + ǫ)c(x ) + (1 − 2β) X S∈S X pS c(z S ) S∈S pS c(S). ✷ We now show how to set the γ-constants in order to maximize the benefits, with the ultimate goal to choose β as large as possible. Sebő [2013] set all γS to 21 . But we designed our r trees in such a way that the trees with small indices have only one edge in narrow cuts, thus trees that have an even number edges in a narrow cut tend to have larger indices. For example, for a narrow cut 11 C with x(C) = 32 , trees with index ≤ 2r have one edge in C (thus it is better to choose a smaller γS for these). On the other hand, the trees that have an even number of edges in C will have an index > r2 , and for those it is better to choose a bigger γS . Therefore we set γS = δ < 12 for the first half of the trees and γS = 1 − δ > 21 for the second half. More precisely: P Lemma 13 Let S1 , . . . , Sr ∈ S and ǫ ≥ 0 such that x = 1r rj=1 χSj satisfies (2), r is even, and for every C ∈ C there exists an h ∈ {1, . . . , r} with hr ≥ 2 − x∗ (C) − ǫ and |C ∩ Sj | = 1 for all j = 1, . . . , h. We set δ := 0.126, and γSj = δ if j ≤ 2r and γSj = 1 − δ otherwise. Choose β such that β 1−2β = 3.327. Then 1 Pr 1 ∗ (6) j=1 bSj ,Ci ≥ 3.327 (2 − x (Ci ) − ǫ) r |{j : |Sj ∩ Ci | even}| r for all i = 0, . . . , ℓ. As the proof is quite technical, we prove this lemma in Section 4. Using Lemma 12 and 13, we can show that the best-of-many Christofides algorithm run on the trees constructed in Theorem 4 and the trees remaining after rounding is a 2−β +ǫ-approximation for β as in Lemma 12 and obtain: Theorem 14 There is a 1.566-approximation algorithm for the s-t-path TSP. P Proof: Let x∗ be an optimal solution for (1) and p a distribution over S with x∗ = S∈S pS χS such that pS > 0 for at most n2 trees S. As mentioned before, x∗ and p canPbe found in 3 polynomial time. Setting p′S := nǫ3 ⌊ nǫ pS ⌋ and p′′S := pS − p′S for S ∈ S, we have S∈S p′′S ≤ nǫ 3 P ′ since |{S ∈ S : pS > 0}| ≤ n2 . By setting r = 2nǫ S∈S pS (an even integer), we can obtain r 3 trees S1 , . .P . , Sr , such that each S ∈ S appears 2⌊ nǫ pS ⌋ times in this list. 1 Let x = r rj=1 χSj . We show that x fulfills property (2). Since all trees (V, S) with pS > 0 have |δ(s) ∩ S| = |δ(t) ∩ S| = 1, we have x(δ(s)) = x(δ(t)) = 1. It remains to show x∗ (F ) − ǫ ≤ x(F ) ≤ x∗ (F ) + ǫ for all F ⊆ E:  X  p′ P S ′ − pS |S ∩ F | x(F ) − x∗ (F ) = p S∈S S∈S S P ′′ X pS − p′′ − (1 − S S∈S pS )pS P = |S ∩ F | 1 − S∈S p′′S S∈S  |S ∩ F | X P P = −p′′S + pS S∈S p′′S 1 − S∈S p′′S S∈S  P P |S∩F | P Since |S∩F | ≤ |S| = n−1 for all S ∈ S, we have x(F )−x∗ (F ) ≤ ( S∈S p′′S ) S∈S (pS 1− p′′ ) ≤ ǫ (n−1) n 1− nǫ S∈S ≤ ǫ. 12 S ǫ P (n−1) |S∩F | n ≥ −ǫ. Similarly, x(F ) − x∗ (F ) ≥ − S∈S p′′S 1−P ′′ ≥ − 1− ǫ p S∈S S n This shows (2). Using Theorem 4 we transform this list to Ŝ1 , . . . , Ŝr fulfilling the statement of the theorem. β = 3.327. Therefore, we can apply Lemma 13 to Ŝ1 , . . . , Ŝr and x and obtain inequality (5) for 1−2β Thus, we will be able to apply Lemma 12 in the analysis. Our algorithm does parity correction for each of Ŝ1 , . . . , Ŝr and for all trees S with p′′S > 0 and returns the best resulting tour. We obtain a bound on the value of the best output by an averaging argument. For that, we use y Ŝj as defined above in (4) to bound the parity correction cost for trees Ŝj and x∗ for other trees. P P P ′ S We have x = 1r rj=1 χŜj = 1r rj=1 χSj = P 1 p′ S∈S pS χ . Therefore, S∈S c(x) = P 1 X p′S c(S) ′ p S∈S S S∈S ≤P S 1 Then, using Lemma 12 in (∗) and (7) in (∗∗),  r  min min c(Ŝj ) + c(y Ŝj ) , min′′ j=1 ≤ X p′S X p′S c(S) 1 r j=1 S∈S = S∈S:pS >0 r X S∈S X pS c(S) ′ S∈S pS S∈S ∗  c(S) + c(x ) =P 1 ′ S∈S pS c(x∗ ). (7)   P  c(Ŝj ) + c(y Ŝj ) + S∈S p′′S c(S) + c(x∗ ) + X p′S S∈S r X 1 Sj r c(y ) + j=1 P S∈S r  X X X  1 Sj p′S = 1+ p′′S c(x∗ ) + r c(y ) P S∈S p′′S c(x∗ ) j=1 S∈S S∈S p′′S c(S) + X X  (∗)   p′S (β + ǫβ)c(x∗ ) + (1 − 2β)c(x) ≤ 1+ p′′S c(x∗ ) + S∈S S∈S (∗∗) ≤  1+ X S∈S  p′′S c(x∗ ) + X p′S (β + ǫβ)c(x∗ ) + (1 − 2β)c(x∗ ) S∈S  X  ≤ 2 − β + ǫβ + p′′S c(x∗ ) S∈S < (2 − β + ǫ)c(x∗ ). P β In the last inequality, we used S∈S p′′S ≤ nǫ . Moreover, 1−2β = 3.327 is equivalent to β = 0.434 + 0.0006. Thus, for ǫ ≤ 0.0006, our approximation ratio is at most 2 − 0.434 = 1.566. 13 3.327 7.654 > ✷ 4 Computing the Benefit In this section we prove Lemma 13. Fix i ∈ {0, . . . , ℓ}. We write ξ := x∗ (Ci ) + ǫ, π := 1r |{j : |Sj ∩ Ci | even}|, and ρ := 1r θi − (2 − x∗ (Ci ) − ǫ). That means ρr is the amount of θi due to rounding up (remember that θi = ⌈r(2 − x∗ (Ci ) − ǫ)⌉). We have π ≤ x(Ci ) − 1 ≤ x∗ (Ci ) + ǫ − 1 = ξ − 1: The first inequality follows from x(Ci ) ≥ 1r (|{Sj : |Ci ∩ Sj | ≥ 2}| + r) ≥ π + 1, the second is due to (2). Clearly, (6) holds if π = 0 or ξ ≥ 2, thus we assume this is not true. Suppose 3.327(2 − ξ) ≤ δ. Consequently, bSj ,Ci = β(2−ξ) 1−2β = 3.327(2 − ξ) for all j with |Sj ∩ Ci | even. Thus, (6) trivially holds and we assume from now on 3.327(2 − ξ) > δ. For convenience of notation, we define µ = min{1 − δ, 3.327(2 − ξ)}. First note that among the trees Sj with j > θi , we either have trees with |Sj ∩ Ci | even or additional trees with |Sj ∩ Ci | = 1, which will guarantee us a certain amount of benefit: If there is an f fraction of trees Sj with |Sj ∩ Ci | ≥ 3, then 3f + (1 − f ) + π ≤ x(Ci ) ≤ x∗ (Ci ) + ǫ = ξ, so . f ≤ ξ−π−1 2 Let us distinguish several cases. Case 1: ξ ≤ 32 . Then θi ≥ 2r , and the first half of the trees have only one edge in Ci . We have 3.327(2 − ξ) ≥ 3.327 · 21 > 1 − δ. Therefore, µ = 1 − δ. Considering all trees, we obtain the following lower bounds for the benefit: |Sj ∩ Ci | j r 2 j≤ r 2 < j ≤ θi θi < j θi < j θi < j fraction of trees 1 1 2,4,... 3,5,... 1 θi r − 1 2 1 3 = 2 2 −ξ+ρ π ≤ − π − 1) 1 ≥ 2 (ξ − π − 1) − ρ 1 2 (ξ bSj ,Ci 1−δ δ µ 0 δ Thus, we can bound the weighted benefit (cf. Figure 4): 1 Pr 1−δ 3 1 j=1 bSj ,Ci ≥ 2 + ( 2 − ξ)δ + πµ + 2 (ξ − π − 1)δ r = = δ+1 2 δ+1 2 − − ξδ 2 ξδ 2 + π(µ − 21 δ) + π(1 − 32 δ) =: g1 (ξ, π)). 14 1 1−δ δ 0 0 1 θi 2 r π 1 Figure 4: On the x-axis, we have the fraction of trees, on the y-axis their benefit. A lower bound for the benefit for trees Sj is given by the grey area for j ≤ θi , the green (dotted) area for |Sj ∩ Ci | even and the blue (lined) area otherwise. To show (6), examine the minimum:   1 Pr g1 (ξ, π) j=1 bSj ,Ci r 3 ≥ min : 1 ≤ ξ ≤ 2, 0 < π ≤ ξ − 1 π(2 − x∗ (Ci ) − ǫ) π(2 − ξ) ) ( ξδ 3 δ+1 − + (ξ − 1)(1 − δ) (∗) 2 2 2 = min : 1 ≤ ξ ≤ 32 (ξ − 1)(2 − ξ) ) ( 2δ − 12 + ξ(1 − 2δ) : ξ ∈ R ≥ 3.327. ≥ min (ξ − 1)(2 − ξ) The first inequality follows from the inequalities derived above. The equality (∗) follows by ξδ  ∂ g1 (ξ,π) 1 δ+1 1 . This is considering the partial derivative of the function in π: ∂π 2 − 2 π(2−ξ) = 2−ξ − π 2 g1 (ξ,π) negative for all 1 ≤ ξ < 2 and δ < 1 and π > 0. Thus the minimum of π(2−ξ) is attained at a point where π is maximal and it suffices to consider the case where π = ξ − 1. The last inequality is matter of computation: We just have to verify that 2δ − 12 + ξ(1 − 2δ) − 3.327(ξ − 1)(2 − ξ) ≥ 0 for δ = 0.126 and all ξ ∈ R. The minimum is attained at ξ ≈ 1.39. Case 2: ξ > 32 . Then θi ≤ 2r because r is even. In this case, less than half of the trees are certain to have only one edge on Ci . Since bSj ,Ci = µ > δ for j ≥ 2r and |Sj ∩ Ci | even, in the worst case those indices j with |Sj ∩ Ci | even are as small as possible. Thus, we obtain the following lower bounds: j j ≤ θi θi < j ≤ min{ 2r , θi + rπ} r 2 < j ≤ θi + rπ θi < j θi < j |Sj ∩ Ci | fraction of trees bSj ,Ci 1 2,4,... 2,4,... 3,5,... 1 2−ξ+ρ min{π, − 23 + ξ − ρ} max{0, 32 − ξ + ρ + π} ≤ 12 (ξ − π − 1) ≥ 21 (ξ − π − 1) − ρ 1−δ δ µ 0 δ 15 1 1−δ µ π δ 0 0 θi r 1 2 1 Figure 5: On the x-axis, we have the fraction of trees, on the y-axis their benefit. A lower bound for the benefit for trees Sj is given by the grey area for j ≤ θi , the green (dotted) area for |Sj ∩ Ci | even and the blue (lined) area for the remaining trees. Using the table we obtain (cf. Figure 5): 1 Pr 3 j=1 bSj ,Ci ≥ (2 − ξ + ρ)(1 − δ) + min{π, − 2 + ξ − ρ}δ r + max{0, 32 − ξ + ρ + π}µ + ( 21 (ξ − π − 1) − ρ)δ ≥ (2 − ξ)(1 − δ) + min{π, − 32 + ξ}δ + max{0, 32 − ξ + π}µ + 12 (ξ − π − 1)δ + ρ(1 − 2δ) ≥ (2 − ξ)(1 − δ) + min{π, − 23 + ξ}δ + max{0, 32 − ξ + π}µ + 12 (ξ − π − 1)δ =: g2 (ξ, π). Now, we distinguish two cases, depending on where the minimum in g2 is attained. Case 2.1: π ≤ ξ − 23 . Then g2 (ξ, π) = (2 − ξ)(1 − δ) + πδ + 2δ (ξ − π − 1) = 2 − 25 δ + ( 23 δ − 1)ξ + 2δ π. As before, we minimize   1 Pr g2 (ξ, π) 3 j=1 bSj ,Ci r 3 ≥ min : < ξ < 2, 0 < π ≤ ξ − 2 π(2 − x∗ (Ci ) − ǫ) π(2 − ξ) 2 ) ( 2 − 52 δ + ( 32 δ − 1)ξ + δ2 (ξ − 32 ) 3 (∗) : 2 <ξ<2 = min (ξ − 32 )(2 − ξ) ( ) 2 − 13 4 δ + (2δ − 1)ξ ≥ min : ξ ∈ R ≥ 3.327 (ξ − 32 )(2 − ξ) 1 ∂ g2 (ξ,π) To show (∗), as before, we consider the partial derivative in π and obtain ∂π π(2−ξ) = 2−ξ −  1 ∂ g2 (ξ,π) 2 − 52 δ + ( 32 δ − 1)ξ . Since 2 − 25 δ + ( 32 δ − 1)ξ = (2 − ξ)(1 − 23 δ) + 2δ > 0, we have ∂π π(2−ξ) < 0 π2 for all ξ < 2 and π > 0, and the minimum is attained for π = ξ − 23 . To verify the last inequality, 3 checking 2− 13 4 δ + (2δ − 1)ξ − 3.327(ξ − 2 )(2− ξ) ≥ 0 for δ = 0.126 and all ξ ∈ R is straightforward. (The minimum is attained at ξ ≈ 1.86.) 16 Case 2.2: π > ξ − 32 . Then g2 (ξ, π) = (2 − ξ)(1 − δ) + (ξ − 23 )δ + ( 23 − ξ + π)µ + 12 (ξ − π − 1)δ = 2 − 4δ + 23 µ + ξ( 52 δ − 1 − µ) + π(µ − 12 δ). Minimization yields 1 Pr j=1 bSj ,Ci r π(2 − x∗ (Ci ) − ǫ)   g2 (ξ, π) 3 3 ≥ min : < ξ < 2, ξ − 2 ≤ π ≤ ξ − 1 π(2 − ξ) 2 ( 2 − 4δ + 32 µ + ξ( 25 δ − 1 − µ) + π(µ − 12 δ) (∗) = min : π(2 − ξ) 3 2 ) < ξ < 2, π ∈ {ξ − 32 , ξ − 1} (∗∗) ≥ 3.327. 1 1 ∂ g2 (ξ,π) 3 To show (∗), again consider the partial derivative in π: ∂π π(2−ξ) = 2−ξ − π 2 2 − 4δ + 2 µ +  ξ( 25 δ − 1 − µ) . Unlike the previous cases, this function can be positive or negative for δ = 0.126 and 32 < ξ < 2 and π > 0. But for fixed ξ, it is either always non-negative or always nonpositive. Therefore the minimum is either attained for maximal or for minimal π, which explains the equality. We now show (∗∗). Let us examine the two possible choices for π. For π = ξ − 32 , we minimize g2 (ξ,ξ− 23 ) , (ξ− 32 )(2−ξ) which is the same function as in Case 2.1 (note that the terms involving µ cancel). For π = ξ − 1, we get: ( 2 − 4δ + 32 µ + ξ( 25 δ − 1 − µ) + (ξ − 1)(µ − 12 δ) : min (ξ − 1)(2 − ξ) ) ( 2 − 72 δ + 12 µ + ξ(2δ − 1) :ξ∈R . ≥ min (ξ − 1)(2 − ξ) If µ = 1 − δ, (8) equals min ξ∈R n5 2 −4δ+ξ(2δ−1) (ξ−1)(2−ξ) o = min ξ∈R n 2δ− 12 +(3−ξ)(1−2δ) (ξ−1)(2−ξ) o 3 2 ) <ξ<2 (8) = min ′ ξ ∈R n 2δ− 12 +ξ ′ (1−2δ) (ξ ′ −1)(2−ξ ′ ) o . We replaced ξ ′ = 3 − ξ, which corresponds to mirroring at 23 . This minimum was already shown to be at least 3.327 in Case 1. If µ = 3.327(2 − ξ), then (8) is equal to ( ) 2 − 27 δ + ξ(2δ − 1) + 12 3.327(2 − ξ) 3 min : 2 <ξ<2 . (ξ − 1)(2 − ξ) To see that this is at least 3.327, it just remains to check that 2 − 27 δ + ξ(2δ − 1) + 21 3.327(2 − ξ) − 3.327(ξ − 1)(2 − ξ) ≥ 0 for δ = 0.126 and all ξ ∈ R. The minimum is attained at ξ ≈ 1.86. This completes the proof of Lemma 13. 17 5 Conclusion The approximation ratio can probably be improved slightly by choosing the γŜj differently, but still depending only on rj . However, using an analysis based on Lemma 12, one cannot obtain a because the benefit can never be more than one and there can better approximation ratio than 14 P9 β 3 ∗ be cuts C with x (C) = 2 and S∈S:|S∩C| even pS = 12 ; therefore 1−2β ≤ 4. Our ratio is already close to this threshold. On the other hand, it is not impossible that the best-of-many Christofides algorithm on a distribution like the one obtained in Theorem 4, or even on an arbitrary distribution, has a better approximation ratio, maybe even 23 . References An, H.-C., Kleinberg, R., and Shmoys, D.B. [2015]: Improving Christofides’ algorithm for the s-t path TSP. Journal of the ACM 62 (2015), Article 34 Asadpour, A., Goemans, M.X., Mądry, A., Oveis Gharan, S., and Saberi, A. [2010]: An O(log n/ log log n)approximation algorithm for the asymmetric traveling salesman problem. Proceedings of the 21st Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2010), 379–389 Christofides, N. [1976]: Worst-case analysis of a new heuristic for the traveling salesman problem. Technical Report 388, Graduate School of Industrial Administration, Carnegie-Mellon University, Pittsburgh 1976 Dantzig, G.B., Fulkerson, D.R., and Johnson, S.M. [1954]: Solution of a large scale traveling salesman problem. Operations Research 2 (1954), 393–410 Edmonds, J. [1965]: The Chinese postman’s problem. Bulletin of the Operations Research Society of America 13 (1965), B-73 Edmonds, J., and Johnson, E.L. [1973]: Matching, Euler tours and the Chinese postman. Mathematical Programming 5 (1973), 88–124 Gao, Z. [2013]: An LP-based 23 -approximation algorithm for the s-t path graph traveling salesman problem. Operations Research Letters 41 (2013), 615–617 Gao, Z. [2015]: On the metric s-t path traveling salesman problem. SIAM Journal on Discrete Mathematics 29 (2015), 1133–1149 Genova, K., and Williamson, D.P. [2015]: An experimental evaluation of the best-of-many Christofides’ algorithm for the traveling salesman problem. arXiv:1506.07776 (extended abstract to appear in ESA 2015) Grötschel, M., Lovász, L., and Schrijver, A. [1981]: The ellipsoid method and its consequences in combinatorial optimization. Combinatorica 1 (1981), 169–197 Held, M., and Karp, R.M. [1970]: The traveling-salesman problem and minimum spanning trees. Operations Research 18 (1970), 1138–1162 18 Hoogeveen, J.A. [1991]: Analysis of Christofides’ heuristic: some paths are more difficult than cycles. Operations Research Letters 10 (1991), 291–295 Oveis Gharan, S., Saberi, A., and Singh, M. [2011]: A randomized rounding approach to the traveling salesman problem. Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer Science (FOCS 2011), 550–559 Sebő, A. [2013]: Eight fifth approximation for TSP paths. In: Integer Programming and Combinatorial Optimization; Proceedings of the 16th IPCO Conference; LNCS 7801 (J. Correa, M.X. Goemans, eds.), Springer 2013, pp. 362–374 Sebő, A., and Vygen, J. [2014]: Shorter tours by nicer ears: 7/5-approximation for graph-TSP, 3/2 for the path version, and 4/3 for two-edge-connected subgraphs. Combinatorica 34 (2014), 597–629 Vygen, J. [2012]: New approximation algorithms for the TSP. OPTIMA 90 (2012), 1–12 Vygen, J. [2015]: Reassembling trees for the traveling salesman. arXiv:1502.03715 19
8cs.DS
Dependent Types for Multi-Rate Flows in Synchronous Programming (System Description) William Blair Hongwei Xi Boston University Massachusetts, USA [email protected] Boston University Massachusetts, USA [email protected] Synchronous programming languages emerged in the 1980s as tools for implementing reactive systems, which interact with events from physical environments and often must do so under strict timing constraints. In this report, we encode inside ATS various real-time primitives in an experimental synchronous language called Prelude, where ATS is a statically typed language with an ML-like functional core that supports both dependent types (of DML-style) and linear types. We show that the verification requirements imposed on these primitives can be formally expressed in terms of dependent types in ATS. Moreover, we modify the Prelude compiler to automatically generate ATS code from Prelude source. This modified compiler allows us to solely rely on typechecking in ATS to discharge proof obligations originating from the need to typecheck Prelude code. Whereas ATS is typically used as a general purpose programming language, we hereby demonstrate that it can also be conveniently used to support some forms of advanced static checking in languages equipped with less expressive types. 1 Introduction Building software that must work reliably is an undeniably challenging process. Yet, engineers can still construct reliable real-time systems that operate under strict temporal requirements. Companies have long adopted model-based methods for designing real-time software[1]. Model-based design methods emphasize automatically generating code from high-level models that formally describe the behavior of a system. In fact, synchronous programming languages helped start a movement towards this paradigm (of model-based automatic code generation) when they were introduced in the 1980s. In synchronous programs, tasks communicate through data flows with the assumption that each data flow produces values instantaneously at every step of a global logical clock. However, the simplicity of a single system-wide global logical clock can be the source of serious limitation in practice. If a system consists of communicating periodic tasks that need to run at different rates, the global logical clock can become an obstacle as the programmer must synchronize all the tasks manually. The process of manual synchronization is likely to be both tedious and error-prone with little support provided by the compiler for preventing subtle timing bugs. As a part of his PhD thesis, Julien Forget developed Prelude[6], a synchronous language that features primitives for describing the real-time behavior of tasks in a multi-rate periodic system. Synchronizing data flows with different clocks in Prelude is still left to the programmer, but the language provides operators to directly modify real-time clocks. For example, a programmer can stretch a clock to accomplish under-sampling or shrink it for over-sampling. The typechecker of Prelude then automatically verifies that all communicating tasks are synchronized. The advantage of this system is that the language offers the ability to easily adjust the temporal behavior of tasks while providing support for verification in the type system. Jeremy Yallop and Damien Doligez (Eds.) ML/OCaml 2015 EPTCS 241, 2017, pp. 36–44, doi:10.4204/EPTCS.241.3 © William Blair & Hongwei Xi This work is licensed under the Creative Commons Attribution License. William Blair & Hongwei Xi 37 In this report, we encode inside ATS the clock calculus presented in Prelude, where ATS is a statically typed programming language with a functional core of ML-style. The type system of ATS is rooted in the Applied Type System framework[11], which gives the language its name. In ATS, both dependent types[13, 12] (of DML-style) and linear types are supported. We assign dependent types to the real-time clocks available in Prelude and then develop functions to capture the behavior of clock transformation operators. With these functions, we can typecheck the automatically generated ATS code from Prelude source (by a modified Prelude compiler) and formally verify that the data flows in the original Prelude source are properly synchronized. A running version of our modified compiler can be found online1 . Although ATS is typically used as a source programming language, we hereby demonstrate that it can also be conveniently used as a target language for the purpose of supporting advanced static checking in a host language equipped with less expressive types. In addition, the reported use of dependent types for reasoning about temporal properties should be interesting in its own regard. 2 Multi-Rate Flows in Prelude An embedded real-time system continuously interacts with its own environment and often must do so under strict timing constraints. For example, a task may periodically sample a sensor and provide data that can in turn be used by the application logic to adjust the output of an actuator. In order to guarantee that the system is responsive, engineers need to demonstrate that the worst case execution time (WCET) for the communication from the sensor, to the application logic, and then to the actuator can never exceed a required deadline. Synchronous programming languages simplify development by abstracting time for the designer and providing well-defined semantics to enable verification and automatic code generation from a model of the system. When a synchronous program is compiled, all of the nodes and flows are translated into a singlestep function that is invoked repeatedly. The time between steps is the period of the system, and the Synchronous Hypothesis states that all the flows must be re-evaluated before the end of each step. It is up to the system designer to ensure that this requirement is met. Several synchronous languages[2, 3] have been developed since the 1980s, with Lustre[4] currently serving as the underlying language powering the SCADE Suite developed by Esterel Systems. This abstraction for time in a synchronous language is of great convenience because it not only alleviates the need for the programmer to worry about low-level timing details but also produces fully deterministic code. However, it makes it difficult at the same time to compose systems where tasks may run at different rates. In order to make tasks execute at the desired rates, the programmer must essentially schedule tasks by hand in terms of a global logical clock, which can be both tedious and error-prone. Prelude addresses the problem of manual synchronization for multi-rate periodic systems by removing the notion of a single system-wide global clock. Instead, it assigns each periodic flow F its own clock C represented as a pair (n, p) where n ∈ N+ (that is, n is a positive integer) and p ∈ Q+ (that is, p is a positive rational number); the integer n is referred to as the period of C and p the phase offset. Note that the period of a flow F is just the inter-arrival time of values produced by F and the phase offset is the initial delay of the flow. The activation time of a flow (that is, its start date) is given by n · p. In Prelude’s clock calculus, every date (that is, a point in time in the program) is required to be an integer so as to simplify schedulability analysis. Therefore, we need to ensure that the date t0 is an integer for every flow. In other words, a clock defined as (n, p) is valid if and only if both n ∈ N+ and n · p ∈ N+ hold. 1 An end-to-end demonstration of our system is available at https://travis-ci.org/wdblair/overture Dependent Types for Multi-Rate Flows 38 imported node database (i: int rate (10, 0)) returns (o: int rate (10, 0)); imported node controller(i: int rate (100, 0); j: int rate (100, 0)) returns (o: int rate (100, 0); p: int rate (100, 0)); sensor i; actuator o; node sampling (i: rate (10, 0)) returns (o: rate (100, 0)) var command: rate (100, 0); var response: rate (10, 0); let (o, command) = controller(i/ˆ10, (0 fby response)/ˆ10); response = database(command*ˆ10); tel Figure 1: A multi-rate periodic system in the Prelude language In Prelude, a generic flow consists of an infinite list of tuples ∀i ∈ N. (vi ,ti ), where vi is the value of the flow produced at time ti . Under the normal synchronous paradigm, every flow produces some value during each tick of the logical clock. There is no longer the notion of a global time scale in Prelude. Instead, every flow is assigned its own clock, which can be lengthened as well as shortened. By modifying a clock, we alter the rate at which values are produced in a program. If a value contained in a flow is undefined for some arbitrary date, we cannot adjust its real-time behavior while expecting deterministic results. As such, it is useful to define a class of data flows that are strictly periodic. Formally speaking, a data flow is strictly periodic if and only if ∃n ∈ N+ .∀i ∈ N. ti+1 − ti = n where n is the period of the clock assigned to the flow and ti is the date when value i is produced by the flow[6]. Intuitively, a strictly periodic data flow produces a value after each period. In Prelude, a set of operators are provided to transform the clocks assigned to strictly periodic flows which greatly facilitates communication between tasks that execute at different rates. In Figure 3 presented in Section 3 we describe how to encode the semantics of these operators into the ATS programming language. As an example, let us consider a simple system where a controller runs with a clock of (100, 0) and interacts with a fast running task that provides critical data from the environment with a clock of (10, 0). Figure 1 implements this system where the controller node sends a command to the database node and receives a response back from it. Observe that the controller executes 10 times slower than the database. If we want the controller node’s output command to be synchronized with the database, we must over-sample it using the *ˆ operator. Likewise, we under-sample the database node’s output response in order to synchronize it with the controller node by using the /ˆ operator. In the rest of the paper, we will define the Prelude operators necessary to implement this program in William Blair & Hongwei Xi 39 node monitor (temperature: int rate (10, 0), fault: bool rate(100, 0)) returns (alert: int rate (100, 0)) let alert = (temperature /ˆ 10) when fault tel Figure 2: The boolean flow alert given here produces data only when fault produces true. ATS so that all synchronization constraints are checked by the type checker. Then, we will describe how we can transform Prelude programs like this one into ATS to verify all communication is synchronized. 3 Assigning Types to Flows in ATS At a high level, Prelude modifies the semantics of traditional synchronous languages by refining the basic flow type to have its own clock as opposed to a global clock. This refinement generates a proof obligation for the compiler as a flow going from one task to another is not necessarily synchronized. We can readily defer this obligation to typechecking in ATS by assigning the following type (written in the concrete syntax of ATS) to strictly periodic flows: abstype SFlow(a: type, n: int, p: rat) Note that int and rat refer to the sorts for integers and rational numbers, respectively. There is another sort nat for natural numbers that will appear later. Given a, n, and p, the type SFlow(a, n, p) is for a strictly periodic flow of values of the type a such that the clock assigned to the flow is of the period n and phase offset p. Formally speaking, a flow F of the type SFlow(a, n, p) consists of the following pairs: ∀i ∈ N.(vi ,ti ) ∈ F where each vi is a (defined) value of the type a and ti = n · (p + i). Flows that are not strictly periodic are also useful in Prelude. For instance, the following type BFlow is for boolean flows: abstype BFlow(a: type, n: int, p: rat) Boolean flows differ from strictly periodic flows by allowing values to be produced only on a subset of the dates given by a flow’s clock. In Figure 2, we define a boolean flow alert to give the value of a temperature sensor only at dates when fault produces a true value. Modifying a flow’s clock affects the values that are included in the flow. In order to determine precisely what values remain in a flow after its clock is modified, Prelude requires that every flow passed to a clock transformation operator be strictly periodic. By the same token, all clock operators yield strictly periodic flows. In Prelude, only integer dates are considered valid. Therefore, any flow’s clock is subject to the requirement that n · p ∈ N holds, where n and p are the period and phase offset of the clock, respectively. If this was not true, then a task’s start time could not be represented as an integer date. When assigning types to a clock transformation operator, we often add a precondition to ensure that only a clock satisfying the requirement can be actually returned by the operator. Dependent Types for Multi-Rate Flows 40 mul div shift fby : : : : ∀a : type ∀n : int ∀p : rat ∀k : nat. (k | n) ⊃ (SFlow(a, n, p), int(k)) → SFlow(a, n/k, p · k) ∀a : type ∀n : int ∀p : rat ∀k : nat. (k | n) ⊃ (SFlow(a, n, p), int(k)) → SFlow(a, n · k, p/k) ∀a : type ∀n : int ∀p : rat ∀k : rat. (n · k ∈ N) ⊃ (SFlow(a, n, p), rat(k)) → SFlow(a, n, p + k) ∀a : type ∀n : int ∀p : rat. (a, SFlow(a, n, p)) → SFlow(a, n, p) Figure 3: Some clock transformation operators and their types The clock transformation operators in Prelude are to be translated into regular functions in ATS. Synchronous flows resemble infinite lazy streams in traditional functional programming languages. For instance, we can introduce an operator cons of the following type: cons : ∀a : type ∀n : int ∀p : rat. (a, SFlow(a, n, p)) → SFlow(a, n, p − 1) Basically, cons adds a given value to the beginning of a given flow to form a new flow that starts one period ahead of the given flow; the first produced value of the new flow is simply the given value, and the rest are those produced by the given flow. As another example, we can introduce an operator tail of the following type: tail : ∀a : type ∀n : int ∀p : rat. SFlow(a, n, p) → SFlow(a, n, p + 1) Given a flow, tail returns another one that simply delays the given flow’s activation by one period. There are four commonly used clock transformation operators in Figure 3 plus the types assigned to them. The mul operator (that is, *ˆ in Prelude) speeds up a clock by shrinking its period; it implements over-sampling where every element in the flow is repeated k times to meet its timing constraint. Likewise, the div operator (that is, /ˆ in Prelude) slows a clock by stretching its period; it implements undersampling where only one element is taken out of every k elements of the input flow. Given a flow and a rational number k, the shift operator shifts the phase offset of the clock assigned to the flow by k, thus delaying the activation of the flow. Given an element and a flow, the fby operator does not modify the flow’s clock at all; it instead simply shifts the flow by one period and sets the first element of the flow with the given one. For instance, fby can be implemented based on cons and shift in the concrete syntax of ATS as follows: implement fby (x, f) = cons(x, shift(f, 1)) Another interesting operator merge is given the following type: ∀a : type ∀n : int ∀p : rat. (SFlow(bool, n, p), SFlow(a, n, p), SFlow(a, n, p)) → SFlow(a, n, p) where bool is the type for booleans. Essentially, merge takes three strictly periodic flows and returns one; for each i, if the value in the first flow at ti is true, then the value in the second flow at ti is picked to be the value of the returned flow at ti ; otherwise, the value in the third flow at ti is picked. Yet another interesting operator when is given the following type: ∀a : type ∀n : int ∀p : rat. (SFlow(bool, n, p), SFlow(a, n, p)) → BFlow(a, n, p) where BFlow is for boolean flows. Given two strictly periodic flows, when returns a boolean flow; for each i, if the value of the first flow is true at ti , then the value of the second flow at ti is picked to be the William Blair & Hongwei Xi 41 value of the returned flow at ti ; otherwise, the returned flow is undefined at ti . Note that when is so far the only operator that does not return a strictly periodic flow. So far we have only thought of communication through flows in Prelude as individual values arriving at dates (that is, points in time). Please recall the simple example presented in Figure 1, where a controller queries a faster data task for information. In order to make the communication between them synchronous, we specified that only one out of every 10 values produced by the faster task is to be read by the controller. Suppose that the controller needs all values produced by the database since it was last invoked. This requires a queuing mechanism in the clock calculus, which is a feature outlined in Prelude’s specification but yet to be supported by the compiler. We can readily add verification support for this feature with dependent types in ATS. Suppose we have a strictly periodic flow F that we want to under-sample by some factor k. We can think of the flow as having values of the type array(a : type, k : int) for arrays that each contain k elements of the type a. This gives us the following type for a divide operation that performs queuing rather than under-sampling. In this way, every time the node that consumes this flow is activated, it has access to all k of the most recent values emitted from the flow. Putting this formally, the type for such an operator is as follows: ∀a : type ∀n : int ∀p : rat ∀k : int. (k > 0) ⊃ (SFlow(a, n, p), int(k)) → SFlow(array(a, k), n · k, p/k) Clearly, this type directly relates the number of accumulated values produced by a given flow to the factor by which the flow’s clock is expected to be stretched. 4 Using ATS as a Target Language With the above outlined operators plus their types, we can write ATS programs that correspond semantically to their Prelude counterparts in terms of the synchronization checks that must be performed. In this work, we modified the Prelude compiler to automatically generate ATS code with type signatures from Prelude source. As we will see in this section, the translation is largely straightforward. Every imported node is declared as an external function with their input and output flows having the same clock, and every node that is actually implemented in Prelude is translated into a function definition in ATS. Probably the biggest issue is to bridge the semantic gap between Prelude, a concurrent declarative language, and ATS, a call-by-value language with an ML-like functional core. This is especially evident when we translate programs where certain flows are passed as arguments before they are actually defined. For instance, please take a careful look at the code presented in Figure 1. This issue cannot be addressed by simply reordering expressions as circular definitions are both legal and common in Prelude. Instead, we address this issue by using a combination of linear types and support for proof terms in ATS. For each locally declared flow, we generate a linear proof term. In order to consume this proof term, we specify a proof function our Prelude compiler must call with two ATS flows that have equal clocks to ensure the Prelude flow they represent is given exactly one clock. In Figure 4 you can see this approach used to verify the response and command flows maintain their specified clock. After modifying the Prelude compiler, we automatically generate ATS code directly from the abstract syntax tree of Prelude source; potential synchronization errors in the Prelude source can be captured by typechecking the ATS code translated from it. For instance, the Prelude example in Figure 1 is translated into the ATS program in Figure 4 for the purpose of typechecking. Note that the lines starting with the keyword prval are theorem-proving code related to the circularly defined flows in this function. Dependent Types for Multi-Rate Flows 42 fun main ( i: SFlow (int, 10, 0) ): (SFlow (int, 100, 0)) = let var response : SFlow (int, 10, 0) prval pfresponse = flow_future_make (response) var command : SFlow (int, 100, 0) prval pfcommand = flow_future_make (command) val response' = (flow_fby (0, response)) val o, command' = controller ((flow_div_clock (i, 10)), (flow_div_clock (response', 10))) val response' = server (flow_mul_clock (command, 10)) prval () = flow_future_elim (pfresponse, response, response') prval () = flow_future_elim (pfcommand, command, command') in (o) end Figure 4: ATS code automatically generated by a modified Prelude compiler 5 Conclusion and Future Work Prelude simplifies the process of composing communicating real-time tasks, that possibly run at different rates, into a fully deterministic system. This desire for rooting out nondeterminism in the development process was present when synchronous languages were first introduced[7] and it continues with methods applied for industrial tools like Matlab’s Simulink[5] and Esterel Technologie’s SCADE which uses the synchronous language Lustre in mission critical software across aerospace, automotive, and industrial applications. There is also research on synchronous languages that addresses the implementation of Mixed Criticality Systems[14], where some tasks may be allowed to violate the Synchronous Hypothesis and miss deadlines. One could imagine extending a language like Prelude to guarantee synchronized deterministic communication between both hard real-time tasks and less critical soft real-time tasks. The Liquid Clocks[10] framework defines a type system for designing new synchronous languages across different models of time and communication. Liquid Clocks uses refinement types and an inference algorithm inspired by Liquid Types[9] to verify time and causality constraints in a network defined by data flows. Our approach differs from Liquid Clocks in that we use the expressive types found in ATS, an existing programming language, to support the implementation of a domain specific synchronous language for multi-rate periodic systems. In this report, dependent types in ATS of DML-style [13, 12] are assigned to the multi-rate flows introduced in Prelude and also to various clock transformation operators on these flows. Note that there is explicit use of quantifiers in the types assigned to these operators, far exceeding what is accomplished in Prelude, where every clock is either fixed to be a constant or can be inferred from a constant one. For instance, a constraint stating that a node must always produce a flow whose clock is twice as fast as its input node can be readily expressed in ATS but cannot currently in Prelude. The basic building block of composition in synchronous languages like Prelude is a unidirectional flow. Yet, in complex embedded systems, individual components may communicate using various dif- William Blair & Hongwei Xi 43 ferent protocols. While these protocols are implemented on top of flows, there is no formal description for them in an architecture design language like Prelude. One avenue for future work is to develop a synchronous domain specific version of ATS that can use typed channels like those found in Session Types[8] to describe both the temporal behavior of the system and what is communicated inside of it. The latter information, for instance, can help verify node implementations in a synchronous program or detect possible errors. 6 Bibliography References [1] A. Benveniste, P. Caspi, S. A. Edwards, N. Halbwachs, P. Le Guernic & R. De Simone (2003): The Synchronous Languages 12 Years Later. Proceedings of the IEEE 91(1), pp. 64–83, doi:10.1109/jproc. 2002.805826. [2] A. Benveniste, P. Le Guernic & C. Jacquemot (1991): Synchronous Programming with Events and Relations: the SIGNAL Language and its Semantics. Science of Computer Programming 16(2), pp. 103–149, doi:10. 1016/0167-6423(91)90001-e. [3] G. Berry & G. Gonthier (1992): The Esterel Synchronous Programming Language. Science of Computer Programming 19(2), pp. 87–152, doi:10.1016/0167-6423(92)90005-v. [4] P. Caspi, D. Pilaud, N. Halbwachs & J. A. Plaice (1987): LUSTRE: a declarative language for real-time programming. In: Proceedings of the 14th Annual ACM Symposium on Principles of Programming Languages, Association for Computing Machinery (ACM), doi:10.1145/41625.41641. [5] Edmund M Clarke & Paolo Zuliani (2011): Statistical model checking for cyber-physical systems. In: Automated Technology for Verification and Analysis, Springer, pp. 1–12, doi:10.1007/ 978-3-642-24372-1_1. [6] Julien Forget, Frédéric Boniol, David Lesens & Claire Pagetti (2008): A Multi-Periodic Synchronous DataFlow Language. In: 2008 11th IEEE High Assurance Systems Engineering Symposium, Institute of Electrical and Electronics Engineers (IEEE), doi:10.1109/hase.2008.47. [7] N. Halbwachs, F. Lagnier & C. Ratel (1992): Programming and verifying real-time systems by means of the synchronous data-flow language LUSTRE. Software Engineering, IEEE Transactions on 18(9), pp. 785–793, doi:10.1109/32.159839. [8] Kohei Honda (1993): Types for dyadic interaction. In: CONCUR’93, Springer, pp. 509–523, doi:10.1007/ 3-540-57208-2_35. [9] Patrick M Rondon, Ming Kawaguci & Ranjit Jhala (2008): Liquid types. In: ACM SIGPLAN Notices, 43, ACM, pp. 159–169, doi:10.1145/1379022.1375602. [10] Jean-Pierre Talpin, Pierre Jouvelot & Sandeep Kumar Shukla (2015): Towards refinement types for timedependent data-flow networks. In: Formal Methods and Models for Codesign, 2015 ACM/IEEE International Conference on, IEEE, pp. 36–41, doi:10.1109/memcod.2015.7340465. [11] Hongwei Xi (2004): Applied Type System (extended abstract). In: post-workshop Proceedings of TYPES 2003, Springer-Verlag LNCS 3085, pp. 394–408, doi:10.1007/978-3-540-24849-1_25. [12] Hongwei Xi (2007): Dependent ML: An approach to practical programming with dependent types. Journal of Functional Programming 17(2), pp. 215–286, doi:10.1017/s0956796806006216. [13] Hongwei Xi & Frank Pfenning (1999): Dependent Types in Practical Programming. In: Proceedings of 26th ACM SIGPLAN Symposium on Principles of Programming Languages, doi:10.1145/292540.292560. 44 Dependent Types for Multi-Rate Flows [14] Eugene Yip, Matthew MY Kuo, Partha S Roop & David Broman (2014): Relaxing the synchronous approach for mixed-criticality systems. In: Real-Time and Embedded Technology and Applications Symposium, 2014 IEEE 20th, IEEE, pp. 89–100, doi:10.1109/rtas.2014.6925993.
6cs.PL
MNRAS 000, 1–13 (2017) Preprint 5 October 2017 Compiled using MNRAS LATEX style file v3.0 Effective Image Differencing with ConvNets for Real-time Transient Hunting Nima Sedaghat1? and Ashish Mahabal2 † 1 Department 2 Center of Computer Science, University of Freiburg, Georges-Koehler-Allee 052, 79110 Freiburg, Germany for Data Driven Discovery, Caltech, 1200 E California Blvd., Pasadena, CA 91125 arXiv:1710.01422v1 [astro-ph.IM] 4 Oct 2017 Accepted XXX. Received YYY; in original form ZZZ ABSTRACT Large sky surveys are increasingly relying on image subtraction pipelines for real-time (and archival) transient detection. In this process one has to contend with varying PSF, small brightness variations in many sources, as well as artifacts resulting from saturated stars, and, in general, matching errors. Very often the differencing is done with a reference image that is deeper than individual images and the attendant difference in noise characteristics can also lead to artifacts. We present here a deep-learning approach to transient detection that encapsulates all the steps of a traditional image subtraction pipeline – image registration, background subtraction, noise removal, psf matching, and subtraction – into a single real-time convolutional network. Once trained the method works lighteningly fast, and given that it does multiple steps at one go, the advantages for multi-CCD, fast surveys like ZTF and LSST are obvious. Key words: Transient — Supernova — Deep Learning — Artificial Inteligence – Convolutional Network – ConvNet 1 INTRODUCTION Time-domain studies in optical astronomy have grown rapidly over the last decade with surveys like ASAS-SN (Pojmański 2014), CRTS (Mahabal et al. 2011; Djorgovski et al. 2011; Drake et al. 2009), Gaia (Gaia Collaboration et al. 2016), Palomar-Quest (Djorgovski et al. 2008), PanSTARSS (Chambers et al. 2016), PTF (Law et al. 2009) etc. to name a few. With bigger surveys like ZTF (Bellm 2014) and LSST (Ivezic et al. 2008) around the corner, there is even more interest in the field. Besides making available vast sets of objects at different cadences for archival studies, these surveys, combined with fast processing and rapid follow-up capabilities, have opened the doors to an improved understanding of sources that brighten and fade rapidly. The real-time identification of such sources - called transients is, in fact, one of the main motivation of such surveys. Examples of transients include extragalactic sources such as the supernovae, and flaring M-dwarf stars within our own Galaxy, to name just two types. The main hurdle is identifying all such varying sources quickly (completeness), and without artifacts (contamination). The identification process is typically done by comparing the latest image (hereafter called the science image), with an older image of the same area of the sky (hereafter called the reference image). The ? E-mail: [email protected] (NS) † E-mail: [email protected] (AAM) © 2017 The Authors reference image is often deeper so that fainter sources are not mistaken as transients in the science image. Some surveys like CRTS convert the images to a catalog of objects using source extraction software (Bertin & Arnouts 1996), and use the catalogs as their discovery domain, comparing brightness of objects detected in the science and reference images. Other surveys like PTF directly difference the reference and science images after proper scaling and look for transients in the difference images. The reference and science images differ in many ways: (1) changes in the atmosphere mean the way light scatters is different at different times. This is characterized by the point spread function (PSF), (2) the brightness of the sky changes depending on the phase and proximity to the moon, (3) the condition of the sky can be different (e.g. very light cirrus), and (4) the noise and depth (detection limit for faintest sources) are typically different for the two images. As a result, image differencing is non-trivial, and along with real transients come through a large number of artifacts per transient. Eliminating these artifacts has been a bottleneck for past surveys, with humans having been often employed to remove them one by one – a process called scanning – in order to shortlist a set of genuine objects for follow-up using the scarce resources available. Here we present an algorithm based on deep learning that almost completely eliminates artifacts, and is nearly complete (or can be made so) in terms of real objects that it finds. In Sec. 2 we describe prior art for image differencing, and on deep learning in astron- 2 Nima Sedaghat Encoder Decoder Figure 1. Our CNN-based encoder-decoder network, TransiNet, produces a difference image without an actual subtraction. It does so through training using a labeled set of transients as the ground-truth. omy. In Sec. 3 we describe the image differencing problem in greater detail, in Sec. 4 we present our method and a generative encoder-decoder network – called TransiNet hereafter – based on convolutional networks (ConvNets or CNNS), in Sec. 5 we detail the experiments we have carried out, and in Sec. 7 we discuss future directions. 2 RELATED WORK For image differencing some of the programs that have been used include Alard & Lupton (1998), Bramich (2008), and PTFIDE (Masci et al. 2017). A recent addition to the list is ZOGY (Zackay et al. 2016) which apparently has lower contamination by more than an order of magnitude. It is to be used with the ZTF pipeline and at least in parts of the LSST pipeline. The main task of such an algorithm is to identify new point sources (convolved by the PSF). The problem continues to be challenging because it has to take in to consideration many complicating factors. Besides maximizing real sources found (true positives), generating as clean an image as possible (fewest false positives) is the quantifiable goal. Please refer to Zackay et al. (2016) for greater detail. Neural networks in their traditional form have been around since as early as 1980s e.g. Rumelhart & Hinton (1986) and LeCun (1985). Such classical architectures have been used in astronomical applications in the past. One famous example is the star-galaxy classifier embedded into the SExtracor package (Bertin & Arnouts 1996). The advent of convolutional neural networks (ConvNets: LeCun et al. (1990, 1998)), followed by the advances in parallel computing hardware (Raina et al. 2009), has started a new era in ‘deep’ convolutional networks, specifically in the areas of image processing and computer vision. The applications span from pixel-level tasks such as de-noising to higher-level tasks such as detecting and recognizing multiple objects in a frame. See, e.g. Krizhevsky et al. (2012) and Simonyan & Zisserman (2014). Researchers in the area of astrophysics have also very recently started to utilize deep learning-based methods to tackle astronomical problems. Deep Learning has already been used for galaxy classification (Hoyle 2016), supernova classification (Cabrera-Vives et al. 2016), light curve classification (Mahabal et al. 2017; Charnock & Moss 2017), identifying bars in galaxies (Abraham 2017), separating Near Earth Asteroids from artifacts in images (Bue 2017), transient selection post image differencing (Morii et al. 2016), Gravitational Wave transient classification (Mukund et al. 2017), and even classifying noise characteristics (Zevin 2017; Abbott 2017; George et al. 2017). One aspect of ConvNets that has not received enough attention in the astrophysical research community, is the ability to generate images as output (Rezende et al. 2014; Bengio et al. 2013). We provide here such a generative model to tackle the problem of contamination in difference images (see Fig. 1) and thereby simplify the transient follow-up process. 3 PROBLEM FORMULATION We cast the transient detection problem as an image generation task. In this approach, the input is composed of a pair of images (generally with different depth, and seeing aka FWHM of the PSF) and the output is an image containing, ideally, only the transient at its correct location and with a proper estimation of the difference in magnitudes. In this work we define a transient as a point source appearing in the second/science input image, and not present in the first/reference image. Such a generative solution as we propose naturally has at its heart registration, noise-removal, sky subtraction, and PSF-matching. In the computer vision literature, this resembles a segmentation task, where one assigns a label to each pixel of an image, e.g. transient vs. non-transient. However our detections include information about the magnitude of the transients and the PSF they are convolved with, in addition to their shape and location. Therefore the pixel-values of the output are real-valued (or are in the same space as the inputs, making it a different problem than a simple segmentation (see Fig. 2). To this end, we introduce an approach that is based on deep-learning, and train a convolutional neural network (ConvNet) to generate the expected output based on the input image pair. MNRAS 000, 1–13 (2017) Transient Hunting 3 back-propagate errors computed based on all different resolutions of the output during training. But in the end and for evaluation purposes, we only consider the full-resolution output. This multi-scale strategy helps the network learn better features with different levels of detail. We use an L1 loss function at each output: E= Figure 2. Examples of the reference (left) and science (center) images. The image on the right is the ground truth output defined for this image pair. It contains the image of a single transient, completely devoid of background and noise. The profile of the transient is the best match to reality our model can produce. We formulate the problem as follows. Let us consider (I1, I2) as the reference-science pair: I1 = I0 ∗ φ1 + S1 + N1 (1) I2 = (I0 + It ) ∗ φ2 + S2 + N2 (2) where I0 is the underlying unconvolved image of the specific region of the sky; φ1 and φ2 represent the PSF models; S1 and S2 are the sky levels and N1 , N2 represent the noise. Note that for the sake of readability, we have illustrated the effect of noise as a simple addition operation. However in reality the noise is ‘applied’ per pixel throughout the workflow. It is the ideal model for the transient, and can be seen as an empty image with an ideal point-source on it. Based on our formulation of the problem, the answer we seek is It ∗ φ2 , which represents the image containing the transient, in the same seeing conditions as the science image. This involves PSF matching for taking the first image from I0 ∗ φ1 to I0 and then to I0 ∗ φ2 , for the subtraction to work. Note that in eqs. (1) and (2), for the sake of clarity, the two images are assumed to be registered. In the real problem that the network is trying to solve, 2 is replaced by: I2 = D{(I0 + It ) ∗ φ2 } + S2 + N2 (3) in which D{} represents spatial inconsistency, which in its simplest form consists of one or more of small rotation, translation and scaling. 4 METHOD We tackle the problem using a deep-learning method, in which an encoder-decoder convolutional neural network is responsible for inferring the desired difference image based on the input pair of images. 4.1 Network Architecture We illustrate TransieNet in Fig. 3. It is a fully-convolutional encoder-decoder architecture inspired by the one introduced in Sedaghat et al. (2017). Ten convolutional layers are responsible for the contraction throughout the encoder, and learn features with varying levels of detail in a hierarchical manner. The expansion component of the network consists of 6 up-convolutional layers which decode the learned features, step by step, and generate estimates of the output with different resolutions along the way. We compute and MNRAS 000, 1–13 (2017) N 1 Õ | yˆn − yn | N (4) n=1 where ŷ and y represent the prediction and the target (ground truth) respectively, and N is the number of samples in each mini-batch – see Sec. 4.3. The reason behind the choice of L1 loss over it’s more popular counterpart, L2 or Euclidean loss, is that the latter introduces more blur into the output, ruining pixel-level accuracy – see Sedaghat et al. (2017), Mathieu et al. (2015). 4.2 Data Preparation Neural networks are in general data-greedy and require a large training dataset. TransiNet is not an exception and in view of the complexity of the problem – and equivalently the architecture – needs a large number of training samples: reference-science image pairs + their corresponding groundtruth images. Real astronomical image pairs with transients are not so publicly available. Difficulty of providing proper transient annotations makes them even scarcer. The best one could do is to manually (or semi-automatically) annotate image pairs, and find smart ways to estimate a close-tocorrect ground truth image: a clean difference image with background-subtracted gradients. Although as explained in Sec. 4.2.1 we implement and prepare such a real training set, it is still too small (∼ 200 samples) and if used as is, the network would easily overfit it. One solution is to use image augmentation techniques, such as spatial transformations, to virtually increase the size of the training set. This trick, though necessary, is still not sufficient in our case with only a few hundred data samples – the network eventually discovers common patterns and overfits to the few underlying real scenes. An alternative solution is to generate a large simulated (aka synthetic) dataset. However, relying only on the synthetic data makes the network learn features based on the characteristics of the simulated examples, making it difficult to transfer the knowledge to the real domain. Our final solution is to feed the network with both types of data: synthetic samples mixed with real astronomical images of sky with approximate annotations. This along with online augmentation, makes a virtually infinite training set, which has the best of both worlds. We describe details about the datasets used and the training strategy in the following sub-sections. 4.2.1 Real Data For real examples we make use of data from the Supernova Hunt project (Howerton 2017) of the CRTS survey. In this project image subtraction is performed on pairs of images of galaxies in search of supernovae. While this may bias the project towards finding supernovae rather than generic transients, that should not affect the end result as we mark 4 Nima Sedaghat Input Conv1 7x7 : 128 Conv2 5x5 : 256 Conv3 5x5 : 512 Conv3_1 Conv4 3x3 : 512 3x3 : 1024 Conv4_1 3x3 : 1024 Conv5 3x3 : 1024 Conv5_1 3x3 : 1024 Conv6 3x3 : 2048 Conv6_1 3x3 : 2048 L1 Loss Output Upconv0 4x4 : 32 Upconv1 4x4 : 64 Upconv2 4x4 : 128 Upconv3 4x4 : 256 Upconv4 4x4 : 512 Upconv5 4x4 : 1024 Figure 3. Our suggested fully convolutional encoder-decoder network architecture. The captions on top/bottom of each layer show the kernel size, followed by the number of feature maps. Each arrow represents a convolution layer with a kernel size of 3x3 and stride and padding values of 1, which preserves the spatial dimensions. Dotted lines represent the skip connections. Low-resolution outputs are depicted on top of each up-convolution layer with the corresponding loss. After each (Up-)convolution layer there is a ReLU layer which is not displayed here. the transients found, and the ground-truth images contain just the transients. If anything, finding such blended point sources should make finding point sources in the field (i.e. away from other sources) easier. Unlike most other surveys, the CRTS images are obtained without a filter, but that too is not something that directly concerns our method. We gathered 214 pairs of publicly available jpeg images from SN Hunt and split this dataset in to training, validation and test subsets of 102, 26 and 86 members respectively. The reference images are typically made by stacking ∼ 20 older images of the same area. The science image is a single 30-second exposure. The pixels are 2 00 .5 × 2 00 .5 and thus comparable or somewhat bigger than the typical PSF. Individual images are 120x120 pixels, and at times not perfectly registered. To prepare the ground truth, we developed an annotation tool. The user needs to roughly define the location of the transient in the science image, by comparing it with the reference image, and put a circular aperture around it. Then the software models the background and subtracts it from the aperture to provide an estimate of the transient’s shape and brightness. Simple annulus-based estimates of the local background (Davis 1987; Howell 1989) or even the recent Aperture Photometry Tool (Laher et al. 2012), are not suitable for most of the samples of this dataset since the transients, often supernovae, naturally overlap their host galaxies. Therefore we use a more complex model and fit a polynomial of degree 8 to a square-shaped neighborhood of size 2r × 2r around the aperture, where r is the radius of the user-defined aperture. Note however that the model fitting is performed only after masking out the aperture, to exclude the effect of the transient itself – the points are literally excluded from model-fitting – rather than being masked and replaced with a value such as zero. This method works reasonably well even when the local background is complex. Fig. 4 illustrates the process. The annotations on real images are not required to be accurate, as the main responsibility of this dataset is to provide the network with real examples of the sky. This lack of accuracy is compensated by the synthetic samples with precise positions. 4.2.2 Synthetic Data To make close-to-real synthetic training samples, we need realistic background images. Existing simulators, such as Skymaker (Bertin 2009), do not yet provide with a diverse set of galaxy morphologies, and therefore are not suitable for our purpose. Instead, we use images from the Kaggle Galaxy Zoo dataset1 , based on the Galaxy Zoo 2 dataset (Willett et al. 2013), for our simulations. To this end, we pick single images as the background image and create a pair of reference-science images based on it. This method also makes us independent of precise physical simulation of the background, allowing us to focus on simulations only at the image level – even for the ‘foreground’, i.e. transients. This may result in some samples that do not exactly resemble a ‘normal’ astronomical scene, in terms of the magnitude and location of the transient, or the final blur of the objects. But that is even better in a learning-based method, as the network will be trained on a more general set of samples, and less prone to over-fitting to specific types of scenes. Fig. 5 illustrates details of this process. We first augment the background image using a random 1 https://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge MNRAS 000, 1–13 (2017) Transient Hunting 5 tively. U shows a normal distribution and N is a 2D normal distributions. T is then a 2D vector and its values show a translation proportional to the dimensions of the image. At the next step simulated transients are added to the science (second) image as ideal point sources, with random locations and magnitudes, to form I0 + It . The transient locations are again sampled from a 2D gaussian distribution. The distribution parameters are adjusted such that transients, although scattered all around the image, happen mostly in the vicinity of galaxies at the center of the image, to resemble real supernovae:     0.1 0  Xt , Yt ∼ N µ = 0, Σ = (7) 0 0.1 In most of our experiments we simulated only a single transient. But in cases where we had more of them, we made sure they were apart from each other by at least half of the bigger dimension of the image. The amplitude of the simulated source is also randomly chosen as:  At ∼ N µ = 10, σ = 0.3 (8) This value, after being convolved with the (sum-normalized) PSF, will constitute the flux of the transient (Ft ). We can select a specific range of At for training – to fine-tune the network – based on the range of transients (and their relative brightening) that we expect to find for a given survey. The two images are then convolved with different gaussian PSFs, generated based on random kernel parameters, with a random eccentricity, limited by a user-defined maximum:  σφ, x ∼ U σφ,m, σφ, M (9) p σφ,y = σφ1, x 1 − ecc2 (10)  ecc ∼ N µ = 0, σ = ecc max (11) Figure 4. An exemplar transient annotation case. From top-left to bottom-right the first two images are the input ref/science pair. Number 3 illustrates the 2r × 2r neighborhood of the transient and on 4 the user-defined aperture is masked out. Image 5 shows the polynomial model fit to the ‘masked-neighborhood’. Note that since the blank aperture is excluded from the fitting process, there’s no dark region in the results. In number 6 the estimated background is subtracted from the masked neighborhood to form a measure of how well the background has been modeled: the more uniform and dark this image is, the better the polynomial has modeled the background. Finally in number 7 the estimated background is subtracted from the neighborhood and the transient stands out. In number 8 the transient is cropped out of 7 using the user-defined aperture. spatial transformation:  R ∼ U 0, 2π    0.03 0  T ∼ N µ = 0, Σ = 0 0.03 (5) In (x, y) = poisson(λ = I(x, y) + S) (6) where poisson is a function returning a sample from a Poisson distribution with the given λ parameter, S is the sky model, and In is the noisy version of input I. here R & T represent rotation and translation (shift) respecMNRAS 000, 1–13 (2017) where σφ,x and σφ,y are the standard deviations of the 2D gaussian function along x and y directions respectively and [σφ,m, σφ, M ] is the range from which σφ,x is uniformly sampled. The PSF is then rotated using a random value, θ φ , uniformly sampled from [0, 2π]. This should also help catch asteroids that would leave a very short streak. Precisely modeling the difference between reference and science images, and adjusting the PSF parameters’ distributions accordingly, would be achievable. However, as stated before, we prefer to keep the training samples as general as possible. Therefore in our experiments [σφ,m, σφ, M ] is set to [2, 5] for both images. These numbers are larger than typically encountered, and real images should fare better. The ecc max value is set to 0.4 and 0.6 for reference and science images, respectively, to model the more isotropically blurred seeing of reference images. The sky and noise levels are different for the reference and science images. We choose to model these difference in our simulations since in contrast to the previous parameters, ignoring them would make learning easier for the network – and that’s exactly what we want to avoid. We model the sky level, S, as a constant value, add it to the image and only after that ‘apply’ the Poisson noise to every pixel: (12) 6 Nima Sedaghat x x Spatial Augmentation + Random PSF #1 Random PSF #2 Random Sky + Poisson Noise Random Sky + Poisson Noise Input #1: Ground Truth: Input #2: Figure 5. The synthetic sample generation procedure. The notations used here are described in Equations (1) and (2). MNRAS 000, 1–13 (2017) Transient Hunting Then we perform a pairwise augmentation (rotation, scaling and translation), such that the two images are not perfectly registered. This forces the network to also learn the task of registration on the fly. The ground truth image is then formed by convolving the ideal transient’s image, It , with the same PSF as applied to the science image. No constant sky value or noise are applied to this image. This way the network learns to predict transient locations and their magnitudes in the same seeing conditions as the science image, in addition to noise removal and sky subtraction. 4.3 Training Details We have two networks, one generic, and the other specialized to the SN Hunt dataset. Each shares the 90K images from Kaggle zoo, with transients inserted to create synthetic science images. The real data currently is just the ∼ 100 SN Hunt image pairs earmarked for training. The zoo and SN Hunt images are further rotated, shifted, and scaled to augment the dataset, and also make the network more robust. Training is done in small batches of 16. We use ADAM for optimization using the Caffe framework (Jia et al. 2014). First 90K iterations are common to both networks. Images used are 140 × 140. Then, for the generic network fine tuning is done using batches that contain 12 CRTS and 4 zoo images. The image size here is 256 × 256 (with images scaled where needed), and the number of iterations is 50K. We put 12 real images and 4 synthetic images in each batch during this second round of training to prevent over-fitting to the small-sized real dataset. We start with a learning rate of 3e−4 and drop it in the second round by a factor of 0.3 every 20K iterations. Running on an NVIDIA GTX 1070 along with 16 CPU cores, the whole training process takes a day and half to complete. For the specialized SN Hunt network, on the other hand, the entire fine-tuning is done using 8K iterations on CRTS images (we emphasize that we are still using just the ∼ 100 image-pairs, modified in many ways). That way it is better at recognizing transients in real data. 4.4 The Attention Trick In this specific type of application, the target images mainly consist of black regions (i.e. zero intensity pixels), and nonzero regions take up only a small number of pixels. Therefore mere use of a simple L1 loss does not generate and propagate big enough error values back to the network, when the network has learned to remove the noise and generates blank images. So the network spends too long a time focusing on generating blank images instead of the desired output, and in some cases fails to even converge. The trick we use to get around this issue is to conditionally boost the error on the interesting regions. The realization of this idea is to simply apply the mapping [0, 1] → [0, K] on the ground-truth pixel values. K represents the boosting factor and we set it to 100 in our experiments. This effectively boosts the error in nonzero regions of the target, virtually increasing the learning rate for those regions only. The output of the network is later downscaled to lie in the normal range. Note that increasing the total learning rate is not an alternative solution, as the network would go unstable and would not even converge. MNRAS 000, 1–13 (2017) 5 7 EXPERIMENTS AND RESULTS We have run TransiNet on samples from CRTS SN Hunt and the Kaggle zoo dataset. The network weights take up about 2GB of memory. Once read, on the NVIDIA GTX 1070 the code runs fast: 39ms per sample, which can be reduced to 14ms if samples are passed to the network as batches of 10. The numbers were calculated by running tests on 10000 images three times. Fig. 6 depicts samples from running TransiNet on the CRTS test subset. The advantage of TransiNet is that the “image differencing” produces a noiseless image ideally consisting of just the transient, and thus robust to artifacts and removes the need for human scanners. With increasing CCD sizes it is much more likely than not that there will be multiple transients in a single image. Since the SN Hunt images, or the zoo images used, rarely have multiple transients, the networks may not be ideal when looking for such cases. However, because of the way the network is trained - with the output as pure PSFlike transients, it is capable of finding multiple transients though we did not train it explicitly with such cases. This is demonstrated in Fig. 7 which depicts an exemplar sample from the zoo subset. Here we introduced four transients, and all were correctly located. Another side effect - a good one - is that the network rejects non-PSF like additions, including Cosmic Rays. In addition to the four transients, we had also inserted 10 single pixel Cosmic Rays in the science image shown in Fig. 7 and all were rejected. An example from the SN Hunt set is shown in Fig. 8 which happens to have two astrophysical objects – the second is likely an asteroid. Here too, the network has detected both transients. Locating new asteroids is as useful as locating transients to help make the asteroid catalog more complete for future linking and position predicting. 5.1 Quantitative Evaluation We provide below quantitative evaluations of TransiNet performance. 5.1.1 Precision-Recall Curve Precision-recall curves are the de-facto evaluation tool for detectors. They capture TruePositives (TP, or ‘hits’, the number of correctly detected objects), FalsePositives (FP, or ‘false alarms’), and FalseNegatives (FN, or ‘misses’, the number of missed real objects) for all possible thresholds. This allows users to either set a fixed threshold, or a dynamic threshold (e.g. 5σ above the background level, or 70% of the max in a difference image etc.) TruePositive TruePositive + FalsePositive TruePositive Recall = TruePositive + FalseNegative Precision = (13) (14) Low-SNR Detections & Blank Outputs The output of TransiNet is an image with real-valued pixels. Therefore each pixel is more likely to contain a non-zero real value, even in the ‘dark’ regions of the image, or when there is no transient to detect at all. Thus, we consider low-SNR detection images 8 Nima Sedaghat Reference Science ZOGY - D ZOGY - (Scorr>5σ) TransiNet Figure 6. Image subtraction examples using ZOGY and TransiNet for a set of CRTS Supernova Hunt images. The first column has the deep reference images, second column contains the science images which have a transient source and are a shallower version of the reference images. The third column contains the ZOGY D images, and the fourth has the ZOGY Scorr images i.e. “the matched filter difference image corrected for source noise and astrometric noise” (Zackay et al. 2016). The fifth column has the thresholded versions of ZOGY SCORR, as recommended in that paper. The sixth column shows the difference image obtained using TransiNet. All images are mapped to the [0,1] range of pixel values, with a gamma correction on the last column for illustration purposes. TransiNet has a better detection accuracy, and is also robust against noise and artifacts. It is possible that ZOGY could be tuned to perform better, and on a different dataset provide superior results - the reason for the comparison here is to simply show that TransiNet does very well. MNRAS 000, 1–13 (2017) Transient Hunting 9 Positives respectively. Now we can rewrite Equations (13) and (14) in a more compact and formal form as: |TP| |TP| + |FP| |TP| Recall = |TP| + |FN| Precision = Figure 7. An exemplar multi-transient case from the zoo dataset. The reference image (left), science image (middle) with 10 singlepixel Cosmic Ray events, indicated by red circles, and four transients, and the network prediction (right) with all transients detected cleanly, and all CRs rejected. Figure 8. An exemplar multi-transient case from the CRTS SN Hunt dataset. The science image (middle) has two transients, and the network prediction (right) finds them both though never trained explicitly to look for multiple transients. as blank images. The outputs of the network (detection images) which have a standard deviation (σ) lower than 0.001 were marked as blank images during our experiments and unconsidered thereafter. Binariziation and Counting of Objects Evaluations at a series of thresholds is the essence of a precision-recall curve and helps reveal low SNR contaminants, while digging for higher completeness (see Fig. 9). The thresholding ( 0 ŷi j < τ Ŷi j = (15) 1 ŷi j ≥ τ results in the binary image, Ŷ , on which we obtain ‘connected’ regions to count detected objects with full connectivity (Fiorio & Gustedt 1996; Wu et al. 2005). For this specific kind of evaluation, we also convert the ground truth image (y) to a similar binary-valued image, Y , using a fixed threshold. Let P be the set of all positives, i.e. the objects in Ŷ , and G the set of all objects in Y . Then we have TP = P ∧ G (16) where ∧ is used here to denote spatial intersection, such that TP is the set of objects in P that have an spatial intersection with a member of G. TP is the set of True Positives. We conversely define TP 0 = G∧P which is of the same cardinality as TP and includes the set of objects in G that have been detected. Then we also have: FP = P − TP (17) FN = G − TP 0 (18) in which FP and FN stand for False Positives and True MNRAS 000, 1–13 (2017) (19) (20) where | · | represents the cardinality of the set. We also define completeness and contamination measures as follows: |TP| = Recall |TP| + |FN | |FP| = 1 − Precision Contamination = |TP| + |FP| Completeness = (21) (22) Figs. 10 and 11 depict the precision-recall curves corresponding to the two versions of TransiNet before and after blanking. Each curve is obtained by sweeping the threshold (τ) in the pixel-value domain. Starting from the minimum (0), Ŷ is set to 1 everywhere, resulting in a 100% recall (everything that is to be found is found) with a close-to-zero precision (too many false positives), which is equivalent to total contamination. But as we increase τ, fewer pixels in Ŷ ‘fire’, generally resulting in a lower recall (some misses) and higher precision (far fewer contaminants) – see Fig. 9. To generate the curves we sampled 101 logarithmically-distributed values for τ from the range [10−4 σ, 100σ], where σ is the standard deviation of the pixel values in each detection image (y). Also the ground truth images were binarized with a fixed threshold of 10−3 . The sharp and irregular behavior of the curve at around 75% of recall on the CRTS dataset, is due to the low contamination levels in the output: transients are detected with a high significance. Contaminants, if any, have a much lower intensity, and their number goes up only when one pushes for high completeness to the lower significance levels. 5.1.2 Relative Magnitude of the Transient Thanks to the freedom in generation of synthetic samples with different parameters, we can evaluate the performance of the network with transients of different magnitudes. However, for this evaluation we use relative magnitudes, as opposed to the absolute intensities used during training. This would make it easier to quantitatively determine the ability of the network to detect faint transients without contamination. In the future we hope to incorporate similar process during training as well. We define the relative magnitude as the difference of magnitudes at the location of transient, with and without the transient: magrel = −2.5 log10 (Frel ) Ft + Flocal Frel = Flocal (23) (24) where Ft is the absolute flux of the transient, and Flocal represents the flux of the background, before having the transient. The latter is measured inside an FWHM-sized square neighborhood around the location of the transient. Fig. 12 depicts the performance of the detector for several relative magnitudes, in terms of the precision-recall curve. With higher visibility, the curve approaches the ideal 10 Nima Sedaghat Ref. Image Sci. Image Raw Det. Image τ= σ 10000 τ= σ 1000 τ= σ 100 τ= σ 10 τ=σ τ = 10σ Figure 9. Visualization of the thresholding process used for generation of precision-recall curves. Each row illustrates exemplar levels of thresholding of a single detection image: first row is chosen from the synthetic subset and the second row from CRTS. Outputs of the network are normally quite clean, and contaminants practically appear only after taking the threshold down below the noise level. This is particularly visible in the second row, where the transient has been of a low magnitude, and so the detection image has a low standard σ deviation (σ). Thus 100 is still too low and below noise level. form. Considering that during the training phase, the network has rarely seen transients with such low magnitudes as the ones in the lower region of this experiment, it is still performing well. We expect it to gain much better results by broadening the range of simulated transient amplitudes during training. 5.1.3 Robustness to Spatial Displacements We analyze the robustness of the TransiNet to pairwise spatial inconsistencies between the science and reference images. That way small rotations, WCS inconsistencies etc. do not give rise to Yin-Yang like ‘features’ and lead to artifacts. To this end, for a subset of image pairs, we exert manual shift, rotation and scaling to one of the images in each pair, and pass them through the network. Fig. 13 depicts the results of these experiments as plots of completeness and contamination vs. the manual perturbation. 5.1.4 Performance of the two TransiNet Networks Table 1 summarizes the testing results for the two TransiNet networks. For new surveys one can start with the generic network, and as events become available, fine-tune the network with specific data. 5.2 Comparison with ZOGY Given the generative, and hence very different, nature of our ‘pipeline’, it is difficult to compare it with direct image differencing pipelines. We have done our best by comparing the output of TransiNet and of ZOGY for synthetic as well as real images. We used the publicly available MATLAB version of ZOGY, and almost certainly we used ZOGY in a sub-optimal fashion. As a result this comparison should be taken only as suggestive. More direct comparisons with real data (PTF and ZTF) are planned for the near-future. Fig. 6 depicts the comparison for a few of the SN Hunt transients. Both pipelines could be run in parallel to choose an ideal set of transients, since the overhead of TransiNet is minuscule. Network Synth/Zoo CRTS/SNH Transients 100 86 TP 100 65 FP 0 1 FN 0 21 Prec. 100.0 98.4 Recall 100.0 75.5 Table 1. Hits and misses for TransiNet for the Synthetic and SN Hunt networks. TransiNet does very well for synthetics. One reason could be that there isn’t enough depth variation in the reference and science images. But for CRTS too the output is very clean for the recall of 76% that it achieves. The lower (than perfect) recall could be due to a smaller sample, larger pixels, large shifts in some of the cutouts etc. Fine tuning with more data can improve performance further. The fixed thresholds used for the synthetic and SN Hunt networks are 40 and 20 respectively. 6 FUTURE DIRECTIONS We have shown how transients can be effectively detected using TransiNet. In using the two networks we described, one with the Kaggle zoo images, and another with CRTS, we cover all broad aspects required, and yet for this method to work with any specific project, e.g. ZTF, appropriate tweaks will be needed, in particular labeled examples from image differencing generated by that survey. Also, the assumptions during simulations can be improved upon by such examples. Using labeled sets from surveys accessible to us is definitely the next step. Since the method works on the large pixels that CRTS has, we are confident that such experiments will improve the performance of TransiNet. The current version produces convolved transients to match the shape and PSF of the science image. One can modify the network to produce just the transient location and leave the determination of other properties to the original science and reference images as they contain more quantitative detail. Further the network could be tweaked to find variable sources too. But for that a much better labeled non-binary training set will be needed. In the same manner, it can also be trained to look for drop-outs, objects that have vanished in recent science images but were present in the corresponding reference images. This is in fact an inverse of the transients problem, and somewhat easier to do. In terms of reducing the number of contaminants even further, one can provide as input not just the pair of science MNRAS 000, 1–13 (2017) Transient Hunting TransiNet (CRTS) 1.0 TransiNet (CRTS) 1.0 1.0 0.8 0.6 0.6 0.2 0.2 0.4 0.6 0.8 1.0 1.0 Contamination 0.8 0.6 0.4 0.2 0.00.0 0.00.0 1.0 Recall TransiNet (CRTS) 0.2 0.4 0.6 Completeness TransiNet (CRTS) 1.0 0.8 1.0 Recall Precision 0.8 0.2 0.4 0.6 0.8 0.4 0.2 0.2 0.4 0.6 0.8 0.4 0.6 0.8 0.4 0.00.0 1.0 Recall TransiNet (CRTS) 1.0 0.6 0.4 0.2 0.2 0.4 0.6 Completeness TransiNet (CRTS) 1.0 Recall Precision 0.8 0.2 0.8 0.00.0 1.0 Completeness TransiNet (Synthetic) 0.8 1.0 Recall Precision 0.8 0.00.0 1.0 0.4 0.4 0.4 0.2 0.2 0.2 0.2 0.00 100 0.00 0.00 1.0 1.0 0.8 40 60 Completeness Contamination 0.4 0.2 20 40 / 60 80 80 / TransiNet (Synthetic) 100 0.6 0.4 0.4 0.2 0.2 20 40 / 60 80 and reference images, but also pairs of the rotated (by 90, 180, 270 degrees) and flipped (about x- and y-axes) versions. The expectation is that the transient will still be detected (perhaps with a slightly different peak, extent), but the weak contaminants, at least those that were possibly conjured by the weights inside the network, will be gone (perhaps replaced by other – similarly weak ones – at a different location), and the averaging of the detections from the set will leave just the real transient. Another way to eliminate inhomogeneities in network weights is to test it with image pairs without any transients. While most image pairs do not have any transient except in a small number of pixels, such a test can help streamline the network better. In order to detect multiple transients, one could cut the image in to smaller parts and provide these submimages for MNRAS 000, 1–13 (2017) 60 / TransiNet (CRTS) 80 0.2 0.4 0.6 0.8 1.0 Completeness TransiNet (Synthetic) Recall Precision 100 0.00 1.0 20 40 60 Completeness Contamination 80 / TransiNet (Synthetic) 100 Completeness Contamination 0.8 0.6 0.4 0.2 0.00 100 Figure 10. Plots showing precision-recall (row 1), completenesscontamination (row 2), and their dependence on threshold (rows 3 and 4) for TransiNet before the blanking is done to remove ultra low-SNR detections (see text). The two columns show CRTS (left) and synthetic (right) subsets. For CRTS a threshold can be picked where 80% transients are detected with little contamination. Not unexpectedly, the performance is better for the synthetic images. 40 0.8 0.6 100 0.00 20 1.0 Completeness Contamination 0.8 0.6 0.00 20 1.0 0.8 0.4 80 0.8 0.2 0.6 60 0.6 Recall TransiNet (Synthetic) 0.4 0.6 / TransiNet (CRTS) 0.4 0.6 0.6 40 0.2 0.8 0.6 20 TransiNet (Synthetic) 0.2 1.0 0.6 1.0 0.00.0 1.0 Recall TransiNet (Synthetic) 0.8 0.00.0 0.4 0.2 Contamination 0.00.0 0.4 Contamination 0.4 Precision 0.8 0.6 Precision 0.8 0.6 0.2 Contamination TransiNet (Synthetic) 0.8 Precision Precision 1.0 11 20 40 / 60 80 100 0.00 20 40 / 60 80 100 Figure 11. The plots are as in Fig. 10, but post-blanking. The plots are now closer to ideal. In this scenario, for CRTS, we never go above a completeness of 80% detections, but all those detections are clean, and the ones we miss are the really low significance ones below the blanking threshold of 0.001. Fig. 9 shows a single transient field for each type at different thresholds. detection. Another possibility is to mask the ‘best’ transient, and rerun the pipeline to look for more transients iteratively until none is left. An easier fix is to train the network for larger images, and for multiple transients in each image pair. Another way to improve the speed of the network is to experiment with the architecture, and if possible obtain a more lightweight network with a smaller footprint that performs equally well. Finally, the current network used jpegs with limited dynamic range as inputs. Using non-lossy FITS images should improve performance of the network. 7 CONCLUSIONS We have introduced a generative method based on convolutional neural networks for image subtraction to detect transients. It is superior to other methods as it has a higher 12 Nima Sedaghat method to other wavelengths like radio and use for surveys including SKA and its path-finders. 1.0 0.8 Precision ACKNOWLEDGEMENTS 0.6 AM was supported in part by the NSF grants AST-0909182, AST-1313422, AST-1413600, and AST-1518308, and by the Ajax Foundation. mag:-0.05 mag:-0.10 mag:-0.20 mag:-0.30 mag:-0.40 mag:-0.50 0.2 0.4 0.4 0.2 0.0 0.0 REFERENCES 0.6 0.8 Recall 1.0 Figure 12. Precision-recall curves for a range of magnitudes. These are for the synthetic transients where we had control over the relative magnitudes. The network misses more transients as the relative magnitude goes lower. This is not unexpected as the network has not seen such faint samples during training. The sharp vertical transitions reflect the clean nature of the detection images. 100 80 60 (%) (%) 100 completeness contamination 80 40 20 00 60 completeness contamination 40 20 1 2 3 shift (pixels) 4 04 5 100 2 1 0 1 2 rotation (degrees) 3 4 completeness contamination 80 hits (%) 3 60 40 20 00.90 0.95 1.00 1.05 scale/zoom 1.10 Figure 13. Robustness of the network to shift (top-left), rotation (top-right), and scaling (bottom) between the reference and science images. Ideally there will be no misalignments, but some can creep in through improper WCS, or changes between runs etc. The peak detectability is close to 75% across the board. completeness at lower thresholds, and at the same time has fewer contaminants. Once the training is done with appropriate labeled datasets, execution on individual images is fast. It can operate on images of any size (after appropriate training), and can be easily incorporated in to real-time pipelines. While we have not explicitly tested the method on high-density fields (e.g. closer to the plane of the Galaxy) it will be possible to get good performance once a corresponding labeled dataset is used for training. We hope surveys like ZTF, LSST as well as those with larger pixels like ASASSN (Shappee et al. 2014), Evryscope (Law et al. 2015) etc. adapt and adopt the method. It is also possible to extend the Abbott B.P. et. al. p., 2017, Phys. Rev. Lett. 118, 221101 Abraham S. et. al. i., 2017, Submitted Alard C., Lupton R. H., 1998, ApJ, 503, 325 Bellm E., 2014, in Wozniak P. R., Graham M. J., Mahabal A. A., Seaman R., eds, The Third Hot-wiring the Transient Universe Workshop. pp 27–33 (arXiv:1410.8185) Bengio Y., Yao L., Alain G., Vincent P., 2013, in Advances in Neural Information Processing Systems. pp 899–907 Bertin E., 2009, Memorie della Societa Astronomica Italiana, 80, 422 Bertin E., Arnouts S., 1996, ] 10.1051/aas:1996164, 117, 393 Bramich D. M., 2008, MNRAS, 386, L77 Bue Brian et al. i., 2017, In Prep. Cabrera-Vives G., Reyes I., FÃűrster F., EstÃl’vez P., Maureira J.-C., 2016, IJCNN, pp 251– Chambers K. C., et al., 2016, preprint, (arXiv:1612.05560) Charnock T., Moss A., 2017, ApJ, 837, L28 Davis L., 1987, National Optical Astronomy Observatories. Revised October Djorgovski S. G., et al., 2008, Astronomische Nachrichten, 329, 263 Djorgovski S. G., et al., 2011, preprint, (arXiv:1110.4655) Drake A. J., et al., 2009, ApJ, 696, 870 Fiorio C., Gustedt J., 1996, Theoretical Computer Science, 154, 165 Gaia Collaboration et al., 2016, A&A, 595, A1 George D., Shen H., Huerta E. A., 2017, preprint, (arXiv:1706.07446) Howell S. B., 1989, Publications of the Astronomical Society of the Pacific, 101, 616 Howerton S. C., 2017, CRTS SNhunt: The First Five Years of Supernova Discoveries Hoyle B., 2016, ] https://doi.org/10.1016/j.ascom.2016.03.006 Ivezic Z., et al., 2008, preprint, (arXiv:0805.2366) Jia Y., Shelhamer E., Donahue J., Karayev S., Long J., Girshick R., Guadarrama S., Darrell T., 2014, arXiv preprint arXiv:1408.5093 Krizhevsky A., Sutskever I., Hinton G. E., 2012, in Advances in Neural Information Processing Systems. pp 1097–1105 Laher R. R., Gorjian V., Rebull L. M., Masci F. J., Fowler J. W., Helou G., Kulkarni S. R., Law N. M., 2012, Publications of the Astronomical Society of the Pacific, 124, 737 Law N. M., et al., 2009, PASP, 121, 1395 Law N. M., et al., 2015, PASP, 127, 234 LeCun Y., 1985, in , Proceedings of Cognitiva 85, Paris, France LeCun Y., Boser B. E., Denker J. S., Henderson D., Howard R. E., Hubbard W. E., Jackel L. D., 1990, in Advances in Neural Information Processing Systems. pp 396–404 LeCun Y., Bottou L., Bengio Y., Haffner P., 1998, 86, 2278 Mahabal A. A., et al., 2011, Bulletin of the Astronomical Society of India, 39, 387 Mahabal A., Sheth K., Gieseke F., Pai A., Djorgovski S. G., Drake A., Graham M., the CSS/CRTS/PTF Collaboration 2017, preprint, (arXiv:1709.06257) MNRAS 000, 1–13 (2017) Transient Hunting Masci F. J., et al., 2017, PASP, 129, 014002 Mathieu M., Couprie C., LeCun Y., 2015, arXiv:1511.05440 Morii M., et al., 2016, PASJ, 68, 104 Mukund N., Abraham S., Kandhasamy S., Mitra S., Philip N. S., 2017, Phys. Rev. D, 95, 104059 Pojmański G., 2014, Contributions of the Astronomical Observatory Skalnate Pleso, 43, 523 Raina R., Madhavan A., Ng A. Y., 2009, in Proceedings of the 26th Annual International Conference on Machine Learning. ACM, pp 873–880 Rezende D. J., Mohamed S., Wierstra D., 2014, arXiv preprint arXiv:1401.4082 Rumelhart D. E., Hinton G. E., 1986, NATURE, 323, 9 Sedaghat N., Zolfaghari M., Brox T., 2017, Technical report, Hybrid Learning of Optical Flow and Next Frame Prediction to Boost Optical Flow in the Wild. arXiv:1612.03777 Shappee B. J., et al., 2014, ApJ, 788, 48 Simonyan K., Zisserman A., 2014, preprint, (arXiv:1409.1556) Willett K. W., et al., 2013, MNRAS, 435, 2835 Wu K., Otoo E., Shoshani A., 2005, Lawrence Berkeley National Laboratory Zackay B., Ofek E. O., Gal-Yam A., 2016, ApJ, 830, 27 Zevin M. e. a., 2017, Class. Quantum Grav 34, 6 This paper has been typeset from a TEX/LATEX file prepared by the author. MNRAS 000, 1–13 (2017) 13
1cs.CV
Approximation Algorithms for Optimal Decision Trees and Adaptive TSP Problems arXiv:1003.0722v3 [cs.DS] 21 Apr 2017 Anupam Gupta∗ Viswanath Nagarajan† R. Ravi‡ Abstract We consider the problem of constructing optimal decision trees: given a collection of tests which can disambiguate between a set of m possible diseases, each test having a cost, and the a-priori likelihood of any particular disease, what is a good adaptive strategy to perform these tests to minimize the expected cost to identify the disease? This problem has been studied in several works, with O(log m)-approximations known in the special cases when either costs or probabilities are uniform. In this paper, we settle the approximability of the general problem by giving a tight O(log m)-approximation algorithm. We also consider a substantial generalization, the adaptive traveling salesman problem. Given an underlying metric space, a random subset S of vertices is drawn from a known distribution, but S is initially unknown—we get information about whether any vertex is in S only when it is visited. What is a good adaptive strategy to visit all vertices in the random subset S while minimizing the expected distance traveled? This problem has applications in routing message ferries in ad-hoc networks, and also models switching costs between tests in the optimal decision tree problem. We give a poly-logarithmic approximation algorithm for adaptive TSP, which is nearly best possible due to a connection to the well-known group Steiner tree problem. Finally, we consider the related adaptive traveling repairman problem, where the goal is to compute an adaptive tour minimizing the expected sum of arrival times of vertices in the random subset S; we obtain a poly-logarithmic approximation algorithm for this problem as well. 1 Introduction Consider the following two adaptive covering optimization problems: • Adaptive TSP under stochastic demands (AdapTSP). A traveling salesperson is given a metric space (V, d) and distinct subsets S1 , S2 , . . . , Sm ⊆ V such that Si appears with probability pi (and ∑i pi = 1). She needs to serve requests at a random subset S of locations drawn from this distribution. However, she does not know the identity of the random subset: she can only visit locations, at which time she finds out whether or not that location is part of the subset S. What adaptive strategy should she use to minimize the expected time to serve all requests in the random set S? • Optimal Decision Trees. Given a set of m diseases, there are n binary tests that can be used to disambiguate between these diseases. If the cost of performing test t ∈ [n] is ct , and we are given the likelihoods {p j } j∈[m] that a typical patient has the disease j, what (adaptive) strategy should the doctor use for the tests to minimize the expected cost to identify the disease? ∗ Computer Science Department, Carnegie Mellon University. and Operations Engineering Department, University of Michigan. ‡ Tepper School of Business, Carnegie Mellon University. † Industrial 1 It can be shown that the optimal decision tree problem is a special case of the adaptive TSP problem: a formal reduction is given in Section 4. In both these problems we want to devise adaptive strategies, which take into account the revealed information in the queries so far (e.g., locations already visited, or tests already done) to determine the future course of action. Such an adaptive solution corresponds naturally to a decision tree, where nodes encode the current “state” of the solution and branches represent observed random outcomes: see Definition 2 for a formal definition. A simpler class of solutions, that have been useful in some other adaptive optimization problems, eg. [DGV08, GM09, BGL+ 12], are non-adaptive solutions, which are specified by just an ordered list of actions. However there are instances for both the above problems where the optimal adaptive solution costs much less than the optimal non-adaptive solution. Hence it is essential that we find good adaptive solutions. The optimal decision tree problem has long been studied, its NP-hardness was shown by Hyafil and Rivest in 1976 [HR77] and many references and applications can be found in [Now11]. There have been a large number of papers providing algorithms for this problem [GG74, CPR+ 11, Now11, GB09].  Lov85,  KPB99,  Das04, AH12,  1 The best results yield approximation ratios of O log pmin and O log(m ccmax ) , where pmin is the minimum min non-zero probability and cmax (resp. cmin ) is the maximum (resp. minimum) cost. In the special cases when the likelihoods {p j } or the costs {ct } are all polynomially bounded in m, these imply an O(log m)-approximation algorithm. However, there are instances (when probabilities and costs are exponential) demonstrating an Ω(m) approximation guarantee for all previous algorithms. On the hardness side, an Ω(log m) hardness of approximation (assuming P 6= NP) is known for the optimal decision tree problem [CPR+ 11]. While the existence of an O(log m)-approximation algorithm for the general optimal decision tree problem has been posed as an open question, it has not been answered prior to this work. Optimal decision tree is also a basic problem in average-case active learning [Das04, Now11, GB09]. In this application, there is a set of n data points, each of which is associated with a + or − label. The labels are initially unknown. A classifier is a partition of the data points into + and − labels. The true classifier h∗ is the partition corresponding to the actual data labels. The learner knows beforehand, a “hypothesis class” H consisting of m classifiers; it is assumed that the true classifier h∗ ∈ H. Furthermore, in the average case model, there is a known distribution π of h∗ over H. The learner wants to determine h∗ by querying labels at various points. There is a cost ct associated with querying the label of each data point t. An active learning strategy involves adaptively querying labels of data points until h∗ ∈ H is identified. The goal is to compute a strategy that minimizes the expectation (over π) of the cost of all queried points. This is precisely the optimal decision tree problem, with points being tests and classifiers corresponding to diseases. Apart from being a natural adaptive routing problem, AdapTSP has many applications in the setting of message ferrying in ad-hoc networks [ZA03, SRJB03, ZAZ04, ZAZ05, HLS10]. We cite two examples below: • Data collection in sparse sensor networks (see eg. [SRJB03]). A collection of sensors is spread over a large geographic area, and one needs to periodically gather sensor data at a base station. Due to the power and cost overheads of setting up a communication network between the sensors, the data collection is instead performed by a mobile device (the message ferry) that travels in this space from/to the base station. On any given day, there is a known distribution D of the subset S of sensors that contain new information: this might be derived from historical data or domain experts. The routing problem for the ferry then involves computing a tour (originating from the base station) that visits all sensors in S, at the minimum expected cost. • Disaster management (see eg. [ZAZ04]). Consider a post-disaster situation, in which usual communication networks have broken down. In this case, vehicles can be used in order to visit locations and assess the damage. Given a distribution of the set of affected locations, the goal here is to route a vehicle that visits all affected locations as quickly as possible in expectation. 2 In both these applications, due to the absence of a direct communication network, the information at any location is obtained only when it is visited: this is precisely the AdapTSP problem. 1.1 Our Results and Techniques In this paper, we settle the approximability of the optimal decision tree problem: Theorem 1 There is an O(log m)-approximation algorithm for the optimal decision tree problem with arbitrary test costs and arbitrary probabilities, where m is the number of diseases. The problem admits the same approximation ratio even when the tests have non-binary outcomes. In fact, this result arises as a special case of the following theorem: Theorem 2 There is an O(log2 n log m)-approximation algorithm for the adaptive Traveling Salesman Problem, where n is the number of vertices and m the number of scenarios in the demand distribution. To solve the AdapTSP problem, we first solve the “isolation problem”, which seeks to identify which of the m scenarios has materialized. Once we know the scenario we can visit its vertices using any constant-factor approximation algorithm for TSP. The high-level idea behind our algorithm for the isolation problem is this— suppose each vertex lies in at most half the scenarios; then if we visit one vertex in each of the m scenarios using a short tour, which is an instance of the group Steiner tree problem1 , we’d notice at least one of these vertices to have a demand; this would reduce the number of possible scenarios by at least 50% and we can recursively run the algorithm on the remaining scenarios. This is an over-simplified view, and there are many details to handle: we need not visit all scenarios—visiting all but one allows us to infer the last one by exclusion; the expectation in the objective function means we need to solve a minimum-sum version of group Steiner tree; not all vertices need lie in less than half the scenarios. Another major issue is that we do not want our performance to depend on the magnitude of the probabilities, as some of them may be exponentially small. Finally, we need to charge our cost directly against the optimal decision tree. All these issues can indeed be resolved to obtain Theorem 2. The algorithm for the isolation problem involves an interesting combination of ideas from the group Steiner [GKR00, CCGG98] and minimum latency TSP [BCC+ 94, CGRT03, FHR07] problems—it uses a greedy approach that is greedy with respect to two different criteria, namely the probability measure and the number of scenarios. This idea is formalized in our algorithm for the partial latency group Steiner (LPGST) problem, which is a key subroutine for Isolation. While this LPGST problem is harder to approximate than the standard group Steiner 2 tree (see Section  2), for which O(log n log m) is the best approximation ratio, we show that it admits a better 2 O(log n), 4 bicriteria approximation algorithm. Moreover, even this bicriteria approximation guarantee for LPGST suffices to obtain an O(log2 n · log m)-approximation algorithm for Isolation. We also show that both AdapTSP and the isolation problem are Ω(log2−ε n) hard to approximate even on tree metrics; our results are essentially best possible on such metrics, and we lose an extra logarithmic factor to go to general metrics, as in the group Steiner tree problem. Moreover, any improvement to the result in Theorem 2 would lead to a similar improvement for the group Steiner tree problem [GKR00, HK03, CP05] which is a long-standing open question. For the optimal decision tree problem, we show that we can use a variant of minimum-sum set cover [FLT04] which is the special case of LPGST on star-metrics. This avoids an O(log2 n) loss in the approximation guarantee, and hence gives us an O(log m)-approximation algorithm which is best possible [CPR+ 11]. Although this variant of min-sum set cover is Ω(log m)-hard to approximate (it generalizes set cover as shown in Section 2), we again give a constant factor bicriteria approximation algorithm, which leads to the O(log m)-approximation for optimal 1 In the group Steiner tree problem [GKR00] the input is a metric (V, d) with root r ∈ V and groups {X ⊆ V } of vertices; the goal is i to compute a minimum length tour originating from r that visits at least one vertex of each group. 3 decision tree. Our result further reinforces the close connection between the min-sum set cover problem and the optimal decision tree problem that was first noticed by [CPR+ 11]. Finally, we consider the related adaptive traveling repairman problem (AdapTRP), which has the same input as AdapTSP, but the objective is to minimize the expected sum of arrival times at vertices in the materialized demand set. In this setting, we cannot first isolate the scenario and then visit all its nodes, since a long isolation tour may negatively impact the arrival times. So AdapTRP (unlike AdapTSP) cannot be reduced to the isolation problem. However, we show that our techniques for AdapTSP are robust, and can be used to obtain: Theorem 3 There is an O(log2 n log m)-approximation algorithm for the adaptive Traveling Repairman Problem, where n is the number of vertices and m the number of scenarios in the demand distribution. Paper Outline: The results on the isolation problem appear in Section 3. We obtain the improved approximation algorithm for optimal decision tree in Section 4. The algorithm for the adaptive traveling salesman problem is in Section 5; Appendix A contains a nearly matching hardness of approximation result. Finally, Section 6 is on the adaptive traveling repairman problem. 1.2 Other Related Work The optimal decision tree problem has been studied earlier by many authors, with algorithms and hardness results being shown by [GG74, HR77, Lov85, KPB99, AH12, Das04, CPR+ 11, CPRS09, GB09]. As mentioned above, the algorithms in these papers gave O(log m)-approximation ratios only when the probabilities or costs (or both) are polynomially-bounded. The early papers on optimal decision tree considered tests with only binary outcomes. More recently, [CPR+ 11] studied the generalization with K ≥ 2 outcomes per test, and gave an O(log K · log m)-approximation under uniform costs. Subsequently, [CPRS09] improved this bound to O(log m), again under uniform costs. Later, [GB09] gave    an algorithm  under arbitrary costs and probabilities, achieving cmax 1 an approximation ratio of O log pmin or O log(m cmin ) . This is the previous best approximation guarantee; see also Table 1 in [GB09] for a summary of these results. We note that in terms of the number m of diseases, the previous best approximation guarantee is only Ω(m). On the other hand, there is an Ω(log m) hardness of approximation for the optimal decision tree problem [CPR+ 11]. Our O(log m)-approximation algorithm for arbitrary costs and probabilities solves an open problem from these papers. A crucial aspect of this algorithm is that it is non greedy. All previous results were based on variants of a greedy algorithm. There are many results on adaptive optimization dealing with covering problems. E.g., [GV06] considered the adaptive set-cover problem; they gave an O(log n)-approximation when sets may be chosen multiple times, and an O(n)-approximation when each set may be chosen at most once. The latter approximation ratio was improved in [MSW07] to O(log2 n log m), and subsequently to the best-possible O(log n)-approximation ratio by [LPRY08], also using a greedy algorithm. In recent work [GK11] generalized adaptive set-cover to a setting termed ‘adaptive submodularity’, and gave many applications. In all these problems, the adaptivity-gap (ratio between optimal adaptive and non-adaptive solutions) is large, as is the case for the problems considered in this paper, and so the solutions need to be inherently adaptive. The AdapTSP problem is related to universal TSP [JLN+ 05, GHR06] and a priori TSP [Jai88, SS08, ST08] only in spirit—in both the universal and a priori TSP problems, we seek a master tour which is shortcut once the demand set is known, and the goal is to minimize the worst-case or expected length of the shortcut tour. The crucial difference is that the demand subset is revealed in toto in these two problems, leaving no possibility of adaptivity—this is in contrast to the slow revelation of the demand subset that occurs in AdapTSP. 4 2 Preliminaries We work with a finite metric (V, d) that is given by a set V of n vertices and distance function d : V ×V → R+ . As usual, we assume that the distance function is symmetric and satisfies the triangle inequality. For any integer t ≥ 1, we let [t] := {1, 2, . . . ,t}. Definition 1 (r-tour) Given a metric (V, d) and vertex r ∈ V , an r-tour is any sequence (r = u0 , u1 , · · · , uk = r) of vertices that begins and ends at r. The length of such an r-tour is ∑ki=1 d(ui , ui−1 ), the total length of all edges in the tour. Throughout this paper, we deal with demand distributions over vertex-subsets that are specified explicitly. A m demand distribution D is specified by m distinct subsets {Si ⊆ V }m i=1 having associated probabilities {pi }i=1 such m that ∑m i=1 pi = 1. This means that the realized subset D ⊆ V of demand-vertices will always be one of {Si }i=1 , m where D = Si with probability pi (for all i ∈ [m]). We also refer to the subsets {Si }i=1 as scenarios. The following definition captures adaptive strategies. Definition 2 (Decision Tree) A decision tree T in metric (V, d) is a rooted binary tree where each non-leaf node of T is labeled with a vertex u ∈ V , and its two children uyes and uno correspond to the subtrees taken if there is demand at u or if there is no demand at u. Thus given any realized demand D ⊆ V , a unique path TD is followed in T from the root down to a leaf. Depending on the problem under consideration, there are additional constraints on decision tree T and the expected cost of T is also suitably defined. There is a (problem-specific) cost Ci associated with each scenario i ∈ [m] that depends on path TSi , and the expected cost of T (under distribution D) is then ∑m i=1 pi · Ci . For example in AdapTSP, cost Ci corresponds to the length of path TSi . Since we deal with explicitly specified demand distributions D, all decision trees we consider will have size polynomial in m (support size of D) and n (number of vertices). Adaptive Traveling Salesman This problem consists of a metric (V, d) with root r ∈ V and a demand distribution D over subsets of vertices. The information on whether or not there is demand at a vertex v is obtained only when that vertex v is visited. The objective is to find an adaptive strategy that minimizes the expected time to visit all vertices of the realized scenario drawn from D. We assume that the distribution D is specified explicitly with a support-size of m. This allows us to model demand distributions that are arbitrarily correlated across vertices. We note however that the running time and performance of our algorithm will depend on the support size. The most general setting would be to consider blackbox access to the distribution D: however, as shown in [Nag09], in this setting there is no o(n)-approximation algorithm for AdapTSP that uses a polynomial number of samples from the distribution. One could also consider AdapTSP under independent demand distributions. In this case there is a trivial constant-factor approximation algorithm, that visits all vertices having non-zero probability along an approximately minimum TSP tour; note that any feasible solution must visit all vertices with non-zero probability as otherwise (due to the independence assumption) there would be a positive probability of not satisfying a demand. Definition 3 (Adaptive TSP) The input is a metric (V, d), root r ∈ V and demand distribution D given by m m m distinct subsets {Si ⊆ V }m i=1 with probabilities {pi }i=1 (where ∑i=1 pi = 1). The goal in AdapTSP is to compute a decision tree T in metric (V, d) such that: • the root of T is labeled with the root vertex r, and 5 • for each scenario i ∈ [m], the path TSi followed on input Si contains all vertices in Si . The objective function is to minimize the expected tour length ∑m i=1 pi · d(TSi ), where d(TSi ) is the length of the tour that starts at r, visits the vertices on path TSi in that order, and returns to r. Isolation Problem This is closely related to AdapTSP. The input is the same as AdapTSP, but the goal is just to identify the unique scenario that has materialized, and not to visit all the vertices in the realized scenario. Definition 4 (Isolation Problem) Given metric (V, d), root r and demand distribution D, the goal in Isolation is to compute a decision tree T in metric (V, d) such that: • the root of T is labeled with the root vertex r, and • for each scenario i ∈ [m], the path TSi followed on input Si ends at a distinct leaf-node of T . The objective is to minimize the expected tour length IsoTime(T ) := ∑m i=1 pi · d(TSi ), where d(TSi ) is the length of the r-tour that visits the vertices on path TSi in that order, and returns to r. The only difference between Isolation and AdapTSP is that the tree path TSi in Isolation need not contain all vertices of Si , and the paths for different scenarios must end at distinct leaf-nodes. In Section 5 we show that any approximation algorithm for Isolation leads to an approximation algorithm for AdapTSP. So we focus on designing algorithms for Isolation. Optimal Decision Tree This problem involves identifying a random disease from a set of possible diseases using binary tests. Definition 5 (Optimal Decision Tree) The input is a set of m diseases with probabilities {pi }m i=1 that sum to one, and a collection {T j ⊆ [m]}nj=1 of n binary tests with costs {c j }nj=1 . There is exactly one realized disease: each disease i ∈ [m] occurs with probability pi . Each test j ∈ [n] returns a positive outcome for subset T j of diseases and returns a negative outcome for the rest [m] \ T j . The goal in ODT is to compute a decision tree Q where each internal node is labeled by a test and has two children corresponding to positive/negative test outcomes, such that for each i ∈ [m] the path Qi followed under disease i ends at a distinct leaf node of Q. The objective is to minimize the expected cost ∑m i=1 pi · c(Qi ) where c(Qi ) is the sum of test-costs along path Qi . Notice that the optimal decision tree problem is exactly Isolation on a weighted star metric. Indeed, given an instance of ODT, consider a metric (V, d) induced by a weighted star with center r and n leaves corresponding to c the tests. For each j ∈ [n], we set d(r, j) = 2j . The demand scenarios are as follows: for each i ∈ [m] scenario i has demands Si = { j ∈ [n] | i ∈ T j }. It is easy to see that this Isolation instance corresponds exactly to the optimal decision tree instance. See Section 4 for an example. So any algorithm for Isolation on star-metrics can be used to solve ODT as well. Useful Deterministic Problems Recall that the group Steiner tree problem [GKR00, HK03] consists of a metric (V, d), root r ∈ V and g groups of vertices {Xi ⊆ V }gi=1 , and the goal is to find an r-tour of minimum length that contains at least one vertex from each group {Xi }gi=1 . Our algorithms for the above stochastic problems rely on solving some variants of group Steiner tree. Definition 6 (Group Steiner Orienteering) The input is a metric (V, d), root r ∈ V , g groups of vertices {Xi ⊆ V }gi=1 with associated profits {φi }gi=1 and a length bound B. The goal in GSO is to compute an r-tour of length at most B that maximizes the total profit of covered groups. A group i ∈ [g] is covered if any vertex from Xi is visited by the tour. 6 An algorithm for GSO is said to be a (β , γ)-bicriteria approximation algorithm if on any instance of the problem, it finds an r-tour of length at most γ · B that has profit at least β1 times the optimal (which has length at most B). Definition 7 (Partial Latency Group Steiner) The input is a metric (V, d), g groups of vertices {Xi ⊆ V }gi=1 with associated weights {wi }gi=1 , root r ∈ V and a target h ≤ g. The goal in LPGST is to compute an r-tour τ that covers at least h groups and minimizes the weighted sum of arrival times over all groups. The arrival time of group i ∈ [g] is the length of the shortest prefix of tour τ that contains an Xi -vertex; if the group is not covered, its arrival time is set to be the entire tour-length. The LPGST objective is termed latency, i.e. latency(τ) = ∑i covered wi · arrival timeτ (Xi ) + ∑i uncovered wi · length(τ). (2.1) An algorithm for LPGST is said to be a (ρ, σ )-bicriteria approximation algorithm if on any instance of the problem, it finds an r-tour that covers at least h/σ groups and has latency at most ρ times the optimal (which covers at least h groups). The reason we focus on a bicriteria approximation for LPGST is that it is harder to approximate than the group Steiner tree problem (see below) and we can obtain a better bicriteria guarantee for LPGST. To see that LPGST is at least as hard to approximate as the group Steiner tree problem, consider an arbitrary instance of group Steiner tree with metric (V, d), root r ∈ V and g groups {Xi ⊆ V }gi=1 . Construct an instance of LPGST as follows. The vertices are V 0 = V ∪ {u} where u is a new vertex. Let L := n2 · maxa,b d(a, b). The distances in metric (V 0 , d 0 ) are: d 0 (a, b) = d(a, b) if a, b ∈ V and d 0 (a, u) = L + d(a, r) if a ∈ V . There are 0 = {u}. The target h = g. The weights are wi = 0 for i ∈ [g] g0 = g + 1 groups with Xi0 = Xi for i ∈ [g] and Xg+1 and wg+1 = 1. Since the distance from r to u is very large, no approximately optimal LPGST solution will visit u. So any such LPGST solution covers all the groups {Xi0 }gi=1 and has latency equal to the length of the solution (as 0 group Xg+1 has weight one and all others have weight zero). This reduction also shows that LPGST on weighted star-metrics (which is used in the ODT algorithm) is at least as hard to approximate as set cover: this is because when metric (V, d) is a star-metric with center r, so is the new metric (V 0 , d 0 ).2 3 Approximation Algorithm for the Isolation Problem Recall that an instance of Isolation is specified by a metric (V, d), a root vertex r ∈ V , and m scenarios {Si }m i=1 with associated probability values {pi }m . The main result of this section is: i=1 Theorem 4 If there is a (4, γ)-bicriteria approximation algorithm for group Steiner orienteering then there is an O(γ · log m)-approximation algorithm for the isolation problem. We prove this in two steps. First, in Subsection 3.1 we show that a (ρ, 4)-bicriteria approximation algorithm for LPGST can be used to obtain an O(ρ · log m)-approximation algorithm for Isolation. Then, in Subsection 3.2 we show that any (4, γ)-bicriteria approximation algorithm for GSO leads to an (O(γ), 4)-bicriteria approximation algorithm for LPGST. Note on reading this section: While the results of this section apply to the isolation problem on general metrics, readers interested in just the optimal decision tree problem need to only consider weighted star metrics (as discussed after Definition 5). In the ODT case, we have the following simplifications (1) a tour is simply a sequence of tests, (2) the tour length is the sum of test costs in the sequence, and (3) concatenating tours corresponds to concatenating test sequences. 2 Recall that group Steiner tree on star-metrics is equivalent to the set cover problem. 7 3.1 Algorithm for Isolation using LPGST Recall the definition of Isolation and LPGST from Section 2. Here we will prove: Theorem 5 If there is a (ρ, 4)-bicriteria approximation algorithm for LPGST then there is an O(ρ · log m)approximation algorithm for Isolation. We first give a high-level description of our algorithm. The algorithm uses an iterative approach and maintains a candidate set of scenarios that contains the realized scenario. In each iteration, the algorithm eliminates a constant fraction of scenarios from the candidate set. So the number of iterations will be bounded by O(log m). In each iteration we solve a suitable instance of LPGST in order to refine the candidate set of scenarios. Single iteration of Isolation algorithm As mentioned above, we use LPGST in each iteration of the Isolation algorithm- we now describe how this is done. At the start of each iteration, our algorithm maintains a candidate set M ⊆ [m] of scenarios that contains the realized scenario. The probabilities associated with the scenarios i ∈ M are not the original pi s but their conditional probabilities qi := ∑ pi p j . The algorithm Partition (given as j∈M Algorithm 1) uses LPGST to compute an r-tour τ such that after observing the demands on τ, the number of scenarios consistent with these observations is guaranteed to be a constant factor smaller than |M|. To get some intuition for this algorithm, consider the simplistic case when there is a vertex u ∈ V located near the root r such that ≈ 50% of the scenarios in M contain it. Then just visiting vertex u would reduce the number candidate scenarios by ≈ 50%, irrespective of the observation at u, giving us the desired notion of progress. However, each vertex may give a very unbalanced partition of M: so we may have to visit multiple vertices before ensuring that the number of candidate scenarios reduces by a constant factor. Moreover, some vertices may be too expensive to visit from r: so we need to carefully take the metric into account in choosing the set of vertices to visit. Addressing these issues is precisely where the LPGST problem comes in. Algorithm 1 Algorithm Partition( hM, {qi }i∈M i ) 1: 2: 3: 4: 5: let g = |M|. For each v ∈ V , define Fv := {i ∈ M | v ∈ Si }, and Dv :=  Fv M \ Fv if |Fv | ≤ g/2 if |Fv | > g/2 for each i ∈ M, set Xi ← {v ∈ V | i ∈ Dv }. run the (ρ, 4)-bicriteria approximation algorithm for LPGST on the instance with metric (V, d), root r, groups {Xi }i∈M with weights {qi }i∈M , and target h := g − 1. let τ := r, v1 , v2 , · · · , vt−1 , r be the r-tour returned.   let {Pk }tk=1 be the partition of M where Pk := return tour τ and the partition {Pk }tk=1 . Dvk \ ∪ j<k Dvj M \ ∪ j<t Dv j if 1 ≤ k ≤ t − 1 if k = t Note that the information at any vertex v corresponds to a bi-partition (Fv , M \ Fv ) of the scenario set M, with scenarios Fv having demand at v and scenarios M \ Fv having no demand at v. So either the presence of demand or the absence of demand reduces the number of candidate scenarios by half (and represents progress). To better handle this asymmetry, Step 1 associates vertex v with subset Dv which is the smaller of {Fv , M \ Fv }; this corresponds to the set of scenarios under which just the observation at v suffices to reduce the number of candidate scenarios below |M|/2 (and represents progress). In Steps 2 and 3, we view vertex v as covering the scenarios Dv . The overall algorithm for Isolation Here we describe how the different iterations are combined to solve Isolation. The final algorithm IsoAlg (given as Algorithm 2) is described in a recursive manner where each “iteration” is a new call to IsoAlg. As mentioned earlier, at the start of each iteration, the algorithm maintains a 8 r Subtrees corresponding to {Pk }4k=1 are constructed recursively. v1 yes Dv1 corresponds to the yes branch no r Dv2 and Dv3 correspond to the no branch. v2 no yes P1 r v3 yes no r P2 r P4 P3 Figure 3.1: Example of decision tree in single iteration using tour τ = (r, v1 , v2 , v3 , r). candidate set M ⊆ [m] of scenarios such that the realized scenario lies in M. Upon observing demands along the tour produced by algorithm Partition, a new set M 0 ⊆ M containing the realized scenario is identified such that the number of candidate scenarios reduces by a constant factor (specifically |M 0 | ≤ 87 · |M|). Then IsoAlg recurses on scenarios M 0 , which corresponds to the next iteration. After O(log m) such iterations the realized scenario would be correctly identified. Algorithm 2 Algorithm IsoAlghM, {qi }i∈M i 1: If |M| = 1, return this unique scenario as realized. 2: run PartitionhM, {qi }i∈M i let τ = (r, v1 , v2 , · · · , vt−1 , r) be the r-tour and {Pk }tk=1 be the partition of M returned. 3: let q0k := ∑i∈Pk qi for all k = 1 . . .t. 4: traverse tour τ and return directly to r after visiting the first (if any) vertex vk∗ (for k∗ ∈ [t −1]) that determines that the realized scenario is in Pk∗ ⊆ M. If there is no such vertex until the end of the tour τ, then set k∗ ← t. q 5: run IsoAlghPk∗ , { q0 i }i∈Pk∗ i to isolate the realized scenario within the subset Pk∗ . ∗ k Note that the adaptive Algorithm IsoAlg implicitly defines a decision tree too: indeed, we create a path (r, v1 , v2 , · · · , vt−1 , vt = r), and hang the subtrees created in the recursive call on each instance hPk , { qq0i }i from the k respective node vk . See also Figure 3.1. Analysis The rest of this subsection analyzes IsoAlg and proves Theorem 5. We first provide an outline of the proof. It is easy to show that IsoAlg correctly identifies the realized scenario after O(log m) iterations: this is shown formally in Claim 10. We relate the objective values of the LPGST and Isolation instances in two steps: Claim 6 shows that LPGST has a smaller optimal value than Isolation, and Claim 8 shows that any approximate LPGST solution can be used to construct a partial Isolation solution incurring the same cost (in expectation). Since different iterations of IsoAlg deal with different sub-instances of Isolation, we need to relate the optimal 9 cost of these sub-instances to that of the original instance: this is done in Claim 9. Recall that the original instance of Isolation is defined on metric (V, d), root r and set {Si }m i=1 of scenarios . IsoAlg works with many sub-instances of the isolation problem. Such an instance J with probabilities {pi }m i=1 pi is specified by a subset M ⊆ [m] which implicitly defines (conditional) probabilities qi = ∑ p j for all i ∈ M. In j∈M other words, J involves identifying the realized scenario conditioned on it being in set M (the metric and root remain the same as the original instance). Let IsoTime∗ (J ) denote the optimal value of any instance J . Claim 6 For any instance J = hM, {qi }i∈M i, the optimal value of the LPGST instance considered in Step 3 of algorithm Partition(J ) is at most IsoTime∗ (J ). Proof: Let T be an optimal decision tree corresponding to Isolation instance J , and hence IsoTime∗ (J ) = IsoTime(T ). Note that by definition of the sets {Fv }v∈V , any internal node in T labeled vertex v has its two children vyes and vno corresponding to the realized scenario being in Fv and M \Fv (respectively); and by definition of {Dv }v∈V , nodes vyes and vno correspond to the realized scenario being in Dv and M \ Dv (now not necessarily in that order). We now define an r-tour σ based on a specific root-leaf path in T . Consider the root-leaf path that at any node labeled v, moves to the child vyes or vno that corresponds to M \ Dv , until it reaches a leaf-node `. Let r, u1 , u2 , · · · , u j denote the sequence of vertices in this root-leaf path, and define r-tour σ = hr, u1 , u2 , · · · , u j , ri. Since T is a feasible decision tree for the isolation instance, there is at most one scenario a ∈ M such that the path TSa traced in T under demands Sa ends at leaf-node `. In other words, every scenario b ∈ M \ {a} gives rise to a root-leaf path TSb that diverges from the root-` path. By our definition of the root-` path, the scenarios that j j diverge from it are precisely ∪k=1 Duk , and so ∪k=1 Duk = M \ {a}. Next, we show that σ is a feasible solution to the LPGST instance in Step 3. By definition of the groups j {Xi }i∈M (Step 2 of Algorithm 1), it follows that tour σ covers groups ∪k=1 Duk . So the number of groups covered is at least |M| − 1 = h, and σ is a feasible LPGST solution. Finally, we bound the LPGST objective value of σ in terms of the isolation cost IsoTime(T ). To reduce notation let u0 = r below. The arrival times in tour σ are:  k k−1 Dus , for k = 1, · · · , j ∑s=1 d(us−1 , us ) if i ∈ Duk \ ∪s=1 arrival timeσ (Xi ) = length(σ ) if i = a k−1 Fix any k = 1, · · · , j. For any scenario i ∈ Duk \ ∪s=1 Dus , the path TSi traced in T contains the prefix labeled k r, u1 , · · · , uk of the root-` path; so d(TSi ) ≥ ∑s=1 d(us−1 , us ) = arrival timeσ (Xi ). Moreover, for scenario a which j is the only scenario not in ∪k=1 Duk , we have d(TSa ) = length(σ ) = arrival timeσ (Xi ). Now by (2.1), latency(σ ) ≤ q · d(T ) = IsoTime(T ) = IsoTime∗ (J ).  ∑i∈M i Si If we use a (ρ, 4)-bicriteria approximation algorithm for LPGST, we get the following claim: Claim 7 For any instance J = hM, {qi }i∈M i, the latency of tour τ returned by Algorithm Partition is at most ρ · IsoTime∗ (J ). Furthermore, the resulting partition {Pk }tk=1 has each |Pk | ≤ 87 |M| for each k ∈ [t], when |M| ≥ 2. Proof: By Claim 6, the optimal value of the LPGST instance in Step 3 of algorithm Partition is at most IsoTime∗ (J ); now the (ρ, 4)-bicriteria approximation guarantee implies that the latency of the solution tour τ is at most ρ times that. This proves the first part of the claim. Consider τ := hr = v0 , v1 , · · · , vt−1 , vt = ri the tour returned by the LPGST algorithm in Step 3 of algorithm Partition; and {Pk }tk=1 the resulting partition. The (ρ, 4)-bicriteria approximation guarantee implies that the |M|−1 h number of groups covered by τ is | ∪t−1 ≥ |M| k=1 Dvk | ≥ 4 = 4 8 (when |M| ≥ 2). By definition of the sets Dv , it 10 holds that |Dv | ≤ |M|/2 for all v ∈ V . Since all but the last part Pt is a subset of some Dv , it holds that |Pk | ≤ |M| 2 for 1 ≤ k ≤ t − 1. Moreover, the set Pt has size |Pt | = |M \ (∪ j<t Dv j )| ≤ 78 |M|. This proves the second part of the claim.  Of course, we don’t really care about the latency of the tour per se, we care about the expected cost incurred in isolating the realized scenario. But the two are related (by their very construction), as the following claim formalizes: Claim 8 At the end of Step 4 of IsoAlghM, {qi }i∈M i, the realized scenario lies in Pk∗ . The expected distance traversed in this step is at most 2ρ · IsoTime∗ (hM, {qi }i∈M i). Proof: Consider the tour τ := hr = v0 , v1 , · · · , vt−1 , vt = ri returned by the Partition algorithm. Recall that visiting any vertex v reveals whether the scenario lies in Dv , or in M \ Dv . In step 4 of algorithm IsoAlg, we traverse τ and one of the following happens: • 1 ≤ k∗ ≤ t − 1. Tour returns directly to r from the first vertex vk (for 1 ≤ k ≤ t − 1) such that the realized scenario lies in Dvk ; here k = k∗ . Since the scenario did not lie in any earlier Dv j for j < k, the definition of Pk = Dvk \ (∪ j<k Dv j ) gives us that the realized scenario is indeed in Pk . • k∗ = t. Tour τ is completely traversed and we return to r. In this case, the realized scenario does not lie in any of {Dvk | 1 ≤ k ≤ t − 1}, and it is inferred to be in the complement set M \ (∪ j<t Dv j ), which is Pt by definition. Hence for k∗ as defined in Step 4 of IsoAlghM, {qi }i∈M i, it follows that Pk∗ contains the realized scenario; this proves the first part of the claim (and correctness of the algorithm). For each i ∈ M, let αi denote the arrival time of group Xi in tour τ; recall that this is the length of the shortest prefix of τ until it visits an Xi -vertex, and is set to the entire tour length if τ does not cover Xi . The construction of partition {Pk }tk=1 from τ implies that αi = ∑kj=1 d(v j−1 , v j ); ∀i ∈ Pk , ∀1 ≤ k ≤ t, and hence latency(τ) = ∑i∈M qi · αi . To bound the expected distance traversed, note the probability that the traversal returns to r from vertex vk (for 1 ≤ k ≤ t − 1) is exactly ∑i∈Pk qi ; with the remaining ∑i∈Pt qi probability the entire tour τ is traversed. Now, using symmetry and triangle-inequality of the distance function d, we have d(vk , r) ≤ ∑kj=1 d(v j−1 , v j ) for all k ∈ [t]. Hence the expected length traversed is at most: ! ! ! ! t k ∑ ∑ qi k=1 i∈Pk t · d(vk , r) + ∑ d(v j−1 , v j ) k ≤ 2· ∑ j=1 ∑ qi i∈Pk k=1 · ∑ d(v j−1 , v j ) j=1 = 2 · ∑ qi · αi , i∈M which is exactly 2 · latency(τ). Finally, by Claim 10, this is at most 2 · ρ · IsoTime∗ (hM, {qi }i∈M i). Now, the following simple claim captures the “sub-additivity” of IsoTime∗ .  Claim 9 For any instance hM, {qi }i∈M i and any partition {Pk }tk=1 of M, ∑tk=1 q0k · IsoTime∗ (hPk , { q0i }i∈Pk i) q k where q0k = ∑i∈Pk qi for all 1 ≤ k ≤ t. 11 ≤ IsoTime∗ (hM, {qi }i∈M i), (3.2) Proof: Let T denote the optimal decision tree for the instance J0 := hM, {qi }i∈M i. For each k ∈ [t], consider instance Jk := hPk , { qq0i }i∈Pk i; a feasible decision tree for instance Jk is obtained by taking the decision tree T k and considering only paths to the leaf-nodes labeled by {i ∈ Pk }. Note that this is a feasible solution since T isolates all scenarios ∪tk=1 Pk . Moreover, the expected cost of such a decision tree for Jk is ∑i∈Pk qq0i · d(TSi ); recall k that TSi denotes the tour traced by T under scenario i ∈ Pk . Hence Opt(Jk ) ≤ ∑i∈Pk qq0i · d(TSi ). Summing over all k parts k ∈ [t], we get t ∑ q0k · Opt(Jk ) k=1 t ≤ qi ∑ q0k · ∑ q0 · d(TS ) i k=1 i∈Pk k = ∑ qi · d(TS ) i = Opt(J0 ), (3.3) i∈M where the penultimate equality uses the fact that {Pk }tk=1 is a partition of M. Given the above claims, we can bound the overall expected cost of the algorithm.  Claim 10 The expected length of the decision tree given by IsoAlghM, {qi }i∈M i is at most: 2ρ · log8/7 |M| · IsoTime∗ (hM, {qi }i∈M i). Proof: We prove this by induction on |M|. The base case of |M| = 1 is trivial, since zero length is traversed. Now consider |M| ≥ 2. Let instance I0 := hM, {qi }i∈M i. For each k ∈ [t], consider the instance Ik := hPk , { qq0i }i∈Pk i, k where q0k = ∑i∈Pk qi . Note that |Pk | ≤ 87 |M| < |M| for all k ∈ [t] by Claim 10 (as |M| ≥ 2). By the inductive hypothesis, for any k ∈ [t], the expected length of IsoAlg(Ik ) is at most 2ρ · log8/7 |Pk | · IsoTime∗ (Ik ) ≤ 2ρ · (log8/7 |M| − 1) · IsoTime∗ (Ik ), since |Pk | ≤ 87 |M|. By Claim 8, the expected length traversed in Step 4 of IsoAlg(I0 ) is at most 2ρ · IsoTime∗ (I0 ). The probability of recursing on Ik is exactly q0k for each k ∈ [t]. So, expected length of IsoAlg(I0 ) ≤ 2ρ · IsoTime∗ (I0 ) + ∑tk=1 q0k · (expected length of IsoAlg(Ik )) ≤ 2ρ · IsoTime∗ (I0 ) + ∑tk=1 q0k · 2ρ · (log8/7 |M| − 1) · IsoTime∗ (Ik ) ≤ 2ρ · IsoTime∗ (I0 ) + 2ρ · (log8/7 |M| − 1) · IsoTime∗ (I0 ) = 2ρ · log8/7 |M| · IsoTime∗ (I0 ) where the third inequality uses Claim 9.  Claim 10 implies that our algorithm achieves an O(ρ log m)-approximation for Isolation. This completes the proof of Theorem 5. 3.2 Algorithm for LPGST using GSO Recall the definitions of LPGST and GSO from Section 2. Here we will prove: Theorem 11 If there is a (4, γ)-bicriteria approximation algorithm for GSO then there is an (O(γ), 4)-bicriteria approximation algorithm for LPGST. We now describe the algorithm for LPGST in Theorem 11. Consider any instance of LPGST with metric (V, d), root r ∈ V , g groups of vertices {Xi ⊆ V }gi=1 having weights {wi }gi=1 , and target h ≤ g. Let ζ ∗ be an optimal tour for the given instance of LPGST: let Lat∗ denote the latency and D∗ the length of ζ ∗ . We assume (without loss of generality) that the minimum non-zero distance in the metric is one. Let parameter a := 54 . Algorithm 3 is the approximation algorithm for LPGST. The “guess” in the first step means the following. We run the algorithm for all choices of l and return the solution having minimum latency amongst those that cover at least h/4 groups. Since 1 < D∗ ≤ n · maxe de , the number of choices for l is at most log (n · maxe de ), and so the algorithm runs in polynomial time. 12 Algorithm 3 Algorithm for LPGST 1: guess an integer l such that al−1 < D∗ ≤ al . 2: mark all groups as uncovered. 3: for i = 1 . . . l do 4: run the (β , γ)-bicriteria approximation algorithm for GSO on the instance with groups {Xi }gi=1 , root r, length bound ai+1 , and profits:  0 for each covered group i ∈ [g] φi := wi for each uncovered group i ∈ [g] 5: 6: 7: 8: 9: 10: 11: 12: let τ (i) denote the r-tour obtained above. mark all groups visited by τ (i) as covered. end for construct tour τ ← τ (1) ◦ τ (2) ◦ · · · ◦ τ (l) , the concatenation of all the above r-tours. Extend τ if necessary to ensure that d(τ) ≥ γ · al (this is only needed for the analysis). run the (β , γ)-bicriteria approximation algorithm for GSO on the instance with groups {Xi }gi=1 , root r, length bound al , and unit profit for each group, i.e. φi = 1 for all i ∈ [g]. let σ denote the r-tour obtained above. output tour π := τ ◦ σ as solution to the LPGST instance. Analysis In order to prove Theorem 11, we will show that the algorithm’s tour covers at least latency O(γ) · Lat∗ . h 4 groups and has Claim 12 The tour τ in Step 9 has length Θ(γ) · D∗ and latency O(γ) · Lat∗ . Proof: Due to the (β , γ)-bicriteria approximation guarantee of the GSO algorithm used in Step 4, the length γ γa3 ∗ of each r-tour τ (i) is at most γ · ai+1 . So the length of τ in Step 8 is at most γ ∑li=1 ai+1 ≤ a−1 al+2 ≤ a−1 D . ∗ ∗ Moreover, the increase in Step 9 ensures that d(τ) ≥ γ · D . Thus the length of τ in Step 8 is Θ(γ) · D , which proves the first part of the claim. The following proof for bounding the latency is based on techniques from the minimum latency TSP [CGRT03, FHR07]. Recall the optimal solution ζ ∗ to the LPGST instance, where d(ζ ∗ ) = D∗ ∈ (al−1 , al ]. For each i ∈ [l], let Ni∗ denote the total weight of groups visited in ζ ∗ by time ai ; note that Nl∗ equals the total weight of the groups covered by ζ ∗ . Similarly, for each i ∈ [l], let Ni denote the total weight of groups visited in τ (1) · · · τ (i) , i.e. by iteration i of the algorithm. Set N0 = N0∗ := 0, and W := ∑gi=1 wi the total weight of all groups. We have: l latency(τ) ≤ i l ∑ (Ni − Ni−1 ) · ∑ γa j+1 i=1 j=1 l = + (W − Nl ) · d(τ) ≤ ∑ (Ni − Ni−1 ) · i=1 γai+2 ∑ ((W − Ni−1 ) − (W − Ni )) · a − 1 i=1 The last inequality uses the bound d(τ) ≤ The latency of the optimal tour ζ ∗ is Lat∗ ≥ ≥ l−1 ∗ ) ∑ ai−1 (Ni∗ − Ni−1 i=1 l−1 ∑ ai−1 i=1 γ l+2 a−1 a γai+2 + (W − Nl ) · d(τ) a−1 l + (W − Nl ) · d(τ) ≤ ∑ (W − Ni ) · i=0 γai+3 a−1 =: T. from above. + (W − Nl∗ ) · D∗  ∗ (W − Ni−1 ) − (W − Ni∗ ) + (W − Nl∗ ) · al−1 13 ≥ 1 l (1 − ) ∑ ai (W − Ni∗ ). a i=0 Consider any iteration i ∈ [l] of the algorithm in Step 4. Note that the optimal value of the GSO instance solved in this iteration is at least Ni∗ − Ni−1 : the ai length prefix of tour ζ ∗ corresponds to a feasible solution to this GSO instance with profit at least Ni∗ − Ni−1 . The GSO algorithm implies that the profit obtained in τ (i) , i.e. Ni − Ni−1 ≥ 41 · (Ni∗ − Ni−1 ), i.e. W − Ni ≤ 43 · (W − Ni−1 ) + 41 · (W − Ni∗ ). Using this, (a − 1) T γ l = ∑ ai+3 · (W − Ni ) i=0 ≤ ≤ a3 ·W + ≤ 1 l i+3 3 l i+3 ∗ a (W − N ) + ∑ ∑ a (W − Ni−1 ) i 4 i=1 4 i=1 3 l a4 · Lat∗ + ∑ ai+3 (W − Ni−1 ) a−1 4 i=1 3a T a4 · Lat∗ + · (a − 1) a−1 4 γ = 3a l−1 i+3 a4 · Lat∗ + ∑ a (W − Ni ) a−1 4 i=0 4 a This implies T ≤ γ · (a−1)2 (1−3a/4) · Lat∗ = O(γ) · Lat∗ since a = 45 . This completes the proof. Claim 13 The tour σ in Step 10 covers at least h 4  groups and has length O(γ) · D∗ . Proof: Since we know that the optimal tour ζ ∗ has length at most al and covers at least h groups, it is a feasible solution to the GSO instance defined in Step 10. So the GSO algorithm ensures that the tour σ has length at most γal = O(γ)D∗ and profit (i.e. number of groups) at least h/4.  Lemma 14 Tour π = τ · σ covers at least h 4 groups and has latency O(γ) · Lat∗ . Proof: Since π visits all the vertices in σ , Claim 13 implies that π covers at least 4h groups. For each group i ∈ [g], let αi denote its arrival time under the tour τ after Step 9—recall that the arrival time αi for any group i that is not covered by τ is set to the length of the tour d(τ). Claim 12 implies that the latency of tour τ, g ∑i=1 wi · αi = O(γ) · Lat∗ . Observe that for each group i that is covered in τ, its arrival time under tour π = τ · σ remains αi . For any group j not covered in τ, its arrival time under τ is d(τ) ≥ γ · al (due to Step 9), and its arrival time under π is d(π) ≤ O(γ) · D∗ = O(1) · d(τ). Hence, the arrival time under π of each group i ∈ [g] is O(1) · αi , i.e., at most a constant factor more than its arrival time in τ. Now using Claim 12 completes the proof.  Finally, Lemma 14 directly implies Theorem 11. Remark: The above approach also leads to an approximation algorithm for the minimum latency group Steiner problem, which is the special case of LPGST when the target h = g. Definition 8 (Minimum Latency Group Steiner) The input is a metric (V, d), g groups of vertices {Xi ⊆ V }gi=1 with associated non-negative weights {wi }gi=1 and root r ∈ V . The goal in LGST is to compute an r-tour that covers all groups with positive weight and minimizes the weighted sum of arrival times of the groups. The arrival time of group i ∈ [g] is the length of the shortest prefix of the tour that contains a vertex from Xi . Note that the objective here is to minimize the sum of weighted arrival times where every group has to be visited. The algorithm for latency group Steiner is in fact simpler than Algorithm 3: we do not need the “guess” l (Step 1) and we just repeat Step 4 until all groups are covered (instead of stopping after l iterations). A proof identical to that in Claim 12 gives: Corollary 15 If there is a (4, γ)-bicriteria approximation algorithm for GSO then there is an O(γ)-approximation algorithm for the latency group Steiner problem. 14 Combined with the (4, O(log2 n)-bicriteria approximation algorithm for GSO (see Section 5.1) we obtain an O(log2 n)-approximation algorithm for LGST. It is shown in [Nag09] that any α-approximation algorithm for LGST can be used to obtain an O(α · log g)-approximation algorithm for group Steiner tree. Thus improving this O(log2 n)-approximation algorithm for latency group Steiner would also improve the best known bound for the standard group Steiner tree problem. 4 Optimal Decision Tree Problem Recall that the optimal decision tree problem consists of a set of diseases with their probabilities (where exactly one disease occurs) and a set of binary tests with costs, and the goal is to identify the realized disease at minimum expected cost. In this section we prove Theorem 1. As noted in Section 2 the optimal decision tree problem (Definition 5) is a special case of Isolation (Definition 4). We recall the reduction for convenience. Given an instance of ODT, consider a metric (V, d) induced c by a weighted star with center r and n leaves corresponding to the tests. For each j ∈ [n], we set d(r, j) = 2j . The demand scenarios are as follows: for each i ∈ [m] scenario i has demands Si = { j ∈ [n] | i ∈ T j }. It is easy to see that this Isolation instance corresponds exactly to the optimal decision tree instance. Figure 4.2 gives an example. Binary tests {A, B, C, D} and diseases {α, β, γ, δ}. cost : prob. 0.1 α 4 10 2 6 A B C D + + − r 1 − 0.1 β + + + − 0.2 γ + − + + 0.6 δ − − − + 3 2 C D 5 A Scenario β is {A, B, C} w.p. 0.1. B Multiway tests {A, B, C} with l = 3 outcomes. Diseases {α, β, γ, δ}. cost : prob. 0.1 α 4 10 2 A B C 0 1 0 r 1 2 (C, 0) 0.1 β 1 2 0 0.2 γ 1 2 2 0.6 δ 2 2 1 5 (A, 0) (A, 1) (C, 2) (C, 1) (A, 2) Dotted edges have zero length. (B, 0) (B, 1) (B, 2) Scenario β is {(A, 1), (B, 2), (C, 0)}. Figure 4.2: Reducing optimal decision tree to Isolation: binary tests (top), multiway tests (bottom). 15 The main observation here is the following: Theorem 16 There is a (1− 1e )-approximation algorithm for the group Steiner orienteering problem on weighted star metrics. Proof: Consider an instance of GSO (Definition 6) on weighted star-metric (V, d) with center r (which is also the root in GSO) and leaves [n], g groups {Xi ⊆ [n]}gi=1 with profits {φi }gi=1 , and length bound B. If for each j ∈ [n], we define set Y j := {i ∈ [g] | j ∈ Xi } of cost c j := d(r,2 j) , then solving the GSO instance is the same as computing a collection K ⊆ [n] of the sets with ∑ j∈K c j ≤ B/2 that maximizes f (K) := ∑{φi | i ∈ ∪ j∈K Y j }. But the latter problem is precisely an instance of maximizing a monotone submodular function over a knapsack constraint (∑ j∈K c j ≤ B/2), for which a (1 − 1e )-approximation algorithm is known [Svi04].  Combining this result with Theorem 4, we obtain an O(log m) approximation algorithm for Isolation on weighted star-metrics and hence ODT. This proves the first part of Theorem 1. Multiway tests. Our algorithm can be easily extended to the generalization of ODT where tests have multiway (instead of binary) outcomes. In this setting (when each test has at most l outcomes), any test j ∈ [n] induces a partition {T jk }lk=1 of [m] into l parts (some of them may be empty), and performing test j determines which part the realized disease lies in. Note that this problem is also a special case of Isolation. As before, consider a metric (V, d) induced by a weighted star with center r and n leaves corresponding to the tests. For each j ∈ [n], c we set d(r, j) = 2j . Additionally, for each j ∈ [n], introduce l copies of test-vertex j, labeled ( j, 1), · · · , ( j, l), at zero distance from each other. The demand scenarios are defined naturally: for each i ∈ [m], scenario i has demands Si = {( j, k) | i ∈ T jk }. See also an example in Figure 4.2. Clearly this Isolation instance is equivalent to the (multiway) decision tree instance. Since the resulting metric is still a weighted star (we only made vertex copies), Theorem 16 along with Theorem 4 implies an O(log m)-approximation for the multiway decision tree problem. This proves the second part of Theorem 1. 5 Adaptive Traveling Salesman Problem Recall that the adaptive TSP (Definition 3) consists of a metric (V, d) with root r ∈ V and demand distribution D, and the goal is to visit all demand vertices (drawn from D) using an r-tour of minimum expected cost. We first show the following simple fact relating this problem to the isolation problem.  Lemma 17 If there is an α-approximation algorithm for Isolation then there is an α + 32 -approximation algorithm for AdapTSP. Proof: We first claim that any feasible solution T to AdapTSP is also feasible for Isolation. For this it suffices to show that the paths TSi 6= TS j for any two scenarios i, j ∈ [m] with i 6= j. Suppose (for a contradiction) that S paths TSi = TS j = π for some i 6= j. By feasibility of T for AdapTSP, path π contains all vertices in Si S j . Since S Si 6= S j , there is some vertex in (Si \ S j ) (S j \ Si ); let u ∈ Si \ S j (the other case is identical). Consider the point where π is at a node labeled u: then path TSi must take the yes child, whereas path TS j must take the no child. This contradicts the assumption TSi = TS j = π. Thus any solution to AdapTSP is also feasible for Isolation; moreover the expected cost remains the same. Hence the optimal value of Isolation is at most that of AdapTSP. Now, using any α-approximation algorithm for Isolation, we obtain a decision tree T 0 that isolates the realized scenario and has expected cost α · Opt, where Opt denotes the optimal value of the AdapTSP instance. This suggests the following feasible solution for AdapTSP: 1. Implement T 0 to determine the realized scenario k ∈ [m], and return to r. 16 2. Traverse a 32 -approximate TSP tour [Chr77] on vertices {r} Sk . S From the preceding argument, the expected length in the first phase is at most α · Opt. The expected length in the S second phase is at most 32 ∑m i=1 pi · Tsp(Si ), where Tsp(Si ) denotes the minimum length of a TSP tour on {r} Si . Note that ∑m i=1 pi · Tsp(Si ) is a lower bound on the optimal AdapTSP value. So we obtain a solution that has expected cost at most (α + 23 )Opt, as claimed.  Therefore, it suffices to obtain an approximation algorithm for Isolation. In the next subsection we obtain a (4, O(log2 n))-bicriteria approximation algorithm for GSO, which combined with Theorem 4 and Lemma 17 yields an O(log2 n · log m)-approximation algorithm for both Isolation and AdapTSP. This would prove Theorem 2. 5.1 Algorithm for Group Steiner Orienteering Recall the GSO problem (Definition 6). Here we obtain a bicriteria approximation algorithm for GSO. Theorem 18 There is a (4, O(log2 n))-bicriteria approximation algorithm for GSO, where n is the number of vertices in the metric. That is, the algorithm’s tour has length O(log2 n) · B and has profit at least 14 times the optimal profit of a length B tour. This algorithm is based on a greedy framework that is used in many maximum-coverage problems: the solution is constructed iteratively where each iteration adds an r-tour that maximizes the ratio of profit to length. In order to find an r-tour (approximately) maximizing the profit to length ratio, we use a slight modification of an existing algorithm [CCGG98]; see Theorem 19 below. The final GSO algorithm is then given as Algorithm 5. Theorem 19 There is a polynomial time algorithm that given any instance of GSO, outputs an r-tour σ having (σ ) 1 Opt profit-to-length ratio φd(σ ) ≥ α · B . Here φ (σ ) and d(σ ) denote the profit and length (respectively) of tour σ , Opt is the optimal value of the GSO instance, B is the length bound in GSO and α = O(log2 n) where n is the number of vertices in the metric. Proof: This result essentially follows from [CCGG98], but requires some modifications which we present here for completeness. We first preprocess the metric to only include vertices within distance B/2 from the root r: note that since the optimal GSO tour cannot visit any excluded vertex, the optimal profit remains unchanged by this. To reduce notation, we refer  to this restricted vertex-set also as V and let |V | = n. We denote the set of all V edges in the metric by E = 2 . We assume (without loss of generality) that every group is covered by some vertex in V ; otherwise the group can be dropped from the GSO instance. By averaging, there is some vertex u ∈ V covering groups of total profit at least n1 ∑gi=1 φi . If Opt ≤ 4n ∑gi=1 φi then the r-tour that just visits vertex u Opt 1 g has profit-to-length ratio at least Opt 4B and is output as the desired tour σ . Below we assume that n ∑i=1 φi < 4 . We use the following linear programming relaxation LPGSO for GSO: max s.t. g ∑i=1 φi · yi x(δ (S)) ≥ yi ∑e∈E de · xe ≤ B 0 ≤ yi ≤ 1 xe ≥ 0 (5.4) ∀S ⊆ V : r 6∈ S, Xi ⊆ S; ∀i ∈ [g] ∀i ∈ [g] ∀e ∈ E It is easy to see that this a valid relaxation of GSO: any feasible GSO solution corresponds to a feasible solution above where the x, y variables are {0, 1} valued. So the optimal value ∑gi=1 φi · yi ≥ Opt. The algorithm is given 17 as Algorithm 4 and uses the following known results: Theorem 20 shows how to round fractional solutions to LPGSO on tree metrics and Theorem 21 shows how to transform an LPGSO solution on general metrics to one on a tree. Theorem 20 ([CCGG98]) There is a polynomial time algorithm that given any fractional solution (x, y) to LPGSO on a tree metric where all variables are integral multiples of N1 , finds a subtree A containing r such ∑e∈E de ·xe that φd(A) (A) ≤ O(log N) · ∑g φ ·y . Here φ (A) and d(A) denote the profit and length (respectively) of subtree A. i=1 i i  Theorem 21 ([FRT04]) There is a polynomial time algorithm that given any metric (V, d) with edges E = V2 and capacity function x : E → R+ , computes a spanning tree T in this metric such that ∑ f ∈T d f ·xT ( f ) ≤ O(log n)· ∑e∈E de · x(e), where xT ( f ) := x(u, v), ∀ f ∈ T. ∑ u,v : f ∈uv path in T Algorithm 4 Algorithm for GSO maximizing profit-to-length ratio. 1: solve the linear program LPGSO to obtain solution (x, y). 2: run the algorithm from Theorem 21 on metric (V, d) with edge-capacities x to obtain a spanning tree T with “new capacities” xT on edges of T . 3: round down each xT (e) to an integral multiple of n13 . 4: for each group i ∈ [g], let y0i be the maximum flow from r to group Xi under capacities xT . 5: run the algorithm from Theorem 20 using variables xT and y0 to obtain subtree A. 6: output an Euler tour σ of the subtree A. By definition of the new edge-capacities xT on edges of T (see Theorem 21) it is clear that the capacity of each cut under xT is at least as much as under x; i.e. ∑e∈δ (S) xT (e) ≥ ∑e∈δ (S) x(e) for all S ⊆ V . For each group i ∈ [g], since capacities x support yi units of flow from r to Xi , it follows that the new capacities xT on tree T also support such a flow. So (xT , y) is a feasible solution to LPGSO on tree T with budget O(log n) · B. In order to apply the rounding algorithm from [CCGG98] for GSO on trees, we need to ensure the technical condition (see Theorem 20) that every variable is an integral multiple of N1 for some N = poly(n). This is the reason behind modifying capacities xT in Step 3. Note that this step reduces the capacity xT (e) of each edge e ∈ T by at most 1 . Since any cut in tree T has at most n edges, the capacity of any cut decreases by at most n12 after Step 3; and n3 by the max-flow min-cut theorem, the maximum flow value for group Xi is y0i ≥ yi − n12 for each i ∈ [g] (in Step 4). Furthermore, since all edge capacities are integer multiples of n13 , so are all the flow values y0i s. So (xT , y0 ) is a feasible solution to LPGSO on tree T (with budget O(log n) · B) that satisfies the condition required in Theorem 20, with N = n3 . Also note that this rounding down does not change the fractional profits much, since g g ∑ φi · y0i i=1 ≥ 1 g ∑ φi · yi − n2 ∑ φi i=1 i=1 ≥ 3 1 g 3 Opt Opt · Opt − 2 ∑ φi ≥ · Opt − ≥ 4 n i=1 4 4n 2 where the second last inequality follows from n1 ∑gi=1 φi ≤ implies that subtree A satisfies: d(A) φ (A) ≤(T heorem 20) ≤(T heorem 21) Opt 4 (by the preprocessing). Now applying Theorem 20 ∑e∈T de · xT (e) ≤(5.5) g ∑i=1 φi · y0i ∑ de · x(e) ≤ O(log N log n) · e∈E Opt O(log N) · 18 (5.5) ∑e∈T de · xT (e) Opt B O(log2 n) · . Opt O(log N) · Finally, since we output an Euler tour of A, the theorem follows.  Remark: A simpler approach in Theorem 19 might have been to use the randomized algorithm from [GKR00] rather than the deterministic algorithm (Theorem 20) from [CCGG98]. This however does not work directly since [GKR00] only yields a random solution A0 with expected length E[d(A0 )] ≤ O(log n) · ∑e∈E de · xe and expected profit E[φ (A0 )] ≥ ∑gi=1 φi · yi . While this does guarantee the existence of a solution with length-to-profit ratio at de ·xe most O(log n) · ∑∑e∈E , it may not find such a solution with reasonable (inverse polynomial) probability. g φ ·y i=1 i i Algorithm The GSO algorithm first preprocesses the metric to only include vertices within distance B/2 from the root r: note that the optimal profit remains unchanged by this. The algorithm then follows a standard greedy approach (see eg. Garg [Gar96]), and is given as Algorithm 5. Algorithm 5 Algorithm for GSO. 1: initialize r-tour τ ← 0/ and mark all groups as uncovered. 2: while length of τ does not exceed α · B do 3: set residual profits:  0 for each covered group i ∈ [g] e φi := φi for each uncovered group i ∈ [g] 4: 5: 6: run the algorithm from Theorem 19 on the GSO instance with profits φe to obtain r-tour σ . if d(σ ) ≤ αB then τ 0 ← τ ◦ σ . if d(σ ) > αB then: (i) ) partition tour σ into at most 2 · d(σ αB paths, each of length at most αB; (ii) let σ 0 denote the path containing maximum profit; (iii) let hr, σ 0 , ri be the r-tour obtained by connecting both end-vertices of path σ 0 to r. (iv) set τ 0 ← τ ∪ hr, σ 0 , ri. set τ ← τ 0 . Mark all groups visited in τ as covered. 8: end while 9: output the r-tour τ. 7: Analysis Let Opt denote the optimal profit of the given GSO instance. In the following, let α := O(log2 n) which comes from Theorem 19. We prove that Algorithm 5 achieves a (4, 2α + 1) bicriteria approximation guarantee, i.e. solution τ has profit at least Opt/4 and length (2α + 1) · B. By the description of the algorithm, we iterate as long as the total length of edges in τ is at most αB. Note that the increase in length of τ in any iteration is at most (α + 1) · B since every vertex is at distance at most B/2 from r. So the final length d(τ) ≤ (2α + 1) · B. This proves the bound on the length. It now suffices to show that the final subgraph τ gets profit at least Opt 4 . At any iteration, let φ (τ) denote the profit of the current solution τ, and d(τ) its length. Since d(τ) > αB upon termination, it suffices to show the following invariant over the iterations of the algorithm:   Opt Opt φ (τ) ≥ min , · d(τ) (5.6) 4 2αB 19 At the start of the algorithm, inequality (5.6) holds trivially since d(τ) = 0 for τ = 0. / Consider any iteration where φ (τ) < Opt/4 at the beginning: otherwise (5.6) trivially holds for the next iteration. The invariant now ensures that d(τ) < αB/2 and hence we proceed further with the iteration. Moreover, in Step 4 the optimal value g ≥ Opt − φ (τ) ≥ 3 · Opt (by considering the optimal tour for of the “residual” GSO instance with profits φe is Opt 4 g ≤ 2α · B/Opt. the GSO instance with profits φ ). By Theorem 19, the r-tour σ satisfies d(σ )/φe(σ ) ≤ α · B/Opt We finish by handling the two possible cases (Steps 5 and 6). • If d(σ ) ≤ αB, then φ (τ 0 ) = φ (τ) + φe(σ ) ≥ Opt 2αB Opt · d(τ) + 2αB · d(σ ) = Opt 2αB · d(τ 0 ). ) 0 • If d(σ ) > αB, then σ is partitioned into at most 2d(σ αB paths of length αB each. The path σ of best profit Opt Opt αB e 0 e 0 has φe(σ 0 ) ≥ 2·d(σ ) φ (σ ) ≥ 4 ; so φ (τ ) ≥ φ (σ ) ≥ 4 . In either case r-tour τ 0 satisfies inequality (5.6), and since τ ← τ 0 at the end of the iteration, the invariant holds for next iteration as well. This completes the proof of Theorem 18. 6 Adaptive Traveling Repairman In this section we consider the adaptive traveling repairman problem (AdapTRP), where given a demand distribution, the goal is to find an adaptive strategy that minimizes the expected sum of arrival times at demand vertices. As in adaptive TSP, we assume that the demand distribution D is specified explicitly in terms of its support. Definition 9 (Adaptive Traveling Repairman) The input is a metric (V, d), root r and demand distribution D m given by m distinct subsets {Si }m i=1 with probabilities {pi }i=1 (which sum to one). The goal in AdapTRP is to compute a decision tree T in metric (V, d) such that: • the root of T is labeled with the root vertex r, and • for each scenario i ∈ [m], the path TSi followed on input Si contains all vertices in Si . The objective function is to minimize the expected latency ∑m i=1 pi · Lat(TSi ), where Lat(TSi ) is the sum of arrival times at vertices Si along path TSi . We obtain an O(log2 n log m)-approximation algorithm for AdapTRP (Theorem 3). The high-level approach here is similar to that for AdapTSP, but there are some important differences. Unlike AdapTSP, we can not directly reduce AdapTRP to the isolation problem: so there is no analogue of Lemma 17 here. The following example illustrates this. Example 1 Consider an instance of AdapTRP on a star-metric with center r and leaves {v, u1 , · · · , un }. Edges √ (r, ui ) have unit length for each i ∈ [n], and edge (r, v) has length n. There are m = n + 1 scenarios: scenario S0 = {v} occurs with 1 − 1n probability; and for each i ∈ [n], scenario Si = {v, ui } occurs with n12 probability. The optimal Isolation value for this instance is Ω(n) and any reasonable solution clearly will not visit vertex v: it appears in all scenarios and hence provides no information. So if we first follow such an Isolation solution, the arrival time for v is Ω(n); since S0 = {v} occurs with 1 − o(1) probability, the resulting expected latency is Ω(n). √ However, the AdapTRP solution that first visits v, and then vertices {u1 , · · · , un } has expected latency O( n). On the other hand, one can not ignore the “isolation aspect” in AdapTRP either. 20 Example 2 Consider another instance of AdapTRP on a star-metric with center r and leaves {vi }ni=1 ∪ {ui }ni=1 . For each i ∈ [n], edge (r, vi ) has unit length and edge (r, ui ) has length n. There are n scenarios: for each i ∈ [n], scenario Si = {vi , ui } occurs with 1n probability. The optimal values for both AdapTRP and Isolation are Θ(n). Moreover, any reasonable AdapTRP solution will involve first isolating the realized scenario (by visiting vertices vi s). Hence, the algorithm needs to interleave the two goals of isolating scenarios and visiting high-probability vertices. This will become clear in the construction of the latency group Steiner instances used by our algorithm (Step 3 in Algorithm 6). Algorithm Outline Although we can not reduce AdapTRP to Isolation, we are still able to use ideas from the Isolation algorithm. The AdapTRP algorithm also follows an iterative approach and maintains a candidate set M ⊆ [m] containing the realized scenario. We also associate conditional probabilities qi := ∑ pi p j for each j∈M scenario i ∈ M. In each iteration, the algorithm eliminates a constant fraction of scenarios from M: so the number of iterations will be O(log m). Each iteration involves solving an instance of the latency group Steiner (LGST) problem: recall Definition 8 and the O(log2 n)-approximation algorithm for LGST (Corollary 15). The construction of this LGST instance is the main point of difference from the Isolation algorithm. Moreover, we will show that the expected latency incurred in each iteration is O(log2 n) · Opt. Adding up the latency over all iterations, would yield an O(log2 n log m)-approximation algorithm for AdapTRP. Using LGST to partition scenarios M In each iteration, the algorithm formulates an LGST instance and computes an r-tour τ using Corollary 15. The details are in Algorithm 6 below. An important property of this tour τ is that the number of candidate scenarios after observing demands on τ will be at most |M|/2 (see Claim 22). Given a candidate set M of scenarios, it will be convenient to partition the vertices into two parts: H consists of vertices which occur in more than half the scenarios, and L := V \ H consists of vertices occurring in at most half the scenarios. In the LGST instance (Step 3 below), we introduce |Si ∩ H| + 1 groups (with suitable weights) corresponding to each scenario i ∈ M. Algorithm 6 PartnLat( hM, {qi }i∈M , {Si }i∈M i ) 1: define Fv := {i ∈ M | v ∈ Si } for each v ∈ V .  n o Fv |M| 2: let L := u ∈ V : |Fu | ≤ 2 , H := V \ L, and Dv := 3: M \ Fv if v ∈ L if v ∈ H define instance G of LGST (Definition 8) on metric (V, d), root r and the following groups: for each scenario i ∈ M, - the main group Xi of scenario i has weight |Si ∩ L|pi and vertices (L ∩ Si ) ∪ (H \ Si ). - for each v ∈ Si ∩ H, group Yiv has weight pi and vertices {v} ∪ (L ∩ Si ) ∪ (H \ Si ). run the LGST algorithm (from Corollary 15) on instance G . let τ := hr, v1 , v2 , · · · , vt−1 , ri be the r-tour  returned.  if 1 ≤ k ≤ t − 1 Dvk \ ∪ j<k Dvj t 5: let {Pk }k=1 be the partition of M where Pk := 4: 6: M \ ∪ j<t Dv j if k = t return tour τ = hr, v1 , v2 , · · · , vt−1 , ri and partition {Pk }tk=1 . Claim 22 When |M| ≥ 2, partition {Pk }tk=1 returned by PartnLat satisfies |Pk | ≤ |M|/2, ∀k ∈ [t]. 21 Proof: For each k ∈ [t − 1], we have Pk ⊆ Dvk and so |Pk | ≤ |M|/2. We now show that |Pt | ≤ 1 which would prove the claim. Let V (τ) = {v1 , . . . , vt−1 } denote the vertices visited in the tour τ output by PartnLat. Consider any S i ∈ Pt : we will show that it is unique. By definition of Pt , we have i 6∈ t−1 k=1 Dvk . By the definition of group Xi and sets Dv s, this means that Xi is not covered by V (τ). Since τ is a feasible solution to G , Xi ’s weight must be zero, i.e. |Si ∩ L| = 0. Thus we have Si ⊆ H. Furthermore, if vk ∈ H \ Si for any k ∈ [t − 1] then i ∈ Dvk , which implies i 6∈ Pt ; so H ∩V (τ) ⊆ Si . Note that each Yiv = {v} ∪ Xi (for v ∈ H ∩ Si = Si ) must be covered by τ, since Yiv s have weight pi > 0. Also since Xi is not covered by V (τ), we must have v ∈ V (τ) for all v ∈ Si . Thus we have Si ⊆ H ∩V (τ), and combined with the earlier observation, H ∩V (τ) = Si . This determines i ∈ M uniquely, and so |Pt | = 1 ≤ |M|/2.  Final AdapTRP algorithm and analysis Given the above partitioning scheme, Algorithm 7 describes the overall AdapTRP algorithm in a recursive manner. Algorithm 7 AdapTRPhM, {qi }i∈M , {Si }i∈M i 1: If |M| = 1, visit the vertices in this scenario using the O(1)-approximation algorithm [FHR07] for deterministic traveling repairman, and quit. 2: run PartnLathM, {qi }i∈M i let τ = (r, v1 , v2 , · · · , vt−1 , r) be the r-tour and {Pk }tk=1 be the partition of M returned. 3: let q0j := ∑i∈Pk qi for all j = 1 . . .t. 4: traverse tour τ and return directly to r after visiting the first vertex vk∗ (for k∗ ∈ [t]) that determines that the realized scenario is in Pk∗ ⊆ M. 5: update the scenarios in Pk∗ by removing vertices visited in τ until vk∗ , i.e. Si0 ← Si \ {v1 , . . . , vk∗ }, 6: for all i ∈ Pk∗ . run AdapTRPhPk∗ , { qq0 i∗ }i∈Pk∗ , {Si0 }i∈Pk∗ i to recursively cover the realized scenario within Pk∗ . k The analysis for this algorithm is similar to that for the isolation problem (Section 3.1) and we follow the same outline. For any sub-instance J of AdapTRP, let Opt(J ) denote its optimal value. Just as in the isolation case (Claim 9), it can be easily seen that the latency objective function is also sub-additive. Claim 23 For any sub-instance hM, {qi }i∈M , {Si }i∈M i and any partition {Pk }tk=1 of M, q ∑tk=1 q0k · Opt(hPk , { q0i }i∈Pk , {Si }i∈Pk i) k ≤ Opt(hM, {qi }i∈M , {Si }i∈M i), (6.7) where q0k = ∑i∈Pk qi for all 1 ≤ k ≤ t. The next property we show is that the optimal cost of the LGST instance G considered in Steps (3)-(4) of Algorithm 6 is not too high. Lemma 24 For any instance J = hM, {qi }i∈M , {Si }i∈M i of AdapTRP, the optimal value of the latency group Steiner instance G in Step 4 of Algorithm PartnLat(J ) is at most Opt(J ). Proof: Let T be an optimal decision tree for the given AdapTRP instance J . Note that any internal node of T , labeled v, has two children corresponding to the realized scenario being in Fv (yes child) or M \ Fv (no child). Now consider the root-leaf path in T (and corresponding tour σ in the metric) which starts at r, and at any internal node v, moves on to the no child if v ∈ L, and moves to the yes child if v ∈ H. We claim that this tour is a feasible solution to G , the latency group Steiner instance G . 22 To see why, first consider any scenario i ∈ M that branched off from path σ in decision-tree T ; let v be the vertex where the tree path of scenario i branched off from σ . If v ∈ L then by the way we defined σ , it follows the “no” child of v, and so v ∈ Si ∩ L. On the other hand, if v ∈ H, then it must be that v ∈ H \ Si (again from the way σ was defined). In either case, v ∈ (Si ∩ L) ∪ (H \ Si ), and hence visiting v covers all groups, associated with scenario i, i.e. Xi and {Yiv | v ∈ Si ∩ H}. Thus σ covers all groups of all the scenarios that branched off it in T . Note that there is exactly one scenario (say a ∈ M) that does not branch off σ ; scenario a traverses σ in T . Since T is a feasible solution for AdapTRP, σ must visit every vertex in Sa . Therefore σ covers all the groups associated with scenario a: clearly {Yav | v ∈ Sa ∩ H} are covered; Xa is also covered unless Sa ∩ L = 0/ (however in that case group Xa has zero weight and does not need to be covered- see Definition 8). Thus σ is a feasible solution to G . We now bound the latency cost of tour σ for instance G . In path σ , let αi (for each i ∈ M) denote the coverage time for group Xi , and βiv (for i ∈ M and v ∈ Si ∩ H) the coverage time for group Yiv . The next claim shows that the latency of σ for instance G is at most Opt(J ). Claim 25 The expected cost of T , Opt(J ) ≥ ∑i∈M pi · |L ∩ Si | · αi + ∑i∈M ∑v∈Si ∩H pi · βiv , which is exactly the latency of tour σ for the latency group Steiner instance G . Proof: Fix any i ∈ M; let σi denote the shortest prefix of σ containing a vertex from Xi . Note that by definition, σi has length αi . We will lower bound separately the contributions of Si ∩ L and Si ∩ H to the cost of T . As all but the last vertex in σi are from (L \ Si ) ∪ (H ∩ Si ), by definition of σ , the path TSi traced in the decision-tree T when scenario i is realized, agrees with this prefix σi . Moreover, no vertex of Si ∩ L is visited before the end of σi . So under scenario Si , the total arrival time for vertices L ∩ Si is at least |L ∩ Si | · αi . Hence Si ∩ L contributes at least pi · |L ∩ Si | · αi towards Opt(J ). Now consider some vertex v ∈ Si ∩ H; let σiv denote the shortest prefix of σ containing a Yiv -vertex. Note that v σi has length βiv , and it is a prefix of σi since Yiv ⊇ Xi . As observed earlier, the path traced in decision tree T under scenario i contains σi : so vertex v is visited (under scenario i) only after tracing path σiv . So the contribution of v  (under scenario i) to Opt(J ) is at least pi · βiv , i.e. the contribution of Si ∩ H is at least ∑v∈Si ∩H pi · βiv Thus we have demonstrated a feasible solution to G of latency at most Opt(J ).  It remains to bound the expected additional latency incurred in Step 4 of Algorithm 7 when a random scenario is realized. Below we assume a ρ = O(log2 n) approximation algorithm for latency group Steiner tree (from Corollary 15). Lemma 26 At the end of Step 4 of AdapTRPhM, {qi }i∈M , {Si }i∈M i, the realized scenario lies in Pk∗ . The expected increase in latency due to this step is at most 2 ρ · Opt(hM, {qi }i∈M , {Si }i∈M i). Proof: The proof that the realized scenario always lies in the Pk∗ determined in Step 4 is identical to that in Claim 8 of the Isolation algorithm, and is omitted. We now bound the expected latency incurred. In the solution τ to the latency group Steiner instance G , define αi as the coverage time for group Xi , ∀i ∈ M; and βiv as the coverage time for group Yiv , ∀i ∈ M and v ∈ Si ∩ H. Let i denote the realized scenario. Suppose that k∗ = ` ≤ t − 1 in Step 4. Then by definition of the parts T Pk s, we have v` ∈ Xi = (Si ∩ L) ∪ (H \ Si ) and Xi {v1 , . . . , v`−1 } = 0. / So the length along τ until v` equals αi . Moreover the total length spent in this step is at most 2 · αi , to travel till v` and then return to r (this uses the symmetry and triangle-inequality properties of the metric). So the latency of any Si -vertex increases by at most this amount. Furthermore we claim that the latency of any v ∈ Si ∩ H increases by at most 2 · βiv : this is clearly true if βiv = αi ; on the other hand if βiv < αi then v is visited before v` and so it only incurs latency βiv . So the increase in latency of Si is at most 2 ∑v∈Si ∩H βiv + 2 · |Si ∩ L| αi . 23 If k∗ = t then by the proof of Claim 22 the realized scenario i satisfies: Si ⊆ H, group Xi is not visited by τ (so αi is undefined), and all of Si is visited by τ. In this case the total latency of Si is ∑v∈Si ∩H βiv which is clearly at most 2 ∑v∈Si ∩H βiv + 2 · |Si ∩ L| αi ; note that |Si ∩ L| = 0 here.   Thus the expected latency incurred in Step 4 is at most 2 ∑i∈M pi · |Si ∩ L| αi + ∑v∈Si ∩H βiv which is twice the latency of τ for the latency group Steiner instance G . Finally, since τ is a ρ-approximate solution to G and using Lemma 24, we obtain the claim.  Finally, combining Claim 22, Lemma 26 and Claim 23, by a proof identical to that of Theorem 4, it follows that the final AdapTRP solution has cost O(log2 n log m) · Opt. This completes the proof of Theorem 3. We note that for the AdapTRP problem on metrics induced by a tree, our algorithm achieves an O(log n log m) approximation ratio (the guarantees in Theorem 18 and Corollary 15 improve by a logarithmic factor on tree metrics). There is also an Ω(log1−ε n)-hardness of approximation the AdapTRP problem on tree metrics [Nag09]. So there is still a logarithmic gap between the best upper and lower bounds for the AdapTRP problem on tree metrics. In going from tree metrics to general, we lose another logarithmic factor in the approximation ratio. 7 Concluding Remarks In this paper, we studied the problem of constructing optimal decision trees; this widely studied problem was previously known to admit logarithmic approximation algorithms for the case of uniform costs or uniform probabilities. The greedy algorithms used in these cases do not extend to the case of non-uniform costs and probabilities, and we gave a new algorithm that seeks to be greedy with respect to two different criteria; our O(log m)-approximation is asymptotically optimal. We then considered a generalization to the adaptive traveling salesman problem, and obtained an O(log2 n log m)-approximation algorithm for this adaptive TSP problem. We also showed that any asymptotic improvement on this result would imply an improved approximation algorithm for the group Steiner tree problem, which is a long-standing open problem. Finally, we gave an O(log2 n log m)approximation algorithm for the adaptive traveling repairman problem— closing the gap between the known upper and lower bounds in this case remains an interesting open problem. Acknowledgments. A preliminary version appeared in the proceedings of the International Colloquium on Automata, Languages and Programming (ICALP), 2010. We thank Ravishankar Krishnaswamy for many useful conversations; the results on the adaptive traveling repairman problem were obtained in joint discussions, and we thank him for permission to include the results here. We also thank the MOR referees for helpful suggestions that improved the presentation of the paper. A. Gupta’s research was supported in part by NSF awards CCF-0448095 and CCF-0729022, and an Alfred P. Sloan Fellowship. R. Ravi’s research was supported in part by NSF grant CCF-0728841. References [AH12] Micah Adler and Brent Heeringa. Approximating optimal binary decision trees. Algorithmica, 62(34):1112–1121, 2012. [BCC+ 94] A. Blum, P. Chalasani, D. Coppersmith, W. R. Pulleyblank, P. Raghavan, and M. Sudan. The minimum latency problem. In Proceedings of the 26th Annual ACM Symposium on Theory of Computing, pages 163–171, 1994. 24 [BGL+ 12] Nikhil Bansal, Anupam Gupta, Jian Li, Julián Mestre, Viswanath Nagarajan, and Atri Rudra. When LP is the cure for your matching woes: Improved bounds for stochastic matchings. Algorithmica, 63(4):733–762, 2012. [CCGG98] Moses Charikar, Chandra Chekuri, Ashish Goel, and Sudipto Guha. Rounding via trees: deterministic approximation algorithms for group Steiner trees and k median. In Proceedings of the 30th Annual ACM Symposium on Theory of Computing, pages 114–123, 1998. [CGRT03] K. Chaudhuri, B. Godfrey, S. Rao, and K. Talwar. Paths, trees, and minimum latency tours. In Proceedings of the 44th Annual IEEE Symposium on Foundations of Computer Science, pages 36– 45, 2003. [Chr77] N. Christofides. Worst-case analysis of a new heuristic for the travelling salesman problem. GSIA, CMU-Report 388, 1977. [CP05] Chandra Chekuri and Martin Pál. A recursive greedy algorithm for walks in directed graphs. In Proceedings of the 46th Annual IEEE Symposium on Foundations of Computer Science, pages 245– 253, 2005. [CPR+ 11] Venkatesan T. Chakaravarthy, Vinayaka Pandit, Sambuddha Roy, Pranjal Awasthi, and Mukesh K. Mohania. Decision trees for entity identification: Approximation algorithms and hardness results. ACM Transactions on Algorithms, 7(2):15, 2011. [CPRS09] Venkatesan Chakaravarthy, Vinayaka Pandit, Sambuddha Roy, and Yogish Sabharwal. Approximating Decision Trees with Multiway Branches. In ICALP, pages 210–221, 2009. [Das04] Sanjoy Dasgupta. Analysis of a greedy active learning strategy. In Advances in Neural Information Processing Systems (NIPS), 2004. [DGV08] Brian C. Dean, Michel X. Goemans, and Jan Vondrák. Approximating the stochastic knapsack problem: The benefit of adaptivity. Math. Oper. Res., 33(4):945–964, 2008. [FHR07] Jittat Fakcharoenphol, Chris Harrelson, and Satish Rao. The k-traveling repairmen problem. ACM Transactions on Algorithms, 3(4), 2007. [FLT04] Uriel Feige, László Lovász, and Prasad Tetali. Approximating min sum set cover. Algorithmica, 40(4):219–234, 2004. [FRT04] Jittat Fakcharoenphol, Satish Rao, and Kunal Talwar. A tight bound on approximating arbitrary metrics by tree metrics. J. Comput. System Sci., 69(3):485–497, 2004. [Gar96] Naveen Garg. A 3-Approximation for the Minimum Tree Spanning k Vertices. In Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, pages 302–309, 1996. [GB09] Andrew Guillory and Jeff Bilmes. Average-Case Active Learning with Costs. In Algorithmic Learning Theory, pages 141–155. Springer Berlin / Heidelberg, 2009. [GG74] M.R. Garey and R.L. Graham. Performance bounds on the splitting algorithm for binary testing. Acta Informatica, 3:347–355, 1974. [GHR06] Anupam Gupta, Mohammad T. Hajiaghayi, and Harald Räcke. Oblivious network design. In SODA ’06: Proceedings of the seventeenth annual ACM-SIAM symposium on Discrete algorithm, pages 970–979, 2006. 25 [GK11] Daniel Golovin and Andreas Krause. Adaptive submodularity: Theory and applications in active learning and stochastic optimization. J. Artif. Intell. Res. (JAIR), 42:427–486, 2011. [GKR00] N. Garg, G. Konjevod, and R. Ravi. A Polylogarithmic Approximation Algorithm for the Group Steiner Tree Problem. Journal of Algorithms, 37(1):66–84, 2000. [GM09] Sudipto Guha and Kamesh Munagala. Multi-armed bandits with metric switching costs. In ICALP, pages 496–507, 2009. [GV06] Michel Goemans and Jan Vondrák. Stochastic covering and adaptivity. In LATIN 2006: Theoretical informatics, volume 3887 of Lecture Notes in Comput. Sci., pages 532–543. Springer, Berlin, 2006. [HK03] Eran Halperin and Robert Krauthgamer. Polylogarithmic inapproximability. In Proceedings of the 35th Annual Symposium on Theory of Computing, pages 585–594, 2003. [HLS10] Ting He, Kang-Won Lee, and Ananthram Swami. Flying in the dark: controlling autonomous data ferries with partial observations. In MobiHoc, pages 141–150, 2010. [HR77] Laurent Hyafil and Ronald L. Rivest. Constructing optimal binary decision trees is NP-complete. Information Processing Lett., 5(1):15–17, 1976/77. [Jai88] Patrick Jaillet. A priori solution of a travelling salesman problem in which a random subset of the customers are visited. Operations Research, 36 (6), 1988. [JLN+ 05] Lujun Jia, Guolong Lin, Guevara Noubir, Rajmohan Rajaraman, and Ravi Sundaram. Universal approximations for TSP, Steiner tree, and set cover. In STOC ’05: Proceedings of the thirty-seventh annual ACM symposium on Theory of computing, pages 386–395, 2005. [KPB99] S. Rao Kosaraju, Teresa M. Przytycka, and Ryan S. Borgstrom. On an Optimal Split Tree Problem. In Proceedings of the 6th International Workshop on Algorithms and Data Structures, pages 157–168, 1999. [Lov85] Donald W. Loveland. Performance bounds for binary testing with arbitrary weights. Acta Inform., 22(1):101–114, 1985. [LPRY08] Zhen Liu, Srinivasan Parthasarathy, Anand Ranganathan, and Hao Yang. Near-optimal algorithms for shared filter evaluation in data stream systems. In SIGMOD ’08: Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 133–146, 2008. [MSW07] Kamesh Munagala, Utkarsh Srivastava, and Jennifer Widom. Optimization of continuous queries with shared expensive filters. In PODS ’07: Proceedings of the twenty-sixth ACM SIGMODSIGACT-SIGART symposium on Principles of database systems, pages 215–224, 2007. [Nag09] Viswanath Nagarajan. Approximation Algorithms for Sequencing Problems. PhD thesis, Tepper School of Business, Carnegie Mellon University, 2009. [Now11] Robert D. Nowak. The geometry of generalized binary search. IEEE Transactions on Information Theory, 57(12):7893–7906, 2011. [SRJB03] Rahul C. Shah, Sumit Roy, Sushant Jain, and Waylon Brunette. Data mules: modeling and analysis of a three-tier architecture for sparse sensor networks. Ad Hoc Networks, 1(2-3):215–233, 2003. 26 [SS08] F. Schalekamp and D. Shmoys. Algorithms for the universal and a priori TSP. Operations Research Letters, 36(1):1–3, 2008. [ST08] David Shmoys and Kunal Talwar. A Constant Approximation Algorithm for the a priori Traveling Salesman Problem. In Proceedings of the 13th International Conference on Integer Programming and Combinatorial Optimization, pages 331–343, 2008. [Svi04] M. Sviridenko. A note on maximizing a submodular set function subject to knapsack constraint. Operations Research Letters, 32:41–33, 2004. [ZA03] Wenrui Zhao and Mostafa H. Ammar. Message ferrying: Proactive routing in highly-partitioned wireless ad hoc networks. In FTDCS, pages 308–314, 2003. [ZAZ04] Wenrui Zhao, Mostafa H. Ammar, and Ellen W. Zegura. A message ferrying approach for data delivery in sparse mobile ad hoc networks. In MobiHoc, pages 187–198, 2004. [ZAZ05] Wenrui Zhao, Mostafa H. Ammar, and Ellen W. Zegura. Controlling the mobility of multiple data transport ferries in a delay-tolerant network. In INFOCOM, pages 1407–1418, 2005. A Hardness of Approximation for AdapTSP We show that AdapTSP is at least as hard to approximate as group Steiner tree. Theorem 27 If there is an α-approximation algorithm for AdapTSP then there is an (α + o(1))-approximation algorithm for group Steiner tree. Hence AdapTSP is Ω(log2−ε n) hard to approximate even on tree metrics. Proof: This reduction is similar to the reduction [CPR+ 11] from set-cover to the optimal decision tree problem; we give a proof in context of AdapTSP for completeness. Consider an arbitrary instance of group Steiner tree on metric (V, d) with root r and groups X1 , · · · , Xg ⊆ V ; let Opt denote its optimal value. Assume without loss of generality that Xi 6= X j for all i 6= j, and the minimum non-zero distance in d is one. We construct an instance of AdapTSP as follows. Let V 0 = V ∪ {s} where s is a new vertex (representing a copy of r), and define metric d 0 on V 0 as:   0 V d(u, v) for u, v ∈ V 0 d (u, v) := , ∀(u, v) ∈ d(u, r) for u ∈ V, v = s 2 There are g + 1 scenarios in the AdapTSP instance: Si := Xi ∪ {s} for i ∈ [g], and Sg+1 := {s}, with probabilities  1 if 1 ≤ i ≤ g gL pi := , 1 1 − L if i = g + 1 Above L  2n · maxu,v d(u, v) is some large value. The root in the AdapTSP instance remains r. Let Opt0 denote the optimal value time of this instance. We will show that (1 − o(1)) · Opt ≤ Opt0 ≤ Opt + 1 which would prove the theorem.  (A) 1 − L1 Opt ≤ Opt0 . Consider the optimal solution to the AdapTSP instance; let σ denote the r-tour traversed by this decision tree under scenario Sg+1 . We now argue that σ is a feasible solution to the group Steiner tree instance, i.e., Opt ≤ d(σ ). Suppose for a contradiction that σ does not visit any Xi -vertex for some i ∈ [g]. Then observe that the r-tour traversed by this decision tree under scenario Si is also σ , since the decision tree can not distinguish scenarios Si and Sg+1 (the only way to do this is by visiting some Xi -vertex). However this 27 violates the requirement that the tour (namely σ ) under scenario Si must visit all vertices Si ⊇ Xi . Finally, we have Opt0 ≥ (1 − L1 ) · d(σ ) ≥ 1 − L1 Opt as required. (B) Opt0 ≤ Opt + 1. Let τ denote an optimal r-tour for the given GST instance, so d(τ) = Opt. Consider the following solution for AdapTSP: 1. Traverse r-tour τ to determine whether or not Xg+1 is the realized scenario. 2. If no demands observed on τ (i.e. scenario Sg+1 is realized), visit vertex s and stop. 3. If some demand observed on τ (i.e. one of scenarios {Si }gi=1 is realized), then visit all vertices in V along an arbitrary r-tour and stop. It is clear that this decision tree is feasible for the AdapTSP instance. For any i ∈ [g + 1], let πi denote the r-tour traversed under scenario Si in the above AdapTSP decision tree. We have d(πg+1 ) = d(τ) ≤ Opt, and d(πi ) ≤ 2n · maxu,v d(u, v) ≤ L for all i ∈ [g]. Thus the resulting AdapTSP objective is at most:   1 1 1− · Opt + g · · L ≤ Opt + 1 L gL  Thus we have the desired reduction. 28
8cs.DS
Rigidity for von Neumann algebras given by locally compact groups and their crossed products by Arnaud Brothier1 , Tobe Deprez2 and Stefaan Vaes2 To appear in Communications in Mathematical Physics. arXiv:1703.09092v2 [math.OA] 6 Dec 2017 Abstract We prove the first rigidity and classification theorems for crossed product von Neumann algebras given by actions of non-discrete, locally compact groups. We prove that for arbitrary free probability measure preserving actions of connected simple Lie groups of real rank one, the crossed product has a unique Cartan subalgebra up to unitary conjugacy. We then deduce a W∗ strong rigidity theorem for irreducible actions of products of such groups. More generally, our results hold for products of locally compact groups that are nonamenable, weakly amenable and that belong to Ozawa’s class S. 1 Introduction and statement of the main results Popa’s deformation/rigidity theory has lead to a wealth of classification, rigidity and structural theorems for von Neumann algebras, and especially for II1 factors arising from countable groups and their actions on probability spaces, through the group von Neumann algebra and the group measure space construction of Murray and von Neumann. We refer to [Po06, Va10a, Io12, Va16] for an introduction to deformation/rigidity theory. The main goal of this article is to prove rigidity and classification theorems for crossed products by actions of non-discrete, locally compact groups. The classification problem for II1 factors M given as crossed products M = L∞ (X) ⋊ Γ for free ergodic probability measure preserving (pmp) actions of countable groups splits into two separate problems: the uniqueness problem for the Cartan subalgebra L∞ (X) and the classification problem for Γ y (X, µ) up to orbit equivalence. Striking progress has been made on both problems. In [OP07], it is proved that for profinite free ergodic pmp actions of the free groups Fn , the crossed product M has a unique Cartan subalgebra up to unitary conjugacy. In [CS11], it was shown that the same holds for profinite actions of nonelementary hyperbolic groups and actually for profinite actions of nonamenable, weakly amenable groups in Ozawa’s class S introduced in [Oz03, Oz04]. For arbitrary free ergodic pmp actions of the same groups, the uniqueness of the Cartan subalgebra was established in [PV11, PV12]. The first goal of this paper is to prove that also for locally compact groups that are nonamenable, weakly amenable and in class S, crossed products M = L∞ (X) ⋊ G by arbitrary free ergodic pmp actions have a unique Cartan subalgebra up to unitary conjugacy. This class of groups includes all rank one simple Lie groups, as well as all locally compact groups that admit a continuous and metrically proper action on a tree, or on a hyperbolic graph (see Proposition 7.1). The precise definition of property (S) goes as follows. 1 University Roma Tor Vergata, Department of Mathematics, Roma (Italy). E-mail: [email protected]. AB is supported by European Research Council Advanced Grant 669240 QUEST. 2 KU Leuven, Department of Mathematics, Leuven (Belgium). E-mails: [email protected] and [email protected]. TD is supported by a PhD fellowship of the Research Foundation Flanders (FWO). SV is supported by European Research Council Consolidator Grant 614195 RIGIDITY, and by long term structural funding – Methusalem grant of the Flemish Government. 1 Definition. Let G be a locally compact group and denote S(G) = {F ∈ L1 (G) | F (g) ≥ 0 for a.e. g ∈ G and kF k1 = 1}. Equip S(G) with the topology induced by the L1 -norm. We say that G has property (S) if there exists a continuous map η : G → S(G) satisfying lim kη(gkh) − g · η(k)k1 = 0 uniformly on compact sets of g, h ∈ G. k→∞ (1.1) By [BO08, Proposition 15.2.3], Ozawa’s class S (see [Oz04]) consists of all countable groups Γ that are exact and that have property (S). Our uniqueness of Cartan theorem can then be stated as follows. In Section 3, we actually prove a more general result, also valid for nonsingular actions (see Theorem 3.1) and thus generalizing the results in [HV12] to the locally compact setting. Theorem A. Let G = G1 × · · · × Gn be a direct product of nonamenable locally compact second countable (lcsc) weakly amenable groups with property (S). Let G y (X, µ) be an essentially free pmp action. Then L∞ (X) ⋊ G has a unique Cartan subalgebra up to unitary conjugacy. To understand Theorem A, note that if G is non-discrete, then L∞ (X) is not a Cartan subalgebra of M , but there is a canonical Cartan subalgebra given by choosing a cross section for G y (X, µ) (see Section 3). We then turn to orbit equivalence rigidity. In Section 4, we prove a cocycle superrigidity theorem for arbitrary cocycles of irreducible pmp actions G1 × G2 y (X, µ) taking values in a locally compact group with property (S). This result is similar to the cocycle superrigidity theorem of [MS04], where the target group is assumed to be a closed subgroup of the isometry group of a negatively curved space. We then deduce that Sako’s orbit equivalence rigidity theorem [Sa09] for irreducible pmp actions G1 × G2 y (X, µ) of nonamenable groups in class S stays valid in the locally compact setting. Recall here that a nonsingular action G1 × G2 y (X, µ) of a direct product group is called irreducible if both G1 and G2 act ergodically. In combination with Theorem A, we deduce the following W∗ strong rigidity theorem. This is the first W∗ strong rigidity theorem for actions of locally compact groups. Theorem B. Let G = G1 × G2 and H = H1 × H2 be unimodular lcsc groups without nontrivial compact normal subgroups. Let G y (X, µ) and H y (Y, η) be essentially free, irreducible pmp actions. Assume that G1 , G2 , H1 , H2 are nonamenable and that H1 , H2 are weakly amenable and have property (S). If p(L∞ (X) ⋊ G)p ∼ = q(L∞ (Y ) ⋊ H)q for nonzero projections p and q, then the actions are conjugate: there exists a continuous group isomorphism δ : G → H and a pmp isomorphism ∆ : X → Y such that ∆(g · x) = δ(g) · ∆(x) for all g ∈ G and a.e. x ∈ X. Fix Haar measures on G and H and denote by Tr the associated normal semifinite trace on the crossed products L∞ (X) ⋊ G and L∞ (Y ) ⋊ H. If the Haar measures are normalized such that δ is measure preserving, then Tr(p) = Tr(q). Also, the isomorphism p(L∞ (X) ⋊ G)p ∼ = q(L∞ (Y ) ⋊ H)q has the explicit form given in Remark 4.3. We deduce Theorem A from a very general structural result on the normalizer NM (A) = {u ∈ U (M ) | uAu∗ = A} of a von Neumann subalgebra A ⊂ M when M is equipped with an arbitrary coaction Φ : M → M ⊗ L(G) of a locally compact weakly amenable group with property (S), see Theorem F below. The main novelty is to show that the main ideas of [PV11] can be made to work in this very general and much more abstract setting, by using several results from the harmonic analysis of coactions. 2 We prove uniqueness of Cartan subalgebras by applying this general result to the canonical coaction Φ : L(R) → L(R) ⊗ L(G) associated with a countable pmp equivalence relation R and a cocycle ω : R → G with values in the locally compact group G. Applying the same general result to the comultiplication ∆ : L(G) → L(G) ⊗ L(G) itself, we obtain the following strong solidity results for locally compact group von Neumann algebras. Recall that a diffuse von Neumann algebra M is called strongly solid if for every diffuse amenable von Neumann subalgebra A ⊂ M that is the range of a normal conditional expectation, the normalizer NM (A)′′ remains amenable. When also the amplification B(ℓ2 (N)) ⊗ M is strongly solid, we say that M is stably strongly solid, see [BHV15]. Theorem C. Let G be a locally compact group with property (S) and assume that L(G) is diffuse. 1. If G is unimodular and weakly amenable, then for every finite trace projection p ∈ L(G), we have that pL(G)p is strongly solid. 2. If G is second countable, if G has the complete metric approximation property (CMAP) and if the kernel of the modular function G0 = {g ∈ G | δ(g) = 1} is an open subgroup of G, then L(G) is stably strongly solid. Note that the von Neumann algebras L(G) appearing in the second part of Theorem C can be of type III. The assumption on G0 being open in the second part of Theorem C is not essential, but it makes the proof much less technical. In all our examples of locally compact groups G with property (S) and with L(G) being nonamenable, the assumption is satisfied. Examples D. Every finite center connected simple Lie group G of real rank one is weakly amenable and has property (S). Every locally compact group G that acts metrically properly on a tree (not necessarily locally finite) has CMAP and property (S). Every locally compact hyperbolic group is weakly amenable and has property (S). References and proofs for these statements are discussed in Section 7. For locally compact groups G acting properly on a tree, [HR16, Theorems C and D] and [Ra15, Theorems E and F] provide criteria ensuring that L(G) is a nonamenable factor. Applying Theorem C, we thus obtain the first examples of nonamenable strongly solid locally compact group von Neumann algebras. In particular, when n, m ∈ Z with 2 ≤ |m| < n and G denotes the Schlichting completion of the Baumslag-Solitar group BS(m, n), then L(G) is strongly solid, nonamenable and of type III|m/n| by combining Theorem C and [Ra15, Theorem G]. Combining Theorem A with [PV08, Proposition 7.1], we also obtain the following first examples of II1 factors having a unique Cartan subalgebra up to unitary conjugacy, but not having a group measure space Cartan subalgebra, in the sense that the countable equivalence relation generated by the unique Cartan subalgebra cannot be written as the orbit equivalence relation of an essentially free group action. Corollary E. Let G = Sp(n, 1) with n ≥ 2 and let G y (X, µ) be any weakly mixing Gaussian action. Put M = L∞ (X) ⋊ G. Then, M is a II∞ factor that has a unique Cartan subalgebra up to unitary conjugacy, but that has no group measure space Cartan subalgebra. In particular, its finite corners pM p are II1 factors with unique Cartan subalgebra, but without group measure space Cartan subalgebra. As explained above, Theorems A and C follow from a general result on normalizers inside tracial von Neumann algebras M that are equipped with a so-called coaction of a locally compact group. Recall that a coaction of a locally compact group G on a von Neumann algebra M is a 3 faithful normal ∗-homomorphism Φ : M → M ⊗ L(G) satisfying (Φ ⊗ id)Φ = (id ⊗ ∆)Φ, where ∆ : L(G) → L(G) ⊗ L(G) is the comultiplication given by ∆(λg ) = λg ⊗ λg for all g ∈ G. Assume that Φ : M → M ⊗ L(G) is a coaction, Tr is a faithful normal semifinite trace on M and p ∈ M is a projection with Tr(p) < ∞. Let A ⊂ pM p be a von Neumann subalgebra. We say that • A can be Φ-embedded if the pM p-bimodule Φ(p)(L2 (M p) ⊗ L2 (G)) given by x · ξ · y = Φ(x)ξ(y ⊗ 1) admits a nonzero A-central vector; • A is Φ-amenable if there exists a nonzero positive functional Ω on Φ(p)(M ⊗ B(L2 (G)))Φ(p) that is Φ(A)-central and satisfies Ω(Φ(x)) = Tr(x) for all x ∈ pM p. Note that the Φ-amenability of A ⊂ pM p is equivalent with the left A-amenability of the pM p-M -bimodule Φ(pM p)Φ(p)(L2 (M ) ⊗ L2 (G))M in the sense of [PV11, Definition 2.3] and this amenability notion for bimodules is a generalization of relative amenability for pairs of von Neumann subalgebras introduced in [OP07, Section 2.2]. The following dichotomy type theorem is a locally compact version of [PV12, Theorem 3.1]. Theorem F. Let G be a locally compact group that is weakly amenable and has property (S). Let (M, Tr) be a von Neumann algebra with a faithful normal semifinite trace and Φ : M → M ⊗L(G) a coaction. Let p ∈ M be a projection with Tr(p) < ∞ and A ⊂ pM p a von Neumann subalgebra. If A is Φ-amenable then at least one of the following statements holds: A can be Φ-embedded or NpM p (A)′′ stays Φ-amenable. Finally, in order to obtain stable strong solidity, we have to replace the normalizer NM (A) by s (A) = {x ∈ M | xAx∗ ⊂ A and x∗ Ax ⊂ A}. Adapting the methods the stable normalizer NM of [BHV15] to the abstract setting of Theorem F, we obtain the following result. Theorem G. If in Theorem F, we add the hypothesis that G has the complete metric approximation property, then in the conclusion, we may replace the normalizer NpM p(A)′′ by the stable s ′′ normalizer NpM p (A) . 2 Proof of Theorem F The proof of Theorem F follows closely the proofs of [PV11, Theorem 5.1] and [PV12, Theorem 3.1]. The main novelty is to develop, in the context of coactions of locally compact groups, a framework in which the main ideas of [PV11, PV12] are applicable. To do this, we need several results from the harmonic analysis of coactions and their crossed products, which were proven for arbitrary locally compact quantum groups in [Va00, BSV02, BS92]. Fix a weakly amenable, locally compact group G with property (S). Denote by Λ(G) the Cowling-Haagerup constant of G, see [CH88]. Also fix a von Neumann algebra M with a faithful normal semifinite trace Tr and a coaction Φ : M → M ⊗ L(G). Let p ∈ M be a projection with Tr(p) < ∞ and A ⊂ pM p a von Neumann subalgebra that is Φ-amenable. Denote by ∆ : L(G) → L(G) ⊗ L(G) the comultiplication, given by ∆(λg ) = λg ⊗ λg . Weak amenability. Denote by A(G) the Fourier algebra of G, defined as the predual of L(G) and identified with a subalgebra of the algebra Cb (G) of bounded continuous functions on G, by identifying ω ∈ L(G)∗ with the function g 7→ ω(λg ). We denote by Ac (G) ⊂ A(G) the subalgebra of compactly supported functions in A(G). By weak amenability of G, using [CH88, Proposition 1.1] and a convexity argument, we can fix a net ηn ∈ Ac (G) such that the 4 associated normal completely bounded maps mn : L(G) → L(G) : mn (x) = (id ⊗ ηn )∆(x) satisfy kmn kcb ≤ Λ(G) < ∞ (id ⊗ mn )(X) → X for all n, and strongly, for all Hilbert spaces H and X ∈ B(H) ⊗ L(G). (2.1) Define the normal completely bounded maps ϕn : M → M : ϕn (x) = (id ⊗ ηn )Φ(x). Using that Φ is a coaction, we get that Φ ◦ ϕn = (id ⊗ mn ) ◦ Φ. Since Φ is faithful, Φ is completely isometric and thus, kϕn kcb ≤ kmn kcb . Since Φ is a homeomorphism for the strong topology on norm bounded subsets, we get that ϕn (x) → x strongly for every x ∈ M . Notations and terminology. Denote K = L2 (M p) ⊗ L2 (G) and view Φ as a normal ∗homomorphism Φ : M → B(K). Also define the normal ∗-antihomomorphism ρ : A → B(K) given by ρ(a)ξ = ξ(a ⊗ 1). Define N = Φ(M ) ∨ ρ(A) as the von Neumann subalgebra of B(K) generated by Φ(M ) and ρ(A). Note that N ⊂ B(L2 (M p)) ⊗ L(G). We also denote by ρ : A → B(L2 (M p)) the ∗-antihomomorphism given by right multiplication. Whenever V is a set of operators on a Hilbert space, we denote by [V] the operator norm closed linear span of V. Denote by N0 ⊂ N the dense C∗ -subalgebra defined as N0 := [Φ(M )ρ(A)]. Write q = Φ(p). We say that a normal completely bounded map ψ : pM p → pM p is adapted if the following two conditions hold. 1. There exists a normal completely bounded map θ : qN q → B(L2 (pM p)) with θ(Φ(x)ρ(a)) = ψ(x)ρ(a) for all x ∈ pM p and a ∈ A. 2. There exist a Hilbert space L, a unital ∗-homomorphism π0 : qN0 q → B(L) and maps V, W : NpM p (A) → L such that Tr(w∗ ψ(x)va) = hπ0 (Φ(x)ρ(a))V(v), W(w)i for all x ∈ pM p, a ∈ A, v, w ∈ NpM p(A) ,  and, defining kVk∞ = sup kV(v)k v ∈ NpM p(A) , we have kVk∞ kWk∞ < ∞ . We denote by kψkadap the infimum of all possible values of Tr(p)−1 kVk∞ kWk∞ . Step 1. Let ω ∈ A(G) and define m : L(G) → L(G) by m = (id ⊗ ω) ◦ ∆. Also define ϕ : M → M by ϕ = (id ⊗ ω) ◦ Φ and, as before, note that (id ⊗ m) ◦ Φ = Φ ◦ ϕ. Put ψ : pM p → pM p : ψ(x) = pϕ(x)p. We claim that ψ is adapted and that kψkadap ≤ kmkcb . To prove step 1, we first prove the following statement: the pM p-A-bimodule weakly contained in the pM p-A-bimodule Φ(pM p)q(L2 (M p) ⊗ L2 (G))A ⊗ 1. pM pL 2 (pM p) A is Using the leg numbering notation for multiple tensor products, we view K′ := q12 (L2 (M ) ⊗ L2 (G) ⊗ L2 (G))q13 as the standard Hilbert space for q(M ⊗ B(L2 (G)))q. The left representation of q(M ⊗ B(L2 (G)))q on K′ is given by left multiplication in tensor positions 1 and 2, while the right representation is given by right multiplication in tensor positions 1 and 3. The Φ-amenability of A then provides a net of vectors ξi ∈ K′ satisfying limhΦ(x)12 ξi , ξi i = Tr(x) and lim kΦ(a)12 ξi − ξi Φ(a)13 k = 0 i i for all x ∈ pM p and a ∈ A. This implies that pM pL2 (pM p)A is weakly contained in the pM p-Abimodule Φ(pM p)12 K′Φ(A)13 . Since the pM p-bimodule Φ(pM p)12 K′Φ(pM p)13 is unitarily conjugate to a multiple of the pM p-bimodule Φ(pM p)q(L2 (M p) ⊗ L2 (G))pM p ⊗ 1, the above weak containment 5 statement is proven. So, we get a unital ∗-homomorphism θ ′ : qN0 q → B(L2 (pM p)) satisfying θ ′ (Φ(x)ρ(a)) = xρ(a) for all x ∈ pM p, a ∈ A. Define the normal completely bounded map θ : qN q → B(L2 (pM p)) by θ(x) = p(id ⊗ ω)(x)p. By construction, θ(Φ(x)ρ(a)) = ψ(x)ρ(a) for all x ∈ pM p and a ∈ A. Since θ(x) = θ ′ (q(id ⊗ m)(x)q) for all x ∈ qN0 q, we get that kθkcb ≤ kmkcb . So, the Stinespring like factorization theorem (see e.g. [BO08, Theorem B.7]) provides a Hilbert space L, a unital ∗-homomorphism π0 : qN0 q → B(L) and bounded operators V0 , W0 : L2 (pM p) → L satisfying θ(x) = W0∗ π0 (x)V0 for all x ∈ qN0 q and kV0 k kW0 k = kθkcb ≤ kmkcb . It now suffices to define V and W by restricting V0 and W0 to NpM p(A) ⊂ L2 (pM p). So we have proved that ψ : pM p → pM p is adapted and that kψkadap ≤ kmkcb . This concludes the proof of step 1. Notations and terminology. We start with a net ηn ∈ Ac (G) such that the associated normal completely bounded maps mn : L(G) → L(G) given by mn = (id ⊗ ηn ) ◦ ∆ satisfy (2.1). Defining ψn : pM p → pM p : ψn (x) = p(id ⊗ ηn )Φ(x)p , we obtain a net of adapted completely bounded maps ψn : pM p → pM p such that ψn (x) → x strongly for all x ∈ pM p and lim supn kψn kadap ≤ Λ(G). We call such a net an adapted approximate identity. We then define κ ≥ 1 as the smallest positive number for which there exists an adapted approximate identity ψn : pM p → pM p with lim supn kψn kadap ≤ κ. We fix such a ψn realizing κ. Since each ψn is adapted, we have normal completely bounded maps θn : qN q → B(L2 (pM p)) satisfying θn (Φ(x)ρ(a)) = ψn (x)ρ(a) for all x ∈ pM p and a ∈ A. We can thus define µn ∈ (qN q)∗ given by µn (T ) = hθn (T )p, pi and satisfying µn (Φ(x)ρ(a)) = Tr(ψn (x)a) for all x ∈ pM p, a ∈ A. For every v ∈ NpM p (A), denote by βv the automorphism of N implemented by right multiplication with v ∗ ⊗ 1 on L2 (M p) ⊗ L2 (G). Note that βv (Φ(x)ρ(a)) = Φ(x)ρ(vav ∗ ) for all x ∈ M and a ∈ A. In particular, βv (q) = q and we also view βv as an automorphism of qN q. Step 2. The functionals µn satisfy the following properties. 1. lim supn kµn k < ∞, 2. limn µn (Φ(x)ρ(a)) = Tr(xa) for all x ∈ pM p, a ∈ A, 3. limn kµn ◦ (βv ◦ Ad Φ(v)) − µn k = 0 for all v ∈ NpM p (A), 4. limn k(Φ(a)ρ(a∗ )) · µn − µn k = 0 for all a ∈ U (A). To prove step 2, one can literally repeat the argument in [Oz10, Proof of Proposition 7] and [PV11, Proof of Proposition 5.4], because for every v ∈ NpM p(A) and every adapted approximate identity ψn : pM p → pM p, the maps x 7→ ψn (xv ∗ )v and x 7→ v ∗ ψn (vx) form again adapted approximate identities. Step 3. There exist positive normal functionals ωn ∈ (qN q)∗ satisfying 1. limn ωn (Φ(x)) = Tr(x) for all x ∈ pM p, 2. limn kωn ◦ (βv ◦ Ad Φ(v)) − ωn k = 0 for all v ∈ NpM p (A), 3. limn ωn (Φ(a)ρ(a∗ )) = Tr(p) for all a ∈ U (A). To prove step 3, choose a weak∗ limit point Ξ ∈ (pN p)∗ of the net µn . We find that Ξ(Φ(x)) = Tr(x) for all x ∈ pM p, that Ξ is invariant under the automorphisms βv ◦ Ad Φ(v) for all v ∈ NpM p(A) and that (Φ(a)ρ(a∗ )) · Ξ = Ξ for all a ∈ U (A). Define Ω1 = |Ξ|. So Ω1 is a 6 positive element of (qN q)∗ satisfying Ω1 ◦(βv ◦Ad Φ(v)) = Ω1 and (Φ(a)ρ(a∗ ))·Ω1 = Ω1 for all v ∈ NpM p (A), a ∈ U (A) . (2.2) Furthermore, we have that | Tr(x)|2 = |Ξ(Φ(x))|2 ≤ kΩ1 k Ω1 (Φ(x∗ x)) for all x ∈ pM p . In order to conclude the proof of step 3, we need to modify Ω1 so that its restriction to Φ(pM p) is given by the trace. We first modify Ω1 so that this restriction is normal and faithful. The bidual of the embedding Φ : pM p → qN q is an embedding Φ∗∗ : (pM p)∗∗ → (qN q)∗∗ . Denote by z ∈ Z((pM p)∗∗ ) the support projection of the natural normal ∗-homomorphism (pM p)∗∗ → pM p given by dualizing the embedding (pM p)∗ ֒→ (pM p)∗ . By construction, for every Ω ∈ (pM p)∗ , the functional Ω(·z) belongs to (pM p)∗ . Write z1 = Φ∗∗ (z). Whenever α ∈ Aut(qN q) satisfies α(Φ(pM p)) = Φ(pM p), the bidual automorphism α∗∗ ∈ Aut((qN q)∗∗ ) satisfies α∗∗ (z1 ) = z1 . In particular, z1 commutes with every unitary in qN q that normalizes Φ(pM p). Since Φ(pM p) ⊂ qN q is regular, it follows that z1 belongs to the center of (qN q)∗∗ . Applying the statement above to the automorphism α = βv ◦Ad Φ(v) and the unitary Φ(a)ρ(a∗ ), it follows that the positive functional Ω2 (·) = Ω1 ( · z1 ) still satisfies the properties in (2.2). By density, we have | Tr(x)|2 ≤ kΩ1 k Ω1 (Φ∗∗ (x∗ x)) for all x ∈ (pM p)∗∗ . Since Tr(x) = Tr(xz) for all x ∈ pM p, we conclude that | Tr(x)|2 ≤ kΩ1 k Ω2 (Φ(x∗ x)) for all x ∈ pM p. In particular, Ω2 ◦ Φ is faithful. Since Ω2 (Φ(x)) = Ω2 (Φ∗∗ (xz)) for all x ∈ pM p, we get that Ω2 ◦ Φ is normal. So, we find a nonsingular T ∈ L1 (pM p)+ such that Ω2 (Φ(x)) = Tr(xT ) for all x ∈ pM p. Since (2.2) holds, we have that T commutes with NpM p(A). For every n ≥ 3, we then define the positive functional Ωn on qN q given by Ωn (·) = Ω2 (Φ((T + 1/n)−1/2 ) · Φ((T + 1/n)−1/2 )) . Each Ωn satisfies the properties in (2.2). Choosing Ω to be a weak∗ -limit point of the sequence Ωn , we have found a positive functional Ω on qN q that satisfies the properties in (2.2) and that moreover satisfies Ω(Φ(x)) = Tr(x) for all x ∈ pM p. Approximating Ω in the weak∗ topology and taking convex combinations, we find a net of positive ωn ∈ (qN q)∗ satisfying the conditions in step 3. Notations and terminology. Choose a standard Hilbert space H for the von Neumann algebra N , which comes with the normal ∗-homomorphism πl : N → B(H), the normal ∗antihomomorphism πr : N → B(H) and the positive cone H+ ⊂ H. For every v ∈ NpM p (A), denote by Wv ∈ U (H) the canonical implementation of βv ∈ Aut(N ). Step 4. There exist vectors ξn ∈ H+ satisfying πl (q)ξn = ξn = πr (q)ξn for all n and 1. limn hπl (Φ(x))ξn , ξn i = Tr(pxp) = limn hπr (Φ(x))ξn , ξn i for all x ∈ M , 2. limn kπl (Φ(v))πr (Φ(v ∗ ))Wv ξn − ξn k = 0 for all v ∈ NpM p(A), 3. limn kπl (Φ(a))ξn − πl (ρ(a))ξn k = 0 for all a ∈ U (A). Note that πl (q)πr (q)H serves as the standard Hilbert space of qN q. Define ξn ∈ πl (q))πr (q)H+ as the canonical implementation of the normal positive functional ωn ∈ (qN q)∗ . The properties of ωn in step 3 translate into the above properties for ξn by the Powers-Størmer inequality. Notations and terminology. Define the coaction Ψ : N → N ⊗ L(G) given by Ψ = id ⊗ ∆. By [Va00, Definition 3.6 and Theorem 4.4], the coaction Ψ has a canonical implementation on H, given by a nondegenerate ∗-homomorphism π : C0 (G) → B(H) satisfying the following natural covariance properties w.r.t. πl , πr and Ψ. 7 Denote by Cλ∗ (G) ⊂ L(G) and Cρ∗ (G) ⊂ R(G) the canonical dense C∗ -subalgebras. We denote by ⊗min the spatial C∗ -tensor product and define V ∈ M (C0 (G) ⊗min Cλ∗ (G)) and W ∈ M (C0 (G) ⊗min Cρ∗ (G)) given by the functions V (g) = λg and W (g) = ρg . Define the unitary operators X, Y ∈ B(H ⊗ L2 (G)) given by X = (π ⊗ id)(V ) and Y = (π ⊗ id)(W ). Denoting by χ : L(G) → R(G) : χ(λg ) = ρ∗g the canonical anti-isomorphism, the covariance properties are then given by (πl ⊗ id)Ψ(x) = X(πl (x) ⊗ 1)X ∗ and (πr ⊗ χ)Ψ(x) = Y (πr (x) ⊗ 1)Y ∗ for all x ∈ N . Formulation of the dichotomy. We are in precisely one of the following cases. • Case 1. For every F ∈ C0 (G), we have that lim supn kπ(F )ξn k = 0. • Case 2. There exists an F ∈ C0 (G) with lim supn kπ(F )ξn k > 0. We prove that in case 1, the von Neumann subalgebra NpM p (A)′′ ⊂ pM p is Φ-amenable and that in case 2, the von Neumann subalgebra A ⊂ pM p can be Φ-embedded. Case 1 – Notations and terminology. Since G has property (S), we have a continuous map Z0 : G → L2 (G) satisfying kZ0 (g)k = 1 for all g ∈ G and lim kZ0 (gkh) − λg (Z0 (k))k = 0 uniformly on compact sets of g, h ∈ G. k→∞ (2.3) For each F ∈ C0 (G), we view Z0 F ∈ C0 (G) ⊗min L2 (G) and in this way, Z0 is an adjointable operator from the C∗ -algebra C0 (G) to the Hilbert C∗ -module C0 (G) ⊗min L2 (G). Define Ve ∈ M (Cλ∗ (G) ⊗min C0 (G)) given by the function g 7→ λg . So, Ve is just the flip of the unitary V defined above. As operators on L2 (G) ⊗ L2 (G), we have ∆(a) = Ve (1 ⊗ a)Ve ∗ for all a ∈ L(G). By [BSV02, Section 5], the closed linear span M0 = [(id ⊗ ω)Φ(x) | x ∈ M, ω ∈ L(G)∗ ] (2.4) is a unital C∗ -subalgebra of M . Also, Φ(M0 ) ⊂ M (M0 ⊗min Cλ∗ (G)) and the restriction of Φ to M0 defines a continuous coaction. In particular, the closed linear span Sl = [Φ(M0 )(1 ⊗ C0 (G))] ⊂ M ⊗ B(L2 (G)) (2.5) is a C∗ -algebra (i.e. the crossed product of M0 and the coaction Φ of Cλ∗ (G), as first defined in [BS92, Définition 7.1]) and M0 = [(id ⊗ ω)Φ(x) | x ∈ M0 , ω ∈ L(G)∗ ] . (2.6) Case 1 – Step 1. We claim that (1 ⊗ Z0 )Φ(x) − (Φ ⊗ id)Φ(x)(1 ⊗ Z0 ) ∈ Sl ⊗min L2 (G) (2.7) for all x ∈ M0 . Note that (2.3), with h = e, can be rephrased as follows: (1 ⊗ λ∗g )Z0 − (λg ⊗ 1)Z0 λ∗g belongs to C0 (G) ⊗min L2 (G), uniformly on compact sets of g ∈ G. This means that for every F ∈ C0 (G), we have ∗ Ve23 (Z0 ⊗ F ) − Ve13 (Z0 ⊗ F )Ve ∗ ∈ [C0 (G) ⊗ L2 (G) ⊗ C0 (G)] and thus, because Ve normalizes C0 (G × G), ∗ e∗ Ve13 V23 (Z0 ⊗ F ) − (Z0 ⊗ F )Ve ∗ ∈ [(C0 (G) ⊗ L2 (G) ⊗ C0 (G))Ve ∗ ] . 8 (2.8) Using that Ve23 and Ve13 commute, we similarly find that Ve23 Ve13 (Z0 ⊗ F ) − (Z0 ⊗ F )Ve ∈ [(C0 (G) ⊗ L2 (G) ⊗ C0 (G))Ve ] . (2.9) By (2.6), it suffices to prove (2.7) for x = (1 ⊗ η ∗ )Φ(y)(1 ⊗ µ) where y ∈ M0 and where η, µ ∈ Cc (G) are viewed as vectors in the Hilbert space L2 (G). Fix F ∈ C0 (G) such that η ∗ F = η ∗ and F µ = µ. Using that Φ is a coaction and that ∆(a) = Ve (1 ⊗ a)Ve ∗ for all a ∈ L(G), we find that ∗ e∗ (Φ ⊗ id)Φ(x) = (1 ⊗ 1 ⊗ 1 ⊗ η ∗ ) Ve34 Ve24 Φ(y)14 Ve24 V34 (1 ⊗ 1 ⊗ 1 ⊗ µ) . Twice using that µ = F µ, it then follows from (2.8) that (Φ ⊗ id)Φ(x) (1 ⊗ Z0 ) ∗ = (1 ⊗ 1 ⊗ 1 ⊗ η ∗ ) Ve34 Ve24 Φ(y)14 (1 ⊗ Z0 ⊗ F ) Ve23 (1 ⊗ 1 ⊗ µ) + T ∗ e e ∗ e = (1 ⊗ 1 ⊗ 1 ⊗ η ) V34 V24 Φ(y)14 (1 ⊗ Z0 ⊗ 1) V23 (1 ⊗ 1 ⊗ µ) + T ∗ = (1 ⊗ 1 ⊗ 1 ⊗ η ∗ ) Ve34 Ve24 (1 ⊗ Z0 ⊗ 1) Φ(y)13 Ve23 (1 ⊗ 1 ⊗ µ) + T (2.10) where the error term T belongs to ∗ [(1 ⊗ 1 ⊗ 1 ⊗ L2 (G)∗ ) Ve34 Ve24 Φ(M0 )14 (1 ⊗ C0 (G) ⊗ L2 (G) ⊗ C0 (G)) Ve23 (1 ⊗ 1 ⊗ L2 (G))] . Using that [Φ(M0 )(1 ⊗ C0 (G))] = Sl = [(1 ⊗ C0 (G))Φ(M0 )], that Ve34 and Ve24 commute, that [Ve (L2 (G) ⊗ C0 (G))] = [L2 (G) ⊗ C0 (G)] and that Ve normalizes C0 (G) ⊗min C0 (G) = C0 (G × G), we get that T belongs to ∗ [(1 ⊗ 1 ⊗ 1 ⊗ L2 (G)∗ ) Ve34 Ve24 (1 ⊗ C0 (G) ⊗ L2 (G) ⊗ C0 (G)) Φ(M0 )13 Ve23 (1 ⊗ 1 ⊗ L2 (G))] = [(1 ⊗ 1 ⊗ 1 ⊗ L2 (G)∗ ) Ve24 (1 ⊗ C0 (G) ⊗ L2 (G) ⊗ C0 (G)) Φ(M0 )13 Ve ∗ (1 ⊗ 1 ⊗ L2 (G))] 2 2 ∗ = [(1 ⊗ C0 (G) ⊗ L (G)) (1 ⊗ 1 ⊗ L (G) ∗ ) Ve23 Φ(M0 )13 Ve23 (1 23 2 ⊗ 1 ⊗ L (G))] = [(1 ⊗ C0 (G) ⊗ L2 (G)) (1 ⊗ 1 ⊗ L2 (G)∗ ) (Φ ⊗ id)Φ(M0 ) (1 ⊗ 1 ⊗ L2 (G))] = [(1 ⊗ C0 (G))Φ(M0 ) ⊗ L2 (G)] = Sl ⊗min L2 (G) . Using that η ∗ = η ∗ F and using (2.9), we can continue the computation in (2.10) and find that (Φ ⊗ id)Φ(x) (1 ⊗ Z0 ) ∗ = (1 ⊗ 1 ⊗ 1 ⊗ η ∗ ) Ve34 Ve24 (1 ⊗ Z0 ⊗ F ) Φ(y)13 Ve23 (1 ⊗ 1 ⊗ µ) + T ∗ = (1 ⊗ 1 ⊗ 1 ⊗ η ∗ ) (1 ⊗ Z0 ⊗ F ) Ve23 Φ(y)13 Ve23 (1 ⊗ 1 ⊗ µ) + T ′ + T = (1 ⊗ Z0 ) (1 ⊗ 1 ⊗ η ∗ ) (Φ ⊗ id)Φ(y) (1 ⊗ 1 ⊗ µ) + T ′ + T = (1 ⊗ Z0 ) Φ(x) + T ′ + T where the error term T ′ belongs to ∗ [(1 ⊗ 1 ⊗ 1 ⊗ L2 (G)∗ ) (1 ⊗ C0 (G) ⊗ L2 (G) ⊗ C0 (G)) Ve23 Φ(M0 )13 Ve23 (1 ⊗ 1 ⊗ L2 (G))] = [(1 ⊗ C0 (G) ⊗ L2 (G)) (1 ⊗ 1 ⊗ L2 (G)∗ ) (Φ ⊗ id)Φ(M0 ) (1 ⊗ 1 ⊗ L2 (G))] = Sl ⊗min L2 (G) . So (2.7) and step 1 are proven. Case 1 – Step 2. Define the ∗-homomorphism ζl : M → B(H) : ζl = πl ◦ Φ and the ∗-antihomomorphism ζr : M → B(H) : ζr = πr ◦ Φ. Define S = [ζl (M0 ) ζr (M0 ) π(C0 (G)) Wv | v ∈ NpM p(A)] . 9 (2.11) Finally, define the isometry Z ∈ B(H, H ⊗ L2 (G)) given by Z = (π ⊗ id)(Z0 ). We claim that S ⊂ B(H) is a C∗ -algebra and that Zζl (x) − (ζl ⊗ id)Φ(x)Z and Zζr (x) − (ζr (x) ⊗ 1)Z belong to S ⊗min L2 (G) (2.12) for all x ∈ M0 . Since ζl : M0 → B(H) and π : C0 (G) → B(H) are covariant w.r.t. the continuous coaction Φ : M0 → M (M0 ⊗min Cr∗ (G)), they induce a nondegenerate representation of the full crossed product. Since G is co-amenable, the canonical homomorphism of the full crossed product onto the reduced crossed product is an isomorphism. The reduced crossed product is given by the C∗ -algebra Sl defined in (2.5). So, we find a nondegenerate ∗-homomorphism θl : Sl → B(H) : θl (Φ(x)(1 ⊗ F )) = ζl (x)π(F ) , for all x ∈ M0 , F ∈ C0 (G). Associated with the coaction Φ : M → M ⊗ L(G), we have the canonical coaction Φop : M op → M op ⊗ R(G) defined as follows. Denote by γ : M → M op : γ(x) = xop the canonical ∗-antiisomorphism. As before, define the ∗-anti-isomorphism χ : L(G) → R(G) = η(λg ) = ρ∗g . Then, Φop ◦ γ = (γ ⊗ χ) ◦ Φ. The corresponding crossed product C∗ -algebra is Sr = [Φop (M0op )(1 ⊗ C0 (G))] ⊂ M op ⊗ B(L2 (G)) . Since also ζr and π are covariant, we similarly find a nondegenerate ∗-homomorphism θr : Sr → B(H) : θr (Φop (xop )(1 ⊗ F )) = ζr (x)π(F ) , for all x ∈ M0 , F ∈ C0 (G). So θl (Sl ) = [ζl (M0 )π(C0 (G))] and θr (Sr ) = [ζr (M0 )π(C0 (G))] and these are C∗ -algebras. Moreover, the unitaries Wv , v ∈ NpM p(A), commute with ζl (M ), ζr (M ) and π(C0 (G)). So, the space S defined in (2.11) is a C∗ -algebra and S = [θl (Sl ) θr (Sr ) Wv | v ∈ NpM p(A)] . Also, θl (Sl ) ⊂ S and θr (Sr ) ⊂ S. Applying to (2.7) the canonical extension of θl ⊗ id to the multiplier algebra, we find the first half of (2.12). In the same way as we proved (2.7), one proves that (1 ⊗ Z0 )Φop (xop ) − (Φop (xop ) ⊗ 1)(1 ⊗ Z0 ) ∈ Sr ⊗min L2 (G) (2.13) for all x ∈ M0 . Applying θr ⊗ id to (2.13), also the second half of (2.12) follows and step 2 is proven. Case 1 – Notations. Write G = NpM p (A) and consider the ∗-algebras CG and D = M ⊗alg M op ⊗alg CG. Define the ∗-homomorphisms Θ : D → B(H) : Θ(x ⊗ y op ⊗ v) = ζl (x) ζr (y) Wv , Θ1 : D → B(H ⊗ L2 (G)) : Θ1 (x ⊗ y op ⊗ v) = (ζl ⊗ id)Φ(x) (ζr (y) Wv ⊗ 1) . Choose a positive functional Ω on B(H) as a weak∗ limit point of the net of vector functionals T 7→ hT ξn , ξn i. The properties of the net ξn established in step 4 above then imply that: Ω(1) = Ω(Θ(p ⊗ pop ⊗ p)) = Tr(p) , Ω(Θ(x ⊗ 1 ⊗ p)) = Tr(pxp) for all x ∈ M , Ω(Θ(v ⊗ (v ∗ )op ⊗ v)) = Tr(p) for all v ∈ G. 10 (2.14) Case 1 – Step 3. Writing C = kΩk Λ(G)2 , we claim that |Ω(Θ(x))| ≤ C kΘ1 (x)k for all x ∈ D . (2.15) Since Wv commutes with π(C0 (G)) for all v ∈ G, we have ZWv = (Wv ⊗ 1)Z for all v ∈ G. Denoting D0 = M0 ⊗alg M0op ⊗alg CG, (2.12) implies that Z ∗ Θ1 (x)Z − Θ(x) ∈ S for all x ∈ D0 . (2.16) Since we are in case 1, we have that Ω(π(F )) = 0 for all F ∈ C0 (G). So, Ω(T ) = 0 for all T ∈ S. It then follows from (2.16) that |Ω(Θ(x))| = |Ω(Z ∗ Θ1 (x)Z)| ≤ kΩk kΘ1 (x)k for all x ∈ D0 . (2.17) To conclude step 3, we now have to approximate as follows an arbitrary x ∈ D by elements in D0 . Take a net ηn ∈ A(G) such that the net mn = (id ⊗ ηn ) ◦ ∆ satisfies (2.1). Define ϕn : M → M by ϕn = (id ⊗ ηn ) ◦ Φ. Note that the image of Θ1 lies in B(H) ⊗ L(G) and that (id ⊗ mn ) ◦ Θ1 = Θ1 ◦ (ϕn ⊗ id ⊗ id). It follows that kΘ1 ((ϕn ⊗ id ⊗ id)(x))k ≤ Λ(G) kΘ1 (x)k for all x ∈ D and all n . Denoting by χ1 : L(G) → L(G) the period 2 anti-automorphism given by χ1 (λg ) = λg−1 , the representation Θ1 is unitarily conjugate to the representation Θ2 : D → B(H ⊗ L2 (G)) : Θ2 (x ⊗ y op ⊗ v) = (ζl (x) ⊗ 1) (ζr ⊗ χ1 )Φ(y) (Wv ⊗ 1) . op op So, writing ϕop m (y ) = (ϕm (y)) , we also find that kΘ1 ((id ⊗ ϕop m ⊗ id)(x))k ≤ Λ(G) kΘ1 (x)k for all x ∈ D and all m . Altogether, we have proved that 2 kΘ1 ((ϕn ⊗ ϕop m ⊗ id)(x))k ≤ Λ(G) kΘ1 (x)k p For every T ∈ B(H), write kT kΩ = Ω(T ∗ T ). Since for all x ∈ D and all n, m . kζl (ϕn (x)) − ζl (x)k2Ω = Tr(p(ϕn (x) − x)∗ (ϕn (x) − x)p) for every x ∈ M , it follows from the Cauchy-Schwarz inequality that for every x ∈ D and every m, op Ω(Θ((id ⊗ ϕop m ⊗ id)(x))) = lim Ω(Θ((ϕn ⊗ ϕm ⊗ id)(x))) . n Similarly, we have Ω(Θ(x)) = lim Ω(Θ((id ⊗ ϕop m ⊗ id)(x))) m for all x ∈ D. Since (ϕn ⊗ ϕop m ⊗ id)(x) ∈ D0 for all n, m, it follows from (2.17) that op |Ω((ϕn ⊗ ϕop m ⊗ id)(x))| ≤ kΩk kΘ1 ((ϕn ⊗ ϕm ⊗ id)(x))k ≤ C kΘ1 (x)k for all n, m. Taking first the limit over n and then over m, we find that (2.15) holds and step 3 is proven. Case 1 – End of the proof. Because of (2.15), we can define a continuous functional Ω1 on the C∗ -algebra [Θ1 (D)] satisfying Ω1 (Θ1 (x)) = Ω(Θ(x)) for all x ∈ D. Since Ω1 (Θ1 (x)∗ Θ1 (x)) = Ω(Θ(x∗ x)) ≥ 0 11 for all x ∈ D, it follows by density that Ω1 is positive. Extend Ω1 to a functional on B(H ⊗ L2 (G)) without increasing its norm. So Ω1 remains positive. Write q1 = Θ1 (p ⊗ pop ⊗ p) = (ζl ⊗ id)Φ(p) (ζr (p) ⊗ 1) . For every v ∈ G, define Uv := Θ1 (v ⊗ (v ∗ )op ⊗ v) = (ζl ⊗ id)(Φ(v))(ζr (v ∗ )Wv ⊗ 1) and note that Uv is a unitary in B(q1 (H ⊗ L2 (G))). By (2.14), these unitaries Uv satisfy Ω1 (Uv ) = Ω1 (Θ1 (v ⊗ (v ∗ )op ⊗ v)) = Ω(Θ(v ⊗ (v ∗ )op ⊗ v)) = Tr(p) = Ω(Θ(p ⊗ pop ⊗ p)) = Ω1 (Θ1 (p ⊗ pop ⊗ p)) = Ω1 (q1 ) . By (2.14), we also have that Ω1 (1 − q1 ) = Ω(1 − Θ(p ⊗ pop ⊗ p)) = 0 and Ω1 ((ζl ⊗ id)(Φ(x))) = Ω(Θ(x ⊗ 1 ⊗ p)) = Tr(pxp) for all x ∈ M . Altogether, we have in particular that Ω1 is Uv -central for every v ∈ G. Define the positive functional Ω2 on q(M ⊗ B(L2 (G)))q given by Ω2 (T ) = Ω1 ((ζl ⊗ id)(T )). Then, Ω2 (Φ(x)) = Tr(x) for all x ∈ pM p. Since for every v ∈ G, the functional Ω1 is Uv -central, while ζr (v ∗ )Wv ⊗ 1 commutes with (ζl ⊗ id)(q(M ⊗ B(L2 (G)))q), we get that Ω2 is Φ(G)-central. Writing P = NpM p (A)′′ , the Cauchy-Schwarz inequality implies that Ω2 is Φ(P )-central. So we have proved that P is Φ-amenable. Proof in Case 2. After passing to a subnet, we may assume that there is an F ∈ C0 (G) such that the net kπ(F )ξn k is convergent to a strictly positive number. Choose a positive functional Ω on B(H) as a weak∗ limit point of the net of vector functionals T 7→ hT ξn , ξn i. Define the C∗ -algebra S1 := θl (Sl ) = [ζl (M0 )π(C0 (G))]. Denote by Ω1 the restriction of Ω to S1′′ . By the properties of the net ξn established in step 4 above, Ω1 (ζl (x)) = Tr(pxp) for all x ∈ M and Ω1 is ζl (A)-central. Also, the restriction of Ω1 to S1 is nonzero. Define δ = kΩ1 |S1 k and put ε = δ(4Λ(G)3 + 2Λ(G)2 + 2)−1 . Since the elements π(F ), with F ∈ Cc (G) and 0 ≤ F ≤ 1, form an approximate identity for S1 , we can fix F ∈ Cc (G) with 0 ≤ F ≤ 1 and Ω1 (π(F )) ≥ δ − ε and |Ω1 (T ) − Ω1 (T π(F ))| < ε kT k for all T ∈ S1 . As above, take a net of completely bounded maps ϕn : M → M such that kϕn kcb ≤ Λ(G) and ϕn (M ) ⊂ M0 for all n and ϕn (x) → x strongly for all x ∈ M . Because Ω1 (ζl (x)) = Tr(pxp) for all x ∈ M , Ω1 (ζl (x)T ζl (y)) = lim Ω1 (ζl (ϕn (x))T ζl (ϕn (y))) (2.18) n for all x, y ∈ M and T ∈ S1′′ . Using the ζl (A)-centrality of Ω1 , we then find, for all a ∈ U (A), δ ≤ Ω1 (π(F )) + ε = Ω1 (ζl (a∗ )π(F )ζl (a)) + ε = lim Re Ω1 (ζl (ϕn (a∗ ))π(F )ζl (ϕn (a))) + ε . n Since ζl (ϕm (a∗ ))π(F )ζl (ϕn (a)) belongs to S1 and has norm at most Λ(G)2 , we get that δ ≤ lim sup Re Ω1 (ζl (ϕn (a∗ ))π(F )ζl (ϕn (a))π(F )) + ε(Λ(G)2 + 1) . n 12 (2.19) We claim that there exists an ω0 ∈ A(G) such that the corresponding completely bounded map ϕ0 : M → M : ϕ0 = (id ⊗ ω0 ) ◦ Φ satisfies kϕ0 kcb ≤ 2Λ(G) and π(F )ζl (x)π(F ) = π(F )ζl (ϕ0 (x))π(F ) for all x ∈ M0 . (2.20) Using θl : Sl → S1 , it suffices to construct ω0 ∈ A(G) such that kϕ0 kcb ≤ 2Λ(G) and (1 ⊗ F )Φ(x)(1 ⊗ F ) = (1 ⊗ F )Φ(ϕ0 (x))(1 ⊗ F ) for all x ∈ M0 . (2.21) Denote by K ⊂ G the (compact) support of F . By [CH88, Proposition 1.1], we can choose ω0 ∈ A(G) such that ω0 (g) = 1 for all g ∈ KK −1 and such that the map m0 = (id ⊗ ω0 ) ◦ ∆ satisfies km0 kcb ≤ 2Λ(G). As operators on L2 (G), we have that F λg F = 0 for all g ∈ G\KK −1 . It follows that F xF = F m0 (x)F for all x ∈ L(G). Writing ϕ0 = (id ⊗ ω0 ) ◦ Φ, we then also have (1 ⊗ F )Φ(x)(1 ⊗ F ) = (1 ⊗ F ) (id ⊗ m0 )(Φ(x)) (1 ⊗ F ) = (1 ⊗ F )Φ(ϕ0 (x))(1 ⊗ F ) . So (2.21) holds and (2.20) is proved. Combining (2.20) and (2.19) and using that ζl (ϕn (a∗ ))π(F )ζl (ϕ0 (ϕn (a))) is an element of S1 with norm at most 2Λ(G)3 , we get that δ ≤ lim sup Re Ω1 (ζl (ϕn (a∗ ))π(F )ζl (ϕ0 (ϕn (a)))π(F )) + ε(Λ(G)2 + 1) n ≤ lim sup Re Ω1 (ζl (ϕn (a∗ ))π(F )ζl (ϕ0 (ϕn (a)))) + δ/2 . n As in (2.18) and using the ζl (A)-centrality of Ω1 , we conclude that δ/2 ≤ Re Ω1 (ζl (a∗ )π(F )ζl (ϕ0 (a))) = Re Ω1 (π(F )ζl (ϕ0 (a)a∗ )) for all a ∈ U (A). For every T ∈ S1′′ and x ∈ M , we have |Ω1 (T ζl (x))|2 ≤ Ω1 (T T ∗ ) Tr(px∗ xp) . So we find a unique η ∈ L2 (M p) such that Ω1 (π(F )ζl (x)) = hxp, ηi for all x ∈ M . It then follows that δ/2 ≤ Rehϕ0 (a)a∗ , ηi for all a ∈ U (A) . Since ω0 ∈ A(G), we can take ξ1 , ξ2 ∈ L2 (G) such that ω0 (g) = hλg ξ1 , ξ2 i for all g ∈ G. It follows that δ/2 ≤ RehΦ(a)(p ⊗ ξ1 )a∗ , η ⊗ ξ2 i for all a ∈ U (A) . Defining ξ3 ∈ Φ(p)(L2 (M p) ⊗ L2 (G)) as the element of minimal norm in the closed convex hull of {Φ(a)(p ⊗ ξ1 )a∗ | a ∈ U (A)}, we conclude that ξ3 satisfies Φ(a)ξ3 = ξ3 a for all a ∈ A and that Rehξ3 , η ⊗ ξ2 i ≥ δ/2. So, ξ3 6= 0 and we have proven that A can be Φ-embedded. 3 Uniqueness of Cartan subalgebras; proof of Theorem A Theorem A is a special case of the following general result. To formulate this result, recall that a nonsingular action G y (X, µ) of a lcsc group G on a standard probability space is called amenable in the sense of Zimmer if there exists a G-equivariant conditional expectation E : L∞ (X × G) → L∞ (X) w.r.t. the action G y X × G given by g · (x, h) = (g · x, gh). 13 Theorem 3.1. Let G = G1 × · · · × Gn be a direct product of lcsc weakly amenable groups with property (S). Let G y (X, µ) be an essentially free nonsingular action. Denote by G◦i the direct product of all Gj , j 6= i. If for every i ∈ {1, . . . , n} and every non-null G-invariant Borel set X0 ⊂ X, the action ◦ Gi y L∞ (X0 )Gi is nonamenable in the sense of Zimmer, then L∞ (X) ⋊ G has a unique Cartan subalgebra up to unitary conjugacy. In particular, L∞ (X) ⋊ G has a unique Cartan subalgebra up to unitary conjugacy when the groups Gi are nonamenable and the action G y (X, µ) is either probability measure preserving or irreducible. Cross section equivalence relations. Theorem 3.1 is proven by using cross section equivalence relations. These were introduced in [Fo74, Co79] and a rather self-contained approach can be found in [KPV13, Section 4.1 and Appendix B]. Let G y (X, µ) be a nonsingular action of a lcsc group G on the standard probability space (X, µ). This means that X is a standard Borel space and that G y X is a Borel action that leaves the measure µ quasi-invariant in the sense that µ(g · U ) = 0 if and only if µ(U ) = 0, whenever U ⊂ X is Borel and g ∈ G. Assume that this action is essentially free, meaning that almost every point x ∈ X has a trivial stabilizer. Since the set of points x ∈ X having a trivial stabilizer is a Borel subset of X, we may equally well assume that the action is really free. A cross section for G y (X, µ) is a Borel subset X1 ⊂ X with the following two properties. • There exists a neighborhood U of e in G such that the map U × X1 → X : (g, x) 7→ g · x is injective. • The subset G · X1 ⊂ X is conull. Note that the first condition implies that the map G × X1 → X : (g, x) 7→ g · x is countableto-one and thus, maps Borel sets to Borel sets. So, the set G · X1 appearing in the second condition is Borel. A partial cross section for G y (X, µ) is a Borel subset X1 ⊂ X satisfying the first condition and satisfying the property that G · X1 is non-null. Given any partial cross section X1 , the equivalence relation R on X1 defined by R = {(y, y ′ ) ∈ X1 × X1 | y ∈ G · y ′ } is Borel and has countable equivalence classes. Also, X1 has a canonical measure class, given by a probability measure µ1 , and this measure µ1 is quasi-invariant under the equivalence relation R. The cross section equivalence relation R on (X1 , µ1 ) is thus a countable nonsingular Borel equivalence relation. By construction, the von Neumann algebra L(R) is canonically isomorphic with q(L∞ (X)⋊G)q, for some projection q, see e.g. [KPV13, Lemma 4.5]. In this way, cross sections define the canonical Cartan subalgebra L∞ (X) ⋊ G. As will become clear in the proof of Theorem 3.1, it is useful to allow µ to be a σ-finite measure and to consider the special case where µ is scaled by the inverse of the modular function of G, meaning that µ(g · U ) = δ(g)−1 µ(U ) for all Borel sets U ⊂ X and all g ∈ G. This covers in particular the case where µ is a G-invariant probability measure and G is unimodular. Fix a right invariant Haar measure λ on G and recall that λ(gU ) = δ(g)−1 λ(U ) for all Borel sets U ⊂ G and g ∈ G. Let µ be a σ-finite measure on X that is scaled by the inverse of the modular function. Let X1 ⊂ X be any partial cross section (and note that the definitions above only depend on the 14 measure class of µ so that taking µ to be σ-finite makes no difference). Then there is a unique σ-finite measure µ1 on X1 such that the following holds: whenever U is a neighborhood of e in G such that the map Ψ : U ×X1 → X : (g, x) 7→ g ·x is injective, we have Ψ∗ (λ|U ×µ1 ) = µ|U ·X1 . This measure µ1 is invariant under the cross section equivalence relation. In the case where G is unimodular and µ is a G-invariant probability measure, we get that µ1 is a finite R-invariant measure. It is then more customary to normalize µ1 , so that µ1 becomes an R-invariant probability measure on X1 and Ψ∗ (λ|U × µ1 ) = covol(X1 ) µ|U ·X1 . The scaling factor covol(X1 ) is called the covolume of X1 . Note that this covolume is proportional to the choice of the Haar measure on G. The coaction Φω associated with a cocycle ω and ω-compactness. Let R be a countable pmp equivalence relation on the standard probability space (X1 , µ1 ). Denote by [R] its full group, i.e. the group of all pmp isomorphisms ϕ : X1 → X1 with the property that (ϕ(x), x) ∈ R for all x ∈ X1 . Denote by [[R]] the full pseudogroup of R, consisting of all partial measure preserving transformations with the property that (ϕ(x), x) ∈ R for all x ∈ dom(ϕ). The tracial von Neumann algebra M = L(R) is generated by the Cartan subalgebra L∞ (X1 ) and the unitary elements uϕ , ϕ ∈ [R], normalizing L∞ (X1 ). Similarly, every ϕ ∈ [[R]] defines a partial isometry uϕ ∈ M . Finally, R is equipped with a natural σ-finite measure and L2 (M ) is naturally identified with L2 (R). Let G be a lcsc group and ω : R → G a cocycle, i.e. a Borel map satisfying ω(x, y) ω(y, z) = ω(x, z) for a.e. (x, y, z) ∈ R(2) , where R(2) = {(x, y, z) ∈ X1 × X1 × X1 | (x, y) ∈ R and (y, z) ∈ R} is equipped with its natural σ-finite measure. Note that ω(x, x) = e for a.e. x ∈ X. We say that a von Neumann subalgebra B ⊂ pM p is ω-compact if for every ε > 0, there exists a compact subset K ⊂ G such that ω kb − PK (b)k2 ≤ ε kbk for all b ∈ B , ω is the orthogonal projection of L2 (R) onto L2 (ω −1 (K)). where PK Given a von Neumann subalgebra B ⊂ M , the same argument as in the proof of [Va10b, Proposition 2.6] implies that the set of projections {p ∈ B ′ ∩ M | p is a projection and Bp is ω-compact } attains its maximum in a unique projection p and that this projection p belongs to NM (B)′ ∩M . We associate to ω the coaction Φω : M → M ⊗ L(G) : Φω (F ) = F ⊗ 1 , Φω (uϕ ) = (uϕ ⊗ 1)Vϕ , for all F ∈ L∞ (X1 ) and all ϕ ∈ [R], where Vϕ ∈ L∞ (X1 ) ⊗ L(G) is given by Vϕ (x) = λω(ϕ(x),x) . We deduce from Theorem F the following result. Theorem 3.2. Let R be a countable pmp equivalence relation on the standard probability space (X1 , µ1 ). Let G be a weakly amenable locally compact group with property (S) and ω : R → G a cocycle. Write M = L(R) and assume that A ⊂ M is a Φω -amenable von Neumann subalgebra with normalizer P = NM (A)′′ . Denote by p ∈ P ′ ∩ M the unique maximal projection such that Ap is ω-compact. Then, P (1 − p) is Φω -amenable. 15 Proof. By Theorem F, we only have to prove the following statement: if p ∈ A′ ∩M is a nonzero projection such that Ap can be Φω -embedded, then there exists a nonzero projection q ∈ A′ ∩M such that q ≤ p and Aq is ω-compact. Since Ap can be Φω -embedded, there exists a nonzero vector ξ ∈ L2 (M ) ⊗ L2 (G) such that Φω (p)ξ = ξ = ξ(p ⊗ 1) and such that Φω (a)ξ = ξ(a ⊗ 1) for all a ∈ Ap. Denote by q the smallest projection in M that satisfies ξ = ξ(q ⊗ 1). Then, q ∈ A′ ∩ M , q ≤ p and q 6= 0. Viewing ξ as affiliated with the W∗ -module M ⊗ L2 (G), we can take the polar decomposition of ξ and find V ∈ M ⊗ L2 (G) satisfying V ∗ V = q and V a = Φω (a)V for all a ∈ Ap. Define G ⊂ [[R]] consisting of all ϕ ∈ [[R]] for which the set {ω(ϕ(x), x) | x ∈ dom(ϕ)} has compact closure in G. For every ϕ ∈ [R] and every ε > 0, we can choose a Borel set U ⊂ X1 with µ1 (X1 \ U ) < ε such that the restriction of ϕ to U belongs to G. Therefore, the linear span of all F uϕ , F ∈ L∞ (X1 ), ϕ ∈ G, defines a dense ∗-subalgebra M0 of M . By construction, ω (x). for every x ∈ M0 , there exists a compact subset K ⊂ G such that x = PK Choose ε > 0. Consider on the W∗ -module M ⊗ L2 (G) the norm k · k2 given by the embedding M ⊗ L2 (G) ⊂ L2 (M ) ⊗ L2 (G), as well as the operator norm k · k∞ . By the Kaplansky density theorem, we can take W ∈ M0 ⊗alg Cc (G) ⊂ M ⊗ L2 (G) such that kW k∞ ≤ 1 and kV − W k2 < ε/3 and kW ∗ W − qk2 < ε/3. For every a ∈ M , we find that kV a− W ak2 ≤ kV − W k2 kak ≤ Therefore, kΦω (a)W −W ak2 ≤ we find that ε kak 3 2ε 3 and kΦω (a)V − Φω (a)W k2 ≤ kak kV − W k2 ≤ ε kak . 3 kak for all a ∈ Ap. Since kW k∞ ≤ 1 and kW ∗ W −qk2 < ε/3, kW ∗ Φω (a)W − aqk2 ≤ ε kak (3.1) for all a ∈ Ap. When ξi ∈ Cc (G) have (compact) supports Ki ⊂ G, then (1 ⊗ ξ2∗ )Φω (x)(1 ⊗ ξ1 ) belongs to L2 (ω −1 (K2 K1−1 )) for all x ∈ M . So because W ∈ M0 ⊗alg Cc (G), we can take a compact subset ω for every x ∈ M . It then follows from K ⊂ G such that W ∗ Φω (x)W belongs to the range of PK ω (3.1) that kPK (aq) − aqk2 ≤ ε kak for all a ∈ Ap. For every element a ∈ Aq, we can choose ω (a) − ak ≤ ε kak for all a1 ∈ Ap with ka1 k = kak and a = a1 q. So we have proved that kPK 2 a ∈ Aq. Since ε > 0 was arbitrary, this means that Aq is ω-compact. In the formulation of Corollary 3.3 below, we make use of the following notion of an amenable pair of group actions, as introduced in [AD81]. Let G be a lcsc group and let G y (Y, η) and G y (X, µ) be nonsingular actions. Assume that p : Y → X is a G-equivariant Borel map such that the measures p∗ (η) and µ are equivalent. Following [AD81, Définition 2.2], the pair (Y, X) is called amenable if there exists a G-equivariant conditional expectation L∞ (Y, η) → L∞ (X, µ). In particular, the action G y (X, µ) is amenable in the sense of Zimmer if and only if the pair (X × G, X) with g · (x, h) = (g · x, gh) is amenable. Corollary 3.3. Let G be a lcsc group and G y (X, µ) an essentially free nonsingular action on the standard σ-finite measure space (X, µ). Assume that the action scales the measure µ by the inverse of the modular function of G. Let (X1 , µ1 ) be a partial cross section with µ1 (X1 ) < ∞ and denote by R the cross section equivalence relation on (X1 , µ1 ), which is a countable equivalence relation with invariant probability measure µ1 (X1 )−1 µ1 . Let H be a weakly amenable locally compact group with property (S) and π : G → H a continuous group homomorphism. Denote by ω : R → H the cocycle given by the composition of π and the canonical cocycle ω0 : R → G determined by ω0 (x′ , x) · x = x′ for all (x′ , x) ∈ R. 16 Let A ⊂ L(R) be a Cartan subalgebra. If A is not ω-compact, then there exists a non-null G-invariant Borel set X0 ⊂ X and a G-equivariant conditional expectation L∞ (X0 × H) → L∞ (X0 ) w.r.t. the action g · (x, h) = (g · x, π(g)h). Proof. Write M = L(R). Assume that the Cartan subalgebra A ⊂ M is not ω-compact. By Theorem 3.2, we can take a nonzero central projection p ∈ Z(M ) such that M p is Φω -amenable. Write p = 1X2 , where X2 ⊂ X1 is an R-invariant Borel set. Put X0 = G·X2 . Then X0 is a nonnull G-invariant Borel set. We prove that there exists a G-equivariant conditional expectation L∞ (X0 × H) → L∞ (X0 ). Since M p is Φω -amenable and since p ∈ L∞ (X1 ) so that Φω (p) = p⊗1, there exists a conditional expectation M p⊗B(L2 (H)) → Φω (M p). Since (X1 , µ1 ) is a partial cross section, we can choose a compact neighborhood K of e in G such that Ψ : K × X1 → X : Ψ(k, x) = k · x is injective. Write N = L∞ (X) ⋊ G and define the coaction Φπ : N → N ⊗ L(H) given by Φπ (F ug ) = F ug ⊗ uπ(g) for all F ∈ L∞ (X), g ∈ G . Define the projection q1 ∈ L∞ (X) given by q1 = 1K·X1 . In [KPV13, Lemma 4.5], an explicit isomorphism q1 N q1 ∼ (3.2) = B(L2 (K)) ⊗ M is constructed. Under this isomorphism, the restriction of Φπ to q1 N q1 is unitarily conjugate with id ⊗ Φω and the projection q2 = 1K·X2 corresponds to 1 ⊗ p. We thus conclude that there exists a conditional expectation q2 N q2 ⊗ B(L2 (H)) → Φπ (q2 N q2 ). Since q0 = 1X0 is the central support of q2 inside N , there also exists a conditional expectation E : N q0 ⊗ B(L2 (H)) → Φπ (N q0 ). We now restrict E to L∞ (X0 × H) ⊂ N q0 ⊗ B(L2 (H)). For all F ∈ L∞ (X0 × H) and F ′ ∈ L∞ (X0 ), we have E(F ) Φπ (F ′ ) = E(F Φπ (F ′ )) = E(F (F ′ ⊗ 1)) = E((F ′ ⊗ 1) F ) = E(Φπ (F ′ )F ) = Φπ (F ′ ) E(F ) . Since L∞ (X0 ) ⊂ N q0 is maximal abelian, it follows that E(F ) ∈ Φπ (L∞ (X0 )) = L∞ (X0 ) ⊗ 1. Define the conditional expectation E0 : L∞ (X0 × H) → L∞ (X0 ) such that E(F ) = E0 (F ) ⊗ 1. Since the action G y L∞ (X0 × H) is implemented by the unitary operators Φπ (ug q0 ), g ∈ G, it follows that E0 is G-equivariant. This concludes the proof of the corollary. Lemma 3.4. Let N be a σ-finite von Neumann algebra. Assume that the Connes-Takesaki continuous core c(N ) = N ⋊σϕ R has at most one Cartan subalgebra up to unitary conjugacy. Then the same holds for N itself. Proof. Let A and B be Cartan subalgebras of N . Denote by EA : N → A and EB : N → B the unique faithful normal conditional expectations. Let z ∈ Z(N ) be a nonzero central projection. Note that z ∈ A ∩ B. The main part of the proof consists in showing that Az ≺N z Bz, where we use the type III variant of Popa’s intertwining relation [Po03, Section 2] as defined in [HV12, Definition 2.4]. Assuming that Az 6≺N z Bz, we deduce a contradiction. By [HV12, Theorem 2.3], there exists a net of unitaries an ∈ U (Az) such that EB (x∗ an y) → 0 ∗-strongly for all x, y ∈ N . (3.3) Choose faithful normal states ϕ on A and ψ on B. Still denote by ϕ and ψ the faithful normal states on N given by ϕ ◦ EA , resp. ψ ◦ EB . The continuous core of N can then be realized as 17 cϕ (N ) = N ⋊σϕ R and as cψ (N ) = N ⋊σψ R. Denote by Θ : cϕ (N ) → cψ (N ) the canonical ∗-isomorphism given by Connes’ Radon-Nikodym theorem. e ⊂ M and B e⊂M e = Θ(A⋊σϕ R) and B e = B ⋊σψ R. Write M = cψ (N ) and note that A Write A are Cartan subalgebras. Denote by Tr the canonical faithful normal semifinite trace on M and e be the trace preserving conditional expectations. Note that E e (x) = EB (x) let EBe : M → B B for all x ∈ N . We prove that EBe (x∗ an y) → 0 ∗-strongly for all x, y ∈ M . (3.4) e once (3.4) is proved, it follows that the Cartan subalgebras Since an is a net of unitaries in Az, e e A and B cannot be unitarily conjugate, contradicting the assumptions of the theorem. So once (3.4) is proved, it follows that Az ≺N z Bz. e is abelian, to prove (3.4), it suffices to prove that Since B lim kEBe (x∗ an y)k2,Tr = 0 for all x, y ∈ M with Tr(x∗ x) < ∞ and Tr(y ∗ y) < ∞. n (3.5) Approximating x, y in k · k2,Tr , it suffices to prove (3.5) for all x, y of the form x = x1 x0 and e with Tr(x∗ x0 ) < ∞ and Tr(y ∗ y0 ) < ∞. But then, y = y1 y0 with x1 , y1 ∈ N and x0 , y0 ∈ B 0 0 EBe (x∗ an y) = x∗0 EBe (x∗1 an y1 ) y0 = x∗0 EB (x∗1 an y1 ) y0 , so that (3.5) follows from (3.3). Thus, (3.4) is proved. As we already explained, it follows that Az ≺N z Bz for every nonzero central projection z ∈ Z(N ). Let z0 ∈ Z(N ) be the maximal central projection such that Az0 and Bz0 are unitarily conjugate inside N z0 . Assume that z0 6= 1 and put z = 1− z0 . By the above, Az ≺N z Bz. By the type III version of Popa’s [Po01, Theorem A.1] proved in [HV12, Theorem 2.5], there exists a nonzero central projection z1 ∈ Z(N )z such that Az1 and Bz1 are unitarily conjugate. This contradicts the maximality of z0 , so that z0 = 1. We are now ready to prove Theorem 3.1. Proof of Theorem 3.1. Take G = G1 × · · · × Gn as in the formulation of the theorem. Let G y (X, µ) be an essentially free nonsingular action and assume that the hypotheses of the theorem hold. We have to prove that N = L∞ (X) ⋊ G has a unique Cartan subalgebra up to unitary conjugacy. By Lemma 3.4, it is enough to prove that the continuous core c(N ) has a unique Cartan subalgebra up to unitary conjugacy. e ⋊ G where The continuous core c(N ) can be realized as a crossed product c(N ) = L∞ (X) e G y (X, µ e) is the Maharam extension given by  e = X × R , g · (x, t) = g · x , t + log(δ(g)) + log(D(g, x)) , de X µ(x, t) = dµ(x) × exp(−t)dt , where δ : G → R+ ∗ is the modular function of G and D is the Radon-Nikodym cocycle for G y (X, µ) determined by Z Z −1 F (x) D(g, x) dµ(x) . F (g · x) dµ(x) = X X e scales the measure µ Note that the action G y X e with δ−1 . e µ Let (X1 , µ1 ) be a cross section for G y (X, e). Denote by R1 the cross section equivalence relation on (X1 , µ1 ). To prove that c(N ) has a unique Cartan subalgebra, it suffices to prove 18 that L∞ (X1 ) is the unique Cartan subalgebra of L(R1 ), up to unitary conjugacy. So it suffices to prove that for every non-null Borel set X2 ⊂ X1 with µ1 (X2 ) < ∞, the restricted equivalence relation R = (R1 )|X2 has the property that L∞ (X2 ) is the unique Cartan subalgebra of L(R) up to unitary conjugacy. Denote by µ2 the restriction of µ1 to X2 . Then (X2 , µ2 ) is a partial e µ cross section for G y (X, e) and µ2 (X2 ) < ∞. Let A ⊂ L(R) be another Cartan subalgebra. Denote by ω : R → G the canonical cocycle determined by ω(x′ , x) · x = x′ for all (x′ , x) ∈ R. We claim that A is ω-compact. Denote by πi : G → Gi the quotient maps and put ωi = πi ◦ ω. To prove the claim that A is ω-compact, it suffices to prove that A is ωi -compact for every i ∈ {1, . . . , n}. Fix such an i and assume that A is not ωi -compact. e0 ⊂ X e and a G-equivariant By Corollary 3.3, we find a non-null G-invariant Borel set X ∞ ∞ e e conditional expectation E0 : L (X0 × Gi ) → L (X0 ) w.r.t. the action g · ((x, t), g′ ) = (g · (x, t), πi (g)g′ ). e given by s · (g, t) = (g, t + s). Note that this Denote by (βs )s∈R the action of R on L∞ (X) action of R commutes with the above G-action. Write p = 1Xe0 and denote by q the smallest e with p ≤ q. Note that q = 1X ×R , where X0 ⊂ X is a (βs )s∈R -invariant projection in L∞ (X) 0 W G-invariant Borel set. Choose sk ∈ R, with s0 = 0, such that q = ∞ k=0 βsk (p). Inductively ∞ e define the G-invariant projections pk ∈ L (X0 ) given by p0 = p and k−1   X βsi −sk (pi ) for all k ≥ 1. pk = p 1 − i=0 By construction, q = P k βsk (pk ). Choosing a point-weak∗ limit point of the sequence En : L∞ (X0 × R × Gi ) → L∞ (X0 × R) : En (F ) = n X k=0  βsk E0 (pk ⊗ 1) (β−sk ⊗ id)(F ) , we obtain a G-equivariant conditional expectation E : L∞ (X0 × R × Gi) → L∞ (X0 × R). Since R is amenable, we can take a mean over R of βs ◦ E ◦ (β−s ⊗ id), so that we may assume that E is G × R-equivariant. The restriction of E to L∞ (X0 )⊗1⊗L∞ (Gi ) then has its image in L∞ (X0 ×R)R = L∞ (X0 )⊗1. So, we find a G-equivariant conditional expectation L∞ (X0 × Gi ) → L∞ (X0 ). Restricting to ◦ L∞ (X0 )Gi ⊗ L∞ (Gi ), we find a Gi -equivariant conditional expectation ◦ ◦ L∞ (X0 )Gi ⊗ L∞ (Gi ) → L∞ (X0 )Gi . ◦ This precisely means that the action Gi y L∞ (X0 )Gi is amenable in the sense of Zimmer, contrary to our assumptions. So the claim that A is ω-compact is proved. Take a compact subset K ⊂ G such that ω (a)k2 ≥ 1/2 for all a ∈ U (A). Since K is compact and ω : R → G is the canonical kPK 2 cocycle, the subset ω −1 (K) ⊂ R is bounded, meaning that ω −1 (K) is the disjoint union of the graphs of finitely many elements ϕi ∈ [[R]], i = 1, . . . , n, in the full pseudogroup of R. But then, writing B = L∞ (X2 ), n X ω kEB (au∗ϕi )k22 kPK (a)k22 = i=1 ω (a)k2 kPK 2 ≥ 1/2 for all a ∈ U (A), it follows that A ≺L(R) B, so that A for all a ∈ L(R). Since and B are unitarily conjugate by [Po01, Theorem A.1]. 19 4 Cocycle and orbit equivalence rigidity; proof of Theorem B Given an irreducible pmp action of G = G1 ×G2 on a standard probability space (X, µ), Monod and Shalom proved in [MS04, Theorem 1.2] a cocycle superrigidity theorem for non-elementary cocycles G × X → H with values in a closed subgroup H < Isom(X) of the isometry group of a “negatively curved” space. It is therefore not surprising that one can also prove a cocycle superrigidity theorem for cocycles with values in a group H satisfying property (S). We do this in Theorem 4.1. Applying cocycle superrigidity to the cocycles given by a stable orbit equivalence between essentially free, irreducible pmp actions G1 × G2 y (X, µ) and H1 × H2 y (Y, η), we obtain the following orbit equivalence strong rigidity theorem (see Theorem 4.2): if G1 and G2 are nonamenable, while H1 and H2 have property (S), the actions must be conjugate. Again, such an orbit equivalence strong rigidity theorem should not come as a surprise: in [Sa09, Theorem 40], Sako proved exactly this result when G1 , G2 and H1 , H2 are countable groups in class S. However, he does not use or prove a cocycle superrigidity theorem. The main novelty of this section is that our approach is surprisingly simple and short. Given lcsc groups G and H and a nonsingular action G y (X, µ), a Borel cocycle ω : G×X → H is a Borel map satisfying ω(gh, x) = ω(g, h · x) ω(h, x) for all g, h ∈ G, x ∈ X . In a measurable context, the slightly more appropriate notion of cocycle is however the following. Denote by M(X, H) the Polish group of Borel functions from X to H, modulo functions equal almost everywhere. The group G acts continuously on M(X, H) by (αg (F ))(x) = F (g−1 · x). Then a cocycle is a continuous map ω : G → M(X, H) : g 7→ ωg satisfying ωgh = αh−1 (ωg ) ωh for all g, h ∈ G . Every Borel cocycle ω gives rise to the cocycle ωg = ω(g, ·). Conversely, every cocycle can be realized by a Borel cocycle after removing from X a G-invariant Borel set of measure zero, see e.g. [Zi84, Theorem B.9]. The (measurable) cocycles ω and ω ′ are called cohomologous if there exists an element ϕ ∈ M(X, H) such that ωg′ = αg−1 (ϕ) ωg ϕ−1 for all g ∈ G . Borel cocycles ω, ω ′ : G×X → H are called cohomologous if there exists a Borel map ϕ : X → H such that ω ′ (g, x) = ϕ(g · x) ω(g, x) ϕ(x)−1 for all g ∈ G, x ∈ X . Again, if two Borel cocycles are measurably cohomologous, then they also are Borel cohomologous on a conull G-invariant Borel set. As in [MS04, Theorem 1.2], the following cocycle superrigidity theorem says that every “nonelementary” cocycle for an irreducible action G1 ×G2 y (X, µ) with values in a group with property (S) is cohomologous to a group homomorphism. In our context, being “non-elementary” is expressed by a non relative amenability property introduced in [AD81] (see the discussion preceding Corollary 3.3). Theorem 4.1. Let G1 , G2 and H be lcsc groups and G1 × G2 y (X, µ) a pmp action with G2 acting ergodically. Assume that H has property (S). Let ω : G1 × G2 × X → H be a cocycle. Then at least one of the following statements holds. 20 e < H such that K is compact and K < H e is normal, 1. There exist closed subgroups K < H e and there exists a continuous group homomorphism δ : G1 → H/K with dense image such that ω is cohomologous to a cocycle ω0 satisfying ω0 (g1 g2 , x) ∈ δ(g1 )K for all gi ∈ Gi and a.e. x ∈ X. 2. With respect to the action G1 y X × H given by g1 ·(x, h) = (g1 ·x, ω(g1 , x) h) and the factor map (x, h) 7→ x, there exists a G1 -equivariant conditional expectation L∞ (X ×H) → L∞ (X). Proof. Throughout the proof, we write G = G1 ×G2 and we view G1 and G2 as closed subgroups of G. We fix a left invariant Haar measure λ on H. We denote by h · ξ the left translation action of H on L2 (H). Formulation of the dichotomy. We are in precisely one of the following situations. 1. There exists no sequence gn ∈ G2 such that ω(gn , ·) → ∞ in measure. More precisely, there exists a compact subset L ⊂ H and an ε > 0 such that for all g ∈ G2 the set {x ∈ X | ω(g, x) ∈ L} has measure at least ε. 2. There exists a sequence gn ∈ G2 such that ω(gn , ·) → ∞ in measure. Case 1. Fix such a compact set L ⊂ H and ε > 0. Define the unitary representation π : G → U (L2 (X × H)) : (π(g)∗ ξ)(x, h) = ξ(g · x, ω(g, x)h) for all g ∈ G, x ∈ X, h ∈ H, ξ ∈ L2 (X × H). Fix a compact subset L0 ⊂ H with λ(L0 ) > 0. Given a Borel set A ⊂ H of finite measure, denote by 1A ∈ L2 (H) the function equal to 1 on A and equal to 0 elsewhere. By our choice of L and ε, we find that hπ(g)∗ (1 ⊗ 1LL0 ), 1 ⊗ 1L0 i ≥ ε λ(L0 ) for all g ∈ G2 . Taking the unique vector of minimal norm in the closed convex hull of {π(g)(1⊗1LL0 ) | g ∈ G2 }, it follows that π admits a nonzero G2 -invariant vector. We thus find a Borel map ξ : X → L2 (H) such that ξ(g2 · x) = ω(g2 , x) · ξ(x) for all g2 ∈ G2 and a.e. x ∈ X, and such that ξ is not zero a.e. Since x 7→ kξ(x)k2 is essentially G2 -invariant and the action G2 y (X, µ) is ergodic, we may assume that kξ(x)k2 = 1 for a.e. x ∈ X. Denote by T ⊂ L2 (H) the unit sphere, defined as T = {ξ0 ∈ L2 (H) | kξ0 k2 = 1}. The left translation action H y T has closed orbits and thus H\T is a well defined Polish space. Since the map x 7→ H · ξ(x) from X to H\T is G2 -invariant, it is constant a.e. So we find a unit vector ξ0 ∈ L2 (H) and a Borel map ϕ : X → H such that ξ(x) = ϕ(x) · ξ0 for a.e. x ∈ X. Replacing ω by the cohomologous cocycle given by (g, x) 7→ ϕ(g · x)−1 ω(g, x) ϕ(x) , we find that ω(g2 , x) · ξ0 = ξ0 for all g2 ∈ G2 and a.e. x ∈ X. Define the closed subgroup K < H given by K = {s ∈ H | s · ξ0 = ξ0 } . Then, K is compact and ω(g2 , x) ∈ K for all g2 ∈ G2 and a.e. x ∈ X. By Zimmer’s theory for compact group valued cocycles (see [Zi75, Section 3]), we may further assume that the restricted cocycle ω2 : G2 × X → K : ω2 = ω|G2 ×X 21 is minimal, in the sense that the associated action G2 y X × K given by g2 · (x, k) = (g2 · x, ω2 (g2 , x)k) is ergodic. Whenever g1 ∈ G1 and g2 ∈ G2 , we have for a.e. x ∈ X ω(g1 , g2 · x) ω(g2 , x) = ω(g1 g2 , x) = ω(g2 g1 , x) = ω(g2 , g1 · x) ω(g1 , x) . (4.1) Fix g1 ∈ G1 . It follows from (4.1) that for all g2 ∈ G2 and a.e. x ∈ X, ω(g1 , g2 · x) ∈ K · ω(g1 , x) · K. Therefore, the map X → K\H/K : x 7→ K · ω(g1 , x) · K is G2 -invariant and thus constant a.e. We then find s ∈ H and Borel maps ϕ, ψ : X → K such that (g1 still being fixed) we have ω(g1 , x) = ϕ(x) s ψ(x) for a.e. x ∈ X. Then (4.1) becomes ϕ(g2 · x) s ψ(g2 · x) ω(g2 , x) = ω(g2 , g1 · x) ϕ(x) s ψ(x) for all g2 ∈ G2 and a.e. x ∈ X. So, the cocycle ω2′ : G2 × X → K : ω2′ (g2 , x) = ψ(g2 · x) ω2 (g2 , x) ψ(x)−1 is cohomologous to ω2 (as cocycles for G2 y X with values in the compact group K) and takes values in K ∩ s−1 Ks. The minimality of ω2 then implies that K ∩ s−1 Ks = K. Making a similar reasoning for the cocycle (g2 , x) 7→ ω(g2 , g1 · x), which by construction is isomorphic with ω2 and thus minimal as well, we also find that K ∩ sKs−1 = K. Defining the closed subgroup H ′ < H given by H ′ := {s ∈ H | sKs−1 = K} , we find that s ∈ H ′ . By construction, K < H ′ is normal. We have proved that for every g1 ∈ G1 , there exists an s ∈ H ′ such that ω(g1 , x) ∈ sK for a.e. x ∈ X. We already had ω(g2 , x) ∈ K for all g2 ∈ G2 and a.e. x ∈ X. So we find a Borel and thus continuous homomorphism δ : G1 → H ′ /K such that ω(g1 g2 , x) ∈ δ(g1 )K for all g1 ∈ G1 , g2 ∈ G2 and e < H ′ as the inverse image of the closure of δ(G1 ), the first statement a.e. x ∈ X. Defining H in the theorem holds. Case 2. Fix a sequence gn ∈ G2 such that ω(gn , ·) → ∞ in measure. Fix a map η : H → S(H) as given by property (S). Define the sequence of Borel maps ηn : X → S(H) : ηn (x) = η(ω(gn , x)−1 ) . By (4.1), we have for all g ∈ G1 that ω(gn , g · x)−1 = ω(g, x) ω(gn , x)−1 ω(g, gn · x)−1 . (4.2) Fix g ∈ G1 and fix ε > 0. Take a compact subset L ⊂ H such that ω(g, x) ∈ L for all x in a set of measure at least 1 − ε. Then take a compact subset L1 ⊂ H such that kη(h1 hh−1 2 ) − h1 · η(h)k1 < ε for all h1 , h2 ∈ L and all h ∈ H \ L1 . Finally take n0 such that for all n ≥ n0 , we have that ω(gn , x)−1 ∈ H \ L1 for all x in a set of measure at least 1 − ε. 22 So, for our fixed g ∈ G1 and for all n ≥ n0 , there exists a Borel set Xn ⊂ X of measure at least 1 − 3ε such that ω(g, x) ∈ L , ω(gn , x)−1 ∈ H \ L1 , ω(g, gn · x) ∈ L for all x ∈ Xn . Applying η to (4.2), we conclude that for our fixed g ∈ G1 and all n ≥ n0 , we have kηn (g · x) − ω(g, x) · ηn (x)k1 < ε for all x ∈ Xn . Since µ(Xn ) ≥ 1 − 3ε, we have proved that for every g ∈ G1 , the sequence of functions x 7→ kηn (x) − ω(g, x)−1 · ηn (g · x)k1 (4.3) converges to zero in measure. View S(H) ⊂ L1 (H) and define the normal conditional expectations Z ∞ ∞ F (x, y) (ηn (x))(y) dy . Pn : L (X × H) → L (X) : (Pn (F ))(x) = H Choose a point-weak∗ limit point P : L∞ (X × H) → L∞ (X). Since the sequence in (4.3) converges to zero in measure, P is a G1 -equivariant conditional expectation. So the second statement in the theorem holds. The cocycle superrigidity theorem 4.1 implies the following orbit equivalence strong rigidity theorem. As mentioned above, for countable groups, the same result was obtained in [Sa09, Theorem 40]. Combining Theorem A and Theorem 4.2, it follows that Theorem B holds. Let G y (X, µ) and H y (Y, η) be essentially free, nonsingular actions of the lcsc groups G, H. We say that these actions are stably orbit equivalent if they admit cross sections such that the associated cross section equivalence relations are isomorphic. Theorem 4.2. Let G = G1 × G2 and H = H1 × H2 be unimodular lcsc groups without nontrivial compact normal subgroups. Assume that G y (X, µ) and H y (Y, η) are essentially free, irreducible pmp actions. Assume that G1 , G2 are nonamenable and that H1 , H2 have property (S). If the actions are stable orbit equivalent, they must be conjugate. More precisely, if (X1 , µ1 ) and (Y1 , η1 ) are cross sections, with cross section equivalence relations R and S, and if π : X1 → Y1 is a nonsingular isomorphism between the equivalence relations R and S, there exist conull R-invariant (resp. S-invariant) Borel sets X2 ⊂ X1 and Y2 ⊂ Y1 , a Borel bijection ∆ : G · X2 → H · Y2 and a continuous group isomorphism δ : G → H such that • X0 = G · X2 and Y0 = H · Y2 are conull Borel sets and ∆∗ (µ) = η, • ∆(g · x) = δ(g) · ∆(x) for all g ∈ G and all x ∈ X0 , • ∆(x) ∈ H · π(x) for all x ∈ X2 , • δ is either of the form δ1 × δ2 where δi : Gi → Hi are isomorphisms, or of the form (g1 , g2 ) 7→ (δ2 (g2 ), δ1 (g1 )) where δ1 : G1 → H2 and δ2 : G2 → H1 are isomorphisms, • normalizing the Haar measures λG and λH such that δ∗ (λG ) = λH , we have covol(X1 ) = covol(Y1 ). 23 Proof. Replacing X and Y by a conull G-invariant, resp. H-invariant, Borel set, we may assume that the Borel actions G y X and H y Y are free. Recall that µ1 is the natural R-invariant probability measure on X1 and that η1 is the natural S-invariant probability measure on Y1 . Normalize the Haar measures λG and λH such that covol(X1 ) = 1 = covol(Y1 ). Take compact neighborhoods U of e in G and V of e in H such that the maps Ψ : U × X1 → X : (k, x) 7→ k · x and Φ : V × Y1 → Y : (l, y) 7→ l · y (4.4) are injective. By the definition of a cross section and its covolume (see page 14), these maps satisfy   Ψ∗ (λG )|U × µ1 = µ|U ·X1 and Φ∗ (λH )|V × η1 = η|V·Y1 . Replacing X1 and Y1 by conull Borel subsets that are invariant under the cross section equivalence relations, we may assume that π : X1 → Y1 is a Borel isomorphism between R and S. Since π∗ (µ1 ) is an S-invariant probability measure on Y1 in the same measure class as η1 , we have π∗ (µ1 ) = η1 . Since G · X1 and H · Y1 are conull and Borel, we may assume that X = G · X1 and Y = H · Y1 . We start by translating the stable orbit equivalence π into a measure equivalence between G and H. This is quite standard: the discrete group case can be found in [Fu98, Section 3], but the locally compact case needs a little bit of care. Choose a Borel map p : X → X1 such that p(x) ∈ G · x for all x ∈ X and p(k · x) = x for all k ∈ U , x ∈ X1 . Extend π to the Borel map ρ : X → Y defined by ρ = π ◦ p. Similarly choose a Borel map q : Y → Y1 and define ρe : Y → X : ρe = π −1 ◦ q. By construction, ρ(G · x) ∈ H · ρ(x) and ρe(H · y) ∈ G · ρe(y) for all x ∈ X and all y ∈ Y . Since the actions G y X and H y Y are free, we have unique Borel cocycles ω : G × X → H : ρ(g · x) = ω(g, x) · ρ(x) for all g ∈ G, x ∈ X, ζ : H × Y → G : ρe(h · y) = ζ(h, y) · ρe(y) for all h ∈ H, y ∈ Y . Since ρe(ρ(x)) ∈ G · x and ρ(e ρ(y)) ∈ H · y for all x ∈ X and all y ∈ Y , we also have unique Borel maps ϕ : X → G : ρe(ρ(x)) = ϕ(x) · x and ψ : Y → H : ρ(e ρ(y)) = ψ(y) · y for all x ∈ X, y ∈ Y . Define the measure preserving Borel actions G × H y X × H and G × H y Y × G given by (g, h) · (x, h′ ) = (g · x, ω(g, x) h′ h−1 ) , (g, h) · (y, g ′ ) = (h · y, ζ(h, y) g′ g −1 ) . It is straightforward to check that  θ : X × H → Y × G : θ(x, h) = h−1 · ρ(x) , ζ(h−1 , ρ(x) ϕ(x)) is a G × H-equivariant Borel map and that θ is a bijection with inverse  θ −1 : Y × G → X × H : θ −1 (y, g) = g−1 · ρe(y) , ω(g −1 , ρe(y)) ψ(y) . Using the maps Ψ and Φ given by (4.4), one checks that θ(Ψ(k, x), l−1 ) = (Φ(l, π(x)), k −1 ) 24 for all k ∈ U , l ∈ V, x ∈ X1 . (4.5) Since Ψ, Φ and π are measure preserving, it follows that the restriction of θ to U · X1 × V −1 is measure preserving. Since the actions of G × H on X × H and Y × G are measure preserving and since the map θ is G×H-equivariant, it follows that the entire map θ is measure preserving. The main part of the proof consists in using Theorem 4.1 to show that the cocycle ω is cohomologous to an isomorphism of groups δ : G → H. Write ω(g, x) = (ω1 (g, x), ω2 (g, x)) ∈ H1 × H2 . Claim. W.r.t. the actions G1 y X × Hi : g1 · (x, hi ) = (g1 · x, ωi (g1 , x) hi ), there is at most one i ∈ {1, 2} for which there exists a G1 -equivariant conditional expectation L∞ (X × Hi ) → L∞ (X). To prove this claim, assume that such a conditional expectation exists for both i = 1, 2. Then there also exists a G1 -equivariant conditional expectation L∞ (X × H) → L∞ (X) w.r.t. the action g1 · (x, h) = (g1 · x, ω(g1 , x) h). Composing with the G-invariant probability measure µ on X, we find a G1 -invariant state on L∞ (X × H). Using θ, it follows that L∞ (Y × G1 × G2 ) admits a G1 -invariant state, w.r.t. the action g1 · (y, g1′ , g2 ) = (y, g1′ g1−1 , g2 ). This implies that G1 is amenable, contrary to our assumptions. So, the claim is proved. Assume that there is no G1 -equivariant conditional expectation L∞ (X × H1 ) → L∞ (X). We prove that the conclusions of the theorem hold with the group isomorphism δ : G → H being of the form δ1 × δ2 . In the case where there is no G1 -equivariant conditional expectation L∞ (X × H2 ) → L∞ (X), we exchange the roles of H1 and H2 and obtain again that the conclusions of the theorem hold with δ being of the form (g1 , g2 ) 7→ (δ2 (g2 ), δ1 (g1 )). Applying Theorem 4.1 to the cocycle ω1 , we find a compact subgroup K1 < H1 , a closed e 1 < H1 with K1 being a normal subgroup of H e 1 , and a continuous group homosubgroup H e 1 /K1 with dense image such that ω1 is cohomologous (as a measurable morphism δ : G1 → H e 1 satisfying ω cocycle) with a cocycle ω e1 : G × X → H e1 (g1 g2 , x) ∈ δ1 (g1 )K1 for all gi ∈ Gi and x ∈ X. In particular, there is an isomorphism between the actions of G on L∞ (X × H1 ) induced by ω1 and ω e1 . Using a K1 -invariant state on L∞ (H1 ), it follows that there is a G2 -equivariant conditional expectation L∞ (X × H1 ) → L∞ (X). Applying the claim above to G2 instead of G1 , it follows that there is no G2 -equivariant conditional expectation L∞ (X × H2 ) → L∞ (X) w.r.t. the action induced by ω2 . We can again apply Theorem 4.1 and altogether, we find compact subgroups Ki < Hi , closed e i < Hi with Ki being a normal subgroup of H e i , and continuous group homomorsubgroups H e i /Ki with dense image such that, writing δ = δ1 × δ2 , K = K1 × K2 , phisms δi : Gi → H e e e H = H1 × H2 , the cocycle ω is cohomologous (as a measurable cocycle) with a cocycle e satisfying ω ω e : G×X →H e (g, x) ∈ δ(g)K for all g ∈ G, x ∈ X. Since the actions of G × H on L∞ (X × H) induced by ω and ω e are isomorphic, we find an e into L∞ (X × H)G . Using θ, we also find such an H-equivariant embedding of L∞ (H/H) e 1 ) ⊗ 1 ⊂ L∞ (H/H) e embedding into L∞ (Y × G)G = L∞ (Y ). Since the elements of L∞ (H1 /H e are H2 -invariant, the irreducibility of the action H y (Y, η) implies that H1 = H1 . We similarly e 2 = H2 . Since we assumed that the groups Hi have no nontrivial compact normal find that H subgroups, we also conclude that K is trivial. We have proved that ω is cohomologous, as a measurable cocycle, with the continuous group homomorphism δ : G → H having dense image. We now prove that δ is bijective. Consider the unitary representation Π of G on L2 (X × H) given by (Π(g)ξ)(x, h) = ξ(g−1 · x, δ(g)−1 h). Combining the map θ and the fact that the cocycle ω is cohomologous with δ, the representation Π is unitarily conjugate to the representation of G on L2 (Y ×G) given by (g·ξ)(y, g′ ) = ξ(y, g′ g). Therefore, Π is a multiple of the regular representation. In particular, Π has C0 -coefficients. So, g 7→ hΠ(g)(1 ⊗ 1D ), 1 ⊗ 1E i = λH (δ(g)D ∩ E) 25 is a C0 -function on G for all compact subsets D, E ⊂ H. It follows that Ker δ is a compact subgroup of G and that δ is proper in the sense that δ(gn ) → ∞ whenever gn → ∞. By assumption, G has no nontrivial compact normal subgroups. So, Ker δ = {e} and δ is injective. Since δ is proper, the image of δ is closed. Since δ has dense image, we conclude that δ is surjective. So we have proved that δ is bijective. Since the Borel cocycles ω and δ are cohomologous as measurable cocycles, they are also cohomologous as Borel cocycles on a conull G-invariant Borel set X0 ⊂ X. Since θ(X0 × H) is a conull G × H-invariant Borel subset of Y × G, it must be of the form Y0 × H. So we can restrict everything to X0 and Y0 , and assume that X0 = X and Y0 = Y . Choose a Borel map γ : X → H such that ω(g, x) = γ(g · x) δ(g) γ(x)−1 for all g ∈ G, x ∈ X. Define the measure preserving Borel bijections θ1 , θ2 : X × H → X × H given by θ1 (x, h) = (x, γ(x)h) and θ2 (x, h) = (δ−1 (h−1 ) · x, h−1 ) . Write θe = θ ◦ θ1 ◦ θ2 . Consider the measure preserving Borel action of G × H on X × H given by (g, h) · (x, h′ ) = (δ−1 (h) · x , h h′ δ(g)−1 ) . Still using the action of G × H on Y × G defined in (4.5), we get that θe is G × H-equivariant. e e) = (∆(x), γe(x)) for all Define the Borel functions ∆ : X → Y and e γ : X → G such that θ(x, x ∈ X. Then,  e h) = θe (e, δ−1 (h−1 )) · (x, e) = (e, δ−1 (h−1 )) · θ(x, e e) = (∆(x), γe(x) δ−1 (h)) θ(x, (4.6) for all x ∈ X and h ∈ H. Since θe is bijective and δ is bijective, also ∆ : X → Y must be bijective. Since θe is H-equivariant, we have ∆(g · x) = δ(g) · ∆(x) for all x ∈ X. Since θe is measure preserving and δ is measure scaling, by (4.6), ∆ must be measure scaling. Since both µ and η are probability measures, it follows that ∆ is measure preserving and thus also that δ is measure preserving. By construction, ∆(x) ∈ H · ρ(x) for all x ∈ X and thus ∆(x) ∈ H · π(x) for all x ∈ X1 . This concludes the proof of the theorem. Remark 4.3. Assume that we are in the situation of Theorem B. Given the more precise description in Theorem 4.2 of the conjugacy between the actions G y (X, µ), H y (Y, η) and its relation to the initial stable orbit equivalence, it follows that up to unitary conjugacy, any ∗-isomorphism π : p(L∞ (X) ⋊ G)p → q(L∞ (Y ) ⋊ H)q is the restriction of a ∗-isomorphism of the form π e : L∞ (X) ⋊ G → L∞ (Y ) ⋊ H : π e(ug F ) = uδ(g) ∆∗ (Ωg F ) for all F ∈ L∞ (X), g ∈ G , where ∆ : X → Y is a pmp isomorphism, δ : G → H is a group isomorphism, ∆(g · x) = δ(g)·∆(x) for all g ∈ G and a.e. x ∈ X, and Ωg ∈ U (L∞ (X)) is a scalar cocycle, i.e. Ωg = Ω(g, ·) where Ω : G × X → T is a Borel map satisfying Ω(gh, x) = Ω(g, h · x) Ω(h, x) for all g, h ∈ G and a.e. x ∈ X. 5 Proof of Theorem G Roughly speaking, Theorem G follows by appropriately combining the proof of [BHV15, Proposition 3.6] with the setup and methods in the proof of Theorem F in Section 2. We start by making a first simplification. We replace M by B(ℓ2 (N)) ⊗ B(ℓ2 (N)) ⊗ M and define the projection e = 1 ⊗ 1 ⊗ p in M . We then replace A by ℓ∞ (N) ⊗ B(ℓ2 (N)) ⊗ A and view it as a von Neumann subalgebra of eM e. We finally replace p by the finite trace projection 26 e00 ⊗ e00 ⊗ p and the coaction Φ by id ⊗ Φ. We are now in the following situation: M is a von Neumann algebra with a faithful normal semifinite trace Tr, e ∈ M is a projection, A ⊂ eM e is a von Neumann subalgebra with Tr |A being semifinite and p ∈ A is a projection of finite s trace. Moreover, by [BHV15, Lemma 3.5], for every x ∈ NpM p (pAp), there exist u ∈ NeM e (A) s and a, b ∈ pAp such that ua = x = bu. In particular, NpM p (pAp)′′ = pNeM e (A)′′ p. Also, for s ∗ ∗ every partial isometry v ∈ NpM p (pAp) with v v = s and vv = t, there exists a u ∈ NeM e (A) such that us = v = tu. Assuming that pAp is Φ-amenable, we have to prove that pAp can be Φ-embedded or that s ′′ ′′ NpM p (pAp) = pNeM e (A) p is Φ-amenable. Write q = Φ(p) and f = Φ(e). Step 1. If u ∈ NeM e (A), then p(A ∪ {u})′′ p is still Φ-amenable. Since pAp is Φ-amenable, there exists a positive functional Ω on q(M ⊗ B(L2 (G)))q that is Φ(pAp)-central and that satisfies Ω(Φ(x)) = Tr(x) for all x ∈ pM p. Denote by E : eM e → A the unique Tr-preserving normal conditional expectation. The functional Ω gives a conditional expectation P : q(M ⊗ B(L2 (G)))q → Φ(pAp) satisfying P (Φ(x)) = Φ(E(x)) for all x ∈ pM p. We have a canonical isomorphism f (M ⊗ B(L2 (G)))f ∼ = B(ℓ2 (N)) ⊗ B(ℓ2 (N)) ⊗ q(M ⊗ B(L2 (G)))q sending Φ(A) onto ℓ∞ (N)⊗B(ℓ2 (N))⊗Φ(pAp). Denoting by E0 : B(ℓ2 (N)) → ℓ∞ (N) the normal conditional expectation, taking E0 ⊗ id ⊗ P , we can extend P to a conditional expectation P : f (M ⊗ B(L2 (G)))f → Φ(A) satisfying P (Φ(x)) = Φ(E(x)) for all x ∈ eM e. For every n ≥ 1, define n Pn : f (M ⊗ B(L2 (G)))f → Φ(A) : Pn (T ) = 1X Φ(uk ) P (Φ(u−k )T Φ(uk )) Φ(u−k ) . n k=1 Note that every Pn is a conditional expectation satisfying Pn (Φ(x)) = Φ(E(x)) for all x ∈ eM e. Define P0 : f (M ⊗ B(L2 (G)))f → Φ(A) as a point-weak∗ limit point of the sequence (Pn )n≥1 . Then P0 is a conditional expectation satisfying P0 (Φ(x)) = Φ(E(x)) for all x ∈ eM e and P0 (Φ(uk )T Φ(u−k )) = Φ(uk )P0 (T )Φ(u−k ) for all T ∈ f (M ⊗ B(L2 (G)))f , k ∈ Z . Define the positive functional Ω0 on q(M ⊗ B(L2 (G)))q given by Ω0 (T ) = Tr(Φ−1 (P0 (T ))), which is well defined because P0 (T ) ∈ Φ(pAp). By construction, Ω0 is Φ(pAp)-central and Ω0 (Φ(x)) = Tr(x) for all x ∈ pM p. Let k ∈ Z and a ∈ A. Put x0 = puk ap. Note that x0 = uk bp where b ∈ A is defined as b = u−k puk a. Using the notation P ′ = Φ−1 ◦ P0 , we have for every T ∈ q(M ⊗ B(L2 (G)))q that Ω0 (Φ(x0 )T ) = Ω0 (Φ(uk bp)T ) = Tr(P ′ (Φ(uk )Φ(bp)T )) = Tr(uk bpP ′ (T Φ(uk ))u−k ) = Tr(P ′ (T Φ(uk ))bp) = Tr(P ′ (T Φ(uk bp))) = Ω0 (T Φ(x0 )) . Since Ω0 (Φ(x)) = Tr(x) for all x ∈ pM p and since the linear span of {puk ap | k ∈ Z, a ∈ A} is strongly dense in p(A ∪ {u})′′ p, the Cauchy-Schwarz inequality implies that Ω0 is p(A ∪ {u})′′ pcentral. This concludes the proof of step 1. 27 Notations. Since G has CMAP, we can fix a net ηn ∈ A(G) such that the normal completely bounded maps mn : L(G) → L(G) given by mn = (id ⊗ ηn ) ◦ ∆ satisfy the properties in (2.1) with Λ(G) = 1. Define ϕn : M → M given by ϕn = (id ⊗ ηn ) ◦ Φ. Since Φ ◦ ϕn = (id ⊗ mn ) ◦ Φ, also kϕn kcb ≤ 1 for all n and ϕn (x) → x strongly for all x ∈ M . Finally, we denote ψn : pM p → pM p : ψn (x) = pϕn (x)p. Whenever Q ⊂ eM e is a von Neumann subalgebra, we denote by NQ the von Neumann subalgebra of B(L2 (M e)) ⊗ L(G) generated by Φ(M ) and ρ(Q), where ρ(a) is given by right multiplication with a ∈ Q. We write N = NA . s ∗ ∗ For every partial isometry v ∈ NpM p (pAp) with s = v v and t = vv , denote by βv : ρ(s)N ρ(s) → ρ(t)N ρ(t) the ∗-isomorphism implemented by right multiplication with v ∗ on L2 (M e) ⊗ L2 (G). Note that βv (Φ(x)ρ(a)) = Φ(x)ρ(vav ∗ ). Define q1 = Φ(p)ρ(p). We still denote by βv the normal, completely contractive map q1 N q1 → q1 N q1 given by T 7→ βv (ρ(s)T ρ(s)). Step 2. Let Q ⊂ eM e be a von Neumann subalgebra such that A ⊂ Q and such that pQp is Φ-amenable. Then there exists a net of functionals µQ n ∈ (q1 NQ q1 )∗ with the following properties. 1. µQ n (Φ(x)ρ(a)) = Tr(ψn (x)a) for all x ∈ pM p and a ∈ pQp. 2. kµQ n k ≤ Tr(p) for all n. To prove step 2, in the same way as in step 1 of the proof of Theorem F, using the Φ-amenability of pQp, we find normal completely contractive maps θn : q1 N q1 → B(L2 (pM p)) satisfying θn (Φ(x)ρ(a)) = ψn (x)ρ(a) for all x ∈ pM p and a ∈ pQp. Composing with the vector functional T 7→ hT p, pi, which has norm Tr(p), the proof of step 1 is complete. Step 3. The positive functionals ωn = |µA n | in (q1 N q1 )∗ satisfy 1. limn ωn (Φ(x)) = Tr(x) for all x ∈ pM p, 2. limn ωn (Φ(a)ρ(a∗ )) = Tr(p) for all a ∈ U (pAp), s 3. for every partial isometry v ∈ NpM p (pAp), we have that limn kωn ◦ βv∗ − ωn ◦ Ad Φ(v)k = 0. Note that, as defined above, the functional ωn ◦βv∗ on q1 N q1 is given by (ωn ◦βv∗ )(Φ(x)ρ(a)) = ωn (Φ(x)ρ(v ∗ av)) for all x ∈ pM p and a ∈ pAp, while the functional ωn ◦ Ad Φ(v) on q1 N q1 is given by (ωn ◦ Ad Φ(v))(Φ(x)ρ(a)) = ωn (Φ(vxv ∗ )ρ(a)). To prove step 3, let Q ⊂ eM e be a von Neumann subalgebra such that A ⊂ Q and such that Q Q Q pQp is Φ-amenable. Define µQ n as in step 2 and denote ωn = |µn |. Since kµn k ≤ Tr(p) for all Q Q Q n and limn µn (q1 ) = Tr(p), we find that limn kµn − ωn k = 0. Whenever a ∈ U (pQp), we get Q ∗ ∗ that limn µQ n (Φ(a)ρ(a )) = Tr(p) and thus also, limn ωn (Φ(a)ρ(a )) = Tr(p). So the first two properties in step 3 are already proven. It also follows that lim k(Φ(a)ρ(a∗ )) · ωnQ − ωnQ k = 0 = lim kωnQ · (Φ(a)ρ(a∗ )) − ωnQ k n n for all a ∈ U (pQp) and thus, lim kΦ(a) · ωnQ − ρ(a) · ωnQ k = 0 = lim kωnQ · Φ(a) − ωnQ · ρ(a)k n n for all a ∈ pQp. 28 (5.1) s ∗ To prove the third property in step 3, fix a partial isometry v ∈ NpM p (pAp) and write s = v v. ′′ Take u ∈ NeM e (A) such that v = us. Define Q = (A ∪ {u}) . By step 1 of the proof, pQp is Φ-amenable. By construction v ∈ pQp. By (5.1), lim kΦ(v ∗ ) · ωnQ · Φ(v) − ρ(v ∗ ) · ωnQ · ρ(v)k = 0 . n Restricting these positive functionals to q1 N q1 , we find the third property in step 3. Notations. Choose a standard Hilbert space H for the von Neumann algebra N , which comes with the normal ∗-homomorphism πl : N → B(H), the normal ∗-antihomomorphism πr : N → B(H) and the positive cone H+ ⊂ H. For every u ∈ NeM e (A), define the automorphism βu of N implemented by right multiplication with u∗ on L2 (M e) ⊗ L2 (G) and denote by Wu ∈ U (H) its canonical implementation. Denote by EZ : pAp → Z(A)p the unique trace preserving conditional expectation (i.e. the center valued trace of pAp). For every projection s ∈ pAp, denote by zs ∈ Z(A)p its central support, which equals the support projection of EZ (s). Denote by P0 ⊂ P(pAp) the set of projections s ∈ pAp for which there exists a δ > 0 such that EZ (s) ≥ δzs . We then denote Ds = (EZ (s))1/2 and we denote by Ds−1 the (bounded) inverse of Ds in Z(A)zs . As in [BHV15, Section 3] and using [BHV15, Lemma 3.9], we can choose a sequence ai ∈ pAp such that ∞ X ai a∗i = Ds zs and ∞ X a∗i ai = Ds−1 s . i=0 i=0 We make once and for all a choice of ai for each s ∈ P0 . We also define T (s) = ∞ X Φ(ai )ρ(a∗i ) ∈ q1 N q1 . i=0 Note that the series defining T (s) is strongly convergent, so that T (s) is a well defined element of q1 N q1 . s ∗ ∗ For every partial isometry v ∈ NpM p (pAp) with s = v v and t = vv , we denote by Wv : πl (ρ(s))πr (ρ(s))H → πl (ρ(t))πr (ρ(t))H the canonical unitary implementation of the ∗-isomorphism βv : ρ(s)N ρ(s) → ρ(t)N ρ(t). Step 4. The canonical implementation ξn ∈ πl (q1 )πr (q1 )H of ωn satisfies the following properties. 1. limn hπl (Φ(x))ξn , ξn i = Tr(pxp) = limn hπr (Φ(x))ξn , ξn i for all x ∈ M , 2. limn kπl (Φ(a))ξn − πr (ρ(a))ξn k = 0 for all a ∈ U (pAp), s ∗ ∗ 3. Whenever v ∈ NpM p (pAp) is a partial isometry such that s = v v and t = vv belong to P0 , we have lim kπl (Φ(v))ξn − πr (T (s)∗ ) πr (Φ(v)) Wv∗ πr (T (t)) ξn k = 0 . (5.2) n The first two properties follow immediately from the first two properties of ωn in step 3. To also deduce the third property from step 3, one can literally apply the proof of [BHV15, Proposition 3.6]. Notations and formulation of the dichotomy. As in the proof of Theorem F, the coaction Ψ : N → N ⊗ L(G) given by Ψ = id ⊗ ∆ has a canonical implementation on the standard Hilbert space H given by a nondegenerate ∗-homomorphism π : C0 (G) → B(H). We again distinguish two cases. 29 • Case 1. For every F ∈ C0 (G), we have that lim supn kπ(F )ξn k = 0. • Case 2. There exists an F ∈ C0 (G) with lim supn kπ(F )ξn k > 0. s ′′ We prove that in case 1, the von Neumann subalgebra NpM p (pAp) ⊂ pM p is Φ-amenable and that in case 2, the von Neumann subalgebra pAp ⊂ pM p can be Φ-embedded. The proof in case 2 is identical to the proof of case 2 in Theorem F, because that part of the proof only relies on the first two properties of the net ξn in step 4. So from now on, assume that we are in case 1. Choose a positive functional Ω on B(H) as a weak∗ limit point of the net of vector functionals T 7→ hT ξn , ξn i. Denote G = NeM e (A). The group G acts on N by the automorphisms βu , u ∈ G. We also consider the diagonal action of G on N ⊗alg N op and denote by D the algebraic crossed product D = (N ⊗alg N op ) ⋊alg G. As a vector space, D = N ⊗alg N op ⊗alg CG and the product and ∗-operation on D are given by (x1 ⊗ y1op ⊗ u1 ) (x2 ⊗ y2op ⊗ u2 ) = x1 βu1 (x2 ) ⊗ (βu1 (y2 )y1 )op ⊗ u1 u2 (x ⊗ y op ∗ ∗ ∗ op ⊗ u) = βu∗ (x ) ⊗ (βu∗ (y )) and ∗ ⊗u . We define the ∗-representations Θ : D → B(H) : Θ(x ⊗ y op ⊗ u) = πl (x) πr (y) Wu , Θ1 : D → B(H) ⊗ L(G) : Θ1 (x ⊗ y op ⊗ u) = (πl ⊗ id)Ψ(x) (πr (y) Wu ⊗ 1) . Define the ∗-subalgebras Ni of N given by N1 = span Φ(M )ρ(A) , N2 = [N1 ] , n N3 = x ∈ N there exist sequences xi ∈ M and ai ∈ A such that X X X X all x∗i xi , xi x∗i , a∗i ai and ai a∗i are bounded i and x = X i i Φ(xi )ρ(ai ) i o i . Each Ni is globally invariant under the automorphisms βu , u ∈ G, and so we have the ∗subalgebras Di ⊂ D defined as Di = (Ni ⊗alg Niop ) ⋊alg G. Note that N1 ⊂ N3 , but that the inclusion N2 ⊂ N3 need not hold. Denote C = Tr(p) = kΩk. We claim that |Ω(Θ(x))| ≤ C kΘ1 (x)k for all x ∈ D3 . (5.3) To prove (5.3), first note that in exactly the same way as we proved (2.15), we get that (5.3) holds for all x ∈ D1 and thus also for all x ∈ D2 by norm continuity. P Whenever xi ∈ M and ai ∈ A are sequences as in the definition of N3 and x = i Φ(xi )ρ(ai ), we can choose a sequence of projections pn ∈ pM p such that pn → p strongly and such that P ∗ for each fixed n, the series pn i xi xi pn is norm convergent. This means that for each n, we have that Φ(pn )x ∈ N2 . Fix x ∈ D3 . Since the automorphisms βu act as the identity on Φ(M ), it follows that we can find a sequence of projections pn ∈ pM p such that pn → p strongly and such that xn = (Φ(pn ) ⊗ 1 ⊗ 1) x (1 ⊗ Φ(pn )op ⊗ 1) ∈ D2 30 for all n. Since Ω(πl (Φ(x))) = Tr(pxp) for all x ∈ M , we get that Ω = Ω · πl (p) and lim kΩ − Ω · πl (Φ(pn )k = 0 . n A similar result holds for πr and thus, lim kΩ − πr (Φ(pn )) · Ω · πl (Φ(pn ))k = 0 . n This implies that Ω(Θ(x)) = limn Ω(Θ(xn )). Since xn ∈ D2 , we know that |Ω(Θ(xn ))| ≤ C kΘ1 (xn )k ≤ C kΘ1 (x)k for all n. So, (5.3) follows. By (5.3), we can define the continuous functional Ω1 on the C∗ -algebra [Θ1 (D3 )] satisfying Ω1 (Θ1 (x)) = Ω(Θ(x)) for all x ∈ D3 . It follows that Ω1 (Θ1 (x)∗ Θ1 (x)) ≥ 0 for all x ∈ D3 and thus, Ω1 is positive. Extend Ω1 to a bounded functional on B(H ⊗ L2 (G)) without increasing its norm. In particular, Ω1 remains a positive functional. s ∗ ∗ Let v ∈ NpM p (pAp) be a partial isometry such that s = v v and t = vv belong to P0 . Using the same notation as in step 4, we define the operator Y (v) ∈ B(H) given by Y (v) = πr (T (s)∗ ) πr (Φ(v)) Wv∗ πr (T (t)) . Note that Y (v) commutes with πl (Φ(M )). Also note that Y (v)∗ = Y (v ∗ ). Take u ∈ NeM e (A) such that v = us. Since Wv = Wu πl (ρ(s)) πr (ρ(s)), we define the element y(v) ∈ D3 given by y(v) = (ρ(s) ⊗ (ρ(s)Φ(v)T (s)∗ )op ⊗ 1) (1 ⊗ 1 ⊗ u∗ ) (1 ⊗ T (t)op ⊗ 1) and note that Θ(y(v)) = Y (v) and Θ1 (y(v)) = Y (v) ⊗ 1. p p For every T ∈ B(H), write kT kΩ = Ω(T ∗ T ). Similarly define kT kΩ1 = Ω1 (T ∗ T ) for all T ∈ B(H ⊗ L2 (G)). Applying (5.2) for v and v ∗ , and using that Y (v ∗ ) = Y (v)∗ , we find that kΘ(Φ(v) ⊗ 1 ⊗ 1 − y(v))kΩ = kπl (Φ(v)) − Y (v)kΩ = 0 and kΘ(Φ(v ∗ ) ⊗ 1 ⊗ 1 − y(v)∗ )kΩ = kπl (Φ(v ∗ )) − Y (v)∗ kΩ = 0 . Then also k(πl ◦ Φ ⊗ id)(Φ(v)) − Y (v) ⊗ 1kΩ1 = kΘ1 (Φ(v) ⊗ 1 ⊗ 1 − y(v))kΩ1 = kΘ(Φ(v) ⊗ 1 ⊗ 1 − y(v))kΩ = 0 , ∗ ∗ k(πl ◦ Φ ⊗ id)(Φ(v )) − Y (v) ⊗ 1kΩ1 = kΘ1 (Φ(v ∗ ) ⊗ 1 ⊗ 1 − y(v)∗ )kΩ1 (5.4) = kΘ(Φ(v ∗ ) ⊗ 1 ⊗ 1 − y(v)∗ )kΩ = 0 . Define the positive functional Ω2 on q(M ⊗ B(L2 (G)))q given by Ω2 = Ω1 ◦ (πl ◦ Φ ⊗ id). Since Y (v) ⊗ 1 commutes with πl (Φ(M )) ⊗ B(L2 (G)), it follows from (5.4) that Ω2 (Φ(v)T ) = s ∗ ∗ Ω2 (T Φ(v)) for every partial isometry v ∈ NpM p (pAp) with v v and vv belonging to P0 . We also have that Ω2 (Φ(x)) = Tr(x) for all x ∈ pM p. Since the linear span of all such partial s ′′ isometries v is k · k2 -dense in P = NpM p (pAp) , it follows that Ω2 is Φ(P )-central. So we have proved that P is Φ-amenable. This concludes the proof of Theorem G. 31 6 Stable strong solidity; proof of Theorem C The following is an immediate consequence of Ozawa’s solidity theorem [Oz03]. Proposition 6.1. Let G be a locally compact group with property (S). Assume that Cr∗ (G) is an exact C∗ -algebra. Then M = L(G) is solid in the sense that for every diffuse von Neumann subalgebra A ⊂ M that is the range of a normal conditional expectation, A′ ∩ M is injective. Proof. By the type III version of Ozawa’s theorem [Oz03], as proved in [VV05, Theorem 2.5], it suffices to prove the Akemann-Ostrand property, meaning that the ∗-homomorphism θ : Cr∗ (G) ⊗alg Cr∗ (G) → B(L2 (G)) : θ(a ⊗ b) = λ(a)ρ(b) + K(L2 (G)) K(L2 (G)) is continuous on the spatial tensor product Cr∗ (G) ⊗min Cr∗ (G). As in the proof of (2.7), property (S) gives rise to an isometry Z0 that is an adjointable operator from C0 (G) to C0 (G) ⊗min L2 (G) with the property that Z0 λ(a) − (λ ⊗ λ)∆(a)Z0 and Z0 ρ(b) − (ρ(b) ⊗ 1)Z0 belong to K(L2 (G)) ⊗min L2 (G) for all a, b ∈ Cr∗ (G). The standard representation of Cr∗ (G) ⊗min Cr∗ (G) on L2 (G) ⊗ L2 (G) is unitarily conjugate to the representation θ1 : Cr∗ (G) ⊗min Cr∗ (G) → B(L2 (G) ⊗ L2 (G)) : θ1 (a ⊗ b) = (λ ⊗ λ)∆(a) (ρ(b) ⊗ 1) . Since θ(a⊗ b) = Z0∗ θ1 (a⊗ b)Z0 + K(L2 (G)) for all a, b ∈ Cr∗ (G), the Akemann-Ostrand property indeed holds. For the proof of Theorem C, we need the following lemma. Lemma 6.2. Let M be a diffuse σ-finite von Neumann algebra and pn ∈ M a sequence of projections such that pn → 1 strongly. Then M is stably strongly solid if and only if pn M pn is stably strongly solid for every n. Proof. Write H = ℓ2 (N) and denote by zn ∈ Z(M ) the central support of pn . Since M is σ-finite, we have B(H) ⊗ pn M pn ∼ = B(H) ⊗ M zn . By [BHV15, Corollary 5.2], we get that pn M pn is stably strongly solid if and only if M zn is stably strongly solid. Since every diffuse von Neumann algebra admits a diffuse amenable (even abelian) von Neumann subalgebra with expectation, it is easy to check that M is stably strongly solid if and only if M zn is stably strongly solid for each n. Proof of Theorem C. First assume that G is unimodular. Fix a Haar measure on G and denote by Tr the associated faithful normal semifinite trace on M = L(G). Fix a projection p ∈ L(G) with Tr(p) < ∞. Assume that G is weakly amenable and has property (S). We have to prove that pM p is strongly solid. So fix a diffuse amenable von Neumann subalgebra A ⊂ pM p. We have to prove that NpM p (A)′′ is amenable. Denote by ∆ : L(G) → L(G) ⊗ L(G) : ∆(λg ) = λg ⊗ λg the comultiplication. View ∆ as a coaction on M , so that we can apply Theorem F. Since A is amenable, we certainly have that A is ∆-amenable. We next prove that A cannot be ∆-embedded. Fix a net an ∈ U (A) such that an → 0 weakly. For every ξ, η ∈ L2 (G), denote by ωξ,η ∈ L(G)∗ the vector functional given by ωξ,η (λg ) = hλg ξ, ηi. Also denote by mξ,η : L(G) → L(G) : mξ,η = (id ⊗ ωξ,η ) ◦ ∆ the associated normal completely bounded map. We claim that mξ,η (an ) → 0 strongly, for all ξ, η ∈ L2 (G). 32 (6.1) Fix ξ, η ∈ L2 (G) and fix µ ∈ L2 (G). To prove (6.1), we must prove that kmξ,η (an )µk → 0. Denote by V ∈ L∞ (G) ⊗ L(G) the unitary given by V (g) = λg for all g ∈ G. For every a ∈ L(G), we have mξ,η (a)µ = (id ⊗ ωξ,η )(∆(a))µ = (ωξ,η ⊗ id)(∆(a))µ = (ωξ,η ⊗ id)(V (a ⊗ 1)V ∗ )µ = (η ∗ ⊗ 1)V (a ⊗ 1)V ∗ (ξ ⊗ µ) . Approximating V ∗ (ξ ⊗ µ) ∈ L2 (G) ⊗ L2 (G) by linear combinations of vectors ξ0 ⊗ µ0 , it suffices to prove that lim k(η ∗ ⊗ 1)V (an ⊗ 1)(ξ0 ⊗ µ0 )k = 0 for all η, ξ0 , µ0 ∈ L2 (G). n Since the operator (η ∗ ⊗ 1)V (1 ⊗ µ0 ) belongs to K(L2 (G)), this last statement indeed holds and the claim in (6.1) is proved. For all µ1 , µ2 ∈ L2 (M p) and for all ξ, η ∈ L2 (G), we have h∆(an ) (µ1 ⊗ ξ) a∗n , µ2 ⊗ ηi = hmξ,η (an )µ1 , µ2 an i . So, (6.1) implies that limh∆(an ) · ξ · a∗n , ξi = 0 for all ξ ∈ L2 (M p) ⊗ L2 (G). n So there is no nonzero vector ξ ∈ L2 (M p) ⊗ L2 (G) satisfying ∆(a)ξ = ξa for all a ∈ A, meaning that A cannot be ∆-embedded. Write P = NpM p (A)′′ . By Theorem F, P is ∆-amenable. Since the P -M -bimodule ∆(P )∆(p)(L 2 (M ) ⊗ L2 (G))M is contained in a multiple of the coarse P -M -bimodule, it follows that P is amenable. So we have proved that pM p is strongly solid. Next assume that G is a locally compact second countable group with CMAP and property (S) such that the kernel G0 of the modular function δ : G → R+ is an open subgroup of G. Fix a left Haar measure on G and denote by ϕ the associated faithful normal semifinite weight on M = L(G). Denote by σ ϕ its modular automorphism group, given by σtϕ (λg ) = δ(g)it λg for all g ∈ G, t ∈ R. So, L(G0 ) lies in the centralizer L(G)ϕ and since G0 ⊂ G is an open subgroup, the restriction of ϕ to L(G0 ) is semifinite. By Lemma 6.2, it is sufficient to prove that pL(G)p is stably strongly solid for each nonzero projection p ∈ L(G0 ) with ϕ(p) < ∞. Fix such a projection p and let A ⊂ pM p be a diffuse amenable von Neumann subalgebra with expectation. Write s ′′ P = NpM p (A) . We have to prove that P is amenable. Denote H = ℓ2 (N) and define M1 = B(H) ⊗ M . Write A0 = B(H) ⊗ A and p1 = 1 ⊗ p. By [BHV15, Lemma 3.4], we have to prove that Np1 M1 p1 (A0 )′′ is amenable. Since G has CMAP, certainly G is exact (see e.g. [BCL16, Corollary E]) and Proposition 6.1 implies that A′ ∩ pM p is amenable. So, A1 := A0 ∨ (A′0 ∩ p1 M1 p1 ) is amenable. Since Np1 M1 p1 (A0 )′′ ⊂ Np1 M1 p1 (A1 )′′ and since this is an inclusion with expectation, it suffices to show that P1 := Np1 M1 p1 (A1 )′′ is amenable. Let e ∈ B(H) be a minimal projection and choose a faithful normal state η on B(H) such that e belongs to the centralizer of η. Also choose a faithful normal state ψ on pM p such that 33 σtψ (A) = A for all t ∈ R. Note that η ⊗ ψ is a faithful normal state on p1 M1 p1 . Then A1 and P1 are globally invariant under σ η⊗ψ and we obtain the canonical inclusions of continuous cores cη⊗ψ (A1 ) ⊂ cη⊗ψ (P1 ) ⊂ cη⊗ψ (p1 M1 p1 ) . Since A′1 ∩ p1 M1 p1 = Z(A1 ), it follows from [BHV15, Lemma 4.1] that cη⊗ψ (P1 ) is contained in the normalizer of cη⊗ψ (A1 ). By Takesaki’s duality theorem [Ta03, Theorem X.2.3], P1 is amenable if and only if its continuous core is amenable. So, it suffices to prove that the normalizer of cη⊗ψ (A1 ) is amenable. Now we can cut down again with the projection e ⊗ 1 and conclude that it is sufficient to prove the following result: for any diffuse amenable B ⊂ pM p with expectation and for every faithful normal state ψ on pM p with σtψ (B) = B for all t ∈ R, the canonical subalgebra cψ (B) of cψ (pM p) has an amenable stable normalizer. Whenever p′ ∈ M ϕ is a projection with p′ ≥ p and ϕ(p′ ) < ∞, we can realize the continuous core of p′ M p′ as πϕ (p′ )Mπϕ (p′ ) where M = cϕ (M ). Let Π : cψ (pM p) → πϕ (p)Mπϕ (p) be the canonical trace preserving isomorphism. Let pn ∈ Lψ (R) be a sequence of projections having finite trace and converging to 1 strongly. Since B is diffuse and using Popa’s intertwining-by-bimodules [Po03, Section 2], it follows from [HU15, Lemma 2.5] that 6≺ Π(pn cψ (B)pn ) Lϕ (R)πϕ (p′ ) (6.2) πϕ (p′ )Mπϕ (p′ ) for all n and all projections p′ ∈ M ϕ with p′ ≥ p and ϕ(p′ ) < ∞. Denote by P the set of these projections p′ and define the ∗-algebra [ M0 := πϕ (p′ )Mπϕ (p′ ) . p′ ∈P There is a unique linear map E : M0 → Lϕ (R) such that for every p′ ∈ P, the restriction of E to πϕ (p′ )Mπϕ (p′ ) is normal and given by E(πϕ (x)λϕ (t)) = ϕ(x)λϕ (t) for all x ∈ p′ M p′ and t ∈ R. Note that this restriction of E can be viewed as ϕ(p′ ) times the unique trace preserving conditional expectation of πϕ (p′ )Mπϕ (p′ ) onto Lϕ (R)p′ . Combining (6.2) and [HI15, Theorem 4.3], in order to prove that cψ (B) has an amenable stable normalizer inside cψ (pM p), it is sufficient to prove the following statement: whenever q ∈ πϕ (p)Mπϕ (p) is a projection of finite trace and A ⊂ qMq is a von Neumann subalgebra that admits a net of unitaries an ∈ U (A) satisfying E(x∗ an y) → 0 strongly, for all x, y ∈ M0 , (6.3) then the stable normalizer of A inside qMq is amenable. Fix such a von Neumann subalgebra A ⊂ qMq and fix a net of unitaries an ∈ U (A) satisfying (6.3). Since ∆ ◦ σtϕ = (σtϕ ⊗ id) ◦ ∆ for all t ∈ R, there is a well defined coaction given by Φ : M → M ⊗ L(G) : Φ(πϕ (x)λϕ (t)) = (πϕ ⊗ id)(∆(x)) (λϕ (t) ⊗ 1) for all x ∈ M , t ∈ R. The M-bimodule Φ(M)(L2 (M) ⊗ L2 (G))M ⊗ 1 is isomorphic with L2 (M) ⊗Lϕ (R) L2 (M) and thus weakly contained in the coarse M-bimodule. Using Theorem G, it only remains to prove that (6.3) implies that A cannot be Φ-embedded. 34 We deduce that A cannot be Φ-embedded from the following approximation result: for all a ∈ Cr∗ (G0 ), ω ∈ L(G)+ ∗ and ε > 0, there exist n ∈ N, elements aj , xj ∈ L(G) and scalars δj > 0 for j ∈ {1, . . . , n} such that σtϕ (aj ) = δjit aj and σtϕ (xj ) = δj−it xj (6.4) for all j ∈ {1, . . . , n} and all t ∈ R, and such that the map Ψ : M → M : Ψ(x) = n X j,k=1  πϕ (x∗j ) E πϕ (paj )∗ xπϕ (pak ) πϕ (xk ) (6.5) is normal and completely bounded, and satisfies  (id ⊗ ω)Φ πϕ (pa)∗ · πϕ (pa) − Ψ cb <ε. (6.6) Already note that (6.4) implies that the right support of paj belongs to P, so that πϕ (paj ) ∈ M0 and the map Ψ is well defined and normal. Assuming that such an approximation exists, we already deduce that A cannot be Φ-embedded. It suffices to prove that limhΦ(an )(µ ⊗ ξ), µan ⊗ ξi = 0 for all µ ∈ L2 (M) and ξ ∈ L2 (G), n because then also limn hΦ(an )η(a∗n ⊗ 1), ηi = 0 for all η ∈ L2 (Mq) ⊗ L2 (G), excluding the existence of a nonzero vector η ∈ L2 (Mq) ⊗ L2 (G) satisfying Φ(a)η = η(a ⊗ 1) for all a ∈ A. Since µ ⊗ ξ can be approximated by vectors of the form Φ(πϕ (a))(µ ⊗ ξ) with a ∈ Cr∗ (G0 ), it suffices to prove that limhΦ(πϕ (a)∗ an πϕ (a))(µ ⊗ ξ), µan ⊗ ξi = 0 for all a ∈ Cr∗ (G0 ), µ ∈ L2 (M) and ξ ∈ L2 (G). n Denoting by ω ∈ L(G)+ ∗ the vector functional implemented by ξ, it is sufficient to prove that (id ⊗ ω)(Φ(πϕ (a)∗ an πϕ (a))) → 0 strongly, for all a ∈ Cr∗ (G0 ), ω ∈ L(G)+ ∗. But this follows by the approximation in (6.6) and because (6.3) implies that Ψ(an ) → 0 strongly. Fixing a ∈ Cr∗ (G0 ), ω ∈ L(G)+ ∗ and ε > 0, it remains to find the approximation (6.6). First take ξ ∈ Cc (G) such that the vector functional ωξ satisfies kω − ωξ k < (1/3) ε kak−2 . It follows that, as maps on M = L(G),   (id ⊗ ω)∆ a∗ p · pa − (id ⊗ ωξ )∆ a∗ p · pa cb < ε . 3 (6.7) Fix F ∈ Cc (G) with 0 ≤ F ≤ 1 and ξ = F ξ. For every x ∈ M and using the unitary V ∈ L∞ (G) ⊗ L(G) as in the first part of the proof, we have   (id ⊗ ωξ )∆ a∗ p x pa = (ωξ ⊗ id)∆ a∗ p x pa = (ξ ∗ ⊗ 1)V (a∗ pxpa ⊗ 1)V ∗ (ξ ⊗ 1) = (ξ ∗ ⊗ 1)V (F a∗ pxp aF ⊗ 1)V ∗ (ξ ⊗ 1) . Since a ∈ Cr∗ (G0 ) ⊂ Cr∗ (G) and F ∈ Cc (G), we get that aF is a compact operator on L2 (G) that commutes with the modular function (viewed as a multiplication operator). 35 So we can approximate aF by a finite rank operator T of the form T = n X µj ξj∗ j=1 where ξj , µj ∈ Cc (G) and δξj = δj ξj , δµj = δj µj , and such that kT k ≤ kaF k ≤ kak and kaF − T k ≤ ε . 3 kak kξk2 Defining m : M → M : m(x) = (id ⊗ ω)(∆(a∗ pxp a)) , m1 : M → M : m1 (x) = (ξ ∗ ⊗ 1)V (T ∗ pxp T ⊗ 1)V ∗ (ξ ⊗ 1) , we get that km − m1 kcb < ε. Defining aj = we get that Z µj (g) λg dg G m1 (x) = and xj = (ωξ,ξj ⊗ id)(V ∗ ) , n X ϕ(a∗j pxp ak )x∗j xk . j,k=1 Both m and m1 commute with the modular automorphism group σ ϕ and thus canonically extend to M = cϕ (M ) by acting as the identity on Lϕ (R). The canonical extension of m equals  (id ⊗ ω)Φ πϕ (pa)∗ · πϕ (pa) , while the canonical extension of m1 equals the map Ψ given by (6.5). Since km − m1 kcb < ε, also (6.6) holds and the theorem is proved. 7 Locally compact groups with property (S) Recall that a compactly generated locally compact group G is said to be hyperbolic if the Cayley graph of G with respect to a compact generating set K ⊂ G satisfying K = K −1 is Gromov hyperbolic, in the sense that the metric d on G defined by ( 0 if g = h, d(g, h) = −1 n min{n ∈ N | h g ∈ K } if g 6= h, turns G into a Gromov hyperbolic metric space. When G is non discrete, this Cayley graph is not locally finite and often the action of G on its Cayley graph is not continuous. By [CCMT12, Corollary 2.6], a locally compact group G is hyperbolic if and only if G admits a proper, continuous, cocompact, isometric action on a proper geodesic hyperbolic metric space. Combining several results from the literature, we have the following list of locally compact groups that are weakly amenable and have property (S). In the formulation of the proposition, graphs are assumed to be simple, non oriented and connected. We always equip their vertex set with the graph metric. A hyperbolic graph is a simple, non oriented, connected graph such that the underlying metric space is Gromov hyperbolic. 36 Proposition 7.1. Let G be a locally compact group. If one of the following conditions holds, then G has the complete metric approximation property and property (S). 1. G is σ-compact and amenable. 2. ([Ha78, Sz91, Oz03]) G admits a continuous action on a (not necessarily locally finite) tree that is metrically proper in the sense that for every vertex x, we have that d(x, g · x) → ∞ when g tends to infinity in G. If one of the following conditions holds, then G is weakly amenable and has property (S). 3. G is compactly generated and hyperbolic. 4. ([Oz03, Oz07]) G admits a continuous proper action on a hyperbolic graph with uniformly bounded degree. 5. ([CH88, Sk88]) G is a real rank one, connected, simple Lie group with finite center. Proof. 1. Since G is amenable, a fortiori G has CMAP. Since G is σ-compact, we can fix an increasing sequence of compact subsets Kn ⊂ G such that the interiors int(Kn ) cover G. We make this choice such that K0 = ∅, Kn = Kn−1 , Kn ⊂ int(Kn+1 ) and Kn Kn Kn ⊂ Kn+1 for all n. Since G is amenable, we can choose ηn ∈ S(G) such that kg · ηn − ηn k1 ≤ 2−n for all n ≥ 0 and all g ∈ Kn . Choose continuous functions Fn : G → [0, 1] such that Fn (g) = 0 for all g ∈ Kn−1 and Fn (g) = 1 for all g ∈ G \ Kn . By convention, F0 (g) = 1 for all g ∈ G. Define the continuous function µ : G → L1 (G)+ : µ(g) = ∞ X Fn (g)ηn . n=0 Note that n ≤ kµ(g)k1 ≤ n + 1 whenever n ≥ 1 and g ∈ Kn \ Kn−1 . Define η : G → S(G) : η(g) = kµ(g)k−1 1 µ(g). Choose ε > 0 and K ⊂ G compact. Take n0 such that K ⊂ Kn0 . Then take n1 > n0 such that 2(2n0 + 6)/n1 < ε. Fix g, k ∈ K and h ∈ G \ Kn1 . We prove that kη(ghk) − g · η(h)k1 < ε. Once this is proved, it follows that G has property (S). Take n ≥ n1 such that h ∈ Kn+1 \ Kn . Since K −1 Kn−1 K −1 ⊂ Kn and KKn+1 K ⊂ Kn+2 , we have ghk ∈ Kn+2 \ Kn−1 . Define γ ∈ L1 (G) given by n X γ= ηk . k=0 By construction, kµ(h) − γk1 ≤ 1 and kµ(ghk) − γk1 ≤ 3. Also, kg · γ − γk1 ≤ 2n0 + n X kg · ηk − ηk k1 ≤ 2n0 + k=n0 n X 2−k ≤ 2(n0 + 1) . k=n0 Altogether, it follows that kg · µ(h) − µ(ghk)k1 ≤ 2n0 + 6. Since kµ(h)k1 ≥ n1 , it follows that kg · η(h) − η(ghk)k1 < ε. 2. Let G = (V, E) be a tree and G y G a continuous metrically proper action. By [BO08, Corollary 12.3.4], the group G has CMAP. For all x, y ∈ V , denote by A(x, y) ⊂ V the (unique) geodesic between x and y. Fix a base point x0 ∈ V . Define the continuous map η : G → Prob(V ) by defining η(g) as the uniform probability measure on A(x0 , g · x0 ). For all g, h, k ∈ G, the symmetric difference between A(x0 , ghk · x0 ) and g · A(x0 , h · x0 ) contains at 37 most d(x0 , g · x0 ) + d(x0 , k · x0 ) elements. Since the action G y G is metrically proper, we have d(x0 , h · x0 ) → ∞ when h tends to infinity in G. It then follows that lim kη(ghk) − g · η(h)k1 = 0 h→∞ uniformly on compact sets of g, k ∈ G. Since the action G y V has compact open stabilizers, there exists a G-equivariant isometric map Prob(V ) → S(G). Composing η with this map, it follows that G has property (S). 3. By [CCMT12, Corollary 2.6], G admits a proper, continuous, cocompact, isometric action on a proper geodesic hyperbolic metric space. By [MMS03, Theorem 21 and Proposition 8], G satisfies at least one of the following three structural properties: G is amenable, or G admits a proper action on a hyperbolic graph with uniformly bounded degree, or G admits closed subgroups K < G0 < G such that G0 is of finite index and open in G, K is a compact normal subgroup of G0 and G0 /K is a real rank one, connected, simple Lie group with finite center. Since we already proved 1, to complete the proof of 3, it suffices to prove 4 and 5 and apply Lemma 7.2 below. 4. Let G = (V, E) be a hyperbolic graph with uniformly bounded degree. Let G y G be a continuous proper action. By [Oz07, Theorem 1], the group G is weakly amenable. The proof of property (S) is almost identical to the proof of [Ka02, Theorem 1.33] and especially the version in [BO08, Theorem 5.3.15]. For completeness, we provide the details here. We use the following ad hoc terminology. Assume that [x′ , y ′ ] ⊂ V is a geodesic. If d(x′ , y ′ ) is even, we call “mid point of [x′ , y ′ ]” the unique point z ∈ [x′ , y ′ ] with d(x′ , z) = d(z, y ′ ) = d(x′ , y ′ )/2. If d(x′ , y ′ ) is odd, we declare two points of [x′ , y ′ ] to be the “mid points of [x′ , y ′ ]”, namely the two points z ∈ [x′ , y ′ ] with d(x′ , z) = (d(x′ , y ′ ) ± 1)/2 and thus d(z, y ′ ) = (d(x′ , y ′ ) ∓ 1)/2. For all x, y ∈ V and k ∈ N, define the nonempty subset A(x, y, k) ⊂ V given by  A(x, y, k) = z ∈ V there exists a geodesic [x′ , y ′ ] ⊂ V with d(x, x′ ) ≤ k and d(y, y ′ ) ≤ k such that z is one of the mid points of [x′ , y ′ ] . Note that A(x, y, k) = A(y, x, k) and A(g · x, g · y, k) = g · A(x, y, k) for all x, y ∈ V , k ∈ N and g ∈ G. Take δ > 0 such that every geodesic triangle in G is δ-thin (see [BO08, Definition 5.3.3]). Define B = sup |{y ∈ V | d(y, x) ≤ 2δ}| . x∈V Since G has uniformly bounded degree, we have that B < ∞. We claim that for all x, y ∈ V with d(x, y) ≥ 4k, we have |A(x, y, k)| ≤ 2(k + 1)B . (7.1) To prove this claim, fix a geodesic [x, y] between x and y and denote by [a, b] ⊂ [x, y] the unique segment determined by d(x, a) = ⌊d(x, y)/2⌋ − k and d(x, b) = ⌈d(x, y)/2⌉ + k . Note that [a, b] contains at most 2(k + 1) vertices. To prove the claim, it thus suffices to show that every z ∈ A(x, y, k) lies at distance at most 2δ from a vertex on [a, b]. Choose a geodesic [x′ , y ′ ] ⊂ V with d(x, x′ ) ≤ k and d(y, y ′ ) ≤ k. Let z be one of the mid points of [x′ , y ′ ]. Since d(x, y) ≥ 4k, the geodesic picture of the five points x, x′ , y, y ′ , z in a tree 38 would look as the following picture on the left. x x y x0 z c y0 e z y′ x′ y y′ x′ In our comparison tree, some of the “small” segments [x, x0 ], [x′ , x0 ], [y, y0 ], [y ′ , y0 ] could be reduced to a single point, but the “large” segment [x0 , y0 ] has length at least 2k. Therefore, in the comparison tree, the mid point z of [x′ , y ′ ] lies on the segment [x0 , y0 ]. We now turn back to segments in the hyperbolic graph G, as in the picture on the right. Denote by c ∈ [x, y] the unique point with d(x, c) = d(x′ , z). By construction, c ∈ [a, b]. To conclude the proof of (7.1), we show that d(c, z) ≤ 2δ. Choose a geodesic [x, y ′ ] and denote by e ∈ [x, y ′ ] the unique point with d(x, e) = d(x′ , z). Applying δ-thinness to the geodesic triangle x, x′ , y ′ , we find that d(z, e) ≤ δ. Then applying δ-thinness to the geodesic triangle x, y, y ′ , we get that d(e, c) ≤ δ. So, d(z, c) ≤ 2δ and the claim in (7.1) is proven. Given a finite subset A ⊂ V , denote by p(A) the uniform probability measure on A. Exactly as in the proof of [BO08, Theorem 5.3.15], define the sequence of maps ηn : V × V → Prob(V ) : ηn (x, y) = 2n 1 X p(A(x, y, k)) . n k=n+1 For finite sets A, B ⊂ V , we have |A ∩ B| 1 kp(A) − p(B)k1 = 1 − . 2 max{|A|, |B|} When d(x, x′ ) ≤ d ≤ k, we have A(x, y, k − d) ⊂ A(x′ , y, k) ⊂ A(x, y, k + d) . Therefore, whenever d(x, x′ ) ≤ d ≤ k, we have |A(x, y, k − d)| 1 kp(A(x, y, k)) − p(A(x′ , y, k))k1 ≤ 1 − . 2 |A(x, y, k + d)| So if d(x, x′ ) ≤ d ≤ n, we use the inequality between arithmetic and geometric mean and get that 2n 1 X |A(x, y, k − d)| 1 ′ kηn (x, y) − ηn (x , y)k1 ≤ 1 − 2 n |A(x, y, k + d)| k=n+1 !1/n 2n Y |A(x, y, k − d)| ≤1− |A(x, y, k + d)| k=n+1 !1/n Qn+d k=n−d+1 |A(x, y, k)| = 1 − Q2n+d . k=2n−d+1 |A(x, y, k)| Using (7.1), it follows that whenever d(x, x′ ) ≤ d ≤ n and d(x, y) ≥ 4(2n + d), we have −2d/n 1 . kηn (x, y) − ηn (x′ , y)k1 ≤ 1 − 2(2n + d + 1)B 2 39 So, for every ε > 0 and every d ∈ N, there exists an n such that kηn (x, y) − ηn (x′ , y)k1 < ε for all x, x′ , y ∈ V with d(x, x′ ) ≤ d and d(x, y) ≥ 4(2n + d). The maps ηn are G-equivariant in the sense that ηn (g · x, g · y) = g · ηn (x, y) for all g ∈ G, x, y ∈ V , and the maps ηn are symmetric in the sense that ηn (x, y) = ηn (y, x) for all x, y ∈ V . Passing to a subsequence, we find a sequence of G-equivariant symmetric maps ηn : V × V → Prob(V ) and a strictly increasing sequence of integers dn ∈ N such that kηn (x, y) − ηn (x′ , y)k1 ≤ 2−n whenever x, x′ , y ∈ V , d(x, x′ ) ≤ n and d(x, y) ≥ dn . By convention, we take d0 = 0 and η0 (x, y) = 12 (δx + δy ). Define the function ρ : N → N given by ρ(n) = max{k ∈ N | dk ≤ n} . Then ρ(0) = 0, ρ is increasing, ρ(n) → ∞ when n → ∞ and |ρ(n) − ρ(m)| ≤ |n − m| for all n, m ∈ N. Using the trick in [BO08, Exercise 15.1.1], define the G-equivariant symmetric map ρ(d(x,y)) 1 + µ : V × V → ℓ (V ) : µ(x, y) = X ηn (x, y) . n=0 Note that kµ(x, y)k1 = 1 + ρ(d(x, y)). For all x, x′ , y ∈ V with ρ(d(x, y)) ≥ d(x, x′ ), we have ρ(d(x,y)) kµ(x, y) − µ(x′ , y)k1 ≤ 2d(x, x′ ) + |ρ(d(x, y)) − ρ(d(x′ , y))| + X kηn (x, y) − ηn (x′ , y)k1 n=d(x,x′ ) ρ(d(x,y)) ′ ≤ 3d(x, x ) + X 2−n ≤ 3d(x, x′ ) + 2 . (7.2) n=d(x,x′ ) Define the G-equivariant symmetric map η : V × V → Prob(V ) : η(x, y) = kµ(x, y)k−1 1 µ(x, y) . Since kµ(x, y)k1 = 1 + ρ(d(x, y)), it follows from (7.2) that kη(x, y) − η(x′ , y)k1 ≤ 2(3d(x, x′ ) + 2) 1 + ρ(d(x, y)) for all x, x′ , y ∈ V with ρ(d(x, y)) ≥ d(x, x′ ). This implies that for every n ∈ N, there exists a κn ∈ N such that 1 kη(x, y) − η(x′ , y)k1 < n ′ ′ for all x, x , y ∈ V with d(x, x ) ≤ n and d(x, y) ≥ κn . Fix a base point x0 ∈ V and define the continuous map γ : G → Prob(V ) = γ(g) = η(x0 , g · x0 ). Since γ(gh) = η(x0 , gh · x0 ) = g · η(g −1 · x0 , h · x0 ) and g · γ(h) = g · η(x0 , h · x0 ) , 40 we find that limh→∞ kγ(gh) − g · γ(h)k1 = 0 uniformly on compact sets of g ∈ G. Since γ(hk) = η(x0 , hk · x0 ) = η(hk · x0 , x0 ) = h · η(k · x0 , h−1 · x0 ) and γ(h) = h · η(x0 , h−1 · x0 ) , we find that kγ(hk) − γ(h)k1 = kη(k · x0 , h−1 · x0 ) − η(x0 , h−1 · x0 )k1 , so that also limh→∞ kγ(hk) − γ(h)k1 = 0 uniformly on compact sets of k ∈ G. As in the proof of 2, there exists a G-equivariant isometric map Prob(V ) → S(G), so that G has property (S). 5. By [CH88], G is weakly amenable. By [Sk88, Proof of Théorème 4.4], G has property (S). In the proof of Proposition 7.1, we used the following stability result for property (S). One can actually prove that property (S) is stable under measure equivalence of locally compact groups, but for our purposes, the following elementary lemma is sufficient. Lemma 7.2. Let G be a locally compact group and K < G0 < G closed subgroups such that K is compact and normal in G0 , and G0 is open and of finite index in G. If G0 /K has property (S), then also G has property (S). Proof. Since L1 (G0 /K) ⊂ L1 (G0 ), we have a G0 -equivariant map S(G0 /K) → S(G0 ). So, property (S) for G0 /K implies property (S) for G0 . Write G as the disjoint union of G0 gi , i = 1, . . . , n. Define the continuous map π : G → G0 given by π(ggi ) = g for all g ∈ G0 and i ∈ {1, . . . , n}. View S(G0 ) ⊂ S(G). Given η0 : G0 → S(G0 ) as in the definition of property (S), define the continuous map n 1 X −1 gi · η0 (π(gi g)) . η : G → S(G) : η(g) = n i=1 One checks that limh→∞ kη(ghk) − g · η(h)k1 = 0 uniformly on compact sets of g, k ∈ G. So, G again has property (S). References [AD81] C. Anantharaman-Delaroche, Action moyennable d’un groupe localement compact sur une algèbre de von Neumann, II. Math. Scand. 50 (1982), 251-268. [BS92] S. Baaj and G. Skandalis, Unitaires multiplicatifs et dualité pour les produits croisés de C∗ -algèbres. Ann. Sci. École Norm. Sup. (4) 26 (1993), 425-488. [BSV02] S. Baaj, G. Skandalis and S. Vaes, Non-semi-regular quantum groups coming from number theory. Comm. Math. Phys. 235 (2003), 139-167. [BHV15] R. Boutonnet, C. Houdayer and S. Vaes, Strong solidity of free Araki-Woods factors. Amer. J. Math., to appear. arXiv:1512.04820 [BCL16] J. Brodzki, C. Cave and K. Li, Exactness of locally compact groups. Adv. Math. 312 (2017), 209-233. [BO08] N.P. Brown and N. Ozawa, C∗ -algebras and finite-dimensional approximations. Graduate Studies in Mathematics 88. American Mathematical Society, Providence, 2008. [CCMT12] P. Caprace, Y. de Cornulier, N. Monod and R. Tessera, Amenable hyperbolic groups. J. Eur. Math. Soc. 17 (2015), 2903-2947. [CS11] I. Chifan and T. Sinclair, On the structural theory of II1 factors of negatively curved groups. Ann. Sci. École Norm. Sup. 46 (2013), 1-34. [Co79] A. Connes, Sur la théorie non commutative de l’intégration. In Algèbres d’opérateurs, Lecture Notes in Mathematics 725, Springer, Berlin, 1979, pp. 19-143. 41 [CH88] [Fo74] [Fu98] [Ha78] [HI15] [HR16] [HU15] [HV12] [Io12] [Ka02] [KPV13] [MMS03] [MS04] [Oz03] [Oz04] [Oz07] [Oz10] [OP07] [Po01] [Po03] [Po06] [PV08] [PV11] [PV12] [Ra15] [Sa09] [Sk88] M. Cowling and U. Haagerup, Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one. Invent. Math. 96 (1989), 507-549. P. Forrest, On the virtual groups defined by ergodic actions of Rn and Zn . Adv. Math. 14 (1974), 271-308. A. Furman, Orbit equivalence rigidity. Ann. of Math. 150 (1999), 1083-1108. U. Haagerup, An example of a nonnuclear C∗ -algebra, which has the metric approximation property. Invent. Math. 50 (1978/79), 279-293. C. Houdayer and Y. Isono, Unique prime factorization and bicentralizer problem for a class of type III factors. Adv. Math. 305 (2017), 402-455. C. Houdayer and S. Raum, Locally compact groups acting on trees, the type I conjecture and non-amenable von Neumann algebras. Preprint. arXiv:1610.00884 C. Houdayer and Y. Ueda, Asymptotic structure of free product von Neumann algebras. Math. Proc. Cambridge Philos. Soc. 161 (2016), 489-516. C. Houdayer and S. Vaes, Type III factors with unique Cartan decomposition. J. Math. Pures Appl. 100 (2013), 564-590. A. Ioana, Classification and rigidity for von Neumann algebras. In European Congress of Mathematics, Kraków, 2012, Eur. Math. Soc., Zürich, 2013, pp. 601-625. V.A. Kaimanovich, Boundary amenability of hyperbolic spaces. In Discrete geometric analysis, Contemp. Math. 347, Amer. Math. Soc., Providence, 2004, pp. 83-111. D. Kyed, H.D. Petersen and S. Vaes, L2 -Betti numbers of locally compact groups and their cross section equivalence relations. Trans. Amer. Math. Soc. 367 (2015), 4917-4956. I. Mineyev, N. Monod and Y. Shalom, Ideal bicombings for hyperbolic groups and applications. Topology 43 (2004), 1319-1344. N. Monod and Y. Shalom, Cocycle superrigidity and bounded cohomology for negatively curved spaces. J. Differential Geom. 67 (2004), 395-455. N. Ozawa, Solid von Neumann algebras. Acta Math. 192 (2004), 111-117. N. Ozawa, A Kurosh-type theorem for type II1 factors. Int. Math. Res. Not. Art. ID 97560 (2006), 21 pp. N. Ozawa, Weak amenability of hyperbolic groups. Groups Geom. Dyn. 2 (2008), 271-280. N. Ozawa, Examples of groups which are not weakly amenable. Kyoto J. Math. 52 (2012), 333-344. N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra. Ann. Math. 172 (2010), 713-749. S. Popa, On a class of type II1 factors with Betti numbers invariants. Ann. of Math. 163 (2006), 809-899. S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, I. Invent. Math. 165 (2006), 369-408. S. Popa, Deformation and rigidity for group actions and von Neumann algebras. In Proceedings of the International Congress of Mathematicians (Madrid, 2006), Vol. I, European Mathematical Society Publishing House, 2007, pp. 445-477. S. Popa and S. Vaes, Cocycle and orbit superrigidity for lattices in SL(n, R) acting on homogeneous spaces. In Geometry, rigidity, and group actions, Chicago Lectures in Math., Univ. Chicago Press, Chicago, 2011, pp. 419-451. S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary actions of free groups. Acta Math. 212 (2014), 141-198. S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary actions of hyperbolic groups. J. Reine Angew. Math. 694 (2014), 215-239. S. Raum, C∗ -simplicity of locally compact Powers groups. To appear in J. Reine Angew. Math. arXiv:1505.07793 H. Sako, Measure equivalence rigidity and bi-exactness of groups. J. Funct. Anal. 257 (2009), 3167-3202. G. Skandalis, Une notion de nucléarité en K-théorie (d’après J. Cuntz). K-Theory 1 (1988), 549-573. 42 [Sz91] [Ta03] [Va00] [Va10a] [Va10b] [Va16] [VV05] [Zi75] [Zi84] G. Szwarc, Groups acting on trees and approximation properties of the Fourier algebra. J. Funct. Anal. 95 (1991), 320-343. M. Takesaki, Theory of operator algebras, II. Encyclopaedia of Mathematical Sciences 125, Springer-Verlag, Berlin, 2003. S. Vaes, The unitary implementation of a locally compact quantum group action. J. Funct. Anal. 180 (2001), 426-480. S. Vaes, Rigidity for von Neumann algebras and their invariants. In Proceedings of the International Congress of Mathematicians (Hyderabad, 2010), Vol. III, Hindustan Book Agency, New Delhi, 2010, pp. 1624-1650. S. Vaes, One-cohomology and the uniqueness of the group measure space decomposition of a II1 factor. Math. Ann. 355 (2013), 661-696. S. Vaes, Amenability versus non amenability: an introduction to von Neumann algebras. In European Congress of Mathematics, Berlin, 2016, Eur. Math. Soc., Zürich, in press. S. Vaes and R. Vergnioux, The boundary of universal discrete quantum groups, exactness and factoriality. Duke Math. J. 140 (2007), 35-84. R.J. Zimmer, Extensions of ergodic group actions. Illinois J. Math. 20 (1976), 373-409. R.J. Zimmer, Ergodic theory and semisimple groups. Monographs in Mathematics 81. Birkhäuser Verlag, Basel, 1984. 43
4math.GR
Molecular simulation of the surface tension of real fluids Stephan Werth,1 Martin Horsch,1, a) and Hans Hasse1 Laboratory of Engineering Thermodynamics, University of Kaiserslautern, Erwin-Schrödinger-Str. 44, 67663 Kaiserslautern, Germany (Dated: 4th August 2016) arXiv:1608.04193v1 [physics.comp-ph] 15 Aug 2016 Molecular models of real fluids are validated by comparing the vapor-liquid surface tension from molecular dynamics (MD) simulation to correlations of experimental data. The considered molecular models consist of up to 28 interaction sites, including Lennard-Jones sites, point charges, dipoles and quadrupoles. They represent 38 real fluids, such as ethylene oxide, sulfur dioxide, phosgene, benzene, ammonia, formaldehyde, methanol and water, and were adjusted to reproduce the saturated liquid density, vapor pressure and enthalpy of vaporization. The models were not adjusted to interfacial properties, however, so that the present MD simulations are a test of model predictions. It is found that all of the considered models overestimate the surface tension. In most cases, however, the relative deviation between the simulation results and correlations to experimental data is smaller than 20 %. This observation corroborates the outcome of our previous studies on the surface tension of 2CLJQ and 2CLJD fluids where an overestimation of the order of 10 to 20 % was found. Keywords: Molecular simulation, surface tension, vapor-liquid equilibrium a) Author to whom correspondence should be addressed: [email protected] 1 I. INTRODUCTION Interfacial properties are important for many applications in process engineering, including processes like absorption, wetting, nucleation, cavitation or foaming. Experimental data on the surface tension are available for pure fluids, but the temperature range is usually limited to ambient conditions1,2 . Hence, it is desirable to have models which allow predicting interfacial properties of pure fluids and mixtures over a wide temperature and pressure range. Molecular modelling and simulation can be used for this purpose if the underlying force fields are accurate3 . In previous work of our group4–16 and recent work by Vrabec and co-workers17–22 a large number of molecular models for real fluids were developed. The molecular model parameters were adjusted to describe the saturated liquid density, vapor pressure and enthalpy of vaporization, which they do well. These models were also used to predict transport properties and they showed very good agreement for the shear viscosity, self diffusion coefficients and thermal conductivity of pure fluids6,18,21–23 and mixtures23–27 . Some of the fluids discussed in the present work have recently been used to develop fundamental equation of state based on molecular simulation as well as experimental data, e.g. ethylene oxide28 , phosgene29 , hexamethyldisiloxane21 and octamethylcyclotetrasiloxane22 . The surface tension was not part of the parameterization and is thus strictly predictive. In previous work systematic evaluations of the surface tension of the two center LennardJones plus point quadrupole (2CLJQ) and the two center Lennard-Jones plus point dipole (2CLJD) molecular model class were conducted30,31 . These models, on average, overestimate surface tension by about 20 % and 12 %, respectively31,32 . Other molecular models which have been adjusted to bulk properties, but not to interfacial properties, exhibit similar deviations33–41 . In the present work, existing molecular models are used straightforwardly. No parameters are changed. By molecular dynamics (MD) simulation, predictions of interfacial properties from bulk properties are obtained. Used in this way, molecular modeling can be compared to other approaches for predicting the surface tension from bulk data, such as phenomenological parachor correlations42–45 , corresponding-states or critical-scaling expressions45,46 , which are also phenomenological correlations, and other molecular methods, e.g. square gradient theory47 and density functional theory48,49 on the basis of molecular equations of state50–53 . In the present work, bulk and interfacial properties of real fluids are determined simultaneously from heterogeneous MD simulations. The simulation results are compared with correlations to experimental data, where available. II. MOLECULAR SIMULATION The molecular models discussed in the present work are taken from previous work of our group4–16 and recent work by Vrabec and co-workers17–22 . The molecular models are internally rigid and consist of several Lennard-Jones sites and superimposed electrostatics. 2 The total potential energy is given by " ( nLJ nLJ " 12  6 # X nei nej j N −1 X N i X X X X 1 qic qjd σijab σijab 4ǫijab U= + − r r 4πǫ0 rijcd ijab ijab a=1 b=1 c=1 d=1 i=1 j=i+1 #) µic µjd µic Qjd + µjd Qic Qic Qjd + 3 · f1 (ωi , ωj ) + · f2 (ωi , ωj ) + 5 · f3 (ωi , ωj ) , 4 rijcd rijcd rijcd (1) where ǫijab and σijab are the Lennard-Jones energy and size parameters, rijab and rijcd are sitesite distances, qic , qjd , µic , µjd , Qic and Qjd are the magnitude of the electrostatic interactions, i.e. the point charges, dipole and quadrupole moments, and fk (ωi , ωj ) are dimensionless angle-dependent expressions in terms of the orientation ωi ,ωj of the point multipoles54 . Thermodynamic properties in heterogeneous systems are very sensitive to a truncation of the intermolecular potential40,55–59 . For dispersive interactions, like the Lennard-Jones potential, various long range correction (LRC) approaches exist which are known to be accurate for planar fluid interfaces60–66 . The simulations in the present work use slab-based LRC techniques based on the density profile64–66 . For polar interactions, LRCs based on Ewald summation are typically used for the simulation of vapor-liquid interfaces61,67–70 . However, a computationally efficient slab-based LRC based on the density profile can also be used for polar molecular models58 . In terms of the thermodynamic results, the different methods deliver a similar degree of accuracy for the two-center Lennard-Jones plus point dipole fluid31,69,70 . Therefore, the slab-based LRC technique described in previous work58 is employed here both for dipolar electrostatic interactions and for dispersion. For the present series of MD simulations, systems were considered where the vapor and liquid phases coexist with each other in direct contact, employing periodic boundary conditions, so that there are two vapor-liquid interfaces which are oriented perpendicular to the y axis. The interfacial tension was computed from the deviation between the normal and the tangential diagonal components of the overall pressure tensor71,72 , i.e. the mechanical route, Z 1 ∞ dy (pN − pT ) . (2) γ= 2 −∞ Thereby, the normal pressure pN is given by the y component of the diagonal of the pressure tensor, and the tangential pressure pT was determined by averaging over x and z components of the diagonal of the pressure tensor. The simulations were performed with the MD code ls1 mardyn 73 in the canonical ensemble with N = 16,000 particles. Further details on the MD simulations are given in the Appendix. III. RESULTS AND DISCUSSION Tab. I gives an overview of the molecular models investigated in the present work. All studied models are rigid, i.e. internal degrees of freedom are no accounted for. The deviations δρ′ and δpS , which are reported for the models in Tab. I, are taken from the corresponding publications and represent relative mean deviations of the simulated values from correlations to experimental data4–22 . The molecular simulations in previous work were performed with the Grand Equilibrium method74 . 3 101 p / MPa 100 10−1 10−2 10−3 10−4 2 3 4 1000 K / T 5 FIG. 1. Vapor pressure curves of ammonia, methanol, sulfur dioxide and benzene. Solid lines represent correlations to experimental data75–78 and symbols are the present simulation results: Sulfur dioxide (◦), benzene (), ammonia (△) and methanol (⋄). Fig. 1 shows the simulation results for the vapor pressure of ammonia, methanol, sulfur dioxide and benzene from the present work which were obtained from heterogeneous simulations; cf. the Appendix (Tab. II) for a complete presentation of the simulated properties. The present results for the saturated densities and the vapor pressure are in very good agreement with experimental data75–78 . However, for low temperatures the uncertainties in the vapor density and vapor pressure are relatively high. This is due to the fact that a low temperatures in many cases on average less than one molecule is in the vapor phase, which yields relatively high statistical uncertainties. Similar findings were obtained for the the other fluids studied in the present work. The simulation results obtained for the vapor pressure and the saturated densities for all studied fluids are reported in Tab. II together with the data for the surface tension. The relative mean deviation δγ between the simulation data and the experimental data reported in Tab. I is calculated in the same way as the deviations for the saturated liquid density and the vapor pressure. It represents the relative mean deviation of the surface tension predicted by the molecular models from Design Institute for Physical Properties (DIPPR) correlations to experimental data v u 2 K  sim u1 X γ (Ti ) − γ exp (Ti ) t , (3) |δγ| = K i γ exp (Ti ) between the triple point temperature and 95 % of the critical temperature. By convention, the sign of δγ is positive if, on average, the model overestimates the surface tension (δγ = + |δγ|) and negative otherwise (δγ = − |δγ|). Underlying experimental surface tension data are usually not available over the entire temperature range. Only for four compounds water, methanol, ammonia, heptafluoropropane - experimental data are available over the entire temperature range. In most cases, the surface tension is measured only up to 373 K and the DIPPR correlation extrapolates these results to the critical point1,2 . 4 γ / mN m−1 50 40 30 20 10 0 100 200 300 400 T /K 500 600 FIG. 2. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 and symbols are the present simulation results: Thiophene (▽), ethylene oxide (◦), dimethyl ether () and heptafluoropropane (△). γ / mN m−1 100 80 60 40 20 0 200 300 400 500 T /K 600 FIG. 3. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 and symbols are the present simulation results: Hydrazine (△), methylhydrazine (▽), 1,1-dimethylhydrazine (⋄), cyanogen chloride (◦) and cyanogen (). No experimental data are available for methylhydrazine and 1,1-dimethylhydrazine. The DIPPR correlations usually agree with available experimental data within 3 %, only for dimethyl sulfide, ortho-dichlorobenzene, heptafluoropropane, cyanogen, decafluorobutane and hexamethyldisiloxane deviations of up to 5 % are reported1 . For three fluids formaldehyde, methylhydrazine and 1,1-dimethylhydrazine - no experimental data are available. The DIPPR correlations do not match the critical temperature for ethylene glycol and formic acid. Therefore, a straight line is used to connect the DIPPR correlation and the critical point of respective fluids. 5 γ / mN m−1 40 30 20 10 0 100 200 300 400 T /K 500 600 FIG. 4. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 and symbols are the present simulation results: Cyclohexanol (△), ethyl acetate (), hexamethyldisiloxane (⋄) and decafluorobutane (◦). Figs. 2 – 9 show the surface tension for all studied fluids as a function of the temperature. The molecular simulation results are compared with DIPPR correlations to experimental data. For formaldehyde, methylhydrazine and 1,1-dimethylhydrazine no experimental data are available. The predictions of the surface tension agree reasonably well with the experimental data. Fig. 10 shows the surface tension predicted by MD simulation as a function of the experimental surface tension calculated by the DIPPR correlation1 for all studied fluids. γ / mN m−1 50 40 30 20 10 0 200 300 400 500 T /K 600 FIG. 5. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 and symbols are the present simulation results: Cyclohexanone (◦), cyclohexane (), formaldehyde (△) and octamethylcyclotetrasiloxane (⋄). No experimental data are available for formaldehyde. 6 The molecular models overestimate the surface tension in all cases. The average deviation between the predictions by the molecular simulation and the experimental data is about 20 %. This is in line with results for the surface tension obtained by molecular simulation in the literature31–41 . Compared to other methods for predicting the surface tension of lowmolecular fluids, molecular modeling and simulation, using models which are adjusted to bulk data, leads to relatively high deviations. Several examples illustrate this: The surface tension of benzene is reproduced with a deviation of |δγ| ≈ 4 % using the corresponding-states (CS) correlation by Sastry and Rao46 , |δγ| = 1 % with the CS correlation by Zuo and Stenby45 , and |δγ| = 8.5 % from a parachor correlation43–45 ; the molecular model from Huang et al.9 has |δγ| = 11.9 %. For cyclohexane, |δγ| ≈ 1 % is obtained following CS by Sastry and Rao46 , |δγ| = 0.7 % with CS by Zuo and Stenby45 , |δγ| = 7.4 % from the parachor correlation43–45 , and |δγ| = 10.8 % with the molecular model from Merker et al.5 . In case of ethyl acetate, a CS correlation yields |δγ| ≈ 7 %46 and density functional theory with the PC-SAFT equation of state49 reaches |δγ| ≈ 4 %, whereas the average relative deviation is |δγ| = 10.3 % for the molecular model from Eckelsbach et al.19 For methanol, the CS correlation by Sastry and Rao46 exhibits almost perfect agreement |δγ| < 1 %; the molecular model by Schnabel et al.13 has |δγ| = 35.3 %. Zuo and Stenby45 reproduce the surface tension of isobutane with an accuracy of |δγ| = 1.6 % using their CS correlation and with |δγ| = 2.5 % using a parachor correlation43–45 ; in contrast, the molecular model for isobutane from Eckl et al.8 exhibits a deviation of |δγ| = 12.5 %. Square gradient theory with the SAFT-VR Mie equation of state, following Garrido et al.47 , typically yields deviations of the order of 2 to 3% for the surface tension of lowmolecular fluids. However, no direct comparison is possible with any of the present results. The anisotropic united atom (AUA) force field79,80 was also adjusted to bulk data only. All results for γ from simulations with AUA models are therefore predictions of interfa- γ / mN m−1 60 40 20 0 100 200 300 400 500 600 700 T /K FIG. 6. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 , the dotted line connects the DIPPR correlation with the critical point of ethylene glycol, and symbols are the present simulation results: Ethylene glycol (◦), orthodichlorobenzene (△), chlorobenzene (▽), benzene (⋄) and phosgene (). 7 100 γ / mN m−1 80 60 40 20 0 200 300 400 500 T /K 600 FIG. 7. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 , the dotted line connects the DIPPR correlation with the critical point of formic acid, and symbols are the present simulation results: Water (◦), formic acid (⋄), methanol (△), dimethylamine (). γ / mN m−1 50 40 30 20 10 0 200 300 400 T /K 500 FIG. 8. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 and symbols are the present simulation results: Dimethyl sulfide (◦), hydrogen cyanide (△) and sulfur dioxide (). cial properties from bulk fluid properties. For benzene, applying the test-area method in Monte Carlo simulations with the AUA-9 sites force field, which was parameterized by Nieto Draghi and collaborators81,82 , Biscay et al.83 report a surface tension which deviates from experimental data by about δγ ≈ +4 %, compared to δγ = +11.9 % for the Huang et al.9 model. For cyclohexane83 , the AUA-9 sites model has δγ ≈ +5 %, whereas for the Merker et al.5 model, δγ = +11.9 % was found in the present work. The AUA-4 model80 underestimates the surface tension of methanol by δγ ≈ −12 %, cf. Biscay et al.84 , which compares 8 γ / mN m−1 60 40 20 0 100 200 300 400 500 600 T /K FIG. 9. Surface tension as a function of the temperature. Solid lines represent DIPPR correlations to experimental data1 and symbols are the present simulation results: Nitromethane (⋄), acetonitrile (△), ammonia (◦) and isobutane (). γ sim / mN m−1 100 80 60 40 20 0 0 20 40 γ corr 60 80 −1 / mN m 100 FIG. 10. Predicted surface tension over the experimental surface tension based on DIPPR correlations1 for the present molecular simulation results. The solid line represents perfect agreement between simulation and experiment, and the dashed line represents a deviation of 20 %. favorably to the Schnabel et al.13 model with δγ = +35.3 %. Overall, the AUA force field is more reliable for predicting the surface tension than the models investigated in the present work85 ; it has roughly the same accuracy as empirical parachor correlations45 . However, this still makes the AUA force field less accurate than empirical CS correlations45,46 and semiempirical square gradient theory47 . Since molecular simulation is also computationally much more expensive than the other approaches, it cannot be recommended to predict interfacial properties from molecular models which were not previously adjusted to or, at least, validated against such data. However, 9 a systematic overestimation of the surface tension has also been observed in density functional theory in combination with physically based equations of state48,86–89 . To account for this overestimation, an empirical correction expression is often employed, which is formally attributed to the presence of capillary waves and decreases the surface tension. Without this correction term, which was adjusted to fit the experimental surface tension values of the n-alkane series48 , density functional theory would deviate from the surface tension of real fluids in a similar way as the molecular models mentioned above. In square gradient theory, the influence parameter, which controls the magnitude of the surface excess free energy, is also adjusted to surface tension data. TABLE I. Molecular models discussed in the present work. The deviations δρ′ and δps as given in respective publications and δγ is the root mean square relative deviation between predictions by the present MD simulations and DIPPR correlations to experimental data1 . For all considered fluids, the molecular models overestimate γ on average. # stands for the number of the corresponding sites in the model. Name Neon Argon Krypton Xenon Methane Acetonitrile Cyclohexane Cyclohexanone Cyclohexanol Ethylene oxide Isobutane Formaldehyde Dimethyl ether Sulfur dioxide Dimethyl sulfide Thiophene Hydrogen cyanide Nitromethane Phosgene Benzene Chlorobenzene Ortho-Dichlorobenzene Cyanogen chloride Cyanogen Heptafluoropropane (R227ea) Ammonia Formic acid Methanol Dimethylamine Ethylene glycol Water Hydrazine Methylhydrazine 1,1-Dimethylhydrazine Ethyl acetate Decafluorobutane Hexamethyldisiloxane Octamethylcyclotetrasiloxane Formula Ne Ar Kr Xe CH4 C2 H3 N C6 H12 C6 H10 O C6 H10 OH C2 H4 O C4 H10 CH2 O C2 H6 O SO2 C2 H6 S C4 H1 0 HCN CH3 NO2 COCl2 C6 H6 C6 H5 Cl C6 H4 Cl2 NCCl C2 N2 C3 HF7 NH3 CH2 O2 CH3 OH C2 H7 N C2 H6 O2 H2 O N2 H4 CH6 N2 C2 H8 N2 C4 H8 O2 C4 F10 C6 H18 OSi2 C8 H24 O2 Si4 CAS RN # LJ # Charges # Dipoles # Quadrupoles δρ′ δps δγ Source 7440-01-9 1 0.6 2.5 20.1 Vrabec et al.16 7440-37-1 1 0.9 2.1 31.9 Vrabec et al.16 7439-90-9 1 0.9 2.1 27.8 Vrabec et al.16 7440-63-3 1 1.4 1.7 46.9 Vrabec et al.16 74-82-8 1 1.0 2.7 19.2 Vrabec et al.16 75-05-8 3 1 0.1 4.7 51.7 Deublein et al.4 110-82-7 6 0.3 1.7 10.8 Merker et al.5 108-94-1 7 1 0.9 2.7 26.0 Merker et al.5 108-93-0 7 3 0.2 3.0 28.3 Merker et al.6 75-21-8 3 1 0.4 1.5 16.6 Eckl et al.7 75-28-5 4 1 1 0.6 4.2 12.5 Eckl et al.8 50-00-0 2 1 0.9 4.3 - Eckl et al.8 115-10-6 3 1 0.4 2.6 18.9 Eckl et al.8 7446-09-5 3 1 1 0.9 4.0 3.4 Eckl et al.8 75-18-3 3 1 2 0.7 4.0 18.1 Eckl et al.8 110-02-1 5 1 1 1.2 3.8 22.4 Eckl et al.8 74-90-8 2 1 1 1.0 7.2 51.9 Eckl et al.8 75-52-5 4 1 1 0.2 18.7 31.5 Eckl et al.8 75-44-5 4 1 1 0.5 2.1 17.2 Huang et al.9 71-43-2 6 6 0.4 3.4 11.9 Huang et al.9 108-90-7 7 1 5 0.9 5.0 17.8 Huang et al.9 95-50-1 8 1 4 0.5 6.4 34.3 Huang et al.9 506-77-4 3 1 1 0.3 2.1 15.3 Miroshnichenko et al.17 460-19-5 4 1 0.6 13.0 2.5 Miroshnichenko et al.17 431-89-0 10 1 1 1.0 1.0 7.2 Eckl et al.10 7664-41-7 1 4 0.7 1.6 36.7 Eckl et al.11 64-18-6 3 4 0.8 5.1 9.5 Schnabel et al.12 67-56-1 2 3 0.6 1.1 35.3 Schnabel et al.13 124-40-3 3 3 0.4 6.2 28.7 Schnabel et al.14 107-21-1 4 6 0.8 4.8 32.6 Huang et al.15 7732-18-5 1 3 1.1 7.2 30.7 Huang et al.15 302-01-2 2 6 0.5 7.6 29.2 Elts et al.18 60-34-4 3 3 0.2 7.0 - Elts et al.18 57-14-7 4 3 1.3 3.7 - Elts et al.18 141-78-6 6 5 0.1 4.6 10.3 Eckelsbach et al.19 355-25-9 14 14 0.5 3.5 10.3 Köster et al.20 107-46-0 9 3 0.5 5.0 12.9 Thol et al.21 556-67-2 16 8 0.5 6.0 10.5 Thol et al.22 10 The unfavorable performance of molecular models, compared to methods which are more abstract physically and less expensive numerically, is explained by the fact that the molecular models are entirely predictive for interfacial properties. All other methods, i.e. density functional theory, square gradient theory, and phenomenological correlations, were adjusted to surface tension data at least indirectly. Moreover, methods which are based on analytical equations of state, including molecular equations of state, fail in the vicinity of the critical point, so that δγ diverges at high temperatures. Renormalization group theory has to be employed in these cases to avoid unphysical behavior49,90,91 . By molecular simulation, the (Ising class) critical scaling behavior of intermolecular pair potentials is correctly captured. Therefore, any fit of force-field parameters to VLE data over a significant temperature range always indirectly adjusts the molecular model to the critical temperature92 . It has been shown before that a better agreement of molecular simulation results with experimental data for the surface tension can be achieved by taking into account experimental data on the surface tension in the parameterization of the molecular models. However, improvements in the quality of the representation of the surface tension have to be traded off against losses in the quality of the representation of bulk fluid properties32,92–94 . IV. CONCLUSION In the present work, the surface tension of real fluids was determined by MD simulation. The surface tension was evaluated for 38 real fluids which were parameterized to reproduce the saturated liquid density, vapor pressure and enthalpy of vaporization. The agreement between the calculated bulk values in heterogeneous simulations and the experimental data is very good. On the basis of well described phase equilibria the surface tension was predicted. The surface tension is consistently overpredicted by the molecular models. On average the deviation is about +20 %. Such overpredictions have been reported before in the literature. The reasons should be investigated in more detail in future research. ACKNOWLEDGEMENT The authors gratefully acknowledge financial support from BMBF within the SkaSim project (grant no. 01H13005A) and from Deutsche Forschungsgemeinschaft (DFG) within the Collaborative Research Center (SFB) 926. They thank Patrick Leidecker for performing some of the present simulations and Gabriela Guevara Carrión, Maximilian Kohns, Yonny Mauricio Muñoz Muñoz and Jadran Vrabec for fruitful discussions. The present work was conducted under the auspices of the Boltzmann-Zuse Society for Computational Molecular Engineering (BZS), and the simulations were carried out on Elwetritsch at the Regional University Computing Center Kaiserslautern (RHRK) under the grant TUKL-MSWS and on SuperMUC at Leibniz Supercomputing Center (LRZ), Garching, within the SPARLAMPE computing project (pr48te). 11 REFERENCES 1 R. L. Rowley, W. V. Wilding, J. L. Oscarson, Y. Yang, N. A. Zundel, T. E. Daubert, and R. P. Danner. DIPPR Information and Data Evaluation Manager for the Design Institute for Physical Properties. AIChE, 2013. Version 7.0.0. 2 DDB - Dortmund Data Bank. 7.3.0.458 (DDB), 1.18.0.122 (Artist). DDBST GmbH, Oldenburg, 2015. 3 A. Ghoufi, P. Malfreyt, and D. J. Tildesley. Chem. Soc. Rev., 45(5):1387–1409, 2016. 4 S. Deublein, P. Metzler, J. Vrabec, and H. Hasse. Mol. Sim., 39(2):109–118, 2012. 5 T. Merker, J. Vrabec, and H. Hasse. Fluid Phase Equilib., 315:77–83, 2012. 6 T. Merker, J. Vrabec, and H. Hasse. Soft Mater., 10(1-3):3–24, 2012. 7 B. Eckl, J. Vrabec, and H. Hasse. Fluid Phase Equilib., 274(1–2):16–26, 2008. 8 B. Eckl, J. Vrabec, and H. Hasse. J. Phys. Chem. B, 112(40):12710–12721, 2008. 9 Y.-L. Huang, M. Heilig, H. Hasse, and J. Vrabec. AIChE J., 52(4):1043–1060, 2011. 10 B. Eckl, Y.-L. Huang, J. Vrabec, and H. Hasse. Fluid Phase Equilib., 260:177–182, 2007. 11 B. Eckl, J. Vrabec, and H. Hasse. Mol. Phys., 106:1039–1046, 2008. 12 T. Schnabel, M. Cortada, J. Vrabec, S. Lago, and H. Hasse. Chem. Phys. Lett., 435:268– 272, 2007. 13 T. Schnabel, A. Srivastava, J. Vrabec, and H. Hasse. J. Phys. Chem. B, 111:9871–9878, 2007. 14 T. Schnabel, J. Vrabec, and H. Hasse. Fluid Phase Equilib., 263:144–159, 2008. 15 Y.-L. Huang, T. Merker, M. Heilig, H. Hasse, and J. Vrabec. Ind. Eng. Chem. Res., 51(21):7428–7440, 2012. 16 J. Vrabec, J. Stoll, and H. Hasse. J. Phys. Chem. B, 105:12126–12133, 2001. 17 S. Miroshnichenko, T. Grützner, D. Staak, and J. Vrabec. Fluid Phase Equilib., 354:286– 297, 2013. 18 E. Elts, T. Windmann, D. Staak, and J. Vrabec. Fluid Phase Equilib., 322–323:79–91, 2012. 19 S. Eckelsbach, T. Janzen, A. Köster, S. Miroshnichenko, Y. M. Muñoz Muñoz, and J. Vrabec. In W. E. Nagel, D. B. Kröner, and M. M. Resch, editors, High Performance Computing in Science and Engineering ’14, pages 645–649. Springer, Berlin/Heidelberg, 2014. 20 A. Köster, P. Nandi, T. Windmann, D. Ramjugernath, and J. Vrabec. Fluid Phase Equilib., 336:104–112, 2012. 21 M. Thol, F. H. Dubberke, G. Rutkai, T. Windmann, A. Köster, R. Span, and J. Vrabec. Fluid Phase Equilib., 416:133–151, 2016. 22 M. Thol, G. Rutkai, A. Köster, F. H. Dubberke, T. Windmann, R. Span, and J. Vrabec. J. Chem. Eng. Data, 61(7):2580–2595, 2016. 23 G. Guevara Carrión, J. Vrabec, and H. Hasse. Int. J. Thermophys., 33:449–468, 2012. 24 S. Pařez, G. Guevara Carrión, H. Hasse, and J. Vrabec. Phys. Chem. Chem. Phys., 15:3985–4001, 2013. 25 G. Guevara Carrión, J. Vrabec, and H. Hasse. J. Chem. Phys., 134:074508, 2011. 26 G. Guevara Carrión, J. Vrabec, and H. Hasse. Fluid Phase Equilib., 316:46–54, 2012. 12 27 G. Guevara Carrión, T. Janzen, Y. M. Muñoz Muñoz, and J. Vrabec. J. Chem. Phys., 144:124501, 2016. 28 M. Thol, G. Rutkai, A. Köster, M. Kortmann, R. Span, and J. Vrabec. Chem. Eng. Sci., 121:87–99, 2015. 29 G. Rutkai and J. Vrabec. J. Chem. Eng. Data, 60:2895–2905, 2015. 30 S. Werth, M. Horsch, and H. Hasse. Fluid Phase Equilib., 392:12–18, 2015. 31 S. Werth, M. Horsch, and H. Hasse. J. Chem. Phys., 144:054702, 2016. 32 S. Werth, K. Stöbener, P. Klein, K.-H. Küfer, M. Horsch, and H. Hasse. Chem. Eng. Sci., 121:110–117, 2015. 33 J.-C. Neyt, A. Wender, V. Lachet, and P. Malfreyt. J. Phys. Chem. B, 115(30):9421–9430, 2011. 34 S. Eckelsbach, S. Miroshnichenko, G. Rutkai, and J. Vrabec. In W. E. Nagel, D. B. Kröner, and M. M. Resch (eds.), High Performance Computing in Science and Engineering ’13, pp. 635–646. Springer, Heidelberg, 2013. 35 C. Avendaño, T. Lafitte, A. Galindo, C. S. Adjiman, G. Jackson, and E. A. Müller. J. Phys. Chem. B, 115(38):11154–11169, 2011. 36 C. Herdes, T. S. Totton, and E. A. Müller. Fluid Phase Equilib., 406:91–100, 2015. 37 N. Ferrando, V. Lachet, J. Pérez Pellitero, A. D. Mackie, and P. Malfreyt. J. Phys. Chem. B, 115(36):10654–10664, 2011. 38 S. Eckelsbach and J. Vrabec. Phys. Chem. Chem. Phys., 17:27195–27203, 2015. 39 S. K. Singh, A. Sinha, G. Deo, and J. K. Singh. J. Phys. Chem. C, 113(17):7170–7180, 2009. 40 R. A. Zubillaga, A. Labastida, B. Cruz, J. C. Martı́nez, E. Sánchez, and J. Alejandre. J. Chem. Theory Comput., 9(3):1611–1615, 2013. 41 C. Caleman, P. J. van Maaren, M. Hong, J. S. Hub, L. T. Costa, and D. van der Spoel. J. Chem. Theory Comput., 8(1):61–74, 2012. 42 S. Sugden. J. Chem. Soc., 125:1177–1189, 1924. 43 J. A. Hugill and A. J. van Welsenes. Fluid Phase Equilib., 29:383–390, 1986. 44 K. A. M. Gasem, P. B. Dulcamara Jr., K. B. Dickson, and R. L. Robinson Jr. Fluid Phase Equilib., 53:39–50, 1989. 45 Y.-X. Zuo and E. H. Stenby. Canad. J. Chem. Eng., 75(6):1130–1137, 1997. 46 S. R. S. Sastry and K. K. Rao. Chem. Eng. J., 59:181–186, 1995. 47 J. M. Garrido, M. M. Piñeiro, F. J. Blas, and E. A. Müller. AIChE J., 62(5):1781–1794, 2016. 48 J. Groß. J. Chem. Phys., 131: 204705, 2009. 49 X. Tang and J. Groß. J. Supercrit. Fluids, 55(2):735–742, 2010. 50 J. Groß and G. Sadowski. Ind. Eng. Chem. Res., 44(4):1244–1260, 2001. 51 J. Groß and J. Vrabec. AIChE J., 52(3): 1194–1204, 2006. 52 J. Vrabec and J. Groß. J. Phys. Chem. B, 112(1):51–60, 2008. 53 C. Avendaño, T. Lafitte, C. S. Adjiman, A. Galindo, E. A. Müller, and G. Jackson. J. Phys. Chem. B, 117(9): 2717–2733, 2013. 54 C. G. Gray and K. E. Gubbins. Theory of Molecular Fluids, Vol. 1: Fundamentals. Clarendon, Oxford, 1984. 55 F. Goujon, P. Malfreyt, and D.-J. Tildesley. J. Chem. Phys., 140:244710, 2014. 13 56 S. Werth, M. Horsch, J. Vrabec, and H. Hasse. J. Chem. Phys., 142:107101, 2015. F. Goujon, P. Malfreyt, and D.-J. Tildesley. J. Chem. Phys., 142:107102, 2015. 58 S. Werth, M. Horsch, and H. Hasse. Mol. Phys., 113(23):3750–3756, 2015. 59 F. Goujon, P. Malfreyt, J.-M. Simon, A. Boutin, B. Rousseau, and A. H. Fuchs. J. Chem. Phys., 121(24):12559–12571, 2004. 60 P. J. in ’t Veld, A. E. Ismail, and G. S. Grest. J. Chem. Phys., 127:144711, 2007. 61 R. E. Isele-Holder, W. Mitchell, and A. E. Ismail. J. Chem. Phys., 137:174107, 2012. 62 G. Mathias, B. Egwolf, M. Nonella, and P. Tavan. J. Chem. Phys., 118(24):10847–10860, 2003. 63 D. Tameling, P. Springer, P. Bientinesi, and A. E. Ismail. J. Chem. Phys., 140:024105, 2014. 64 M. Mecke, J. Winkelmann, and J. Fischer. J. Chem. Phys., 107(21):9264–9270, 1997. 65 J. Janeček. J. Phys. Chem. B, 110(12):6264–6269, 2006. 66 S. Werth, G. Rutkai, J. Vrabec, M. Horsch, and H. Hasse. Mol. Phys., 112(17):2227–2234, 2014. 67 J. W. Perram, H. G. Peterson, and S. W. de Leeuw. Mol. Phys., 65(4):875–893, 1988. 68 R. W. Hockney and J. W. Eastwood. Computer Simulations Using Particles. Taylor & Francis, Bristol, PA, USA, 1988. 69 S. G. Moore, M. J. Stevens, and G. S. Grest. Phys. Rev. E, 91:022309, 2015. 70 S. Enders, H. Kahl, M. Mecke, and J. Winkelmann. J. Mol. Liq., 115(1):29–39, 2004. 71 J. P. R. B. Walton, D. J. Tildesley, J. S. Rowlinson, and J. R. Henderson. Mol. Phys., 48(6):1357–1368, 1983. 72 J. G. Kirkwood and F. P. Buff. J. Chem. Phys., 17(3):338–343, 1949. 73 C. Niethammer, S. Becker, M. Bernreuther, M. Buchholz, W. Eckhardt, A. Heinecke, S. Werth, H.-J. Bungartz, C. W. Glass, H. Hasse, J. Vrabec, and M. Horsch. J. Chem. Theory Comput., 10(10):4455–4464, 2014. 74 J. Vrabec and H. Hasse. Mol. Phys., 100(21):3375–3383, 2002. 75 R. Tillner-Roth, F. Harms-Watzenberg, and H. D. Baehr. DKV-Tagungsbericht, 20:167– 181, 1993. 76 A. Polt, B. Platzer, and G. Maurer. Chem. Tech. (Leipzig), 44(6):216–224, 1992. 77 K. M. de Reuck and R. J. B. Craven. Methanol, International Thermodynamic Tables of the Fluid State - 12. Blackwell Scientific, London, 1993. 78 E. W. Lemmon and R. Span. J. Chem. Eng. Data, 51:785–850, 2006. 79 S. Toxvaerd. J. Chem. Phys., 93(6):4290–4295, 1990. 80 Ph. Ungerer, C. Beauvais, J. Delhommelle, A. Boutin, B. Rousseau, and A. H. Fuchs. J. Chem. Phys., 112(12):5499–5510, 2000. 81 P. Bonnaud, C. Nieto Draghi, and Ph. Ungerer. J. Phys. Chem. B, 111(14):3730–3741, 2007. 82 C. Nieto Draghi, P. Bonnaud, and Ph. Ungerer. J. Phys. Chem. C, 111(43):15686–15699, 2007. 83 F. Biscay, A. Ghoufi, V. Lachet, and P. Malfreyt. Phys. Chem. Chem. Phys., 11(29):6132– 6147, 2009. 84 F. Biscay, A. Ghoufi, V. Lachet, and P. Malfreyt. J. Phys. Chem. C, 115(17):8670–8683, 2011. 57 14 85 F. Biscay, A. Ghoufi, F. Goujon, V. Lachet, and P. Malfreyt. J. Phys. Chem. B, 112(44):13885–13897, 2008. 86 F Llovell, A. Galindo, F. J. Blas, and G. Jackson. J. Chem. Phys., 133:024704, 2010. 87 P. I. Teixeira and M. M. Telo da Gama. J. Phys.: Condens. Matter, 3(1):111–125, 1991. 88 J. Winkelmann. J. Phys.: Condens. Matter, 13(21):4739–4768, 2001. 89 P. Frodl and S. Dietrich. Phys. Rev. E, 48(5):3741–3759, 1993. 90 E. Forte, F. Llovell, L. F. Vega, J. P. M. Trusler, and A. Galindo. J. Chem. Phys., 134:154102, 2011. 91 E. Forte, F. Llovell, J. P. M. Trusler, and A. Galindo. Fluid Phase Equilib., 337:274–287, 2013. 92 K. Stöbener, P. Klein, M. Horsch, K. Küfer, and H. Hasse. Fluid Phase Equilib., 411:33–42, 2016. 93 K. Stöbener, P. Klein, S. Reiser, M. Horsch, K.-H. Küfer, and H. Hasse. Fluid Phase Equilib., 373:100–108, 2014. 94 S. Werth, K. Stöbener, M. Horsch, and H. Hasse. Mol. Phys., 2016, accepted. 95 D. Fincham. Mol. Sim., 8(3–5):165–178, 1992. 96 S. Werth, S. V. Lishchuk, M. Horsch, and H. Hasse. Physica A, 392(10):2359–2367, 2013. 97 J. M. Prausnitz, R. N. Lichtenthaler, and E. G. de Azevedo. Molecular Thermodynamics of Fluid-Phase Equilibria. Pearson Education, New Jersey, 1998. 15 TABLE II: Molecular simulation results for the vapor-liquid equilibrium of the pure components from the present work. The numbers in parentheses indicate the statistical uncertainties of the last decimal digits. T K Acetonitrile 300 370 440 510 Cyclohexane 280 335 390 445 510 Cyclohexanone 250 315 380 445 510 Cyclohexanol 300 375 450 525 600 Ethylene oxide 180 245 310 375 440 Isobutane 120 185 250 315 380 Formaldehyde 180 235 290 345 ′ ′′ ps MPa ρ mol l−1 ρ mol l−1 γ mN m−1 0.005(4) 0.137(38) 0.839(90) 2.72(14) 18.885(1) 17.013(20) 14.810(21) 11.84(9) 0.002(1) 0.053(12) 0.307(58) 1.09(15) 41.9(15) 27.1(6) 15.7(30) 5.9(3) 0.008(6) 0.049(8) 0.270(33) 0.855(34) 2.387(93) 9.355(3) 8.750(4) 8.093(3) 7.325(13) 6.117(26) 0.003(1) 0.017(3) 0.090(9) 0.273(11) 0.834(41) 29.2(12) 22.3(13) 15.5(10) 9.5(7) 3.5(4) 0.000(0) 0.002(2) 0.020(6) 0.148(22) 0.558(37) 9.902(8) 9.327(4) 8.741(5) 8.118(11) 7.423(14) 0.000(0) 0.001(1) 0.006(1) 0.042(5) 0.146(9) 50.6(38) 40.5(12) 30.8(21) 22.3(13) 14.8(6) 0.000(0) 0.009(7) 0.122(20) 0.633(24) 2.093(91) 9.700(22) 9.028(4) 8.277(3) 7.422(13) 6.325(20) 0.000(0) 0.003(1) 0.034(8) 0.165(12) 0.575(31) 40.3(62) 32.0(13) 22.9(4) 14.0(7) 5.7(11) 0.000(0) 0.006(4) 0.331(68) 1.304(57) 4.77(40) 23.230(4) 21.325(6) 19.323(27) 16.851(13) 13.39(13) 0.000(0) 0.004(3) 0.145(30) 0.495(27) 2.16(34) 48.1(33) 38.6(23) 25.0(13) 14.3(19) 3.7(11) 0.000(0) 12.654(6) 0.000(0) 37.9(13) 0.001(1) 11.539(3) 0.001(1) 26.7(8) 0.067(17) 10.383(23) 0.032(7) 17.4(5) 0.584(10) 8.983(28) 0.263(16) 9.1(7) 2.32(12) 7.112(41) 1.13(12) 2.3(4) 0.001(1) 0.030(9) 0.331(54) 1.66(13) 16 31.984(54) 29.228(5) 26.222(77) 22.594(39) 0.001(1) 0.017(5) 0.155(20) 0.782(66) 51.6(14) 36.3(15) 24.2(14) 12.7(9) T K Dimethyl ether 205 260 315 370 Sulfur dioxide 220 265 310 355 400 Dimethyl sulfide 190 260 330 400 470 Thiophene 250 320 390 460 530 Hydrogen cyanide 280 315 340 375 410 Nitromethane 260 330 400 470 540 Phosgene 160 225 290 355 420 ′ ′′ ps MPa ρ mol l−1 ρ mol l−1 γ mN m−1 0.016(8) 0.205(39) 0.95(13) 3.34(35) 17.016(5) 15.415(53) 13.563(13) 10.94(12) 0.016(4) 0.100(35) 0.442(38) 1.70(31) 30.0(12) 19.8(13) 10.9(10) 3.6(6) 0.013(8) 0.106(39) 0.65(17) 1.95(46) 4.65(26) 24.771(11) 22.923(25) 20.932(52) 18.48(12) 14.87(28) 0.007(4) 0.049(14) 0.278(74) 0.83(19) 2.30(25) 39.8(29) 29.6(27) 20.0(14) 11.1(9) 3.6(10) 0.001(1) 0.036(14) 0.342(26) 1.70(15) 5.32(33) 19.996(5) 18.300(17) 16.471(18) 14.315(56) 10.85(26) 0.000(0) 0.022(14) 0.134(16) 0.635(65) 2.60(57) 48.6(16) 36.1(10) 23.6(7) 11.8(13) 2.1(8) 0.002(1) 0.035(16) 0.263(37) 1.12(11) 3.13(29) 13.107(7) 12.168(6) 11.166(7) 10.001(16) 8.458(65) 0.001(1) 0.013(4 ) 0.084(9) 0.342(35) 1.01(16) 43.6(22) 33.3(7) 22.4(13) 13.6(10) 5.0(8) 0.038(21) 0.122(12) 0.343(32) 0.949(50) 2.03(10) 25.756(38) 24.015(12) 22.589(23) 20.50(12) 17.766(69) 0.021(11) 0.063(10) 0.153(23) 0.431(41) 1.00(12) 34.0(11) 25.5(21) 19.2(12) 13.2(17) 7.6(7) 0.001(1) 0.030(7) 0.109(22) 0.826(69) 2.84(26) 19.569(16) 18.014(6) 16.259(52) 14.337(25) 11.70(20) 0.001(1) 0.014(6) 0.044(15) 0.266(32) 1.06(15) 55.3(37) 41.6(34) 28.3(14) 16.5(15) 6.0(9) 0.000(0) 0.007(7) 0.133(34) 0.937(36) 3.35(33) 17.047(11) 15.578(10) 14.047(7) 12.311(20) 9.873(48) 0.000(0) 0.004(3) 0.058(8) 0.367(8) 1.46(22) 49.2(32) 34.8(17) 23.1(21) 12.5(12) 3.7(7) 17 T K Benzene 290 350 410 470 530 Chlorobenzene 250 335 420 505 590 Ortho-Dichlorobenzene 375 470 565 660 Cyanogen chloride 280 315 350 385 420 Cyanogen 260 290 320 350 380 Heptafluoropropane 200 250 300 350 Ammonia 220 260 300 340 380 ps MPa ′ ρ mol l−1 ′′ ρ mol l−1 γ mN m−1 0.007(7) 11.208(13) 0.003(3) 33.2(13) 0.092(23) 10.403(7) 0.033(3) 24.3(7) 0.423(47) 9.523(3) 0.137(16) 15.7(10) 1.441(67) 8.450(19) 0.468(22) 8.5(11) 3.337(92) 6.884(36) 1.21(13) 2.6(9) 0.004(3) 10.320(8) 0.002(2) 46.7(28) 0.011(6) 9.466(11) 0.005(3) 33.3(19) 0.122(13) 8.567(23) 0.037(2) 21.9(8) 0.79(14) 7.520(30) 0.222(44) 11.7(9) 2.91(8) 5.992(20) 0.94(7) 3.2(9) 0.003(2) 0.097(17) 0.701(89) 2.79(14) 8.472(5) 7.631(3) 6.647(16) 5.240(33) 0.001(1) 0.026(6) 0.173(28) 0.807(80) 36.1(51) 23.4(8) 12.6(15) 3.0(7) 0.079(21) 0.312(36) 0.86(17) 1.86(43) 3.74(36) 19.777(13) 18.490(46) 17.08(39) 15.27(7) 12.92(9) 0.036(8) 0.135(9) 0.361(83) 0.79(23) 1.86(54) 28.3(21) 21.6(16) 14.9(21) 9.4(17) 4.2(7) 0.094(31) 0.349(95) 0.79(30) 2.17(34) 4.40(5) 18.030(33) 16.866(58) 15.546(99) 13.820(16) 11.15(28) 0.049(21) 0.168(25) 0.36(17) 1.04(26) 2.89(40) 21.9(24) 16.6(12) 11.5(16) 5.7(6) 1.2(6) 0.002(1) 10.364(13) 0.001(1) 0.063(10) 9.381(8) 0.031(2) 0.466(39) 8.215(3) 0.213(21) 1.776(68) 6.531(20) 0.938(46) 21.9(10) 14.8(19) 7.7(9) 1.8(6) 0.029(2) 0.233(38) 1.029(80) 3.080(66) 7.23(15) 55.1(19) 40.1(36) 26.2(13) 15.2(25) 4.8(1) 18 41.985(19) 38.927(14) 35.470(30) 31.410(26) 25.572(21) 0.015(1) 0.116(23) 0.467(39) 1.421(43) 3.86(21) T K Formic acid 300 360 420 480 540 Methanol 245 320 395 470 Dimethylamine 210 260 310 360 410 Ethylene glycol 380 480 580 680 Water 300 375 450 525 600 Hydrazine 280 360 440 520 Methylhydrazine 270 345 420 495 570 ps MPa ′ ′′ ρ mol l−1 ρ mol l−1 γ mN m−1 0.014(10) 0.072(9) 0.333(28) 1.14(17) 3.20(25) 26.192(5) 24.478(9) 22.607(11) 20.36(17) 17.45(14) 0.007(2) 0.038(14) 0.169(20) 0.543(61) 1.57(18) 39.3(29) 30.2(23) 22.0(15) 15.3(13) 6.7(15) 0.001(1) 0.045(17) 0.58(13) 3.61(14) 26.175(9) 23.958(6) 21.317(36) 17.318(91) 0.001(1) 0.020(7) 0.359(49) 1.61(26) 36.0(16) 27.1(16) 14.0(20) 6.0(14) 0.003(1) 0.033(11) 0.257(24) 1.04(10) 2.65(74) 16.653(10) 15.523(4) 14.279(7) 12.888(6) 10.814(77) 0.002(1) 0.016(6) 0.109(10) 0.407(45) 1.07(50) 35.2(13) 27.3(4) 19.0(2) 11.8(9) 5.3(18) 0.005(3) 0.107(60) 1.144(52) 5.45(58) 16.867(29) 15.527(25) 13.759(14) 10.906(97) 0.002(1) 0.028(11) 0.268(9) 1.44(19) 65.6(21) 42.3(36) 23.6(31) 6.2(12) 0.006(5) 0.069(16) 0.63(10) 3.31(16) 10.43(20) 56.348(12) 52.650(12) 48.475(11) 43.412(11) 36.485(46) 0.002(1) 0.024(8) 0.31(13) 0.94(6) 3.28(3) 94.0(21) 75.1(13) 56.9(19) 34.1(21) 13.7(20) 0.000(0) 0.028(10) 0.358(24) 2.24(64) 32.321(8) 29.955(5) 27.220(18) 24.063(80) 0.000(0) 0.010(5) 0.153(19) 1.16(33) 96.7(23) 72.6(13) 49.4(30) 29.9(20) 0.009(5) 0.031(13) 0.420(64) 1.93(21) 5.21(22) 19.384(10) 17.944(6) 16.358(48) 14.598(16) 12.21(12) 0.004(3) 0.012(3) 0.137(40) 0.573(51) 1.73(13) 57.3(13) 43.1(16) 29.7(12) 16.7(8) 5.7(9) 19 T K 1,1-Dimethylhydrazine 260 315 370 425 480 Ethyl acetate 190 265 340 415 490 Decafluorobutane 260 310 360 Hexamethyldisiloxane 210 280 350 420 490 Octamethylcyclotetrasiloxane 310 355 420 485 550 ′ ′′ ps MPa ρ mol l−1 ρ mol l−1 γ mN m−1 0.003(2) 0.045(13) 0.290(14) 1.03(13) 2.97(19) 13.783(6) 12.833(4) 11.806(6) 10.603(3) 9.026(33) 0.001(1) 0.018(6) 0.101(9) 0.344(40) 1.10(12) 35.5(21) 27.6(16) 19.4(9) 11.9(4) 4.6(4) 0.000(0) 11.726(4) 0.000(0) 42.2(19) 0.002(1) 10.709(18) 0.001(1) 29.9(6) 0.063(12) 9.636(10) 0.024(3) 19.7(9) 0.497(39) 8.379(6) 0.170(6) 10.9(12) 2.078(53) 6.584(19) 0.831(37) 3.0(7) 0.077(13) 0.357(6) 1.329(67) 6.848(12) 0.038(7) 11.8(17) 6.021(28) 0.157(5) 6.5(6) 4.852(77) 0.657(43) 1.7(3) 0.000(0) 0.002(1) 0.047(15) 0.299(20) 1.161(26) 5.211(31) 4.788(5) 4.327(8) 3.771(10) 3.020(25) 0.000(0) 0.001(0) 0.017(4) 0.097(10) 0.421(10) 0.000(0) 0.004(2) 0.044(14) 0.217(15) 0.699(12) 3.135(14) 2.970(12) 2.709(8) 2.403(12) 1.985(27) 0.000(0) 18.6(33) 0.001(0) 14.6(10) 0.013(4) 10.1(11) 0.061(4) 5.7(2) 0.211(7) 1.9(3) 26.3(17) 18.4(9) 12.1(11) 6.7(4) 1.8(2) MOLECULAR SIMULATION DETAILS The equation of motion was solved by a leapfrog integrator95 with a time step of ∆t = 1 fs. The elongation of the simulation volume normal to the interface was 30 nm and the thickness of the liquid film in the center of the simulation volume was 15 nm to account for finite size effects96 . The elongation in the other spatial directions was at least 10 nm. The equilibration was executed for 500,000 time steps. The production was conducted for 2,500,000 time steps to reduce statistical uncertainties. The statistical errors were estimated to be three times the standard deviation of five block averages, each over 500,000 time steps. The saturated densities and vapor pressures were calculated as an average over the respective phases excluding the area close to the interface, i.e. the area where the first derivative of the density with respect to the y coordinate deviated from zero significantly. 20 The cutoff radius was set to 17.5 Å and a center-of-mass cutoff scheme was employed. The Lennard-Jones interactions were corrected with a slab-based long range correction based on the density profile66 . Electrostatic long-range interactions were approximated by a resulting effective molecular dipole and corrected with a slab-based long range correction based on the density profile58 . The quadrupolar interactions do not need a long-range correction as they decay by r −10 , cf. Prausnitz et al.97 21
5cs.CE
Universal Approximation Using Shuffled Linear Models arXiv:1308.6498v2 [math.DS] 19 Nov 2013 Laurens Bliek Faculty of Electrical Engineering, Mathematics, and Computer Science, Delft University of Technology, Mekelweg 4, 2628 CD Delft, The Netherlands Abstract This paper proposes a specific type of Local Linear Model, the Shuffled Linear Model (SLM), that can be used as a universal approximator. Local operating points are chosen randomly and linear models are used to approximate a function or system around these points. The model can also be interpreted as an extension to Extreme Learning Machines with Radial Basis Function nodes, or as a specific way of using Takagi-Sugeno fuzzy models. Using the available theory of Extreme Learning Machines, universal approximation of the SLM and an upper bound on the number of models are proved mathematically, and an efficient algorithm is proposed. Keywords: neural networks, system identification, nonlinear systems 1. Introduction The approximation of nonlinear functions or systems is a problem that appears in many disciplines [1] and that can be tackled by various approaches. One of these approaches is to decompose the problem into smaller problems and use a combination of simple solutions that only work for a small part of the problem. The combination of these local models should then approximate the global function or system. The Local Linear Model (LLM) is a key example of this strategy, where a dynamical system is partitioned into multiple operating regimes, and for each regime a simple linear model is used to model the system. If the goal is to approximate a function, not to model a system, the use of artificial neural networks has been a popular approach. One of the most popular algorithms for approximating a function with neural networks is the backpropagation algorithm. The Extreme Learning Machine (ELM) algorithm as proposed by Huang et al. [2], is an alternative algorithm which Preprint submitted to Neurocomputing February 6, 2018 outperforms the backpropagation algorithm in many aspects. Instead of tuning the parameters of the hidden layer and output layer, the hidden layer parameters are initialised randomly and kept fixed while the output weights are trained by a linear least-squares method. Since very efficient algorithms are available for solving linear least-squares problems, the ELM algorithm is an efficient method for training a feedforward neural network. Although ELMs can use nodes with many functions that do not at all resemble the activation functions observed in biological neurons [3], which makes ELM a very general approach, to the author’s knowledge they have not yet been extended in such a way that they include LLMs. The activation functions can be very general, but the hidden network nodes typically take one of two forms: either additive nodes, or Radial Basis Function (RBF) nodes are used, while the output is a linear combination of these. Unfortunately, with this architecture, this type of network can not be made equivalent to a LLM as defined in this paper. To prove that the same ELM algorithm does work for LLMs, an extension has to be made. The ELM algorithm is extended in this paper to include the use of LLMs, and the approximation capabilities are analysed and proved mathematically. Since ELMs use randomly initialised hidden layer parameters, the linear models used in the LLM in this paper will be accurate around a randomly chosen local point. This can be seen as shuffling a deck of cards and handing them to the players in a cardgame, where the cards consist of parameters that indicate the location and impact factors of the local points, while the players are the linear models. The proposed model in this paper is therefore called Shuffled Linear Model (SLM). Note that not the linear models themselves, but the localities where they are most valid are chosen randomly. Extending the ELM algorithm to include LLMs gives several advantages, like a better physical interpretation and the potential to identify nonlinear systems, making the algorithm more fit for system identification. Besides this, the extension could be used in the area of fuzzy modelling. Fuzzy models are a popular approach to model nonlinear systems, and can be used to combine both rigid mathematical concepts and vaguer linguistic concepts. One of the most popular fuzzy models is the Takagi-Sugeno model [4] (TSM). TSMs use linear consequences in their fuzzy rules, making them equivalent to LLMs under some circumstances. This is also shortly analysed in this paper, but the main point of this paper is the extension of the ELM algorithm. The paper has the following structure: in Section 2, a summary of the ELM approach will be given. Section 3 contains the proposed model with 2 the definition of the SLM. Section 4 provides proof of the approximation capabilities of the SLM and contains the proposed algorithm. Section 5 contains several ways to interpret the SLM. Conclusions are given in Section 7. 2. ELM summary The ELM architecture is similar to a feedforward neural network with one hidden layer. The output neurons have a linear activation function, while the hidden neurons have a nonlinear activation function. The main difference with traditional neural networks is in the tuning of the parameters: the hidden layer parameters are initialised randomly and remain fixed, while only the output weights are tuned. This leads to a linear least-squares problem. Although the original ELM used neural alike hidden nodes, ELMs have been generalised to include many activation functions that are not neuron alike [5]. The most common example is the Radial Basis Function (RBF) 2 g(a, b, x) = e−b||x−a|| , with parameters a ∈ Rn , b ∈ R>0 , and input x ∈ Rn . In this paper, only ELMs with RBF nodes are considered. For a network with h hidden nodes, the output y ∈ Rm of the ELM is: y= h X βi g(ai , bi , x) = i=1 h X βi gi (x), (1) i=1 with βi ∈ Rm the output weight vector connecting the i-th hidden node with the m output nodes. For N inputs xj , this can be written in matrix notation as Y = GB, (2) where and   g1 (x1 ) . . . gh (x1 )   .. .. G=  . . g1 (xN ) . . . gh (xN ) N ×h   y1T   Y =  ...  , T yN N ×m (3)   β1T   B =  ...  . T βh h×m 3 (4) If a similar N × m-matrix Y ∗ consisting of target output values y ∗ is constructed, the goal is to minimise the sum of square errors: E= N X ||yj − yj∗ ||2 = ||Y − Y ∗ ||2F = ||GB − Y ∗ ||2F , (5) j=1 where ||·||F denotes the Frobenius norm. The output weights B that minimise E, can be found by using the Moore-Penrose pseudo-inverse [6], here denoted as †. This also makes sure that the norm of the output weights are minimised. Then the optimal solution is: B̂ = G† Y ∗ . (6) 3. Proposed model The model proposed in this paper is an extension of the ELM model. To define the model, two other definitions are given first: Definition 1. A Linear Model (LM) is a function of the form LMi (x) = αi x + βi , with αi an m × n-matrix and βi ∈ Rm . Definition 2. A Local Linear Model (LLM) is a weighted sum Ph of LMs, where the model weights pi can depend on the input: LLM (x) = i=1 pi (x)LMi (x) P = hi=1 pi (x)(αi x + βi ). Now the proposed model will be defined: Definition 3. A Shuffled Linear Model (SLM) is a LLM with RBF model 2 weights pi (x) = gi (x) = e−bi ||x−ai || , where the parameters ai ∈ Rn and bi ∈ R>0 are drawn from aP continuous probability distribution (independent 2 of the input): SLM (x) = hi=1 e−bi ||x−ai || (αi x + βi ). Note that the SLM is equivalent to the ELM if all αi are zero. It is therefore expected that, for nonzero αi , the approximation capabilities of the SLM are better than those of the ELM if the number of local models is equal to the number of hidden nodes h. This is what will be proved in this paper. 4 4. Approximation capabilities of SLM and proposed algorithm In this section it will be proved that a SLM, as an extension of an ELM, can act as a universal approximator, and an upper bound on the number of models will be given. First, it will be proved that the SLM can act as a universal approximator. After that, several lemmas will be proved that lead to the main theorem of this paper, where Lemma 2 is used in the proof of Lemma 3, and Lemmas 1 and 3 are used in the proof of the main theorem. This section concludes with a learning algorithm for the SLM. 4.1. Universal approximation with a SLM The result of the first theorem follows directly from the approximation capabilities of the ELM: Theorem 1. SLMs are universal approximators. That is, given a continuous target function f : Rm → Rn and a SLM with free parameters αi and βi , i = 1 . . . h, then with probability one the parameters αi and βi can be chosen R in such a way that limh→∞ Rn ||SLM (x) − f (x)||2 dx = 0. Proof. Choosing αi = 0, i = 1 . . . h, the SLM is equivalent to an ELM. For an ELM, the parameters βi can be determined using linear regression to guarantee universal approximation [5, Thm. II.1]. The interesting part is, of course, whether the extension from ELM to SLM gives any improvements. It turns out that a SLM can also be trained using linear regression with all free parameters, just like an ELM. But for the same approximation accuracy, it needs less local models than the number of hidden nodes in the ELM. This is simply due to the fact that a SLM has more free parameters. The remainder of this section is devoted to the proof of these results, after which an algorithm for training the SLM will be presented. 4.2. Necessary lemmas Before presenting the next theorem, some lemmas will be proved. Most proofs are similar to those of the ELM theorems [2, 5]. Lemma 1. Given a SLM with inputs xj and desired outputs yj∗ , j = 1 . . . N , P ∗ 2 the error E = N j=1 ||SLM (xj ) − yj || can be minimised using linear regression. 5 Proof. Let γi = [αi βi ] be the m × (n + 1)-matrix of the free parameters for model i, and let Γ = [γ1 . . . γh ]T be the (n + 1)h × m-matrix of all free xj parameters. Let also zj = , then the output of the SLM can be written 1 as: h X SLM (x) = gi (x)γi z. (7) i=1 Now the error E can be rewritten as follows: E = N X ||SLM (xj ) − yj∗ ||2 j=1 N h X X || gi (xj )γi zj − yj∗ ||2 = j=1 i=1   g1 (xj )zj   .. ∗ 2 = ||ΓT   − yj || . j=1 gh (xj )zj N X (8) = ||ΓT K T − Y ∗ T ||2F , = ||KΓ − Y ∗ ||2F , with and   y1∗ T   , Y ∗ =  ...  ∗ T yN N ×m (9)  g1 (x1 )z1T . . . gh (x1 )z1T   .. .. . K=  . . T T g1 (xN )zN . . . gh (xN )zN N ×(n+1)h (10)  This sum of squared errors can be minimised using linear regression, for example by using the minimum norm least-square solution Γ̂ = K † Y ∗ . This completes the proof. 6 (11) Lemma 2. Given N distinct samples (xj , yj∗ ) ∈ Rn ×Rm , if a ∈ Rn is chosen randomly from a continuous probability distribution, then with probability one we have ||xj − a|| = 6 ||xj 0 − a||, for j 6= j 0 , j, j 0 = 1 . . . N . Proof. The proof is an adaptation to one of the proofs of Huang and Babri [7]. Consider the set V (xj , xj 0 ) = {a ∈ Rn : ||xj − a|| = ||xj 0 − a||}. This set is a hyperplane in Rn . Since this is an (n−1)-dimensional surface in Rn , and there S 0 are only finite j, j , the union U = j6=j 0 V (xj , xj 0 ) is a finite union of (n − 1)dimensional surfaces for j 6= j 0 , j, jR0 = 1 . . . N . So, for any probability density function f , we have P (a ∈ U ) = U f (a)da = 0. Therefore, the probability that ||xj − a|| = ||xj 0 − a|| is 0 for randomly chosen a, and the result of the lemma follows. Lemma 3. For a SLM , matrix K from equation (10) has full rank with probability one if the matrix Z = [z1 . . . zN ]T has full rank. Proof. For i = 1 . . . h, and k = 1 . . . n + 1, the (i × k)-th column of K is given 2 by [gi (x1 )z1k . . . gi (xN )zNk ]T , where gi (x) = g(ai , bi , x) = e−bi ||x−ai || . Let I = (u, v) be any interval from R, with u < v. Let c : I → RN be the curve defined as c(bi ) = [gi (x1 ) . . . gi (xN )]T , (12) seen as a function of bi , for bi ∈ I. Using a similar proof as that of Tamura and Tateishi [8], it can be proved by contradiction that c(bi ) does not belong to a subspace with dimension less than N . Suppose it does, then there exists a vector w 6= 0 orthogonal to this subspace: T w (c(bi ) − c(u)) = N X wj g(ai , bi , xj ) − wT c(u) = 0. (13) j=1 Without loss of generality, assume wN 6= 0, then N −1 X wj 1 T w c(u) − g(ai , bi , xj ). g(ai , bi , xN ) = wN wN j=1 (14) On both sides of this equation are functions of bi . Taking the s-th partial derivative to bi on both sides gives the following equation: N −1 X wj ∂ s ∂s g(a , b , x ) = − g(ai , bi , xj ), i i N ∂bsi wN ∂bsi j=1 7 s = 1, 2, . . . . (15) Now, g is infinitely differentiable w.r.t. bi , with ||2 ∂s g(ai , bi , xj ) ∂bsi = (−||xj − ai ||2 )s e−bi ||xj −ai . Using this fact and the result of Lemma 2, it follows that with probability one, Equation (15) actually contains an infinite number of equations that are linear in parameters wj , for s = 1, 2, . . .. The number of free parameters wj , however, is N . This gives a contradiction. Hence, c(bi ) does not belong to a subspace with dimension less than N . Since the above contradiction holds for bi from any interval I ⊆ R, it is possible to choose h values b1 , b2 , . . . , bh from any continuous probability distribution over R, such that the following matrix G has full rank: G = [c(b1 ) c(b2 ) . . . c(bh )]. (16) This is the same N × h matrix as the hidden layer output matrix from equation 3. The matrix Z = [z1 . . . zN ]T is an N × (n + 1)-matrix. Let M = G ⊗ Z be the Kronecker product of these matrices. Then M is an N 2 × (n + 1)h-matrix. Since both G and Z have full rank, and for matrices the general rule rank(A ⊗ B) = rank(A)rank(B) holds, the matrix M has rank r = min(h, N ) min(n + 1, N ). Now, matrix K appears actually inside matrix M . For matrix M , we have:   g1 (x1 )z1T . . . gh (x1 )z1T .. ..     . .  T T   g1 (x1 )zN . . . gh (x1 )zN    .. .. . (17) M =G⊗Z = . .    g (x )z T . . . g (x )z T   1 N 1 h N 1    .. ..   . . T T g1 (xN )zN . . . gh (xN )zN The N × (n + 1)h-matrix K from equation (10) follows from this matrix by removing rows until only the rows with indices (j − 1)(N + 1) + 1 are left, for j = 1, . . . , N . It follows that the rank of matrix K is equal to min(N, r) = min(N, (n + 1)h), so matrix K has full rank. 4.3. Improvement of the SLM as an extension of an ELM Using the Lemmas that appeared in this section, the main theorem of this paper, which is similar to that of Huang [2, Thm. 2.2] can be presented. 8 From the proofs of the Lemmas, the following definitions are needed:   xj zj = , 1 Z = [z1 . . . zN ]T . (18) (19) Theorem 2. Given any  > 0, and given N distinct samples (xj , yj∗ ) ∈ Rn × Rm , if matrix Z has full rank, then there exists h ≤ N/(n + 1) such that aPSLM with h local models can be trained using linear regression to get ∗ 2 <  with probability one. E= N j=1 ||SLM (xj ) − yj || Proof. The error E can actually become zero by choosing h = (n + 1)N . Since Z has full rank, from Lemma 3 it follows that matrix K also has full rank, with probability one. Then K is also invertible, since the size of K is N ×(n+1)h = N ×N . The solution showed at the end of the proof of Lemma 1 is now actually equal to Γ̂ = K −1 Y ∗ , and the error E = ||KΓ − Y ∗ ||2F is zero. Since zero error can be achieved by choosing h = (n + 1)N , there always exists a h ≤ (n + 1)N to let the error be as small as desired. Note that the condition that Z has full rank can be easily satisfied, for example by adding noise to the training samples. Theorem 2 shows that the SLM needs less local models than the number of hidden nodes in an ELM. This does not necessarily imply a decrease in computation time, since the matrix of which a pseudo-inverse needs to be computed, has size N ×(n+1)h, not size N ×h. An ELM with h hidden nodes and a SLM with h/(n+1) local models are therefore comparable in both approximation capabilities and in computational efficiency. The construction of the matrix for which a pseudoinverse needs to be computed, however, can be done more efficiently due to the block structure of K, so there is some gain in computational efficiency for a large number of models or hidden neurons. There might also exist a more efficient way to calculate the pseudo-inverse of this block matrix. It is questionable whether it is desired to use less nodes or models in an approach that uses random nodes and models, since the model performance should not depend highly on parameters that are chosen randomly. Still, there are several advantages of the SLM, compared to an ELM. As mentioned above, the matrix K can be constructed efficiently because it contains a block structure, and maybe in the future an efficient method to calculate the pseudo-inverse of this matrix could be found. Besides this, the 9 SLM is not as black-box as an ELM because the output matrix Γ actually shows the direct (linear) relations between input and output near several operating points of the input space, whereas the output weights of an ELM only show relations between the less interpretable hidden layer and the outputs. Finally, the SLM allows for a clear physical interpretation due to its similarities with local linear models from the area of system identification, and its similarities with fuzzy inference models; see Sections 5.3 and 5.4. 4.4. Proposed algorithm The proposed SLM learning algorithm is shown in Figure 1. After running the output of the SLM for input x is given by SLM (x) =  the algorithm,  g1 (x)z   x  . T  .. Γ  .  , with z = 1 gh̃ (x)z SLM Algorithm 1. Given N input-output pairs (xj , yj∗ ) ∈ Rn × Rm , randomly generate parameters ai ∈ Rn and bi ∈ R>0 from a continuous probability distribution independent of the input. 2 2. Choose the number of models h and let gi (x) = e−bi ||x−ai || for i = 1, . . .  , h.  xj 3. For zj = , calculate the N × (n + 1)h-matrix 1   g1 (x1 )z1T . . . gh (x1 )z1T   .. .. K= . . . T T g1 (xN )zN . . . gh (xN )zN ∗ T 4. Put the desired outputs yj∗ in matrix Y ∗ = [y1∗ . . . yN ] and calculate Γ = K † Y ∗ by using the Moore-Penrose pseudo-inverse. Figure 1: Proposed SLM algorithm. 10 5. Interpretation of the model Several interpretations of the SLM are possible. These interpretations will be described in the following subsections. 5.1. ELM interpretation The main interpretation of the SLM in this paper is the extension of an ELM with RBF nodes. The output weights βi of an ELM can be seen as constant functions, since they do not depend directly on the input. After training an ELM, the output weights are the same for every possible input of the network. In the SLM extension, the output weights are not constants, but functions of the input. Instead of multiplying the outputs of the hidden layer gi (x) with constant output weights βi , they are multiplied with a linear function of the input: αi x+βi . This also changes the interpretation of output weight to linear model, and from hidden layer output to model weight. Since only the parameters in the hidden layer are fixed, a SLM has more free parameters than an ELM. The vector βi is of size m and the matrix αi is of size m×n. If h1 is the number of local models in a SLM and h2 the number of hidden neurons in an ELM, this implies that the SLM has h1 · m · (n + 1) free parameters, while an ELM has only h2 · m free parameters. A logical conclusion is that an ELM needs n + 1 times as many neurons as the number of local models in a SLM, for similar approximation capabilities. This is true under the conditions shown in Theorem 2. 5.2. RBF interpretation A SLM can be interpreted as a RBF network, but there are some important differences. In the first place, it is a generalisation of the RBF network. Several generalisations of RBF networks exist [9, 10], but the one used in this paper is similar to the one proposed by Hunt, Haas and Murray-Smith [11]. Using the notation from this paper, a standard RBF network can be deP noted as y = hi=1 βi gi (x), while a generalised RBF network is denoted as P y = hi=1 θi (x)gi (x), letting the output weights depend on the input 1 . Using θi (x) = αi x + βi , where the output weights are a linear function of the input, this is equivalent to a SLM. 1 The generalised RBF as defined by Hunt, Haas and Murray-Smith, differs from a standard RBF network in three aspects. Only one of these aspects is considered here, namely that the output weights can depend on the input. 11 In the second place, a SLM differs from a RBF network in the way the hidden layer parameters are chosen. It is customary in RBF networks to let the location of the centers of the Gaussians ai be the same as some of the training samples xj , or else to determine them by some sort of clustering algorithm. In a SLM, both the centers ai as well as the widths bi of the Gaussians are taken from a continuous probability distribution, and are therefore independent of the training samples 2 . This has the advantage that no prior information about the training samples is necessary, and that it takes practically no time to determine the parameters. 5.3. TSM interpretation The generalised RBF network has been shown to be equivalent to the Takagi-Sugeno fuzzy model [4] under some circumstances [11]. The TakagiSugeno model (TSM) consists of a number of fuzzy if-then rules with a fuzzy premise and a linear consequence, e.g. Ri : if x1 is A1 ∧ x2 is A2 ∧ · · · ∧ xm is Am then θi (x) = αi x + βi (20) The circumstances under which the TSM is equivalent to the generalised RBF network, and therefore also to the SLM, are: 1) the number of fuzzy rules is equal to the number of RBF units, or local models for the SLM, 2) the membership functions within each rule are all Gaussians, and 3) the operator used for the fuzzy ∧ is multiplication. Using multiplication for the if-then part as well, and letting gik (x) denote the fuzzy membership function of the fuzzy set Ak of fuzzy rule Ri , this result is straightforward: T SM (x) = h Y n X gik (xk )(αi x + βi ) i=1 k=1 = h Y n X 2 e−bi (xk −aik ) (αi x + βi ) i=1 k=1 2 The probability distribution of the parameters ai and bi can be any continuous probability distribution, for example the normal or the uniform distribution. However, the distribution of bi should be such that P (bi ≤ 0) = 0, since bi ∈ R>0 . In theory this is the only constraint, but in practice there might be more constraints due to rounding errors in the numerical evaluations of the Gaussians. 12 = = h X i=1 h X e−bi Pn k=1 (xk −aik )2 (αi x + βi ) (21) 2 e−bi ||x−ai || (αi x + βi ) i=1 = SLM (x). A SLM can therefore be seen as a TSM with randomly generated Gaussian membership functions for the premises of the fuzzy rules. This equivalence relation gives the following result: Corollary 1. Theorem 2 also holds for the Takagi-Sugeno fuzzy model with h fuzzy rules, if the membership functions for each rule are Gaussians with parameters chosen randomly from continuous probability distributions, and the operator used for the fuzzy ∧ is multiplication. 5.4. LLM interpretation LLMs like in definition 2 usually appear in the context of nonlinear dynamical systems, where a complex (nonlinear) system is approximated by taking a weighted sum of simpler (linear) models. These linear models can be seen as linearisations of the nonlinear system around an operating point, and the weights are used to describe the transitions between the operating points. The problem of using a LLM to approximate a nonlinear dynamical system, belongs to the area of system identification. Several overviews of identification techniques for local models exist [12, 13]. The input of such a dynamical LLM consists of past inputs and outputs or states of the system, while the output of the dynamical LLM should give the next output or state of the system, e.g. st+1 = f (st , ut ), (22) where s is the output or state, and u the control input of the dynamical system. If the function f is differentiable, it can be linearised around an operating point (s(i) , u(i) ), to get: f (st , ut ) ≈ f (s(i) , u(i) ) + ∂f (i) (i) (s , u )(st − s(i) ) ∂st ∂f (i) (i) (s , u )(ut − u(i) ) ∂ut = Ai st + Bi ut + Oi , + 13 (23) where Ai and Bi are the partial derivatives of f w.r.t. s and u respectively, and Oi = f (s(i) , u(i) ) − Ai s(i) − Bi u(i) . The local linear model is then a weighted sum of these linearisations: LLM (st , ut ) = h X gi (st , ut )(Ai st + Bi ut + Oi ), (24) i=1 where gi is a scalar-valued function that represents the validity of the local model. The relationship between these kinds of dynamical LLMs and TSMs has also been investigated [14]. Going out of the area of dynamical systems, if the input of the function f is denoted as x rather than (st , ut ), the LLM of equation 24 is equal to the one from Definition 2. Now, for a SLM, the weights are Gaussian functions of the input, but the parameters of the Gaussians (i.e. their centers and widths, or the operating points and regimes) are chosen randomly. This implies that the local models are supposed to be good approximations of the global system around randomly chosen operating points, while the traditional approach is to use a gradient-descent type algorithm to determine the operating points and regimes. Theorem 2 shows that even with randomly chosen operating points and regimes, it is possible to approximate any function with any desired accuracy, as long as the number of local models h is high enough, and that this can be done with a direct algorithm. 6. Experiments and results As an example, in this section the SLM algorithm is compared with the ELM algorithm for modelling the following dynamical system: dx1 = x2 , dt dx2 = λ(x21 − 1)x2 − x1 . dt (25) These are the equations of a Van der Pol oscillator. It is a nonlinear system with a limit cycle, a phenomenon that does not occur in linear systems. The experiment is implemented in Matlab, on a computer with a 2.80 GHz processor. 14 6.1. Data The model inputs are the x1 and x2 values of the Van der Pol oscillator with λ = 1 for timesteps t = 1, . . . , 1000, while the desired model outputs are the x1 and x2 values for timesteps t + 1. The values are computed with the Euler forward method. The data is seperated in three phases: the learning phase, the generalisation phase, and the simulation phase. In each phase, the dynamical system is run for 10 different initial conditions, giving 104 data samples. 6.2. Models used The data is tested on a SLM with 100 local models. The parameters ai are drawn from a zero-mean normal distribution with variance 2, while the parameters bi are drawn from a uniform distribution over (0, 1). The ELM uses the same distributions for the parameters, but with 300 hidden neurons since h(n + 1) = 300 in this case (see Section 5.1 for a comparison between the number of hidden neurons in an ELM and the number of local models in a SLM). The experiment consists of three phases: a learning phase, a generalisation phase and a simulation phase. In the learning phase, the SLM is trained using the algorithm from Figure 1, and the ELM is trained using the ELM algorithm proposed by Huang et al. [2]. In the generalisation phase, only the inputs of the system are used, and the models have to produce the outputs using the model parameters obtained during the training phase. In the simulation phase, only the initial conditions of the system are given, and the models have to use the model output of the previous timestep as an input for future timesteps, also using the model parameters obtained during the training phase. 6.3. Results and discussion The whole process has been repeated 100 times. The average results are shown in Table 1, where MSE stands for Mean Squared Error. The output of the training phase of the last run is shown in Figure 2, for the SLM case only. The ELM case showed similar results, as well as the generalisation and simulation phases for both algorithms. From the results we see that the SLM algorithm takes less computation time. As expected, this is not due to the calculation of the pseudo-inverse, which has actually become slower in the SLM algorithm (the reason for this is 15 unclear). The construction of the matrix that is to be inverted has been done more efficiently in the SLM approach by making use of its block structure. The errors during all three phases are similar in both methods, while the number of local models for the SLM was three times less than the number of hidden neurons for the ELM. This is in line with the theory presented in this paper. Table 1: Comparing the errors and computation time of the SLM and ELM algorithm for the modelling of a Van der Pol oscillator. Total computation time Computation time for pseudo-inverse MSE during training MSE during generalisation MSE during simulation Average over 100 SLM tests 9.0952 sec. Average over 100 ELM tests 17.8737 sec. 1.8038 sec. 1.5514 sec. 2.3872 · 10−11 2.2645 · 10−10 1.3108 · 10−11 3.1529 · 10−9 7.6113 · 10−5 7.5266 · 10−5 7. Conclusions An extension to Extreme Learning Machines (ELM) was proposed in this paper, together with an algorithm and theorems concerning the approximation capabilities of the Random Local Linear Model (SLM). Under certain conditions, the SLM can achieve similar approximation capabilities as the ELM while using less local models than the number of hidden nodes of an ELM. Although this is not necessarily an advantage, the SLM allows for a clear physical interpretation and can be run more efficiently than an ELM when the number of local models and hidden nodes is high. The SLM is similar to Local Linear Models from the area of system identification and, under some circumstances, to the well-known Takagi-Sugeno Model from the area of fuzzy modelling, for which an upper bound on the number of randomly generated rules is given. This also makes them more fit 16 Figure 2: Modelling a Van der Pol oscillator with a Random Local Linear Model. The black dots are the locations of the centers of the Gaussians (parameters ai ), which are distributed randomly. Around each black dot, a certain linear model is valid. Adding the linear models together in a nonlinear fashion results in a global model that accurately approximates the nonlinear dynamics of the oscillator. to be used for modelling dynamical systems. Compared to these techniques, the training of a SLM can be done more efficiently, since the operating points or fuzzy rules can be chosen randomly and independent from the input, which implies that less prior knowledge is required. The training algorithm then uses linear regression, which can be performed much faster than most other parameter estimation methods. The theorems and proofs of this paper could be extended to more general cases. However, the link with system identification and fuzzy modelling techniques already makes sure that the SLM could be an alternative to some very popular techniques, and it allows much room for further investigation. 17 Acknowledgment This research was performed at Almende B.V. in Rotterdam, the Netherlands. The author would like to thank Anne van Rossum and Dimitri Jeltsema for their supervision during this research, and Giovanni Pazienza for providing valuable feedback on this paper. References [1] O. Nelles, Nonlinear system identification: from classical approaches to neural networks and fuzzy models, Springer, 2001. [2] G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning machine: theory and applications, Neurocomputing 70 (1) (2006) 489–501. [3] G.-B. Huang, L. Chen, C.-K. Siew, Universal approximation using incremental constructive feedforward networks with random hidden nodes, Neural Networks, IEEE Transactions on 17 (4) (2006) 879–892. [4] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, Systems, Man and Cybernetics, IEEE Transactions on (1) (1985) 116–132. [5] G.-B. Huang, L. Chen, Convex incremental extreme learning machine, Neurocomputing 70 (16) (2007) 3056–3062. [6] C. R. Rao, S. K. Mitra, Generalized inverse of a matrix and its applications, J. Wiley, New York. [7] G.-B. Huang, H. A. Babri, Upper bounds on the number of hidden neurons in feedforward networks with arbitrary bounded nonlinear activation functions, Neural Networks, IEEE Transactions on 9 (1) (1998) 224–229. [8] S. Tamura, M. Tateishi, Capabilities of a four-layered feedforward neural network: four layers versus three, Neural Networks, IEEE Transactions on 8 (2) (1997) 251–255. [9] S. A. Billings, H.-L. Wei, M. A. Balikhin, Generalized multiscale radial basis function networks, Neural Networks 20 (10) (2007) 1081–1094. 18 [10] F. Fernández-Navarro, C. Hervás-Martı́nez, J. Sanchez-Monedero, P. A. Gutiérrez, Melm-grbf: a modified version of the extreme learning machine for generalized radial basis function neural networks, Neurocomputing 74 (16) (2011) 2502–2510. [11] K. J. Hunt, R. Haas, R. Murray-Smith, Extending the functional equivalence of radial basis function networks and fuzzy inference systems, Neural Networks, IEEE Transactions on 7 (3) (1996) 776–781. [12] R. Murray-Smith, T. A. Johansen, Multiple model approaches to nonlinear modeling and control (1997). [13] V. Verdult, Non linear system identification: a state-space approach, Twente University Press, 2002. [14] T. A. Johansen, R. Shorten, R. Murray-Smith, On the interpretation and identification of dynamic takagi-sugeno fuzzy models, Fuzzy Systems, IEEE Transactions on 8 (3) (2000) 297–313. Laurens Bliek received the B.Sc. degree in applied mathematics at Delft University of Technology in 2011, and is currently graduating for the M.Sc. degree of the same study. His minor, B.Sc thesis, and many of his electives, had to do with artificial intelligence and neural networks. His research interests include neural networks, dynamical systems theory and optimisation. 19
9cs.NE
1 On the Secure and Reconfigurable Multi-Layer Network Design for Critical Information Dissemination in the Internet of Battlefield Things (IoBT) arXiv:1801.09986v1 [eess.SP] 26 Jan 2018 Muhammad Junaid Farooq, Student Member, IEEE and Quanyan Zhu, Member, IEEE Abstract—The Internet of things (IoT) is revolutionizing the management and control of automated systems leading to a paradigm shift in areas such as smart homes, smart cities, health care, transportation, etc. The IoT technology is also envisioned to play an important role in improving the effectiveness of military operations in battlefields. The interconnection of combat equipment and other battlefield resources for coordinated automated decisions is referred to as the Internet of battlefield things (IoBT). IoBT networks are significantly different from traditional IoT networks due to battlefield specific challenges such as the absence of communication infrastructure, heterogeneity of devices, and susceptibility to cyber-physical attacks. The combat efficiency and coordinated decision-making in war scenarios depends highly on real-time data collection, which in turn relies on the connectivity of the network and information dissemination in the presence of adversaries. This work aims to build the theoretical foundations of designing secure and reconfigurable IoBT networks. Leveraging the theories of stochastic geometry and mathematical epidemiology, we develop an integrated framework to quantify the information dissemination among heterogeneous network devices. Consequently, a tractable optimization problem is formulated that can assist commanders in cost effectively planning the network and reconfiguring it according to the changing mission requirements. Index Terms—Battlefield, epidemics, internet of things, multiplex networks, Poisson point process, random geometric graph. I. I NTRODUCTION The Internet of things (IoT) is an emerging paradigm that allows the interconnection of devices which are equipped with electronic sensors and actuators [2]. It allows for a higher level of situational awareness and effective automated decisions without human intervention. The concept has proven to be extremely useful in applications such as smart homes, energy management, smart cities, transportation, health care, and other domains [3]. Recently, there is an interest in the Preliminary results have been presented at the 15th Intl. Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt 2017) [1]. This research is partially supported by a DHS grant through Critical Infrastructure Resilience Institute (CIRI), grants CNS-1544782 and SES1541164 from National Science of Foundation (NSF), and grant de-ne0008571 from the Department of Energy (DOE). The statements made herein are solely the responsibility of the authors. Muhammad Junaid Farooq and Quanyan Zhu are with the Department of Electrical & Computer Engineering, Tandon School of Engineering, New York University, Brooklyn, NY, USA, E-mails: {mjf514, qz494}@nyu.edu. Fig. 1: A typical IoBT network with heterogeneous battlefield things and random enemy attacks. Battlefield devices interact with other devices within their communication range in a D2D manner for exchange of mission critical information. The link shape illustrates the piece of information being shared. defence community to leverage the benefits enabled by the IoT to improve the combat efficiency in battlefields and effectively manage war resources [4], [5]. This emerging area of using IoT technology for defence applications is being referred to as the Internet of battlefield things (IoBT) [6], [7]. Fig. 1 illustrates a typical battlefield comprising of heterogeneous objects, such as soldiers, armoured vehicles, and aircrafts, that communicate with each other amidst cyber-physical attacks from the enemy. The IoBT has the potential to completely revolutionize modern warfare by using data to improve combat efficiency as well as reduce damages and losses by automated actions while reducing the burden on human war-fighters. Currently, the command, control, communications, computers, intelligence, surveillance and reconnaissance (C4 ISR) systems use millions of sensors deployed on a range of platforms to provide situational awareness to military commanders and troops, on the ground, seas, and in the air [8]. However, the real power lies in the interconnection of devices and sharing of sensory information that will enable humans to make useful sense of the massive, complex, confusing, and potentially deceptive ocean of information [9]. In the battlefield scenarios, the communications between strategic war assets such as aircrafts, warships, armoured vehicles, ground stations, and soldiers can lead to improved coordination, which can be enabled by the IoBT [10]. However, to become a reality, this vision will have to overcome several technical limitations of current information systems and networks. Most civilian IoT applications such as smart homes and cities are infrastructure based, where the devices are connected to each other and the internet via an access point or gateway. In the battlefield scenario, the communication infrastructure such as cellular networks or base stations may not be available. Hence, the battlefield things need to exploit deviceto-device (D2D) communications [11] to share information with other things1 . Therefore, the information sharing can be affected by the physical parameters of the network such as the transmission power of the things, the number of deployed things, their locations, and the flexibility of communication with other types of things. In addition to these factors, another impediment in the connectivity of battlefield things is the susceptibility to cyber-physical attacks. The information exchange between things may be affected by several factors such as jamming of radio frequency (RF) channels, physical attacks on infrastructure, node failures due to attacks on power sources, or lack of power, etc. Since the analytics and automated decisions in an IoBT network rely on the real-time data provided by the sensors deployed in the battlefield, we need to ensure the spread of information in the networks with a certain level of reliability and security to make accurate decisions. Although the IoBT has to do to with much more beyond the connectivity of battlefield things, such as digital analytics and automated response, which allows the systems to react more quickly and precisely than humans; however, the connectivity aspect is vital in allowing the IoBT systems to unleash their full potential. It is ideal if we can achieve complete situational awareness and perfect information spreading throughout the network. However, due to limited available resources, incurred costs (capital and operational), and vulnerability to attacks, it is not practical and hence requires a cost-effective, secure and reconfigurable network design. Therefore, in this paper, we first characterize the information dissemination in an IoBT network under vulnerabilities, and use it to design the network parameters to reliably achieve mission specific performance goals with minimal amount of resources. We then present a reconfigurable framework that adapts with the changing connectivity situation of the network. A. Related Work Stochastic geometry (SG) based models have been successfully used in the modeling and analysis of traditional wireless networks such as cellular networks [12] and ad-hoc wireless sensor networks [13]. These models accurately capture the effect of spatial distribution of network devices on their connectivity and performance. However, they lack the capability of analysing the dynamics of information dissemination in communication networks. One area of research related to this work is the study of probabilistic message broadcasting and percolation in random graphs [13], [14]. Research in this direction has focused on identifying the conditions for the existence of giant connected components, and the analysis of connectivity. On the other hand, mathematical epidemiology [15] has been studied extensively for analyzing the spread of viruses in computers, rumours in humans, and infectious diseases in biological networks [16]–[18]. Although the epidemic models 1 We use the terms “things” and “devices” interchangeably to refer to battlefield things throughout the paper. are useful in capturing the diffusion of pathogens or rumors in a population, they do not take the geometry of the network into account and hence, cannot give meaningful insights in physically deployed communication networks. This research aims to investigate the dynamic decision-making and control of information dissemination and the design of multi-layer communication networks. Some attempts have been made to study the dynamical information dissemination in wireless networks such as [19], [20]. However, they focus on traditional wireless adhoc networks, which do not capture the unique characteristics of IoBT networks. Moreover, there are few descriptive models available in literature for designing IoT networks, most of which are developed for civilian applications [21], and are not applicable to IoT networks over battlefields. Since IoBT networks are envisioned to exploit D2D communications and devices may be equipped with multiple radios for tactical communications, the information dissemination is directly associated with the communication characteristics of network devices as well as their associated deployment densities. Hence, it is imperative to develop an integrated design framework that can capture the specific features of IoBT networks. IoBT networks are composed of heterogeneous devices which may connect to more than one communication network based on the hardware capabilities. This requires a multi-layer network model that is able to incorporate the connectivity of such devices. In epidemiology literature, the multi-layer networks have been studied in the context of communicationcontact layer networks [22]. These are networks in which there is physical contact between the agents in one layer that allows for disease spreading and communication contact on the other layer, e.g., via telephone or social network, that allows for awareness spreading. The diffusion in such multiplex networks [23] is asymmetric and the interplay between the networks has been studied in [24]. The dynamical process of multiple diseases spreading in multiplex networks is provided in [25]. However, our focus is on symmetric multilayer communication networks that have different types of information propagating simultaneously in different network layers. Moreover, the physical connectivity in each layer has a certain structure based on the communication characteristics of the devices as opposed to complex networks that have nontrivial topological features. In this paper, we develop a specialized framework for information dissemination over IoBT networks that utilizes the connectivity structure in the multilayer heterogeneous network by making use of the existing works in epidemiology. Some initial results of the developed framework have been presented in [1]. However, the results are based on the assumption of reciprocity of communication between devices of one layer that are under the influence of the devices of other network layers. In this work, we lift this assumption and present a practical network model consisting of multi-radio devices that forms a multiplex communication network. We believe that such network models are well-suited to battlefield communications. B. Contributions In this paper, we develop a SG based model to characterize the connectivity of IoBT networks in terms of the degree distribution of the devices. We then use an epidemic spreading model to capture the dynamic diffusion of multiple messages within the network of devices at the equilibrium state. The resulting integrated open-loop system model is used as a basis for reconfiguring the network parameters to ensure a missiondriven information spreading profile in the network. This paper attempts to bridge the gap between the spatial stochastic models for wireless networks and the dynamic diffusion models in contact-based biological networks to derive new insights that aid in the planning and design of secure and reliable IoBT networks for mission critical information dissemination. The developed framework, with some modifications, is also applicable to the more general class of heterogeneous ad-hoc networks. The main contributions of this paper are summarized as follows: • • • • A novel multiplex network model for IoBT networks is proposed that helps in characterizing the intra-layer and network-wide connectivity of heterogeneous battlefield devices by considering the spatial randomness in their locations. A tractable framework is developed for quantification of simultaneous information dissemination in the multiplex IoBT network based on mathematical epidemiology that considers the perceived level of threat to the network from cyber-physical attacks. Approximate closed form results relating the proportion of informed devices at equilibrium and the network parameters are provided. An optimization problem is formulated that can assist military commanders in identifying the physical network parameters that are required in order to sufficiently secure the network from the perceived attacks. It can also help in reconfiguring existing networks to achieve a desired level of communication reliability. A detailed investigation of the developed integrated framework is provided for particular battlefield missions and the effect of threat level and performance thresholds is studied. II. S YSTEM M ODEL In this section, we first describe the geometry of the IoBT network and propose a bi-layer abstraction model using tools from stochastic geometry. Then, we develop a dynamic model to characterize information dissemination in the heterogeneous IoBT network based on mathematical epidemiology. A. Network Geometry We consider uniformly deployed heterogeneous battlefield things in R2 that can be abstracted as a Poisson Point Process (PPP)2 [26] with intensity λ devices/km2 , referred to as 2 While the devices can be placed more strategically according to their characteristics and utility, however, due to potential mobility and difficulty of tracking network topology, these optimal locations may not be known or fixed, which justifies the PPP assumption. Φ = {Xi , Ti }i≥1 , where Xi and Ti represents the location and type of the ith device respectively. We assume that the network is composed of two types of devices, i.e., Ti ∈ {1, 2}, ∀i ≥ 1. The first type of devices, i.e., Ti = 1, referred to as type-I devices, are equipped with two radio interfaces. The second type of devices, i.e., Ti = 2, referred to as type-II devices, have only one radio interface that is compatible with type-I devices. In the first network layer, only type-I devices can communicate with other type-I devices, while in the second layer, both typeI and type-II devices can communicate with each other due to the availability of common radios. Assuming that each device can be of type-I with probability p, i.e., P(Ti = 1) = p, ∀i ≥ 1, then the set of active devices in the first network layer can be represented by a PPP Φ1 = {Xi ∈ Φ : Ti = 1} with intensity λ1 = pλ, that is obtained by an independent thinning of the original point process Φ. On the other hand, since all the devices in the second network layer can communicate with each other, so the active devices can be represented by a PPP Φ2 = Φ, with intensity λ2 = λ. Note that the devices can move independently and we assume that the placement of devices as a result of mobility remains uncorrelated, i.e., can be represented by a new realization of the original PPP, thus resulting in a quasi-static network. This type of network configuration is particularly wellsuited to IoBT networks, where type-I devices may correspond to ground stations or armored vehicles equipped with multiple types of radios, while type-II devices may correspond to soldiers equipped with single radio smart mobile devices. The equipped radios on the devices are characterized by their transmission power or equivalently, the communication range rm in meters, m ∈ {1, 2}. Type-I devices have two radios with transmission ranges r1 and r2 , respectively, while typeII devices have one radio with transmission range r2 . The communication range of the radios is tunable in the interval min max min min max , ∀m ∈ {1, 2} ≥ rm ≥ 0 and rm ,rm ], where rm [rm and we assume that r2 ≤ r1 . Due to the absence of traditional communication infrastructure such as base stations, the devices are only able to communicate using D2D communications, i.e., devices x, y ∈ Φm are connected to each other in network layer m if kx − yk ≤ rm , m ∈ {1, 2}, where k.k represents the Euclidean distance. Similarly, devices x ∈ Φ1 and y ∈ Φ\Φ1 can communicate with each other only if kx−yk ≤ r2 . Hence, the communication links between devices in each layer can be modeled using a random geometric graph (RGG) [27] with a connection radius of rm , m ∈ {1, 2}. Each of the layers of the multiplex network has a different connectivity that depends on the respective device densities and the communication ranges. An illustrative representation of the network model is provided in Fig. 2 that shows the connectivity between type-I devices in the first layer and the connectivity between all devices in the second layer. In the subsequent subsection, we analyze the connectivity of the devices in each layer, referred to as intralayer connectivity and the connectivity of the overall network, referred to as network-wide connectivity. Type 1 device Type 2 device m=1 r1 m=2 r2 Fig. 2: Heterogeneous IoBT network decomposed into virtual connectivity layers. The blue region illustrates the communication reach of type-I devices in layer 1 while the red region illustrates the communication reach of type-I and typeII devices in layer 2. B. Network Connectivity In this subsection, we describe the connectivity between the heterogeneous devices in an IoBT network. The connectivity of devices can be classified into intra-layer and network-wide connectivity, which are explained as follows: 1) Intra-layer Connectivity: Within a particular network layer m, the active3 devices can communicate with each other if they are within a distance of rm of each other. Referring to Fig. 1, the dotted lines between battlefield vehicles represent the connectivity of the devices in the first layer. Similarly, the solid lines between soldiers and vehicles represent the connectivity of the devices in the second layer. The set of communication neighbours of a typical device x ∈ Φm in layer m, can be expressed as Nm (x) = {y ∈ Φm : kx − yk ≤ rm }. The connectivity of the RGG formed by devices in layer m is characterized by the degree of the devices, denoted by Km , which is defined as the number of neighbours of each device, i.e., Km = |Nm (x)|, where |.| represents the set cardinality. Since the network is spatially distributed as a PPP, the degree of each device in the RGG is a Poisson random variable [13]. Therefore, the resulting intra-layer degree distribution of a typical device can be expressed by the following lemma. Lemma 1. The intra-layer degree of a typical device in each network layer is distributed as follows: ( 2 (1 − p) + pe−λ1 πr1 , if k = 0, (1) P(K1 = k) = 2 k 2 (λ πr ) if k > 0, pe−λ1 πr1 1 k! 1 , (λ2 πr22 )l , l ≥ 0, (2) l! for sufficiently large λ1 , λ2 , r1 and r2 . The average degree of a typical device in the two network layers can be expressed as follows: E[K1 ] = pλ1 πr12 , (3) 2 P(K2 = l) = e−λ2 πr2 E[K2 ] = λ2 πr22 . Proof. See Appendix A. (4) From Fig. 2, it is clear that the average degree, or equivalently the connectivity of devices in each layer depends on the density of the deployed devices as well as the communication range. The joint probability distribution of the connectivity of a typical device in each layer is denoted by P(K1 = k, K2 = l). 3 Type-I devices are active in both network layers while type-II devices are only active in the second network layer. 2) Network-wide Connectivity: If the two network layers are collapsed together to form a single virtual network such that both layers reinforce the connectivity of the devices, then the connectivity is characterized in terms of the combined degree denoted by Kc . The combined-layer degree of a typical device x ∈ Φ is defined as Kc = |N1 (x)| + |N2 (x)|. Since the degree of the devices in each layer is Poisson distributed, the combined-layer degree follows a Poisson mixture distribution expressed by the following lemma. Lemma 2. The average combined-layer degree of a typical device can be expressed as follows: E[Kc ] = pλ1 πr12 + λ2 πr22 . (5) Proof. See Appendix B In the following section, we describe the information dissemination over such bi-layer networks that have the above mentioned connectivity profile. III. I NFORMATION D ISSEMINATION Each type of device in the IoBT network generates data that needs to be propagated to other devices of the same type and/or different types of devices depending on the role of that device. There are certain pieces of information that needs to be shared among the same type of devices, e.g., soldiers need to communicate information with other soldiers and similarly commanding units might also share information amongst them. Henceforth, we refer to the information sharing in each network layer as intra-layer information dissemination. On the other hand, some information might be important for all network nodes such as network health monitoring data or network discovery beacons. This is henceforth referred to as network-wide information dissemination. We assume a time slotted system where the duration of each slot is τ s. During each time slot, the informed devices broadcast (i) information to their neighbors at a rate of γ. Let Ps ∈ [0, 1] be the average probability that the transmitted information type i ∈ {1, 2, c} by a typical device is successfully received by its neighbors, referred to as the success probability, and δ ∈ [0, 1] be the probability that the communication will be affected by cyber-physical attacks. Since the the event of successful transmission due to interference from other devices and the event of a cyber-physical attack are independent, therefore the effective probability of a successful transmission (i) can be expressed as (1−δ)Ps . Consequently, the information spreading rate between devices, denoted by α(i) , i ∈ {1, 2, c}, can be expressed as follows: α(i) = γ(1 − δ)Ps(i) , (6) We refer to δ as the threat level as it signifies the perceived risk in information transmission between devices. Without loss of generality4 , we can select the contact rate γ = 1, so effectively, α(i) is the probability of successful information transmission between devices. In essence, α(i) can be interpreted as the desired security level from the perceived threats to the communication network. The probability of successful transmission can be computed by setting a threshold on the received signalto-interference-plus-noise-ratio (SINR) at a typical device. 4 There is no loss of generality since τ can be made arbitrarily small. Several techniques can be used from SG literature to accurately characterize the SINR based success probability depending on the medium access protocol used [28]. We use a generalized (i) representation of the success probability Ps = gi (p, λ, r) in terms of the densities and communication ranges of the devices, where the function gi (.) is assumed to be monotone in its arguments. The parameter δ can capture a broad range of cyber-physical threats in IoBT networks. Different methods can be used to assess the threat level in battlefields due to jamming, physical attacks, and other adversarial actions based on historical data and/or statistical models of attack types, some of which are explored in existing works such as [29] and [30]. For instance, to model jamming attacks, the parameter δ can be based on the SINR, in which case the RGG becomes an interference graph [31]. To tackle physical network attacks such as targeted attacks, the parameter δ can be based on the density of device deployment, the connectivity of devices, or the type of devices. Furthermore, an integrated metric can also be developed that can simultaneously capture a multitude of threats. However, developing such a comprehensive metric is beyond the scope of the current work and we use a generic threat level for illustrative purposes. Over time, the adversarial attacks may compromise or negatively impact significant portion of the network connectivity. So there is a need for a resilient framework that can be reconfigured to recover from the lost connectivity by cyber-physical attacks. Another important aspect of the information dissemination process is to account for the information annihilation at each time step. There are several reasons that a device may not broadcast the information that it has received from another device in the previous time slot such as limited buffer capacity and misclassifying information as unimportant. However, the most important factor is to ensure propagating the most recent information in the network. This dynamical information spreading process in an IoBT network can be formalized using the susceptible-infected-susceptible (SIS) model [32], which is well studied in mathematical epidemiology. The challenge is that information propagates over a topology in wireless communication networks, while the classical SIS model does not deal with topological constraints. Furthermore, we deal with simultaneous information dissemination in multiple network layers which presents additional challenges. In the following subsection, we describe the dynamics of the information dissemination process. A. Information Dynamics sion. To model this behaviour and explain the dynamics of information dissemination across the IoBT network, we exploit the susceptible-infected-susceptible (SIS) model [32] from mathematical epidemiology. The information dissemination is directly related to the degree of the devices in the network, which in turn depends on the physical network parameters. Since the network is random with potentially a large number of devices, we use the degree based mean-field approach, in which all devices are considered to be statistically equivalent in terms of the degree and the analysis is done on a typical device. Therefore, the information dissemination dynamics of the system can be written in terms of the degree of the devices as follows [33]: (c) dIk (t) (c) (c) = −µIk (t) + α(c) kUk (t)Θ(c) , (7) dt (c) where Ik (t) denotes the proportion of devices with degree k that are in state Ik , i.e., informed with network-wide informa(c) (c) tion, at time t, Uk (t) = 1 − Ik (t) denotes the proportion of degree k devices that are in state Uk . The superscript (c) refers to the combined-layer network signifying network-wide information dissemination. The first term in (7) explains the annihilation of information with time, i.e., the informed devices return to the uninformed sate at a rate of µ. This ensures that at each time step, only the most recent information is propagated in the network as a particular piece of information is discarded after being retransmitted multiple times depending to the annihilation rate. For equilibrium analysis, we can assume that µ = 1 as the effect of the annihilation can be captured by the effective spreading rate. The second term accounts for the creation of informed devices due to the spreading. The rate of increase in the density of informed devices with degree k is directly proportional to the degree, the probability of successful transmission of information α, the proportion of uninformed (c) devices with degree k, i.e., Uk (t), and the average probability that a neighbour of a device with degree k is informed, denoted by Θ(c) . The probability Θ(c) can be computed as P (c) neighbour = k 0 |Kc = k)Ik0 (t), where P(k 0 |Kc = k) k0 P(Kc denotes the probability that a neighbour of a typical device with Kc = k has a degree Kcneighbour = k 0 . Since the network is PPP, the degrees are uncorrelated, i.e., E[Kcneighbour Kc ] = E[Kcneighbour ]E[Kc ], so we can effectively write P(Kcneighbour = (c) k0 P(Kcneighbour =k0 ) (c) Ik0 (t). Hence, by approk 0 |Kc = k)Ik0 (t) = E[Kc ] priately renaming dummy variables, Θ(c) can be expressed as follows [32]: X kP(Kc = k) (c) Θ(c) = Ik (t), (8) E[Kc ] k≥0 In this section, we present the dynamics of information dissemination across the IoBT network. For ease of explanation, we first describe the dynamics of a single message being propagated in the network, followed by the dynamics of two messages simultaneously spreading in the network. 1) Single Message Propagation: If all the devices in the network disseminate the same message from one device to another in a broadcast manner during each time slot using all the available radio interfaces, then each device with degree k can either be in an uninformed state (Uk ) or an informed state (Ik ) depending on the success of information transmis- where E[Kc ] is provided in Section II-B. Note that another expression for the rate of change in the uninformed devices, dU (c) (t) i.e., kdt , can also be written; however, it is useless since (c) (c) Uk (t) depends directly on Ik (t) and thus (7) completely describes the dynamics of network-wide information dissemination. 2) Multiple Message Propagation: In general, there may be multiple messages or pieces of information spreading in the IoBT network at any particular time. However, for the bilayer network model, we assume that there are two messages UUk,l k σ1 UUk,l l σ2 UUk,l IUk,l UIk,l IUk,l UI k,l characterizes the eventual spread of information in the network which is independent of time. Although, with the changes in network topology and other network configurations, the actual information spread might be different; however, the equilibrium state provides us with a reasonable understanding of the system behavior. For the single message dissemination dI k σ1 UIk,l l σ2 IU k,l II k,l II k,l II k,l Fig. 3: State transition diagram for the simultaneous diffusion of two different messages in the IoBT network. At each time instant, a change in status of only one of the messages is allowed. The arrows are labeled with the transition probabilities. propagating in the network, i.e., one in each network layer. Therefore, each device with a degree k in the first layer and l in the second layer can be in one of the four possible dynamical states, i.e., uninformed of both messages (U Uk,l ), informed of message 1 but uninformed of message 2 (IUk,l ), uninformed of message 1 but informed of message 2 (U Ik,l ), and informed of both messages (IIk,l ). These state variables denote the proportion of devices in the network that are in that particular state. To model the coupled dynamics of this information diffusion process, we can make use of the SISSIS interaction model [25] from mathematical epidemiology. A state transition diagram is given by Fig. 3. Notice that a change in status of one of the messages is allowed at each time instant. The set of differential equations describing the state evolution are given by (9) - (12), where Θ1 and Θ2 are the effective probabilities of a completely uninformed device, i.e., in state U Uk,l , to get informed by message 1 and message 2 respectively. They can be evaluated as follows: 1 X Θ1 = P(K1 = k, K2 = l)k (IUk,l,σ1 ,σ2 (t) + E[K1 ] k,l IIk,l,σ1 ,σ2 (t)) , (13) Θ2 = 1 X P(K1 = k, K2 = l)l (U Ik,l,σ1 ,σ2 (t) + E[K2 ] k,l IIk,l,σ1 ,σ2 (t)) , (14) Since U Uk,l (t) + U Ik,l (t) + IUk,l (t) + IIk,l (t) = 1, so in essence, only three of the four differential equations are linearly independent. The particular quantity of interest is IIk,l (t) as it signifies the proportion of devices that are informed by both messages simultaneously at any particular time. This may be crucial in the IoBT setting where the devices require information of both layers to make accurate decisions. Note that the developed dynamics can be easily extended to the case of multiple message propagation, which will require a larger state space; however, in this paper, we restrict ourselves to the case of only two network layers and messages. B. Steady State Analysis We are interested in determining the steady state or equilibrium of the information dissemination process since it (c) (t) case, we impose the stationarity condition, i.e., set kdt = 0. It results in the following expression: α(c) kΘ(c) (α(c) ) (c)∗ Ik = . (15) 1 + α(c) kΘ(c) (α(c) ) Notice that Θ(c) (α(c) ) is now a constant that depends on α(c) . Now, (8) and (15) present a system of equations that needs to be solved self-consistently to obtain the solution for Θ(c) (α(c) ) (c)∗ and Ik . For the multiple message propagation case, we can first write the reduced system in terms of the three independent states by substituting U Uk,l (t) = 1 − IUk,l (t) − U Ik,l (t) − IIk,l (t) to obtain (16), (17), and (18). dIU (t) k,l = 0, Again, using the stationarity condition, i.e., dt dIIk,l (t) = 0, and = 0, we obtain the following dt expressions: α(1) kΘ1 ∗ IUk,l = , (19) (1 + α(1) kΘ1 )(1 + α(2) lΘ2 ) α(2) lΘ2 ∗ U Ik,l = , (20) (1 + α(1) kΘ1 )(1 + α(2) lΘ2 )     α(1) kΘ1 α(2) lΘ2 ∗ IIk,l = , (21) 1 + α(1) kΘ1 1 + α(2) lΘ2 Now, the equations eqs. (19) to (21) and eqs. (13) and (14) need to be solved self-consistently to obtain the equilibrium ∗ solution of the quantity of interest, i.e., IIk,l . In the subsequent section, we present the methodology to obtain the solution to the dynamical information spreading process for the IoBT network and develop a framework that can assist in the planning and efficient design of such networks. dU Ik,l (t) dt IV. M ETHODOLOGY In this section, we first present a solution to the dynamical information spreading system developed in Section III and then use it for the efficient design of IoBT networks for mission-specific battlefield applications. A. Equilibrium Solution In order to find the equilibrium solution for the single network-wide message propagation case, we need to solve the self-consistent system expressed in (8) and (15). In fact, it reduces to obtaining a solution to the following fixed-point system: α(c) kΘ(c) (α(c) ) 1 X kP(Kc = k) Θ(c) (α(c) ) = . E[Kc ] 1 + α(c) kΘ(c) (α(c) ) k≥0 (22) An obvious solution for this fixed-point system is Θ(c) (α(c) ) = 0; however, it is noninformative. The existence of a nonzero solution is stated in the following theorem: Theorem 1. The fixed point equation in (22) relating to the information dissemination dynamics may have at least one dU Uk,l (t) = −(α(1) kΘ1 + α(2) lΘ2 )U Uk,l (t) + IUk,l (t) + U Ik,l (t), dt dIUk,l (t) = α(1) kΘ1 U Uk,l (t) − (α(2) lΘ2 + 1)IUk,l (t) + IIk,l (t), dt dU Ik,l (t) = α(2) lΘ2 U Uk,l (t) − (α(1) kΘ1 + 1)U Ik,l (t) + IIk,l (t), dt dIIk,l (t) = α(1) kΘ1 U Ik,l (t) + α(2) lΘ2 IUk,l (t) − 2IIk,l (t), dt dIUk,l (t) = α(1) kΘ1 − (α(1) kΘ1 + α(2) lΘ2 + 1)IUk,l (t) − α(1) kΘ1 U Ik,l (t) − (α(1) kΘ1 − 1)IIk,l (t), dt dU Ik,l (t) = α(2) lΘ2 − α(2) lΘ2 IUk,l (t) + (α(1) kΘ1 + α(2) lΘ2 + 1)U Ik,l (t) − (α(2) lΘ2 − 1)IIk,l (t), dt dIIk,l (t) = α(1) kΘ1 U Ik,l (t) + α(2) lΘ2 IUk,l (t) − 2IIk,l (t). dt solution in the domain Θ(c) (α(c) ) > 0 depending on the value of α(c) . The condition for this bifurcation to hold is α(c) ≥ E[Kc ] E[Kc2 ] . This bifurcation point is unique in the domain 0 < Θ(c) (α(c) ) ≤ 1. Proof. See Appendix C. Obtaining a closed form solution for the fixed point system in (22) for a PPP setting is not possible due to the complex form of P(Kc = k). Hence, an approximate solution can be obtained using the following theorem: Theorem 2. If a nonzero solution exists for the information E[Kc ] spreading dynamics in (8) and (15), i.e., α(c) ≥ E[K 2 , then c] for E[Kc ]  1, a lower bound approximation of the solution can be expressed as follows:  + 1 (c) (c) Θ (α ) ≈ 1 − (c) , (23) α E[Kc ] where (x)+ represents max(0, x). Proof. See Appendix D. As shown in Appendix D, Theorem 2 provides a lower bound for the exact solution and becomes a tight approximation for E[Kc ]  1. Moreover, the approximation has also resulted in an increase of the critical information spreading 1 threshold to α(c) ≥ E[K to ensure that Θ(c) ≥ 0. This is c] also asymptotically accurate as the original condition can be E[Kc ] 1 2 written as α(c) ≥ E[K , where σK is the 2] = σ2 c c Kc E[Kc ]+ E[K c] 1 approaches E[Kc ] as E[Kc ] variance of Kc . It becomes large. The relaxed condition is formally expressed by the following corollary. Corollary 1. The condition for obtaining an approximate nonzero equilibrium for the information dissemination dynamics is given as follows: 1 α(c) ≥ . (24) E[Kc ] In the IoBT network, the physical interpretation of E[Kc ] is that the average number of communication neighbours of a device in the combined network and hence, it is reasonable to assume that E[Kc ] ≥ 1 due to the potential high density of devices in IoBT networks. Therefore, the solution presented in Theorem 2 is indeed a good approximation to the actual solution. The corresponding solution for the density of informed (9) (10) (11) (12) (16) (17) (18) devices can be obtained using (15). In the case of multiple message propagation, we need to solve the self-consistent system of equations defined by the equations eqs. (19) to (21) and eqs. (13) and (14). It reduces to solving the fixed-point equations given in (25) and (26). Notice that Θ1 and Θ2 are independent of each other and their fixed point equations are similar to (22). Hence, the existence and uniqueness of the fixed point can be proved under similar conditions. Using a similar approach to the solution of the single message propagation case, a lower bound approximate solution is provided by the following theorem: Theorem 3. If a nonzero solution exists for the information spreading dynamics for the case two simultaneous message E[K1 ] E[K2 ] propagation, i.e., α(1) ≥ E[K and α(2) ≥ E[K 2 2 , then for 1] 2] E[K1 ]  1 and E[K2 ]  1, a lower bound approximation of the solution can be expressed as follows:  + 1 , (27) Θ1 (α) ≈ 1 − (1) α E[K1 ]  + 1 Θ2 (α) ≈ 1 − (2) . (28) α E[K2 ] Proof. See Appendix E. Similar to Corollary 1, the condition for obtaining an approximate nonzero equilibrium solution is α(i) ≥ 1/E[Ki ], i ∈ {1, 2}. The corresponding solution for the proportion of devices that are informed both messages, i.e., IIk,l can be obtained using (21) which turns out to be the following: (1)∗ (2)∗ ∗ IIk,l = Ik × Il . (29) This result is interesting and useful as it can be easily generalized for the case of multiple connectivity layers, i.e., types of devices, and multiple messages propagating simultaneously. B. Secure and Reconfigurable Network Design Once the equilibrium point for information dissemination has been determined, the next step is to design the IoBT network to achieve mission specific goals while efficiently using battlefield resources. In essence, the network design implies tuning the knobs of the network, which in the case of IoBT networks are the transmission ranges and the node deployment densities of the different types of battlefield things. However, changing the physical parameters may have an impact on the Θ2 = 1 X P(K1 = k, K2 = l)l E[K2 ] k,l cost and hence, the goal is to ensure a certain information spreading profile in the network while deploying the minimum number of devices and using the minimum transmit power. Let r = [r1 r2 ]T represent the vector of communication ranges of each of the type-I and type-II devices in the IoBT network. The minimum density of the devices in the network, determined by the mission requirements, is denoted by λmin , λmin ≥ 0. The maximum deployment density of the devices, defined by the capacity of the available devices, is denoted by λmax , λmax ≥ λmin . Similarly, the tunable transmission range limits of the devices can be expressed as rmin = [r1min r2min ]T , min rm ≥ 0, ∀m ∈ {1, 2}, and rmax = [r1max r2max ]T , min max , m ∈ {1, 2}. If w = [w1 w2 ]T such that rm ≥ rm P 2 m=1 wm = 1 represents the weight vector corresponding to the relative capital cost of deploying a type-I and type-II device respectively, and c represents the unit operational power cost signifying the importance of network power consumption, then a cost function for the network with device densities λ and transmission ranges r can be expressed as follows: C(p, λ, r) = w1 pλ + w2 (1 − p)λ + c(pλr1η + λr2η ), (30) where η denotes the path-loss exponent5 . The first term represents the total deployment cost per unit area of all the network devices while the second term represents the total energy cost per unit area of operating all the devices with transmission range r. The weights w can depend on several factors such as the time required for deployment, the monetary cost involved, or the number of devices available in stock, etc. We can then formulate the secure and reconfigurable network design problem as follows: minimize C(p, λ, r) (31) p,λ,r subject to (c)∗ (c) Ik ≥ Tk , ∀k ≥ 0, ∗ IIk,l min ≥ Tk,l , ∀k ≥ 0, l ≥ 0, p r min max ≤p≤p max ≤r≤r min ,λ , (32) (33) max ≤λ≤λ , (34) (c) Tk where ∈ (0, 1), k ≥ 0 are the minimum desired proportions of degree k devices that are informed with a single network-wide message, Tk,l ∈ (0, 1), k ≥ 0, l ≥ 0 are the desired proportion of devices with degree K1 = k and K2 = l that are informed with both messages simultaneously, and pmin and pmax are the minimum and maximum fractions of devices respectively that are of type-I. The cost function C is a convex nondecreasing function of p, λ, and r. However, it is an infinite dimension optimization problem due to a constraint on each degree class of devices. To be able to solve this problem, (c) we need to select a desired mapping Tk for the combined degree Kc = k and similarly a mapping Tk,l for devices 5 The power consumption of a device with communication range r m is η proportional to rm . (25) (26) 1 (c)∗ k,l !   2 1 X α(1) kΘ1 + α(1) klΘ1 Θ2 α(1) kΘ1 = , P(K = k)k 1 E[K1 ] (1 + α(1) kΘ1 )(1 + α(2) lΘ2 ) 1 + α(1) kΘ1 k !   2 α(2) lΘ2 α(2) lΘ2 + α(2) klΘ1 Θ2 1 X P(K2 = l)l = . E[K2 ] (1 + α(1) kΘ1 )(1 + α(2) lΘ2 ) 1 + α(2) lΘ2 l Proportion of Informed Devices, Ik 1 X Θ1 = P(K1 = k, K2 = l)k E[K1 ] 0.9 0.8 E[K c] = 5, 10, 20, 100 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 10 20 30 40 50 60 70 80 90 100 Degree of a typical device, k Fig. 4: Information dissemination profiles for varying average degree of devices (α = 0.3). with joint intra-layer degrees K1 = k, K2 = l, ∀k, l ≥ 0. Further investigation reveals that for a fixed α, the information dissemination can only have restricted trajectories based on the average device degree as shown in Fig. 4. Hence, the threshold mappings cannot be defined arbitrarily as they might not be achievable. To handle this problem, we can express the constraints in (32), assuming that α(c) satisfies the condition in Corollary 1, as follows: 1   , ∀k ≥ 0, E[Kc ] ≥ (35) (c) (c) (c) α − Tk / k(1 − Tk ) (c) Now, we need to satisfy (35) for each (k, Tk ) pair, and from Fig. 4, it is clear that there is no incentive to choose a threshold profile that is different from one of the possible trajectories. Therefore, specifying the threshold for a single value of the degree is sufficient to characterize the entire trajectory. Since Kc is a random variable with a distribution centered around E[Kc ], it is plausible to set a threshold on the proportion of devices with the average device degree, i.e., using the pair (c) (E[Kc ], TE[Kc ] ). It results in the following single constraint instead of the infinite set of constraints in (32): 1 E[Kc ] ≥ . (36) (c) α(c) (1 − TE[Kc ] ) (c) Note that this constraint implies α(c) E[Kc ] ≥ 1/(1 − TE[Kc ] ), (c) which satisfies the condition in Corollary 1, ∀ TE[Kc ] ∈ [0, 1], thus validating our assumption. For the second set of con∗ straints in (33), due to the decomposability of IIk,l as shown (i)∗ in (29) and the fact that Ik ∈ [0, 1], i ∈ {1, 2}, the constraints can be separated as follows: α(1) kΘ1 α(2) lΘ2 ≥ Tk,l , ≥ Tk,l , ∀k, l ≥ 0. (37) (1) 1 + α kΘ1 1 + α(2) lΘ2 Moreover, the thresholds can be replaced by simply Tk and Tl instead of Tk,l as they are constants. Using a similar approach as before, we replace the set of infinite constraints in (33) by the following two constraints: 1 , i ∈ {1, 2}. (38) E[Ki ] ≥ (i) α (1 − TE[Ki ] ) Algorithm 1 Secure and Reconfigurable Network Design 1: 2: (c) For brevity, we henceforth denote TE[Kc ] , TE[K1 ] , and TE[K2 ] as simply T (c) , T (1) , and T (2) . Furthermore, we denote 1 1 1 (c) , α(1) (1−T , T (1) , (1) ) , and α(2) (1−T (2) ) by T α(c) (1−T (c) ) and T (2) respectively corresponding to the desired minimum (i) success probabilities Ps leading to αi , i ∈ {1, 2, c}. The original optimization problem can then be rewritten as follows: minimize w1 pλ + w2 (1 − p)λ + c(pλr1η + λr2η ), (39) 3: 4: 5: p,λ,r subject to p2 λπr12 + λπr22 ≥ T (c) , 2 p λπr12 ≥ T (1) , λπr22 ≥ T (2) , min max (40) (41) (42) 6: rmin ≤ r ≤ rmax . (43) It is important to note that the conditions for existence of nonzero equilibrium in Corollary 1 and the ones resulting from Theorem 3 are implicitly incorporated into the constraints and do not need to be imposed separately. This implies that if a feasible solution to the optimization problem exists, then there exists a nonzero equilibrium solution to the information dissemination dynamics. The objective and constraints are nondecreasing smooth functions of the optimization variables. It is clear that the objective and constraints are convex in the feasible solution space. Therefore, the problem can be solved using standard convex optimization techniques [34]. The secure and resilient framework for mission critical information dissemination in IoBT networks is provided in Algorithm 1. At the beginning of the mission, the central network planner obtains the optimal physical network parameters by solving the optimization problem in eqs. (39) to (43), and accordingly deploys the devices with appropriate transmission powers. The central planner then periodically analyzes the connectivity situation of the network with a reconfigurability interval denoted by tr . The reconfigurability interval could range from several hours to days depending on the mission requirements. Several techniques may be employed to estimate the connectivity, and consequently the information dissemination in the network. The effect of physical damage can be measured by physically monitoring the network with the aid of robotic systems such as in [35]. On the other hand, the effect of cyber attacks may be estimated by running discovery tests on the network to assess the reachability of nodes. Based on these results, estimates of the information dissemination thresholds can be obtained, i.e., T̂ (1) , T̂ (2) , and T̂ (c) . If there is a significant drop in the estimated threshold as compared to the desired one or there is a change in the required security level, then the optimization needs to be re-computed. The new optimized parameters help in identifying the additional deployment needed for each type of devices and/or the reconfiguration of their transmission powers to achieve the desired information dissemination in 7: p ≤p≤p , λmin ≤ λ ≤ λmax , 8: 9: 10: 11: 12: 13: At epoch, i.e., t = 0; Initialize requirements for information dissemination, i.e., T (1) , T (2) , and T (c) and the anticipated threat level δ. init init init Obtain optimal network parameters λinit 1 = p λ , λ2 = init init init λ , r1 , and r2 by solving the optimization problem in eqs. (39) to (43) and accordingly deploy the devices with the appropriate communication ranges. repeat if t = ζtr , ζ ∈ Z+ then Obtain an estimate of the density of active devices λ̂ = [λ̂1 , λ̂2 ] and use the initial communication ranges to estimate the prevailing information dissemination level T̂ (1) , T̂ (2) , and T̂ (c) . Re-evaluate the desired security level in response to the threats and accordingly update the parameter δ̂. if |T (1) − T̂ (1) | ≥  or |T (2) − T̂ (2) | ≥  or |T (c) − T̂ (c) | ≥  or δ̂ 6= δ, then Recompute the optimization problem in eqs. (39) to (43) to obtain the new set of optimal parameters ← λnew , r1new , and r2new . ← pnew λnew , λnew λnew 2 1 Deploy additional required λnew − λ̂ devices in the network and reconfigure transmission powers to achieve the required transmission ranges rnew . end if end if t ← t + 1; until End of mission. the IoBT network. In the following section, we investigate the behaviour of the optimal solutions under varying threat levels and mission specific performance thresholds. V. R ESULTS In this section, we provide the results obtained by testing the developed framework under different battlefield missions. We assume a bi-layer IoBT network comprising of type-I and type-II devices. The first type of devices is assumed to be commanders and the second type is assumed to be followers. The assumption yields a simple yet natural network configuration in a battlefield, e.g., being composed of soldiers and distributed commanding units. The allowable physical parameter ranges of the respective devices are selected to be as follows: the minimum and maximum device deployment density is selected as λmin = 1 km−2 and λmax = 15 km−2 respectively, the minimum and maximum fraction of type-I devices pmin = 0 and pmax = 0.4 respectively, the minimum and maximum communication ranges of devices in the first layer, r1min = 100 m and r1max = 2000 m respectively, and the minimum and maximum communication ranges of devices in the second layer, r2min = 10 m and r2max = 800 m. The parameters imply that the active devices in the first layer, i.e., the commanding units have a higher allowable transmission range than the followers. In practice, the limits can be based on tactical requirements of the missions. For simplicity, we assume that sufficient number of channels are available and for the considered densities and communication ranges of the devices, and the MAC protocol is able to effectively mitigate interference in the communication resulting in a constant success probability. Further, we assume that the desired success (i) probabilities Ps = 1, ∀i ∈ {1, 2, c} implying perfect success of transmissions. Consequently, α(1) = α(2) = α(c) and is henceforth referred to as α. The weights representing the relative deployment cost are chosen to be w1 = 100 and w2 = 50, which implies that the deployment cost of the commanding units is twice as much as the follower units. The unit cost of power is selected to be c = 100 that can be adjusted according to the importance of each mission and the path-loss exponent η = 4. A. Mission Scenarios In the battlefields, there can be several types of missions such as intelligence, surveillance, encounter battle, espionage, reconnaissance, etc. In our results, we will focus particularly on the two most common mission scenarios, i.e., intelligence and encounter battle. Both of them have completely different requirements in terms of the information flow in the network, which are described as follows: 1) Intelligence: In the intelligence mission, the goal is to provide commanders with the information from a range of sources to assist them in operational or campaign planning. This requires a high network-wide dissemination of a single message, while the condition for simultaneously informed devices may not be stringent. In essence, the commander network must reinforce the follower network to achieve a high network-wide message propagation. Hence, to emulate such an intelligence mission, we select the following set of information spreading thresholds: T (1) = T (2) = 0.6, T (c) = 0.8. The optimal physical parameters obtained for the intelligence mission against increasing threat level δ are shown in Fig. 5. There are several interesting observations in the intelligence mission. A general trend is that the required transmission ranges and deployment densities increases as the threat level increases. Consequently, the cost function, which signifies the deployment and operation cost of the network, also increases as shown in Fig. 5(c). Fig. 5(a) and Fig. 5(b) show that the transmission range of the commanders is always higher than the followers while the densities of the followers is higher than that of the commanders. This observation makes sense as the followers equipped with sensors should be more in the total number while the commander network should have a larger influence area to be able to gather information for the intelligence mission. Another important observation is that the framework tends to increase the deployment density of the devices first before increasing their transmission ranges. It is due to a high cost of power consumption that tends to force the devices to minimize the transmission ranges. 2) Encounter Battle: In the encounter battle or meeting engagement scenario, there is a contact between the battling forces. In such situations, commanders need to act quickly to gain advantage over the opponents. This requires devices to be informed simultaneously about the information disseminated in both network layers. Hence, there is a need for strong coordination among commanders as well as followers. Additionally, the common status information sharing between all network devices must be sufficiently high to ensure accurate decision-making. Therefore, we set the following information spreading thresholds: T (1) = T (2) = 0.8 and T (c) = 0.6. The resulting optimal parameters against the changing threat level are presented in Fig. 6. In contrast with the intelligence mission, the framework utilizes all the available devices at a much lower threat level and is also forced to increase the transmission ranges to full capacity despite the high power cost. The cost function for the encounter battle in Fig. 6(c) is significantly higher than the intelligence mission in Fig. 5(c) due to more stringent information spreading criteria which requires more transmission power and higher device density. Next, we study the impact of the information spreading thresholds on the optimal network parameters. In Fig. 7, we fix the intra-layer information dissemination threshold to T (1) = T (2) = 0.5 and observe the change in parameters for varying network-wide information dissemination threshold T (c) . Notice that the curves in Fig. 7(a) and Fig. 7(b) are mostly flat except for very high values of T (c) . This implies that there is no need for additional resources to achieve a higher network-wide information dissemination, which reaffirms the fact that the intra-layer information dissemination is a much stricter condition than the network-wide information dissemination. Similarly, in Fig. 8, we fix the network-wide information dissemination threshold to T (c) = 0.5 and observe the change in optimal parameters with a change in intralayer information dissemination threshold. In this case, the device densities are increased successively followed by the communication ranges as the threshold for the proportion of simultaneously informed devices is increased. VI. C ONCLUSION In this paper, we have presented a generic framework for secure and reconfigurable design of IoT empowered battlefield networks. The framework provides a tractable approach to tune the physical network parameters to achieve the desired realtime data dissemination among different types of battlefield devices according to the assigned missions. It takes into account the perceived threat level from the opponent as well as the costs involved in the deployment and operation of combat equipment to provide a robust and cost effective design of communication networks in battlefields which can be highly useful in military planning. Optimized network parameters are provided for the two typical mission scenarios of intelligence and encounter battle in which the desired information spreading objectives are different from one another. Results show that the mission goals can be achieved by either appropriately changing the deployment density of combat units or by changing their transmission powers or both in response T (1) = 0.6, T (2) = 0.6, T (c) = 0.8 T (1) = 0.6, T (2) = 0.6, T (c) = 0.8 2000 λ1 r2 λ2 Device Intensity (devices/km2 ) 1600 1400 1200 1000 800 600 400 0.1 8000 7000 10 Cost Function Transmission Range (m) 1800 T (1) = 0.6, T (2) = 0.6, T (c) = 0.8 9000 15 r1 5 6000 5000 4000 3000 2000 1000 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.9 0.2 0.3 Threat level, δ 0.4 0.5 0.6 0.7 0.8 0 0.1 0.9 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Threat level, δ Threat level, δ (a) Transmission ranges of devices against vary- (b) Deployment density of devices against varying threat level. ing threat level. (c) Cost function against varying threat level. Fig. 5: Optimal network parameters for the intelligence mission. T (1) = 0.8, T (2) = 0.8, T (c) = 0.6 T (1) = 0.8, T (2) = 0.8, T (c) = 0.6 2000 λ1 1600 1400 1200 1000 800 600 400 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 14 λ2 10000 12 Cost Function r2 Device Intensity (devices/km2 ) Transmission Range (m) 1800 T (1) = 0.8, T (2) = 0.8, T (c) = 0.6 12000 16 r1 10 8 8000 6000 4000 6 2000 4 2 0.1 0.2 0.3 Threat level, δ 0.4 0.5 0.6 0.7 0.8 0 0.1 0.9 0.2 0.3 Threat level, δ 0.4 0.5 0.6 0.7 0.8 0.9 Threat level, δ (a) Transmission ranges of devices against vary- (b) Deployment density of devices against varying threat level. ing threat level. (c) Cost function against varying threat level. Fig. 6: Optimal network parameters for the encounter mission. T (1) = T (2) = 0.5 T (1) = T (2) = 0.5 2000 16 r1 λ1 r2 Device Intensity (devices/km2 ) Transmission Range (m) 1800 1600 1400 δ = 0.85, 0.8, 0.75, 0.7 1200 1000 800 δ = 0.85, 0.8, 0.75, 0.7 600 14 λ2 12 10 8 δ = 0.65, 0.7, 0.75, 0.8 6 4 δ = 0.65, 0.7, 0.75, 0.8 400 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Network-wide information dissemination threshold, T (c) (a) Transmission ranges of devices against varying network-wide information dissemination threshold. 2 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Network-wide information dissemination threshold, T (c) (b) Deployment density of devices against varying network-wide information dissemination threshold. Fig. 7: Network parameter variation with changes in combined-layer information dissemination threshold. to changing mission requirements. Moreover, the network can be reconfigured according to the periodic assessment of lost connectivity or changes in security requirements to meet the desired mission goals. A PPENDIX A P ROOF OF L EMMA 1 To evaluate the degree distribution of the devices in the first network layer, we proceed as follows: 2 X P(K1 = k) = P(K1 = k|Ti = j)P(Ti = j), (44) j=1 Since type-II devices are not active in the first network layer, so we know that P(K1 = 0|Ti = 2) = 1 and P(K1 > 0|Ti = 2) = 0. However, type-I devices can communicate with other type-I devices in the first network layer. Since the active devices in the first layer are also represented as a PPP, the degree of the RGG formed among the active devices is Poisson distributed [13], i.e., P(K1 = k|Ti = 1) ∼ 2 Poisson(λ1 πr12 ). It follows that P(K1 = 0|Ti = 1) = e−λ1 πr1 2 e−λ1 πr1 (λ πr 2 )k 1 1 , k > 0. Finally, and P(K1 > 0|Ti = 1) = k! since P(Ti = 1) = p and P(Ti = 2) = 1 − p, we can substitute the corresponding expressions for K1 = 0 and K1 > 0 in (44) to obtain the result in (1). Similarly, for the distribution of K2 , T (c) = 0.5 T (c) = 0.5 2000 15 r1 λ1 r2 Device Intensity (devices/km2 ) Transmission Range (m) 1800 δ = 0.8, 0.7, 0.6, 0.5 1600 1400 1200 1000 800 δ = 0.8, 0.7, 0.6, 0.5 600 400 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Intra-layer information dissemination threshold, T1 , T2 (a) Transmission ranges of devices against varying intra-layer information dissemination threshold. λ2 10 δ = 0.65, 0.7, 0.75, 0.8 5 δ = 0.65, 0.7, 0.75, 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Network-wide information dissemination threshold, T (c) (b) Deployment density of devices against varying intralayer information dissemination threshold. Fig. 8: Network parameter variation with changes in intra-layer information dissemination threshold. Type-I device we know that P(K2 = k) ∼ Poisson(λ2 πr22 ) regardless of Ti . This directly leads to the result in (2). Using these probability distributions, it is straightforward to derive the expected degree in both layers given by (3) and (4). Type-II device r1 r2 O Region A A PPENDIX B P ROOF OF L EMMA 2 The distribution of the combined degree can be evaluated as follows: 2 X P(Kc = k) = P(Kc = k|Ti = j)P(Ti = j), (45) Region B Fig. 9: Combined degree of a typical device located at the origin. A PPENDIX C P ROOF OF T HEOREM 1 j=1 It is clear that P(Kc = k|Ti = 2) ∼ Poisson(λ2 πr22 ) as type-II devices are only connected in the second network layer. However, if a typical device is of type-I, then the degree needs to be more carefully evaluated. We have provided an illustration in Fig. 9 to aid in characterizing the degree. The degree of a type-I device placed at the origin can be expressed as Kc = 2N1A + N2A + N1B , where N1A and N2A represent the number of devices of type-I and typeII respectively in the green shaded circular region A. N1B represents the number of type-I devices in the blue shaded hollow circular region B. Note that type-I and type-II devices are distributed according to independent PPPs with intensity pλ and (1−p)λ. It results in the fact that N1A , N2A , and N1B are independent random variables following a Poisson(pλπr22 ), Poisson((1−p)λπr22 ), and Poisson(pλπ(r12 −r22 )) respectively. The sum Kc is not composed of independent terms, however, due to N1A which has a multiplicity of 2. Therefore, P(Kc = k|Ti = 1) = P(N1A = k2 ) ∗ P(N2A = k) ∗ P(N1B = k), where ∗ represents the convolution operator. Therefore, the probability distribution of the combined degree can be obtained by substituting the above developed expressions into (45) and using the fact that P(Ti = 1) = p and P(Ti = 2) = 1 − p. The average combined layer degree can be obtained as E(Kc ) = p (2E(N1A ) + E(N2A ) + E(N1B ))+ (1 − p)(λπr22 ) = pλ1 πr12 + λ2 πr22 . To prove that the fixed point equation described by (22) has a unique solution in the domain Θ(c) > 0, we make use of the Banach fixed-point theorem (or contraction mapping theorem) [36]. We prove that the functional   1 Kc2 α(c) Θ(α(c) ) (c) F (Θ(α )) = E (46) E[Kc ] 1 + Kc α(c) Θ(α(c) ) experiences a contraction for all Θ(α(c) ) ∈ [0, 1]. More precisely, we prove that |F (Θ1 (α(c) )) − F (Θ2 (α(c) ))| ≤ c|Θ1 (α(c) ) − Θ2 (α(c) )| for any Θ1 (α(c) ), Θ2 (α(c) ) ∈ [0, 1], where 0 ≤ c < 1. The fact that the constant c is strictly less than 1 implies that the functional is contracted. The proof is as follows: F (Θ1 (α(c) )) − F (Θ2 (α(c) )) = (47)   Kc2 α(c) Θ1 (α(c) ) 1 E − E[Kc ] 1 + Kc α(c) Θ1 (α(c) )   1 Kc2 α(c) Θ2 (α(c) ) E , E[Kc ] 1 + Kc α(c) Θ2 (α(c) ) Θ1 (α(c) ) − Θ2 (α(c) ) = × E[Kc ]   Kc2 α(c) E . (48) (1 + Kc α(c) Θ1 (α(c) ))(1 + Kc α(c) Θ2 (α(c) )) To complete the proof, we need to show that   1 Kc2 α(c) < 1, E E[Kc ] (1 + Kc α(c) Θ1 (α(c) ))(1 + Kc α(c) Θ2 (α(c) )) (49) 0.8 Exact Solution Approximation E[K i] = 12.57 0.7 0.6 E[K i] = 6.28 Θ (α ) Kc2 α(c) . (1+Kc α(c) Θ1 (α(c) ))(1+Kc α(c) Θ2 (α(c) )) 1 0.9 0.5 It can be Let g(Kc ) = E[K ] = 3.14 0.4 proved that g(Kc ) is concave for Kc ≥ 0 by showing that 0.3 g 00 (Kc ) < 0, ∀Kc ≥ 0. Therefore, using Jensen’s inequal0.2 ity [37], we can say that E[g(Kc )] ≤ g(E[Kc ]), with equality 0.1 iff Kc is deterministic. It follows that 0   0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 Kc2 α(c) Information spreading rate, α E ≤ (c) (c) (c) (c) E[Kc ] (1 + Kc α Θ1 (α ))(1 + Kc α Θ2 (α )) Fig. 10: Accuracy of Jensen’s lower bound. E[Kc ]α(c) , (1 + E[Kc ]α(c) Θ1 (α(c) ))(1 + E[Kc ]α(c) Θ2 (α(c) )) A PPENDIX D 1 . = P ROOF OF T HEOREM 2 Θ1 (α(c) ) + Θ2 (α(c) ) + E[Kc ]α(c) Θ1 (α(c) )Θ2 (α(c) ) + E[Kc1]α(c) Obtaining the nonzero solution for the fixed point equation (50) The expression in (50) is strictly less than 1 only if the (22) in closed form is not possible since we need to solve the following equation for Θ(α(c) ): following condition is satisfied: ∞ Kc2 α(c) 1 X Θ1 (α(c) ) + Θ2 (α(c) ) + E[Kc ]α(c) Θ1 (α(c) )Θ2 (α(c) )+ P(Kc = k), (55) 1= E[Kc ] 1 + Kc α(c) Θ(α(c) ) 1 k=0 > 1. (51) E[Kc ]α(c) where P (Kc = k) is difficult to obtain in closed form. The condition in (51) depends on the relative magnitudes of Therefore, we resort to finding an approximation for the the quantities E[Kc ] and α(c) . Regardless, it reveals that we solution which is asymptotically accurate. Let g(Kc ) = Kc2 α(c) need to exclude the values of Θ(α(c) ) that are too close to . Since g 00 (Kc ) > 0, ∀Ki ≥ 0, so g(Ki ) is 1+Kc α(c) Θ(α(c) ) (c) zero. For sufficiently large values of Θ(α ), it is clear from a convex function for K ≥ 0. Using Jensen’s inequality, we i (51), that F (Θ(α(c) )) is indeed a contraction with respect to can say that E[g(K )] ≥ g(E[K ]), with equality only if K c c c the absolute value metric. Hence, by the contraction mapping is deterministic. This implies the following: (c)   theorem, F (Θ(α )) has a unique fixed point in the domain Kc2 α(c) E[Kc ]2 α(c) Θ(α(c) ) > 0. E ≥ . (56) (c) (c) 1 + Kc α Θ(α ) 1 + E[Kc ]α(c) Θ(α(c) ) The nonzero equilibrium solution can be obtained by solvTherefore, we can write (55) as follows: ing the following equation: E[Kc ]α(c)   , (57) 1≥ 1 Kc2 α(c) 1 + E[Kc ]α(c) Θ(α(c) ) 1= E . (52) E[Kc ] 1 + Kc α(c) Θ(α(c) ) which leads to the final solution, h i Kc2 α(c) 1 (c) 1 Let h(Θ(α )) = E[Kc ] E 1+Kc α(c) Θ(α(c) ) . We need to find Θ(α(c) ) ≥ 1 − (c) . (58) α E[K (c) c] a solution to the equation h(Θ(α )) = 1 in the domain Using our prior knowledge that Θ(α(c) ) ≥ 0, we need to 0 < Θ(α(c) ) ≤ 1. It is clear that h(Θ(α(c) )) is monotonically (c) (c) decreasing for Θ(α ) > 0. Therefore, it is sufficient to show ensure that α E[Kc ] ≥ 1. In general, the complete solution 1 (c) that h(0) > 1 and h(1) < 1 for a unique nonzero solution to can be expressed as Θ(α ) ≥ max(0, 1 − α(c) E[Kc ] ). To (c) exist for the equation h(Θ(α )) = 1. This result is proved measure the accuracy of this bound, we solve the fixed-point equation exactly using an fixed-point iteration and compare below: h i 2 the results for different values of α(c) and E[Kc ] = λi πr2 . 1 E[K ] c h(0) = E Kc2 α(c) = α(c) , (53) We choose a fixed r = 0.2 km and λ = [25, 50, 100] km−2i , i i E[Kc ] E[Kc ]   which results in E[K ] = [3.14, 6.28, 12.57]. A plot of the 2 (c) i 1 Kc α h(1) = E results is provided in Fig. 10. It can be observed that the lower E[Kc ] 1 + Kc α(c) bound obtained using Jensen’s inequality is tight for all values   (c) Kc α(c) 1 1 E Ki E[Kc ] = 1. (54) of α when E[Ki ]  1. = < E[Kc ] E[Kc ] 1 + Kc α(c) Kc α(c) A PPENDIX E In (54), the inequality follows from the fact that 1+K (c) < cα P ROOF OF T HEOREM 3 (c) 1, ∀K > 0, α > 0. A nonzero solution to (52) exists only i c E[Kc ] if h(0) ≥ 1, which implies that α(c) ≥ E[K 2 . This is exactly c] the critical spreading rate, also known as epidemic threshold, in the SIS model [18]. The objective is to obtain a fixed point solution to the equations (25) and (26). Since the equations are decoupled, we can employ the same approach used for Theorem 2 to show the existence and uniqueness of the fixed point. Consequently, a lower bound approximation of the solution for each fixedpoint equation can be obtained using the Jensen’s inequality as in Appendix D. The condition for existence of a solution E[K1 ] E[K2 ] is α(c) ≥ E[K and α(c) ≥ E[K 2 , which can be written as 2]   12 ] E[K ] E[K ] α(c) ≥ max E[K12 ] , E[K22 ] . 1 2 R EFERENCES [1] M. J. Farooq and Q. Zhu, “Secure and reconfigurable network design for critical information dissemination in the internet of battlefield things (IoBT),” in 15th Int. Symp. Model. Optim. Mobile Ad Hoc Wireless Netw. (WiOpt 2017), Paris, France, May 2017. [2] D. Evans, “The Internet of Things: How the Next Evolution of the Internet Is Changing Everything,” Cisco Inc., White Paper, 2011. [3] The Internet Of Things: Mapping the Value Beyond the Hype. Mckinsey Global Institute, 2015. [4] D. Zheng and W. Carter, Leveraging the Internet of Things for a more Efficient and Effective Military. Lanham, MD, USA: Rowman & Littlefield, 2015. [5] M. Tortonesi, A. Morelli, M. Govoni, J. Michaelis, N. Suri, C. Stefanelli, and S. Russell, “Leveraging internet of things within the military network environment-challenges and solutions,” in IEEE 3rd World Forum on Internet of Things (WF-IoT 2016), RESTON, VA, USA, Dec. 2016, pp. 111–116. [6] “The Internet of Things For Defense,” Wind River Systems, 2015. [7] A. Kott, A. Swami, and B. J. West, “The internet of battle things,” Computer, vol. 49, no. 12, Dec. 2016. [8] N. A. Stanton and C. Baber, Modelling Command and Control: Event Analysis of Systemic Teamwork. CRC Press, 2008. [9] P. Fraga-Lamas, T. M. Fernndez-Carams, M. Surez-Albela, L. Castedo, and M. Gonzlez-Lpez, “A review on internet of things for defense and public safety,” Sensors, vol. 16, no. 10, 2016. [10] N. Suri, M. Tortonesi, J. Michaelis, P. Budulas, G. Benincasa, S. Russell, C. Stefanelli, and R. Winkler, “Analyzing the applicability of internet of things to the battlefield environment,” in Intl. Conf. Mil. Commun. and Inf. Sys. (ICMCIS 2016), Brussels, Belgium, May 2016. [11] O. Bello and S. Zeadally, “Intelligent device-to-device communication in the internet of things,” IEEE Syst. J., vol. 10, no. 3, pp. 1172–1182, Sept 2016. [12] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59, no. 11, pp. 3122–3134, Nov. 2011. [13] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic geometry and random graphs for the analysis and design of wireless networks,” IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1029–1046, Sep. 2009. [14] P. Balister, A. Sarkar, and B. Bollobás, Percolation, Connectivity, Coverage and Colouring of Random Geometric Graphs. Berlin, Heidelberg: Springer Berlin Heidelberg, 2008, pp. 117–142. [15] F. Brauer, P. van den Driessche, and J. Wu, Eds., Mathematical Epidemiology. Germany: Springer Berlin Heidelberg, 2008. [16] A. L. Lloyd and R. M. May, “How viruses spread among computers and people,” Science, New Series, vol. 292, no. 5520, pp. 1316–1317, May 2001. [17] Y. Moreno, M. Nekovee, and A. F. Pacheco, “Dynamics of rumor spreading in complex networks,” Phys. Rev. E, vol. 69, p. 066130, Jun. 2004. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevE.69. 066130 [18] R. Pastor-Satorras and A. Vespignani, “Epidemic spreading in scale-free networks,” Phys. Rev. Lett., vol. 86, pp. 3200–3203, Apr. 2001. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevLett.86.3200 [19] N. C. Valler, B. A. Prakash, H. Tong, M. Faloutsos, and C. Faloutsos, “Epidemic spread in mobile ad hoc networks: Determining the tipping point,” in Proc. 10th Intl. IFIP TC 6 Conf. Netw. - Volume Part I. Berlin, Heidelberg: Springer-Verlag, 2011. [20] E. Yanmaz, “Epidemic propagation in overlaid wireless networks,” in IEEE Global Commun. Conf. (GLOBECOM 2008), New Orleans, LA, USA, Nov. 2008. [21] A. B. M. Gharbieh, H. ElSawy and M.-S. Alouini, “Spatiotemporal stochastic modeling of IoT enabled cellular networks: Scalability and stability analysis,” in IEEE Global Commun. Conf. (GLOBECOM 2016), Washington D.C, USA, Dec. 2016. [22] Q.-H. Liu, W. Wang, M. Tang, and H.-F. Zhang, “Impacts of complex behavioral responses on asymmetric interacting spreading dynamics in multiplex networks,” Sci. Rep., Nature, vol. 6, no. 25617, 2016. [23] C. Granell, S. Gómez, and A. Arenas, “Dynamical interplay between awareness and epidemic spreading in multiplex networks,” Phys. Rev. Lett., vol. 111, p. 128701, Sep. 2013. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevLett.111.128701 [24] W. Wang, M. Tang, H. Yang, Y. Do, Y.-C. Lai, and G. Lee, “Asymmetrically interacting spreading dynamics on complex layered networks,” Sci. Rep., Nature, vol. 4, May. [25] J. Sanz, C.-Y. Xia, S. Meloni, and Y. Moreno, “Dynamics of interacting diseases,” Phys. Rev. X, vol. 4, p. 041005, Oct. 2014. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevX.4.041005 [26] D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic geometry and its applications, ser. Wiley series in probability and mathematical statisitics. Chichester, W. Sussex, New York: Wiley, 1987. [27] J. Dall and M. Christensen, “Random geometric graphs,” Phys. Rev. E, vol. 66, p. 016121, Jul 2002. [28] F. Baccelli and B. Baszczyszyn, “Stochastic geometry and wireless networks: Volume i theory,” Foundations and Trends in Networking, vol. 3, no. 34, pp. 249–449, 2010. [29] A. Marttinen, A. M. Wyglinski, and R. Jntti, “Statistics-based jamming detection algorithm for jamming attacks against tactical MANETs,” in IEEE Mil. Commun. Conf. (MILCOM 2014), Baltimore, USA, Oct 2014. [30] R. Sanchez, J. Evans, and G. Minden, “Networking on the battlefield: challenges in highly dynamic multi-hop wireless networks,” in IEEE Mil. Commun. Conf. (MILCOM 1999), 1999. [31] O. Dousse, M. Franceschetti, N. Macris, R. Meester, and P. Thiran, “Percolation in the signal to interference ratio graph,” Journal of Applied Probability, vol. 43, no. 2. [32] R. Pastor-Satorras, C. Castellano, P. Van Mieghem, and A. Vespignani, “Epidemic processes in complex networks,” Rev. Mod. Phys., vol. 87, pp. 925–979, Aug. 2015. [33] J. Marro and R. Dickman, Nonequilibrium Phase Transitions in Lattice Models. New York, USA: Cambridge University Press, 1999, ch. The contact process. [34] L. V. Stephen P. Boyd, Convex Optimization. UK: Cambridge University Press, 2004. [35] P. Reverdy and D. E. Koditschek, “Mobile robots as remote sensors for spatial point process models,” in IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS 2016), Daejeon, Korea, Oct. 2016, pp. 2847– 2852. [36] J. K. Hunter and B. Nachtergaele, Applied Analysis. World Scientific, 2001, ch. The Contraction Mapping Theorem. [37] Z. Cvetkovski, Inequalities: Theorems, Techniques and Selected Problems, Springer, Berlin, Heidelberg, 2012, ch. Convexity, Jensen’s Inequality, pp. 69–77. Muhammed Junaid Farooq received the B.S. degree in electrical engineering from the School of Electrical Engineering and Computer Science (SEECS), National University of Sciences and Technology (NUST), Islamabad, Pakistan, the M.S. degree in electrical engineering from the King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia, in 2013 and 2015, respectively. Then, he was a Research Assistant with the Qatar Mobility Innovations Center (QMIC), Qatar Science and Technology Park (QSTP), Doha, Qatar. Currently, he is a PhD student at the Tandon School of Engineering, New York University (NYU), Brooklyn, New York. His research interests include modeling, analysis and optimization of wireless communication systems, cyber-physical systems, and the Internet of things. He was the recipient of the President’s Gold Medal for the best academic performance from the National University of Sciences and Technology (NUST). Quanyan Zhu (S’04, M’12) received B. Eng. in Honors Electrical Engineering from McGill University in 2006, M.A.Sc. from University of Toronto in 2008, and Ph.D. from the University of Illinois at Urbana-Champaign (UIUC) in 2013. After stints at Princeton University, he is currently an assistant professor at the Department of Electrical and Computer Engineering, New York University. He is a recipient of many awards including NSERC Canada Graduate Scholarship (CGS), Mavis Future Faculty Fellowships, and NSERC Postdoctoral Fellowship (PDF). He spearheaded and chaired INFOCOM Workshop on Communications and Control on Smart Energy Systems (CCSES), and Midwest Workshop on Control and Game Theory (WCGT). His current research interests include Internet of things, cyber-physical systems, security and privacy, and system and control.
3cs.SY
Dominance analysis of linear complementarity systems arXiv:1802.00284v1 [cs.SY] 1 Feb 2018 F. A. Miranda-Villatoro∗ Abstract— The paper extends the concepts of dominance and p-dissipativity to the non-smooth family of linear complementarity systems. Dominance generalizes incremental stability whereas p-dissipativity generalizes incremental passivity. The generalization aims at an interconnection theory for the design and analysis of switching and oscillatory systems. The approach is illustrated by a detailed study of classical electrical circuits that switch and oscillate. I. INTRODUCTION Dominance analysis and p-dissipativity were recently introduced in [1], [2] to extend the application of dissipativity theory to the analysis of multistable and oscillatory systems. The approach is differential, that is, based on the analysis of linearized dynamics along trajectories, in the spirit of contraction theory [3], convergence analysis [4], or differential stability analysis [5]. It is particularly adapted to systems whose linearization can be easily parametrized, such as Lure systems that interconnect linear time-invariant systems with static nonlinearities [6]. The present paper investigates how to extend this analysis to nonlinear circuits modeled as linear complementarity systems: models that consist of linear time-invariant systems augmented with a static complementarity constraint. The modeling framework of linear complementarity systems has proven very useful to analyze systems whose nonlinear dynamical behavior arises from non-smooth constraints [7], [8], [9], [10], [11]. They find applications in a number of fields including mechanical systems with unilateral constraints [12], electrical circuits with diodes [13], and mathematical programming [14]. Linear complementarity systems provide an attractive framework for dominance analysis because they are general enough to model switching and oscillatory behaviors often encountered in the presence of non-smooth constraints and specific enough to lead to tractable analysis. In particular, the passivity property of complementarity constraints has proven central to analyze linear complementarity systems in the framework of dissipativity theory [15], [9]. Linear complementarity systems hence offer an ideal platform for the application of p-dissipativity theory to switching and oscillatory behaviors. To account for the non-smoothness of linear complementarity systems, the differential analysis of [1], [2] has to be replaced by incremental analysis [16], [17], [18]. Incremental analysis studies how increments between The research leading to these results has received funding from the European Research Council under the Advanced ERC Grant Agreement Switchlet n. 670645. ∗ University of Cambridge, Department of Engineering, Trumpington Street, CB2 1PZ, Cambridge, United Kingdom. Emails: [email protected], [email protected], [email protected] F. Forni∗ R. Sepulchre∗ trajectories evolve in time, whereas differential analysis only considers linearized trajectories, that it, infinitesimal increments. In the context of linear complementarity systems, the difference is technical rather than conceptual: the main results of [1], [2] extend with no difficulty to the incremental setting imposed by non-smooth constraints. This paper is organized as follows. In Section II we briefly review the modeling of linear complementarity systems and its core passivity property. Section III is dedicated to the property of dominance and p-dissipativeness in the incremental setting. The example Section IV illustrates the potential of p-dissipativity theory to analyze classical switching and oscillatory circuits. The papers ends with conclusions in Section V. II. L INEAR COMPLEMENTARITY SYSTEMS AND INCREMENTAL PASSIVITY A. Linear complementarity systems A linear complementarity system [15] consists of a linear dynamical system subject to a complementarity constraint   ẋ = Ax + Bu + Bv (1) y = Cx + Du   0 ≤ u ⊥ y ≥ 0, where x ∈ Rn is the state variable, u ∈ Rm and y ∈ Rm are the so-called complementary variables, and v ∈ Rm is an additional control input. The matrices A, B, C, and D are constant and of the appropriate dimensions. The complementarity condition 0 ≤ u(t) ⊥ y(t) ≥ 0 is a compact representation of the following three conditions: i) m u ∈ Rm + , ii) y ∈ R+ , and iii) hu(t), y(t)i = 0. A solution of the linear complementarity system (1) is any tuple (x, u, y, v) such that x : R+ → Rn is an absolutely continuous function and (x, u, y, v) satisfies (1) for almost all forward times t ∈ R+ . In general, we assume that the initial conditions x(0) = x0 are such that the complementarity conditions hold. This implies the absence of jumps in the initial condition and in the complementarity variables [9]. B. Incremental passivity A cornerstone in the analysis of linear complementarity systems is to observe that the complementarity relation R⊥ = {(y, ζ) ∈ Rm × Rm |0 ≤ y ⊥ −ζ ≥ 0} (2) defines an incrementally passive relation. We recall the definition and the proof of that property. Definition 1: A relation R ⊆ Rm × Rm is incrementally passive if for any (y1 , ζ1 ) ∈ R, and any (y2 , ζ2 ) ∈ R the inequality hy1 − y2 , ζ1 − ζ2 i ≥ 0 (3) holds. Proposition 2: R⊥ is incrementally passive. Proof: Take (y1 , ζ1 ) ∈ R⊥ and (y2 , ζ2 ) ∈ R⊥ . Then, hy1 − y2 , ζ1 − ζ2 i = h−y2 , ζ1 i + hy1 , −ζ2 i + hy1 , ζ1 i + hy2 , ζ2 i = hy2 , −ζ1 i + hy1 , −ζ2 i ≥ 0. | {z } | {z } | {z } | {z } =0 =0 ≥0 ≥0 An alternative description of the relation R⊥ is via the multivalued map ζ = ϕ⊥ (y) ∈ {ζ ∈ Rm |0 ≤ y ⊥ −ζ ≥ 0} (4) which leads to the feedback representation in Figure 1. This Lure type representation of linear complementarity Fig. 1. systems calls for an analysis rooted in passivity theory: the linear complementary system is incrementally passive as the negative feedback interconnection of a linear passive system Σ with an incrementally passive relation. Passivity of the linear part is a standard assumption in the literature on linear complementarity systems; it guarantees existence and uniqueness of solutions for (1). Details can be found in [9], [8], [19] based on the following additional assumption, which ensures well-posedness of the closed loop in the presence of the throughput term D, [9]. Assumption 3: The linear part of (1) is a minimal realization, it is passive and the matrix   B (5) D + D⊤ V̇ (x) ≤ hy, ui y (6) where the derivative of the quadratic storage V := xT P x, P = P T > 0, is computed along the linear dynamics. Passivity and incremental passivity coincide for linear systems. In fact, for any pair of trajectories xi , outputs yi and inputs ui , the incremental dynamics characterized by the variables ∆x = x1 − x2 , ∆y = y1 − y2 , ∆u = u1 − u2 satisfies V̇ (∆x) ≤ h∆y, ∆ui . |x1 (t) − x2 (t)| ≤ β(|x1 (0) − x2 (0)|) ∀t ≥ 0 for any pair of trajectories x1 (·), x2 (·) of (1). Furthermore, if the passive inequality (6) is strict, the resulting closed loop becomes incrementally asymptotically stable; its trajectories converge towards each other lim |x1 (t) − x2 (t)| = 0 . t→∞ In this case, an equilibrium point is necessarily unique and globally stable. The concept of incremental stability [20] is analog to the concept of differential stability in the theory of contraction [3] or convergent systems [4]. But it does not require any differentiability of the system dynamics. C. Beyond linear complementarity relations For the purpose of this paper, the linear complementarity condition can be replaced by any incrementally passive static relation. Figure 2 provides an illustration of incrementally passive memoryless nonlinearities. Those multivalued maps are widely used for modeling electronic circuits. For example, the three graphs in Figure 2 represent the ideal currentvoltage characteristic of a diode, of a zener diode, and of an array of diodes [21], [22]. Denoting (w, z) ∈ Ri any pair Block diagram of the linear complementarity system (1) has full column rank. Passivity of the linear part of (1) from u to y reads incrementally stable, that is, there exists a nondecreasing function β such that (7) Since the negative feedback interconnection of incrementally passive systems is incrementally passive [15], the closed loop linear complementarity system is incrementally passive. For any constant input v, the resulting closed loop is thus Fig. 2. Popular examples of incrementally passive relations. (w, z) that belongs to the i-th relation in Figure 2, we extend the class of linear complementarity systems to the family of systems of the form   ẋ = Ax + Bu + Bv (8) y = Cx   (y, −u) ∈ Ri . Following the approach of linear complementarity systems, for a solution of (8) we mean any tuple (x, u, y, v) such that x : R+ → Rn is an absolutely continuous function and (x, u, y, v) satisfies (8) for almost all forward times t ∈ R+ . Indeed, the closed loop (8) allows for the block diagram representation in Figure 1, with ϕ⊥ replaced by ϕRi , the static multivalued map associated to the relation Ri . Thus, for any passive relation Ri , the closed loop (8) is incrementally passive. We remark that incremental passivity guarantees existence and uniqueness of solutions also for general maximal monotone static multi-valued maps, [8]. III. D OMINANCE AND p- DISSIPATIVITY A. Dominance Dominance was recently introduced in [1], [2], [6] as a generalization of incremental stability for smooth nonlinear systems. Motivated by dominance analysis of linear complementarity systems, we extend the definition of dominance in a non smooth setting, replacing differential analysis by incremental analysis as in the previous section. For the sake of simplicity in this section, and the rest of the paper, we consider generic pairs of trajectories x1 (·) and x2 (·), and we adopt the notation ∆x = x1 − x2 to denote their mismatch. ∆ẋ = ẋ1 − ẋ2 is defined for almost every t by the right-hand side of (8) computed for x1 and x2 , respectively. A similar notation is adopted for inputs ∆u = u1 − u2 and outputs ∆y = y1 − y2 . Finally, we say that a symmetric matrix P has inertia {p, 0, n − p} when it has p negative eigenvalues and n − p positive eigenvalues. Definition 4: The nonsmooth system (8) is p-dominant with rate γ ≥ 0 if there exist a matrix P = P ⊤ with inertia {p, 0, n − p} and a constant ε ≥ 0 such that for any pair trajectories of (8),     ∆ẋ 0 P ∆ẋ ≤ 0. (9) ∆x P 2γP + εI ∆x Strict p-dominance holds for ε > 0. Note that dominance is just incremental stability if P is positive definite, which corresponds to p = 0. But we are interested in the generalization corresponding to p = 1 and to p = 2. For smooth closed systems ẋ = f (x), (9) is equivalent to the linear matrix inequality inequality ∂f (x)⊤ P + P ∂f (x) + 2γP + εI ≤ 0 ∀x ∈ Rn (10) where ∂f (x) denotes the Jacobian of f at x. In the linear case, f (x) = Ax, (10) implies the existence of an invariant splitting such that Rn = En ⊕ En−p , where Ep is the pdimensional eigenspace associated to the dominant modes of A (eigenvalues of A whose real part is larger than −γ), and En−p is the (n − p)-dimensional eigenspace associated to the transient modes of A (i.e., the eigenvalues of A whose real part is smaller than −γ). Roughly speaking, in the nonlinear case the property of dominance forces the asymptotic behavior to be p-dimensional [2, Theorem 2], as shown by the analysis of the linearized flow in [2, Theorem 1]. The following theorem extends [2, Theorem 2] to the non-smooth case. Theorem 5: Assume v constant and suppose that all the trajectories of (13) are bounded. Let Ω(x) be the set of all ω-limit points of x and let (13) be strictly p-dominant with rate γ ≥ 0. Then, the flow on the Ω(x) is topologically equivalent to the flow of a p-dimensional system. Proof: Consider any pair of trajectories x1 (·) and x2 (·), define the increment ∆x(·), and consider the quadratic form V (∆x(t)) = ∆x(t)T P ∆x(t). From (9), d V (∆x(t)) ≤ −2γV (∆x(t)) − εk∆x(t)k2 , dt therefore d 2γt e V (∆x(t)) ≤ −εe2γt k∆x(t)k2 . dt By integration, Z t e−2γ(t−τ )k∆x(τ )k2 dτ V (∆x(t)) ≤ e−2γt V (∆x(0)) − ε 0 (11) Let x̄1 and x̄2 be two different points of Ω(x) and define ∆x̄ = x̄1 − x̄2 . Note that both x̄1 and x̄2 are accumulation points of a suitable trajectory, therefore ∆x̄ 6= 0 and (11) implies V (∆x̄ ) < 0. (12) Let HP , VP be the eigenspaces of P associated to the p negative eigenvalues of P , and n − p positive eigenvalues of P , respectively. Let Π : Rn → HP be the projection onto HP along VP . We claim that Π restricted to Ω is one-toone. In fact, assume by contradiction that for x̄1 , x̄2 ∈ Ω(x), x̄1 6= x̄2 implies Π(∆x̄ ) = 0, it follows that ∆x̄ ∈ VP and therefore V (∆x̄ ) > 0 which contradicts (12). Hence, Π restricted to Ω(x) is one-to-one. Now, for each constant input v, consider the equivalent representation of the system (8) based on the differential inclusion ẋ ∈ Fv (x) . (13) Using the results above, if y ∈ ΠΩ(x) then there exists a unique initial condition z(0) ∈ Ω(x) such that y = Πz(0) and the flow Πz(t) in HP is generated by the vector field Gv (y) = ΠFv (Π−1 y), y ∈ Ω(x) (14) which is p-dimensional. Theorem 5 shows that the asymptotic behavior of a strict p-dominant system is strongly constrained for small values of p. Corollary 6: Let the assumptions of Theorem 5 hold. In addition, assume that solutions of (8) are unique. Then all solutions asymptotically converge to 1) a unique equilibrium point, if p = 0. 2) an equilibrium point, if p = 1. 3) an equilibrium point, a set of equilibrium points and connecting arcs, or a limit cycle, if p = 2. Under uniqueness of solutions, distinct trajectories cannot intersect. For p = 1, the asymptotic dynamics are onedimensional, forcing bounded trajectories to converge to some fixed point. Uniqueness of solutions is also sufficient for guaranteeing the validity of the Poincaré-Bendixson Theorem, see e.g., [23, Theorem 5.3]. Hence, under the assumption of uniqueness of solutions, a 2-dominant system with a compact limit set that contains no equilibrium point has a closed orbit. B. Incremental p-dissipativity C. p-Passivity of linear complementarity systems Dissipativity theory is an interconnection theory for stability analysis. In the same way, p-dissipativity is an interconnection theory for dominance analysis [2], [6]. It mimics standard dissipativity theory in the differential/incremental setting but relies on quadratic storage functions that have a prescribed inertia. Definition 7: A nonsmooth system (8) is p-dissipative with rate γ ≥ 0 and incremental supply w : Rm × Rm → R  ⊤    ∆y Q L⊤ ∆y w(∆u, ∆y) := (15) ∆u L R ∆u As in the classical theory, p-passivity is p-dissipativity for the particular supply rate  ⊤    ∆y 0 I ∆y w(∆u, ∆y) := (17) ∆u I 0 ∆u if there exists a matrix P = P ⊤ with inertia {p, 0, n − p} and ε ≥ 0 such that     ∆ẋ 0 P ∆ẋ ≤ w(∆y, ∆u) (16) ∆x P 2γP + εI ∆x for any pair of trajectories. Strict p-dissipativity holds for ε > 0. 0-dissipativity coincides with the classical concept of incremental dissipativity. p-dissipativity allows for an interconnection theory for non-smooth systems, as clarified by the next theorem. The simplest example is given by the closed loop in Figure 1. Theorem 8: Let Σ1 and Σ2 be (strict) p1 and p2 dissipative respectively, both with rate γ ≥ 0 and supplies   i ⊤   ∆y Qi Li ∆y i wi (∆ui , ∆y i ) = , i = 1, 2. ∆ui L⊤ Ri ∆ui i The negative feedback interconnection u1 = −y 2 + v 1 , u2 = y 1 + v 2 of Σ1 and Σ2 is (strict) (p1 + p2 )-dissipative with respect to the input v := [v 1 , v 2 ]⊤ and the output y := [y 1 , y 2 ]⊤ , with incremental supply given by   ⊤ L1 R2    ⊤ Q1⊤+ R2 −L1 + L2 ∆y  + L2 Q2 + R1 −R1 L2   ∆y . −L1 ⊤ ∆v  R1 0  ∆v L1 −R1⊤ 0 R2 R2 L⊤ 2 and rate γ ≥ 0. In addition, if   Q1 + R2 −L1 + L⊤ 2 ≤0 −L⊤ Q2 + R1 1 + L2 then the interconnection is (strictly) (p1 + p2 )-dominant. Proof: The proof follows by standard arguments of dissipativity theory. See also [1]. Classical dissipativity theory provides a tool to analyze stable systems, that is, 0-dominant systems, as interconnection of dissipative open systems, that is, 0-dissipative systems. Theorem 8 generalizes this conclusion: 1-dominant systems can be analyzed as interconnections of 0-dissipative systems with a 1-dissipative system; 2-dominant systems can be analyzed as interconnections of 0-dissipative systems with a 2-dissipative system, or as interconnections of two 1-dissipative systems. We have seen in Section II that the linear complementarity relation R⊥ is 0-passive (since 0-passivity and incremental passivity coincide). Also the static nonlinearities Ri in Section II-C are 0-passive. Thus, from Theorem 8, the nonsmooth system (8) is the negative feedback loop of a 0passive static nonlinearity with a linear system, whose degree of p-passivity determines the degree of passivity of the closed loop. The degree of passivity of the linear part restricts the asymptotic behavior of the system. We observe that for linear systems of the form ẋ = Ax + Bu, y = Cx, the inequality (16) with (17) reduces to the simple feasibility test AT P + P A + 2γP ≤ −εI P B = CT for some matrix P = P T with inertia {p, 0, n − p}; a numerically tractable condition. Also, p-passivity has a frequency domain characterization based on the rate-shifted transfer function G(s − γ) = C(sI − (A + γI))−1 B, [6]: Proposition 9: A linear system is p-passive if and only if the following two conditions hold, 1) ℜ {G(jω − γ)} > 0, for all, ω ∈ R ∪ {+∞}. 2) G(s − γ) has p poles on the right-hand side of the complex plane. Frequency domain conditions prove useful in capturing the limits of the theory and for the selection of the systems parameters, as shown in the next section. IV. S WITCHING AND OSCILLATING LCS CIRCUITS A. The operational amplifier is 0-passive A model of the operational amplifier is shown in Figure 3: a first order model [24], with additional voltage saturation limits, implemented via ideal diodes, to take into account the physical limitations of any op-amp device. Note that the right-most element ×1 in the model denotes a buffer, which decouples the output current from the internal circuit. We make the usual assumption of infinite input impedance Zin = +∞ and 0 output impedance Zout = 0. The model in Figure 3 is described by the linear complementarity system α 1 1 xa + VE − (ID1 (t) − ID2 ) (18a) ẋa = − Ra Ca Ca Ca V0 = xa (18b) 0 ≤ −V0 + E1 ⊥ ID1 ≥ 0, (18c) 0 ≤ V0 + E2 ⊥ ID2 ≥ 0. (18d) where we assume that the voltage sources E1 and −E2 are constant and Ei > 0, i = 1, 2. The model is derived via Kirchhoff’s laws together with the complementarity conditions for the diodes [21]. xa denotes the voltage across the capacitor Ca . The input is set to VE and the output is V0 . For completeness, we write (18) in the standard linear complementarity form (1) by taking VE = V1 − V2 , Fig. 4. Block diagram of the operational amplifier (18) contains an op-amp, whose model is given by (18), and a linear network Σc represented by ( 2 ẋ1 = − RR11R+R x1 + R21C1 ν1 2 C1 Σc : Schmitt (20) y1 (t) = −x1 , Fig. 3. A nonsmooth model of an operational amplifier where x1 is the voltage across the capacitor C1 . Their u = [ID1 , ID2 , E1 , E2 ]⊤ , v = [−αV1 , −αV2 , 0, 0]⊤ , A = − Ra1Ca , B = C1a [−1, 1, 0, 0],     −1 0 0 1 0 1  0 0 0 1    C=  0  , and D = 0 0 0 0 . 0 0 0 0 0 A representation of (18) is the negative feedback interconnection of a strictly 0-passive system with a 0-passive relation R defined by (18b),(18c), and (18d), linking voltage V0 and the difference of diode currents IDD = ID1 − ID2 . The relations 0 ≤ −V0 (t) + E1 and 0 ≤ V0 (t) + E2 imply −E2 ≤ V0 (t) ≤ E1 . Thus, the output of the op-amp V0 (t) = xa (t) always belongs to the interval [−E2 , E1 ]. For V0 (t) = E1 , (18c) and (18d) imply that ID1 (t) ≥ 0 and ID2 (t) = 0, respectively, that is, IDD (t) ≥ 0. Similarly, V0 (t) = −E2 implies IDD (t) ≤ 0 and V0 (t) ∈ (−E2 , E1 ) implies IDD (t) = 0. Hence, R corresponds to the 0-passive relation represented in Figure 2.ii). Its associated multivalued function ϕR maps the voltage V0 into   (−∞, 0], V0 (t) = −E2 IDD = ϕR (V0 ) ∈ {0}, −E2 < V0 (t) < E1   [0, +∞), V0 (t) = E1 . The operational amplifier model (18) is thus given by the feedback loop in Figure 4, combining the 0-passive linear system Σa with matrices A = − Ra1Ca , B = C1a , and C = 1, with the 0-passive nonlinearity ϕR . The transfer function of Σa reads G(s) = 1 Ca s+ 1 Ra C a . (19) and Proposition 9 guarantees strict 0-passivity with rate γ ∈ [0, Ra1Ca ). By Theorem 8, the op-amp is thus a strictly 0passive device from VE to V0 with rate γ ∈ [0, Ra1Ca ). B. Positive feedback amplifier: multistable Schmitt trigger The positive feedback interconnection of the op-amp with an additional passive network leads to 1-passive circuits. For instance, the Schmitt trigger circuit represented in Figure 5 Fig. 5. Schmitt Trigger circuit formed as the positive feedback interconnection of the circuit in Figure 3 with a RC network interconnection is characterized by the positive feedback identity VE = x1 , ν1 = V0 . (21) Positive feedback loops of two 0-passive systems are not 0passive. But Σc is also strictly 1-passive from the input ν1 2 to the output y1 with rate γ ∈ ( RR11R+R , +∞), which allows 2 C1 to rewrite (21) as negative feedback VE = −y1 , ν1 = V0 . (22) Indeed, the positive feedback in (21) is equivalent to the negative feedback between a strictly 0-passive system, the opamp, and a strictly 1-passive system, Σc . Thus, by selecting the circuit parameters to satisfy 1 R1 + R2 < , R1 R2 C1 Ra Ca (23) Theorem 8 guarantees that the Schmitt trigger is strictly 12 , 1 ). The closed loop is passive with rate γ ∈ ( RR11R+R 2 C 1 Ra C a thus strictly 1-dominant. An interpretation of (23) is that the linear circuit Σc must have a slower dynamics than the op-amp dynamics, to determine a dominant behavior of dimension 1. Mathematically, (23) guarantees the existence of a common rate γ ≥ 0 for which op-amp and Σc are respectively 0-passive and 1-passive, as required by Theorem 8. A 1-passive circuit can be multistable. Bounded trajectories of a strictly 1-dominant system necessarily converge to a fixed point. Boundedness of trajectories follows from the saturation of the op-amp voltage, which essentially “opens the loop” for large overshoots. To enforce multistability, we look for circuit parameters that guarantee the existence of at least one unstable equilibrium point. The condition αR1 1 < , (24) Ra R1 + R2 makes the zero-equilibrium unstable. The parameters in Table I satisfies (23) and (24), enabling bistability. The value of resistance Ra and capacitance Ca have been taken from [24], to ensure good matching between simulations and the behavior of a real op-amp component. Figure 6 shows the trajectories of the system from two different initial conditions (xa (0), x1 (0)) ∈ {(−2, −2), (2, 2)}. Fig. 7. Relaxation Oscillator realized as the mixed positive and negative feedback of two RC networks and an op-amp For a2 = 0, there is no negative feedback and the system reduces to the Schmitt trigger. For a2 6= 0, the negative feedback loop either stabilizes the closed loop system, typically for parameter values that guarantee 0-passivity of G(s), or induces oscillations, typically for parameter values that guarantee 2-passivity of G(s). For instance, G(jω − γ) has positive real part if Fig. 6. Schmitt trigger trajectories from two different initial conditions. R1 = 1KΩ R2 = 1KΩ Ra = 1M Ω α = 0.1 C1 = 100µF Ca = 15.9nF E1 = 12V E2 = 12V TABLE I PARAMETER VALUES FOR THE SIMULATION OF THE S CHMITT TRIGGER . C. Mixed feedback amplifier: relaxation oscillator The interconnection of the op-amp with two slow passive networks, one with negative and one with positive feedback, leads to 2-passive circuits. For instance, the circuit in Figure 7 is a typical architecture for the generation of relaxation oscillations. It is derived from the Schmitt trigger through the addition of a slow network Σd , in negative feedback, represented by ( 4 ẋ2 = − RR33R+R x2 + R41c2 ν2 4 C2 Σd : (25) y2 = x2 , where x2 is the voltage across the capacitor C2 . The interconnection of the op-amp and of the two networks Σc and Σd is given by the mixed positive/negative feedback VE = x1 − x2 , ν1 = ν2 = V0 . (26) which can be written as a standard negative feedback loop using the aggregate output y = y1 + y2 , for instance VE = −y = −y1 − y2 , 1 R2 C 1 , 1 R4 C 2 , ν1 = ν2 = V0 . (27) R1 +R2 R2 R1 C 1 Taking a1 = a2 = b1 = and b2 = R3 +R4 , the aggregate transfer function from V to y reads 0 R3 R4 C 2 G(s) = − a1 a2 (a2 − a1 )s + a2 b1 − a1 b2 + = . s + b1 s + b2 (s + b1 )(s + b2 ) (28) a2 (b2 − γ) − a1 (b1 − γ) > 0 a2 (b1 − γ) − a1 (b2 − γ) > 0 Hence, by Proposition 9, if     1 a2 b1 − γ b2 − γ 0 < γ < min b1 , b2 , , > max , Ra Ca a1 b2 − γ b1 − γ then G(s) is strictly 0-passive with rate γ. The overall closed loop has a globally asymptotically stable fixed point. In contrast, for   a2 b1 − γ b2 − γ 1 , < min , 0 < max{b1 , b2 } < γ < Ra Ca a1 b2 − γ b1 − γ G(s) is strictly 2-passive with rate γ. Thus, by Theorem 8, the overall closed-loop is 2-dominant with rate γ. Indeed, γ divides the fast op-amp dynamics from the dominant two dimensional slow dynamics of the linear networks. We conclude the section with a numerical simulation. The parameters of Table I together with R3 = 3.3KΩ, R4 = 1KΩ and C2 = 200µF satisfy the conditions above, thus guarantee strict 2-dominance the closed loop with rate γ = 25. The fixed point in 0 is unstable but all trajectories remain bounded (the constraint xa ∈ [−E2 , E1 ] implies that x1 and x2 remain bounded), which enforces oscillations, as shown in Figure 8. V. CONCLUSIONS We extended the concept of dominance to the analysis of nonsmooth linear complementarity systems. The extension mimics the smooth case when uniqueness of solutions is assumed. The approach is based on the interconnection theory of dissipativity. It opens the way to the analysis of switching or oscillatory circuits with no restriction on the dimension of the state-space. The potential of the approach was illustrated with a detailed analysis of well-known circuits Fig. 8. Trajectories of the relaxation oscillator in Figure 7. xa (·) – thin/black line; x1 (·) – thick/gray line; x2 (·) – dashed line. based on op-amps, predicting multistable and oscillatory behaviors, while providing margins on the circuit parameters to enable such behaviors. R EFERENCES [1] F. Forni and R. Sepulchre, “A dissipativity theorem for p-dominant systems,” in 56th IEEE Conference on Decision and Control, Melbourne, Australia, 2017. [2] ——, “Differential dissipativity theory for dominance analysis,” 2017, submitted to IEEE Transactions on Automatic Control, http://arxiv.org/abs/1710.01721. [3] W. Lohmiller and J.-J. E. Slotine, “On contraction analysis for nonlinear systems,” Automatica, vol. 34, no. 6, pp. 683–696, 1998. [4] A. Pavlov, N. Van De Wouw, and H. Nijmeijer, “Convergent systems: analysis and synthesis,” in Control and observer design for nonlinear finite and infinite dimensional systems. Springer, 2005, pp. 131–146. [5] F. Forni and R. Sepulchre, “A differential Lyapunov framework for contraction analysis,” IEEE Transactions on Automatic Control, vol. 59, no. 3, pp. 614–628, 2014. [6] F. A. Miranda-Villatoro, F. Forni, and R. Sepulchre, “Analysis of Lur’e dominant systems in the frequency domain,” 2017, submitted to Automatica, http://arxiv.org/abs/1710.01645. [7] B. Brogliato, “On the control of non-smooth complementarity dynamical systems,” Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, vol. 359, no. 1789, pp. 2369–2383, 2001. [8] ——, “Some perspectives on the analysis and control of complementarity systems,” IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 918–935, 2003. [9] M. K. Çamlıbel, W. Heemels, and J. Schumacher, “On linear passive complementarity systems,” European Journal of Control, vol. 8, no. 3, pp. 220–237, 2002. [10] M. K. Çamlıbel, P. J. S, and J. L. Shen, “Lyapunov stability of linear complementarity systems,” SIAM Journal on Optimization, vol. 17, no. 4, pp. 1056–1101, 2006. [11] W. Heemels and B. Brogliato, “The complementarity class of hybrid dynamical systems,” European Journal of Control, vol. 9, no. 2, pp. 322–360, 2003. [12] B. Brogliato, Nonsmooth Mechanics, ser. Communications and Control Engineering. Springer, 2016. [13] V. Acary, O. Bonnefon, and B. Brogliato, Nonsmooth modeling and simulation for switched circuits, ser. Lecture Notes in Electrical Engineering. Springer, 2011. [14] A. Nagurney, Network economics: A variational inequality approach, ser. Advances in Computational Economics. Springer, 1999. [15] A. J. van der Schaft, L2-Gain and Passivity Techniques in Nonlinear Control, ser. Lecture Notes in Control and Information Sciences. Springer, 1996. [16] G. B. Stan and R. Sepulchre, “Analysis of interconnected oscillators by dissipativity theory,” IEEE Transactions on Automatic Control, vol. 52, no. 2, pp. 256–270, 2007. [17] A. Pavlov and M. L, “Incremental passivity and output regulation,” Systems & Control Letters, vol. 57, pp. 400–409, 2008. [18] T. Liu, D. J. Hill, and Z. J, “Incremental-dissipativity-based synchronization of interconnected systems,” in Proceeding of the 18th World Congress, Milano, Italy, 2011, pp. 8890–8895. [19] F. A. Miranda-Villatoro and F. Castaños, “Robust output regulation of strongly linear passive systems with multivalued maximally monotone controls,” IEEE Transactions on Automatic Control, vol. 62, no. 1, pp. 238–249, 2017. [20] D. Angeli, “A Lyapunov approach to incremental stability properties,” IEEE Transactions on Automatic Control, vol. 47, pp. 410–421, 2000. [21] V. Acary and B. Brogliato, “Implicit Euler numerical scheme and chattering-free implementation of sliding mode systems,” Systems & Control Letters, vol. 59, no. 5, pp. 284–293, 2010. [22] Y. J. Lootsma, A. J. van der Schaft, and M. K. Çamlıbel, “Uniqueness of solutions of linear relay systems,” Automatica, vol. 35, pp. 467–478, 1999. [23] K. Ciesielski, “The Poincaré-Bendixson theorems for two dimensional semiflows,” Topological Methods in Nonlinear Analysis, vol. 3, pp. 163–178, 1994. [24] J. Karki, “Effect of parasitic capacitance in op amp circuits,” Texas Instruments, Tech. Rep. SLOA013A, 2000.
3cs.SY
arXiv:1301.1468v3 [math.AC] 18 Oct 2013 Annihilators of Artinian modules compatible with a Frobenius map Mordechai Katzman1 Department of Pure Mathematics, University of Sheffield, Hicks Building, Sheffield S3 7RH, United Kingdom Wenliang Zhang2 Department of Mathematics, University of Nebraska, 203 Avery Hall, Lincoln, NE 68588, USA Abstract In this paper we consider Artinian modules over power series rings endowed with a Frobenius map. We describe a method for finding the set of all prime annihilators of submodules which are preserved by the given Frobenius map and on which the Frobenius map is not nilpotent. This extends the algorithm by Karl Schwede and the first author, which solved this problem for submodules of the injective hull of the residue field. The Matlis dual of this problem asks for the radical annihilators of quotients of free modules by submodules preserved by a given Frobenius near-splitting, and the same method solves this dual problem in the F -finite case. Keywords: Frobenius map, Frobenius splitting 1. Introduction This paper describes an algorithm for finding the annihilators of submodules of Artinian modules which are preserved by a given Frobenius map. Throughout this paper R will denote a ring of formal power series over a field K of prime characteristic p, m will denote its maximal ideal, and E = ER (R/m) will denote the injective hull of its residue field. The Frobenius map sending r ∈ R to its pth power will be denoted f , and f e will be its eth iteration. Given any R-module M and e ≥ 0, we may endow M with a new R-module structure e given by r · m = rp m for all r ∈ R and m ∈ M and we denote this new module F∗e M . An Email addresses: [email protected] (Mordechai Katzman), [email protected] (Wenliang Zhang) 1 The first author thankfully acknowledges support from EPSRC grant EP/I031405/1. 2 The second author is supported in part by NSF Grant DMS #1068946. Preprint submitted to Elsevier June 19, 2017 eth Frobenius map on M is an element of HomR (M, F∗e M ), or, equivalently, an additive map e φ : M → M such that φ(rm) = rp φ(m) for all r ∈ R and m ∈ M . Given such a Frobenius map φ ∈ HomR (M, F∗e M ) we call an R-submodule N ⊆ M φ-compatible if φ(N ) ⊆ F∗e N . When discussing the case e = 1, we shall drop the e from the notation above. The aim of this paper is to find the set of radical annihilators of all φ-compatible submodules of a given Artinian R-module, or, equivalently (cf. Proposition 2.1 below), given a φ ∈ HomR (E α , F∗ E α ) for some positive integer α, to find all radical annihilators of Rsubmodules N ⊆ E α which satisfy φ(N ) ⊆ F∗ N . We shall accomplish this under the assumption that this φ restricts to a non-zero map on N : in this case the set of radical annihilators is shown to be finite and given by the intersection of all prime ideals in it (cf. [S, Corollary 3.11] and [EH, Section 3].) This extends the results in [KS] which describes an algorithm for producing such sets of annihilators when α = 1. We shall first describe this algorithm from a more algebraic point of view than in [KS] and comment on why it cannot be directly extended for α > 1. We shall then give a description of Frobenius maps on E α in terms of certain matrices and finally we will produce an algorithm which works recursively on α, in which the case α = 1 treated in [KS] provides the foundation. This paper is organized as follows. Section 2 introduces the notion of Frobenius maps and studies Frobenius maps of Artinian modules using the properties of a version of Matlis duality which keeps track of the Frobenius maps: these are the functors ∆e and Ψe described there. In section 3 we generalize two operations which were originally introduced in the context of Frobenius splittings and Frobenius maps in the injective hulls of residue fields, namely, e the Ie (−) operation (denoted [1/p ] by some authors) and the ⋆-closure. These are extended from operations on ideals to operations on submodules of free modules, and some of their properties are studied here, e.g., their behaviour under localization. Section 4 reviews the algorithm in [KS] for finding prime annihilators of submodules of the injective hull of the residue field stable under a given Frobenius map, and presents a proof for its main ingredient in algebraic language. The main section of this paper, section 4 generalizes the Katzman-Schwede algorithm to deal with prime annihilators of general Artinian modules endowed with a Frobenius map. The main result, Theorem 5.5, yields an algorithm which is described in detail in section 6. We also carry out two calculations following the algorithm to illustrate its use. Finally, section 7 translates the previous results into the language of Frobenius nearsplittings of free modules in the case where we work over an F -finite ring: in this setup Frobenius maps and near-splittings are dual notions. 2 2. Frobenius maps of Artinian modules and their stable submodules In this section we describe all Frobenius maps on Artinian R-modules. We may think of eth Frobenius maps as left-module structures over the following skew-commutative rings i pe R[Θ; f e ]: as an R-module it is the free module ⊕∞ i=0 RΘ and we extend the rule Θr = r Θ for all r ∈ R to a (non-commutative!) multiplicative structure on R[Θ; f e ]. Now given an eth Frobenius map φ on an R-module M , we can turn it into a left R[Θ; f e ]-module by extending the rule Θm = φ(m) for all m ∈ M . The fact that this gives M the structure of a left R[Θ; f e ]-module is simply because for all r ∈ R and m ∈ M , e e Θ(rm) = φ(rm) = rp φ(m) = rp Θm = (Θr)m. Conversely, if M is a left R[Θ; f e ]-module, then Θ : M → M is an eth Frobenius map. Recall the definition of the eth Frobenius functor : the tensoring (−) → F∗e R ⊗R (−) defines a functor from the category of R-modules to the category of F∗e R-modules. We may now identify the rings R and F∗e R and to obtain eth Frobenius functor FRe (−) from the category of R-modules to itself. Following [K1] we shall refer to the category of Artinian R[Θ; f e ]-modules C and the category D of R-linear maps M → FRe (M ) where M is a finitely generated R-module, and a b where a morphism between M − → FRe (M ) and N − → FRe (N ) is a commutative diagram of R-linear maps µ /N . M a  FRe (M ) b e FR (µ)  / FR (N ) We also refer to the mutually inverse functors ∆e : C → D and Ψe : D → C also introduced in [K1]. These are extensions of Matlis duality functors (−)∨ = HomR (−, E) which, additionally, keep track of Frobenius actions and are defined as follows. Given an R[Θ; f e ]-module M , ∆e (M ) is defined (functorially) as Matlis dual of the R-linear map F∗e R ⊗R M → M given by r ⊗ m 7→ rΘm where (F∗e R ⊗R M )∨ is identified with FRe (M ∨ ) (cf. [L, Lemma 4.1].) Given an R-linear map in De , one can reverse the steps of the construction of ∆e and obtain functorially an Artinian module with a Frobenius map defined on it. As before, we will suppress e from the notation when e = 1. Given an Artinian R-module M we can embed M in E α for some α and extend this inclusion to an exact sequence At 0 → M → E α −−→ E β → . . . 3 ∼ HomR (Rα , Rβ ) is a β × α matrix with entries in R. Proposiwhere At ∈ HomR (E α , E β ) = tion 2.1 below shows that the Frobenius maps on M are restrictions of Frobenius maps on E α and those can be described in terms of the following canonical Frobenius map T : E α → E α . − Since R is regular local, E is isomorphic to the module of inverse polynomials K[[x− 1 , . . . , xd ]] where x1 , . . . , xd are minimal generators of the maximal ideal of R (cf. [BS, §12.4].) Thus E 1 d has a natural R[T ; f ]-module structure additively extending T (λx−α . . . x−α ) = λp x1−pα1 . . . x1−pαd 1 1 for λ ∈ K and α1 , . . . , αd > 0. We can further extend this to a natural R[T ; f ]-module structure on E α given by     a1 T a1  .   .     T  ..  =  ..  . aα T aα Proposition 2.1. Let M = ker At be an Artinian R-module where A is a α × β matrix with e entries in R. Let e ≥ 1 and let B be the set of α×α matrices which satisfy Im BA ⊆ Im A[p ] . e [pe ] For a given eth Frobenius map on M , ∆ (M ) ∈ HomR (Coker A, Coker A ) and is given by multiplication by a matrix B in B and, conversely, any such B defines an R[Θ; f e ]-module structure on M which is given by the restriction to M of the Frobenius map φ : E α → E α defined by φ(v) = B t T e (v) where T is the natural Frobenius map on E α . A Proof. Matlis duality gives an exact sequence Rβ − → Rα → M ∨ → 0 hence e ∆e (M ) ∈ HomR (M ∨ , FRe (M ∨ )) = HomR (Coker A, Coker A[p ] ). e Let ∆e (M ) be the map φ : Coker A → Coker A[p ] . In view of Theorem 3.1 in [K1] we only need to show that any such R-linear map is given by multiplication by a matrix B in B, and that any such B defines an element in ∆e (M ). e The freeness of Rα enables us to lift the map φ : Coker A → Coker A[p ] to a map φ′ : Rα → Rα given by multiplication by some α × α matrix B in B. Conversely, any such e matrix B defines a map φ : Coker A → Coker A[p ] , and Ψe (φ) is a Frobenius map on M as described in the statement of the proposition. In the rest of the paper we shall consider Frobenius actions Θ = U T on E α and R[Θe ; f e ] submodules M ⊆ E α . The proposition above shows that for any such M there is a V ⊆ Rα e such that M = annE α V t := {z ∈ E α | V t z = 0} and U V ⊆ V [p ] . This will be henceforth used extensively and implicitly. For simplicity we adopt the following notation: given any V ⊆ Rα we define E (V ) = annE α V t . 3. Extending the ⋆-closure The purpose of this section is to extend the ⋆-closure operation as first defined in section 5 of [K1]. Definition 3.1. Let e ≥ 0. 4 e (a) Given any matrix (or vector) A with entries in R, we define A[p ] to be the matrix obtained from A by raising its entries to the pe th power. e (b) Given any submodule K ⊆ Rα , we define K [p ] to be the R-submodule of Rα generated e by {v [p ] | v ∈ K}. The theorem below extends the Ie (−) operation defined on ideals in [K1, Section 5] and e in [BMS, Definition 2.2] (where it is denoted (−)[1/p ] ) to submodules of free R-modules. Theorem 3.2. Let e ≥ 1. (a) Given a submodule K ⊆ Rα there exists a minimal submodule L ⊆ Rα for which e K ⊆ L[p ] . We denote this minimal submodule Ie (K). (b) Let U be a α × α matrix with entries in R and let V ⊆ Rα . The set of all submodules e W ⊆ Rα which contain V and which satisfy U W ⊆ W [p ] has a unique minimal element. e Proof. Let L be the intersection of all submodules M ⊆ Rα for which K ⊆ M [p ] . Proposie tion 5.3 in [K1] implies that K ⊆ L[p ] and clearly, L is minimal with this property. To prove (b) we carry out a construction similar to that in [K1, section 5]. Define inductively V0 = V and Vi+1 = I1 (U Vi ) + Vi for all i ≥ 0. The sequence {Vi }i≥0 must stabilize to some submodule W = Vj ⊆ Rα . Since W = I1 (U W ) + W , I1 (U W ) ⊆ W and U W ⊆ W [p] . Let Z be any submodule of Rα containing V for which U Z ⊆ Z [p] . We show by induction on i that Vi ⊆ Z for all i ≥ 0. Clearly, V0 = V ⊆ Z, and if for some i ≥ 0, Vi ⊆ Z then U Vi ⊆ U Z ⊆ Z [p] hence I1 (U Vi ) ⊆ Z and Vi+1 ⊆ Z. This shows that W ⊆ Z. Definition 3.3. With notation as in Theorem 3.2, we call the unique minimal submodule in 3.2(b) the star closure of V with respect to U and denote it V ⋆U . The effective calculation of the ⋆-closure boils down to the calculation of Ie , and this is a straightforward generalization of the calculation of Ie for ideals. To do so, we first note that if R is a free Rp -module with free basis B (e.g., when dimKp K < ∞), then every element P [pe ] v ∈ Rα can be expressed uniquely in the form v = b∈B ub b where ub ∈ Rα for all b ∈ B. Proposition 3.4. Let e ≥ 1. (a) For any submodules V1 , . . . , Vℓ ⊆ Rn , Ie (V1 + · · · + Vℓ ) = Ie (V1 ) + · · · + Ie (Vℓ ). (b) Assume that R is a free Rp -module with free basis B (e.g., when dimKp K < ∞). Let v ∈ Rα and let X [pe ] v= ub b b∈B be the unique expression for v where ub ∈ Rα for all b ∈ B. Then Ie (Rv) is the submodule W of Rα generated by {ub | b ∈ B}. Proof. The proof of this proposition is a straightforward modification of the proofs of propositions 5.2 and 5.6 in [K1] and Lemma 2.4 in [BMS]. 5 Clearly, Ie (V1 + · · · + Vℓ ) ⊇ Ie (Vi ) for all 1 ≤ i ≤ ℓ, hence Ie (V1 + · · · + Vℓ ) ⊇ Ie (V1 ) + · · · + Ie (Vℓ ). On the other hand e e e (Ie (V1 ) + · · · + Ie (Vℓ ))[p ] = Ie (V1 )[p ] + · · · + Ie (Vℓ )[p ] ⊇ V1 + · · · + Vℓ and the minimality of Ie (V1 + · · · + Vℓ ) implies that Ie (V1 + · · · + Vℓ ) ⊆ Ie (V1 ) + · · · + Ie (Vℓ ) and (a) follows. e Clearly v ∈ W [p ] , and so Ie (Rv) ⊆ W . On the other hand, let W be a submodule of Ps e [pe ] Rα such that v ∈ W [p ] . Write v = i=1 ri wi for ri ∈ R and wi ∈ W for all 1 ≤ i ≤ s, P pe and for each such i write ri = b∈B rbi b where rbi ∈ R for all b ∈ B. Now ! s X [pe ] X X pe [pe ] b rbi wi ub b = v = b∈B b∈B i=1  P [pe ] s pe [pe ] r w and since these are direct sums, we compare coefficients and obtain ub = i=1 bi i Ps for all b ∈ B and so ub = ( i=1 rbi wi ) for all b ∈ B hence ub ∈ W for all b ∈ B. Lemma 3.5 (cf. [M]). Let S ⊂ R be a multiplicative set, and let W ⊆ Rα . For all e ≥ 1, Ie (S−1 W ) exists and equals S−1 Ie (W ). e e Proof. We first note that Ie (S−1 W ∩ Rα )[p ] ⊇ S−1 W ∩ Rα hence S−1 Ie (S−1 W ∩ Rα )[p ] ⊇ S−1 (S−1 W ∩ Rα ) = S−1 W . [pe ] Let W1 ⊆ Rα be another module for which S−1 W1 ⊇ S−1 W ; we have [pe ] [pe ] S−1 W1 ∩ Rα = S−1 W1 ∩ Rα ⊇ S−1 W ∩ Rα  hence Ie (S−1 W ∩Rα ) ⊆ S−1 W1 ∩Rα and S−1 Ie (S−1 W ∩Rα ) ⊆ S−1 S−1 W1 ∩ Rα = S−1 W1 . We can now conclude that Ie (S−1 W ) exists and equals S−1 Ie (S−1 W ∩ Rα ). We now have Ie (S−1 W ) = S−1 Ie (S−1 W ∩ Rα ) ⊇ S−1 Ie (W ), and we finish the proof by showing that Ie (S−1 W ) ⊆ S−1 Ie (W ). This last inclusion is equivalent to S−1 W ⊆ [pe ] e S−1 Ie (W ) and this follows from the fact that W ⊆ Ie (W )[p ] . The existence of the Ie (−) operation in localizations of Rα allows us to define ⋆U operations on submodules of these localizations in an identical way to its definition for submodules of Rα . We shall use these later in Section 5. Throughout the rest of this section we fix a Frobenius map Θ = U t T : E α → E α where U is an α × α matrix with entries in R. Recall that, given any V ⊆ Rα we use E (V ) to denote annE α V t . We will collect some properties of E (V ) which will be used later in Section 5. Lemma 3.6 (cf. Theorem 4.7 in [K1]). The R[Θ; f ]-submodule Z = {a ∈ E α | Θe a = 0} is given by   e−1 e−2 E Ie (Im U [p ] U [p ] . . . U ) . 6 Proof. Write Z = E (W ) for some R-submodule W ⊆ Rα . We may view E α and Z as R[Θe ; f e ]-modules and an application ∆e to the inclusion Z ⊆ E α gives a commutative diagram with exact rows Rα U [p e−1 ] U [p e−2 ] / Rα /W U [p ···U /0 e−1 ]  / Rα /W [pe ]  Rα U [p e−2 ] ···U /0 where the right-most vertical map is zero. We deduce that W is the smallest submodule of e−1 e−2 Rα for which the rightmost vertical map is zero, i.e., W = Ie (Im U [p ] U [p ] · · · U ). Lemma 3.7 (cf. Theorem 4.8 in [K3]). Let K ⊆ Rα and assume that E (K) is an R[Θ; f ]module. The R[Θ; f ]-module M = {z ∈ E α | Θz ∈ E (K)} is E (I1 (U K)). Proof. Since E (K) is an R[Θ; f ]-module, E (K) ⊆ M . Write M = E (V ) for some V ⊆ Rα and apply ∆1 to the short exact sequence 0 → E (K) → E (V ) → E (V ) / E (K) → 0 of R[Θ; f ]-modules to obtain the following commutative diagram with exact rows 0 / K/V 0  / K [p] /V [p] / Rα /V U / Rα /K U  / Rα /V [p] /0 U  / Rα /K [p] /0 and V is the smallest submodule of Rα on which the leftmost vertical map vanishes, i.e., V = I1 (U K). Lemma 3.8. Let E (W ) be a R[Θ; f ]-submodule, where W ⊆ Rα . Write J = (0 :R E (W )) = α c (0 :R Rα /W) and  let Q be an associated prime of J. There   exists an R-submodule W ⊆ R c is a R[Θ; f ]-submodule and (0 :R E W c ) = (0 :R Rα /W c ) = Q. such that E W √ Proof. Let q1 ∩ · · · ∩ qs be a minimal primary decomposition of J with Q = q1 . Pick an c = (W :Rα a) := {v ∈ Rα | av ∈ W }. It is a ∈ R for which (J : a) = Q. and write W α c straightforward to verify that (0 :R R /W ) = (J : a) = Q and since U W ⊆ W [p] , we have c ⊆ ap−1 U W ⊆ ap−1 W [p] ⊆ W [p] ap U W c ⊆ (W [p] :Rα ap ) = (W :Rα a)[p] . and so U W Next, we want to introduce the following terminology. Definition 3.9. Let Θ = U t T : E α → E α (where U is a α × α matrix with entries in R) be a Frobenius map. We shall call an ideal Θ-special (or just special if Θ is understood) if it is an annihilator of an R[Θ; f ]-submodule of E α . Equivalently, an ideal is Θ-special if it is the annihilator of Rα /W where U W ⊆ W [p] . A Θ-special prime ideal shall be referred to as being Θ-special prime. 7 A basic fact concerning special primes is the following. ⋆U Lemma 3.10. Let P ⊂ R be a special prime and V = (P Rα ) largest R[Θ; f ]-module whose annihilator is P . ⊆ Rα . Then E (V ) is the Proof. The construction of the ⋆-closure guarantees that E (V ) is an R[Θ; f ]-module, and this is clearly annihilated by P . If E (W ) is another R[Θ; f ]-module annihilated by P then ⋆U P Rα ⊆ W and (P Rα ) ⊆ W ⋆U = W and hence E (V ) ⊇ E (W ) and the annihilator of both is P . 4. The case α = 1 In this section we describe an algorithm for finding all submodules of E (P ) ⊂ E which are preserved by a given Frobenius map Θ = uT (u ∈ R), under the assumptions that P ⊂ R is prime and that the restriction of Θ : E → E to E (P ) is not the zero map. This algorithm is essentially the one described in [KS], however, we present it here in terms of R[Θ; f ]-submodules of E rather than in terms of Frobenius splittings and we do so in more algebraic language. Fix u ∈ R and Θ = uT throughout the rest of this section. Theorem 4.1 (cf. section 4 in [KS]). Let P ⊂ Q be prime Θ-special ideals, write S = R/P . Let J ⊆ R be an ideal whose image in S defines its singular locus. (a) If (P [p] : P )Q ⊆ Q[p] then J ⊆ Q. (b) If (P [p] : P )Q * Q[p] then (uR + P [p] : (P [p] : P )) ⊆ Q. (c) Assume further that R is F -finite. If the restriction Θ to E (P ) is not the zero map, then (uR + P [p] : (P [p] : P )) ) P . Proof. Write ERQ = ERQ (RQ /QRQ ). Note that RQ is regular; let Te be the natural Frobenius on ERQ . e are e = ESQ (SQ /QSQ ) = annER P RQ and note that the Frobenius maps on E Write E Q [p] given by (P R : P RQ )Te and that the Frobenius maps on annER QRQ ⊂ annER P RQ = Q Q Q e are given by (QR[p] : QRQ )Te (cf. Proposition 4.1 in [K1]). E Q [p] [p] If (a), then (P RQ : P RQ ) ⊆ (QRQ : QRQ ), i.e., annERQ QRQ is an S[θ]-submodule of e e Now if J * Q, SQ is regular and E e is a simple S[τ ]-module E for all Frobenius maps θ on E. e e where τ : E → E is the natural Frobenius map, hence Q = P or Q = R, a contradiction. If (b), pick any c ∈ (uR + P [p] : (P [p] : P )). We have c(P [p] : P )Q ⊆ (uR + P [p] )Q ⊆ Q[p] and since (P [p] : P )Q * Q[p] we conclude c is a zero-divisor on R/Q[p] and c ∈ Q. To prove (c) we follow [F] and identify the S-module (P [p] : P )/P [p] with HomS (F∗ S, S) and u with a non-zero ψ ∈ HomS (F∗ S, S). We define C to be the S-submodule of HomS (F∗ S, S) generated by ψ. Now HomS (F∗ S, S) is a rank-one F∗ S-module (cf. [F, Lemma 1.6]) and hence there exists a non-zero c ∈ S which multiplies HomS (F∗ S, S) into C, and hence c multiplies (P [p] : P ) into uR + P [p] . 8 To turn this theorem into an algorithm, one would start with a given special prime P and find all special primes Q ) P for which there is no special prime strictly between P and Q. We shall henceforth refer to such special prime Q as minimally containing P . Corollary 4.2. (a) Any prime containing I1 (uR) is a special prime. (b) Let P be a special prime which does not contain I1 (uR). The set of special primes minimally containing P is finite. (c) Let P be a special prime such that uT is not nilpotent on E (P ). The set of special primes minimally containing P is finite. Proof. The first statement follows from Lemma 3.6 with e = 1, Θ = uT : if P ⊇ I1 (uR), the restriction of uT to E (P ) is zero. For P as in (b), any special prime ideal Q minimally containing P is either among the finitely many special primes not containing I1 (uR) or a special prime which contains I1 (uR) + P ) P , and in the latter case it is among the minimal primes of I1 (uR) + P . For (c) note that if uT is not nilpotent on E (P ), the restriction of uT to E (P ) is not zero, and Lemma 3.6 shows that P does not contain I1 (uR) A by-product of Theorem 4.1 and Corollary 4.2 is the algorithm described in [KS, section 3] which produces in the F -finite case all Θ-special primes P for which the restriction of Θ to E (P ) is not the zero map. As stated in the introduction, the aim of this paper is to extend this algorithm and produce the prime annihilators of submodules of E α preserved by a given Frobenius map which restricts to a non-zero map, and we shall do so in the subsequent sections. It might be instructive at this point to see why Theorem 4.1 is not useful when α > 1: while parts (a) and (b) of the Theorem hold in this extended generality, part (c) of the Theorem fails. The problem with (c) is that the module HomS (F∗ S α , S α ) is usually not cyclic when α > 1. 5. The case α > 1 The main aim of this section is to extend Corollary 4.2 to the case α > 1 and to obtain as a byproduct an algorithm for finding all special primes P with the property that for some R[Θ; f ]-submodule M ⊆ E α with (0 :R M ) = P , the restriction of Θ to M is not nilpotent. Theorem 5.1. The set of all special primes P with the property that for some R[Θ; f ]submodule M ⊆ E α with (0 :R M ) = P , the restriction of Θ to M is not zero, is finite. We will prove this theorem by induction on α; the case α = 1 being the content of Corollary 4.2. We shall assume henceforth in this section that α > 1 and that the theorem holds for α − 1 and that, additionally, as in the case α = 1, there is an effective way of finding the finitely many special primes in question. t We should note that, given a non-zero Frobenius action U ′ T : E (W ′ ) → E (W ′ ) with special prime Q = (0 : Rα−1 /W ′ ), the induction hypothesis gives us an effective method for 9 t finding this Q: for any other U ′ T -special prime P ⊂ Q, P Rα−1 t W ′ and hence the restriction of U ′ T to   ⋆U ′  ⋆U ′  ⊃ E QRα−1 E P Rα−1 ⋆U ′ t T ⊂ QRα−1 ⋆U ′ t T ⊆ is not nilpotent. Now we can enumerate all these special primes P starting with P = 0 and ascending recursively to bigger special primes until all such special primes are listed. For the rest of this section, we will fix a Frobenius map Θ = U t T : E α → E α (where U is a α × α matrix with entries in R) and we wish to find all the special primes with respect to Θ. The following lemma is our starting point of finding special primes Q ⊇ P when a special prime P is given. Lemma 5.2. Let Q be a special prime minimally containing the special prime P . Let ⋆U a ∈ Q \ P and write V = ((P + aR)Rα ) then Q is among the minimal primes of Rα /V . Proof. We have ⋆U (P Rα ) and so ⋆U ⊆ V ⊆ (QRα )     ⋆U ⋆U E (QRα ) ⊆ E (V ) ⊆ E (P Rα ) and looking at the annihilators of these we get P ⊆ (0 :R Rα /V ) ⊆ Q and we deduce that Q contains a minimal prime of Rα /V . This minimal prime is also special by Lemma 3.8 and since Q minimally contains P , this minimal prime must equal Q. Next, we want to treat a (crucial) special case: the α-th column of U is entirely zero. To this end, we need the following lemma, which will enable us to reduce the rank of U by one when we handle the aforementioned special case. Lemma 5.3. Assume α > 1. Let Q be a special prime, and let W ⊆ Rα be such that U W ⊆ W [p] and (0 :R Rα /W ) = Q. Let a ∈ / Q and let X be an invertible α × α matrix with entries in the localization Ra . Let ν ≫ 0 be such that U ′ = aν X [p] U X −1 has entries in R t and let W ′ = XWa ∩ Rα . Write Θ′ = U ′ T . Then (a) Q is a minimal prime of (0 :R Rα /W ′ ), [p] t (b) U ′ W ′ ⊆ W ′ and hence Q is U ′ T -special, and e (c) if the restriction of Θe to E (W ) is not zero, nor is the restriction Θ′ to E (W ′ ), Proof. We have (0 :R Rα /W ′ )a = (0 :Ra Raα /XWa ) = (0 :Ra Raα /Wa ) = (0 :R Rα /W )a = QRa and (a) follows. 10 For (b) consider the commutative diagram U Raα /Wa / Raα /Wa[p] X [p] X   X [p] UX/ −1 α [p] Raα /XWa Ra /X [p] Wa and compute U ′ W ′ = aν X [p] U X −1 (XWa ∩ Rα ) ⊆ (aν X [p] U X −1 XWa ) ∩ Rα ⊆ [p] (X [p] Wa[p] ) ∩ Rα = (XWa )[p] ∩ Rα = (XWa ∩ Rα )[p] = W ′ . The second statement in (b) now follows from (a) and Lemma 3.8. Lemma 3.6 shows that the restriction of Θe to E (W ) is not zero if and only if e−1 Ie (U [p ] e−2 U [p ] · · · U Rα ) 6⊆ W, i.e., if and only if e−1 U [p ] e−2 U [p ] e · · · U Rα 6⊆ W [p ] . The same Lemma shows that to prove (c) we need to verify that this implies that Ie (U ′ [pe−1 ] U′ [pe−2 ] · · · U ′ Rα ) 6⊆ W ′ , i.e., that U′ ν pν ···a Write b = a a U′ e e−1 and if bX [p ] U [p ] pe−1 ν e U′ [pe−2 ] · · · U ′ Rα 6⊆ W ′ ] U′ e−2 ] · · · U X −1 Rα ⊆ W ′ [p e−1 ] [pe−2 ] e−2 U [p [pe ] . . We calculate e−1 U [p X [p ] U [p [pe−1 ] ] e e−1 · · · U ′ = bX [p ] U [p · · · U Raα [pe ] ] e−2 U [p ] · · · U X −1 we may localize at a to obtain e e−1 = bX [p ] U [p ⊆ Wa′ e [p ] ] e−2 U [p ] · · · U X −1 Raα [pe ] = (XWa ) e e = X [p ] Wa[p ] e−1 e−2 e e hence U [p ] U [p ] · · · U Raα ⊆ Wa [p ] , and since a is not a zero-divisor on Rα /W [p e−1 e−2 e deduce U [p ] U [p ] · · · U Rα ⊆ W [p ] , contradicting our assumption. ] we We are now in position to deal with the following crucial special case. Proposition 5.4. Assume that the αth column of U is zero. Then Theorem 5.1 holds and there exists an effective method for finding the special primes P with the property that for some R[Θ; f ]-submodule M ⊆ E α with (0 :R M ) = P , the restriction of Θ to M is not zero. 11 Proof. It suffices to show that, when given a special prime P , we can always find all special primes Q that minimally contains P (since we can always start with the special prime (0)). To this end, assume that P is a special prime. Let π : Rα → Rα−1 be the projection onto the first α − 1 coordinates, let U0 be the submatrix of U consisting of its first α − 1 rows and columns. Define the Frobenius map Θ0 : E α−1 → E α−1 given by Θ0 = U0t T . Let Q be a special prime minimally containing P , and let W = QRα ⋆U so that U W ⊆ W [p] and (0 :R Rα /W ) = Q. Our proof consists of a number of steps. (a) E (P Rα ) being an R[Θ; f ]-module is equivalent to P Im U ⊆ P[p] Rα , and this implies that all entries in U are in (P [p] : P ). This shows that E P Rα−1 is an R[Θ0 ; f ]-module and that P is Θ0 -special.  (b) Consider the case when the action of Θ0 on E P Rα−1 is nilpotent. Pick e ≥ 1 so  e−1 e−2 that the restriction of Θe0 to E P Rα−1 is zero. Consider the matrix U [p ] U [p ] · · · U : denote its last row (g1 , . . . , gα−1 , 0) and note that its top left (α − 1) × (α − 1) submatrix [pe−1 ] is U0 [pe ] [pe−2 ] U0 [pe ] · · · U0 and our assumption implies that the entries of this matrix are in e−1 t e−2 t P ⊂Q so the action of Θe = U [p ] U [p ] · · · U t on E (W ) is the same as the action of a matrix Ue whose first α − 1 rows are zero and its last row is (g1 , . . . , gα−1 , 0). Define L as the union of L0 = QRα L1 L2 = = Ie (Ue QRα ) + QRα Ie (Ue Ie (Ue QRα ) + Ue QRα ) + Ie (Ue QRα ) + QRα = Ie (Ue QRα ) + QRα and the stable value at L1 defines an R[Ue T e ; f e ]-module E (L1 ) whose annihilator is Q. e Now Ue Q ⊆ Q[p] so gi Q ⊆ Q[p ] for all 1 ≤ i ≤ α − 1, hence Q is gi T e -special for all 1 ≤ i ≤ α − 1. One of these gi T e must restrict to a non-zero map on E (P ) otherwise e gi ∈ P [p ] for all 1 ≤ i ≤ α and then the restriction of Θe to E (P Rα ) would be zero. We can now find all such Q using the algorithm in section 5 of [KS]. This finishes our step (b). Let τ ⊂ R be the intersection of the finite set of Θ0 -special prime ideals minimally containing P and write the submodule Nil(E α−1 ) := {z ∈ E α−1 | Θe0 z = 0 for some e ≥ 0} as E (K) where K ⊆ Rα−1 . Write J = (0 :R Rα−1 /π(W )). Note that J ⊇ Q and hence J ) P. (c) Let K0 = Rα−1 and define recursively Kj+1 = I1 (U0 Kj ) for all j ≥ 0. Then clearly K ⊆ K0 . Lemma 3.7 implies that I1 (U K) = K, so if we assume inductively that K ⊆ Kj , then K = I1 (U K) ⊆ I1 (U Kj ) = Kj+1 . α−1 ⋆U0 and hence that τ K ⊆ π(W ); and we reason as follows. (d) We claim that τ K ⊆ (JR √ ) e Lemma 3.8 shows that τ ⊆ J, and so for all large e ≥ 0 we have τ [p ] ⊆ J and hence also  e τ [p ] K ⋆U0 ⊆ JRα−1 ⋆U0 12 ⊆ π(W )⋆U0 = π(W ). e We compute τ [p ] K L0 = L1 = L2 = .. . Le = .. . e ⋆U0 as the union of τ [p ] K   e e−1 e−1 I1 U0 τ [p ] K + L0 = τ [p ] I1 (U0 K) + L0 = τ [p ] K1 + L0   e−2 e−2 e−1 I1 U0 τ [p ] K1 + L1 = τ [p ] I1 (U0 K1 ) + L1 = τ [p ] K2 + L1 τ Ke + Le−1 ⋆U0 [pe ] and from (c) we deduce that τ K ⊆ τ K K . e ⊆ Le ⊆ τ  (e) If the action of Θ0 on E P Rα−1 is not nilpotent, then we claim that τ K 6⊆ P Rα−1  and U0 τ K 6⊆ P [p] Rα−1 . The action of Θ0 on E P Rα−1 being not nilpotent is equivalent α−1 to K 6⊆ P Rα−1 . Since τ 6⊆ P we obtain τ K 6⊆ P Rα−1 . If U0 τ K ⊆ P [p] , then R [p] α−1 α−1 α−1 U0 K ⊆ P R , K = I1 (U0 K) ⊆ P R , and the action of Θ0 on E P R is nilpotent. This completes our step (e). For any v = (w1 , . . . , wα−1 , wα )t ∈ Rα , we define w = (w1 , . . . , wα−1 , 0)t and for any V ⊆ Rα let V denote {v | v ∈ V }. Let ι : Rα−1 → Rα−1 ⊕ R be the natural inclusion ι(v) = v ⊕ 0. Note that V = ι(π(V )). ⋆U ⋆U (f ) We claim I1 (U ι(τ K)) ⊆ W and I1 (U ι(τ K)) 6⊆ P Rα . Define W1 = {w ∈ W | π(w) ∈ τ K} and note that (d) implies that π(W1 ) = τ K. We have W1 ⊆ W hence W1⋆U ⊆ W ⋆U = W ; also W1⋆U = I1 (U W1 )⋆U + W1 and U W1 = U W1 = U ι(τ K) hence ⋆U I1 ((U ι(τ K)) ⊆ W1⋆U ⊆ W . If I1 (U ι(τ K))⋆U ⊆ P Rα then U0 τ K = π (U ι(τ K)) ⊆ P [p] Rα−1 , in contradiction to (e). ⋆U (g) Let M ′ be a matrix whose columns generate I1 ((U ι(τ K)) ⊆ W and choose an entry a in it which is not in P . ⋆U If a ∈ Q, Lemma 5.2 shows that Q is among the minimal primes of ((P + Ra)Rα ) , and we are done (with finding such Q). If a ∈ / Q, we can apply Lemma 5.3 with the matrix X with entries in Ra such that eα ∈ W ′ = XWa ∩ R where eα is the vector (0, 0, . .. , 0, 1)t ∈ Rα . Now Rα /W ′ ∼ = Rα−1 /π(W ′ ), α−1 ′ ′t Q is an associated prime of 0 :R R /π(W ) , and Q is U T -special, with U ′ as defined in Lemma 5.3. t Now Q is special for the Frobenius map obtained by the restriction of U ′ T to E (W ′ ); part (b) and Lemma 5.3(c) shows that this map is not nilpotent and we can apply the induction hypothesis to find Q. This finishes our last step (g) and hence the proof of our proposition. Our next theorem provides an effective algorithm to find all special primes with the property that for some R[Θ; f ]-submodule M ⊆ E α with (0 :R M ) = P , the restriction of Θ to M is not zero. 13 Theorem 5.5. Let P be a special prime, let Q be a special prime minimally containing P . ⋆ Let M be a matrix whose columns generate (P Rα ) . (I) Assume that Im M ) P Rα . Then either (a) All entries of M are in Q, hence there is such an entry q ∈ Q \ P , and Q is among the minimal primes of (0 :R Rα /((P + qR)Rα )⋆ ), or (b) There exists an entry of M which is not in Q, and Q is a special prime of a Frobenius action on E α−1 . (II) Assume that Im M = P Rα . There exist an a1 ∈ R \ P , a g ∈ (P [p] : P ) and an α × α matrix V such that for some µ > 0, aµ1 U ≡ gV modulo P [p] . Write d = det V . Either (a) d ∈ P , and Q can be obtained as a special prime of a Frobenius action on E α−1 , (b) d ∈ Q \ P , and Q is among the minimal primes of (0 :R Rα /((P + dR)Rα )⋆ ), or (c) d ∈ / Q and Q can be obtained as a special prime of the Frobenius action on gT on E. [p] Proof. Choose WQ ⊆ Rα such that U WQ ⊆ WQ and such that (0 :R Rα /WQ ) = Q. Assume first that we are in case (I). If (a) we can choose an entry q of M such that q ∈ Q\ ⋆ P . Now Lemma 5.2 shows that Q is among the minimal primes of 0 :R Rα / ((P + qR)Rα ) . Assume now that we are in case (I)(b), i.e., assume the existence of an entry of M not in ⋆ ⋆ Q. Note that with WQ as above we must have WQ ⊇ (QRα ) ⊇ (P Rα ) = Im M and if we choose a matrix MQ whose columns generate WQ , we see that MQ contains an entry not in Q. We now apply Lemma 5.3 with W replaced by WQ : there exists an invertible α×α matrix X with entries in Ra such that XWQ contains the elementary vector eα := (0, . . . , 0, 1)t , and with U ′ and W ′ as in the lemma, we obtain Q as a special prime of the Frobenius t action (U ′ ) on E (W ′ ). We note that Rα /W ′ ∼ = Rα−1 /W ′′ where W ′′ is the projection ′ of W onto its first α − 1 coordinates, hence Q is a special prime of the Frobenius action (U ′ )t on E (W ′′ ). We may now apply the induction hypothesis and deduce that we have an effective method of finding Q. Assume henceforth case (II) and note that U P Rα ⊆ P [p] Rα implies that the entries of U are in (P [p] : P ). Write S = R/P ; recall that (P [p] : P )/P [p] ∼ = HomS (F∗ S, S) is an S-module of rank one, so we can find an element a1 ∈ R \ P such that the localization of [p] (P [p] : P )/P [p] at a1 is generated by one element g/1 + Pa1 as an Sa1 -module (and hence / Q. also as an Ra1 -module). If a1 ∈ Q, we may construct Q as in case (I)(a), so assume a1 ∈ We can now write aµ1 U = gV + V ′ for some µ ≥ 0 and α × α matrices V and V ′ t with entries in R and P [p] , respectively. Now the restriction of the Frobenius map V ′ T to E (P Rα ) is zero, and hence so is its restriction to E (WQ ), and we may, and do, replace V ′ with the zero matrix without affecting any issues. Write d = det V and distinguish between three cases: (1) Assume d ∈ P . Working in the fraction field F of S we can find an invertible matrix X with entries in F such that the last column of V X −1 is zero. We can now find an element a2 ∈ R \ P such that the entries of X and X −1 are in the localization Ra2 ; write a = a1 a2 . We now apply Lemma 5.3 to deduce, using the Lemma’s notation, that Q is a special t prime of the Frobenius map U ′ T where U ′ = aν X [p] U X −1 , Q = (0 :R Rα /W ′ ), and (U ′ )t T : E (W ′ ) → E (W ′ ) is not nilpotent. We note that the last column of U ′ is zero and hence we can produce Q using Proposition 6.2. 14 (2) If d ∈ Q \ P , we may construct Q as in case (I)(a). (3) Finally we may assume that d ∈ / Q. We now apply Lemma 5.3 with a = a1 d, W = α ⋆U (QR ) and X = Iα , the the α × α identity matrix. With the notation of that ⋆U ⋆U Lemma, we have W ′ = (QRα )a ∩ Rα . We now explicitly compute (QRα )a as the union of the sequence L0 = QRaα L1 L2 = = .. . I1 (U QRaα ) + QRaα = I1 (gV QRaα ) + QRaα = I1 (gQRα )a + QRaα I1 (gL1 ) + L1 ⋆gIα and we compare this to (QRα )a explicitly computed as the union of the sequence L′0 L′1 = = QRaα I1 (gQRaα ) + QRaα = I1 (gQRα )a + QRaα L′2 = .. . I1 (gL′1 ) + L′1 where we used Lemma 3.5 in the third equalities for L1 and L′1 . The fact that L1 = L′1 ⋆U ⋆gI implies that Li = L′i for all i ≥ 1 and hence (QRα )a = (QRα )a α . Now W ′ = ⋆U ⋆gI (QRα )a ∩ Rα = (QRα )a α ∩ Rα , Lemma 5.3 implies that Q is a minimal prime of α ′ R /W and hence Q is (gIα )T -special. We can now deduce that Q is gT -special, and we find Q using the case α = 1, provided that P is also gT -special and gT : E (P ) → E (P ) is non zero. The former follows from the fact that g ∈ (P [p] : P ) and the latter from the fact that the image of g + P [p] after localization at the fraction field of S generates a non-zero module, hence g ∈ / P [p] . Theorem 5.5 also finishes the induction step of the proof of Theorem 5.1. For the sake of completeness, we end with a proof of Theorem 5.1. Proof of Theorem 5.1. We will use induction on α. When α = 1, our theorem has been proved in Section 4. Assume that α > 1 and our theorem has been established for α − 1. Let P be a special prime (we can always start with P = (0)) and let M be a matrix ⋆ whose columns generate (P Rα ) , then there will be two cases: (I). Im M ) P Rα ; (II). Im M = P Rα . As proved in Theorem 5.5, in either cases there are only finitely many special primes Q minimally containing P (by our induction hypothesis). Since only finitely many special primes are produced at each step and the number of steps is bounded by the dimension of R, there are only finitely special primes with the desired property. 6. The algorithm in action and two calculations We first piece together all the results of the previous sections into an explicit algorithm. 15 Input • A ring R = K[x1 , . . . , xn ] where K is a field of prime characteristic p, and • A α × α matrix U with entries in R such that U t T is not a nilpotent Frobenius map on E α . Output The list A of all U t T -special primes Q with the property that the Frobenius map on E (QRα ) is not nilpotent. Initialize A = {0} , B = ∅. Execute the following If α = 1 use the algorithm described in [KS] to find the desired special primes, put these in A, output it, and stop. While A 6= B, pick any P ∈ A \ B, write (P Rα )⋆U as the image of a matrix M and do the following: (1) If there is an entry a of M which is not in P then (1a) add to A the minimal primes of the annihilator of Rα /((P + aR)Rα )⋆U , and (1b) find an α × α invertible matrix X with entries in Ra such that Im M Ra contains the αth elementary vector, choose ν ≫ 0 such that U ′ = X [p] U X −1 has entries in R, let U0 be the submatrix of U consisting of its first α − 1 rows and columns, find recursively the special primes of U0T T , add to A those which are also U T T -special. (2) If (P Rα )⋆U = (P Rα ) find a1 ∈ R \ P , g ∈ (P [p] : P ), α × α matrix V and µ > 0 such that aµ1 ≡ gV modulo P [p] . Compute d = det V . (2a) If d ∈ P , find an element a2 ∈ R \ P and an invertible matrix with entries in Ra2 such that the last column of U X −1 is zero. Find ν ≫ 0 such that the entries of U1 (a1 a2 )ν X [p] U X −1 are in R. Let U0 be the submatrix of U1 consisting of its first α − 1 rows and columns.  (2a)(i) If the restriction of (U0t T )e to E P Rα−1 is zero for some e ≥ 0, write the [pe ] [pe−1 ] last row of U1 U1 . . . U1 as (g1 , g2 , . . . , gα−1 , 0), find all gi T -special primes as in [KS] and add to A those which are also U t T -special.  (2a)(ii) If the restriction of U0t T to E P Rα−1 is not nilpotent, compute recursively all U0t T -special primes minimally containing P and their intersection τ , find K ⊆ Rα−1 such that E (K) is the module of U0t T -nilpotent ⋆U elements, compute I1 (U1 ι(τ K)) 1 and write this as the image of a matrix M ′ . Find a entry a in M ′ not in P . Now 16 • Add to A all the minimal primes of Rα /(P + aR)⋆U1 which are also U t T -special. • Find an invertible matrix X with entries in Ra such that Im M ′ Ra contains the αth elementary vector. Find ν ≫ 0 such that U2 = aν X [p] U1 X −1 has entries in R. Let U3 be the submatrix of U2 consisting of its first α − 1 rows and columns. Compute recursively all U3t T -special primes and add those which are also U t T -special to A. (2b) If d ∈ / P , add to A the minimal primes of the annihilator of Rα ((P + dR)Rα )⋆U . (2c) If d ∈ / P , use the algorithm described in [KS] to find the gT -special primes and add to A those which are also U t T -special. (3) Add P to B. Output A and stop. We now apply the algorithm to the calculation of special primes in two examples. The first, illustrates the trick of reducing α = 2 to a calculation with α = 1, and the second illustrates a case where Theorem 5.5(II)(a), and hence, Proposition 6.2, needs to be applied. 6.1. First example Let R = Z/2Z[x, y, z] and let U= x3 + y 3 + z 3 x(y 2 + z 2 ) xy 2 z 5 x3 ! We start with the special prime P0 = 0: to find the special primes minimally containing P0 , we apply Theorem 5.5 and find ourselves in case (II) with g = 1, U = V , and d = det U = x2 (x(y 3 + z 2 ) + x4 + y 2 z 5 (y 2 + z 2 )). We look for special primes containing d as in Theorem 5.5(II)(b): !  y z 0 x 2 ⋆U = Im . dR 0 0 x y+z ⋆U The annihilator of R2 / dR2 has a unique minimal prime P1 = (x, y + z)R, hence P1 is a special prime. We look for special primes not containing d as in Theorem 5.5(II)(c): these would contain g = 1, hence there aren’t any. Next we find special primes Q containing P1 . We compute !  x y z 0 0 2 ⋆U = Im ) P1 R2 P1 R 0 0 0 x y+z 17 and we are in case (I) of Theorem 5.5. We first consider the cases y ∈ Q and z ∈ Q which give: ⋆U = (x, y, z)R2 (x, y, z)R2 and we obtain the special prime P2 = (x, y, z)R. Assume now that y ∈ / Q, let ! 1/y 0 X= 0 1 and compute [2] ′ U = yX U X −1 ! x3 + y 3 + z 3 xyz 5 xy 2 (y 2 + z 2 ) x3 y = . We now deduce that Q must be a x3 y-special prime; these are computed with the algorithm in [KS] to be xR, yR and (x, y)R. We compute annR R/(xR2 )⋆U = P1 , annR R/(yR2 )⋆U = P2 , and annR R/((x, y)R2 )⋆U = P2 so xR, yR and (x, y)R are not U -special. We conclude that the set of special primes is {P0 , P1 , P2 }. 6.2. Second example Let R = Z/2Z[x, y, z], f = x3 + y 3 + z 3 , g = x2 + z 4 and define ! xf yf U= . xg yg We start with the special prime P0 = 0, and find the special primes Q minimally containing P0 by following Theorem 5.5(II)(a) as follows: with ! 1 −y/x −1 X = 0 1 we compute ′ [2] U = xX U X −1 = x2 f + y 2 g x2 g t 0 0 ! . The U t -special primes either contain x or are U ′ -special primes. To find the former we find ⋆U that P1 = (x, z)R is the only minimal prime of the annihilator of R2 /(xR2 ) and we add it to the list of U t -special primes. t We now find the U ′ -special primes using Proposition as follows. Write U0 = (x2 f + y 2 g) and compute K = R⋆U0 = (x, y)R ∩ (x, z 2 )R ∩ (x3 , y, z)R and since this is not zero, U0 T is t not nilpotent on E (P0 ) and we can proceed to find the U ′ -special primes using Proposition 6.2(g). We compute the set of U0 T -special primes to be {τ R, (x, y)R, (x, z)R} where τ = (y 2 z 4 + x2 (x3 y 3 + z 3 + y 2 ))R; the intersection of these is τ R. We now compute ! y z x 0 ⋆U0 I1 (U ι(τ K)) = Im ; 0 0 y+z x 18 We now look for special primes which contain x, and were considered above, and those Q which do not contain x. For the latter we apply Lemma 5.3 with X being the identity, and deduce that Q is U0 T -special which we found above. To test whether any of τ R, (x, y)R, and (x, z)R is U0t T -special we compute (τ R2 )⋆U0 = (x, y, z)R2 , ((x, y)R2 )⋆U0 = (x, y, z)R2 , and ((x, z)R2 )⋆U0 = (x, y, z)R2 so none of this is U0t T -special. We conclude that the only U0t T -special is 0. We iterate the algorithm again, this time with the special U t T -prime P = P1 = (x, z)R. We compute ! x y z 0 0 2 ⋆U (P1 R ) = Im ) P1 R2 0 0 0 x z and with Theorem 5.5(I) we look for special U t T -primes minimally containing P1 among those containing y and those not containing y. The former must contain the minimal primes of the annihilator or ⋆U R2 /(yR + P1 ) = R2 /(x, y, z)R2 hence the only such special prime is P2 = (x, y, z)R. For those special primes which do not contain y, and application of Lemma 5.3 (with X as the identity matrix) shows that that those special primes are yg special. These special primes can be computed to be yR, (y, x + z 2 )R and (x + z 2 ). Only the last excludes y and none contain P1 so none yield new U t T -special primes. We conclude that the only U t T -special primes are P0 = 0, P1 = (x, z) and P2 = (x, y, z). 7. Connections with Frobenius near splittings Recall that a Frobenius near-splitting of R is an element φ of HomR (F∗ R, R), and, if, additionally, φ(1) = 1, we call φ a Frobenius splitting. The results in [KS] used in this paper were mainly in terms of ideals I ⊆ R stable under a given Frobenius near-splitting, that is, ideals I ⊆ R for which φ(F∗ I) ⊆ I, and later in section 6.2 there a connection was established with submodules stable under a given Frobenius map on E. Here we go the other way around and, having established a method for finding the radical annihilators of submodules of E α , we show how this method finds the annihilators of certain modules stable under the following generalization of Frobenius near-splittings. Definition 7.1. A Frobenius near-splitting of Rα is an element φ of HomR (F∗ Rα , Rα ). Given such a Frobenius near-splitting φ, we call a submodule V ⊆ Rα φ-compatible if φ(F∗ V ) ⊆ V . 19 Thus if V ⊆ Rα is φ-compatible we have a commutative diagram φ F∗ Rα α  / Rα  / Rα /V φ F∗ R /F∗ V and if we take Matlis duals we obtain Hom(F∗ Rα , E) o ∨ O φ ? Hom(F∗ Rα /F∗ V, E)∨o φ EO α (1) ? E (V ) The following establishes in the F -finite case the connection between the annihilators of submodules of E α fixed under a Frobenius map and the annihilators of quotients of Rα by submodules compatible under a Frobenius near-splitting. Proposition 7.2. Assume that R is F -finite. Let φ ∈ HomR (F∗ Rα , Rα ) and let V ⊆ Rα be a φ-compatible submodule. Then φ∨ , the Matlis dual of φ, is a Frobenius map on E α with the property that φ∨ (E (V )) ⊆ E (V ) and the annihilator of Rα /V coincides with that of E (V ). Hence the method of Theorem 5.1 finds all radical annihilators of quotients Rα /V for φ-compatible submodules V for which φ(F∗ Rα ) 6⊆ V . Proof. Since R is F -finite, F∗ R is a free R-module. In this case one has a natural isomorphism HomR (F∗ R, E) ∼ = F∗ E and diagram (1) is identified with α o F∗ E O φ∨ ? annF∗ E α F∗ V t o φ∨ EO α (2) ? annE α V t We now recall that Rα /V and its Matlis dual have the same annihilator and to establish the last claim we note that the restriction of φ∨ to annE α V t is zero precisely when the map F∗ Rα /F∗ V → Rα /V vanishes, i.e., when φ(F∗ Rα ) ⊆ V . The correspondence between Frobenius near-splittings and Frobenius maps in the nonF -finite case may be far more complicated (for example, cf. section 4 in [K2].) We shall assume for the rest of this section that R is F -finite and we will exhibit a more explicit connection between Frobenius map and near-splittings as follows. We can choose a free basis for F∗ R = K[[x1 , . . . , xd ]] over R which contains F∗ xp−1 · · · xp−1 and we 1 d p−1 p−1 let π ∈ HomR (F∗ R, R) denote the projection onto the free summand F∗ x1 · · · xd . 20 Proposition 7.3. Let α be a positive integer and write Φ ∈ HomR (F∗ Rα , Rα ) for the direct sum of α copies of π. (a) Any φ ∈ HomR (F∗ Rα , Rα ) has the form φ = Φ ◦ F∗ U where U is an α × α matrix with entries if R. (b) For any φ = Φ ◦ F∗ U as in (a) and any F∗ R-submodule F∗ V ⊆ F∗ Rα , φ(F∗ V ) = I1 (U V ). Proof. Since HomR (F∗ Rα , Rα ) = HomR (F∗ R, R)α×α we need to show (a) holds for α = 1. Let B a K-basis for F∗ K which contains 1 ∈ K; we obtain a free basis {F∗ bxγ11 · · · xγdd | b ∈ B, 0 ≤ γ1 , . . . , γd < p} for the free R-module F∗ R. As an R-module, HomR (F∗ R, R) is generated by the projections πb,γ1 ,...,γd onto the free summands F∗ bxγ11 · · · xγdd ; moreover, 1 d πb,γ1 ,...,γd = π ◦ F∗ b−1 xp−γ · · · xp−γ hence every element in HomR (F∗ R, R) is given by 1 d π ◦ F∗ u for some u ∈ R. It is enough to establish (b) when V is generated by one element v ∈ F∗ Rα . The equality Φ(F∗ RU v) = I1 (U Rv) in (b) now follows from the fact that both sides are obtained as the outcome of the same calculation. Given an n × n matrix U with entries in R, Proposition 7.3 exhibits (in the F-finite case) a correspondence between submodules E(V ) of E n fixed under the Frobenius map U t T and submodules V of Rn for which the near-splitting Φ ◦ F∗ U composed with the quotient map Rn → Rn /V does not vanish. In addition, Rn /V has the same annihilator as its Matlis dual E(V ). Thus, if one is given a the near-splitting Φ ◦ F∗ U , finding all prime annihilators of submodules V of Rn for which the near-splitting Φ ◦ F∗ U composed with the quotient map Rn → Rn /V does not vanish, is equivalent to finding all annihilators of submodules E(V ) of E n fixed under the Frobenius map U t T , and this we can do by applying the algorithm described in section 6, provided that the resulting Frobenius map is not nilpotent. We conclude by re-interpreting the examples from section 6 in the context of Frobenius near-splittings. Example 7.4. Let R = Z/2Z[x, y, z], let  3 x + y3 + z 3 U= x(y 2 + z 2 ) xy 2 z 5 x3  and consider the Frobenius near-splitting φ = (π ⊕ π) ◦ U . In the first example of section 6 we found special primes P0 = 0, P1 = (x, y + z)R and P2 = (x, y, z)R. These now give us the following three submodules V1 = 0,    x y z 0 0 2 ⋆U = Im V2 = P1 R 0 0 0 x y+z ⋆U = (x, y, z)R2 of F∗ R2 which are compatible with φ and such that and V3 = (x, y, z)R2 2 R /Vi has annihilator Pi for i = 0, 1, 2. 21 Example 7.5. Let R = Z/2Z[x, y, z], f = x3 + y 3 + z 3 , g = x2 + z 4 , define   xf yf U= xg yg and consider the Frobenius near-splitting φ = (π ⊕ π) ◦ U . 6 we found special primes P0 = 0, P1 = (x, z) and P2 = following three submodules V1 = 0,  ⋆U x y z = Im V2 = P1 R2 0 0 0 In the second example of section (x, y, z). These now give us the 0 x 0 z  ⋆U and V3 = (x, y, z)R2 = (x, y, z)R2 of F∗ R2 which are compatible with φ and such that 2 R /Vi has annihilator Pi for i = 0, 1, 2. References [BMS] Manuel Blickle, Mircea Mustaţă and Karen E. Smith. Discreteness and rationality of F-thresholds. Michigan Mathematical Journal 57 (2008), 4361. [BS] M. P. Brodmann and R. Y. Sharp. Local cohomology: an algebraic introduction with geometric applications. Cambridge Studies in Advanced Mathematics, 60, Cambridge University Press, Cambridge, 1998. [EH] F. Enescu and M. Hochster. The Frobenius structure of local cohomology. Algebra Number Theory 2 (2008), no. 7, 721–754. [F] R. Fedder. F -purity and rational singularity. Transactions of the AMS, 278 (1983), no. 2, 461–480. [K1] M. Katzman. Parameter test ideals of Cohen Macaulay rings. Compositio Mathematica, 144 (2008), 933–948. [K2] M. Katzman. Some properties and applications of F -finite F -modules. Journal of Commutative Algebra, 3 (2011), 225–241. [K3] M. Katzman. Frobenius maps on injective hulls and their applications to tight closure. Journal of London Mathematical Society, 81 (2010), 589-607. [KS] M. Katzman and K. Schwede. An algorithm for computing compatibly Frobenius split subvarieties. Journal of Symbolic Computation 47 (2012), 996–1008. [L] G. Lyubeznik. F -modules: applications to local cohomology and D-modules in characteristic p > 0. J. Reine Angew. Math. 491 (1997), pp. 65–130. [M] S. Murru. PhD Thesis, University of Sheffield. 22 [S] R. Y. Sharp. Graded annihilators of modules over the Frobenius skew polynomial ring, and tight closure. Transactions of the AMS 359 (2007), no. 9, 4237–4258. 23
0math.AC
arXiv:1102.4528v3 [cs.CE] 17 Mar 2011 March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics MODELLING THE DYNAMICS OF THE WORK-EMPLOYMENT SYSTEM BY PREDATOR-PREY INTERACTIONS N. S. COSTA SERPA UNIP - Universidade Paulista, Instituto de Ciências Exatas e Tecnologia SGAS Quadra 913, s/no - Conjunto B - Asa Sul - Brası́lia - DF, Brasil, CEP 70390-130 [email protected], [email protected] J. R. STEINER UnB - Universidade de Brası́lia, Instituto de Fı́sica CP 04455, Brası́lia - DF, Brasil, CEP 70919-970 [email protected] The broad application range of the predator-prey modelling enabled us to apply it to represent the dynamics of the work-employment system. For the adopted period, we conclude that this dynamics is chaotic in the beginning of the time series and tends to less perturbed states, as time goes by, due to public policies and hidden intrinsic system features. Basic Lotka-Volterra approach was revised and adapted to the reality of the study. The final aim is to provide managers with generalized theoretical elements that allow to a more accurate understanding of the behavior of the work-employment system. Keywords: predator-prey model; chaotic dynamics; work-employment system; perturbed states; evolution. ”Economies possess general ecosystem properties, such as dynamism, evolution, integrity, stability and resilience. Economies are inextricably embedded in larger natural ecosystems, and exchange flows of materials and energy with natural systems”. S. Farber & D. Bradley 1. Introduction The concept of equilibrium in the predator-prey population dynamics has its origins in the works of Lotka 1 and Volterra 2 , making the base of several theoretical models of interaction among species, with applications in the context of ecological systems and wildlife management (Caughley and Sinclair, 1994 13 ; Earn et al., 2000 14 ; Blasius et al., 1999 3 ), including fluctuating environments (Collie and Spencer, 1997 15 ). Particularly, Blasius et al. were very benefited by previous works on phase 1 March 21, 2011 1:22 WSPC/INSTRUCTION FILE 2 Econophysics Nilo Sylvio Costa Serpa synchronization phenomena in coupled chaotic systems (Rosenblum et al., 1996 7 ). Also, there are generalized approaches of the Lotka-Volterra model (Tu and Wilman, 1992 16 ). There is special interest on the problem of the growth of two populations conflicting with one another, known as ”the problem of Volterra” 8 . The generality of type predator-prey models becomes possible to abstract them from their early ecological roots, and, by analogy, to apply them on a large range of mathematical modelling problems. Recently it befell to us to investigate the work-employment dynamics in Brazil by the implementation of a type predator-prey model. It is not the case of a classical application of the standard Lotka-Volterra competition modelling, since the variables storage dissimilarities among population registers, not the populations themselves; this is so because the existing cyclic pattern which entangles the work-employment evolution is much more easily discerned by dissimilar modelling. Besides, while is tempting to discard the standard Lotka-Volterra model as too simplistic, there is a real situation where the detailed and complex dissimilar modelling holds undeniable utility. The aim of this article is to show that the workemployment system in Brazil admits a predator-prey modelling. The model started from a simple correlation suggested by the superposition of two time series, one to the number of employed workers, and other to the number of active employers. The entanglement of the two series is analyzed with the aid of the clustering methodology. In the present article, the abstract analogy with a natural ecosystem is the work market; the employer or corporation plays the rule of the predator; the prey is represented by the worker. The assumption of employers as predators has many reasons: • The number of employers is much smaller than the number of workers (as in linx-hare, lion-zebra, crocodile-gnu and cheetah-baboon predator-prey relationships). • The employers handle the number of workers depending on the profit margins which they want, heavily controlling the levels of unemployment. • Employers often fail to collect the so-called ”Guarantee Fund for Time in Service” related to their workers, causing irregularities that hinder the realization of labor rights in the event of contract termination agreement. Moreover, even without a contract being rescinded, the nonpayment of the Guarantee Fund blocks the employee to request it to buy a home. • Employers often fail to inform the CAGED (Cadastro Geral de Empregados e Desempregados)a on the movement of their staff, resulting in bureaucratic disarray that blocks the access to some labor rights in the event of loss of working papers; furthermore, the fail to inform the CAGED affects the labor statistics. • Many employers absorb the workforce without formal contracts, depriving workers of their entitlements. a CAGED is a great database updated monthly and containing all workers and employers in Brazil. March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 3 • However the scenario is changing in Brazil, many employers still do not make investments in quality, training employees and providing opportunities for better wages by productivity and capacity. The result is a high turnover, low wages and a working class that never achieves a status of participation on the employer’s profits. In sum, we have a highly non-egalitarian society which does not prioritize education and manpower quality. Socially speaking, the predator is who foment social discrepancies and, in present case, they are employers (public or private). That is the point here. From this point of view, we have not a model based on political preferences, as some watchers would think, but based on hard facts. There is no influence of maniqueist ideas, but simply findings. Paraphrasing Lawson, the offer of labor posts depends on the investment decisions of capitalists (employers), and the investment decisions of capitalists depend on the existing offer of manpower 10 . Thus, since employers impose financial directions on the work market, often indifferent to the zeal for the rights of the working class, thus creating distortions and fueling widespread dissatisfaction, it is reasonable to consider the adopted viewpoint in constructing the model. Obviously, in a more egalitarian society, firmly established on ethics and education, economically viable and with eyes in the future, it would be required a comprehensive review of the basic premises assumed. Also, in face of the facts itemized above, employers appear in many situations to dampen fluctuations in employee populations. 2. Observational premises The data sources were the CAGED, and the RAIS (Relação Anual de Informações Sociais), two of the more important corporative databases of the Brazilian Ministry of Work and Employment. First of all we looked for harmonic patterns in the behavior of the working class and active employers. Figure 1 clearly presents the referred pattern in the chosen period (1996-2008). We defined active employer or company as one that presents variation in the number of admitted workers from a certain month to another. Secondly, we verify the presence of lags between the harmonic patterns of both series. This feature is also clearly visible in figure 1, with relaxation of the system all along the last fifty months. Table 1 shows a segment of the CAGED for the first year of the time series considered in present study March 21, 2011 1:22 WSPC/INSTRUCTION FILE Nilo Sylvio Costa Serpa workers employers 0 50 time(months) 100 150 3 e+05 2.4 e+07 2.6 e+07 4 e+05 2.8 e+07 3.0 e+07 workers employers 5 e+05 3.2 e+07 3.4 e+07 6 e+05 3.6 e+07 4 Econophysics Figure 1: harmonic patterns from workers (full line) and active employers (dashed line). The phase difference is clear in most of the peaks and valleys, especially in the fifty more recent months. Table 1: aggregate data Period Balance workers 01/1996 -12.626 02/1996 -4.094 03/1996 10.003 04/1996 118.918 05/1996 172.930 06/1996 115.028 07/1996 68.920 08/1996 46.937 09/1996 88.964 10/1996 19.466 11/1996 -15.899 12/1996 -258.516 of the CAGED database. Workers Active employers 23.743.110 336.946 23.739.016 334.349 23.749.019 359.248 23.867.937 349.273 24.040.867 344.896 24.155.895 326.832 24.224.815 338.098 24.271.752 339.688 24.360.716 327.525 24.380.182 329.672 24.364.283 308.009 24.105.767 298.644 March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics Active employers Workers 5 Balance workers Figure 1-a: the relationship of workers, active employers and balance of workers. The original Lotka-Volterra predator-prey model makes several simplifying assumptions; it was adapted in many ways, as we may see in the vast literature on predatorprey interactions from which the authors selected some of the more important publications (see references). In many cases the former assumptions are relaxed or rearranged to fit some particular dynamics. There is no doubt that the predatorprey model adapts, at least conceptually, the work-employment scenario, as it is very reasonable to admit that employers and workers are, after all, two populations conflicting with one another. Since the formulations of Karl Marx - with the analysis of the conflicts between worker and employer, and of the entailments of such conflicts with structural elements formed by the connections among politic, social and economic plains - we understand that conflicts belong to the dialectic nature of the capitalist work relations. Also Richard Goodwin, in a different approach, found in Marx arguments conceptual similarities that lead to his predator-prey dynamic model 11 . March 21, 2011 1:22 WSPC/INSTRUCTION FILE 6 Econophysics Nilo Sylvio Costa Serpa 3. Methods and simulations 3.1. Formal representation The pertinent information such a model can render is whether or not population abundances tend to an equilibrium at which both, workers and employers, will coexist and survive with few conflicts. The adopted model was based on the approach of Blasius and Stone 3 about oscillations with Uniform Phase Evolution and Chaotic Amplitudes (UPCA). An ecological UPCA model was chosen because it is suitable for treatment of two correlated periodic phenomena with chaotic amplitude varying in time, however, almost constant amplitude-independent frequencies as hinted in Figure 1. We applied the model to investigate the complex temporal phase synchronization in work-employment system at Brazil. In other words, our approach was based on the theory of synchronization of chaotic oscillations, defined in the most general case 7 as the entanglement between the phases of two coupled systems, while the amplitudes stay over chaotic regime in time. The real entanglement of phases in predator-prey relationships is much more complex than one could suppose at first. As pointed out by Holling, predator-prey interactions have shown that, in some cases of population densities, predators may control the number of preys, but this is not true to all density cases 4 . However in Figure 1 populations grow and fluctuate in a net periodic way, the tests on the tangle between these populations goes far beyond the simplicity of that plot, as the reader can see in Figure 1-a. The balance of workers is a finite difference that indicates whether the workemployment scenario is favorable or not to the employee, and to what degree; it reflects the potential of the Mean (work market) to sustain the system evolving, by analogy with the available natural resources in an ecological niche. In other words, if the balance is negative we believe that there is a retraction of work market to support occupied posts and, virtually, active employers. It is a variable that provides additional qualitative (positive or negative sign) information. Figure 1-a shows the active employers-workers-balance of workers relationship, illustrating the potential oscillations in these populations. The Lotka-Volterra-Serpa differential equations consider a state vector formed by the product between a geometric tracing vector and a dissimilarity vector of observables. We use pairwise dissimilarities between registers instead of the quantities directly stored in data base just to compute the relative differences which are really cyclic all along the time series. That is the basic difference from the classical Lotka-Volterra approach and the name was given by fellow researchers in honor of the author, Nilo Sylvio Costa Serpa. The dissimilarity matrix was made from the monthly populations of active employers and workers, and from the monthly balance of workers. This matrix computes all the pairwise dissimilarities (or ”distances”) between observations in the CAGED data set by means of the DAISY algorithm 12 , that executes mixed measurements. We assume that the pairwise differences computed from the time registers determine the dynamical configuration of the system. Thereby, our model has three differential equations, one for the dissimilarity of balance of workers u (resources), one for the dissimilarity of March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 4 0 2 workers/employers 6 8 7 0 10 20 30 40 50 60 70 time Figure 2: UPCA oscillations of employer (grey) and worker (black) dissimilarities with initial time-delayed growth (α2 = 1.0). workers v (preys) and one for the dissimilarity of active employers w (predators), such as  u̇ = a.u.uo − α1 .f1 (u.uo , v.vo , k1 )    v̇ = − b.v.vo + α1 .f1 (u.uo , v.vo , k1 )− (1)  −α2 .f2 (v.vo , w.wo , k2 )   ẇ= − c.(w.wo − w† ) + α2 .f2 (v.vo , w.wo , k2 ), where f1 and f2 represent either the Lotka-Volterra term fi (x, y) = xy or the Holling type II term fi (x, y) = xy/(1 + ki x), and w† denotes the minimum level of existent employers dissimilarity when there is scarcity of workpower or doldrums. The overdot indicates differentiation with respect to time. The quantities a, b and c are growth rates of balance of workers, workers, and employers respectively. The Holling type II functional response 4 introduces a decelerating intake rate related to the assumption that the dissimilarities of active employers is limited by its capacity March 21, 2011 1:22 WSPC/INSTRUCTION FILE Nilo Sylvio Costa Serpa 3 2 0 1 workers/employers 4 5 8 Econophysics 0 10 20 30 40 50 60 70 time Figure 3: UPCA oscillations of employer (grey) and worker (black) dissimilarities with initial time-delayed growth (α2 = 1.4). to regulate dissimilarities of workers. Comparing to the original equations applied by Blasius,   u̇ = a.u − α1 .f1 (u, v, k1 ) v̇ = − b.v + α1 .f1 (u, v, k1 ) − α2 .f2 (v, w, k2 )  ẇ= − c.(w − w† ) + α2 .f2 (v, w, k2 ), the reader can see that we merge variables u, v, z with u0 , v0 , z0 and this changes the final geometric form of the system integration according to de observed data. Figure 5-b shows an example of UPCA oscillations for the last system. In our model, the positive constant k2 measure the carrying capacity of employers to absorb manpower; the positive constant k1 measure the carrying capacity of the work market to accept new workers. As pointed out by Kenneth Arrow and colleagues 6 , March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 3 2 0 1 workers/employers 4 9 0 10 20 30 40 50 60 70 time Figure 4: UPCA oscillations of employer (grey) and worker (black) dissimilarities with initial time-delayed growth (α2 = 1.8). ”Carrying capacities in nature are not fixed, static, or simple relations. They are contingent on technology, preferences, and the structure of production and consumption. They are also contingent on the everchanging state of interactions between the physical and biotic environment. A single number for human carrying capacity would be meaningless because the consequences of both human innovation and biological evolution are inherently unknowable”. So, it is interesting to investigate what would happen for work-employment scenarios with very low carrying capacity of employers to create new posts. For instance, if we take k2 = 0 we means that the employers are saturated or temporarily disabled to open new posts, although k1 remains greater than zero, that is, the work market remains in flux, but mainly due to high turn over. The contents of the differential March 21, 2011 1:22 WSPC/INSTRUCTION FILE Nilo Sylvio Costa Serpa 3 2 0 1 workers/employers 4 10 Econophysics 0 10 20 30 40 50 60 70 time Figure 5: UPCA oscillations of employer (grey) and worker (black) dissimilarities with initial time-delayed growth (α2 = 2.0). variables were defined by means of the vectors, V (u < −ω[1], v < −ω[2], w < −ω[3]) , U (uo < −Π[δ, 1], vo < −Π[δ, 2], wo < −Π[δ, 3]) , (2) being ω[j] ( j in [1, 2, 3]) the geometric components describing the state trajectories and Π[δ, 2j] the j dissimilarity components between observations with δ registers calculated according to the definition of Euclidian distance, that is, Π[δ, j] ∝ d X κ=1 |xδκ − xjκ | r !1/r , (3) where d is the phase space dimensionality; xδκ and xjκ are, respectively, the κ − th components of the δ − th and j − th registers. March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 3 2 0 1 workers/employers 4 11 0 10 20 30 40 50 60 70 time Figure 5-a: UPCA oscillations of employer (grey) and worker (black) dissimilarities with initial time-delayed growth and k2 = 0.00008 (α2 = 1.8). 3.2. The equilibrium condition No changes with respect to time in populations and resources, or in their dissimilarities, are represented by setting the equations equal to zero. So, the equilibrium condition for the system (1) requires,    a.u.uo − α1 .f1 (u.uo , v.vo , k1 ) = 0  −b.v.vo + α1 .f1 (u.uo , v.vo , k1 )− (4)  −α 2 .f2 (v.vo , w.wo , k2 ) = 0   −c.(w.wo − w† ) + α2 .f2 (v.vo , w.wo , k2 ) = 0 Introducing the term of Lotka-Volterra (the simplest one), the equilibrium - where the dissimilarities of active employers, resources and workers do not vary - leads to,   a.u.uo − α1 .uo .vo .u.v = 0 −b.v.vo + α1 .uo .vo .u.v − α2 .vo .wo .v.w = 0 (5)  −c.(w.wo − w† ) + α2 .vo .wo .v.w = 0 March 21, 2011 1:22 WSPC/INSTRUCTION FILE 12 Econophysics Nilo Sylvio Costa Serpa Figure 5-a: simple UPCA oscillations by Blasius et al. for variables u and v. From the first equation we obtain, whereas population dissimilarity shall be never equal to zero, u.(a.uo − α1 .uo .vo .v) = 0, (6) a.uo − α1 .uo .vo .v = 0, (7) a a.uo = . α1 .uo .vo α1 .vo (8) v= Taking the third equation we get, w= − c.w.wo + c.w† + α2 .vo .wo .v.w = 0, (9) w (−c.wo + α2 .vo .wo .v) = −c.w† , (10) −c.w† c.w† = . − (c.wo − α2 .vo .wo .v) c.wo − α2 .vo .wo .v (11) Finally, taking the second equation it follows, α1 .uo .vo .u.v=α2 .vo .wo .v.w + b.v.vo , u= (12) α2 .vo .wo .v.w + b.v.vo , α1 .uo .vo .v (13) α2 .wo .w + b . α1 .uo (14) u= Substituting the value of w given by equality (11), it comes,   1 c.w† , +b u = α2 .wo . c.wo − α2 .vo .wo .v α1 .uo (15) March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 13 u=   1 α2 .c.w† , +b c − α2 .vo .v α1 .uo (16) u= α2 .c.w† + b. (c − α2 .vo .v) . α1 .uo . (c − α2 .vo .v) (17) Now, applying the result set by expression (8), we have,   α1 .c − α2 .a 1 † u = α2 .c.w + b. , α1 uo . (α1 .c − α2 .a) u= α1 .α2 .c.w† + b. (α1 .c − α2 .a) . α1 .uo . (α1 .c − α2 .a) (18) (19) For the term of Holling type II, that is, f1 = u.u0 .v.v0 / (1 + k1 .u.u0 ) , f2 = v.v0 .w.w0 / (1 + k2 .v.v0 ) , the equilibrium sets,  a.u.uo − α1 .u.u0 .v.v0 / (1 + k1 .u.u0 ) = 0    −b.v.vo + α1 .u.u0 .v.v0 / (1 + k1 .u.u0 ) −  −α 2 .v.v0 .w.w0 / (1 + k2 .v.v0 ) = 0   −c.(w.wo − w† ) + α2 .v.v0 .w.w0 / (1 +k2 .v.v0 ) = 0 (20) (21) The system becomes much more complex. Nevertheless, we fix k2 = 0, meaning that the employers are temporarily disabled to open new posts, which determines the solution as it follows. The second equation leads to, w= α1 .u.u0 / (1 + k1 .u.u0 ) − b . α2 .w0 (22) The first equation gives, a = α1 .v.v0 / (1 + k1 .u.u0 ) , (23) a. (1 + k1 .u.u0 ) = α1 .v.v0 , (24) 1 + k1 .u.u0 = u= α1 .v.v0 , a α1 .v.v0 − a . u0 .a.k1 (25) (26) Finally, the third equation sets v as, α2 .v.v0 .w.w0 / 1 +k2 .v.v0 ) = c.(w.wo − w† ), v= c.(w.wo − w† ) . α2 .v0 .w.w0 (27) (28) March 21, 2011 1:22 WSPC/INSTRUCTION FILE 14 8 12 workers Nilo Sylvio Costa Serpa 10 14 Econophysics 6 4 6 4 2 companies 10 8 2 0 0 0 50 100 150 200 time 12 10 workers 6 14 4 companies 8 10 Figure 6: the system in three-dimensional phase space. 2 8 6 4 2 0 0 0 50 100 150 200 time Figure 7: another vision of the same system. 4. Discussion Insecurity generated by the employers and the work market, over many years of restructuring and adjustment to globalization as the Nineties, led to an exacerbation of individualism. Competition became more dominant than cooperation. Fear and instability certainly promoted higher levels of depression, while millions of peo- March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 6 4 0 2 companies 8 10 15 0 2 4 6 8 10 12 workers Figure 8: workers and employers in phase space with an attracting central curve. ple reached the productive age, pressing more and more the formal work market. Moreover, due to the neo-liberal offensive (1980-1990) to the resumption of the ”Theory of Human Capital”, backed by Gary Becker, the wage differences came to be seen as responsibility of the worker himself, who thus became the chief culprit for all economic ills. The spreading of innovations due to technological changes (the convergence of information technology and telecommunications and the continuing cost reductions) greatly affected the rate of hiring as a function of the resizing of the human effectives face to an accelerated automation. Also, Security in all the countries that care about their retirees do not support the significant increase in life expectancy, discouraging retirements and thus retaining jobs occupied for much longer. All these factors contributed to a more chaotic work-employment scenario in Brazil from mid-1990 until about 2002 and they corroborate the interpretation of the plots. Many simulations are performed with vector V describing the geometry of the dynamics (theoretical), intrinsic to the nature of the system itself, and vector U storaging the measurable content (observational) of the dynamics. For the more realistic Holling type II functional response, with k2 = 0, figures 2, 3, 4 and 5 show graphics of the general propensity of the system to less chaotic states accordingly the values 1.0, 1.4, 1.8 and 2.0 of the population interaction parameter α2 . Figure 5-a shows similar patterns for k2 = 0.00008. Note that in all these figures there is a transition state to the phase of greater relaxation of the system. The great phase transition occurs close to the middle of the time series, from which the system goes on a relaxation stage. Comparing figures 1 and 5, the best fit realy seems to be for α2 = 2.0 with phase transition around the zone between 50 and 100 months in figure 1. Figures 6, 7, and 8 show the system’s phase space for α2 = 1.0. The exact March 21, 2011 1:22 WSPC/INSTRUCTION FILE 16 Econophysics Nilo Sylvio Costa Serpa fitting of this parameter is a matter of more studies, since predator-prey stability could come from many factors at play, such as migration of manpower, employers self-limiting their staffs, seasonal recruitment due to foreseen peaks of sales, and so on, in the same way as pointed out by Andrzej Pekalski 5 in biological systems. 5. Conclusion This study identified a predator-prey pattern in the work-employment system. The central equations applied are modified versions of those studied by Blasius et al, introducing pairwise dissimilarities between observations as a way to identify cyclic behaviors in the system. For carrying capacity k2 = 0, simulations showed trajectories with initial chaotic behavior and tendency to final states of greater equilibrium, which means that, even with zero carrying capacity of employers, the dynamics of the work-employment system in Brazil has demonstrated, during the past 12-15 years, propensity to adjusting to less chaotic regimes. In part, this is due to the great migration in recent years to solidary economy and autonomous work. Present article analyzed the system dynamics under the interpretation of the population interaction parameter α2 as an index of the convergence between the aims of employers and workers or as an index of the equilibrium created by lower demand for formal employment. Of course, interactions and proliferation of workers and employers are not easily controlled, but the captivation of manpower is a variable that may be monitored and partially governed by public policies. Moreover, as Grafton and Echenique 9 pointed out, we have the problem to choose the more appropriate ecological model to describe the situation in study. In our opinion, as more is learned about the system, we must change adaptively the model in order to lower uncertainty. Far from exhausting the subject, it is clear that in its general form the model represents reasonably well the dynamics of predator-prey for the work-employment system in Brazil. We hope this work will serve as inspiration for further studies in order to compare the application of the model in different countries. 6. Acknowledgements The authors acknowledge Dr. Sergio Alves Cotia for their comments on the manuscript and for his integral support to this work. ♦♦♦ References 1. Lotka, A. 1925, Elements of Physical Biology, Baltimore, Williams and Wilkins. 2. Volterra, V. 1931, Leçons sur la Thorie Mathmatiques de la Lutte pour la Vie, Paris. March 21, 2011 1:22 WSPC/INSTRUCTION FILE Econophysics 17 3. Blasius, B. and Stone, L. 1999, Chaotic Waves and Phase Synchronization in spatially extended ecological systems in D. S. Broomhead, E. A. Luchinskaya, P. V. E. McClintock and T. Mullin, ed., Stochaos: Stochastic and Chaotic Dynamics in the Lakes, American Institute of Physics, Woodbury, NY, 5ior02:221-225. 4. Holling, C. S. 1959, The components of predation as revealed by a study of smallmammal predation of the European pine sawfly, Canadian Entomologist 91: 293-320. 5. Pekalski, A. 2004, A Short Guide To Predator-Prey Lattice Models, Computer Simulations, Computing in Science & Engineering. 6. Arrow, K. et al. 1995, Economic Growth, Carrying Capacity, and the Environment, Science, vol. 268. 7. Rosenblum, M. et al. 1996, Phys. Rev. Lett., 76, 1804. 8. Davis, H. 1960, Introduction to Nonlinear Differential and Integral Equations, United States Atomic Energy Commission, Washington, DC. 9. Grafton, R. and Silva-Echenique, J. 1997, How to Manage Nature? Strategies, PredatorPrey Models, and Chaos, Marine Resource Economics, vol. 12, 127143. 10. Lawson, T. 2004, Reorienting Economics: On Heterodox Economics, the Mata and the Use of Mathematics in Economics, Journal of Economic Methodology 11: 329340. · · · Lawson, T. 2005, The (Confused) State of Equilibrium Analysis in Modern Economics: an Explanation, Journal of Post Keynesian Economics 27: 423444. · · · Lawson, T. 2006, The Nature of Heterodox Economics, Cambridge Journal of Economics 30: 483505. 11. Goodwin, R. 1967, A Growth Cycle, in Feinstein, C.H. (ed.) Socialism, Capitalism and Economic Growth, Cambridge, Cambridge University Press. 12. Kaufman, L. and Rousseeuw, P. 1990, Finding Groups in Data: An Introduction to Cluster Analysis, Wiley, New York. 13. Caughley, G. and Sinclair, A. 1994, Wildlife Ecology and Management, Blackwell Science Ltd, Melbourne. 14. Earn, D. et al. 2000, A simple model for complex dynamical transitions in epidemics, Science 287, 667670. 15. Spencer, P. and Collie, J. 1997, Patterns of population variability in marine fish stocks, Fisheries Oceanography, 6:188204. 16. Tu, P. and Wilman, E. 1992, A Generalized Predator-Prey Model: Uncertainty and Management, Journal of Environmental Economics and Management 23, 123-138.
5cs.CE
Estimation of Mean Residual Life arXiv:1707.02484v2 [math.ST] 11 Jul 2017 W. J. Hall Department of Biostatistics University of Rochester Rochester, NY e-mail: [email protected] Jon A. Wellner∗ Department of Statistics University of Washington Seattle, WA 98195-4322 e-mail: [email protected] url: http://www.stat.washington.edu/jaw/ Abstract: Yang [1978b] considered an empirical estimate of the mean residual life function on a fixed finite interval. She proved it to be strongly uniformly consistent and (when appropriately standardized) weakly convergent to a Gaussian process. These results are extended to the whole half line, and the variance of the the limiting process is studied. Also, nonparametric simultaneous confidence bands for the mean residual life function are obtained by transforming the limiting process to Brownian motion. AMS 2000 subject classifications: Primary 62P05, 62N05, 62G15;Secondary 62G05, 62E20. Keywords and phrases: life expectancy, consistency, limiting Gaussian process, confidence bandsconsistency, confidence bands, life expectancy, limiting Gaussian process. Contents 1 2 3 4 5 6 7 Introduction and summary . . . . . . . . . . . . . . . . . . . . Convergence on R+ ; covariance function of the limiting process Alternative sufficient conditions; V ar[Z(x)] as x → ∞. . . . . . Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Confidence bands for e. . . . . . . . . . . . . . . . . . . . . . . Illustration of the confidence bands. . . . . . . . . . . . . . . . Further developments . . . . . . . . . . . . . . . . . . . . . . . 7.1 Confidence bands and inference . . . . . . . . . . . . . . . 7.2 Censored data . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Median and quantile residual life functions . . . . . . . . . 7.4 Semiparametric models for mean and median residual life 7.5 Monotone and Ordered mean residual life functions . . . . 7.6 Bivariate residual life . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ∗ Supported . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 3 7 8 9 11 12 12 15 15 15 16 16 16 in part by NSF Grant DMS-1104832 and the Alexander von Humboldt Foundation 1 imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 2 1. Introduction and summary This is an updated version of a Technical Report, Hall and Wellner [1979], that, although never published, has been referenced repeatedly in the literature: e.g., Csörgő and Zitikis [1996], Berger et al. [1988], Csörgő et al. [1986], Hu et al. [2002], Kochar et al. [2000], Qin and Zhao [2007]. Let X1 , . . . , Xn be a random sample from a continuous d.f. F on R+ = [0, ∞) with finite mean µ = E(X), variance σ 2 ≤ ∞, and density f (x) > 0. Let F = 1 − F denote the survival function, let Fn and Fn denote the empirical distribution function and empirical survival function respectively, and let Z ∞ e(x) ≡ eF (x) ≡ E(X − x|X > x) = F dI/F (x), 0≤x<∞ x denote the mean residual life function or life expectancy function at age x. We use a subscript F or F on e interchangeably, and I denotes the identity function and Lebesgue measure on R+ . A natural nonparametric or life table estimate of e is the random function ên defined by  Z ∞ Fn dI/Fn (x) 1[0,Xnn ) (x) ên (x) = x where Xnn ≡ max1≤i≤n Xi ; that is, the average, less x, of the observations exceeding x. Yang [1978b] studied ên on a fixed finite interval 0 ≤ x ≤ T < ∞. She proved that ên is a strongly uniformly consistent estimator of e on [0, T ], and that, when properly centered and normalized, it converges weakly to a certain limiting Gaussian process on [0, T ]. We first extend Yang’s (1978) results to all of R+ by introducing suitable metrics. Her consistency result is extended in Theorem 2.1 by using the techniques of Wellner [1977, 1978]; then her weak convergence result is extended in Theorem 2.2 using Shorack [1972] and Wellner [1978]. It is intuitively clear that the variance of ên (x) is approximately σ 2 (x)/n(x) where σ 2 (x) = V ar[X − x|X > x] is the residual variance and n(x) is the number of observations exceeding x; the formula would be justified if these n(x) observations were a random sample of fixed sized n(x) from the conditional distribution P (·|X > x). Noting that Fn (x) = n(x)/n → F (x) a.s., we would then have nV ar[ên (x)] = nσ 2 (x)/n(x) → σ 2 (x)/F (x). Proposition 2.1 and Theorem 2.2 validate this (see (2.4) below): the variance of the limiting distribution of n1/2 (ên (x) − e(x)) is precisely σ 2 (x)/F (x). In Section 3 simpler sufficient conditions for Theorems 2.1 and 2.2 are given and the growth rate of the variance of the limiting process for large x is considered; these results are related to those of Balkema and de Haan [1974]. Exponential, Weibull, and Pareto examples are considered in Section 4. In Section 5, by transforming (and reversing) the time scale and rescaling the state space, we convert the limit process to standard Brownian motion on the unit interval (Theorem 5.1); this enables construction of nonparametric simultaneous confidence bands for the function eF (Corollary 5.2). Application to survival data of guinea pigs subject to infection with tubercle bacilli as given by Bjerkedal [1960] appears in Section 6. We conclude this section with a brief review of other previous work. Estimation of the function e, and especially the discretized life-table version, has been considered by Chiang; see pages 630-633 of Chiang [1960] and page 214 of Chiang [1968]. (Also see Chiang [1968], page 189, for some early history of the subject.) The basis for marginal inference (i.e. √ at a specific age x) is that the estimate ên (x) is approximately normal with estimated standard error Sk / k, where k = nFn (x) is the observed number of observations beyond x imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 3 and Sk is the sample standard deviation of those observations. A partial justification of this is in Chiang [1960], page 630, (and is made precise in Proposition 5.2 below). Chiang [1968], page 214, gives the analogous marginal result for grouped data in more detail, but again without proofs; note the solumn Sêi in his Table 8, page 213, which is based on a modification and correction of a variance formula due to Wilson [1938]. We know of no earlier work on simultaneous inference (confidence bands) for mean residual life. A plot of (a continuous version of) the estimated mean residual life function of 43 patients suffering from chronic gramulocytic leukemia is given by Bryson and Siddiqui [1969]. Gross and Clark [1975] briefly mention the estimation of e in a life - table setting, but do not discuss the variability of the estimates (or estimates thereof). Tests for exponentiality against decreasing mean residual life alternatives have been considered by Hollander and Proschan [1975]. 2. Convergence on R+ ; covariance function of the limiting process Let {an }n≥1 be a sequence of nonnegative numbers with an → 0 as n → ∞. For any such sequence and a d.f. F as above, set bn = F −1 (1 − an ) → ∞ as n → ∞. Then, for any function f on R+ , define f ∗ equal to f for x ≤ bn and 0 for x > bn : f ∗ (x) = f (x)1[0,bn ] (x). Let kf kba ≡ supa≤x≤b |f (x)| and write kf k if a = 0 and b = ∞. R1 Let H(↓) denote the set of all nonnegative, decreasing functions h on [0, 1] for which 0 (1/h)dI < ∞. Condition 1a. There exists h ∈ H(↓) such that M1 ≡ M1 (h, F ) ≡ sup R∞ Condition 1b. There exists h ∈ H(↓) for which R∞ h(F )dI/h(F (x)) < ∞. e(x) x R∞ Since 0 < h(0) < ∞ and e(0) = E(X) < ∞, Condition 1a implies that 0 h(F )dI < ∞. Also note that h(F )/h(0) is a survival function on R+ and that the numerator in Condition 1a is simply eh(F )/h(0) ; hence Condition 1a may be rephrased as: there exists h ∈ H(↓) such that M1 ≡ keh(F )/h(0) /eF k < ∞. x 0 h(F )dI < ∞ and keh(F )k < ∞. Bounded eF and existence of a moment of order greater than 1 is more than sufficient for Condition 1b (see Section 3). Theorem 2.1. Let an = αn−1 log log n with α > 1. If Condition 1a holds for a particular h ∈ H(↓), then ρh(F )e/F (ê∗n , e∗ )   |ên (x) − e(x)|F (x) : x ≤ bn →a.s. 0 as n → ∞. ≡ sup h(F (x))e(x) (2.1) If Condition 1b holds, then ρ1/F (ê∗n , e∗ ) ≡ sup{|ên (x) − e(x)|F (x) : x ≤ bn } →a.s. 0 as n → ∞. (2.2) The metric in (2.2) turns out to be a natural one (see Section 5); that in (2.1) is typically stronger. imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 4 Proof. First note that for x < Xnn F (x) ên (x) − e(x) = Fn (x) ( − R∞ x (Fn − F )dI F (x) ) e(x) + (Fn (x) − F (x)) . F (x) Hence ρh(F )e/F (ê∗n , e∗ ) ≤ ≤ →a.s. F Fn bn F Fn bn 0 0 ( sup x | R∞ (Fn − F )dI| |Fn (x) − F (x)| + sup h(F (x))e(x) h(F (x)) x x ) · ρh(F ) (Fn , F )(M1 + 1) 0 using Condition 1a, Theorem 1 of Wellner [1977] to show ρh(F ) (Fn , F ) →a.s. 0 a.s., and Theorem 2 of Wellner [1978] to show that lim supn kF /Fn kb0n < ∞ a.s.. Similarly, using Condition 1b,  Z ∞ b  F n sup | (Fn − F )dI| + sup e(x)|Fn (x) − F (x)| ≤ ρ1/F (ê∗n , e∗ ) Fn 0 x x x Z ∞  bn F ≤ · ρh(F ) (Fn , F ) h(F )dI + keh(F )k Fn 0 0 →a.s. 0. To extend Yang’s weak convergence results, we will use the special uniform empirical processes Un of the Appendix of Shorack [1972] or Shorack and Wellner [1986] which converge to a special Brownian bridge process U in the strong sense that ρq (Un , U) →p 0 as n → ∞ for q ∈ Q(↓), the set of all continuous functions on [0, 1] which are monotone decreasing on [0, 1] and R 1 −2 dI < ∞. Thus Un = n1/2 (Γn − I) on [0, 1] where Γn is the empirical d.f. of special uniform (0, 1) 0 q random variables ξ1 , . . . , ξn . Define the mean residual life process on R+ by  Z ∞  1 1/2 1/2 1/2 n (ên (x) − e(x)) = − n (Fn − F )dI + e(x)n (Fn (x) − F (x)) Fn (x) x   Z ∞ 1 d Un (F )dI + e(x)Un (F (x)) − = Γn (F (x)) x ≡ Zn (x), 0 ≤ x < F −1 (ξnn ) where ξnn = max1≤i≤n ξi , and Zn (x) ≡ −n1/2 e(x) for x ≥ F −1 (ξnn ). Thus Zn has the same law as n1/2 (ên −e) and is a function of the special process Un . Define the corresponding limiting process Z by  Z ∞  1 Z(x) = − U(F )dI + e(x)U(F (x)) , 0 ≤ x < ∞. (2.3) F (x) x If σ 2 = V ar(X) < ∞ (and hence under either Condition 2a or 2b below), Z is a mean zero Gaussian process on R+ with covariance function described as follows: imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 5 Proposition 2.1. Suppose that σ 2 = V ar(X) < ∞. For 0 ≤ x ≤ y < ∞ Cov[Z(x), Z(y)] = F (y) σ 2 (y) V ar[Z(y)] = F (x) F (y) (2.4) where 2 σ (t) ≡ V ar[X − t|X > t] = R∞ t (x − t)2 F (x) F (t) − e2 (t) is the residual variance function; also Cov[Z(x)F (x), Z(y)F (y)] = V ar[Z(y)F (y)] = F (y)σ 2 (y). Proof. It suffices to prove (2.5). Let Z′ ≡ ZF ; from (2.3) we find Cov[Z′ (x), Z′ (y)] = e(x)e(y)F (x)F (y) − e(x) Z (2.5) ∞ F (x)F (z)dz y Z ∞ − e(y) (F (y ∧ z) − F (y)F (z))dz x Z ∞Z ∞ (F (z ∧ w) − F (z)F (w))dzdw. + x y Expressing integrals over (x, ∞) as the sum of integrals over (x, y) and (y, ∞), and recalling the defining formula for e(y), we find that the right side reduces to Z ∞Z ∞ (F (z ∧ z) − F (z)F (w)dzdw − e2 (y)F (y)F (y) y y = Z ∞ y (t − y)2 dF (t) − F (y)e2 (y) = F (y)σ 2 (y) which, being free of x, is also V ar[Z′ (y)]. As in this proposition, the process Z is often more easily studied through the process Z′ = ZF ; such a study continues in Section 5. Study of the variance of Z(x), namely σ 2 (x)/F (x), for large x appears in Section 3. Condition 2a. σ 2 < ∞ and there exists q ∈ Q(↓) such that R∞ q(F )dI/q(F (x)) < ∞. M2 ≡ M2 (q, F ) ≡ sup x e(x) x R∞ Since 0 < q(0) < ∞ and e(0) = E(X) < ∞, Condition 2a implies that 0 q(F )dI < ∞; Condition 2a may be rephrased as: M2 ≡ keq(F )/q(0) /eF k < ∞ where eq(F )/q(0) denotes the mean residual life function for the survival function q(F )/q(0). R∞ Condition 2b. σ 2 < ∞ and there exists q ∈ Q(↓) such that 0 q(F )dI < ∞. Bounded eF and existence of a moment of order greater than 2 is more than sufficient for 2b (see Section 3). imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 6 Theorem 2.2. (Process convergence). Let an → 0, nan → ∞. If Condition 2a holds for a particular q ∈ Q(↓), then ρq(F )e/F (Z∗n , Z∗ )   |Zn (x) − Z(x)|F (x) : x ≤ bn →p 0 ≡ sup q(F (x))e(x) as n → ∞. (2.6) If Condition 2b holds, then ρ1/F (Z∗n , Z∗ ) ≡ sup{|Zn (x) − Z(x)|F (x) : x ≤ bn } →p 0 as n → ∞. (2.7) Proof. First write  Zn (x) − Z(x) =  F (x) − 1 Z1n (x) + (Z1n (x) − Z(x)) Γn (F (x)) where Z1n (x) ≡ 1 F (x) ∞  Z − x  Un (F )dI + e(x)Un (F (x)) , 0 ≤ x < ∞. Then note that, using Condition 2a, ρq(F )e/F (Z1n , 0) R∞ Un (F )dI x + ρq (Un , 0) q(F (x))e(x) x ρq (Un , 0){M2 + 1)} = Op (1); ≤ sup ≤ | n →p 0 by Theorem 0 of Wellner (1978) since nan → ∞; and, again using Condition 2a, that kI/Γn − 1k1−a 0 that R∞ | x (Un (F ) − U(F ))dI| 1 ρq(F )e/F (Zn , Z) ≤ sup + ρq (Un , U) q(F (x))e(x) x ≤ ρq (Un , U){M2 + 1} →p 0. Hence 1−a n I −1 ρq(F )e/F (Z1n , 0) + ρq(F )e/F (Z1n , Z) Γn 0 = op (1)Op (1) + op (1) = op (1). ρq(F )e/F (Z∗n , Z∗ ) ≤ Similarly, using Condition 2b ρ1/F (Z1n , 0) ≤ sup ≤ sup ∞ Un (F )dI + sup e(x)|Un (F (x))| x Z ∞  ρq (Un , 0) q(F )dI + keq(F )k = Op (1), x ≤ ρ1/F (Z1n , Z) Z x 0 Z ∞ (Un (F ) − U(F ))dI + sup e(x)|Un (F (x)) − U(F (x))| x Z ∞  ≤ ρq (Un , U) q(F )dI + keq(F )k →p 0, x x 0 imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 7 and hence 1−a ρ1/F (Z∗n , Z∗ ) n I ρ1/F (Z1n , 0) + ρ1/F (Z1n , Z) −1 Γn 0 = op (1)Op (1) + op (1) = op (1). ≤ 3. Alternative sufficient conditions; V ar[Z(x)] as x → ∞. Our goal here is to provide easily checked conditions which will imply the somewhat cumbersome Conditions 2a and 2b; similar conditions also appear in the work of Balkema and de Haan [1974], and we use their results to extend their formula for the residual coefficient of variation for large x ((3.1) below). This provides a simple description of the behavior of V ar[Z(x)], the asymptotic variance of n1/2 (ên (x) − e(x)) as x → ∞. Condition 3. E(X r ) < ∞ for some r > 2. Condition 4a. Condition 3 and limx→∞ d dx (1/λ(x)) = c < ∞ where λ = f /F , the hazard function. Condition 4b. Condition 3 and lim supx→∞ {F (x)1+γ /f (x)} < ∞ for some r−1 < γ < 1/2. Proposition 3.1. If Condition 4a holds, then 0 ≤ c ≤ r−1 , Condition 2a holds, and the squared residual coefficient of variation tends to 1/(1 − 2c): 1 σ 2 (x) = . x→∞ e2 (x) 1 − 2c (3.1) lim If Condition 4b holds, then Condition 2b holds. Corollary 3.1. Condition 4a implies V ar[Z(x)] ∼ e2 (x) (1 − 2c)−1 F (x) as x → ∞. γ Proof. Assume 4a. Choose γ between r−1 and 1/2; define a d.f. G on R+ by G = F and note that g/G = γf /F = γλ. By Condition 3 xr F (x) R→ 0 as x → ∞ and hence xγr G(x) → 0 as x → ∞. Since γr > 1, ∞ G has a finite mean and therefore eG (x) = x GdI/G(x) is well-defined. Set η = 1/λ = F /f , and note that η(x)G(x) → 0 as x → ∞. (If lim sup η(x) < ∞, then it holds trivially; otherwise, η(x) → ∞ (because of 4a) and lim η(x)G(x) = lim(η(x)/x)(xG(x)) = lim η ′′ (x)xG(x) = 0 by 4a and L’Hopital. Thus by L’Hopital’s rule 0 ≤ lim η(x)G(x) η(x) = lim R ∞ eG (x) x GdI η(x)g(x) − G(x)η ′ (x) G(x) ′ = γ − lim η (x) = γ − c by 4a. = lim Thus c ≤ γ for any γ > r−1 and it follows that c ≤ r−1 . It is elementary that c ≥ 0 since η = 1/λ is nonnegative. imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 8 Choose q(t) = (1 − t)γ . Then γ − c > 0, q ∈ Q(↓), and to verify 2a it now suffices to show that lim(η(x)/eF (x)) = 1 − c < ∞ since it then follows that lim eG (x) η(x)/eF (x) 1−c = lim = < ∞. eF (x) η(x)/eG (x) γ−c By continuity and eG (0) < ∞, 0 < eF (0) < ∞, this implies Condition 2a. But r > 2 implies that xF (x) → 0 as x → ∞ so η(x)F (x) → 0 and hence lim ηF (x) η(x) = lim R ∞ = lim(1 − η ′ (x)) = 1 − c. eF (x) F dI x That (3.1) holds will now follows from results of Balkema and de Haan [1974], as follows: Their Corollary to Theorem 7 implies that P (λ(t)(X − t) > x|X > t) → e−x if c = 0 and → (1 + cx)−1/c if c > 0. Thus, in the former case, F is in the domain of attraction of the Pareto residual life distribution and its related extreme value distribution. Then Theorem 8(a) implies convergence of the (conditional) mean and variance of λ(t)(X − t) to the mean and variance of the limiting Pareto distribution, namely (1 − c)−1 and (1 − c)−2 (1 − 2c)−1 . But the conditional mean of λ(t)(Xt ) is simply λ(t)e(t), so that λ(t) ∼ (1 − c)−1 /e(t) and (3.1) now follows. R∞ γ If Condition 4b holds, let q(F ) = F again. Then 0 q(F )dI < ∞, and it remains to show that lim sup{e(x)F (x)γ } < ∞. This follows from 4b by L’Hopital. Similarly, sufficient conditions for Conditions 1a and 1b can be given: simply replace “2” in Condition 3 and “1/2” in Condition 4b with “1”, and the same proof works. Whether (3.1) holds when r in Condition 3 is exactly 2 is not known. 4. Examples. The typical situation, when e(x) has a finite limit and Condition 3 holds, is as follows: e ∼ F /f ∼ f /(−f ′ ) as x → ∞ (by L’Hopital), and hence 4b, 2b, and 1b hold; also η ′ ≡ (F /f )′ = [(F/f )(−f /f ′)] − 1 → 0 (4a with c = 0, and hence 2a and 1a hold), σ(x) ∼ e(x) from (3.1), and V ar[Z] ∼ e2 /F ∼ (F /f )2 /F ∼ 1/(−f ′ ). We treat three examples, not all ‘typical’, in more detail. Example 4.1. (Exponential). Let F (x) = exp(−x/θ), x ≥ 0, with 0 < θ < ∞. Then e(x) = θ for all x ≥ 0. Conditions 4a and 4b hold (for all r, γ ≥ 0) with c = 0, so Conditions 2a and 2b hold by Proposition 3.1 with q(t) = (1 − t)1/2−δ , 0 < δ < 1/2. Conditions 1a and 1b hold with h(t) = (1 − t)1−δ , 0 < δ < 1. Hence Theorems 2.1 and 2.2 hold where now Z 1 U(F (x)) 1 U d Z(x) = − dI = θB(ex/θ ), 0 ≤ x < ∞ 1 − (x) 1 − F (x) F (x) 1 − I R1 and B is standard Brownian motion on [0, ∞). (The process B1 (t) = U(1 − t) − 1−t (U/(1 − I))dI, 0 ≤ t ≤ 1, is Brownian motion on [0, 1]; and with B2 (x) ≡ xB1 (1/x) for 1 ≤ x ≤ ∞, Z(x) = θB2 (1/F (x)) = θB2 (ex/θ ).) Thus, in agreement with (2.4), Cov[Z(x), Z(y)] = θ2 e(x∧y)/θ , 0 ≤ x, y < ∞. d An immediate consequence is that kZ∗n F k →d kF k = θ sup0≤t≤1 |B1 (t)|; generalization of this to other F ’s appears in Section 5. (Because of the “memoryless” property of exponential F , the results for this example can undoubtedly be obtained by more elementary methods.) imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 9 Example 4.2. (Weibull). Let F (x) = exp(−xθ ), x ≥ 0, with 0 < θ < ∞. Conditions 4a and 4b hold (for all r, γ > 0) with c = 0, so Conditions 1 and 2 hold with h and q as in Example 1 by Proposition 3.1. Thus Theorems 2.1 and 2.2 hold. Also, e(x) ∼ θ−1 x1−θ as x → ∞, and hence V ar[Z(x)] ∼ θ−2 x2(1−θ) exp(xθ ) as x → ∞. Example 4.3. (Pareto). Let F (x) = (1 + cx)−1/c , x ≥ 0, with 0 < c < 1/2. Then e(x) = (1 − c)−1 (1 + cx), and Conditions 4a and 4b hold for r < c−1 and γ ≥ c (and c of 4a is c). Thus Proposition 3.1 holds with r > 2 and c > 0 and V ar[Z(x)] ∼ c2+(1/c) (1 − c)−2 (1 − 2c)−1 x2+(1/c) as x → ∞. Conditions 1 and 2 hold with h and q as in Example 1, and Theorems 2.1 and 2.2 hold. If instead 1/2 ≤ c < 1, then E(X) < ∞ but E(X 2 ) = ∞, and 4a and 4b hold with 1 < r < 1/c ≤ 2 and γ ≥ c. Hence Condition 1 and Theorem 2.1 hold, but Condition 2 (and hence our proof of Theorem 2.2) fails. If c ≥ 1, then E(X) = ∞ and e(x) = ∞ for all x ≥ 0. Not surprisingly, the limiting process Z has a variance which grows quite rapidly, exponentially in the exponential and Weibull cases, and as a power (> 4) of x in the Pareto case. 5. Confidence bands for e. We first consider the process Z′ ≡ ZF on R+ which appeared in (2.5) of Proposition 2.1. Its sample analog Z′n ≡ Zn Fn is easily seen to be a cumulative sum (times n−1/2 ) of the observations exceeding x, each centered at x+ e(x); as x decreases the number of terms in the sum increases. Moreover, the corresponding increments apparently act asymptotically independently so that Z′n , in reverse time, is behaving as a cumulative sum of zero-mean independent increments. Adjustment for the non-linear variance should lead to Brownian motion. Let us return to the limit version Z′ . The zero-mean Gaussian process Z′ has covariance function Cov[Z′ (x), Z′ (y)] = V ar[Z′ (x ∨ y)] (see (2.5)); hence, when viewed in reverse time, it has independent increments (and hence Z′ is a reverse martingale). Specifically, with Z′′ (s) ≡ Z′ (− log s), Z′′ is a zero-mean Gaussian process on [0, 1] with independent increments and V ar[Z′′ (s)] = V ar[Z′ (− log s)] ≡ τ 2 (s). Hence τ 2 is increasing in s, and, from (2.5), τ 2 (s) = F (− log s)σ 2 (− log s). (5.1) Now τ 2 (1) = σ 2 (0) = σ 2 , and τ 2 (0) = lim F (− log ǫ)σ 2 (− log ǫ) = lim F (x)σ 2 (x) = 0 x→∞ ǫ↓0 since 0 ≤ F (x)σ 2 (x) ≤ F (x)E(X 2 |X > x) = Z ∞ x y 2 dF (y) → 0. Since f (x) > 0 for all x ≥ 0, τ 2 is strictly increasing. Let g be the inverse of τ 2 ; then τ 2 (g(t)) = t, g(0) = 0, and g(σ 2 ) = 1. Define W on [0, 1] by W(t) ≡ σ −1 Z′′ (g(σ 2 t)) = σ −1 Z′ (− log g(σ 2 t)). (5.2) Theorem 5.1. W is standard Brownian motion on [0, 1]. Proof. W is Gaussian with independent increments and V ar[W(t)] = t by direct computation. Corollary 5.1. If (2.7) holds, then ′ ′ ρ(Zn∗ , Z ∗ ) ≡ sup |Zn (x)Fn (x) − Z(x)F (x)| = op (1) x≤bn and hence kZn Fn kb0n →d kZF k = σkWk10 as n → ∞. imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 10 Proof. By Theorem 0 of Wellner [1978] kFn /F − 1kb0n →p 0 as n → ∞, and this together with (2.7) implies the first part of the statement. The second part follows immediately from the first and (5.2). Replacement of σ 2 by a consistent estimate Sn2 (e.g. the sample variance based on all observations), and of bn by b̂n = F−1 n (1 − an ), the (n − m)−th order statistic with m = [nan ], leads to asymptotic confidence bands for e = eF : Corollary 5.2. Let 0 < a < ∞, and set dˆn (x) ≡ n−1/2 aSn /Fn (x). If (2.7) holds, Sn2 →p σ 2 , and nan / log log n ↑ ∞, then, as n → ∞   P ên (x) − dˆn (x) ≤ e(x) ≤ ên (x) + dˆn (x) for all 0 ≤ x ≤ b̂n → Q(a) (5.3) where Q(a) ≡ P (kWk10 < a) = ∞ X k=−∞ (−1)k {Φ((2k + 1)a) − Φ((2k − 1)a)} = 1 − 4{Φ(a) − Φ(3a) + Φ(5a) − · · · } and Φ denotes the standard normal d.f. Proof. It follows immediately from Corollary 5.1 and Sn →p σ > 0 that kZn Fn kb0n /Sn →d kZF k/σ = kWk10 . Finally bn may be replaced by b̂n without harm: letting cn = 2 log log /(nan ) → 0 and using Theorem 4S of d 1/2 −1 −1 Wellner [1978], for τ > 1 and all n ≥ N (ω, τ ), b̂n ≡ F−1 (Γ−1 ({1 + τ cn }(1 − n (1 − an ) = F n (1 − an )) ≤ F an )) w.p. 1. This proves the convergence claimed in the corollary; the expression for Q(a) is well-known (e.g. see Billingsley [1968], page 79). The approximation 1 − 4Φ(a) for Q(a) gives 3-place accuracy for a > 1.4. A short table appears below: Table 1 Q(a) for selected a a 2.807 2.241 1.960 Q(a) .99 .95 .90 a 1.534 1.149 0.871 Q(a) .75 .50 .25 Thus, choosing a so that Q(a) = β, (5.3) provides a two-sided simultaneous confidence band for the function e with confidence coefficient asymptotically β. In applications we suggest taking an = n−1/2 so that b̂n is the (n − m)−th order statistic with m = [n1/2 ]; we also want m large enough for an adequate central limit effect, remembering that the conditional life distribution may be quite skewed. (In a similar fashion, one-sided asymptotic bands are possible, but they will be less trustworthy because of skewness.) Instead of simultaneous bands for all real x one may seek (tighter) bands on e(x) for one or two specific x−values. For this we can apply Theorem 2.2 and Proposition 2.1 directly. We first require a consistent estimator of the asymptotic variance of n1/2 (ên (x) − e(x)), namely σ 2 (x)/F (x). Proposition 5.1. Let 0 ≤ x < ∞ be fixed and let Sn2 (x) be the sample variance of those observations exceeding x. If Condition 3 holds then Sn2 (x)/Fn (x) →a.s. σ 2 (x)/F (x). imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 11 Proof. Since Fn (x) →a.s. F (x) > 0 and Sn2 (x) = 2 R∞ x (y − x)Fn (y)dy Fn (x) − ê2n (x), R∞ R∞ it suffices to show that x yFn (y)dy →a.s. x yF (y)dy. Let h(t) = (1 − t)γ+1/2 and q(t) = (1 − t)γ with R∞ r−1 < γ < 1/2 so that h ∈ H(↓), q ∈ Q(↓), and 0 q(F )dI < ∞ by the proof of Proposition 3.1. Then, Z ∞ Z ∞ Z ∞ yFn (y)dy − yF (y)dy ≤ ρh(F ) (Fn , F ) Ih(F )dI →a.s. 0 x x 0 by Theorem 1 of Wellner (1977) since Z ∞ Z Ih(F )dI = 0 ∞ 0 (I 2 F )1/2 q(F )dI < ∞. By Theorem 2.2, Propositions 2.1 and 5.1, and Slutsky’s theorem we have: Proposition 5.2. Under the conditions of Proposition 5.1, 1/2 dn (x) ≡ n1/2 (ên (x) − e(x))Fn (x)/Sn (x) →d N (0, 1) as n → ∞. This makes feasible an asymptotic confidence interval for e(x) (at this particular fixed x). Similarly, for x < y, using the joint asymptotic normality of (dn (x), dn (y)) with asymptotic correlation {F (y)σ 2 (y)/F (x)σ 2 (x)}1/2 estimated by {Fn (y)Sn2 (y)/Fn (x)Sn2 (x)}1/2 , an asymptotic confidence ellipse for (e(x), e(y)) may be obtained. 6. Illustration of the confidence bands. We illustrate with two data sets presented by Bjerkedal [1960] and briefly mention one appearing in Barlow and Campo (1975). Bjerkedal gave various doses of tubercle bacilli to groups of 72 guinea pigs and recorded their survival times. We concentrate on Regimens 4.3 and 6.6 (and briefly mention 5.5, the only other complete data set in Bjerkedal’s study M); see Figures 1 and 2 below. First consider the estimated mean residual life ên , the center jagged line in each figure. Figure 1 has been terminated at day 200; the plot would continue approximately horizontally, but application of asymptotic theory to this part of ên , based only the last 23 survival times (the last at 555 days), seems unwise. Figure 2 has likewise been terminated at 200 days, omitting only nine survival times (the last at 376 days); the graph of ên would continue downward. The dashed diagonal line is X − x; if all survival times were equal, say µ, then the residual life function would be (µ − x)+ , a lower bound on e(x) near the origin. More specifically, a Maclaurin expansion yields e(x) = µ + (µf0 − 1)x + (1/2){(2µf0 − 1)f0 + f0′ }x2 + o(x2 ) where f0 = f (0), f0′ = f ′ (0), if f ′ is continuous at 0, or e(x) = µ − x + µd r x + o(xr ) r! imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 12 if f (s) (0) = 0 for s < (r − 1) (≥ 0) and = d for s = r − 1 (if f (r−1) is continuous at 0). It thus seems likely from Figures 1 and 2 that in each of these cases either f0 = 0 and f0′ > 0 or f0 is near 0 (and f0′ ≥ 0). Also, for large x, e(x) ∼ 1/λ(x), and Figure 1 suggests that the corresponding λ and e have finite positive limits, whereas the e of Figure 2 may eventually decrease (λ increase). We know of no parametric F that would exhibit behavior quite like these. The upper and lower jagged lines in the figures provide 90% (asymptotic) confidence bands for the respective e’s, based on (5.3). At least for Regimen 4.3, a constant e (exponential survival) can be rejected. The vertical bars at x = 0, x = 100, and x = 200 in Figure 1, and at 0, 50, and 100 in Figure 2, are 90% (asymptotic) pointwise confidence intervals on e at the corresponding x−values (based on Proposition 5.2). Notice that these intervals are not much narrower than the simultaneous bands early in the survival data, but are substantially narrower later on. A similar graph for Regimen 5.5 (not presented) is somewhat similar to that in Figure 2, with the upward turn in ên occurring at 80 days instead of at 50, and a possible downward turn at somewhere around 250 days (the final death occurring at 598 days). A similar graph was prepared for the failure data on 107 right rear tractor brakes presented by Barlow and Campo [1975], page 462. It suggests a quadratic decreasing e for the first 1500 to 2000 hours (with f (0) at or near 0 but f ′ (0) definitely positive), with X = 2024, and with a possibly constant or slightly increasing e from 1500 or so to 6000 hours. The e for a gamma distribution with λ = 2 and α = .001 (e(x) = α−1 (αx+2)/(αx+1) with α = .001) fits reasonably well – i.e. is within the confidence bands, even for 25% confidence. Note that this is in excellent agreement with Figures 2.1(b) and 3.1(d) of Barlow and Campo [1975]. Bryson and Siddiqui’s (1969) data set was too small (n = 43) for these asymptotic methods, except possibly early in the data set.) 7. Further developments The original version of this paper, Hall and Wellner [1979], ended with a one-sentence sketch of two remaining problems: “Confidence bands on the difference between two mean residual life functions, and for the case of censored data, will be presented in subsequent papers.” Although we never did address these questions ourselves, others took up these further problems. Our aim in this final section is to briefly survey some of the developments since 1979 concerning mean residual life, including related studies of median residual life and other quantiles, as well as developments for censored data, alternative inference strategies, semiparametric models involving mean or median residual life, and generalizations to higher dimensions. For a review of further work up to 1988 see Guess and Proschan [1988]. 7.1. Confidence bands and inference Csörgő et al. [1986] gave a further detailed study of the asymptotic behavior of the mean residual life process as well as other related processes including the Lorenz curve. Berger et al. [1988] developed tests and confidence sets for comparing two mean residual life functions based on independent samples from the respective populations. These authors also gave a brief treatment based on comparison of median residual life, to be discussed in Subsection 7.3 below. Csörgő and Zitikis [1996] introduced weighted metrics into the study of the asymptotic behavior of the mean residual life process, thereby avoiding the intervals [0, xn ] changing with n involved in our Theorems 2.1 and 2.2, and thereby provided confidence bands and intervals for eF in the right tail. Zhao and Qin [2006] introduced empirical likelihood methods to the study of the mean residual life function. They obtained confidence intervals and confidence bands for compact sets [0, τ ] with τ < τF ≡ inf{x : F (x) = 1}. imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 13 Fig 1: 90% confidence bands for mean residual life; Regimen 4.3 imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 14 Fig 2: 90% confidence bands for mean residual life; Regimen 6.6 imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 15 7.2. Censored data Yang [1978a] initiated the study of estimated mean residual life under random right censorship. She used an estimator F̂n which is asymptotically equivalent to the Kaplan - Meier estimator and considered, in particular, the case when X is bounded and stochastically smaller than the censoring variable C. In this √ case she proved that n(ê(x) − e(x)) converges weakly (as n → ∞) to a Gaussian process with mean zero. Csörgő and Zitikis [1996] give a brief review of the challenges involved in this problem; see their page 1726. Qin and Zhao [2007] extended their earlier study (Zhao and Qin [2006]) of empirical likelihood methods to this case, at least for the problem of obtaining pointwise confidence intervals. The empirical likelihood methods seem to have superior coverage probability properties in comparison to the Wald type intervals which follow from our Proposition 5.2. Chaubey and Sen [1999] introduced smooth estimates of mean residual life in the uncensored case. In Chaubey and Sen [2008] they introduce and study smooth estimators of eF based on corresponding smooth estimators of F = 1 − F introduced by Chaubey and Sen [1998]. 7.3. Median and quantile residual life functions Because mean residual life is frequently difficult, if not impossible, to estimate in the presence of rightcensoring, it is natural to consider surrogates for it which do not depend on the entire right tail of F . Natural replacements include median residual life and corresponding residual life quantiles. The study of median residual life was apparently initiated in Schmittlein and Morrison [1981]. Characterization issues and basic properties have been investigated by Gupta and Langford [1984], Joe and Proschan [1984b], and Lillo [2005]. Joe and Proschan [1984a] proposed comparisons of two populations based on their corresponding median (and other quantile) residual life functions. As noted by Joe and Proschan, “Some results differ notably from corresponding results for the mean residual life function”. Jeong et al. [2008] investigated estimation of median residual life with right-censored data for one-sample and two-sample problems. They provided an interesting illustration of their methods using a long-term follow-up study (the National Surgical Adjuvant Breast and Bowel Project, NSABP) involving breast cancer patients. 7.4. Semiparametric models for mean and median residual life Oakes and Dasu [1990] investigated a characterization related to a proportional mean residual life model: eG = ψeF with ψ > 0. Maguluri and Zhang [1994] studied several methods of estimation in a semiparametric regression version of the proportional mean residual life model, e(x|z) = exp(θT z)e0 (x) where e(x|z) denotes the conditional mean residual life function given Z = z. Chen et al. [2005] provide a nice review of various models and study estimation in the same semiparametric proportional mean residual life regression model considered by Maguluri and Zhang [1994], but in the presence of right censoring. Their proposed estimation method involves inverse probability of censoring weighted (IPCW) estimation methods (Horvitz and Thompson [1952]; Robins and Rotnitzky [1992]). Chen and Cheng [2005] use counting process methods to develop alternative estimators for the proportional mean residual life model in the presence of right censoring. The methods of estimation considered by Maguluri and Zhang [1994], Chen et al. [2005], and Chen and Cheng [2005] are apparently inefficient. Oakes and Dasu [2003] consider information calculations and likelihood based estimation in a two-sample version of the proportional mean residual life model. Their calculations suggest that certain weighted ratio-type estimators may achieve asymptotic efficiency, but a definitive answer to the issue of efficient estimation apparently remains unresolved. Chen and Cheng [2006] proposed an alternative additive semiparametric regression model involving mean residual life. Ma and Yin [2010] considered a large family of semiparametric regression models which includes both the additive model proposed by Chen and Cheng [2006] and the proportional mean residual life model considered by earlier imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 16 authors, but advocated replacing mean residual life by median residual life. Gelfand and Kottas [2003] also developed a median residual life regression model with additive structure and took a semiparametric Bayesian approach to inference. 7.5. Monotone and Ordered mean residual life functions Kochar et al. [2000] consider estimation of eF subject to the shape restrictions that eF is increasing or decreasing. The main results concern ad-hoc estimators that are simple monotizations of the basic nonparametric empirical estimators ên studied here. These authors show that the nonparametric maximum likelihood estimator does not exist in the increasing MRL case and that although the nonparametric MLE exists in the decreasing MRL case, the estimator is difficult to compute. Ebrahimi [1993] and Hu et al. [2002] study estimation of two mean residual life functions eF and eG in one- and two-sample settings subject to the restriction eF (x) ≤ eG (x) for all x. Hu et al. [2002] also develop large sample confidence bands and intervals to accompany their estimators. 7.6. Bivariate residual life Jupp and Mardia [1982] defined a multivariate mean residual life function and showed that it uniquely determines the joint multivariate distribution, extending the known univariate result of Cox [1962]; see Hall and Wellner [1981] for a review of univariate results of this type. See Ma [1996, 1998] for further multivariate characterization results. Kulkarni and Rattihalli [2002] introduced a bivariate mean residual life function and propose natural estimators. Remarks: This revision was accomplished jointly by the authors in 2011 and 2012. The first author passed away in October 2012. Section 7 only covers further developments until 2012. A MathSciNet search for “mean residual life” over the period 2011 - 2017 yielded 148 hits on 10 July 2017. References Balkema, A. A. and de Haan, L. (1974). Residual life time at great age. Ann. Probability, 2, 792–804. Barlow, R. E. and Campo, R. (1975). Total time on test processes and applications to failure data analysis. In Reliability and fault tree analysis (Conf., Univ. California, Berkeley, Calif., 1974), pages 451–481. Soc. Indust. Appl. Math., Philadelphia, Pa. Berger, R. L., Boos, D. D., and Guess, F. M. (1988). Tests and confidence sets for comparing two mean residual life functions. Biometrics, 44(1), 103–115. Billingsley, P. (1968). Convergence of probability measures. John Wiley & Sons Inc., New York. Bjerkedal, T. (1960). Acquisition of resistance in guinea pigs infected with different doses of virulent tubercle bacilli. Amer. Jour. Hygiene, 72, 130–148. Bryson, M. C. and Siddiqui, M. M. (1969). Some criteria for aging. J. Amer. Statist. Assoc., 64, 1472–1483. Chaubey, Y. P. and Sen, A. (2008). Smooth estimation of mean residual life under random censoring. In Beyond parametrics in interdisciplinary research: Festschrift in honor of Professor Pranab K. Sen, volume 1 of Inst. Math. Stat. Collect., pages 35–49. Inst. Math. Statist., Beachwood, OH. Chaubey, Y. P. and Sen, P. K. (1998). On smooth estimation of hazard and cumulative hazard functions. In Frontiers in probability and statistics (Calcutta, 1994/1995), pages 91–99. Narosa, New Delhi. Chaubey, Y. P. and Sen, P. K. (1999). On smooth estimation of mean residual life. J. Statist. Plann. Inference, 75(2), 223–236. The Seventh Eugene Lukacs Conference (Bowling Green, OH, 1997). imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 17 Chen, Y. Q. and Cheng, S. (2005). Semiparametric regression analysis of mean residual life with censored survival data. Biometrika, 92(1), 19–29. Chen, Y. Q. and Cheng, S. (2006). Linear life expectancy regression with censored data. Biometrika, 93(2), 303–313. Chen, Y. Q., Jewell, N. P., Lei, X., and Cheng, S. C. (2005). Semiparametric estimation of proportional mean residual life model in presence of censoring. Biometrics, 61(1), 170–178. Chiang, C. L. (1960). A stochastic study of the life table and its applications: I. probability distributions of the biometric functions. Biometrics, 16, 618–635. Chiang, C. L. (1968). Introduction to Stochastic Processes in Biostatistics. Wiley, New York. Cox, D. R. (1962). Renewal theory. Methuen & Co. Ltd., London. Csörgő, M. and Zitikis, R. (1996). Mean residual life processes. Ann. Statist., 24(4), 1717–1739. Csörgő, M., Csörgő, S., and Horváth, L. (1986). An asymptotic theory for empirical reliability and concentration processes, volume 33 of Lecture Notes in Statistics. Springer-Verlag, Berlin. Ebrahimi, N. (1993). Estimation of two ordered mean residual lifetime functions. Biometrics, 49(2), 409–417. Gelfand, A. E. and Kottas, A. (2003). Bayesian semiparametric regression for median residual life. Scand. J. Statist., 30(4), 651–665. Gross, A. J. and Clark, V. A. (1975). Survival distributions: reliability applications in the biomedical sciences. Wiley, New York. Guess, F. and Proschan, F. (1988). Mean residual life: theory and application. In Handbook of Statistics: Quality Control and Reliability, volume 7, pages 215–224. North-Holland, Amsterdam. Gupta, R. C. and Langford, E. S. (1984). On the determination of a distribution by its median residual life function: a functional equation. J. Appl. Probab., 21(1), 120–128. Hall, W. J. and Wellner, J. (1981). Mean residual life. In Statistics and related topics (Ottawa, Ont., 1980), pages 169–184. North-Holland, Amsterdam. Hall, W. J. and Wellner, J. A. (1979). Estimation of mean residual life. Technical report, University of Rochester. Hollander, M. and Proschan, F. (1975). Tests for the mean residual life. Biometrika, 62(3), 585–593. Horvitz, D. G. and Thompson, D. J. (1952). A generalization of sampling without replacement from a finite universe. J. Amer. Statist. Assoc., 47, 663–685. Hu, X., Kochar, S. C., Mukerjee, H., and Samaniego, F. J. (2002). Estimation of two ordered mean residual life functions. J. Statist. Plann. Inference, 107(1-2), 321–341. Statistical inference under inequality constraints. Jeong, J.-H., Jung, S.-H., and Costantino, J. P. (2008). Nonparametric inference on median residual life function. Biometrics, 64(1), 157–163, 323–324. Joe, H. and Proschan, F. (1984a). Comparison of two life distributions on the basis of their percentile residual life functions. Canad. J. Statist., 12(2), 91–97. Joe, H. and Proschan, F. (1984b). Percentile residual life functions. Oper. Res., 32(3), 668–678. Jupp, P. E. and Mardia, K. V. (1982). A characterization of the multivariate Pareto distribution. Ann. Statist., 10(3), 1021–1024. Kochar, S. C., Mukerjee, H., and Samaniego, F. J. (2000). Estimation of a monotone mean residual life. Ann. Statist., 28(3), 905–921. Kulkarni, H. V. and Rattihalli, R. N. (2002). Nonparametric estimation of a bivariate mean residual life function. J. Amer. Statist. Assoc., 97(459), 907–917. Lillo, R. E. (2005). On the median residual lifetime and its aging properties: a characterization theorem and applications. Naval Res. Logist., 52(4), 370–380. Ma, C. (1996). Multivariate survival functions characterized by constant product of mean remaining lives and hazard rates. Metrika, 44(1), 71–83. Ma, C. (1998). Characteristic properties of multivariate survival functions in terms of residual life imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017 Hall & Wellner/Mean Residual Life 18 distributions. Metrika, 47(3), 227–240. Ma, Y. and Yin, G. (2010). Semiparametric median residual life model and inference. Canad. J. Statist., 38(4), 665–679. Maguluri, G. and Zhang, C.-H. (1994). Estimation in the mean residual life regression model. J. Roy. Statist. Soc. Ser. B , 56(3), 477–489. Oakes, D. and Dasu, T. (1990). A note on residual life. Biometrika, 77(2), 409–410. Oakes, D. and Dasu, T. (2003). Inference for the proportional mean residual life model. In Crossing boundaries: statistical essays in honor of Jack Hall , volume 43 of IMS Lecture Notes Monogr. Ser., pages 105–116. Inst. Math. Statist., Beachwood, OH. Qin, G. and Zhao, Y. (2007). Empirical likelihood inference for the mean residual life under random censorship. Statist. Probab. Lett., 77(5), 549–557. Robins, J. M. and Rotnitzky, A. (1992). Recovery of information and adjustment for dependent censoring using surrogate markers. In AIDS Epidemiology, Methodological Issues, pages 297–331. Birkhauser, Boston. Schmittlein, D. C. and Morrison, D. G. (1981). The median residual lifetime: a characterization theorem and an application. Oper. Res., 29(2), 392–399. Shorack, G. R. (1972). Functions of order statistics. Ann. Math. Statist., 43, 412–427. Shorack, G. R. and Wellner, J. A. (1986). Empirical Processes with Applications to Statistics. Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics. John Wiley & Sons Inc., New York. Wellner, J. A. (1977). A Glivenko-Cantelli theorem and strong laws of large numbers for functions of order statistics. Ann. Statist., 5(3), 473–480. Wellner, J. A. (1978). Limit theorems for the ratio of the empirical distribution function to the true distribution function. Z. Wahrsch. Verw. Gebiete, 45(1), 73–88. Wilson, E. B. (1938). The standard deviation of sampling for life expectancy. J. Amer. Statist. Assoc., 33, 705–708. Yang, G. (1978a). Life expectancy under random censorship. Stochastic Processes Appl., 6(1), 33–39. Yang, G. L. (1978b). Estimation of a biometric function. Ann. Statist., 6(1), 112–116. Zhao, Y. and Qin, G. (2006). Inference for the mean residual life function via empirical likelihood. Comm. Statist. Theory Methods, 35(4-6), 1025–1036. imsart-generic ver. 2012/08/31 file: MRL-revived4c.tex date: July 12, 2017
10math.ST
Discourse-Based Objectives for Fast Unsupervised Sentence Representation Learning Yacine Jernite Samuel R. Bowman Department of Computer Science Department of Linguistics New York University and Center for Data Science [email protected] New York University [email protected] David Sontag Department of EECS Massachussets Institute of Technology [email protected] arXiv:1705.00557v1 [cs.CL] 23 Apr 2017 Abstract This work presents a novel objective function for the unsupervised training of neural network sentence encoders. It exploits signals from paragraph-level discourse coherence to train these models to understand text. Our objective is purely discriminative, allowing us to train models many times faster than was possible under prior methods, and it yields models which perform well in extrinsic evaluations. 1 Task 1: ORDER classifier Task 2: CONJUNCTION classifier Task 3: NEXT classifier Sentence encoder Sentence encoder Sentence(s) 1 Sentence(s) 2 Figure 1: We train a sentence encoder (shown as two copies with shared parameters) on three discourse-based objectives over unlabeled text. Introduction Modern artificial neural network approaches to natural language understanding tasks like translation (Sutskever et al., 2014; Cho et al., 2014), summarization (Rush et al., 2015), and classification (Yang et al., 2016) depend crucially on subsystems called sentence encoders that construct distributed representations for sentences. These encoders are typically implemented as convolutional (Kim, 2014), recursive (Socher et al., 2013), or recurrent neural networks (Mikolov et al., 2010) operating over a sentence’s words or characters (Zhang et al., 2015; Kim et al., 2016). Most of the early successes with sentence encoder-based models have been on tasks with ample training data, where it has been possible to train the encoders in a fully-supervised end-to-end setting. However, recent work has shown some success in using unsupervised pretraining with unlabeled data to both improve the performance of these methods and extend them to lower-resource settings (Dai and Le, 2015; Kiros et al., 2015; Bajgar et al., 2016). This paper presents a set of methods for unsupervised pretraining that train sentence encoders to recognize discourse coherence. When reading text, human readers have an expectation of coher- ence from one sentence to the next. In most cases, for example, each sentence in a text should be both interpretable in context and relevant to the topic under discussion. Both of these properties depend on an understanding of the local context, which includes both relatively knowledge about the state of the world and the specific meanings of previous sentences in the text. Thus, a model that is successfully trained to recognize discourse coherence must be able to understand the meanings of sentences as well as relate them to key pieces of knowledge about the world. Hobbs (1979) presents a formal treatment of this phenomenon. He argues that for a discourse (here, a text) to be interpreted as coherent, any two adjacent sentences must be related by one of a few set kinds of coherence relations. For example, a sentence might be followed by another that elaborates on it, parallels it, or contrasts with it. While this treatment may not be adequate to cover the full complexity of language understanding, it allows Hobbs to show how identifying such relations depends upon sentence understanding, coreference resolution, and commonsense reasoning. Recently proposed techniques (Kiros et al., 2015; Ramachandran et al., 2016) succeed in ex- ploiting discourse coherence information of this kind to train sentence encoders, but rely on generative objectives which require models to compute the likelihood of each word in a sentence at training time. In this setting, a single epoch of training on a typical (76M sentence) text corpus can take weeks, making further research difficult, and making it nearly impossible to scale these methods to the full volume of available unlabeled English text. In this work, we propose alternative objectives which exploit much of the same coherence information at greatly reduced cost. In particular, we propose three fast coherencebased pretraining tasks, show that they can be used together effectively in multitask training (Figure 1), and evaluate models trained in this setting on the training tasks themselves and on standard text classification tasks.1 We find that our approach makes it possible to learn to extract broadly useful sentence representations in hours. 2 Related Work This work is inspired most directly by the Skip Thought approach of Kiros et al. (2015), which introduces the use of paragraph-level discourse information for the unsupervised pretraining of sentence encoders. Since that work, three other papers have presented improvements to this method (the SDAE of Hill et al. 2016, also Gan et al. 2016; Ramachandran et al. 2016). These improved methods are based on techniques and goals that are similar to ours, but all three involve models that explicitly generate full sentences during training time at considerable computational cost. In closely related work, Logeswaran et al. (2016) present a model that learns to order the sentences of a paragraph. While they focus on learning to assess coherence, they show positive results on measuring sentence similarity using their trained encoder. Alternately, the FastSent model of Hill et al. (2016) is designed to work dramatically more quickly than systems like Skip Thought, but in service of this goal the standard sentence encoder RNN is replaced with a low-capacity CBOW model. Their method does well on existing semantic textual similarity benchmarks, but its insensitivity to order places an upper bound on its performance in more intensive extrinsic language understanding tasks. 1 All code, resources, and models involved in these experiments will be made available upon publication. Sentence Pair Label Relation A strong one at that. Then I became a woman. Y elaboration I saw flowers on the ground. I heard birds in the trees. N list It limped closer at a slow pace. Soon it stopped in front of us. N spatial I kill Ben, you leave by yourself. I kill your uncle, you join Ben. Y time Table 1: The binary ORDER objective. Discourse relation labels are provided for the reader, but are not available to the model. Looking beyond work on unsupervised pretraining: Li and Hovy (2014) and Li and Jurafsky (2016) use representation learning systems to directly model the problem of sentence order recovery, but focus primarily on intrinsic evaluation rather than transfer. Wang and Cho (2016) train sentence representations for use as context in language modeling. In addition, Ji et al. (2016) treat discourse relations between sentences as latent variables and show that this yields improvements in language modeling in an extension of the document-context model of Ji et al. (2015). Outside the context of representation learning, there has been a good deal of work in NLP on discourse coherence, and on the particular tasks of sentence ordering and coherence scoring. Barzilay and Lapata (2008) provide thorough coverage of this work. 3 Discourse Inspired Objectives In this work, we propose three objective functions for use over paragraphs extracted from unlabeled text. Each captures a different aspect of discourse coherence and together the three can be used to train a single encoder to extract broadly useful sentence representations. Binary Ordering of Sentences Many coherence relations have an inherent direction. For example, if S1 is an elaboration of S0 , S0 is not generally an elaboration of S1 . Thus, being able to identify these coherence relations implies an ability to recover the original order of the sentences. Our first task, which we call ORDER, consists in taking pairs of adjacent sentences from text data, switching their order with probability 0.5, and training a model to decide whether they have been switched. Table 1 provides some examples of Context No, not really. I had some ideas, some plans. But I never even caught sight of them. Candidate Successors 1. 2. 3. 4. 5. There’s nothing I can do that compares that. Then one day Mister Edwards saw me. I drank and that was about all I did. And anyway, God’s getting his revenge now. He offered me a job and somewhere to sleep. Table 2: The NEXT objective. Sentence Pair Label He had a point. For good measure, I pouted. (Still) RETURN It doesn’t hurt at all. It’s exhilarating. (In fact) The waterwheel hammered on. There was silence. (Otherwise) STRENGTHEN CONTRAST Table 3: The CONJUNCTION objective. Discourse relation labels are shown with the text from which they were derived. this task, along with the kind of coherence relation that we assume to be involved. It should be noted that since some of these relations are unordered, it is not always possible to recover the original order based on discourse coherence alone (see e.g. the flowers / birds example). Next Sentence Many coherence relations are transitive by nature, so that any two sentences from the same paragraph will exhibit some coherence. However, two adjacent sentences will generally be more coherent than two more distant ones. This leads us to formulate the NEXT task: given the first three sentences of a paragraph and a set of five candidate sentences from later in the paragraph, the model must decide which candidate immediately follows the initial three in the source text. Table 2 presents an example of such a task: candidates 2 and 3 are coherent with the third sentence of the paragraph, but the elaboration (3) takes precedence over the progression (2). Conjunction Prediction Finally, information about the coherence relation between two sentences is sometimes apparent in the text (Miltsakaki et al., 2004): this is the case whenever the second sentence starts with a conjunction phrase. To form the CONJUNCTION objective, we create a list of conjunction phrases and group them into nine categories (see supplementary material). We then extract from our source text all pairs of sentences where the second starts with one of the listed conjunctions, give the system the pair without the phrase, and train it to recover the conjunction category. Table 3 provides examples. 4 Experiments In this section, we introduce our training data and methods, present qualitative results and comparisons among our three objectives, and close with quantitative comparisons with related work. Experimental Setup We train our models on a combination of data from BookCorpus (Zhu et al., 2015), the Gutenberg project (Stroube, 2003), and Wikipedia. After sentence and word tokenization (with NLTK; Bird, 2006) and lower-casing, we identify all paragraphs longer than 8 sentences and extract a NEXT example from each, as well as pairs of sentences for the ORDER and CONJUNCTION tasks. This gives us 40M examples for ORDER, 1.4M for CONJUNCTION, and 4.1M for NEXT. Despite having recently become a standard dataset for unsupervised learning, BookCorpus does not exhibit sufficiently rich discourse structure to allow our model to fully succeed—in particular, some of the conjunction categories are severely under-represented. Because of this, we choose to train our models on text from all three sources. While this precludes a strict apples-toapples comparison with other published results, our goal in extrinsic evaluation is simply to show that our method makes it possible to learn useful representations quickly, rather than to demonstrate the superiority of our learning technique given fixed data and unlimited time. We consider three sentence encoding models: a simple 1024D sum-of-Words (CBOW) encoding, a 1024D GRU recurrent neural network (Cho et al., 2014), and a 512D bidirectional GRU RNN (BiGRU). All three use FastText (Joulin et al., 2016) pre-trained word embeddings2 to which we apply a Highway transformation (Srivastava et al., 2015). The encoders are trained jointly with three bilinear classifiers for the three objectives (for the NEXT examples, the three context sentences are encoded separately and their representations are 2 https://github.com/facebookresearch/fastText/ blob/master/pretrained-vectors.md CONJUNCTION ORDER NEXT CBOW joint GRU joint BiGRU joint 42.8 39.5 45.1 56.6 54.3 58.3 27.7 25.9 30.2 BiGRU single 45.5 57.1 29.3 Table 4: Intrinsic evaluation results. Grant laughed and complied with the suggestion. Pauline stood for a moment in complete bewilderment. Her eyes narrowed on him, considering. Helena felt her face turn red hot. Her face remained expressionless as dough. Table 5: The nearest neighbors for a sentence. concatenated). We perform stochastic gradient descent with AdaGrad (Duchi et al., 2011), subsampling CONJUNCTION and NEXT by a factor of 4 and 6 respectively (chosen using held-out accuracy averaged over all three tasks on held out data after training on 1M examples). In this setting, the BiGRU model takes 8 hours to see all of the examples from the BookCorpus dataset at least once. For ease of comparison, we train all three models for exactly 8 hours. Intrinsic and Qualitative Evaluation Table 4 compares the performance of different training regimes along two axes: encoder architecture and whether we train one model per task or one joint model. As expected, the more complex bidirectional GRU architecture is required to capture the appropriate sentence properties, although CBOW still manages to beat the simple GRU (the slowest model), likely by virtue of its substantially faster speed, and correspondingly greater number of training epochs. Joint training does appear to be effective, as both the ORDER and NEXT tasks benefit from the information provided by CONJUNC TION . Early experiments on the external evaluation also show that the joint BiGRU model substantially outperforms each single model. Table 5 and the supplementary material show nearest neighbors in the trained BiGRU’s representation space for a random set of seed sentences. We select neighbors from among 400k held-out sentences. The encoder appears to be especially sensitive to high-level syntactic structure. Extrinsic Evaluation We evaluate the quality of the encoder learned by our system, which we call DiscSent, by using the sentence representations it produces in a variety of sentence classifi- Model Time MSRP TREC SUBJ FastSent FastSent+AE1 ≈13h 72.2 71.2 76.8 80.4 88.7 88.8 SDAE1 SDAE+embed1 192h 76.4 73.7 77.6 78.4 89.3 90.8 SkipT biGRU2 SkipT GRU2 SkipT+feats2 336h 71.2 73.0 75.8 89.4 91.4 92.2 92.5 92.1 93.6 Ordering model3 Ordering+embed3 +embed+SkipT3 48h 72.3 74.0 74.9 – – – – – – DiscSent biGRU DiscSent+unigram DiscSent+embed 8h 71.6 72.5 75.0 81.0 87.9 87.2 88.6 92.7 93.0 1 Table 6: Text classification results, including training time. +embed lines combine the sentence encoder output with the sum of the pretrained word embeddings for the sentence. +unigram lines do so using embeddings learned for each target task without pretraining. +feats varies by task. References: 1 Hill et al. (2016) 2 Kiros et al. (2015) 3 Logeswaran et al. (2016) cation tasks. We follow the settings of Kiros et al. (2015) on paraphrase detection (MSRP; Dolan et al., 2004), subjectivity evaluation (SUBJ; Pang and Lee, 2004) and question classification (TREC; Voorhees, 2001). Overall, our system performs comparably with the SDAE and Skip Thought approaches with a drastically shorter training time. Our system also compares favorably to the similar discourseinspired method of Logeswaran et al. (2016), achieving similar results on MSRP in a sixth of their training time. 5 Conclusion In this work, we introduce three new training objectives for unsupervised sentence representation learning inspired by the notion of discourse coherence, and use them to train a sentence representation system in competitive time, from 6 to over 40 times shorter than comparable methods, while obtaining comparable results on external evaluations tasks. We hope that the tasks that we introduce in this paper will prompt further research into discourse understanding with neural networks, as well as into strategies for unsupervised learning that will make it possible to use unlabeled data to train and refine a broader range of models for language understanding tasks. References Ondrej Bajgar, Rudolf Kadlec, and Jan Kleindienst. 2016. Embracing data abundance: BookTest dataset for reading comprehension. CoRR abs/1610.00956. Regina Barzilay and Mirella Lapata. 2008. Modeling local coherence: An entity-based approach. Computational Linguistics 34. Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov, Richard S. Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Skip-thought vectors. In NIPS 2015, Montreal, Quebec, Canada. Jiwei Li and Eduard H. Hovy. 2014. A model of coherence based on distributed sentence representation. In EMNLP 2014, Doha, Qatar. Steven Bird. 2006. NLTK: the natural language toolkit. In ACL 2006, Sydney, Australia. Jiwei Li and Dan Jurafsky. 2016. Neural net models for open-domain discourse coherence. CoRR abs/1606.01545. Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In EMNLP 2014, Doha, Qatar. Lajanugen Logeswaran, Honglak Lee, and Dragomir R. Radev. 2016. Sentence ordering using recurrent neural networks. CoRR abs/1611.02654. Andrew M. Dai and Quoc V. Le. 2015. Semisupervised sequence learning. In NIPS 2015, Montreal, Quebec, Canada. Tomas Mikolov, Martin Karafiát, Lukás Burget, Jan Cernocký, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model. In INTERSPEECH 2010, Makuhari, Chiba, Japan. Bill Dolan, Chris Quirk, and Chris Brockett. 2004. Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources. In COLING 2004, Geneva, Switzerland. Eleni Miltsakaki, Rashmi Prasad, Aravind K. Joshi, and Bonnie L. Webber. 2004. The penn discourse treebank. In LREC 2004, Lisbon, Portugal. John C. Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research 12. Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In ACL 2004, Barcelona, Spain.. pages 271–278. Zhe Gan, Yunchen Pu, Ricardo Henao, Chunyuan Li, Xiaodong He, and Lawrence Carin. 2016. Unsupervised learning of sentence representations using convolutional neural networks. CoRR abs/1611.07897. Prajit Ramachandran, Peter J. Liu, and Quoc V. Le. 2016. Unsupervised pretraining for sequence to sequence learning. CoRR abs/1611.02683. Felix Hill, Kyunghyun Cho, and Anna Korhonen. 2016. Learning distributed representations of sentences from unlabelled data. In NAACL 2016, San Diego, California, USA. Jerry R. Hobbs. 1979. Coherence and coreference. Cognitive Science 3. Yangfeng Ji, Trevor Cohn, Lingpeng Kong, Chris Dyer, and Jacob Eisenstein. 2015. Document context language models. CoRR abs/1511.03962. Yangfeng Ji, Gholamreza Haffari, and Jacob Eisenstein. 2016. A latent variable recurrent neural network for discourse relation language models. CoRR abs/1603.01913. Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2016. Bag of tricks for efficient text classification. CoRR abs/1607.01759. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In EMNLP 2014, Doha, Qatar. Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. 2016. Character-aware neural language models. In AAAI 2016, Phoenix, Arizona, USA. Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. In EMNLP 2015, Lisbon, Portugal. Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, Christopher Potts, et al. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP 2013, Seattle, Washington, USA. Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. 2015. Highway networks. CoRR abs/1505.00387. Bryan Stroube. 2003. Literary freedom: project gutenberg. ACM Crossroads 10. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In NIPS 2014, Montreal, Quebec, Canada. Ellen M. Voorhees. 2001. Overview of the TREC 2001 question answering track. In TREC 2001, Gaithersburg, Maryland, USA. Tian Wang and Kyunghyun Cho. 2016. Larger-context language modelling with recurrent neural network. In ACL 2016, Berlin, Germany. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alexander J. Smola, and Eduard H. Hovy. 2016. Hierarchical attention networks for document classification. In NAACL 2016, San Diego, California, USA. Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In NIPS 2015, Montreal, Quebec, Canada. Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In ICCV 2015, Santiago, Chile. Supplement to: Discourse-Based Objectives for Fast Unsupervised Sentence Representation Learning Table 7 lists the conjunction phrases and groupings used. Table 8 (next page) shows the Euclidean nearest neighbors of a sample of sentences in our representation space. addition again furthermore also moreover besides in addition finally further result accordingly consequently hence thus therefore specific namely specifically notably that is for example contrast anyway contrarily however conversely instead nonetheless nevertheless in contrast otherwise rather compare likewise similarly strengthen indeed in fact Table 7: Grouping of conjunctions. time meanwhile next then now thereafter return still recognize undoubtedly certainly His main influences are Al Di, Jimi Hendrix, Tony, JJ Cale, Malmsteen and Paul Gilbert. The album features guest appearances from Kendrick Lamar, Schoolboy Q, 2 Chainz, Drake, Big. The production had original live rock, blues, jazz, punk, and music composed and arranged by Steve and Diane Gioia. There are 6 real drivers in the game: Gilles, Richard Burns, Carlos Sainz, Philippe, Piero, and Tommi. Other rappers that did include Young Jeezy, Lil Wayne, Freddie Gibbs, Emilio Rojas, German rapper and Romeo Miller. Grant laughed and complied with the suggestion. Pauline stood for a moment in complete bewilderment. Her eyes narrowed on him, considering. Helena felt her face turn red hot. Her face remained expressionless as dough. Items can be selected specifically to represent characteristics that are not as well represented in natural language. Cache manifests can also use relative paths or even absolute urls as shown below. Locales can be used to translate into different languages, or variations of text, which are replaced by reference. Nouns can only be inflected for the possessive, in which case a prefix is added. Ratios are commonly used to compare banks, because most assets and liabilities of banks are constantly valued at market values. A group of generals thus created a secret organization, the united officers’ group, in order to oust Castillo from power. The home in Massachusetts is controlled by a private society organized for the purpose, with a board of fifteen trustees in charge. A group of ten trusted servants men from the family were assigned to search the eastern area of the island in the area. The city is divided into 144 administrative wards that are grouped into 15 boroughs. each of these wards elects a councillor. From 1993 to 1994 she served as US ambassador to the United Nations commission on the status of women. As a result of this performance, Morelli’s play had become a polarizing issue amongst Nittany Lion fans. In the end, Molly was deemed to have more potential, eliminating Jaclyn despite having a stellar portfolio. As a result of the Elway connection, Erickson spent time that year learning about the offense with Jack. As a result of the severe response of the czarist authorities to this insurrection, had to leave Poland. Another unwelcome note is struck by the needlessly aggressive board on the museum which has already been mentioned. Zayd Ibn reported , “we used to record the Quran from parchments in the presence of the messenger of god.” Daniel Pipes says that “primarily through “the protocols of the Elders of Zion”, the whites spread these charges to [. . . ]” Sam wrote in “” (1971) that Howard’s fiction was “a kind of wild West in the lands of unbridled fantasy.” said , the chancellor “elaborately fought for an European solution” in the refugee crisis, but this was “out of sight”. Robert , writing for “The New York Post”, states that, “in Mellie , the show has its most character [. . . ]” Many “Crimean Goths” were Greek speakers and many Byzantine citizens were settled in the region called [. . . ] The personal name of “Andes”, popular among the Illyrians of southern Pannonia and much of Northern Dalmatia [. . . ] is identified by the Chicano as the first settlement of the people in North America before their Southern migration [. . . ] The range of “H.” stretches across the Northern and Western North America as well as across Europe [. . . ] The name “Dauphin river” actually refers to two closely tied communities; bay and some members of Dauphin river first nation. She smiled and he smiled in return. He shook his head and smiled broadly. He laughed and shook his head. He gazed at her in amazement. She sighed and shook her head at her foolishness. The jury returned a verdict of not in the Floyd cox case, in which he was released immediately. The match lasted only 1 minute and 5 seconds, and was the second quickest bout of the division. His results qualified him for the Grand Prix final, in which he placed 6th overall. The judge stated that the prosecution had until march 1, 2012, to file charges. In November, he reached the final of the Ruhr Open, but lost 42̆0130 against Murphy. Here was at least a slight reprieve. The monsters seemed to be frozen in time. This had an impact on him. That was all the sign he needed. So this was disturbing as hell. Table 8: Nearest neighbor examples
9cs.NE
FINITE ELEMENT MODEL UPDATING USING RESPONSE SURFACE METHOD Tshilidzi Marwala School of Electrical and Information Engineering University of the Witwatersrand P/Bag 3, Wits, 2050, South Africa [email protected] This paper proposes the response surface method for finite element model updating. The response surface method is implemented by approximating the finite element model surface response equation by a multi-layer perceptron. The updated parameters of the finite element model were calculated using genetic algorithm by optimizing the surface response equation. The proposed method was compared to the existing methods that use simulated annealing or genetic algorithm together with a full finite element model for finite element model updating. The proposed method was tested on an unsymmetrical H-shaped structure. It was observed that the proposed method gave the updated natural frequencies and mode shapes that were of the same order of accuracy as those given by simulated annealing and genetic algorithm. Furthermore, it was observed that the response surface method achieved these results at a computational speed that was more than 2.5 times as fast as the genetic algorithm and a full finite element model and 24 times faster than the simulated annealing. proximation model, which traditionally is a polynomial and therefore is less expensive to evaluate. This makes RSM very useful to FE model updating because optimizing the FE model to match measured data to FE model generated data is a computationally expensive exercise. Furthermore, the calculation of the gradients that are essential when traditional optimization methods, such as conjugate gradient methods, are used is computationally expensive and often encounters numerical problems such as ill-conditioning. RSM tends to be immune to such problems when used for FE model updating. This is largely because RSM solves a crude approximation of the FE model rather than the full FE model which is of high dimensional order. The multi-layer perceptron (MLP)6 is used to approximate the response equation. The RSM is particularly useful for optimizing systems that are evolving as a function of time, a situation that is prevalent in model-based fault diagnostics found in the manufacturing sector. To date, RSM has been used extensively to optimize complex models and processes7,8. In summary, the RSM is used because of the following reasons: (1) the ease of implementation that includes low computational time; (2) the suitability of the approach to the manufacturing sector where model-based methods are often used to monitor structures that evolve as a function of time. FE model updating has been used widely to detect damage in structures9. When implementing FE updating methods for damage identification, it is assumed that the FE model is a true dynamic representation of the structure and this is achieved through FE model updating. This means that changing any physical parameter of an element in the FE model is equivalent to introducing damage in that region. There are two approaches that are used in FE updating: direct methods and iterative Introduction Finite element (FE) models are widely used to predict the dynamic characteristics of aerospace structures. These models often give results that differ from the measured results and therefore need to be updated to match the measured data. FE model updating entails tuning the model so that it can better reflect the measured data from the physical structure being modeled1. One fundamental characteristic of an FE model is that it can never be a true reflection of the physical structure but it will forever be an approximation. FE model updating fundamentally implies that we are identifying a better approximation model of the physical structure than the original model. The aim of this paper is to introduce updating of finite element models using Response Surface Method (RSM)2. Thus far, the RSM method has not been used to solve the FE updating problem1. This new approach to FE model updating is compared to methods that use simulated annealing (SA) or genetic algorithm (GA) together with full FE models for FE model updating. FE model updating methods have been implemented using different types of optimization methods such as genetic algorithm and conjugate gradient methods3-5. Levin and Lieven5 proposed the use of SA and GA for FE updating. RSM is an approximate optimization method that looks at various design variables and their responses and identify the combination of design variables that give the best response. The best response, in this paper, is defined as the one that gives the minimum distance between the measured data and the data predicted by the FE model. RSM attempts to replace implicit functions of the original design optimization problem with an ap* Associate Professor 1 methods1. Direct methods, which use the modal properties, are computationally efficient to implement and reproduce the measured modal data exactly. Furthermore, they do not take into account the physical parameters that are updated. Consequently, even though the FE model is able to predict measured quantities, the updated model is limited in the following ways: it may lack the connectivity of nodes - connectivity of nodes is a phenomenon that occurs naturally in finite element modeling because of the physical reality that the structure is connected; the updated matrices are populated instead of banded - the fact that structural elements are only connected to their neighbors ensures that the mass and stiffness matrices are diagonally dominated with few couplings between elements that are far apart; and there is a possible loss of symmetry of the systems matrices. Iterative procedures use changes in physical parameters to update FE models and produce models that are physically realistic. Iterative methods that use modal properties and the RSM for FE model updating are implemented in this paper. The FE models are updated so that the measured modal properties match the FE model predicted modal properties. The proposed RSM updating method is tested on an unsymmetrical H-shaped structure. and stiffness matrices. The frequency response functions (FRFs) are defined as the ratio of the Fourier transformed response to the Fourier transformed force. The FRFs may be expressed in receptance and inertance form. On the one hand, receptance expression of the FRF is defined as the ratio of the Fourier transformed displacement to the Fourier transformed force. On the other hand, inertance expression of the FRF is defined as the ratio of the Fourier transformed acceleration to the Fourier transformed force. The inertance FRF (H) may be written in terms of the modal properties by using the modal summation equation as follows10: N − ω 2φkiφli H kl ( ω ) = ∑ (3) 2 2 i =1 − ω + 2ζ i ω i ωj + ω i Equation 3 is an FRF due to excitation at position k and response measurement at position l, ω is the frequency point, ω i is the ith natural frequency, N is the number of modes and ζi is the damping ratio of mode i. The excitation and response of the structure and Fourier transform method10 can be used to calculate the FRFs. Through equation 3 and a technique called modal analysis10, the natural frequencies and mode shapes can be indirectly calculated from the measured FRFs. The modal properties of a dynamic system depend on the mass and stiffness matrices of the system as indicated by equation 2. Therefore, the measured modal properties can be reproduced by the FE model if the correct mass and stiffness matrices are identified. FE model updating is achieved by identifying the correct mass and stiffness matrices. The correct mass and stiffness matrices, in the light of the measured data, can be obtained by identifying the correct moduli of elasticity for various sections of the structure under consideration1. In this paper, to correctly identify the moduli of elasticity of the structure, the following cost function that measures the distance between measured data and FE model calculated data, is minimized: Mathematical Background In this study, modal properties, i.e. natural frequencies and mode shapes, are used as a basis for FE model updating. For this reason these parameters are described in this section. Modal properties are related to the physical properties of the structure. All elastic structures may be described in terms of their distributed mass, damping and stiffness matrices in the time domain through the following expression10: [ M ]{ X ' ' } + [ C ]{ X ' } + [ K ]}{ X } = { F } (1) where [M], [C] and [K] are the mass, damping and stiffness matrices respectively, and {X}, {X′} and {X′′} are the displacement, velocity and acceleration vectors respectively while {F} is the applied force vector. If equation 1 is transformed into the modal domain to form an eigenvalue equation for the ith mode, then10:  ω m − ω calc E = ∑ γ i  i m i ωi i =i  N N ( 2   ...   + β ∑ 1 − diag( MAC({ φ }icalc ,{ φ }im )) (4) ) i 2 ( −ω i [ M ] + jω i [ C ] + [ K ]){ φ }i = { 0 } (2) th where j = − 1 , ω i is the i complex eigenvalue, with its imaginary part corresponding to the natural frequency ωi, { 0 } is the null vector and { φ }i is the ith complex mode shape vector with the real part corresponding to the normalized mode shape {φ}i. From equation 2, it may be deduced that the changes in the mass and stiffness matrices cause changes in the modal properties of the structure. Therefore, the modal properties can be identified through the identification of the correct mass Here m is for measured, calc is for calculated, N is the number of modes; γ i is the weighting factor that measures the relative distance between the initial estimated natural frequencies for mode i and the target frequency of the same mode; the parameter β is the weighting function on the mode shapes; the MAC is the modal assurance criterion11; and the diag(MAC)i stands for the ith diagonal element of the MAC matrix. The MAC is a measure of the correlation between two sets of mode shapes of the same dimension. In equation 4 the first 2 part has a function of ensuring that the natural frequencies predicted by the FE model are as close to the measured ones as possible while the second term ensures that the mode shapes between measurements and those predicted by the FE model are correlated. When two sets of mode shapes are perfectly correlated then the MAC matrix is an identity matrix. The updated model is evaluated by comparing the natural frequencies and mode shapes from the FE models before and after updating to the measured ones. Initial Conditions Updating Parameters Updating Objective Updating Space Generation of surface response data using the FE model Response Surface Method Functional approximation RSM method is a procedure that operates by generating a response for a given input. The inputs are the parameters to be updated and the response is the error between the measured data and the FE model generated data. Then an approximation model of the input parameters and the response, called a response surface equation, is constructed. As a consequence of this, the optimization method operates on the surface response. This equation is usually simple and not computationally intensive as opposed to a full FE model. RSM has other advantages such as the ease of implementation through parallel computation and the ease at which parameter sensitivity can be calculated. The proposed RSM consists of these essential components: (1) the response surface approximation equation; and (2) the optimization procedure. There are many techniques that have been used for response surface approximation such as polynomial approximation12 and neural networks13. A multi-layer perceptron is used as a response surface approximation equation6. Further understanding of different approaches to response surface approximation may be found in the literature14-19. In this paper, MLP is used because it has been successfully used to solve complicated regression problems. The details of the MLP are described in the next section. The second component of the RSM is the optimization of the response surface. There are many types of optimization methods that can be used to optimize the response surface equation and these include the gradient based methods20 and evolutionary computation methods21. The gradient based methods have a shortcoming of identifying local optimum solutions while evolutionary computing methods are better able to identify global optimum solution. As a result of the global optimum advantage of evolutionary methods, in this study the GA is used to optimize the response surface equation. The manner in which the RSM is implemented is shown in Figure 1. In this figure it shown that the RSM is implemented by following these steps: 1) Setting initial conditions which are: updating parameters, updating objective, which is in equation 4, Global optimization Functional evaluation at the optimum solution on the FE model Updating criteria satisfied? N Replace the worst surface response data with the optimum point and its FE response Y Stop Figure 1. The flowchart of the RSM. Here N stands for no and Y stands for yes. and updating space. 2) The FE model is then used to generate sample response surface data 3) MLP is used to approximate the response surface approximation equation from the data generated in Step 2. 4) GA is used to find a global optimum solution. 5) The new optimum solution is used to evaluate the response from the full FE model. 6) If the optimum solution does not satisfy the objective, then the new optimum and the corresponding FE model calculated response replaces the candidate with the worst response in data set generated in Step 2 and then steps 3 to 5 are repeated. If the objective is satisfied then stop and the optimum solution becomes the ultimate solution. Step 6 ensures that the simulation always operates in the region of the optimum solution. The next section 3 sic dimensionality of the data. Models of this form can approximate any continuous function to arbitrary accuracy if the number of hidden units M is sufficiently large. The relationship between the output y, representing error between the model and measured data, and input, x, representing updating parameters may be written as follows6: describes an MLP, which is used for functional approximation. Multi-layer Perceptron Multi-layer perceptron is a type of neural networks which used in the present study. This section gives the over-view of the MLP in the context of functional approximation. The MLP is viewed in this paper as parameterized graphs that make probabilistic assumptions about data. Learning algorithms are viewed as methods for finding parameter values that look probable in the light of the data. Supervised learning is used to identify the mapping function between the updating parameters (x) and the response (y). The response is calculated using equation 4. The reason why the MLP is used is because it provides a distributed representation with respect to the input space due to cross-coupling between input, hidden and output layers. The MLP architecture contains a hyperbolic tangent basis function in the hidden units and linear basis functions in the output units6. A schematic illustration of the MLP is shown in Figure 2. This network architecture contains hidden units and output units and has one hidden layer. The bias parameters in the first layer are shown as weights from an extra M   d  y k =  ∑ wkj( 2 ) tanh ∑ w (ji1 ) x i + w (j 01 )  + wk( 02 )  (5)  i =1   j =1  Here, w (ji1 ) and w(ji2 ) indicate weights in the first and second layers, respectively, going from input i to hidden unit j, M is the number of input units, d is the number of output units while w (j 01 ) indicates the bias for the hidden unit j. Training the neural network identifies the weights in equations 5 and a cost function must be chosen to identify these weights. A cost function is a mathematical representation of the overall objective of the problem. The main objective, this is used to construct a cost function, is to identify a set of neural network weights given updating parameters and the error between the FE model and the measured data. If the training set D = { x k , t k }kN=1 is used and assuming that the targets t are sampled independently given the inputs xk and the weight parameters, wkj, the cost function, E, may be written as follows using the sum-of-square error func6 tion : Output Units N zM The sum-of-square error function is chosen because it 6 has been found to be suited to regression problems . In equation 6, N is the number of training examples and K is the number of output units. In this paper, N is equal to 150, while K is equal to 1. Before the MLP is trained, the network architecture needs to be constructed by choosing the number of hidden units, M. If M is too small, the MLP will be insufficiently flexible and will give poor generalization of the data because of high bias. However, if M is too large, the neural network will be unnecessarily flexible and will give poor generalization due to a phenomenon known as over-fitting caused by high variance. In this study, we choose M such that the number of weights is at most fewer than the number of response data. This is in line with the basic mathematical principle which states that in order to solve a set of equations with n variables you need at least n independent data points. The next section describes the GA, which is a method that is used to solve for the optimum solution of the response surface approximation equation. Hidden Units bias xd (6) n =1 k =1 z0 z1 K E = ∑ ∑ { t nk − y nk } 2 yc y1 x1 x0 Input Units Figure 2. Feed-forward network having two layers of adaptive weights input having a fixed value of x0=1. The bias parameters in the second layer are shown as weights from an extra hidden unit, with the activation fixed at z0=1. The model in Figure 2 is able to take into account the intrinsic di- Genetic Algorithms GA was inspired by Darwin’s theory of natural evolution. Genetic algorithm is a simulation of natural evolu- 4 tion where the law of the survival of the fittest is applied to a population of individuals. This natural optimization method is used to optimize either the response surface approximation equation or the error between the FE model and the measured data. GA is implemented by generating a population and creating a new population by performing the following procedures: (1) crossover; (2) mutation; (3) and reproduction. The details of these procedures can be found in Holland21 and Goldberg22. The crossover operator mixes genetic information in the population by cutting pairs of chromosomes at random points along their length and exchanging over the cut sections. This has a potential of joining successful operators together. Arithmetic crossover technique22 is used in this paper. Arithmetic crossover takes two parents and performs an interpolation along the line formed by the two parents. For example if two parents p1 and p2 undergo crossover, then a random number a which lies in the interval [0,1] is generated and the new offsprings formed are p1(a-1) and pa. The next section describes simulated annealing which is used to update an FE model using a FE model. Simulated Annealing Simulated Annealing is a Monte Carlo method that is used to investigate the equations of state and frozen states of n degrees of freedom system23. SA was inspired by the process of annealing where objects, such as metals, re-crystallize or liquids freeze. In the annealing process the object is heated until it is molten, then it is slowly cooled down such that the metal at any given time is approximately in thermodynamic equilibrium. As the temperature of the object is lowered, the system becomes more ordered and approaches a frozen state at T=0. If the cooling process is conducted insufficiently or the initial temperature of the object is not sufficiently high, the system may become quenched forming defects or freezing out in metastable states. This indicates that the system is trapped in a local minimum energy state. The process that is followed to simulate the annealing process was proposed by Metropolis et al.24 and it involves choosing the initial state with energy Eold (see equation 4) and temperature T and holding T constant and perturbing the initial configuration and computing Enew at the new state. If Enew is lower than Eold, then accept the new state, otherwise if the opposite is the case then accept this state with a probability of exp -(dE/T) where dE is the change in energy. This process can be mathematically represented as follows: Mutation is a process that introduces to a population, new information. Non-uniform mutation22 was used and it changes one of the parameters of the parent based on a non-uniform probability distribution. The Gaussian distribution starts with a high variance and narrows to a point distribution as the current generation approaches the maximum generation. Reproduction takes successful chromosomes and reproduces them in accordance to their fitness functions. In this study normalized geometric selection method was used22. This method is a ranking selection function which is based on the normalized geometric distribution. Using this method the least fit members of the population are gradually driven out of the population. The basic genetic algorithm was implemented in this paper as follows: if E new < E old accept state E new  E − E old  (7)  else accept E new with probability exp new T   This processes is repeated such that the sampling statistics for the current temperature is adequate, and then the temperature is decreased and the process is repeated until a frozen state where T=0 is achieved. SA was first applied to optimization problems by Kirkpatrick, et al. 23. The current state is the current updating solution, the energy equation is the objective function in equation 4, and the ground state is the global optimum solution. 1) Randomly create an initial population of a certain size. 2) Evaluate all of the individuals in the population using the objective function in equation 4. 3) Use the normalized geometric selection method to select a new population from the old population based on the fitness of the individuals as given by the objective function. Example: Asymmetrical H-structure An unsymmetrical H-shaped aluminum structure shown in Figure 3 was used to validate the proposed method. This structure was also used by Marwala and Heyns4 as well as Marwala25. This structure had three thin cuts of 1mm that went half-way through the crosssection of the beam. These cuts were introduced to elements 3, 4 and 5. The structure with these cuts was used so that the initial FE model gives data that are far from the measured data and, thereby test the proposed proce- 4) Apply some genetic operators, non-uniform mutation and arithmetic crossover, to members of the population to create new solutions. 5) Repeat steps 2-6, which is termed one generation, until a certain fixed number of generations has been achieved 5 dure on a difficult FE model updating problem. The structure was suspended using elastic rubber bands. The structure was excited using an electromagnetic shaker and the response was measured using an accelerometer. The structure was divided into 12 elements. It was excited at a position indicated by double-arrows, in Figure 3, and acceleration was measured at 15 positions indicated by single-arrows in Figure 3. The structure was tested freely-suspended, and a set of 15 frequency response functions were calculated. A roving accelerometer was used for the testing. The mass of the accelerometer was found to be negligible compared to the mass was run 150 times to generate the data for functional approximation. The MLP implemented had 12 input variables corresponding to the 12 elements in the FE model, 8 hidden units and one output unit corresponding to the error in equation 4. As described before, the MLP had a hyperbolic tangent activation function in the hidden layer and linear activation function in the output layer. The RSM functional approximation via the MLP was evaluated 10 times (iterations) each time using the GA to calculate the optimum point and evaluating this optimum point on the FE model and then storing the previous optimum point in the data set for the current A 200mm 400mm B 600mm y Cross-section indicated by line AB 9.8mm x 32.2mm Figure 3. Irregular H-shaped structure of the structure. The number of measured coordinates functional approximation. The scaled conjugate gradiwas 15. ent method was used to train the MLP, primarily beThereafter, the finite element model was constructed cause of its computational efficiency27. The initial funcusing the Structural Dynamics Toolbox26. The FE tional approximation was obtained by training the MLP model used Euler-Bernoulli beam elements. The FE for 150 training cycles and on a subsequent functional model contained 12 elements. The moduli of elasticity of these elements were used as updating parameTable 1. Results showing measured frequencies, the initial freters. When the FE updating was implemented the quencies and the frequencies obtained when the FE model is moduli of elasticity was restricted to vary from updated using the RSM, SA and GA 10 10 -2 6.00x10 to 8.00x10 N.m . The weighting factors, Measured Initial Frequencies Frequencies Frequencies in the first term in equation 4, were calculated for Freq Freq from RSM from SA from GA each mode as the square of the error between the (Hz) (Hz) Updated Updated Updated measured natural frequency and the natural frequency Model (Hz) Model (Hz) Model (Hz) calculated from the initial model and the weighting 53.9 56.2 52.2 54.0 53.9 function for the second term in equation 4 was set to 117.3 127.1 118.4 118.8 120.1 0.75. When the RSM, SA and GA were implemented 228.4 209.4 209.7 211.3 for model updating the results shown in Table 1 were 208.4 254.0 263.4 251.1 253.8 253.4 obtained. 445.1 452.4 432.7 435.8 438.6 On implementing the proposed RSM, the FE model 6 and that from the initial model was 8.4%. When the RSM was used, this error was reduced to 0.9% while using SA it was reduced to 1.3% and using the GA it was reduced to 2.4%. The error of the third natural frequencies between the measured data and the initial FE model was 9.6%. When the RSM was used, this error was reduced to 0.5% while using SA reduced it to 0.6% and using the GA and a full FE model reduced it to 1.4%. The error between the fourth measured natural frequency and that from the initial model was 3.7%. When the RSM was used for FE updating, this error was reduced to 1.1% while using the SA reduced it to 0.1% and using the GA and a full FE model reduced it to 0.2%. The error between the fifth measured natural frequency and that from the initial model was 1.6%. When the RSM was used, this error was increased to 2.8% while using SA increased it to 2.1% and using the GA and a full FE model the error was reduced to 1.5%. Overall, the SA gave the best results with an average error, calculated over all the five natural frequencies, of 0.9% followed by the GA with an average error of 1.1% and then RSM with an average error of 1.7%. All the three methods on average improved when compared to the average error between the initial FE model and the measured data, which was 5.5%. The updated FE models implemented were 9 also validated on the mode shapes they pre8 dicted. To make this assessment possible the 7 MAC11 was used. The mean of the diagonal of 6 the MAC vector was used to compare the mode 5 shapes predicted by the updated and initial FE models to the measured mode shapes. The av4 erage MAC calculated between the mode shapes 3 from an initial FE model and the measured 2 mode shapes was 0.8394. When the average 1 MAC was calculated between the measured data 0 and data obtained from the updated FE models, it was observed that the RSM, SA and GA up1 2 3 4 5 6 7 8 9 10 11 12 dated FE model gave the improved average of Element the diagonal of the MAC matrix of 0.8413, Initial RSM GA SA 0.8430 and 0.8419, respectively. Therefore, the Figure 4. Graph showing the initial moduli of elasticity and the SA gave the best MAC followed by the GA moduli of elasticity obtained when the FE model is updated using which was followed by the RSM. However, the RSM, GA and SA. Here e10 indicates 10 to the power 10 and these differences in accuracies of the MAC and the units are Nm-2 natural frequencies were not significant. The computational time taken to run the complete RSM method was 46 CPU seconds, while measured natural frequency and that from the initial FE the SA and a full FE model took 19 CPU minutes to run model, which was obtained when the modulus of elasticand the GA and a full FE model took 117 CPU seconds. 10 -2 ity of 7.00x10 N.m was assumed, was 4.3%. When The RSM was found to be faster than the GA which was the RSM was used for FE updating, this error was rein turn much faster than the SA which was faster that the duced to 3.1% while using SA it was reduced to 0.2% GA. On implementing the RSM, 160 FE model evaluaand using the GA approach it was reduced to 0%. The tions were made, while on implementing the SA 19485 error between the second measured natural frequency FE model evaluations were made and on implementing Modulus of Elasticity (e10) approximation, where the data set had the previous optimum solution added to it, used 5 training cycles. On using the RSM, the MLP was only initialized once. The GA was implemented on a population size of 50 and 200 generations. The normalized geometric distribution was implemented with a probability of selecting the best candidate set to 8%, mutation rate of 0.3% and crossover rate of 60%. When SA and a full FE model was implemented for FE updating, the scale of the cooling schedule was set to 4 and the number of individual annealing runs was set to 3. When the simulation was run, the first run involved 7008 FE model calculations, in the second run 6546 FE model calculations and in the third run 5931 FE model calculations were made. On implementing the GA and a full FE model, the same options as those that were used in the implementation of the RSM were used. The results showing the moduli of elasticity of the initial FE model, RSM updated FE model, SA updated FE model and GA updated FE model are shown in Figure 4. Table 1 shows the measured natural frequencies, initial natural frequencies and natural frequencies obtained by the RSM, SA and GA updated FE models. The error between the first 7 the GA 10000 FE model calculations were made. In this paper, a simple FE model with 39 degrees of freedom is updated. It can, therefore, be concluded that if the FE model had several thousand degrees of freedom, the RSM will be substantially faster than the other methods. This conclusion should be understood in the light of the fact that FE models usually have many degrees of freedom. changes in their vibration characteristics: a literature review” Los Alamos National Laboratory Report LA13070-MS, 1996. 10 Ewins, D.J., Modal testing: theory and practice, Research Studies Press, Letchworth, U.K, 1995. 11 Allemang, R.J. and Brown, D.L., “A correlation coefficient for modal vector analysis” Proceedings of the 1st International Modal Analysis Conference, 1-18, 1982. 12 Sacks, J., Welch, W.J., Mitchell, T.J., and Wynn, H.P., “Design and analysis of computer experiments” Statistical Science, 4(4), 1989, pp. 409-435. 13 Varaajan, S., Chen, W., and Pelka, C.J., “Robust concept exploration of propulsion systems with enhanced model approximation” Engineering Optimization, 32(3), 2000, pp. 309-334. 14 Giunta, A. A., and Watson, L. T., “A Comparison of Approximation Modeling Techniques: Polynomial Versus Interpolating Models,” AIAA-98-4758, American Institute of Aeronautics and Astronautics, Inc., 1998, pp.392-401. 15 Koch, P. N., Simpson, T. W., Allen, J. K., and Mistree, F., 1999, “Statistical Approximations for Multidisciplinary Design Optimization: The Problem of Size,” Journal of Aircraft, 36(1), 1999, pp. 275-286. 16 Jin, R., Chen, W., and Simpson, T., 2000, “Comparative Studies of Metamodeling Techniques under Multiple Modeling Criteria,” 8th AIAA/NASA/USAF/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Long Beach, CA, September 6-8, 2000. 17 Lin, Y., Krishnapur, K., Allen, J. K., and Mistree, F., 2000, “Robust Concept Exploration in Engineering Design: Metamodeling Techniques and Goal Formulations,” Proceedings of the 2000 ASME Design Engineering Technical Conferences, DETC2000/DAC14283, September 10-14, 2000, Baltimore, Maryland. 18 Wang, G.G., 2003, “Adaptive response surface method using inherited Latin hypercube design points,” Transactions of the ASME, Journal of Mechanical Design,125, pp. 210-220. 19 Simpson, T. W., Peplinski, J. D., Koch, P. N., and Allen, J. K., “Metamodels for Computer-based Engineering Design: Survey and Recommendations,” Engineering with Computers, 17, 2001, pp. 129-150. 20 Fletcher, R., Practical Methods of Optimization. 2nd edition, New York: Wiley, 1987. 21 Holland, J, Adaptation in Natural and Artificial Systems, Ann Arbor: University of Michigan Press, 1975. 22 Goldberg, D.E., Genetic algorithms in search, optimization and machine learning, Addison-Wesley, Reading, MA, 1989. 23 Metropolis, N, Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., and Teller, E., “Equations of state Conclusion In this study, RSM is proposed for FE model updating. The proposed RSM was implemented within the framework of the MLP for functional approximation and GA for optimization of the MLP response surface function. This procedure was compared to the GA and SA. When these techniques were tested on the unsymmetrical H-shaped structure, it was observed that the RSM was faster than the SA and GA without much compromise on the accuracy of the predicted modal properties. Acknowledgment The author would like to thank Stefan Heyns, the now National Research Foundation as well as the AECI, Ltd for their assistance in this work. References 1 Friswell, M.I., and Mottershead, J.E., Finite element model updating in structural dynamics, Kluwer Academic Publishers Group, Norwell, MA, 1995, pp. 1-286. 2 Montgomery, D.C., Design and analysis of experiments, 4th Edition, John Wiley and Sons, NY, 1995, Chapter 14. 3 Marwala, T., “Finite element model updating using wavelet data and genetic algorithm” Journal of Aircraft, 39(4), 2002, pp. 709-711. 4 Marwala, T., and Heyns, P.S., “A multiple criterion method for detecting damage on structures” AIAA Journal, 195(2), 1998, pp. 1494-1501. 5 Levin, R.I. and Lieven, N.A.J, “Dynamic finite element model updating using simulated annealing and genetic algorithms” Mechanical Systems and Signal Processing, 12(1), pp. 91-120. 6 Bishop, C.M., Neural Networks for Pattern Recognition. Oxford: Clarendon, 1996. 7 Lee, S.H., Kim, H.Y., and Oh, S.I., “Cylindrical tube optimization using response surface method based on stochastic process” Journal of Materials Processing Technology, 130-131(20), 2002, pp. 490-496. 8 Edwards, I.M., and Jutan, A., “Optimization and control using response surface methods” Computers & Chemical Engineering, 21(4), 1997, pp. 441-453. 9 Doebling, S.W., Farrar, C.R., Prime, M.B., and Shevitz, D.W., “Damage identification and health monitoring of structural and mechanical systems from 8 26 calculations by fast computing machines,” Journal of Chemical Physics, 21, 1953, pp. 1087-1092. 24 Kirkpatrick, S., Gelatt, C.D., and Vecchi, M.P., “Optimization by simulated annealing,” Science, 220, 1983, pp. 671-680. 25 Marwala, T, A multiple criterion updating method for damage detection on structures. University of Pretoria Masters Thesis, 1997. Balmès, E., Structural Dynamics Toolbox User’s Manual Version 2.1, Scientific Software Group, Sèvres, France, 1997. 27 Møller, M. “A scaled conjugate gradient algorithm for fast supervised learning” Neural Networks, vol. 6, 1993, pp. 525-533. 9
5cs.CE
arXiv:1605.04059v1 [math.ST] 13 May 2016 The lq consistency of the Dantzig Selector for Cox’s Proportional Hazards Model Kou Fujimori and Yoichi Nishiyama Waseda University Abstract The Dantzig selector for the proportional hazards model proposed by D.R. Cox is studied in a high-dimensional and sparse setting. We prove the lq consistency for all q ≥ 1 of some estimators based on the compatibility factor, the weak cone invertibility factor, and the restricted eigenvalue for certain deterministic matrix which approximates the Hessian matrix of log partial likelihood. Our matrix conditions for these three factors are weaker than those of previous researches. 1 Introduction The proportional hazards model proposed by Cox (1972) is widely used in survival analysis. Using the counting process approach, Andersen and Gill (1982) proved that the maximum partial likelihood estimator satisfies the consistency and the asymptotic normality when the number of covariates p is fixed. In this paper, we deal with a high-dimensional and sparse setting, i.e., the case where p is much larger than n and the number S of non-zero components of the true parameter is relatively small. In this setting, the maximum partial likelihood estimator may not behave well, so we need to consider other estimation procedures. For example, the penalized methods such as LASSO (Tibshirani (1997), Huang et al. (2013), Bradic et al. (2013), among others) and SCAD (Fan and Li (2002) and Bradic et al. (2011)) are studied by many researchers. Instead of the penalized methods, in this paper we apply a relatively new method called the Dantzig selector for the proportional hazards model. The Dantzig selector (DS) was proposed by Candés and Tao (2007) in a high-dimensional and sparse setting for the linear regression model: Y = Zβ + ǫ. The Dantzig selector returns the estimator β̂ for β defined as follows: β̂ := arg min kβk1 , β∈C C := {β ∈ Rp : |Z j T (Y − Zβ)| ≤ λ}, 1 where k · k1 is l1 -norm, Z j denotes the j-th column of the design matrix Z and λ ≥ 0 is a suitable constant. Assume that ǫ is a Gaussian error term for simplicity. When λ = 0, the DS returns to the MLE. For λ > 0, the DS searches for the sparsest β within the given distance of the MLE. Notice that this method has a good potential to be applied for other models. Antoniadis et al. (2010) proposed the Survival Dantzig selector (SDS), which is an application of the DS for the proportional hazards model, and studied the l2 consistency of the SDS under the UUP condition which is the matrix condition used by Candés and Tao (2007). The purpose of this paper is to prove the lq consistency for q ≥ 1 under some weaker matrix conditions. Our way of proofs are similar to those of Huang et al. (2013) for LASSO in the proportional hazards model. They introduced the compatibility factor, the weak invertibility factor, and the restricted eigenvalue for Hessian matrix of log partial likelihood. Since the Hessian is a random matrix, these factors are random variables. They also derived some conditions to treat them as deterministic constants. In contrast, we will define these factors for a deterministic matrix, which approximates the Hessian matrix to clarify the proofs. This paper is organized as follows. The settings for the proportional hazards model and an estimation procedure are given in Section 2. In Section 3, we introduce some matrix conditions to derive the consistency. Main results are presented in Section 4. Throughout this paper, for every q ∈ [1, ∞], we denote by k · kq the lq -norm of Rp -vector, which is defined as follows:  kvkq =  p X j=1  q1 |vj |q  , q < ∞; kvk∞ = sup |vj |. 1≤j≤p Moreover, for a given m × n matrix A, where m, n ∈ N, we define kAk∞ by kAk∞ := sup sup |Aji |, 1≤i≤m 1≤j≤n where Aji denotes the (i, j)-component of the matrix A. 2 2.1 Model set up and estimation procedure Model set up Let T be a survival time, and C a censoring time, which are positive real-valued random variables on a probability space (Ω, F , P ). The survival time T is assumed to be conditionally independent of a censoring time C given the multidimensional covariate Z. For every n ∈ N, we observe the data {(Xi , Di , Zi )}ni=1 , which are i.i.d. copies of (X, D, Z), where D = 1{T ≤C} . Assume that the dimension of covariates p = pn depends on n. In particular, we assume that 2 pn ≫ n. Let Ni (t) = 1{Xi ≤t, Di =1} be counting processes based on these data, and Yi (t) = 1{Xi ≥t} at risk processes for every i = 1, 2, . . . , n. Assume that the sample paths t ❀ Ni (t), i = 1, 2, . . . , n, have no simultaneous jumps. We introduce the filtration {Ft }t≥0 defined by Ft := σ{Ni (u), Yi (u), 0 ≤ u ≤ t, Zi , i = 1, 2, . . . , n}. Assume that Ni has the intensity λi of the form λi (t, Zi ) = Yi (t)α0 (t) exp(ZiT β0 ), where α0 is a baseline hazard function which is regarded as a nuisance parameter, and β0 ∈ Rpn is the unknown true parameter. We are interested in the estimation problem of β0 in a high dimensional and sparse setting, i.e., in the case where β0 ∈ Rpn has S non-zero components. Using the Doob-Meyer decomposition theorem, we have that the processes Z t Mi (t) := Ni (t) − λi (u, Zi )du, i = 1, 2, . . . , n, 0 are independent {Ft }t≥0 square-integrable martingales. Their predictable quadratic variations are given by Z t hMi , Mi i(t) = λi (u, Zi )du. 0 We fix the time interval [0, τ ]. Following Cox (1972), we introduce the log partial likelihood defined by Cn (β) := n Z X i=1 where Sn0 (β, u) := n P i=1 0 τ {ZiT β − log(Sn0 (β, u))}dNi (u), Yi (u) exp(ZiT β). Furthermore, we define the pn dimen- sional vector Un (β), and the pn × pn matrix Jn (β) as   n Z 1X τ Sn1 ˙ Un (β) := ln (β) = Zi − 0 (β, u) dNi (u), n i=1 0 Sn " #  1 ⊗2 Z Sn 1 τ Sn2 (β, u) − (β, u) dN̄ (u), Jn (β) := −l̈n (β) = n 0 Sn0 Sn0 where N̄ (u) := n P i=1 and Sn2 (β, u) := Ni (u), ln (β) := Cn (β)/n, Sn1 (β, u) := n P i=1 n P i=1 Yi (u) exp(ZiT β)Zi , Yi (u) exp(ZiT β)Zi⊗2 . Note that Jn (β0 ) is a nonnegative definite matrix. In particular, for the true parameter β = β0 and for all j = 3 1, 2, . . . , pn , we can see that Unj (β0 ), j = 1, 2, . . . , pn are the terminal values of martingales {Unj (β0 , t)}t∈[0,τ ] given by: n Unj (β0 , t) 1X = n i=1 Z t 0 Zij  Sn1j − 0 (β0 , u) dMi (u). Sn Hereafter, we assume the following conditions. Assumption 2.1. (i) There exists a positive constant K1 such that sup |Zij | < K1 . i,j (ii) The baseline hazard function α0 is integrable on [0, τ ], i.e., Z τ α0 (t)dt < ∞. 0 (iii) There exist R-valued function s0n (β, t), Rpn -valued function s1n (β, t), and pn × pn matrix-valued function s2n (β, t) such that for l = 0, 1, 2, sup sup β∈Rpn t∈[0,τ ] 1 l S (β, t) − sln (β, t) n n ∞ →p 0. (iv) For l = 0, 1, 2, the functions (β, t) 7→ sln (β, t) are uniformly continuous. Moreover, they satisfy the following conditions: lim sup sup sup ksln (β, t)k∞ < ∞, n→∞ β∈Rpn t∈[0,τ ] lim inf infp n→∞ β∈R n inf s0n (β, t) > 0. t∈[0,τ ] (v) Assume that S is a constant not depending on n. Define the pn × pn matrix In (β) of rank S by #  1 ⊗2 Z τ" 2 sn sn In (β) := (β, u) − (β, u) s0n (β0 , u)α0 (u)du. s0n s0n 0 When we compute s0n , s1n , s2n only with respect to the components of the true S dimensional vector β0 , we have the S × S sub-matrix In (β0 ) of In (β0 ). We assume that In (β0 ) is positive definite, and In (β) for all β ∈ Rpn are nonnegative definite matrices. 4 2.2 Estimation procedure We define the estimator β̂n of β0 as β̂n := arg min kβk1 , (1) β∈Bn where Bn := {β ∈ Rpn : kUn (β)k∞ ≤ γ}, and γ ≥ 0 is a suitable constant. We call the estimator β̂n the Dantzig Selector for Proportional Hazards model (DSfPH). Note that this estimator is called the Survival Dantzig Selector (SDS) by Antoniadis et al. (2010). 3 Matrix conditions In this section, we will discuss some matrix conditions to derive the theoretical results for DSfPH β̂n . Hereafter, we write T0 for the support of β0 , i.e., T0 := {j : β0j 6= 0}. To begin with, we introduce the following three factors (A), (B) and (C), all of which are used by Huang et al. (2013) for LASSO in Cox’s proportional hazards model. Definition 3.1. For every index set T ⊂ {1, 2, · · · , pn } and h ∈ Rpn , hT is a R|T | dimensional sub-vector of h constructed by extracting the components of h corresponding to the indices in T . Define the set CT by CT := {h ∈ Rpn : khT c k1 ≤ khT k1 }. We introduce the following three factors. (A) Compatibility factor 1 κ(T0 ; In (β0 )) := inf 06=h∈CT0 1 S 2 (hT In (β0 )h) 2 . khT0 k1 (B) Weak cone invertibility factor 1 Fq (T0 ; In (β0 )) := inf 06=h∈CT0 1 S q (hT In (β0 )h) 2 , khT0 k1 khkq q ≥ 1. (C) Restricted eigenvalue 1 RE(T0 ; In (β0 )) := 5 inf 06=h∈CT0 (hT In (β0 )h) 2 . khk2 As mentioned in the Introduction, Huang et al. (2013) defined these factors for the random matrix Jn (β0 ), and derived some conditions to treat them as deterministic constants. On the other hand, we define them not for Jn (β0 ), but for the deterministic matrix In (β0 ), since we will prove that kIn (β0 ) − Jn (β0 )k∞ = op (1) in Section 4 of this paper. There exist other matrix conditions for DSfPH such as the UUP condition, which is used in Candés and Tao (2007) and Antoniadis et al. (2010). To discuss the relationship between the UUP condition and our conditions, let us introduce some objects. Note that there exists a matrix A such that AT A = In (β), because In (β) is a nonnegative definite matrix. Given a index set T ⊂ {1, 2, . . . , pn }, we write AT for the pn ×|T | matrix constructed by extracting the columns of A corresponding to the indices in T . The restricted isometry constant δN (In (β0 )) is the smallest quantity such that (1 − δN (In (β0 )))khk22 ≤ kAT hk22 ≤ (1 + δN (In (β0 )))khk22 , for all T ⊂ {1, 2, . . . , pn } with |T | ≤ N , where N ≤ pn is an integer, and all h ∈ R|T | . The restricted orthogonality constant θS,S ′ (In (β0 )) is the smallest quantity such that |(AT h)T AT ′ h′ | ≤ θS,S ′ (In (β0 ))khk2 kh′ k2 for all disjoint sets T, T ′ ⊂ {1, 2, . . . , pn } with |T | ≤ S, |T ′ | ≤ S ′ , where S+S ′ ≤ ′ pn and all vectors h ∈ R|T | and h′ ∈ R|T | . For δ2S (In (β0 )) and θS,2S (In (β0 )), the UUP condition is described that 0 < 1 − δ2S (In (β0 )) − θS,2S (In (β0 )). In addition, we introduce another factor φ2S (T0 ; In (β0 )) by 1 φ2S (T0 ; In (β0 )) := where DT0 ,T :=  inf T ⊃T0 , |T |≤2S, h∈DT0 ,T (hT In (β0 )h) 2 , khT k2  h ∈ CT0 : khT c k∞ ≤ min |hj | , j∈T \T0 T ⊃ T0 . Define that minj∈T \T0 |hj | = ∞ when T = T0 . The next lemma provides the asymptotic relationship between the UUP condition and a condition for φ2S (T0 ; In (β0 )). The proof is an adaptation of that in van de Geer and Bühlmann (2007), so it is omitted. Lemma 3.2. If lim inf n→∞ {1 − δ2S (In (β0 )) − θS,2S (In (β0 ))} > 0, then it holds that lim inf n→∞ φ2S (T0 ; In (β0 )) > 0. Noting that khT0 k21 ≤ SkhT0 k22 , we have that κ(T0 ; In (β0 )) ≥ φ2S (T0 ; In (β0 )), which implies that lim inf κ(T0 ; In (β0 )) ≥ lim inf φ2S (T0 ; In (β0 )) > 0. n→∞ n→∞ 6 Noting also that khT0 kq1 ≥ khT0 kqq for all q ≥ 1, we can see that κ(T0 ; In (β0 )) ≤ √ 2 SRE(T0 ; In (β0 )), and κ(T0 ; In (β0 )) ≤ Fq (T0 ; In (β0 )). We thus have that the factors (A), (B) and (C) are strictly positive when n is large if lim inf n→∞ {1 − δ2S (In (β0 )) − θS,2S (In (β0 ))} > 0. So we will assume in our main theorems that these three factors are “asymptotically positive”, in the sense that lim inf κ(T0 ; In (β0 )) > 0, n→∞ lim inf Fq (T0 ; In (β0 )) > 0 n→∞ or lim inf RE(T0 ; In (β0 )) > 0, n→∞ to prove the consistency of DSfPH. 4 Main result In this section, we will prove the consistency of DSfPH β̂n . To do this, we will prepare three lemmas. Lemma 4.1 below states that the true parameter β0 is an element of Bn appearing in (1) with large probability when the sample size n is large. Lemma 4.1. Put γ = γn,pn = K2 log(1 + pn )/nα , where 0 < α ≤ 1/2 and K2 > 0 are constants. If pn = O(nξ ) for some ξ > 1 or if log pn = O(nζ ) for some 0 < ζ < α, then it hold that lim P (kUn (β0 )k∞ ≥ γn,pn ) = 0 n→∞ and that γn,pn → 0 as n → ∞. Proof. Note that n Unj (β0 , τ ) where 1X = n i=1 Z 0 τ " n X {Zij k=1 − # Zkj }wk (β0 , u) exp(ZkT β0 )Yk (u) . wk (β0 , u) = P n exp(ZlT β0 )Yl (u) dMi (u), l=1 We use Lemma 2.1 from van de Geer (1995). To do this, we shall evaluate ∆Unj (β0 , u), u ∈ [0, τ ], and hUnj (β0 , ·), Unj (β0 , ·)iτ . Since the jumps of Mi do not occur at the same time and are all of magnitude 1, it holds that " n # n Z 1X u X j j j |∆Un (β0 , u)| = {Zi − Zk }wk (β0 , u) dMi (u) n i=1 u− k=1 n ≤ X 1 sup |Zij − Zij | wk (β0 , u) n i,j,k k=1 2K1 . ≤ n 7 On the other hand, we have " n #2 n Z 1 X τ X j j = 2 {Zi − Zk }wk (β0 , u) dhMi iu n i=1 0 k=1 " n #2 n Z 1 X τ X j j {Zi − Zk }wk (β0 , u) exp(ZiT β0 )Yi (u)α0 (u)du n2 i=1 0 k=1 n Z τ X 1 j j 2 sup |Z − Z | exp(ZiT β0 )Yi (u)α0 (u)du k n2 i,j,k i 0 i=1 Z τ 4K12 j n exp(S sup |Zi |kβ0 k∞ |) α0 (u)du n2 i,j 0 K3 , n is a positive constant. We now use the Lemma 2.1 from van de Geer hUnj (β0 , ·), Unj (β0 , ·)iτ = ≤ ≤ ≤ where K3 (1995): P (|Unj (β0 , τ )|   K3 j j j ≥ γn,pn ) = P |Un (β0 , τ )| ≥ γn,pn , hUn (β0 , ·), Un (β0 , ·)iτ ≤ n ! 2 γn,pn  . ≤ 2 exp − 2K1 2 n γn,pn + Kn3 Write k · kψ for the Orlicz norm with respect to ψ(x) = ex − 1. We apply Lemma 2.2.10 from van der Vaart and Wellner (1996) to deduce that there exists a constant L > 0 depending only on ψ such that ! r K3 2K1 j log(1 + pn ) + log(1 + pn ) . max |Un (β0 , τ )| ≤ L 1≤j≤pn n n ψ Using Markov’s inequality, we have that P (kU (β0 )k∞ ≥ γn,pn ) = P ( max |Unj (β0 , τ )| ≥ γn,pn ) 1≤j≤pn      j max |Un (β0 , τ )| γ 1≤j≤pn n,pn  ≥ ψ  ≤ P ψ  k max |Unj (β0 , τ )|kψ k max |Unj (β0 , τ )|kψ 1≤j≤pn ≤ ≤  ψ  γn,pn k max |Unj (β0 , τ )|kψ 1≤j≤pn 1≤j≤pn −1  −1   γn,pn  ψ q    K3 1 L 2K n log(1 + pn ) + n log(1 + pn ) In our settings, the right-hand side of this inequality converges to 0. 8 ✷ Next we will show that Jn (β0 ) is approximated by In (β0 ). Lemma 4.2. The random sequence ǫn defined by ǫn := kJn (β0 ) − In (β0 )k∞ converges in probability to 0. Proof. Define the pn × pn matrices hn (β0 , t) and Hn (β0 , t) for t ∈ [0, τ ] by ⊗2 s1n (β0 , t), s0n  1 ⊗2 S2 Sn Hn (β0 , t) := n0 (β0 , t) − (β0 , t). Sn Sn0 s2 hn (β0 , t) := n0 (β0 , t) − sn  Note that the matrices In (β0 ) and Jn (β0 ) can be written in this form: Z τ hn (β0 , u)s0 (β0 , u)α0 (u)du, In (β0 ) = 0 Z τ dN̄ (u) . Hn (β0 , u) Jn (β0 ) = n 0 Pn Put M̄ (u) = i=1 Mi (u). Then, it holds that kJn (β0 ) − In (β0 )k∞ ≤ (I) + (II) + (III), where Z τ dN̄ (u) , (I) = kHn (β0 , u) − hn (β0 , u)k∞ n 0  0  Z τ Sn (β0 , u) (II) = hn (β0 , u) − s0 (β0 , u) α0 (u)du, n 0 ∞ Z 1 τ (III) = hn (β0 , u)dM̄ (u) . n 0 ∞ Since the process t ❀ N̄ (t)/n has bounded variation uniformly in n, Assumption 2.1 implies that (I) = op (1) and (II) = op (1). Moreover, it follows from Assumption 2.1 that hn (β0 , u) is uniformly bounded. So we obtain that (III) = op (1) by the same way as the proof of Lemma 4.1. ✷ The next lemma is used to control Un (β̂) − Un (β0 ) and Jn (β0 ). See Huang et al. (2013) and Hjort and Pollard (1993) for the proofs. Lemma 4.3. Define that ηh = maxi,j |hT Zi − hT Zj |, for h ∈ Rpn . Then for all β ∈ Rpn , it holds that e−ηh hT J(β)h ≤ hT [U (β + h) − U (β)] ≤ eηh hT J(β)h. Now, we are ready to prove the main result of this paper. Theorem 4.4 below provides the l2 consistency of DSfPH. 9 Theorem 4.4. Under the assumption of Lemma 4.1 and Assumption 2.1, if lim inf n→∞ RE(T0 ; In (β0 )) > 0, then it holds that   K4 γn,pn 2 lim P kβ̂n − β0 k2 ≥ = 0, n→∞ RE 2 (T0 ; In (β0 )) − ǫn where K4 is a positive constant and ǫn = kIn (β0 ) − Jn (β0 )k∞ = op (1). In particular, kβ̂n − β0 k2 →p 0. Proof. It is sufficient to prove that kUn (β0 )k∞ ≤ γn,pn implies kβ̂n − β0 k22 ≤ K4 γn,pn . RE 2 (T0 ; In (β0 )) − ǫn By the construction of the estimator, we have kU (β̂n )k∞ ≤ γn,pn , which implies that kUn (β̂n ) − Un (β0 )k∞ ≤ kUn (β̂n )k∞ + kUn (β0 )k∞ ≤ 2γn,pn . Note that h := β̂ − β0 ∈ CT0 , since it holds that X X c | + 0 ≥ kβ0 + hk1 − kβ0 k1 = |hT0j (|β0j + hT0j | − |β0j |) j∈T0c ≥ X j∈T0c j∈T0 |hT0c j | − X j∈T0 |hT0j | = khT0c k1 − khT0 k1 . Notice moreover that khk1 ≤ kβ̂n k1 + kβ0 k1 ≤ 2kβ0 k1 by the definition of β̂n . Now, we use Lemma 4.3 for h to deduce that hT Jn (β0 )h ≤ eηh hT [Un (β̂n ) − Un (β0 )] ≤ exp(max |hT Zi − hT Zj |) · 2γn,pn khk1 i,j ≤ exp(4K1 kβ0 k1 ) · 4γn,pn kβ0 k1 =: K4 γn,pn . Thus it holds that hT In (β0 )h ≤ |hT (In (β0 ) − Jn (β0 ))h| + hT Jn (β0 )h ≤ ǫn hT h + K4 γn,pn = ǫn kβ̂n − β0 k22 + K4 γn,pn . By the definition of the restricted eigenvalue, we have that RE 2 (T0 ; In (β0 )) ≤ ≤ hT In (β0 )h kβ̂n − β0 k22 ǫn kβ̂n − β0 k22 + K4 γn,pn kβ̂n − β0 k22 10 . Noting that RE 2 (T0 ; In (β0 )) > 0, we obtain that kβ̂n − β0 k22 ≤ K4 γn,pn . 0 ; In (β0 )) − ǫn RE 2 (T ✷ To derive the l1 consistency and the lq consistency of DSfPH, we shall use the compatibility factor and the weak cone invertibility factor, respectively. Theorem 4.5. Under the assumptions of Lemma 4.1 and Assumption 2.1, if lim inf n→∞ κ(T0 ; In (β0 )) > 0, then the following (i) and (ii) hold true. (i) It holds that  lim P kβ̂n − β0 k1 ≥ n→∞ 4K5 Sγn,pn κ2 (T0 ; In (β0 )) − 4Sǫn  = 0, where K5 is a positive constant. In particular, kβ̂n − β0 k1 →p 0. (ii) It holds for any q > 1 that 1 lim P n→∞ 1 2S q ǫn 2K5 S q γn,pn 2K5 Sγn,pn kβ̂n − β0 kq ≥ + · 2 Fq (T0 ; In (β0 )) κ (T0 ; In (β0 )) − 2Sǫn Fq (T0 ; In (β0 )) In particular, kβ̂n − β0 kq →p 0. Proof. It follows from the proof of Theorem 4.4 that hT Jn (β0 )h ≤ K5 γn,pn kβ̂n − β0 k1 . Noting that kbk22 ≤ kbk21 for all b ∈ Rpn , we have that hT In (β0 )h ≤ ǫn kβ̂n − β0 k22 + K5 γn,pn kβ̂n − β0 k1 ≤ ǫn kβ̂n − β0 k21 + K5 γn,pn kβ̂n − β0 k1 . The definition of κ(T0 ; In (β0 )) implies that ShT In (β0 )h khT0 k21 Sǫn khk21 + K5 Sγn,pn khk1 ≤ . khT0 k21 κ2 (T0 ; In (β0 )) ≤ Since khk1 ≤ 2khT0 k1 , this yields the conclusion in (i). On the other hand, using the weak cone invertibility factor for every q ≥ 1, we have that 1 1 S q ǫn khk21 + S q K5 γn,pn khk1 , Fq (T0 ; In (β0 )) ≤ khT0 k1 khkq 11 ! = 0. which implies that 1 1 2S q ǫn kβ̂ − β0 k1 + 2S q K5 γn,pn kβ̂n − β0 kq ≤ . Fq (T0 ; In (β0 )) Using the l1 bound derived above, we obtain the conclusion in (ii). ✷ Acknowledgements. The second author’s work was supported by Grant-inAid for Scientific Research (C), 15K00062, from Japan Society for the Promotion of Science. References [1] Andersen, P.K. and Gill, R.D. (1982). Cox’s regression model for counting processes: a large sample study. Ann. Statist. 10,no.4, 1100-1120. [2] Antoniadis, A., Fryzlewicz, P. and Letué, F. (2010). The Dantzig selector in Cox’s proportional hazards model. Scand. J. Stat. 37, no.4, 531-552. [3] Bickel, P.J., Ritov, Y. and Tsybakov, A.B. (2009). Simultaneous analysis of lasso and Dantzig selector. Ann. Statist. 37 (2009), no. 4, 1705-1732. [4] Bradic, J. Fan, J. and Jiang, J. (2011). Regularization for Cox’s proportional hazards model with NP-dimensionality. Ann. Statist. 39, no.6, 30923120. [5] Candés, E.J. and Tao, T. (2005). Decoding by linear programming. IEEE Trans. Inform. Theory 51, no. 12, 4203-4215. [6] Candés, E. and Tao, T. (2007). The Dantzig selector: statistical estimation when p is much larger than n. Ann. Statist. 35, no.6, 2313-2351. [7] Cox, D.R. (1972). Regression models and life tables (with discussion). J. Roy. Statist. Soc. Ser B 34 187-220. [8] Fan, J. and Li, R. (2002). Variable selection for Cox’s proportional hazards model and frailty model. Ann. Statist. 30, no.1 74-99. [9] Gautier, E. and Tsybakov, A.B. (2014). High-dimensional instrumental variables regression and confidence sets. arXiv:1105.2454 [math.ST] [10] Hjort, N.L. and Pollard, D. (1993). Asymptotics for minimisers of convex processes. arXiv:1107.3806 [math.ST] [11] Huang, J., Sun, T., Ying, Z., Yu, Y. and Zhang, C-H. (2013). Oracle inequalities for the LASSO in the Cox model. Ann. Statist. 41, no.3, 11421165. 12 [12] Koltchinskii, V. (2009). The Dantzig selector and sparsity oracle inequalities. Bernoulli 15, no. 3, 799-828. [13] Rosembaum, M. and Tsybakov, A.B. (2013). Improved matrix uncertainty selector. From probability to statistics and back: high-dimensional models and processes, 276-290, Inst. Math. Stat. (IMS) Collect., 9, Inst. Math. Statist., Beachwood, OH. [14] Shorack, G.R. and Wellner, J.A. (1986). Empirical Processes with Applications to Statistics. Wiley Series in Probability and Mathematical Statistics. John Wiley and Sons, Inc., New York. [15] Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. J. Roy. Statist. Soc. Ser B. 58, no.1, 267-288. [16] Tibshirani, R. (1997). The lasso method for variable selection in the Cox model. Stat. Med. 16 385-395. [17] van de Geer, S. (1995). Exponential inequalities for martingales, with application to maximum likelihood estimation for counting processes. Ann. Statist. 23, no.5, 1779-1801. [18] van de Geer, S.A. and Bühlmann, P. (2009). On the conditions used to prove oracle results for the Lasso. Electron. J. Stat. 3, 1360-1392. [19] van der Vaart, A.W. and Wellner, J.A. (1996). Weak Convergence and Empirical Processes. With Applications to Statistics. Springer Series in Statistics. Springer-verlag, New York. [20] Ye, F. and Zhang, C-H. (2010). Rate minimaxity of the Lasso and Dantzig selector for the lq loss in lr balls. J. Mach. Learn. Res. 11, 3519-3540. [21] Zhang, H.H. and Lu, W. (2007). Adaptive Lasso for Cox’s proportional hazards model. Biometrika 94, no.3, 691-703. 13
10math.ST
DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS arXiv:1605.02548v4 [math.GT] 1 Dec 2016 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS Abstract. In this survey, we study representations of finitely generated groups into Lie groups, focusing on the deformation spaces of convex real projective structures on closed manifolds and orbifolds, with an excursion on projective structures on surfaces. We survey the basics of the theory of character varieties, geometric structures on orbifolds, and Hilbert geometry. The main examples of finitely generated groups for us will be Fuchsian groups, 3-manifold groups and Coxeter groups. Contents 1. Introduction 2. Character varieties 3. Geometric structures on orbifolds 4. A starting point for convex projective structures 5. The existence of deformations or exotic structures 6. Real projective surfaces 7. Convex projective Coxeter orbifolds 8. Infinitesimal deformations 9. Infinitesimal duality to complex hyperbolic geometry 10. Convex projective structures on 3-manifolds References 1 4 5 9 13 15 23 29 32 33 34 1. Introduction The goal of this paper is to survey the deformation theory of convex real projective structures on manifolds and orbifolds. Some may prefer to speak of discrete subgroups of the group PGLd+1 (R) of projective transformations of the real projective space RPd which preserve a properly convex open subset of RPd , and some others prefer to speak of Hilbert geometries. Some motivations for studying this object are the following: S. Choi was supported by the Mid-career Researcher Program through the NRF grant NRF2013R1A1A2056698 funded by the MEST. G.-S. Lee was supported by the DFG research grant “Higher Teichmüller Theory” and he acknowledges support from the U.S. National Science Foundation grants DMS 1107452, 1107263, 1107367 “RNMS: GEometric structures And Representation varieties” (the GEAR Network). L. Marquis acknowledges support from the French ANR programs Finsler and Facets. The authors thank Lizhen Ji, Athanase Papadopoulos and Shing-Tung Yau for the opportunity to publish this survey in “Handbook of Group Actions”. Finally, we would like to thank the referee and Athanase Papadopoulos for carefully reading this paper and suggesting several improvements. 1 2 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS 1.1. Hitchin representations. Let S be a closed surface of genus g ⩾ 2 and let Γ be the fundamental group of S. There is a unique irreducible representation im ∶ PSL2 (R) → PSLm (R), up to conjugation. A representation ρ ∶ Γ → PSLm (R) is called a Hitchin representation if there are a continuous path ρt ∈ Hom(Γ, PSLm (R)) and a discrete, faithful representation τ ∶ Γ → PSL2 (R) such that ρ0 = ρ and ρ1 = im ○ τ ∶ Γ → PSLm (R). The space Hitm (Γ) of conjugacy classes of Hitchin representations of Γ in PSLm (R) has a lot of interesting properties: Each connected component is homeomorphic to an open ball of dimension 2(g − 1)(m2 − 1) (see Hitchin [65]), and every Hitchin representation is discrete, faithful, irreducible and purely loxodromic (see Labourie [78]). When m = 3, the first author and Goldman [35] showed that each Hitchin representation preserves a properly convex domain1 of RP2 . In other words, Hit3 (Γ) is the space C(S) of marked convex real projective structures on the surface S. To understand the geometric properties of Hitchin representations, Labourie [78] introduced the concept of Anosov representation. Later on, Guichard and Wienhard [63] studied this notion for finitely generated Gromov-hyperbolic groups. For example, if M is a closed manifold whose fundamental group is Gromov-hyperbolic, then the holonomy representations of convex real projective structures on M are Anosov. 1.2. Deformations of hyperbolic structures. Let M be a closed hyperbolic manifold of dimension d ⩾ 3, and let Γ = π1 (M ). By Mostow rigidity, up to conjugation, there is a unique faithful and discrete representation ρgeo of Γ in PSOd,1 (R). The group PSOd,1 (R) is canonically embedded in PGLd+1 (R). We use the same notation ρgeo ∶ Γ → PGLd+1 (R) to denote the composition of ρgeo ∶ Γ → PSOd,1 (R) with the canonical inclusion. Now, there is no reason that ρgeo is the unique faithful and discrete representation of Γ in PGLd+1 (R), up to conjugation. In fact, there are examples of closed hyperbolic manifold M of dimension d such that Γ admits discrete and faithful representations in PGLd+1 (R) which are not conjugate to ρgeo (see Theorem 5.2). We can start looking at the connected component Ben(M ) of the space of representations of Γ into PGLd+1 (R) containing ρgeo , up to conjugation. The combination of a theorem of Koszul and a theorem of Benoist implies that every representation in Ben(M ) is discrete, faithful, irreducible and preserves a properly convex domain Ω of RPd .2 At the moment of writing this survey, there is no known necessary and sufficient condition on M to decide if Ben(M ) consists of exactly one element, which is the hyperbolic structure. There are infinitely many closed hyperbolic 3-manifolds M such that Ben(M ) is the singleton (see Heusener-Porti [64]), and there are infinitely many closed hyperbolic 3-orbifolds M such that Ben(M ) is homeomorphic to an open k-ball, for any k ∈ N (see Marquis [84]). 1.3. Building blocks for projective surfaces. Let S be a closed surface. We might wish to understand all possible real projective structures on S, not necessarily only the convex one. The first author showed that convex projective structures are the main building blocks to construct all possible projective structures on the surface S (see Theorem 6.9). 1We 2The abbreviate a connected open set to a domain. action of Γ on Ω is automatically proper and cocompact for general reasons. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 3 1.4. Geometrization. Let Ω be a properly convex domain of RPd , and let Aut(Ω) be the subgroup of PGLd+1 (R) preserving Ω. There is an Aut(Ω)-invariant metric dΩ on Ω, called the Hilbert metric, that make (Ω, dΩ ) a complete proper geodesic metric space, called a Hilbert geometry. We will discuss these metrics in Section 4.2. The flavour of the metric space (Ω, dΩ ) really depends on the geometry of the boundary of Ω. For example, on the one hand, the interior of an ellipse equipped with the Hilbert metric is isometric to the hyperbolic plane, forming the projective model of the hyperbolic plane, and on the other hand, the interior of a triangle is isometric to the plane with the norm whose unit ball is the regular hexagon (see de la Harpe [48]). Unfortunately, Hilbert geometries are almost never CAT(0): A Hilbert geometry (Ω, dΩ ) is CAT(0) if and only if Ω is an ellipsoid (see Kelly-Straus [70]). However, the idea of Riemmanian geometry of non-positive curvature is a good guide towards the study of the metric properties of Hilbert geometry. An irreducible symmetric space X = G/K is “Hilbertizable” if there exist a properly convex domain Ω of RPd for some d and an irreducible representation ρ ∶ G → PSLd+1 (R) such that ρ(G) acts transitively on Ω and the stabilizer of a point of Ω is conjugate to K. The symmetric spaces for PSOd,1 (R), PSLm (K) for K = R, C, H, and the exceptional Lie group E6,−26 are exactly the symmetric spaces that are Hilbertizable (see Vinberg [103, 104] or Koecher [74]). Nevertheless we can ask the following question to start with: “Which manifold or orbifold M can be realized as the quotient of a properly convex domain Ω by a discrete subgroup Γ of Aut(Ω)?” If this is the case, we say that M admits a properly convex real projective structure. In dimension 2, the answer is easy: a closed surface S admits a convex projective structure if and only if its Euler characteristic is non-positive. The universal cover of a properly convex projective torus is a triangle, and a closed surface of negative Euler characteristic admits a hyperbolic structure, which is an example of a properly convex projective structure. In dimension greater than or equal to 3, no definite answer is known; see Section 5 for a description of our knowledge. To arouse the reader’s curiosity we just mention that there exist manifolds which admit a convex real projective structure but which cannot be locally symmetric spaces. 1.5. Coxeter groups. A Coxeter group is a finitely presented group that “resembles” the groups generated by reflections; see Section 7 for a precise definition, and de la Harpe [47] for a beautiful invitation. An important object to study Coxeter group, denoted W , is a representation ρgeo ∶ W → GL(V ) introduced by Tits [24]. The representation ρgeo , in fact, preserves a convex domain of the real projective space P(V ). For example, Margulis and Vinberg [83] used this property of ρgeo to show that an irreducible Coxeter group is either finite, virtually abelian or large.3 3A group is large if it contains a subgroup of finite index that admits an onto morphism to a non-abelian free group. 4 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS From our point of view, Coxeter groups are a great source for building groups acting on properly convex domains of P(V ). Benoist [14, 15] used them to construct the first example of a closed 3-manifold that admits a convex projective structure Ω/Γ such that Ω is not strictly convex, or to build the first example of a closed 4-manifold that admits a convex projective structure Ω/Γ such that (Ω, dΩ ) is Gromov-hyperbolic but not quasi-isometric to the hyperbolic space (see Section 5). 2. Character varieties All along this article, we study the following kind of objects: ● a finitely generated group Γ which we think of as the fundamental group of a complete real hyperbolic manifold/orbifold or its siblings, ● a Lie group G which is also the set of real points of an algebraic group G, and ● a real algebraic set Hom(Γ, G). We want to understand the space Hom(Γ, G). First, the group G acts on Hom(Γ, G) by conjugation. We can notice that the quotient space is not necessarily Hausdorff since the orbit of the action of G on Hom(Γ, G) may not be closed. But the situation is not bad since each orbit closure contains at most one closed orbit. Hence, a solution to the problem is to forget the representations whose orbits are not closed. Let us recall the characterization of the closedness of the orbit: Lemma 2.1 (Richardson [97]). Assume that G is the set of real points of a reductive4 algebraic group defined over R. Let ρ ∶ Γ → G be a representation. Then the orbit G ⋅ ρ is closed if and only if the Zariski closure of ρ(Γ) is a reductive subgroup of G. Such a representation is called a reductive representation. Define R(Γ, G) = Hom(Γ, G)/G and χ(Γ, G) = {[ρ] ∈ R(Γ, G) ∣ ρ is reductive}. These spaces are given with the quotient topology and the subspace topology, respectively. Theorem 2.2 (Topological, geometric and algebraic viewpoint, Luna [81, 82] and Richardson-Slodowy [98]). Assume that G and ρ are as in Lemma 2.1. Then → ● There exists a unique reductive representation Ð ρ ∈ G ⋅ ρ, up to conjugation. ● The space χ(Γ, G) is Hausdorff and it is identified with the Hausdorff quotient of R(Γ, G). ● The space χ(Γ, G) is also a real semi-algebraic variety which is the GIT-quotient5 of the action of G on Hom(Γ, G). The real semi-algebraic Hausdorff space χ(Γ, G) is called the character variety of the pair (Γ, G). Remark 2.3. Theorem 2.2 is not explicitly stated in Luna’s work. The statement can be founded in Remark 7.3.3 of Richardson-Slodowy [98] and is also proved by independent method via the action of the reductive group G on the affine variety Hom(Γ, G). See also Section 3 of Bergeron [18]. 4An algebraic group is reductive if its unipotent radical is trivial. is the abbreviation for the Geometric Invariant Theory; see the lecture notes of Brion [26] for information on this subject. 5GIT DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 5 A baby example. The space χ(Z, G) is the set of semi-simple elements of G modulo conjugation. ● If G = SLm (C), then χ(Z, G) = Cm−1 . ● If G = SL2 (R), then χ(Z, G) is a circle {eiθ ∣ 0 ⩽ θ < 2π} with two half-lines that are glued on the circle at the points {1} and {−1}. 3. Geometric structures on orbifolds In this section, we recall the vocabulary of orbifolds and of geometric structures on orbifolds. The reader can skip this section if he or she is familiar with these notions. A classical reference are Thurston’s lecture notes [100]. See also Goldman [55], Choi [34], BoileauMaillot-Porti [22]. For the theory of orbifolds itself, we suggest the article of Moerdijk-Pronk [90], the books of Adem-Leida-Ruan [1] and of Bridson-Haefliger [25]. 3.1. Orbifolds. An orbifold is a topological space which is locally homeomorphic to the quotient space of Rd by a finite subgroup of Diff(Rd ), the diffeomorphism group of Rd . Here is a formal definition: A d-dimensional orbifold O consists of a second countable, Hausdorff space XO with the following additional structure: (1) A collection of open sets {Ui }i∈I , for some index I, which is a covering of XO and is closed under finite intersections. (2) To each Ui are associated a finite group Γi , a smooth action of Γi on an open subset ̃i in Rd and a homeomorphism φi ∶ U ̃i /Γi → Ui . U (3) Whenever Ui ⊂ Uj , there are an injective homomorphism fij ∶ Γi → Γj and a smooth ̃ij ∶ U ̃ij (γx) = fij (γ)φ ̃ij (x) ̃i → U ̃j equivariant with respect to fij , i.e. φ embedding φ ̃ for γ ∈ Γi and x ∈ Ui , such that the following diagram commutes: ̃i U ̃ij φ ̃i /Γi U ̃j /fij (Γi ) U ̃j /Γj U φi Ui ̃j U ⊂ φj Uj (4) The collection {Ui } is maximal relative to the conditions (1) – (3). This additional structure is called an orbifold structure, and the space XO is the underlying ̃ij is uniquely determined up to space of O. Here, it is somewhat important to realize that φ compositions of elements of Γi and Γj . Example. If M is a smooth manifold and Γ is a subgroup of Diff(M ) acting properly discontinuously on M , then the quotient space M /Γ has an obvious orbifold structure. An orbifold is said to be connected, compact or noncompact according to whether the underlying space is connected, compact or noncompact, respectively. 6 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS A smooth map between orbifolds O and O′ is a continuous map f ∶ XO → XO′ satisfying that ̃ /Γ of x in O and U ′ ≈ Ũ′ /Γ′ of f (x) for each x ∈ O there are coordinate neighborhoods U ≈ U ̃→U ̃′ in O′ such that f (U ) ⊂ U ′ and the restriction f ∣U can be lifted to a smooth map f̃ ∶ U which is equivariant with respect to a homomorphism Γ → Γ′ . Note that the homomorphism Γ → Γ′ may not be injective nor surjective. An orbifold-diffeomorphism between O and O′ is a smooth map O → O′ with a smooth inverse map. If there is an orbifold-diffeomorphism between O and O′ , we denote this by O ≈ O′ . An orbifold S is a suborbifold of an orbifold O if the underlying space XS of S is a subset ̃ /Γ) of s in O of XO and for each point s ∈ XS , there are a coordinate neighborhood U = φ(U ̃ preserved by Γ such that V = φ(Ṽ /(Γ∣ ̃ )) is a coordinate and a closed submanifold Ṽ of U V neighborhood of s in S. Here, since the submanifold Ṽ is preserved by Γ, we denote by Γ∣Ṽ the group obtained from the elements of Γ by restricting their domains and codomains to Ṽ . Note that the restriction map Γ → Γ∣Ṽ may not be injective (see also Borzellino-Brunsden [23]). For example, let σx (resp. σy ) be the reflection in the x-axis Lx (resp. y-axis Ly ) of R2 . Then Lx /⟨σy ⟩ and Ly /⟨σx ⟩ are suborbifolds of R2 /⟨σx , σy ⟩, however both maps ⟨σx , σy ⟩ → ⟨σy ⟩ and ⟨σx , σy ⟩ → ⟨σx ⟩ are not injective. Our definition is more restrictive than Adem-LeidaRuan’s [1] and less restrictive than Kapovich’s [69], however, this definition seems to be better for studying decompositions of 2-orbifolds along 1-orbifolds. 3.2. (G, X)-orbifolds. Let X be a real analytic manifold and let G be a Lie group acting analytically, faithfully and transitively on X. An orbifold is a (G, X)-orbifold if Γi is a sub̃ij is locally an element of G (c.f. the definition ̃i is an open subset of X, and φ group of G, U of orbifold). A (G, X)-manifold is a (G, X)-orbifold with Γi trivial. A (G, X)-structure on an orbifold O is an orbifold-diffeomorphism from O to a (G, X)-orbifold S. Here are some examples: Let Ed be the d-dimensional Euclidean space and let Isom(Ed ) be the group of isometries of Ed . Having an (Isom(Ed ), Ed )-structure (or Euclidean structure) on a manifold O is equivalent to having a Riemannian metric on O of sectional curvature zero. We can also define a spherical structure or a hyperbolic structure on O and give a similar characterization for each structure. Let Ad be the d-dimensional affine space and let Aff(Ad ) be the group of affine transformations, i.e. transformations of the form x ↦ Ax + b where A is a linear transformation of Ad and b is a vector in Ad . An (Aff(Ad ), Ad )-structure (or affine structure) on an orbifold O is equivalent to a flat torsion-free affine connection on O (see Kobayashi-Nomizu [73]). Similarly, a (PGLd+1 (R), RPd )-structure (or real projective structure) on O is equivalent to a projectively flat torsion-free affine connection on O (see Eisenhart [49]). 3.3. A tool kit for orbifolds. To each point x in an orbifold O is associated a group Γx ̃ /Γ this is the isomorphism called the isotropy group of x: In a local coordinate system U ≈ U ̃ . The set {x ∈ XO ∣ Γx ≠ {1}} is class of the stabilizer Γx̃ ⩽ Γ of any inverse point x̃ of x in U the singular locus of O. In general, the underlying space of an orbifold is not even a manifold. However, in dimension two, it is homeomorphic to a surface with/without boundary. Moreover, the singular DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 7 locus of a 2-orbifold can be classified into three families because there are only three types of finite subgroups in the orthogonal group O2 (R) : ● Mirror: R2 /Z2 when Z2 acts by reflection. ● Cone points of order n ⩾ 2: R2 /Zn when Zn acts by rotations of angle 2π n . ● Corner reflectors of order n ⩾ 2: R2 /Dn when Dn is the dihedral group of order 2n generated by reflections in two lines meeting at angle πn . ̃i to be an open set in the closed half-space Rd+ of In the definition of an orbifold, if we allow U Rd , then we obtain the structure of an orbifold with boundary. To make a somewhat redundant remark, we should not confuse the boundary ∂O of an orbifold O with the boundary ∂XO of the underlying space XO , when XO is a manifold with boundary. Example. A manifold M with boundary ∂M can have an orbifold structure in which ∂M becomes a mirror, i.e. a neighborhood of any point x in ∂M is orbifold-diffeomorphic to Rd /Z2 such that Z2 acts by reflection. Notice that the singular locus is then ∂M and the boundary of the orbifold is empty. Given a compact orbifold O, we can find a cell decomposition of the underlying space XO such that the isotropy group of each open cell is constant. Define the orbifold Euler characteristic to be 1 χ(O) ∶= ∑(−1)dim ci . ∣Γ(ci )∣ ci Here, ci ranges over the cells and ∣Γ(ci )∣ is the order of the isotropy group Γ(ci ) of any point in the relative interior of ci . A covering orbifold of an orbifold O is an orbifold O′ with a continuous surjective map p ∶ XO′ → XO between the underlying spaces such that each point x ∈ XO lies in a coordinate ̃ /Γ and each component Vi of p−1 (U ) is orbifold-diffeomorphic to U ̃ /Γi neighborhood U ≈ U with Γi a subgroup of Γ. The map p is called a covering map. Example. If a group Γ acts properly discontinuously on a manifold M and Γ′ is a subgroup of Γ, then M /Γ′ is a covering orbifold of M /Γ with covering map M /Γ′ → M /Γ. In particular, M is a covering orbifold of M /Γ. Even if it is more delicate than for manifold, we can define the universal covering orbifold ̃ with covering of an orbifold O: A universal covering orbifold of O is a covering orbifold O ′ ′ ̃ ̃ ̃′ → O, there map p ∶ O → O such that for every covering orbifold O with covering map p ∶ O ̃→O ̃′ which satisfies the following commutative diagram: is a covering map q ∶ O ̃ O q ̃′ O p O p′ It is important to remark that every orbifold has a unique universal covering orbifold (up to orbifold-diffeomorphism). The orbifold fundamental group π1orb (O) of O is the group of ̃ deck transformations of the universal covering orbifold O. 8 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS Example. If Γ is a cyclic group of rotations acting on the sphere S2 fixing the north and south poles, then the orbifold S2 /Γ is a sphere with two cone points. Therefore its orbifold fundamental group is Γ, even though the fundamental group of the sphere, which is the underlying space of S2 /Γ, is trivial. An orbifold O is good if some covering orbifold of O is a manifold. In this case, the ̃ is a simply connected manifold and the group π orb (O) acts universal covering orbifold O 1 ̃ In other words, a good orbifold is simply a manifold M with properly discontinuously on O. a properly discontinuous group action on M . Moreover, we have the following good news: Theorem 3.1 (Chapter 3 of Thurston [100]). Every (G, X)-orbifold is good. 3.4. Geometric structures on orbifolds. We will discuss the deformation space of geometric structures on an orbifold O as Goldman [55] exposed the theory for manifolds. Suppose that M and N are (G, X)-orbifolds. A map f ∶ M → N is a (G, X)-map if, for each pair of charts φ ∶ Ũi /Γi → Ui ⊂ M from the (G, X)-orbifold structure of M and ψ ∶ Ṽj /Γj → Vj ⊂ N from the (G, X)-orbifold structure of N , the composition ψ −1 ○ f ○ φ restricted to φ−1 (Ui ∩ f −1 (Vj )) lifts to the restriction of an element of G on the inverse image ̃i of φ−1 (Ui ∩ f −1 (Vj )). in U Recall a (G, X)-structure on an orbifold O is an orbifold-diffeomorphism f from O to a (G, X)-orbifold S. Two (G, X)-structures f ∶ O → S and f ′ ∶ O → S ′ on O are equivalent if the map f ′ ○ f −1 ∶ S → S ′ is isotopic to a (G, X)-map from S to S ′ (in the category of orbifold). The set of equivalence classes of (G, X)-structures on O is denoted by Def(O). There is a topology on Def(O) informally defined by stating that two (G, X)-structures f and f ′ are close if the map f ′ ○ f −1 ∶ S → S ′ is isotopic to a map close to a (G, X)-map. Below is a formal definition. The construction of the developing map and the holonomy representation of manifolds extends to orbifolds without difficulty; see Goldman [55] for manifolds and Choi [34] for ̃→X orbifolds. For a (G, X)-orbifold O, there exists a pair (D, ρ) of an immersion D ∶ O orb orb and a homomorphism ρ ∶ π1 (O) → G such that for each γ ∈ π1 (O), the following diagram commutes: ̃ D X O γ ̃ O ρ(γ) D X We call D a developing map and ρ a holonomy representation of O. Moreover if (D′ , ρ′ ) is another such pair, then there exists g ∈ G such that D′ = g ○ D and ρ′ (γ) = gρ(γ)g −1 for each γ ∈ π1orb (O). In other words, a developing pair (D, ρ) is uniquely determined up to the action of G: (1) g ⋅ (D, ρ(⋅)) = (g ○ D, gρ(⋅)g −1 ), for each g ∈ G. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 9 Consider the space ̃ → X is an immersion Def ′2 (O) = {(D, ρ) ∣ D ∶ O equivariant with respect to a homomorphism ρ ∶ π1orb (O) → G}/ ∼ . Here (D, ρ) ∼ (D′ , ρ′ ) if D′ = D ○ι̃ for the lift ι̃ of an isotopy ι ∶ O → O satisfying γ ○ι̃ = ι̃○γ for every γ ∈ π1orb (O). We topologize this space naturally using the C r -compact-open topology, r ⩾ 2, before taking the quotient, and denote by Def 2 (O) the quotient space of Def ′2 (O) by the action of G (see Equation (1)). We can define a map µ ∶ Def 2 (O) → Def(O) from [(D, ρ)] to a (G, X)-structure on O ̃ by D and taking the orbifold by pulling back the canonical (G, X)-structure on X to O quotient. The inverse map is derived from the construction of the developing pair, hence µ is a bijection. This gives a topology on Def(O). 3.5. Ehresmann-Thurston principle. One of the most important results in this area is the following theorem first stated for closed manifolds. However, it can be easily generalized to closed orbifolds. There exist many proofs of this theorem for manifolds; see Canary-EpsteinGreen [27], Lok [80] following John Morgan, Bergeron-Gelander [19], Goldman [55]. For a proof for orbifolds, see Choi [31], which is a slight modification of the proof for manifolds. Suppose that G is the real points of a reductive algebraic group defined over R. A representation ρ ∶ π1orb (O) → G is stable when ρ is reductive and the centralizer of ρ is finite.6 Denote by Homst (π1orb (O), G) the space of stable representations. It is shown in JohnsonMillson [66] that this is an open subset of Hom(π1orb (O), G) and that the action of G on Homst (π1orb (O), G) is proper. Denote by Def st 2 (O) the space of (G, X)-structures on O whose holonomy representation is stable. Theorem 3.2 (Ehresmann-Thurston principle). Let O be a closed orbifold. Then the map Def ′2 (O) → Hom(π1orb (O), G) and st orb Def st 2 (O) → Hom (π1 (O), G)/G induced by (D, ρ) → ρ are local homeomorphisms. This principle means that sufficiently nearby (G, X)-structures are completely determined by their holomony representations. 4. A starting point for convex projective structures 4.1. Convexity in the projective sphere or in the projective space. Let V be a real vector space of dimension d + 1. Consider the action of R∗+ on V by homothety, and the projective sphere S(V ) = (V ∖ {0}) / R∗+ = {rays of V }. Of course, S(V ) is the 2-fold cover of the real projective space P(V ). The canonical projection map V ∖ {0} → S(V ) is denoted by S. A convex cone C is sharp if C does not contain an affine line. A subset C of S(V ) is convex (resp. properly convex) if the subset S−1 (C) ∪ {0} of V is a convex cone (resp. sharp convex 6A representation ρ ∈ Hom(π1orb (O), G) being stable is equivalent to the fact that the image of ρ is not contained in any parabolic subgroup of G (see Johnson-Millson [66]) 10 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS cone). Given a hyperplane H of S(V ), we call the two connected components of S(V ) ∖ H affine charts. An open set Ω ⫋ S(V ) is convex (resp. properly convex) if and only if there exists an affine chart A such that Ω ⊂ A (resp. Ω ⊂ A) and Ω is convex in the usual sense in A. A properly convex set Ω is strictly convex if every line segment in ∂Ω is a point. All these definitions can be made for subset of P(V ). The projective space is more common but the projective sphere allows to get rid of some technical issues. It will be clear from the context whether our convex domain is inside S(V ) or P(V ). The group SL± (V ) of linear automorphisms of V of determinant ±1 is identified to the group of automorphisms of S(V ). A properly convex projective structure on an orbifold O is a (PGL(V ), P(V ))-structure (or a (SL± (V ), S(V ))-structure) whose developing map is a diffeomorphism onto a properly convex subset of RPd (or Sd ). We refer the reader to Section 1 of Marquis [88] to see the correspondence between properly convex (PGL(V ), P(V ))-structures and properly convex (SL± (V ), S(V ))-structures (see also p.143 of Thurston [99]). From now on, C(O) will denote the space of properly convex projective structures on an orbifold O. This is a subspace of the deformation space of real projective structures on O. 4.2. Hilbert geometries. On every properly convex domain Ω, there exists a distance dΩ on Ω defined using the cross ratios: take two points x ≠ y ∈ Ω and draw the line between them. This line intersects the boundary ∂Ω of Ω in two points p and q. We assume that x is between p and y. If [p ∶ x ∶ y ∶ q] denotes the cross ratio of p, x, y, q, then the following formula defines a metric (see Figure 1): 1 dΩ (x, y) = log ([p ∶ x ∶ y ∶ q]), for every x, y ∈ Ω. 2 This metric gives to Ω the same topology as the one inherited from S(V ). The metric space (Ω, dΩ ) is complete and the closed balls in Ω are compact. The group Aut(Ω) acts on Ω by isometries, and therefore properly. p+ v y p p− q x Ω Figure 1. The Hilbert metric This metric is called the Hilbert metric and it can also be defined by a Finsler norm on the tangent space Tx Ω at each point x of Ω: Let v be a vector of Tx Ω. The quantity DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 11 d ∣ dt t=0 dΩ (x, x + tv) defines a Finsler norm FΩ (x, v) on Tx Ω. Let us choose an affine chart A containing Ω and a Euclidean norm ∣ ⋅ ∣ on A. If p+ (resp. p− ) is the intersection point of ∂Ω with the half-line determined by x and v (resp. −v), and ∣ ab ∣ is the distance between two points a, b of A (see Figure 1), then we obtain FΩ (x, v) = d ∣v∣ ⎛ 1 1 ⎞ . ∣ dΩ (x, x + tv) = + − dt t=0 2 ⎝ ∣xp ∣ ∣xp+ ∣ ⎠ The regularity of the Finsler norm is the same as the regularity of the boundary ∂Ω. The Finsler structure gives rise to an absolutely continuous measure µΩ with respect to Lebesgue measure, called the Busemann measure. If Ω is an ellipsoid, then (Ω, dΩ ) is the projective model of the hyperbolic space. More generally, if Ω is round, i.e. strictly convex with C 1 -boundary, then the metric space (Ω, dΩ ) exhibits some hyperbolic behaviour even though (Ω, dΩ ) is not Gromov-hyperbolic.7 A properly convex domain Ω is a polytope if and only if (Ω, dΩ ) is bi-Lipschitz to the Euclidean space [101]. If Ω is the projectivization of the space of real positive definite symmetric m × m matrices, then Aut(Ω) = SLm (R). Convex projective structures are therefore a special kind of (G, X)-structures, whose golden sisters are hyperbolic structures and whose iron cousins are higher-rank structures, i.e. the (G, G/K)-structures where G is a semi-simple Lie group without compact factor of (real) rank ⩾ 2, K is the maximal compact subgroup of G and G/K is the symmetric space of G. The above fact motivated an interest in convex projective structures. However, it is probable that this is not the main reason why convex projective structures are interesting. The main justification is the following theorem. 4.3. Koszul-Benoist’s theorem. Recall that the virtual center of a group Γ is the subset in Γ consisting of the elements whose centralizer is of finite index in Γ. It is easy to check that the virtual center of a group is a subgroup, and the virtual center of a group is trivial if and only if every subgroup of finite index has a trivial center. Theorem 4.1 (Koszul [75], Benoist [13]). Let O be a closed orbifold of dimension d admitting a properly convex real projective structure. Suppose that the group π1orb (O) has a trivial virtual center. Then the space C(O) of properly convex projective structures on O corresponds to a union of connected components of the character variety χ(π1orb (O), SL±d+1 (R)). Proof. Let Γ = π1orb (O) and G = SL±d+1 (R). It was proved by Koszul [75] that the space of holonomy representations of elements of C(O) is open in χ(Γ, G). The closedness follows from Benoist [13] together with the following explanation. Assume that ρ ∈ χ(Γ, G) is an algebraic limit of a sequence of holonomy representations ρn ∈ χ(Γ, G) of elements of C(O). This means that ρn (Γ) acts on a properly convex domain Ωn such that Ωn /ρn (Γ) is orbifolddiffeomorphic to O. Then the representation ρ is discrete, faithful and irreducible and ρ(Γ) acts on a properly convex domain Ω by Theorem 1.1 of Benoist [13]. We need to show that 7If Ω is strictly convex and ∂Ω is of class C 2 with a positive Hessian, then (Ω, dΩ ) is Gromov-hyperbolic (see Colbois-Verovic [41]), but unfortunately the convex domain we are interested in will be at most round, except the ellipsoid. See also the necessary and sufficient condition of Benoist [11]. 12 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS the quotient orbifold Q ∶= Ω/ρ(Γ) is orbifold-diffeomorphic to O: The openness implies that there is a neighbourhood U of ρ in χ(Γ, G) consisting of holonomy representations of elements of C(Q). Since ρn ∈ U for sufficiently large n, there is a properly convex domain Ω′n such that Ω′n /ρn (Γ) is diffeomorphic to Q. Proposition 3 of Vey [102] implies that Ω′n = Ωn (see also Theorem 3.1 of Choi-Lee [38] and Proposition 2.4 of Cooper-Delp [42]). Since Q ≈ Ω′n /ρn (Γ) = Ωn /ρn (Γ) ≈ O, the orbifold Q is orbifold-diffeomorphic to O, completing the proof that ρ is the holonomy representation of an element of C(O).  Remark 4.2. The closedness of Theorem 4.1 was proved by Choi-Goldman [35] in dimension d = 2, and by Kim [72] in the case that Γ is a uniform lattice of PSO3,1 (R). The condition that the virtual center of a group is trivial is transparent as follows: Proposition 4.3 (Corollary 2.13 of Benoist [13]). Let Γ be a discrete subgroup of PGLd+1 (R). Suppose that Γ acts on a properly convex domain Ω in RPd and Ω/Γ is compact. Then the following are equivalent. ● The virtual center of Γ is trivial. ● Every subgroup of finite index of Γ has a finite center. ● Every subgroup of finite index of Γ is irreducible in PGLd+1 (R).8 ● The Zariski closure of Γ is semisimple. ● The group Γ contains no infinite nilpotent normal subgroup. ● The group Γ contains no infinite abelian normal subgroup. 4.4. Duality between convex real projective orbifolds. We start with linear duality. Every sharp convex open cone C of a real vector space V gives rise to a dual convex cone C ⋆ = {ϕ ∈ V ⋆ ∣ ϕ∣C∖{0} > 0}. It can be easily verified that C ⋆ is also a sharp convex open cone and that C ⋆⋆ = C. Hence, the duality leads to an involution between sharp convex open cones. Now, consider the “projectivization” of linear duality: If Ω is a properly convex domain of P(V ) and CΩ is the cone of V over Ω, then the dual Ω⋆ of Ω is P(CΩ⋆ ). This is a properly convex domain of P(V ⋆ ). In a more intrinsic way, the dual Ω⋆ is the set of hyperplanes H of V such that P(H) ∩ Ω = ∅. Since there is a correspondence between hyperplanes and affine charts of the projective space, the dual Ω⋆ can be defined as the space of affine charts A of RPd such that Ω is a bounded subset of A. The second interpretation offers us a map Ω⋆ → Ω: namely to A ∈ Ω⋆ we can associate the center of mass of Ω in A. The map is well defined since Ω is a bounded convex domain of A. In fact, Vinberg showed that this map is an analytic diffeomorphism (see Vinberg [103] or Goldman [56]), so we call it the Vinberg duality diffeomorphism. Finally, we can bring the group in the playground. Recall that the dual representation ∶ Γ → PGL(V ⋆ ) of a representation ρ ∶ Γ → PGL(V ) is defined by ρ⋆ (γ) = t ρ(γ −1 ), i.e. the dual projective transformation of ρ(γ). All the constructions happen in projective geometry, therefore if a representation ρ preserves a properly convex domain Ω, then the ρ⋆ 8We call Γ strongly irreducible if every finite index subgroup of Γ is irreducible. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 13 dual representation ρ⋆ preserves the dual properly convex domain Ω⋆ . Even more, if we assume the representation ρ to be discrete, then the Vinberg duality diffeomorphism induces a diffeomorphism between the quotient orbifolds Ω/ρ(Γ) and Ω⋆ /ρ⋆ (Γ). 5. The existence of deformations or exotic structures 5.1. Bending construction. Johnson and Millson [66] found an important class of deformations of convex real projective structures on an orbifold. The bending construction was introduced by Thurston to deform (PSO2,1 (R), H2 )-structures on a surface into (PSO3,1 (R), S2 )structures, i.e. conformally flat structures, and therefore in particular to produce quasiFuchsian groups. This was extended by Kourouniotis [76] to deform (PSOd,1 (R), Hd )-structures on a manifold into (PSOd+1,1 (R), Sd )-structures. Johnson and Millson indicated several other deformations, all starting from a hyperbolic structure on an orbifold. However we will focus only on real projective deformations. Just before that we stress that despite the simplicity of the argument, the generalization is not easy. Goldman and Millson [58], for instance, show that there exists no non-trivial9 deformation of a uniform lattice of SUd,1 into SUd+1,1 . Let sld+1 (R) be the Lie algebra of SLd+1 (R), and let O be a closed properly convex projective orbifold. Suppose that O contains a two-sided totally geodesic suborbifold Σ of codimension 1. For example, all the hyperbolic manifolds obtained from standard arithmetic lattices of PSOd,1 (R), up to a finite cover, admit such a two-sided totally geodesic hypersurface (see Section 7 of Johnson-Millson [66] for the construction of standard arithmetic lattices). Let Γ = π1orb (O) and A = π1orb (Σ). Recall that the Lie group SLd+1 (R) acts on the Lie algebra sld+1 (R) by the adjoint action. Lemma 5.1 (Johnson-Millson [66]). Let ρ ∈ Hom(Γ, SLd+1 (R)). Suppose that ρ(A) fixes an element x1 in sld+1 (R) and that x1 is not invariant under Γ. Then Hom(Γ, SLd+1 (R)) contains a non-trivial curve (ρt )t∈(−ε,ε) , ε > 0, with ρ0 = ρ, i.e. the curve is transverse to the conjugation action of SLd+1 (R). Theorem 5.2 (Johnson-Millson [66], Koszul [75], Benoist [12, 13]). Suppose that a closed hyperbolic orbifold O contains r disjoint totally geodesic suborbifolds Σ1 , . . . , Σr of codimensionone. Then the dimension of the space C(O) at the hyperbolic structure is greater than or equal to r. Moreover, the bending curves lie entirely in C(O) and all the properly convex structures on O are strictly convex. We say that a convex domain Ω of the real projective space is divisible (resp. quasi-divisible) if there exists a discrete subgroup Γ of Aut(Ω) such that the action of Γ on Ω is cocompact (resp. of finite Busemann covolume). Theorem 5.3 (Johnson-Millson [66], Koszul [75], Benoist [12]). For every integer d ⩾ 2, there exists a non-symmetric divisible strictly convex domain of dimension d. Remark 5.4. Kac and Vinberg [67] made the first examples of non-symmetric divisible convex domains of dimension 2 using Coxeter groups (see Section 7). Remark 5.5. The third author with Ballas [86, 8] extended Theorem 5.3 to non-symmetric quasi-divisible (not divisible) convex domains. 9A deformation of a representation ρ ∶ Γ → G is trivial if it is a conjugation of ρ in G. 14 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS 5.2. The nature of the exotics. Now we have seen the existence of non-symmetric divisible convex domains in all dimensions d ⩾ 2. We first remark that so far all the divisible convex domains built are round if they are not the products of lower dimensional convex domains. So we might want to know if we can go further, specially if we can find indecomposable divisible convex domains that are not strictly convex. The first result in dimension 2 is negative:10 5.2.1. Dimension 2 : Kuiper-Benzécri’s Theorem. Theorem 5.6 (Kuiper [77], Benzécri [17], Marquis [87]). A quasi-divisible convex domain of dimension 2 is round, except the triangle. The next result in dimension 3 is positive: 5.2.2. Dimension 3 : Benoist’s Theorem. First, we can classify the possible topology for closed convex projective 3-manifold or 3-orbifold. Theorem 5.7 (Benoist [14]). If a closed 3-orbifold O admits an indecomposable11 properly convex projective structure, then O is topologically the union along the boundaries of finitely many 3-orbifolds each of which admits a finite-volume hyperbolic structure on its interior. Second, these examples do exist. Theorem 5.8 (Benoist [14], Marquis [84], Ballas-Danciger-Lee [7]). There exists an indecomposable divisible properly convex domain Ω of dimension 3 which is not strictly convex. Moreover, every line segment in ∂Ω is contained in the boundary of a properly embedded triangle.12 At the time of writing this survey, Theorem 5.7 is valid only for divisible convex domain. However, Theorem 5.8 is true also for quasi-divisible convex domains which are not divisible (see Marquis [89]). The presence of properly embedded triangles in the convex domain is related to the existence of incompressible Euclidean suborbifolds on the quotient orbifold. Benoist and the third author made examples using Coxeter groups and a work of Vinberg [105]. We will explore more this technique in Section 7. Ballas, Danciger and the second author [7] found a sufficient condition under which the double of a cusped hyperbolic three-manifold admits a properly convex projective structure, to produce the examples. In order to obtain the quasi-divisible convex domains in [89], the third author essentially keeps the geometry of the cusps. In other words, the holonomy of the cusps preserves an ellipsoid as do the cusps of finite-volume hyperbolic orbifolds. From this perspective, there is a different example whose cusps vary in geometry. Theorem 5.9 (Ballas [6], Choi [33]). There exists an indecomposable quasi-divisible (necessarily not divisible) properly convex domain Ω of dimension 3 which is not strictly convex nor with C 1 -boundary such that the quotient is homeomorphic to a hyperbolic manifold of finite volume. 10The word “positive/negative” in this subsection reflects only the feeling of the authors. convex projective orbifold is indecomposable if its holonomy representation is strongly irreducible. 12A simplex ∆ in Ω is properly embedded if ∂∆ ⊂ ∂Ω. 11A DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 15 More precisely, the example of Ballas is an explicit convex projective deformation of the hyperbolic structure on the figure-eight knot complement. Note that the first author gave such an example in Chapter 8 of [33] earlier without verifying its property. We do not survey results about cofinite volume action of discrete groups on Hilbert geometries, and we refer the reader to: Marquis [87, 89], Crampon-Marquis [46], Cooper-LongTillmann [45], Choi [33]. 5.2.3. Orbifolds of dimension 4 and beyond. Until now, there are only three sources for nonsymmetric divisible convex domains of dimension d ⩾ 4: from the “standard” bending of Johnson-Millson [66], from the “clever” bending of Kapovich [68], and using Coxeter groups. The last method was initiated by Benoist [14, 15] and extended by the three authors [40]: Theorem 5.10 (Benoist [14], Choi-Lee-Marquis [40]). For d = 4, . . . , 7, there exists an indecomposable divisible convex domain Ω of dimension d which is not strictly convex nor with C 1 -boundary such that Ω contains a properly embedded (d−1)-simplex. Moreover, the quotient is homeomorphic to the union along the boundaries of finitely many d-orbifolds each of which admits a finite-volume hyperbolic structure on its interior. Other examples were built by Benoist in dimension 4 and by Kapovich in every dimension: Theorem 5.11 (Benoist [15], Kapovich [68]). For d ⩾ 4, there exists a divisible convex domain Ω of dimension d such that (Ω, dΩ ) is Gromov-hyperbolic but it is not quasi-isometric to a symmetric space. In particular, Ω is strictly convex. However it is not quasi-isometric to the hyperbolic space Hd . The three authors recently construct somehow different examples: Theorem 5.12 (Choi-Lee-Marquis [39]). For d = 4, . . . , 7, there exists an indecomposable divisible convex domain Ω of dimension d which is not strictly convex nor with C 1 -boundary such that Ω contains a properly embedded (d − 2)-simplex but does not contain a properly embedded (d − 1)-simplex. 6. Real projective surfaces Another motivation for studying convex projective structures is that these structures are just the right building blocks for all projective structures on closed surfaces. 6.1. Affine and projective structures on tori. 6.1.1. Classification of affine surfaces. Compact affine surfaces are topologically restrictive: Theorem 6.1 (Benzécri [17]). If S is a compact affine surface with empty or geodesic13 boundary, then χ(S) = 0. In the early 1980s, Nagano and Yagi [93] classified the affine structures on a torus and an annulus with geodesic boundary; see Benoist [10] for a modern viewpoint and Baues [9] for an extensive survey on this topic. Let Q be the closed positive quadrant of R2 and let U be the closed upper-half plane. An elementary affine annulus is the quotient of Q ∖ {0} by the group generated by a diagonal 13An arc of a projective (or affine) surface is geodesic if it has a lift which is developed into a line segment. 16 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS linear automorphism with all eigenvalues greater than 1 or the quotient of U ∖ {0} by the group generated by a linear automorphism ( λ0 µλ ) with λ > 1. It is indeed an affine annulus with geodesic boundary. An affine torus A is complex if its affine structure comes from a (C ∖ {0}, C∗ )-structure (see page 112 of Thurston [99]). Theorem 6.2 (Nagano-Yagi [93]). If A is a compact affine surface with empty or geodesic boundary, then one of the following holds : ● The universal cover of A is either a complete affine space, a half-affine space, a closed parallel strip or a quadrant. ● The surface A is a complex affine torus. ● The surface A is decomposed into elementary affine annuli along simple closed geodesics. 6.1.2. Classification of projective tori. Let γ be an element of SL3 (R). A matrix γ is positive hyperbolic if γ has three distinct positive eigenvalues, that is, γ is conjugate to ⎛ λ 0 0 ⎞ ⎜ 0 µ 0 ⎟ ⎝ 0 0 ν ⎠ (λµν = 1 and 0 < λ < µ < ν). A matrix γ is planar if γ is diagonalizable and it has only two distinct eigenvalues. A matrix γ is quasi-hyperbolic if γ has only two distinct positive eigenvalues and it is not diagonalizable, that is, γ is conjugate to (2) ⎛ λ 1 0 ⎞ ⎜ 0 λ 0 ⎟ ⎝ 0 0 λ−2 ⎠ (0 < λ ≠ 1). A matrix γ is a projective translation (resp. parabolic) if γ is conjugate to ⎛ 1 1 0 ⎞ ⎛ 1 1 0 ⎞ ⎜ 0 1 0 ⎟ (resp. 0 1 1 ). ⎝ 0 0 1 ⎠ ⎝ 0 0 1 ⎠ These types of matrices represent all conjugacy classes of non-trivial elements of SL3 (R) with positive eigenvalues. Let ϑ be a positive hyperbolic element of SL3 (R) with eigenvalues λ < µ < ν. It is easy to describe the action of ϑ on the projective plane. This action preserves three lines meeting at the three fixed points. The fixed point r (resp. s, a) corresponding to the eigenvector for λ (resp. µ, ν) is said to be repelling (resp. saddle, attracting). Let △ be a triangle with vertices r, s, a.14 An elementary annulus of type I is one of the two15 real projective annuli (△ ∪ ar○ ∪ as○ )/⟨ϑ⟩ and (△ ∪ ar○ ∪ rs○ )/⟨ϑ⟩ (see Figure 2). These annuli are in fact compact since we can find a compact fundamental domain. We call the image of ar○ in the annulus a principal closed geodesic, and the image of as○ and the image of rs○ weak closed geodesics. 14To be precise, in RP2 (resp. S2 ) there are four (resp. eight) different choices for △ and a unique choice (resp. two choices) for each r, s, a. However, once the triangle △ in S2 is chosen, the points r, s, a of S2 are uniquely determined. Of course, this choice does not affect the later discussion. 15 If P = △∪as○ ∪rs○ , then the space P /⟨ϑ⟩ is not Hausdorff: Let x be a point of as○ , and let y be a point of ○ rs . For any neighborhood U, V of x, y, respectively, in P , there exists N ∈ N such that ϑ−N (U) ∩ ϑN (V) ≠ ∅. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 17 a a r r s a r a s a r a Figure 2. Elementary annuli of type I and type II Now, let ϑ be a quasi-hyperbolic element conjugate to the matrix (2) with λ > 1. In this case, it is easier to give a description of ϑ in S2 than in RP2 , hence we work in S2 . The fixed point a (resp. r) corresponding to the eigenvalue λ (resp. λ−2 ) is attracting (resp. repelling). Let C be the line on which the action of ϑ is parabolic, and let β be the invariant segment in C with endpoints a and −a, the antipodal point of a, such that, for each point x ∈ β ○ , the sequence ϑn (x) converges to a. Let L be an open lune bounded by β, ar, −ar. An elementary annulus of type II is one of two real projective annuli (L ∪ ar○ ∪ −ar○ )/⟨ϑ⟩ and (L ∪ β ○ ∪ ar○ )/⟨ϑ⟩ (see Figure 2). These annulli are also compact with geodesic boundary. We call the image of ar○ and of −ar○ in the annulus a principal closed geodesic, and the image of β ○ a weak closed geodesic. By pasting the boundaries of finitely many compact elementary annuli, we obtain an annulus or a torus. The gluing of course requires that boundaries are either both principal or both weak, and their holonomies are conjugate to each other. Theorem 6.3 (Goldman [52]). If T is a projective torus or a projective annulus with geodesic boundary, then T is an affine torus or T can be decomposed into elementary annuli. 6.2. Automorphisms of convex 2-domain. A 2-orbifold O is of finite-type if the underlying space of O is a surface of finite-type16 and the singular locus of O is a union of finitely many suborbifolds of dimension 0 or 1. An element γ of π1orb (O) is peripheral if γ is isotopic to an element of π1orb (O ∖ (C ∪ ∂O)) for every compact subset C of O ∖ ∂O. 16A surface of finite-type is a compact surface with a finite number of points removed. 18 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS Let C be a properly convex subset of S2 and let Γ be a discrete subgroup of SL3 (R) acting properly discontinuously on C. A closed geodesic g in S = C/Γ is principal when the holonomy γ of g is positive hyperbolic or quasi-hyperbolic, and the lift of g is the geodesic segment in C connecting the attracting and repelling fixed points of γ. In addition, if γ is positive hyperbolic, then g is said to be h-principal. The following theorem generalizes well-known results of hyperbolic structures on surfaces, and is essential to understand the convex real projective 2-orbifolds. The nonorientable orbifold version exists but is a bit more complicated to state (see Choi-Goldman [36]). Theorem 6.4 (Kuiper [77], Choi [28], Marquis [87]). Let S = C/Γ be an orientable properly convex real projective 2-orbifold of finite type with empty or geodesic boundary, for a properly convex subset C ⊂ S2 . We denote by Ω the interior of C. Suppose that Ω is not a triangle. (1) An element has finite order if and only if it fixes a point in Ω. (2) Each infinite-order element γ of Γ is positive hyperbolic, quasi-hyperbolic or parabolic. (3) If an infinite-order element γ is nonperipheral, then γ is positive hyperbolic and a unique closed geodesic g realizes γ. Moreover, g is principal and any lift of g is contained in Ω. If γ is represented by a simple closed curve in S, then g is also simple. (4) A peripheral positive hyperbolic element is realized by a unique principal geodesic g, and either all the lifts of g are contained in Ω or all the lifts of g are in C ∩ ∂Ω. (5) A quasi-hyperbolic element γ is peripheral and is realized by a unique geodesic g in C. Moreover, g is principal and any lift of g is in C ∩ ∂Ω. (6) A parabolic element γ is peripheral and is realized by the projection of E ∖ {p} where E is a γ-invariant ellipse whose interior is inside Ω and p ∈ ∂Ω is the unique fixed point of γ. Note that in the fourth item, the closed geodesics homotopic to g may not be unique. 6.3. Convex projective structures on surfaces. Let S be a compact surface with or without boundary. When S has boundary, in general, the holonomy of a convex projective structure on S does not determine the structure. More precisely, there exists a convex projective structure whose holonomy preserves more than one convex domain. Therefore we need to make some assumptions on the convex projective structure in order to avoid this pgb problem hence we consider the subspace, denoted by C (S), in C(S) of convex projective structures on S for which each boundary component is a principal geodesic. Theorem 6.5 (Goldman [57]). If S is a closed surface of genus > 1, then the space C(S) is homeomorphic to an open cell of dimension −8χ(S). The following two propositions illustrate the proof of Theorem 6.5. Recall that for a Lie group G, a fiber bundle π ∶ P → X is G-principal if there exists an action of G on P such that G preserves the fibers and acts simply transitively on them. Let c be a non-peripheral simple closed curve in S. The complement of c in S can be connected, or a disjoint union of S1 and S2 . In the first case, the completion S ′ of S ∖ c has pgb two boundary components c′1 and c′2 corresponding to c. Denote by Cc (S ′ ) the subspace of pgb structures in C (S ′ ) satisfying that the holonomies of c′1 and c′2 are both positive hyperbolic and conjugate to each other. In the second case, the completion of each Si , i = 1, 2, has DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 19 a boundary component corresponding to c. Denote by Cc (S1′ ) ⊠c Cc (S2′ ) the subspace of pgb pgb structures (P1 , P2 ) in C (S1′ ) × C (S2′ ) whose holonomies corresponding to c are positive hyperbolic and conjugate to each other. pgb pgb Proposition 6.6 (Goldman [57], Marquis [85]). Let S be a compact surface with or without boundary such that χ(S) < 0, and let c be a non-peripheral simple closed curve in S. If S ∖ c pgb pgb is connected (resp. a disjoint union of S1 and S2 ), then the forgetful map C (S) → Cc (S ′ ) pgb pgb pgb (resp. C (S) → Cc (S1′ ) ⊠c Cc (S2′ )) is an R2 -principal fiber bundle. Proposition 6.6 should be compared with Lemma 5.1. Since the holonomy γ of a nonperipheral simple closed curve c is positive hyperbolic, the centralizer of γ is 2-dimensional. The first gluing parameter is the twist parameter like in hyperbolic geometry, and the second gluing parameter is the bending parameter we obtain in view of Lemma 5.1. Next, we need to understand the convex projective structures on a pair of pants. Assume that γ is an element of SL3 (R) with positive eigenvalues. We denote by λ(γ) the smallest eigenvalue of γ and by τ (γ) the sum of the two other eigenvalues, and we call the pair (λ(γ), τ (γ)) the invariant of γ. The map [γ] ↦ (λ(γ), τ (γ)) is a homeomorphism between the space of conjugacy classes of positive hyperbolic or quasi-hyperbolic elements γ of SL3 (R) and the space 2 1 R ∶= {(λ, τ ) ∈ R2 ∣ 0 < λ < 1, √ ⩽ τ ⩽ λ + 2 } λ λ (see Figure 3). Note that (1, 2) ∉ R corresponds to the conjugacy class of parabolic elements. 20 15 10 5 0.0 0.2 0.4 0.6 0.8 1.0 Figure 3. The region R between the graphs y = √2 x and y = x + x12 Let S be a compact surface with n (oriented) boundary components ci , i = 1, . . . , n. Let us define a map pgb R∂S ∶ C (S) → Rn from each structure to the n-tuples of invariants (λ(γi ), τ (γi )) of the holonomy γi of ci . Proposition 6.7 (Goldman [57], Marquis [85]). If P is a pair of pants, then the map R∂P ∶ Cpgb (P ) → R3 is an R2 -principal fiber bundle, and the interior of Cpgb (P ) is exactly the space 20 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS of convex projective structures with h-principal geodesic boundary. In particular it is an open cell of dimension 8. Now consider a surface S of finite type with ends. A properly convex projective structure on S is of relatively finite volume if for each end of S, there exists an end neighborhood pgb pgb V such that µS (V) < ∞. Let Cgf (S) denote the subspace in C (S) of convex projective structures with principal geodesic boundary and of relatively finite volume. The following theorem generalizes to such structures: Theorem 6.8 (Fock-Goncharov [50], Theorem 3.7 of Marquis [85]). Let S be a surface of finite-type. If b is the number of boundary components of S and p is the number of ends of S, pgb then the space Cgf (S) is a manifold with corner homemorphic to R16g−16+6p+7b × [0, 1]b , and pgb the interior of Cgf (S) is exactly the space of structures with h-principal geodesic boundary and of relatively finite volume. 6.4. The convex decomposition of projective surfaces of genus > 1. Now let us consider compact real projective surfaces with geodesic boundary. Theorem 6.9 (Choi [28, 29]). Let S be a compact projective surface with geodesic boundary. If χ(S) < 0, then S can be decomposed along simple closed geodesics into convex projective subsurfaces with principal geodesic boundary and elementary annuli. The annuli with principal geodesic boundary whose holonomies are positive hyperbolic (resp. quasi-hyperbolic) were classified by Goldman [52] (resp. Choi [29]). If A is an annulus with quasi-hyperbolic principal geodesic boundary, then by Proposition 5 of Choi [30] only one boundary can be identified with a boundary of a convex real projective surface with principal geodesic boundary. Hence, if a compact real projective surface S has quasi-hyperbolic holonomy for a closed curve, then S must have a boundary. Remark 6.10. Given a convex projective surface S = Ω/Γ with a h-principal boundary and an elementary annulus A, we can obtain a new projective surface S ′ by identifying the respective boundary components of S and A. The surface S ′ is still convex since the union of Ω and the triangles given by the universal cover of A is convex. 6.5. Projective structures on a closed surface of genus > 1. Let S be a closed surface of genus greater than 1 and let P(S) denote the space of real projective structures on S. For each connected component of P(S), any two elements share the same decomposition up to isotopy, given by Theorem 6.9. Let S(S) denote the collection of mutually disjoint isotopy classes of non-trivial simple closed curves, and let F2+,even be the set of elements of the free semigroup on two generators whose word-lengths are even. In [52] Goldman constructs a map P(S) → S(S) × F2+,even that describes the gluing patterns of elementary annuli of type I (with h-principal geodesic boundary) in a projective structure P on S. Finally, by removing all the annuli from P and reattaching it, we obtain a convex projective structure on S. Theorem 6.11 (Choi [28, 29], Goldman [57]). Let S be a closed surface with χ(S) < 0. Then each fiber of the map P(S) → S(S) × F2+,even can be identified with C(S). In particular, P(S) is homeomorphic to a disjoint union of countably many open cells of dimension 16(g − 1). DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 21 6.6. Convex projective orbifolds of negative Euler characteristic. Every compact 2orbifold Σ is obtained from a surface with corners by making some arcs mirrors and putting cone-points and corner-reflectors in a locally finite manner. An endpoint of a mirror arc can be either in a boundary component of Σ or a corner-reflector in Σ that should be an endpoint of another mirror arc. Moreover, the smooth topology of a 2-orbifold is determined by the underlying topology of the surface with corners, the number of cone-points of order q ∈ N ∖ {0, 1}, the number of corner-reflectors of order r ∈ N ∖ {0, 1}, and the boundary patterns of the mirror arcs. A full 1-orbifold is a segment with two mirror endpoints. Let Σ be a compact 2-orbifold with m cone points of order qi , i = 1, . . . , m, and n corner-reflectors of order rj , j = 1, . . . , n, and nΣ boundary full 1-orbifolds. The orbifold Euler characteristic of Σ is m χ(Σ) = χ(XΣ ) − ∑ (1 − i=1 1 1 n 1 1 ) − ∑ (1 − ) − nΣ . qi 2 j=1 rj 2 This is called the generalized Riemman-Hurwitz formula (see Section 3.3 for the definition of the orbifold Euler characteristic). Theorem 6.12 (Thurston [100]). Let Σ be a compact 2-orbifold of negative orbifold Euler characteristic with the underlying space XΣ . Then the deformation space T (Σ) of hyperbolic structures on Σ is a cell of dimension −χ(XΣ ) + 2m + n + 2nΣ where m is the number of conepoints, n is the number of corner-reflectors and nΣ is the number of boundary full 1-orbifolds. In order to understand the deformation spaces of convex projective structures on surfaces, we saw that it is important to study the convex projective structures on a pair of pants, which is the most “elementary” surface. Similarly, in the case of 2-orbifolds, we should firstly understand the elementary 2-orbifolds. Let us discuss the process of splitting and sewing of 2-orbifolds. Note that orbifolds always have a path-metric. For example, we can define a notion of Riemannian metric on an orb̃i /Γi , there exist Γi -invariant Riemannian ifold Σ, i.e. for all coordinate neighborhoods Ui ≈ U ̃i which are compatible each other (see Choi [34]). Let c be a simple closed curve metrics on U or a full 1-orbifold in the interior17 of a 2-orbifold Σ and let Σ̂ be the completion of Σ ∖ c with respect to the path-metric induced from one on Σ. We say that Σ̂ is obtained from the splitting of Σ along c. Conversely, if ĉ is the union of two boundary components of Σ̂ corresponding to c, then Σ is obtained from sewing Σ̂ along ĉ. An elementary 2-orbifold is a compact 2-orbifold of negative orbifold Euler characteristic which we cannot split further along simple closed curves or full 1-orbifolds into suborbifolds. We assume in this subsection that our orbifolds are of negative orbifold Euler characteristic. The following is the classification of elementary 2-orbifolds (see Figure 4). Arcs with/without dotted arcs next to them indicate boundary/mirror components, respectively, and black points indicate singular points. We can obtain the orbifolds (Pj), j = 2, 3, 4, from changing the boundary components of (P1) to cusps and then to elliptic points, considering them as 17The interior of an orbifold Σ is Σ ∖ ∂Σ. 22 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS hyperbolic surfaces with singularities. For each j = 1, . . . , 4, the orbifold (Dj) (resp. (Aj)) is the quotient orbifold of (Pj) by an order-two involution preserving (resp. switching a pair of) boundary components or cone-points. Note that the underlying space of (P1) is a pair of pants, the ones of (P2), (A1) and (A2) are closed annuli, the ones of (P3), (A3), (A4) and (D1)-(D4) are closed disks, and the one of (P4) is a sphere. (P1) (P2) (P3) (A1) (A2) (A3) (D1) (D2) (D3) (P4) (A4) (D4) Figure 4. Elementary orbifolds (P1) (P2) (P3) (P4) A pair of pants (χ = −1). An annulus with a cone-point of order p (χ = p1 − 1). A disk with two cone-points of orders p, q (χ = p1 + 1q − 1). A sphere with three cone-points of order p, q, r (χ = p1 + 1q + 1r − 1). (A1) (A2) (A3) (A4) An annulus with a boundary circle, a boundary arc and a mirror arc (χ = − 21 ). 1 An annulus with a boundary circle and a corner-reflector of order p (χ = 2p − 12 ). A disk with a boundary arc, a mirror arc and a cone-point of order p (χ = p1 − 12 ). 1 A disk with a corner-reflector of order p and a cone-point of order q (χ = 2p + 1q − 12 ). (D1) A disk with three mirror arcs and three boundary arcs (χ = − 21 ). (D2) A disk with a corner-reflector of order p at which two mirror arcs meet, one more 1 mirror arc and two boundary arcs (χ = 2p − 12 ). 1 1 (D3) A disk with two corner-reflectors of order p, q, and a boundary arc (χ = 2p + 2q − 21 ). 1 1 + 2q + (D4) A disk with three corner-reflectors of order p, q, r and three mirror arcs (χ = 2p 1 1 2r − 2 ). Let Σ be a properly convex real projective 2-orbifold. A geodesic full 1-orbifold c in Σ is ̃ of Σ such that the double cover of c in Σ ̃ is a simple h-principal if there is a double cover Σ closed geodesic and is h-principal. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 23 Assume that every boundary component of Σ is h-principal. Let c be an oriented boundary component of Σ. We know from Section 6.3 that if c is homeomorphic to a circle, then the space Inv(c) of projective invariants of c is homeomorphic to R○ . However, if c is a full 1orbifold, then Inv(c) = R∗ because in that case the space C(c) of convex projective structures on the Coxeter 1-orbifold c is parametrised by the Hilbert length of c in the universal cover of c. Denoting by B(Σ) the set of boundary components of Σ, we define Inv(∂Σ) ∶= ∏ Inv(c) and Inv(∅) = {∗} is a singleton. c∈B(Σ) Proposition 6.13 (Choi-Goldman [36]). Let S be an elementary orbifold in Figure 4. The pgb map C (S)○ → Inv(∂S) is a fibration of an n-dimensional open cell over the k-dimensional open cell with l-dimensional open cell fiber. We list (n, k, l) below. (P1): (8, 6, 2) (Goldman [57]). (P2): (6, 4, 2) if there is no cone-point of order 2. Otherwise (4, 4, 0). (P3): (4, 2, 2) if there is no cone-point of order 2. Otherwise (2, 2, 0). (P4): (2, 0, 2) if there is no cone-point of order 2. Otherwise (0, 0, 0). (A1): (4, 3, 1). (A2): (3, 2, 1) if there is no corner-reflector of order 2. Otherwise (2, 2, 0). (A3): (2, 1, 1). (A4): (1, 0, 1) if there is no corner-reflector of order 2. Otherwise (0, 0, 0). (D1): (4, 3, 1). (D2): (3, 2, 1) if there is no corner-reflector of order 2. Otherwise (2, 2, 0). (D3): (2, 1, 1) if there is no corner-reflector of order 2. Otherwise (1, 1, 0). (D4): (1, 0, 1) if there is no corner-reflector of order 2. Otherwise (0, 0, 0). Finally, we can describe the deformation space of convex projective structures on closed 2-orbifolds. Theorem 6.14 (Choi-Goldman [36]). Let Σ be a closed 2-orbifold with χ(Σ) < 0. Then the space C(Σ) of convex projective structures on Σ is homeomorphic to a cell of dimension −8χ(XΣ ) + (6m − 2m2 ) + (3n − n2 ) where XΣ is the underlying space of Σ, m is the number of cone-points, m2 is the number of cone-points of order 2, n is the number of corner-reflectors, and n2 is the number of corner-reflectors of order 2. 7. Convex projective Coxeter orbifolds 7.1. Definitions. 7.1.1. Coxeter group. Let S be a finite set and denote by ∣S∣ the cardinality of S. A Coxeter matrix on S is an ∣S∣ × ∣S∣ symmetric matrix M = (Mst )s,t∈S with diagonal entries Mss = 1 and other entries Mst ∈ {2, 3, . . . , ∞}. The pair (S, M ) is called a Coxeter system. To a Coxeter system (S, M ) we can associate a Coxeter group W = WS : it is the group generated by S with the relations (st)Mst = 1 for all (s, t) ∈ S × S such that Mst ≠ ∞. If 24 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS (S, M ) is a Coxeter system, then for each subset T of S we can define the Coxeter subsystem (T, MT = (Mst )s,t∈T ). The Coxeter group WT can be thought of as a subgroup of WS since the canonical map from WT to WS is an embedding. We stress that the last sentence is not an obvious statement and it is, in fact, a corollary of Theorem 7.1. The Coxeter graph of a Coxeter system (S, M ) is the labelled graph where the set of vertices is S, two vertices s and t are connected by an edge st if and only if Mst ≠ 2, and the label of the edge st is Mst . A Coxeter system (S, M ) is irreducible when its Coxeter graph is connected. It is a little abusive but we also say that the Coxeter group W is irreducible. 7.1.2. Coxeter orbifolds. We are interested in d-dimensional Coxeter orbifolds whose underlying space is homeomorphic to a d-dimensional polytope18 P minus some faces, and whose singular locus is the boundary of P made up of mirrors. For the sake of clarity, facets are faces of codimension 1, ridges are faces of codimension 2 and proper faces are faces different from P and ∅. Choose a polytope P and a Coxeter matrix M on the set S of facets of P such that if two facets s and t are not adjacent,19 then Mst = ∞. When two facets s, t are adjacent, the ridge s ∩ t of P is said to be of order Mst . The first objects we obtain are the Coxeter system (S, M ) and the Coxeter group W . We now build an orbifold whose fundamental group is W and whose underlying topological space is the starting polytope P minus some faces: For each proper face f of P , let Sf = {s ∈ S ∣ f ⊂ s}. If Wf ∶= WSf is an infinite Coxeter group then the face f is said to be undesirable. Let P̂ be the orbifold obtained from P with undesirable faces removed, with facets as mirrors, with the remaining ridges s ∩ t as corner reflectors of orders Mst . We call P̂ a Coxeter d-orbifold. We remark that a Coxeter d-orbifold is closed if and only if for each vertex v of P , the Coxeter group Wv is finite. For example, let P be a polytope in X = Sd , Ed or Hd with dihedral angles submultiples of π. The uniqueness of the reflection across a hyperplane of X allows us to obtain a Coxeter (Isom(X), X)-orbifold P̂ from P . 7.1.3. Deformation spaces. Recall that C(P̂ ) denotes the deformation space of properly convex real projective structures on the Coxeter orbifold P̂ , that is, the space of projective structures on P̂ whose developing map is a diffeomorphism onto a properly convex subset in RPd . 7.2. Vinberg’s breakthrough. In this subsection we give a description of Vinberg’s results in his article [105]. An alternative treatment is given in Benoist’s notes [16]. 7.2.1. Groundwork. Let V be the real vector space of dimension d + 1. A projective reflection (or simply, reflection) σ is an element of order 2 of SL± (V ) which is the identity on a hyperplane H. All reflections are of the form σ = Id − α ⊗ b for some linear functional α ∈ V ⋆ and some vector b ∈ V with α(b) = 2. Here, the kernel of α is the subspace H of fixed points of σ and b is the eigenvector corresponding to the eigenvalue −1. 18We implicitly assume that all the polytopes and polygons are convex. facets s and t are adjacent if s ∩ t is a ridge of P . 19Two DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 25 Let P be a d-polytope in S(V ) and let S be the set of facets of P . For each s ∈ S, choose a reflection σs = Id − αs ⊗ bs with αs (bs ) = 2 which fixes s. By making a suitable choice of signs, we assume that P is defined by the inequalities αs ⩽ 0, s ∈ S. Let Γ ⊂ SL± (V ) be the group generated by all these reflections (σs )s∈S and let P̊ be the interior of P . A pair (P, (σs )s ) is called a projective Coxeter polytope if the family {γ P̊ }γ∈Γ is pairwise disjoint. The ∣S∣ × ∣S∣ matrix A = (Ast )s,t∈S , Ast = αs (bt ), is called the Cartan matrix of a projective Coxeter polytope P . For each reflection σs , the linear functional αs and the vector bs are defined up to transformations αs ↦ λs αs and bs ↦ λ−1 s bs with λs > 0. Hence the Cartan matrix of P is defined up to the following equivalence relation: two matrices A and B are equivalent if A = ΛBΛ−1 for a diagonal matrix Λ having positive entries. This implies that for every s, t ∈ S, the number Ast Ats is an invariant of the projective Coxeter polytope P . 7.2.2. Vinberg’s results. Vinberg proved that the following conditions are necessary and sufficient for P to be a projective Coxeter polytope: (V1) Ast ⩽ 0 for s ≠ t, and Ast = 0 if and only if Ats = 0. (V2) Ass = 2; and for s ≠ t, Ast Ats ⩾ 4 or Ast Ats = 4 cos2 ( mπst ), mst ∈ N ∖ {0, 1}. The starting point of the proof is that for every two facets s and t of P , the automorphism σs σt has to be conjugate to one of the following automorphisms of V /U with U = ker(αs ) ∩ ker(αt ): λ 0 1 1 cos θ − sin θ [ ] (λ > 0), [ ] or [ ] (θ = m2πst ). −1 0 λ 0 1 sin θ cos θ In the third case we call σs σt a rotation of angle θ. To a projective Coxeter polytope P , we can associate the Coxeter matrix M = (Mst )s,t∈S with the set S of facets of P such that Mst = mst if σs σt is a rotation of angle m2πst , and Mst = ∞ otherwise. Now, from the Coxeter system (S, M ) and the polytope P , we obtain the Coxeter group W and projective Coxeter orbifold P̂ . Eventually, we are also interested in the subgroup Γ of SL± (V ) generated by all the reflections across the facets of P . Theorem 7.1 (Tits [24], Vinberg [105]). Let P be a projective Coxeter polytope. Then the following are true : (1) The morphism σ ∶ W → Γ given by σ(s) = σs is an isomorphism. (2) The group Γ is a discrete subgroup of SL± (V ). (3) The union of tiles C ∶= ∪γ∈Γ γP is convex. (4) The group Γ acts properly discontinuously on Ω, the interior of C, hence the quotient Ω/Γ is a convex real projective Coxeter orbifold. (5) An open face f of P lies in Ω if and only if the Coxeter group Wf is finite. Remark 7.2. Tits proved Theorem 7.1 (without the fifth item) assuming that P is a simplex and the Cartan matrix A of P is symmetric and Ast Ats ⩽ 4 for all facets s, t of P . The statements can be found in Chapter 5 (Theorem 1 of Section 4 and Proposition 6 of Section 6) of [24]. The final version is due to Vinberg [105]. 26 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS 7.3. Convex projective Coxeter 2-orbifolds. In the previous section, we explain the deformation space C(Σ) of properly convex projective structures on a closed 2-orbifold Σ of negative orbifold Euler characteristic (see Theorem 6.14). As a special case, if Σ is a closed projective Coxeter 2-orbifold, then the underlying space of Σ is a polygon and Σ does not contain cone-points. Let v+ be the number of corner reflectors of order greater than 2, and let T (Σ) be the Teichmüller space of Σ. Goldman [52] showed that C(Σ) is homeomorphic to an open cell of dimension −8 + 3v − (v − v+ ) = v+ − 2 + 2(v − 3) = v+ − 2 + 2 dim T (Σ). 7.4. Hyperbolic Coxeter 3-orbifolds. The Coxeter 3-orbifolds which admit a finite-volume hyperbolic structure have been classified by Andreev [2, 3]. A polytope is naturally a CW complex. A CW complex arising from a polytope is called a combinatorial polytope. We abbreviate a 3-dimensional polytope to a polyhedron. Let G be a combinatorial polyhedron and (∂G)∗ be the dual CW complex of the boundary ∂G. A simple closed curve γ is called a k-circuit if it consists of k edges of (∂G)∗ . A circuit γ is prismatic if all the edges of G intersecting γ are disjoint. Theorem 7.3 (Andreev [2, 3]). Let G be a combinatorial polyhedron, and let {si }ni=1 be the set of facets of G. Suppose that G is not a tetrahedron and non-obtuse angles θij ∈ (0, π2 ] are given at each edge sij = si ∩ sj of G. Then the following conditions (A1)–(A4) are necessary and sufficient for the existence of a compact hyperbolic polyhedron P which realizes20 G with dihedral angle θij at each edge sij . (A1) If si ∩ sj ∩ sk is a vertex of G, then θij + θjk + θki > π. (A2) If si , sj , sk form a prismatic 3-circuit, then θij + θjk + θki < π. (A3) If si , sj , sk , sl form a prismatic 4-circuit, then θij + θjk + θkl + θli < 2π. (A4) If G is a triangular prism with triangular facets s1 and s2 , then θ13 + θ14 + θ15 + θ23 + θ24 + θ25 < 3π. The following conditions (F1)–(F6) are necessary and sufficient for the existence of a finitevolume hyperbolic polyhedron P which realizes G with dihedral angle θij ∈ (0, π2 ] at each edge sij . (F1) If si ∩ sj ∩ sk is a vertex of G, then θij + θjk + θki ⩾ π. (F2) (resp. (F3) or (F4)) is the same as (A2) (resp. (A3) or (A4)). (F5) If si ∩ sj ∩ sk ∩ sl is a vertex of G, then θij + θjk + θkl + θli = 2π. (F6) If si , sj , sk are facets such that si and sj are adjacent, sj and sk are adjacent, and si and sk are not adjacent but meet at a vertex not in sj , then θij + θjk < π. In both cases, the hyperbolic polyhedron is unique up to hyperbolic isometries. 7.5. Convex projective Coxeter 3-orbifolds. 7.5.1. Restricted deformation spaces. A point of C(P̂ ) gives us a projective Coxeter polytope (P0 , (σs )s ), well defined up to projective automorphisms. We can focus on the subspace of C(P̂ ) with a projectively fixed underlying polytope P0 . This subspace is called the restricted deformation space of P̂ and denoted by CP0 (P̂ ). 20There is an isomorphism φ ∶ G → P such that the given angle at each edge e of G is the dihedral angle at the edge φ(e) of P . DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 27 Let P̂ be a Coxeter 3-orbifold. We now give a combinatorial hypothesis on P̂ , called the “orderability”, which allows us to say something about the restricted deformation space CP0 (Pˆ) of P̂ . A Coxeter 3-orbifold P̂ is orderable if the facets of P̂ can be ordered so that each facet contains at most three edges that are edges of order 2 or edges in a facet of higher index. Let e (resp. f , e2 ) be the number of edges (resp. facets, edges of order 2) of P , and let k(P ) be the dimension of the group of projective automorphisms of P . Note that k(P ) = 3 if P is tetrahedron, k(P ) = 1 if P is the cone over a polygon other than a triangle, and k(P ) = 0 otherwise. Theorem 7.4 (Choi [32]). Let P̂ be a Coxeter 3-orbifold such that C(P̂ ) ≠ ∅. Suppose that P̂ is orderable and that the Coxeter group π1orb (P̂ ) is infinite and irreducible. Then every restricted deformation space CP (P̂ ) is a smooth manifold of dimension 3f − e − e2 − k(P ). A simplicial polyhedron21 is orderable. By Andreev’s theorem, hyperbolic triangular prisms are orderable. However the cube and the dodecahedron do not carry any orderable Coxeter orbifold structure, since the lowest index facet in an orderable orbifold must be triangular. 7.5.2. Truncation polyhedra. Andreev’s theorem gives the necessary and sufficient conditions for the existence of a closed or finite-volume hyperbolic Coxeter 3-orbifold. We can think of analogous questions for closed or finite-volume properly convex projective Coxeter orbifolds. The third author [84] completely answered the question of whether or not a Coxeter 3orbifold P̂ admits a convex projective structure assuming that the underlying space P is a truncation polyehedron: a truncation d–polytope is a d-polytope obtained from the d-simplex by iterated truncations of vertices. For example, a triangular prism is a truncation polyhedron. However the cube and the dodecahedron are not truncation polyhedra. A prismatic 3-circuit of P̂ formed by the facets r, s, t is bad if 1 1 1 + + ⩾ 1 and 2 ∈ {Mrs , Mst , Mtr }. Mrs Mst Mtr Let e+ be the number of edges of order greater than 2 in P̂ . Theorem 7.5 (Marquis [84]). Let P̂ be a Coxeter 3-orbifold arising from a truncation polyhedron P . Assume that P̂ has no bad prismatic 3-circuits. If P̂ is not a triangular prism and e+ > 3, then C(Pˆ) is homeomorphic to a finite union of open cells of dimension e+ − 3. Moreover, if P̂ admits a hyperbolic structure, then C(Pˆ) is connected. The third author actually provided an explicit homeomorphism between C(Pˆ) and the union of q copies of Re+ −3 when P̂ is a Coxeter truncation 3-orbifold. Moreover, the integer q can be computed easily in terms of the combinatorics and the edge orders. 7.6. Near the hyperbolic structure. 21A simplicial polyhedron is a polyhedron whose facets are all triangles. 28 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS 7.6.1. Restricted deformation spaces. The first and second authors and Hodgson [37] described the local restricted deformation space for a class of Coxeter orbifolds arising from ideal hyperbolic polyhedra, i.e. polyhedra with all vertices on ∂H3 . Note that a finite volume hyperbolic Coxeter obifold is unique up to hyperbolic isometries by Andreev’s theorem [2, 3] or Mostow-Prasad rigidity theorem [92, 95]. Theorem 7.6 (Choi-Hodgson-Lee [37]). Let P be an ideal hyperbolic polyhedron whose dihedral angles are all equal to π3 . If P is not a tetrahedron, then at the hyperbolic structure the restricted deformation space CP (P̂ ) is smooth and of dimension 6. 7.6.2. Weakly orderable Coxeter orbifolds. The first and second authors [38] found a “large” class of Coxeter 3-orbifolds whose local deformation spaces are understandable. A Coxeter 3-orbifold P̂ is weakly orderable if the facets of P can be ordered so that each facet contains at most 3 edges of order 2 in a facet of higher index. Note that Greene [60] gave an alternative (cohomological) proof of the following theorem. Theorem 7.7 (Choi-Lee [38], Greene [60]). Let P̂ is a closed hyperbolic Coxeter 3-orbifold. If P̂ is weakly orderable, then at the hyperbolic structure C(P̂ ) is smooth and of dimension e+ − 3. For example, if P is a truncation polyhedron, then P̂ is always weakly orderable. The cube is not a truncation polyhedron, but every closed hyperbolic Coxeter 3-orbifold arising from the cube is weakly orderable. On the other hand, there exist closed hyperbolic Coxeter 3orbifolds arising from the dodecahedron which are not weakly orderable. However, almost all the closed hyperbolic Coxeter 3-orbifolds arising from the dodecahedron are weakly orderable: Theorem 7.8 (Choi-Lee [38]). Let P be a simple22 polyhedron. Suppose that P has no prismatic 3-circuit and has at most one prismatic 4-circuit. Then lim m→∞ ∣{weakly orderable, closed hyperbolic Coxeter 3-orbifolds P̂ with edge order ⩽ m}∣ ∣{closed hyperbolic Coxeter 3-orbifolds P̂ with edge order ⩽ m}∣ =1 A result similar to Theorem 7.7 is true for higher dimensional closed Coxeter orbifolds P̂ whose underlying polytope P is a truncation polytope: Theorem 7.9 (Choi-Lee [38], Greene [60]). If P̂ be a closed hyperbolic Coxeter orbifold arising from a truncation polytope P , then at the hyperbolic structure C(P̂ ) is smooth and of dimension e+ − d. We remark that if P̂ is not weakly orderable, then Theorem 7.7 is not true anymore: Let m be a fixed integer greater than 3. Consider the compact hyperbolic Coxeter polyhedron π . OthP1 shown in Figure 5 (A). Here, if an edge is labelled m, then its dihedral angle is m π erwise, its dihedral angle is 2 . Let P̂1 be the corresponding hyperbolic Coxeter 3-orbifold. Then e+ − 3 = 0, but C(P̂1 ) = R (see Choi-Lee [38]). Of course P̂1 is not weakly orderable, since every facet in P̂1 contains four edges of order 2. There is also a compact hyperbolic Coxeter 4-polytope P2 such that C(P̂2 ) is not homeomorphic to a manifold. The underlying polytope P2 is the product of two triangles and the Coxeter graph of P̂2 is shown in Figure 5 (B). 22A polyhedron P is simple if each vertex of P is adjacent to exactly three edges. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS m 29 5 5 m 5 m (b) (a) Figure 5. (A) Coxeter 3-polytope P1 (B) Coxeter 4-polytope P2 The space C(P̂2 ) is homeomorphic to the following solution space (see Choi-Lee [38]): √ √ √ S = {(x, y) ∈ (R+ )2 ∣ 8x − (5 + 5)y − (6 − 2 5)xy − (5 + 5)x2 y + 8xy 2 = 0}, which is pictured in Figure 6, and hence C(P̂2 ) is not a manifold. Here the singular point (1, 1) ∈ S corresponds to the hyperbolic structure in C(P̂2 ). 2.0 1.5 1.0 0.5 0.0 0.0 Figure 6. 8x − (5 + 0.5 1.0 1.5 2.0 √ √ √ 5)y − (6 − 2 5)xy − (5 + 5)x2 y + 8xy 2 = 0 8. Infinitesimal deformations 8.1. Rigidity or deformability. We first discuss some general theory. Definition 8.1. A representation ρ ∶ Γ → G is locally rigid if the G-orbit of ρ in Hom(Γ, G) contains a neighborhood of ρ in Hom(Γ, G). Otherwise, ρ is locally deformable. If ρ is locally deformable, then there exists a sequence of representations ρn ∶ Γ → G converging to ρ such that ρn is not conjugate to ρ. We emphasise that ρn has no reason to be discrete or faithful even if ρ is so. Definition 8.2. Two representations ρ, ρ′ ∶ Γ → G are of the same type if for all γ ∈ Γ, ρ(γ) and ρ′ (γ) have the same type in the Jordan decomposition. A discrete faithful representation ρ ∶ Γ → G is globally rigid if every discrete faithful representation in Hom(Γ, G) whose type is the same as ρ is conjugate to ρ. For example, two representations ρ, ρ′ ∶ Γ → PGL(2, R) are of the same type if and only if for each γ ∈ Γ, ρ(γ) and ρ′ (γ) are both hyperbolic, parabolic or elliptic. 30 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS 8.2. What is an infinitesimal deformation? In this subsection, we explore the tangent space to a representation. In order to do that, we will combine differential geometry with algebraic geometry. Given a semi-algebraic set X and a point x ∈ X, we say that X is a smooth manifold of dimension r at x if there is an open neighborhood U of x such that the subset U ∩ X is a smooth r-manifold. Such a point x is said to be smooth. Assume now that X is an algebraic set. We can define the Zariski tangent space at any point x ∈ X. If X is a smooth manifold of dimension r at x, then the Zariski tangent space at x is of dimension at least r. Conversely, if the Zariski tangent space at x is of dimension r and there is a smooth r-manifold in X containing x, then X is a smooth manifold of dimension r at x. A point x ∈ X is said to be singular if there is a Zariski tangent vector which is not tangent to a smooth curve in X. 8.3. First order. Assume that ρt is a smooth path in Hom(Γ, G), i.e. for each γ ∈ Γ, a path ρt (γ) in G is smooth. Then there exists a map u1 ∶ Γ → g such that ∀γ ∈ Γ, ρt (γ) = exp(t u1 (γ) + O(t2 )) ρ0 (γ). Since ρt is a homomorphism, i.e. ρt (γ δ) = ρt (γ)ρt (δ), it follows that u1 is a 1-cocycle. Conversely, if u1 ∶ Γ → g is a 1-cocyle, then ρt is a homomorphism up to first order. This computation motivates the following: Given a representation ρ ∶ Γ → G, we define the space of 1-cocycles Γ → g: Z 1 (Γ, g)ρ = {u1 ∶ Γ → g ∣ u1 (γδ) = u1 (γ) + Adρ(γ) u1 (δ), ∀γ, δ ∈ Γ}. Moreover, since the Zariski tangent space to an algebraic variety is the space of germs of paths satisfying the equations up to first order, the Zariski tangent space TρZar Hom(Γ, G) to Hom(Γ, G) at ρ can be identified with the space of 1-cocyles Z 1 (Γ, g)ρ via the following: ∀γ ∈ Γ, d ρt (γ)ρ0 (γ −1 )∣t=0 = u1 (γ). dt Eventually we want to understand the tangent space to the character variety, hence we need to figure out which cocycles come from the conjugation. We introduce the space of 1-coboundaries: B 1 (Γ, g)ρ = {v1 ∶ Γ → g ∣ ∃ u0 ∈ g such that v1 (γ) = Adρ(γ) u0 − u0 }. Every coboundary v1 , in fact, is tangent to the conjugation path ρt = exp(−tu0 ) ρ0 exp(tu0 ). The first cohomology group with coefficients in g twisted by the adjoint action of ρ is H 1 (Γ, g)ρ = Z 1 (Γ, g)ρ /B 1 (Γ, g)ρ . Basically, we explain that the map TρZar Hom(Γ, G) → Z 1 (Γ, g)ρ is an isomorphism. In addition, under this isomorphism, the Zariski tangent vectors coming from the G-conjugation of ρ exactly correspond to the coboundaries. Definition 8.3. A representation ρ ∶ Γ → G is infinitesimally rigid if H 1 (Γ, g)ρ = 0. The following theorem motivates the terminology. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 31 Theorem 8.4 (Weil’s rigidity theorem [107]). If ρ is infinitesimally rigid, then ρ is locally rigid. A nice presentation of Theorem 8.4 can be found in Besson [20]. Weil, Garland and Raghunathan also computed the group H 1 (Γ, g)ρ in a number of important cases and showed that it is often trivial. Theorem 8.5 (Weil [106], Garland-Raghunathan [51], Raghunathan [96]). Suppose that G is a semi-simple group without compact factor and Γ is an irreducible lattice of G. Denote by i ∶ Γ → G the canonical representation and let ρ ∶ G → H be a non-trivial irreducible representation of G into a semi-simple group H. ● If H 1 (Γ, g)i ≠ 0, then – either g = so2,1 (R) = su1,1 = sl2 (R), – or g = so3,1 (R) = sl2 (C) and Γ is a non-uniform lattice. ● If H 1 (Γ, h)ρ○i ≠ 0 and Γ is a uniform lattice, then g = sod,1 (R) or sud,1 . Moreover, if we write h = V1 ⊕ ⋯ ⊕ Vr the decomposition of the g-semi-simple module h into simple modules, then the highest weight of each Vi is a multiple of the highest weight of the standard representation. 8.4. Higher order. Given a 1-cocycle u1 ∈ Z 1 (Γ, g)ρ , i.e. a Zariski tangent vector to the representation variety, we may ask if u1 is integrable, i.e. the tangent vector to a smooth deformation. We can start with the simplest investigation: Is the 1-cocycle u1 integrable up to second order, i.e. the tangent vector to a smooth deformation up to order 2? Writing the expression ∀γ ∈ Γ, ρt (x) = exp(t u1 (γ) + t2 u2 (γ) + O(t3 )) ρ0 (γ) and using the Baker-Campbell-Hausdorff formula, we see that ρt is a homomorphism up to second order if and only if ∀γ, δ ∈ Γ, 1 1 u2 (γ) − u2 (γδ) + Adρ0 (γ) u2 (δ) = [Adρ0 (γ) u1 (δ), u1 (γ)] ∶= [u1 , u1 ](γ, δ). 2 2 Hence, the 1-cocycle u1 is integrable up to second order if and only if the 2-cocycle [u1 , u1 ] ∈ is a 2-coboundary. We could ask the same question for the third order and so on. We would find a sequence of obstructions, which are all in H 2 (Γ, g)ρ . In other words, for each n ⩾ 2, if we let Z 2 (Γ, g) Vn ∶= {u1 ∈ Z 1 (Γ, g)ρ ∣ ok (u1 ) = 0 for every k = 2, . . . , n − 1}, then there exists a map on ∶ Vn → H 2 (Γ, g)ρ such that the 1-cocycle u1 is integrable up to order n if and only if the obstructions ok (u1 ) = 0 for all k = 2, . . . , n. The story ends with a good news. Recall that G = GR and RJtK is the ring of formal power series. A formal deformation of ρ ∶ Γ → G is a representation ρ̃ ∶ Γ → GRJtK whose evaluation at t = 0 is ρ. A 1-cocycle u1 is, by definition, the formal tangent vector to a formal deformation (or simply formally integrable) if and only if the obstructions on (u1 ) = 0 for all n ⩾ 2. A priori, this does not imply that u1 is the tangent vector to a smooth deformation, but this is in fact true: Theorem 8.6 (Artin [4]). If a 1-cocycle u1 is formally integrable, then u1 is integrable. 32 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS 8.5. Examples in hyperbolic geometry. The world of hyperbolic geometry offers a lot of interesting behaviors. Assume that M is a hyperbolic d-dimensional manifold with or without boundary and Γ is the fundamental group of M . 8.5.1. Hyperbolic surfaces. A lot is known on representations of surface groups, and the story about surface groups is different from higher dimensional manifold groups, which we will eventually concentrate on. Hence, we refer the readers to their favourite surveys on surface group representations (see, for example, Goldman [53, 54], Labourie [79], Guichard [62]). 8.5.2. Finite-volume hyperbolic manifolds. If d ⩾ 3 and M has finite volume, then the famous Mostow-Prasad rigidity theorem [92, 95] states that the holonomy ρ of M is globally rigid. This (conjugacy class of) representation is the geometric representation of Γ. However this does not imply that ρ is locally rigid. Indeed, the geometric representation might be deformed to non-faithful or non-discrete representations.23 It is a theorem of Thurston for dimension d = 3 and of Garland and Raghunathan for dimension d ⩾ 4 that ρ is locally deformable if and only if d = 3 and M has a cusp. This wonderful exception in the local rigidity of finite-volume hyperbolic manifolds is the starting point of the Thurston hyperbolic Dehn surgery theorem. Theorem 8.7 (Thurston [100], Garland-Raghunathan [51]). The holonomy representation of a finite-volume hyperbolic manifold M of dimension d ⩾ 3 is infinitesimally rigid except if d = 3 and M is not compact. In the exceptional case, the geometric representation is a smooth point of the character variety of dimension twice the number of cusps. Bergeron and Gelander [19] gave an alternative proof of Theorem 8.7 using the MostowPrasad rigidity. 8.5.3. Hyperbolic manifolds with boundary. We also wish to cite this beautiful theorem which pushes this kind of question beyond the scope of finite-volume manifolds. Theorem 8.8 (Kerckhoff-Storm [71]). The holonomy representation of a compact hyperbolic manifold with totally geodesic boundary of dimension d ⩾ 4 is infinitesimally rigid. 9. Infinitesimal duality to complex hyperbolic geometry We now return to the original interest of this survey: convex projective structures on manifolds. From the point of view of representations, our problem is to understand deformations ρt ∶ π1 (M ) → SLd+1 (R) from the holonomy ρ0 ∶ π1 (M ) → PSOd,1 (R) of the hyperbolic structure on M into representations in SLd+1 (R). Suppose that M is a finite-volume hyperbolic manifold of dimension d ⩾ 3 and Γ is the fundamental group of M . We have seen that there exists a unique discrete faithful representation ρgeo of Γ into PSOd,1 (R), up to conjugation. If G is a Lie group and i ∶ PSOd,1 (R) → G is a representation, then we call the conjugacy class [i ○ ρgeo ] the hyperbolic point of the character variety χ(Γ, G) and we denote it again by ρgeo . We abuse a little bit of notation here, since we ignore i, but in the following i will always be the canonical inclusion. 23It is easy to see that every discrete and faithful representations of Γ are of the same type. DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 33 Complex hyperbolic geometry can help us to understand local deformations into SLd+1 (R). Indeed, complex hyperbolic geometry is “dual” to Hilbert geometry, however, only at the hyperbolic point and at the infinitesimal level. Remark 9.1. The groups SLd+1 (R) and SUd,1 are non-compact real forms of the complex algebraic group SLd+1 (C) that both contains the real algebraic group PSOd,1 (R). Moreover, the Lie algebra sld+1 (R) splits as (3) sld+1 (R) = sod,1 ⊕ o where o is the orthogonal complement to sod,1 in sld+1 (R) with respect to the Killing form of sld+1 (R), and the adjoint action of PSOd,1 (R) preserves the decomposition (3). Hence to study the cohomology group H 1 (M, sld+1 (R))ρ , we just have to understand H 1 (M, o)ρ , since the cohomology group H 1 (M, sod,1 )ρ is well known. But, since the Lie algebra sud,1 = sod,1 ⊕io, we can find H 1 (M, o)ρ using complex hyperbolic geometry (see Heusener-Porti [64], CooperLong-Thistlethwaite [44] for more details). Remark 9.1 evolves into the following theorem: Theorem 9.2 (Cooper-Long-Thistlethwaite [44]). Let M be a closed hyperbolic manifold, and Γ = π1 (M ). Then the hyperbolic point ρgeo in χ(Γ, SLd+1 (R)) is smooth if and only if the corresponding hyperbolic point in χ(Γ, SUd,1 ) is smooth. Moreover, in that case, the dimensions of the two character varieties at the hyperbolic point are the same. Theorem 9.3 (Guichard [61]). Let M be a closed hyperbolic manifold, and Γ = π1 (M ). In χ(Γ, SLd+1 (C)), hence in particular in χ(Γ, SLd+1 (R)) and χ(Γ, SUd,1 ), the representations close to the hyperbolic point ρgeo are faithful and discrete. The local pictures of the character variety for G = SLd+1 (R) and SUd,1 are therefore the same; however the global pictures can be very different. The work of Morgan-Shalen [91], Bestvina [21] and Paulin [94] shows that the space of discrete and faithful representations of Γ in SUd,1 is compact (if d ⩾ 2), but this can be false if G = SLd+1 (R). 10. Convex projective structures on 3-manifolds 10.1. Computing character varieties. Cooper, Long and Thistlethwaite [43] investigated the local structure of the variety χ(Γ, SL4 (R)) at the hyperbolic point ρ ∶= ρgeo when Γ is the fundamental group of one of the first 4500 closed hyperbolic 3-manifolds with 2-generator groups in the Hodgson-Weeks census: http://www.math.uic.edu/t3m/SnapPy/censuses.html We summarize their conclusions about the character variety χ(Γ, SL4 (R)) around the hyperbolic point: ● 4439 points, i.e. H 1 (Γ, SL4 (R))ρ = 0. ● 9 singular points, i.e. H 1 (Γ, SL4 (R))ρ ≠ 0 but no Zariski tangent vector is integrable. ● 43 smooth curves. ● 7 smooth surfaces. ● 1 singular surface such that H 1 (Γ, SL4 (R))ρ is 3-dimensional. ● 1 singular 3-variety, which has two 3-dimensional branches meeting in a curve. First, we should mention that these computations are mostly done in floating-point mode, hence this summary is a very good speculation but not a statement. Second, the authors checked rigorously their speculations on certain character varieties of this list. 34 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS Remark 10.1. If Γ is the fundamental group of a closed hyperbolic 3-manifold, then the first obstruction o2 (u1 ) = 0, for every Zariski tangent vector u1 at the hyperbolic point ρ (see Cooper-Long-Thistlethwaite [44]). Indeed, first, the infinitesimal rigidity of the closed hyperbolic 3-manifold in PSO3,1 (R) implies that u1 ∈ H 1 (Γ, o). Second, since [o, o] ⊂ so3,1 , it follows that o2 (u1 ) is not only an element of H 2 (Γ, sl4 ) but also an element of H 2 (Γ, so3,1 ). Finally, we know from Poincaré duality that H 2 (Γ, so3,1 ) = H 1 (Γ, so3,1 ) = 0. The singularities of some varieties are therefore more than quadratic, since for example the manifold “Vol3” is locally rigid even if its Zariski tangent space is one-dimensional. Compare to the result of Goldman-Millson [59] that the singularity at a reductive representation is at most quadratic if Γ is the fundamental group of a Kähler manifold. 10.2. Infinitesimal rigidity relative to the boundary. Heusener and Porti made use of a relative version of the infinitesimal rigidity for finite-volume hyperbolic 3-manifolds M in order to obtain the infinitesimal rigidity for some Dehn fillings of M . Let M be a 3-manifold with a boundary whose interior N carries a finite-volume complete hyperbolic metric, and let ρ ∶= ρgeo be the holonomy representation of N . We say that M (or N ) is infinitesimally rigid relative to the boundary if the map H 1 (M, G)ρ → H 1 (∂M, G)ρ is injective. Roughly speaking, at the infinitesimal level, every deformation must change the geometry of the cusp. The combination of the following two theorems shows in particular that infinitely many closed hyperbolic 3-manifolds are locally rigid in G = SL4 (R). Theorem 10.2 (Theorem 1.4 of Heusener-Porti [64]). Infinitely many Dehn fillings of a non-compact hyperbolic 3-manifold of finite volume which is infinitesimally rigid relative to the boundary are infinitesimally rigid. Theorem 10.3 (Heusener-Porti [64]). There exist non-compact hyperbolic 3-manifolds of finite volume which are infinitesimally rigid relative to the boundary. Remark 10.4. For example, the figure-eight knot complement and the Whitehead link complement are infinitesimally rigid relative to the boundary. A finite-volume non-compact hyperbolic 3-manifold which contains an embedded totally geodesic closed hypersurface is not infinitesimally rigid relative to the boundary. This raises the following question: Can we find a (topological) characterization of finite-volume hyperbolic 3-manifolds which are infinitesimally rigid relative to the boundary? An answer even for hyperbolic knot or link complements would already be quite nice. Surprisingly, the technique of Heusener and Porti, which is extended by Ballas, also produces deformations. A slope is a curve in the boundary, and a slope γ is rigid if the map H 1 (M, o)ρ → H 1 (γ, o)ρ is non-trivial. Theorem 10.5 (Heusener-Porti [64], Ballas [5]). Infinitely many (generalized ) Dehn fillings of an amphichiral knot whose longitude is a rigid slope are deformable. References [1] A. Adem, J. Leida, and Y. Ruan. Orbifolds and stringy topology, volume 171 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 2007. 5, 6 [2] E. Andreev. Convex polyhedra in Lobačevskiı̆ spaces. Mat. Sb. (N.S.), 81 (123):445–478, 1970. 26, 28 [3] E. Andreev. Convex polyhedra of finite volume in Lobačevskiı̆ space. Mat. Sb. (N.S.), 83 (125):256–260, 1970. 26, 28 DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 35 [4] M. Artin. On the solutions of analytic equations. Invent. Math., 5:277–291, 1968. 31 [5] S. Ballas. Deformations of noncompact projective manifolds. Algebr. Geom. Topol., 14(5):2595–2625, 2014. 34 [6] S. Ballas. Finite volume properly convex deformations of the figure-eight knot. Geom. Dedicata, 178:49– 73, 2015. 14 [7] S. Ballas, J. Danciger, and G.-S. Lee. Convex projective structures on non-hyperbolic three-manifolds. arXiv:1508.04794. 14 [8] S. Ballas and L. Marquis. Properly Convex Bendings of Hyperbolic Manifolds. In preparation. 13 [9] O. Baues. The deformation of flat affine structures on the two-torus. In Handbook of Teichmüller theory. Vol. IV, volume 19 of IRMA Lect. Math. Theor. Phys., pages 461–537. Eur. Math. Soc., Zürich, 2014. 15 [10] Y. Benoist. Nilvariétés projectives. Comment. Math. Helv., 69(3):447–473, 1994. 15 [11] Y. Benoist. Convexes hyperboliques et fonctions quasisymétriques. Publ. Math. Inst. Hautes Études Sci., (97):181–237, 2003. 11 [12] Y. Benoist. Convexes divisibles. I. In Algebraic groups and arithmetic, pages 339–374. Tata Inst. Fund. Res., Mumbai, 2004. 13 [13] Y. Benoist. Convexes divisibles. III. Ann. Sci. École Norm. Sup. (4), 38(5):793–832, 2005. 11, 12, 13 [14] Y. Benoist. Convexes divisibles. IV. Structure du bord en dimension 3. Invent. Math., 164(2):249–278, 2006. 4, 14, 15 [15] Y. Benoist. Convexes hyperboliques et quasiisométries. Geom. Dedicata, 122:109–134, 2006. 4, 15 [16] Y. Benoist. Five lectures on lattices in semisimple Lie groups. In Géométries à courbure négative ou nulle, groupes discrets et rigidités, volume 18 of Sémin. Congr., pages 117–176. Soc. Math. France, Paris, 2009. 24 [17] J.-P. Benzécri. Sur les variétés localement affines et localement projectives. Bull. Soc. Math. France, 88:229–332, 1960. 14, 15 [18] M. Bergeron. The topology of nilpotent representations in reductive groups and their maximal compact subgroups. Geom. Topol., 19(3):1383–1407, 2015. 4 [19] N. Bergeron and T. Gelander. A note on local rigidity. Geom. Dedicata, 107:111–131, 2004. 9, 32 [20] G. Besson. Calabi-Weil infinitesimal rigidity. In Géométries à courbure négative ou nulle, groupes discrets et rigidités, volume 18 of Sémin. Congr., pages 177–200. Soc. Math. France, Paris, 2009. 31 [21] M. Bestvina. Degenerations of the hyperbolic space. Duke Math. J., 56(1):143–161, 1988. 33 [22] M. Boileau, S. Maillot, and J. Porti. Three-dimensional orbifolds and their geometric structures, volume 15 of Panoramas et Synthèses [Panoramas and Syntheses]. Société Mathématique de France, Paris, 2003. 5 [23] J. Borzellino and V. Brunsden. On the notions of suborbifold and orbifold embedding. Algebr. Geom. Topol., 15(5):2789–2803, 2015. 6 [24] N. Bourbaki. Éléments de mathématique. Fasc. XXXIV. Groupes et algèbres de Lie. Chapitre IV: Groupes de Coxeter et systèmes de Tits. Chapitre V: Groupes engendrés par des réflexions. Chapitre VI: systèmes de racines. Actualités Scientifiques et Industrielles, No. 1337. Hermann, Paris, 1968. 3, 25 [25] M. Bridson and A. Haefliger. Metric spaces of non-positive curvature, volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999. 5 [26] M. Brion. Introduction to actions of algebraic groups. Les cours du CIRM, 1, 2010. 4 [27] R. Canary, D. Epstein, and P. Green. Notes on notes of Thurston. In Fundamentals of hyperbolic geometry: selected expositions, volume 328 of London Math. Soc. Lecture Note Ser., pages 1–115. Cambridge Univ. Press, Cambridge, 2006. 9 [28] S. Choi. Convex decompositions of real projective surfaces. I. π-annuli and convexity. J. Differential Geom., 40(1):165–208, 1994. 18, 20 [29] S. Choi. Convex decompositions of real projective surfaces. II. Admissible decompositions. J. Differential Geom., 40(2):239–283, 1994. 20 [30] S. Choi. Convex decompositions of real projective surfaces. III. For closed or nonorientable surfaces. J. Korean Math. Soc., 33(4):1139–1171, 1996. 20 [31] S. Choi. Geometric structures on orbifolds and holonomy representations. Geom. Dedicata, 104:161–199, 2004. 9 36 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS [32] S. Choi. The deformation spaces of projective structures on 3-dimensional Coxeter orbifolds. Geom. Dedicata, 119:69–90, 2006. 27 [33] S. Choi. The convex real projective orbifolds with radial or totally geodesic ends: The closedness and openness of deformations. ArXiv e-prints, Nov. 2010. 14, 15 [34] S. Choi. Geometric structures on 2-orbifolds : exploration of discrete symmetry, volume 27 of MSJ Memoirs. Mathematical Society of Japan, Tokyo, 2012. 5, 8, 21 [35] S. Choi and W. Goldman. Convex real projective structures on closed surfaces are closed. Proc. Amer. Math. Soc., 118(2):657–661, 1993. 2, 12 [36] S. Choi and W. Goldman. The deformation spaces of convex RP2 -structures on 2-orbifolds. Amer. J. Math., 127(5):1019–1102, 2005. 18, 23 [37] S. Choi, C. Hodgson, and G.-S. Lee. Projective deformations of hyperbolic Coxeter 3-orbifolds. Geom. Dedicata, 159:125–167, 2012. 28 [38] S. Choi and G.-S. Lee. Projective deformations of weakly orderable hyperbolic Coxeter orbifolds. Geom. Topol., 19(4):1777–1828, 2015. 12, 28, 29 [39] S. Choi, G.-S. Lee, and L. Marquis. Convex projective generalized Dehn fillings. 15 [40] S. Choi, G.-S. Lee, and L. Marquis. Deformation space of truncation Coxeter polytope. In preparation. 15 [41] B. Colbois and P. Verovic. Hilbert geometry for strictly convex domains. Geom. Dedicata, 105:29–42, 2004. 11 [42] D. Cooper and K. Delp. The marked length spectrum of a projective manifold or orbifold. Proc. Amer. Math. Soc., 138(9):3361–3376, 2010. 12 [43] D. Cooper, D. Long, and M. Thistlethwaite. Computing varieties of representations of hyperbolic 3manifolds into SL(4, R). Experiment. Math., 15(3):291–305, 2006. 33 [44] D. Cooper, D. Long, and M. Thistlethwaite. Flexing closed hyperbolic manifolds. Geom. Topol., 11:2413–2440, 2007. 33, 34 [45] D. Cooper, D. Long, and S. Tillmann. On convex projective manifolds and cusps. Adv. Math., 277:181– 251, 2015. 15 [46] M. Crampon and L. Marquis. Finitude géométrique en géométrie de Hilbert. Ann. Inst. Fourier (Grenoble), 64(6):2299–2377, 2014. 15 [47] P. de la Harpe. An invitation to Coxeter groups. In Group theory from a geometrical viewpoint (Trieste, 1990), pages 193–253. World Sci. Publ., River Edge, NJ, 1991. 3 [48] P. de la Harpe. On Hilbert’s metric for simplices. In Geometric group theory, Vol. 1 (Sussex, 1991), volume 181 of London Math. Soc. Lecture Note Ser., pages 97–119. Cambridge Univ. Press, Cambridge, 1993. 3 [49] L. Eisenhart. Non-Riemannian geometry, volume 8 of American Mathematical Society Colloquium Publications. American Mathematical Society, Providence, RI, 1990. Reprint of the 1927 original. 6 [50] V. Fock and A. Goncharov. Moduli spaces of convex projective structures on surfaces. Adv. Math., 208(1):249–273, 2007. 20 [51] H. Garland and M. Raghunathan. Fundamental domains for lattices in (R-)rank 1 semisimple Lie groups. Ann. of Math. (2), 92:279–326, 1970. 31, 32 [52] W. Goldman. Affine manifolds and projective geometry on surfaces. 1977. Thesis (B.A.)–Princeton University. 17, 20, 26 [53] W. Goldman. The symplectic nature of fundamental groups of surfaces. Adv. in Math., 54(2):200–225, 1984. 32 [54] W. Goldman. Representations of fundamental groups of surfaces. In Geometry and topology (College Park, Md., 1983/84), volume 1167 of Lecture Notes in Math., pages 95–117. Springer, Berlin, 1985. 32 [55] W. Goldman. Geometric structures on manifolds and varieties of representations. In Geometry of group representations (Boulder, CO, 1987 ), volume 74 of Contemp. Math., pages 169–198. Amer. Math. Soc., Providence, RI, 1988. 5, 8, 9 [56] W. Goldman. Projective geometry on manifolds. note available from the author, 1988. 12 [57] W. Goldman. Convex real projective structures on compact surfaces. J. Differential Geom., 31(3):791– 845, 1990. 18, 19, 20, 23 [58] W. Goldman and J. Millson. Local rigidity of discrete groups acting on complex hyperbolic space. Invent. Math., 88(3):495–520, 1987. 13 DEFORMATIONS OF CONVEX REAL PROJECTIVE MANIFOLDS AND ORBIFOLDS 37 [59] W. Goldman and J. Millson. The deformation theory of representations of fundamental groups of compact Kähler manifolds. Inst. Hautes Études Sci. Publ. Math., (67):43–96, 1988. 34 [60] R. Greene. The deformation theory of discrete reflection groups and projective structures. ProQuest LLC, Ann Arbor, MI, 2013. Thesis (Ph.D.)–The Ohio State University. 28 [61] O. Guichard. Groupes plongés quasi isométriquement dans un groupe de Lie. Math. Ann., 330(2):331– 351, 2004. 33 [62] O. Guichard. Aspects topologiques et géométriques des représentations de groupes de surfaces. 2011. 32 [63] O. Guichard and A. Wienhard. Anosov representations: domains of discontinuity and applications. Invent. Math., 190(2):357–438, 2012. 2 [64] M. Heusener and J. Porti. Infinitesimal projective rigidity under Dehn filling. Geom. Topol., 15(4):2017– 2071, 2011. 2, 33, 34 [65] N. Hitchin. Lie groups and Teichmüller space. Topology, 31(3):449–473, 1992. 2 [66] D. Johnson and J. Millson. Deformation spaces associated to compact hyperbolic manifolds. In Discrete groups in geometry and analysis (New Haven, Conn., 1984), volume 67 of Progr. Math., pages 48–106. Birkhäuser Boston, Boston, MA, 1987. 9, 13, 15 [67] V. Kac and É. Vinberg. Quasi-homogeneous cones. Mat. Zametki, 1:347–354, 1967. 13 [68] M. Kapovich. Convex projective structures on Gromov-Thurston manifolds. Geom. Topol., 11:1777– 1830, 2007. 15 [69] M. Kapovich. Hyperbolic manifolds and discrete groups. Modern Birkhäuser Classics. Birkhäuser Boston, Inc., Boston, MA, 2009. Reprint of the 2001 edition. 6 [70] P. Kelly and E. Straus. Curvature in Hilbert geometries. Pacific J. Math., 8:119–125, 1958. 3 [71] S. Kerckhoff and P. Storm. Local rigidity of hyperbolic manifolds with geodesic boundary. J. Topol., 5(4):757–784, 2012. 32 [72] I. Kim. Rigidity and deformation spaces of strictly convex real projective structures on compact manifolds. J. Differential Geom., 58(2):189–218, 2001. 12 [73] S. Kobayashi and K. Nomizu. Foundations of differential geometry. Vol. I. Wiley Classics Library. John Wiley & Sons, Inc., New York, 1996. Reprint of the 1963 original, A Wiley-Interscience Publication. 6 [74] M. Koecher. The Minnesota notes on Jordan algebras and their applications, volume 1710 of Lecture Notes in Mathematics. Springer-Verlag, 1999. 3 [75] J.-L. Koszul. Déformations de connexions localement plates. Ann. Inst. Fourier (Grenoble), 18:103–114, 1968. 11, 13 [76] C. Kourouniotis. Deformations of hyperbolic structures. Math. Proc. Cambridge Philos. Soc., 98(2):247– 261, 1985. 13 [77] N. Kuiper. On convex locally-projective spaces. In Convegno Internazionale di Geometria Differenziale, Italia, 1953, pages 200–213. Edizioni Cremonese, Roma, 1954. 14, 18 [78] F. Labourie. Anosov flows, surface groups and curves in projective space. Invent. Math., 165(1):51–114, 2006. 2 [79] F. Labourie. Lectures on representations of surface groups. Zurich Lectures in Advanced Mathematics. European Mathematical Society (EMS), Zürich, 2013. 32 [80] W. Lok. Deformations of locally homogeneous spaces and Kleinian groups. PhD thesis, Columbia University, New York, 7 1984. 9 [81] D. Luna. Sur certaines opérations différentiables des groupes de Lie. Amer. J. Math., 97:172–181, 1975. 4 [82] D. Luna. Fonctions différentiables invariantes sous l’opération d’un groupe réductif. Ann. Inst. Fourier (Grenoble), 26(1):ix, 33–49, 1976. 4 [83] G. Margulis and É. Vinberg. Some linear groups virtually having a free quotient. J. Lie Theory, 10(1):171–180, 2000. 3 [84] L. Marquis. Espace des modules de certains polyèdres projectifs miroirs. Geom. Dedicata, 147:47–86, 2010. 2, 14, 27 [85] L. Marquis. Espace des modules marqués des surfaces projectives convexes de volume fini. Geom. Topol., 14(4):2103–2149, 2010. 19, 20 [86] L. Marquis. Exemples de variétés projectives strictement convexes de volume fini en dimension quelconque. Enseign. Math. (2), 58(1-2):3–47, 2012. 13 38 SUHYOUNG CHOI, GYE-SEON LEE, AND LUDOVIC MARQUIS [87] L. Marquis. Surface projective convexe de volume fini. Ann. Inst. Fourier (Grenoble), 62(1):325–392, 2012. 14, 15, 18 [88] L. Marquis. Around groups in Hilbert geometry. In Handbook of Hilbert geometry, volume 22 of IRMA Lect. Math. Theor. Phys., pages 207–261. Eur. Math. Soc., Zürich, 2014. 10 [89] L. Marquis. Coxeter group in Hilbert geometry, Aug. 2014. 14, 15 [90] I. Moerdijk and D. Pronk. Orbifolds, sheaves and groupoids. K-Theory, 12(1):3–21, 1997. 5 [91] J. Morgan and P. Shalen. Valuations, trees, and degenerations of hyperbolic structures. I. Ann. of Math. (2), 120(3):401–476, 1984. 33 [92] G. Mostow. Quasi-conformal mappings in n-space and the rigidity of hyperbolic space forms. Inst. Hautes Études Sci. Publ. Math., (34):53–104, 1968. 28, 32 [93] T. Nagano and K. Yagi. The affine structures on the real two-torus. I. Osaka J. Math., 11:181–210, 1974. 15, 16 [94] F. Paulin. Topologie de Gromov équivariante, structures hyperboliques et arbres réels. Invent. Math., 94(1):53–80, 1988. 33 [95] G. Prasad. Strong rigidity of Q-rank 1 lattices. Invent. Math., 21:255–286, 1973. 28, 32 [96] M. Raghunathan. On the first cohomology of discrete subgroups of semisimple Lie groups. Amer. J. Math., 87:103–139, 1965. 31 [97] R. Richardson. Conjugacy classes of n-tuples in Lie algebras and algebraic groups. Duke Math. J., 57(1):1–35, 1988. 4 [98] R. Richardson and P. Slodowy. Minimum vectors for real reductive algebraic groups. J. London Math. Soc. (2), 42(3):409–429, 1990. 4 [99] W. Thurston. Three-dimensional geometry and topology. Vol. 1, volume 35 of Princeton Mathematical Series. Princeton University Press, Princeton, NJ, 1997. Edited by Silvio Levy. 10, 16 [100] W. Thurston. The geometry and topology of three-manifolds. http://library.msri.org/nonmsri/ gt3m, Electronic version 1.1, March 2002. 5, 8, 21, 32 [101] C. Vernicos. Lipschitz characterisation of polytopal Hilbert geometries. Osaka J. Math., 52(1):215–235, 2015. 11 [102] J. Vey. Sur les automorphismes affines des ouverts convexes saillants. Ann. Scuola Norm. Sup. Pisa (3), 24:641–665, 1970. 12 [103] É. Vinberg. The theory of homogeneous convex cones. Trudy Moskov. Mat. Obšč., 12:303–358, 1963. 3, 12 [104] É. Vinberg. Structure of the group of automorphisms of a homogeneous convex cone. Trudy Moskov. Mat. Obšč., 13:56–83, 1965. 3 [105] È. Vinberg. Discrete linear groups that are generated by reflections. Izv. Akad. Nauk SSSR Ser. Mat., 35:1072–1112, 1971. 14, 24, 25 [106] A. Weil. On discrete subgroups of Lie groups. II. Ann. of Math. (2), 75:578–602, 1962. 31 [107] A. Weil. Remarks on the cohomology of groups. Ann. of Math. (2), 80:149–157, 1964. 31 Department of Mathematical Sciences, KAIST, Republic of Korea E-mail address: [email protected] Mathematisches Institut, Ruprecht-Karls-Universität Heidelberg, Germany E-mail address: [email protected] IRMAR, University of Rennes, France E-mail address: [email protected]
4math.GR
Constructing unlabelled lattices arXiv:1609.08255v2 [math.CO] 18 Feb 2018 Volker Gebhardt, Stephen Tawn 18 February 2018 Abstract We present an improved orderly algorithm for constructing all unlabelled lattices up to a given size, that is, an algorithm that constructs the minimal element of each isomorphism class relative to some total order. Our algorithm employs a stabiliser chain approach for cutting branches of the search space that cannot contain a minimal lattice; to make this work, we grow lattices by adding a new layer at a time, as opposed to adding one new element at a time, and we use a total order that is compatible with this modified strategy. The gain in speed is between one and two orders of magnitude. As an application, we compute the number of unlabelled lattices on 20 elements. 1 Introduction Enumerating all isomorphism classes of unlabelled lattices, in the sense of systematically constructing a complete list of isomorphism classes up to a certain size threshold, is a difficult combinatorial problem. The number un of isomorphism classes of unlabelled lattices on n elements grows exponentially in n [KL71, KW80], as does the typical size of an isomorphism class. Indeed, the largest value of n for which un has been published previously is n = 19 [JL15, Slo]. When trying to enumerate combinatorial objects modulo isomorphism, one typically faces the problem that the number and the size of the isomorphism classes are so large that trying to weed out isomorphic objects through explicit isomorphism tests is out of the question. Instead, an orderly algorithm is needed, that is, an algorithm that traverses the search space in such a way that every isomorphism class is encountered exactly once. A general strategy for the construction of isomorphism classes of combinatorial objects using canonical construction paths was described in [McK98]. Orderly algorithms for enumerating isomorphism classes of unlabelled lattices, as well as special subclasses of unlabelled lattices, were given in [HR02, JL15]. The fastest published method currently is the one described in [JL15]; the computations of u18 and u19 reported in [JL15] took 26 hours respectively 19 days on 64 CPUs. Both authors acknowledge support under UWS grant 20721.81112. acknowledges support under the Spanish Project MTM2013-44233. MSC-class: 05A15 (Primary) 06A07, 05-04 (Secondary) Keywords: unlabelled lattice, enumeration, orderly algorithm 1 Volker Gebhardt The algorithms described in [HR02, JL15] follow a similar strategy: (i) A total order <wt on all labelled lattices of a given size is defined. (ii) Starting from the (unique) <wt -minimal lattice on 2 elements, the <wt -minimal labelled representative of each isomorphism class of unlabelled lattices with at most n elements is constructed using a depth first search, where the children of a parent lattice are obtained by adding a single new element covering the minimal element of the parent lattice. (iii) For each parent lattice, one child is obtained for every choice for the covering set of the added element that yields a labelled lattice that is <wt -minimal in its isomorphism class of unlabelled lattices. It is the test for <wt -minimality in step (iii) that takes most of the time: While there are some necessary conditions that are easy to verify, ensuring that the newly constructed labelled lattice is indeed <wt -minimal in its isomorphism class requires checking the candidate covering set of the added element against all possible relabellings of the elements of the existing lattice; details are given in Section 2. Basically, one has a certain permutation group that acts on a configuration space of covering sets, and one must verify that a given candidate is minimal in its orbit. It turns out that the elements of a <wt -minimal labelled lattice are arranged by levels (cf. Section 2), and thus it is tempting to construct and test candidate covering sets of a new element level by level, exploiting the levellised structure for a divide-and-conquer approach; such an approach promises two advantages: (i) The orbit of the restriction of a candidate covering set to a given level is potentially much smaller than the orbit of the complete covering set. (ii) The entire branch of the search space that corresponds to the candidate configuration of covers on the given level can potentially be discarded in a single test. However, we shall see that the constructions from [HR02, JL15] do not adapt well to this levellised approach: In order to make the levellised approach work, we need to modify the depth-first-search to add one level at a time as opposed to one element at a time, and we need to modify the total order to be level-major. The structure of the paper is as follows: In Section 2, we recall some results from [HR02, JL15] that are needed later, and we interpret the total order used in [HR02, JL15] as row-major. In Section 3, we describe our new construction using a level-major order and prove the results required to establish its correctness. In Section 4, we remark on implementation details and compare the performance of our new approach to that of those published in [HR02, JL15]. We thank the Institute for Mathematics at the University of Seville (IMUS) for providing access to a 64-node 512 GB RAM computer that was used for an earlier version of this paper, Peter Jipsen and Nathan Lawless for providing the code for the algorithm used in [JL15] for timing comparisons, and Jukka Kohonen for suggesting to include a discussion of graded lattices. 2 Background We start by giving a brief summary of the algorithms from [HR02, JL15]. We refer to these sources for details. Definition 1. A finite poset L is an n-poset, if the elements of L are labelled 0, 1, . . . , n − 1, where 0 is a lower bound of L and 1 is an upper bound of L. 2 An n-poset that is a lattice is called an n-lattice. To avoid confusion with the numerical order of integers, we denote the partial order of an n-poset L by ⊑L and ⊒L , or simply ⊑ and ⊒ if the poset is obvious. Notation 2. Assume that L is an n-poset. For a ∈ L, we define the shadow of a as ↓ a = ↓L a = {x ∈ L : x ⊑L a} and the shadeSof a as ↑ a = ↑L a = {x ∈ L : x ⊒ SL a}. For A ⊆ L, we define ↑ A = ↑L A = a∈A ↑L a as well as ↓ A = ↓L A = a∈A ↓L a. We say that a ∈ L has depth dep(a) = depL (a) = p, if the maximum length of any chain from 1 to a in L is p + 1. Given a non-negative integer k, we call levk (L) = {a ∈ L : depL (a) = k} the k-th level of L. We say that L is levellised, if depL (i) ≤ depL (j) holds for all 0 < i ≤ j < n. For a, b ∈ L, we write a ≺ L b (or simply a ≺ b) if a is covered by b in L, that is, a ⊑ L b holds andca ⊑ L c x ⊑ L b implies x = a or x = b. We denote the covering set of a ∈ L by a = L a = {x ∈ L : a ≺ x}, and we say that a ∈ L is an atom in L if 0 ≺L a holds. If L is a lattice and a, b ∈ L, we denote the least common upper bound of a and b in L by a ∨ L b (or simply a ∨ b), and the greatest common lower bound of a and b in L by a ∧ L b (or simply a ∧ b). 2.1 Canonical representatives The idea of an orderly algorithm is to construct all those lattices that are minimal, with respect to a suitable total order, in their isomorphism class. In this section, we recall the total order used in [HR02, JL15] and some of its properties. Definition 3. Let L be an n-poset. (a) For A ⊆ L, we define wtL (A) = P 2j . j∈A (b) For i ∈ L, we define wtL (i) = wtL c L  i .  (c) We define wt(L) = wtL (2), wtL (3), . . . , wtL (n − 1) . Ordering n-lattices lexicographically with respect to wt(L), we obtain a total order <wt on the set of all n-lattices. Remark 4. An n-poset L is completely defined by its covering relation. Indeed, the upper bound 1 is not covered by any element, and it covers precisely those elements that are not covered by any other element. Similarly, the lower bound 0 covers no element, and it is covered precisely by those elements that do not cover any other element. Thus, L is completely described by specifying the pairs (i, j), for 1 < i, j < n, for which i ≺ j holds. The latter information can be interpreted as an (n − 2) by (n − 2) matrix over F2 , and the total order <wt from Definition 3 amounts to a row-major lexicographic order on the associated matrices; cf. Figure 1. Theorem 5 ([HR02, Theorem 1]). If L is a <wt -minimal n-lattice and one has 0 < i ≤ j < n, then depL (i) ≤ depL (j) holds. Corollary 6. If L is a <wt -minimal n-lattice and one has 0 < i and i ≺ j, then j < i holds. 3 Remark 7. Theorem 5 and Corollary 6 say that a <wt -minimal n-lattice L is levellised, that is, that the non-minimal levels of L are filled by elements labelled in their numerical order; cf. Figure 2. Remark 8. Any relabelling π ∈ Sym({0, . . . , n − 1}) of the elements of an nlattice L acts on each level of L, as depL (π(i)) = depL (i) holds for every i ∈ L. In particular, if L is a levellised lattice with the elements labelled as in Figure 2, then the relabelled lattice π(L) is levellised if and only if π ∈ Sym({0}) × Sym({1}) × Sym({2, . . . , a2 − 1})× × Sym({a2 , . . . , a3 − 1}) × · · · × Sym({ak , . . . , n − 1}) holds. Theorem 9 ([HR02, Theorem 2]). If L is a <wt -minimal n-lattice, one has wtL (2) ≤ wtL (3) ≤ . . . ≤ wtL (n − 1) . Remark 10. Theorem 9 says that the rows of the matrices describing the covering relation of a <wt -minimal n-lattice L (cf. Figure 1) are sorted in nondecreasing order with respect to a (right-to-left) lexicographic order on the rows. 2.2 Incremental construction The algorithms from [HR02, JL15] work by traversing a tree of <wt -minimal n-lattices in a depth-first manner; the root of the tree is the unique 2-lattice, e is a descendant of the <wt -minimal n-lattice L, if L e is and an (n + 1)-lattice L e is <wt -minimal. obtained from L by adding a new cover of 0 (labelled n) and L e is determined by L and the covering set of the new element n; The lattice L the possible choices for the latter can be characterised effectively. Definition 11 ([HR02]). If L is an n-lattice, a non-empty antichain A ⊆ L\{0} is called a lattice-antichain for L, if a∧L b ∈ {0}∪(↑L A) holds for any a, b ∈ ↑L A. Remark 12. To test the condition in Definition 11, it is clearly sufficient to verify that a ∧L b ∈ ↑L A holds for those pairs (a, b) that are minimal in the set  (a, b) ∈ (↑L A) × (↑L A) : a ∧L b 6= 0 with respect to the product partial order in L × L. 2 n−1 n−1 2 Figure 1: Interpreting the order <wt as row-major lexicographic order on the matrices specifying the covering relation. Note that adding a column for 1 would not affect the order: An entry in this column is determined by the other entries in the same row, and it is checked after those in the lexicographic comparison. 4 Theorem 13 ([HR02, Lemma 2]). Let L be an n-lattice. A subset A ⊆ L \ {0} is a lattice-antichain for L, if and only if L is a subposet c of an (n + 1)-lattice LA in which 0 ≺LA n (that is, n is an atom in LA ) and LA n = A hold. Remark 14. In the situation of Theorem 13, it is clear that the pair (L, A) uniquely determines LA and vice versa. Moreover, the covering relation of LA is obtained from the covering relation of L by (i) adding the pair (0, n); (ii) adding all pairs of the form (n, a) for a ∈ A; and (iii) removing all pairs of the form (0, a) for a ∈ A that are present. As mentioned in Remark 4, the covers of 0 need not be stored explicitly; in this case, only step (ii) is needed. Indeed, this definition of LA makes sense for any n-poset L and any A ⊆ L. It is obvious from the definitions that one has wtLA (n) = wtL (A). Moreover, for 1 ≤ i < n, we have i 6⊑LA n and thus wtLA (i) = wtL (i) and depLA (i) = depL (i). The following two results are consequences of Remark 14 and Theorem 9. Corollary 15 ([HR02, §3]). If L is an n-lattice and A is a lattice-antichain for L such that LA is <wt -minimal, then L is <wt -minimal. Corollary 16 ([HR02, §5]). If L is an n-lattice and A is a lattice-antichain  for L such that LA is <wt -minimal, then one has A ∩ levk−1 (L) ∪ levk (L) 6= ∅ for k = depL (n − 1). 2.3 Testing for canonicity In the light of Corollary 16, there are two cases to consider for testing whether the descendant LA of a <wt -minimal n-lattice L defined by a lattice-antichain A for L with depL (n − 1) = k is <wt -minimal: (A) A ∩ levk (L) 6= ∅, that is, depLA (n) = k + 1 (B) A ∩ levk (L) = ∅ 6= A ∩ levk−1 (L), that is, depLA (n) = k Case (A): In this case, the new element n forms a separate level of LA ; cf. Figure 3(a). As LA is levellised by construction and the non-minimal elements of L correspond to the levels 0, . . . , k of LA , any relabelling π of LA for which π(LA ) is levellised must fix n and induce a relabelling of L by Remark 8. By Remark 14 1 2 3 ··· a2 −1 a2 a1 +1 ··· a3 −1 ··· ··· ··· ak ak +1 ··· ··· n−1 0 Figure 2: A levellised n-lattice; the dashed lines separate the levels. 5 and the definition of the lexicographic order, wt(π(LA )) < wt(LA ) implies wt(π(L)) ≤ wt(L). By Corollary 15, the latter implies wt(π(L)) = wt(L) and thus π(L) = L. To test whether LA is <wt -minimal, it is thus sufficient to check that  (1) wtL (A) = min wtL (π(A)) : π ∈ Stab(L) × Sym({n}) ∼ = Stab(L) holds, again using Remark 14. Case (B): In this case, the new element n is added to the lowest existing non-trivial level of L; cf. Figure 3(b). Thus, a relabelling π of LA for which π(LA ) is levellised need not fix n and induce a relabelling of L. It will, however, induce a relabelling of the lattice L′ induced by the levels 0, . . . , k − 1, k + 1 of L (or LA ), and one has π(L′ ) = L′ by the same arguments as in the previous case. If the lowest non-trivial level of LA contains the elements ak , . . . , n, checking  whether LA is <wt -minimal means testing that wtLA (ak ), . . . , wtLA (n) is lexicographically minimal in its the group Stab(L′ )× Sym({ak , . . . , n}). corbit under ′ Observe that one has LA i ⊆ L for ak ≤ i ≤ n, and that the action of π ∈ Stab(L′ ) × Sym({ak , . . . , n}) not only modifies the individual weights wtLA (i) (by relabelling the elements of L′ ), but also permutes their positions in the sequence (by acting on {ak , . . . , n}). 2.4 Vertically indecomposable lattices Definition 17. An n-lattice L is vertically decomposable, if there exists an element i ∈ L \ {0, 1} that is comparable to every other element of L. Otherwise, L is vertically indecomposable. One can speed up the construction by restricting to lattices that are vertically indecomposable; a straightforward recursion makes it possible to recover all lattices from the vertically indecomposable ones. We refer to [HR02, §5] for details. 3 An improved algorithm  The test for minimality of wtL (A) or wtLA (ak ), . . . , wtLA (n) in their orbit under the acting permutation group is the most time consuming part of the construction and thus an obvious target for improvement. In Section 3.1, we sketch the basic idea for a more efficient algorithm, but we will see that the construction of [HR02, JL15] has to be modified to make this idea work. We describe our modified construction in Section 3.2. L \ {0} ( 1 ··· ··· L′ \ {0} ··· ak ··· n−1 n ( 1 ··· ak−1 ··· ak 0 ··· ··· ak −1 ··· n−1 n 0 (a) Case (A). (b) Case (B). Figure 3: The lattice LA obtained by adding a new cover n of 0 to the n-lattice L. 6 3.1 Stabiliser chain approach Case (A) from Section 2.3 suggests a possible approach, namely the use of a standard technique from computational group theory: stabiliser chains. Since Stab(L) preserves each level of L by Remark 8 and L is levellised, it is tempting to construct and test lattice-antichains level by level: Defining Sk := Stab(L) as well as Ad := A ∩ levd (L) and Sd−1 := Sd ∩ Stab(Ad ) for d = k, . . . , 1, condition (1) is equivalent to the following condition:  for d = k, . . . , 1 (2) wtL (Ad ) = min wtL (π(Ad )) : π ∈ Sd The sets Ad for d = k, . . . , 1 can be constructed and tested one at a time; this offers two advantages: Firstly, if the test at level d fails, the levels d − 1, . . . , 1 don’t have to be constructed; an entire branch of the search space is discarded in one step. Secondly, even if thePtest succeeds on all levels, the cost of testing k condition (2) is proportional to d=1 |ASd d |, and thus in general much smaller Q k than |AStab(L) | = d=1 |ASd d |, which is the cost of testing condition (1) directly. However, when trying to use a similar approach for Case (B), we run into problems: We must compare   wtLA (ak ), . . . , wtLA (n) = wtL (ak ), . . . , wtL (n − 1), wtL (A) lexicographically to its images under the elements of the acting permutation group Stab(L′ ) × Sym({ak , . . . , n}), but if A has only been constructed partially, wtLA (n) = wtL (A) is not completely determined; in the interpretation of Remark 4 and Figure 1, the leftmost entries of the last row of the binary matrix corresponding to LA are undefined. The elements of the group Stab(L′ ) × Sym({ak , . . . , n}) can permute the rows of this matrix, so the position of the undefined entries will vary. Clearly, the lexicographic comparison of the two matrices must stop once it reaches an entry that is undefined in one of the matrices being compared; in this situation, the order of the two matrices cannot be decided on the current level. The problem is that the position in the matrix at which the lexicographic comparison must stop depends on the relabelling that is applied (cf. Figure 4). A consequence of this is that the subset of elements of Stab(L′ )×Sym({ak , . . . , n}) for which the parts of the matrices that can be compared are equal does not form a subgroup, so applying a stabiliser chain approach is not possible. 3.2 Levellised construction The analysis at the end of the preceding section indicates that the problem is that possible relabellings can swap an element whose covering set is only partially determined with an element whose covering set is completely determined, or in other words, that we add a new element to an existing level of L. The idea for solving this problem is simple: Rather than adding one element at a time, possibly to an already existing level, we only ever add an entire level at a time; that way, the problem of adding elements to an existing level is avoided. To make the stabiliser chain approach work in this setting, we must use a total order that compares parts of the covering sets in the same order in which they are constructed; that is, we have to compare the entries of the matrices describing the covering relations in level-major order. 7 Notation 18. For a levellised n-lattice L with n > 2 and depL (n − 1) = k, let L′ denote the lattice induced by the levels 0, . . . , k − 1, k + 1 of L, that is, the lattice obtained from L by removing its last non-trivial level. Note that L′ is a levellised n′ -lattice for some n′ < n. The total order we are about to define uses the partition of the covering set of each element according to levels. Definition 19. Given a levellised n-lattice L with depL (n − 1) = k, an element cd c  i ∈ L \ {0} and an integer d ∈ {1, . . . , k − 1}, we define L i = L i ∩ lev d (L). Definition 20. Using induction on n, we define a relation < on the set of levellised n-lattices that are isomorphic as unlabelled lattices as follows: Given n > 2 and levellised n-lattices L1 and L2 that are isomorphic as unlabelled lattices, we say L1 < L2 , if one of the following holds: • L′1 < L′2 • L′1 = L′2 = L′ and, denoting depL1 (n − 1) = depL2 (n − 1) = k and levk (L1 ) = levk (L2 ) = {ak , . . . , n − 1}, there exist ℓ ∈ {1, . . . , k − 1} as well as i ∈ {ak , . . . , n − 1} such that both of the following hold: cd  cd  – wtL′ L2 j if d > ℓ, or d = ℓ and j ∈ {ak , . . . , i − 1} L1 j = wtL′ cℓ  cℓ  – wtL′ L2 i L1 i < wtL′ Remark 21. The relation from Definition 20 corresponds to a level-major lexicographic comparison of the binary matrices describing the covering relations of L1 and L2 as illustrated in Figure 5. As L1 and L2 are levellised, both matrices are lower block triagonal, with the blocks defined by the levels. Moreover, as L1 and L2 are isomorphic as unlabelled lattices, the block structures of both matrices are identical. Notice also that the matrix describing the covering relation of L′1 (respectively L′2 ) is obtained from that of L1 (respectively L2 ) by removing the lowest row and the rightmost column of blocks. Lemma 22. The relation < from Definition 20 is a total order on the set of levellised n-lattices that are isomorphic as unlabelled lattices. ak ak n n 2 ak −1 2 ak −1 Figure 4: Lexicographic comparison of a lattice LA obtained from a partially constructed lattice-antichain A (left) and a relabelling (right); only the relevant parts of the matrices are shown. Thick lines indicate the boundaries between levels. The parts of the lattice-antichain not yet constructed are shown in grey. 8 Proof. Verifying transitivity is routine. Trichotomy holds, as for s ∈ {1, 2} and i ∈ {2, . . . , n − 1} with depL1 (i) = depL2 (i) = D, one has k Ls i=  D−1 S cd   i    d=1 Ls      {1} ; if ; if D−1 S cd Ls d=1 D−1 S cd Ls d=1 i 6= ∅ i=∅ cd  cd  c c whence L1 i = L2 i holds if and only if one has wtL′ L2 i L1 i = wtL′ for d = 1, . . . , D − 1. Definition 23. An n-lattice L is canonical if L is levellised and <-minimal among all levellised n-lattices that are isomorphic to L as unlabelled lattices. Lemma 24. Every isomorphism class of unlabelled lattices on n elements contains a unique canonical n-lattice. Proof. The set of representatives that are levellised n-lattices is clearly nonempty and finite, and < is a total order on this set by Lemma 22. The following Theorem 26 is an analogue of Corollary 15. Lemma 25. If L is a levellised (n + 1)-lattice for n > 1, then L = L \ {n} is a levellised n-lattice and depL (i) = depL (i) holds for i = 1, . . . , n − 1. Moreover, for a, b ∈ L one has the following identities:  a ∧L b ; if a ∧L b 6= n a ∨L b = a ∨L b and a ∧L b = 0 ; if a ∧L b = n Proof. As L is levellised, 1 ≤ i < n implies depL (i) ≤ depL (n), so i 6⊑L n, and hence depL (i) = depL (i). As L is levellised, so is L. A routine verification shows the identities for a ∨L b and a ∧L b; in particular, L is an n-lattice. Theorem 26. If L is a canonical n-lattice, then L′ is canonical. 2 n−1 n−1 2 Figure 5: Interpreting the order < as level-major lexicographic order on the matrices specifying the covering relation. Thick lines indicate the boundaries between levels. The entries of the matrix shown in grey are zero. 9 Proof. Iterated application of Lemma 25 shows that L′ is levellised. Assume that π ′ is a relabelling of L′ such that π ′ (L′ ) is levellised and one has π ′ (L′ ) < L′ . We can trivially extend π ′ to a relabelling π of L such that ′ π(L) = π ′ (L′ ) < L′ , contradicting the assumption that L is canonical. Theorem 26 means that we can again construct a tree of canonical n-lattices in a depth-first manner; the root of the tree is the unique 2-lattice, and an (n + m)e is a descendant of a canonical n-lattice L, if L e is obtained from L by lattice L e is canonical. adding m new covers of 0 (labelled n, . . . , n + m − 1) and L e The lattice L is determined by L and the covering sets of the new elements n, . . . , n + m − 1; the possible choices for the latter can again be characterised effectively using lattice-antichains, although this time, extra compatibility conditions are needed. The following Theorem 29, a generalisation of Theorem 13, makes this precise. Notation 27. Given m ∈ N+, ann-poset L, and An , . . . , An+m−1 ⊆ L \ {0}, let LAn ,...,An+m−1 = · · · LAn · · · An+m−1 denote the (n + m)-poset obtained c i = Ai from L by adding m new atoms n, . . . , n + m − 1 with LAn ,...,A n+m−1 for i = n, . . . , n + m − 1. (See Remark 14.) Lemma 28. Let L be a levellised n-poset with depL (n − 1) = k, let m ∈ N+ , and let Ai ⊆ L \ {0} for i = n, . . . , n + m − 1. The following are equivalent: e = LAn ,...,An+m−1 is a levellised (n+m)-poset with dep e (a) = depL (a) ≤ k (A) L L for 1 ≤ a < n and depLe (n) = . . . = depLe (n + m − 1) = k + 1. (B) Ai ∩ levk (L) 6= ∅ holds for n ≤ i < n + m. Proof. As L is levellised, one has depL (a) ≤ depL (n − 1) = k for all a ∈ L \ {0}, and induction using Remark 14 shows that depLe (a) = depL (a) ≤ k holds for all a ∈ L \ {0}, which implies depLe (i) ≤ k + 1 for i = n, . . . , n + m − 1 by construction. Thus, for any i = n, . . . , n + m − 1, one has depLe (i) = k + 1 if c  e e and only if ∅ = 6 e i ∩ levk (L) = Ai ∩ levk (L) = Ai ∩ levk (L) holds. Finally, L e is as L is levellised, depLe (n) = . . . = depLe (n + m − 1) = k + 1 implies that L levellised. Theorem 29. Let L be a levellised n-lattice with depL (n − 1) = k, let m ∈ N+ , and let Ai ⊆ L \ {0} for i = n, . . . , n + m − 1. The following are equivalent: e = LAn ,...,An+m−1 is a levellised (n + m)-lattice, dep e (a) = depL (a) ≤ k (A) L L for 1 ≤ a < n, and depLe (n) = . . . = depLe (n + m − 1) = k + 1. (B) (i) Ai ∩ levk (L) 6= ∅ holds for n ≤ i < n + m; (ii) Ai is a lattice-antichain for L for n ≤ i < n + m; and (iii) if a, b ∈ (↑L Ai ) ∩ (↑L Aj ) for n ≤ i < j < n + m, then a ∧L b 6= 0. Proof. We use induction on m. In the case m = 1, condition (B)(iii) is vacuous e = LAn is a lattice if and only if An is a and, by Theorem 13, the poset L lattice-antichain for L. Together with Lemma 28, the claim for m = 1 is shown. Let m > 1 and consider L◦ = LAn ,...,An+m−2 and A = An+m−1 ; we have e L = (L◦ )A . By Lemma 28, we can assume depLe (a) = depL◦ (a) = depL (a) ≤ k 10 for 1 ≤ a < n and depLe (n) = . . . = depLe (n + m − 1) = depL◦ (n) = . . . = depL◦ (n + m − 2) = k + 1. In particular, ↑Le A = ↑L◦ A = ↑L A holds. Also, for i = n, . . . , n + m − 1 and a ∈ L, we have i ⊑Le a if and only if a ∈ ↑L Ai holds. First assume that (A) holds. Induction using Lemma 25 shows that the sets An , . . . , An+m−2 are latticeantichains for L, and a, b ∈ (↑L Ai ) ∩ (↑L Aj ) for n ≤ i < j < n + m − 1 implies a∧L b 6= 0. Further, by Theorem 13, the set A is a lattice-antichain for L◦ , which means that for a, b ∈ ↑L◦ A = ↑L A, one has a ∧L◦ b ∈ {0} ∪ ↑L◦ A = {0} ∪ ↑L A. Let a, b ∈ ↑L A. If a ∧L b 6= 0 holds, we have a ∧L b = a ∧L◦ b by repeated application of Lemma 25, whence a ∧L b ∈ ↑L A. Thus, A is a lattice-antichain for L. On the other hand, if a ∧L b = 0 holds, then n + m − 1 is a maximal e and the lattice property of L e then implies common lower bound of a and b in L, that i 6⊑Le a, that is a ∈ / ↑L Ai , holds for all n ≤ i < n + m − 1. In particular, have condition (B)(iii). Together with Lemma 28, we have thus shown (B). Now assume that (B) holds. By induction, L◦ is a levellised (n + m − 1)-lattice. Let a, b ∈ ↑L◦ A = ↑L A. If we have a ∧L b 6= 0, then a ∧L◦ b = a ∧L b ∈ ↑L A = ↑L◦ A holds, using Lemma 25 and the fact that A is a lattice-antichain for L by assumption. On the other hand, a ∧L b = 0 implies a ∧L◦ b = 0, as by assumption, there is no i ∈ {n, . . . , n+m−2} such that a, b ∈ ↑L Ai holds. Thus, A is a lattice-antichain e is a lattice by Theorem 13. Together with Lemma 28, we have for L◦ , whence L thus shown (A). Notation 30. Let L be a levellised n-lattice with depL (n − 1) = k, let m ∈ N+ , and let Ai ⊆ L \ {0} for i = n, . . . , n + m − 1. For d = 1, . . . , k we define LW dL (An , . . . , An+m−1 ) as the sequence    wtL An ∩ levd (L) , . . . , wtL An+m−1 ∩ levd (L) and we define the sequence LW L (An , . . . , An+m−1 ) as the concatenation of LW kL (An , . . . , An+m−1 ), . . . , LW 1L (An , . . . , An+m−1 ) in this order. Theorem 31. Let L be a levellised n-lattice with depL (n − 1) = k, let m ∈ N+ , and assume that Ai ⊆ L \ {0} for i = n, . . . , n + m − 1 satisfy condition (B) from Theorem 29. Then LAn ,...,An+m−1 is a canonical (n + m)-lattice if and only if: (i) L is canonical; and (ii) the sequence LW L (An , . . . , An+m−1 ) is lexicographically minimal under the action of Stab(L) × Sym({n, . . . , n + m − 1}) given by      π LW L (An , . . . , An+m−1 ) = LW L π Aπ−1 (n) , . . . , π Aπ−1 (n+m−1) for π ∈ Stab(L) × Sym({n, . . . , n + m − 1}). Proof. As any π ∈ Stab(L)×Sym({n, . . . , n+m−1}) induces a relabelling of the e = LAn ,...,An+m−1 , elements of L, the action is well-defined. Moreover, defining L cd Theorem 29 implies Le i = Ai ∩ levd (L) for i = n, . . . , n + m − 1 and 1 ≤ d ≤ k. e ′ is canonical, so (i) e is canonical. By Theorem 26, L = L First assume that L holds. If (ii) does not hold, there exist π ∈ Stab(L) × Sym({n, . . . , n + m − 1}) as well as ℓ ∈ {1, . . . , k} and i ∈ {n, . . . , n + m − 1} such that one has 11    – wtL π Aπ−1 (j) ∩ levd (L) = wtL Aj ∩ levd (L) if d > ℓ, or d = ℓ and j ∈ {n, . . . , i − 1}; and    – wtL π Aπ−1 (i) ∩ levℓ (L) < wtL Ai ∩ levℓ (L) . b=L By Theorem 29, L π (A ),...,π(Aπ−1 (n+m−1) ) is a levellised (n + m)-lattice, and L = L = L holds by construction. Hence, the above conditions mean e is not canonical, contradicting the assumption. Thus (ii) holds. that L b′ π −1 (n) e′ e is not canonical, there is a relabelling π of L e such that L b = π(L) e Conversely, if L b e e is levellised and L < L holds. As π actson the levels of L it induces a relabelling b ′ = π(L) e ′=π L e ′ = π(L), which is levellised as well. If (i) of L, and we have L b<L e implies that one has L b′ = L e′ = L holds, we cannot have π(L) < L, so L and there exist ℓ ∈ {1, . . . , k} as well as i ∈ {n, . . . , n + m − 1} such that both of the following hold: cd  cd  – wtL b j = wtL e j if d > ℓ, or d = ℓ and j ∈ {n, . . . , i − 1} L L – wtL cℓ  cℓ  b i < wtL ei . L L As we have cd b L j = π(Aπ−1 (j) ) ∩ levd (L) for j = n, . . . , n + m − 1 and 1 ≤ d ≤ k, the above conditions imply that      LW L π Aπ−1 (n) , . . . , π Aπ−1 (n+m−1) = π LW L (An , . . . , An+m−1 ) b ′ = L, is lexicographically smaller than LW L (An , . . . , An+m−1 ). Since π(L) = L we have π ∈ Stab(L) × Sym({n, . . . , n + m − 1}), so (ii) does not hold. Corollary 32. Let L be a levellised n-lattice with depL (n − 1) = k, let m ∈ N+ , and assume that Ai ⊆ L \ {0} for i = n, . . . , n + m − 1 satisfy condition (B) from Theorem 29. Then LAn ,...,An+m−1 is a canonical (n + m)-lattice if and only if: (i) L is canonical; and (ii) for d = k, . . . , 1, the sequence LW dL (An , . . . , An+m−1 ) is lexicographically minimal under the action of Sd given by      π LW dL (An , . . . , An+m−1 ) = LW dL π Aπ−1 (n) , . . . , π Aπ−1 (n+m−1) for π ∈ Sd , where we define Sk = Stab(L) × Sym({n, . . . , n + m − 1}) and Sd−1 = Sd ∩ Stab LW dL (An , . . . , An+m−1 ) for d = k, . . . , 2. Proof. Since the sequence LW L (An , . . . , An+m−1 ) is the concatenation of the sequences LW kL (An , . . . , An+m−1 ), . . . , LW 1L (An , . . . , An+m−1 ) in this order, condition (ii) from Theorem 31 is equivalent to condition (ii) of this corollary. Remark 33. Corollary 32 makes it possible to construct the lattice-antichains An , . . . , An+m−1 level by level: The comparison at step d in condition (ii) only involves the elements of An , . . . , An+m−1 that live on the level d of L. In particular, the benefits of using stabiliser chains mentioned in Section 3.1 apply: 12 (i) If the test at level d fails, the levels d − 1, . . . , 1 do not have to be constructed; an entire branch of the search space is discarded in one step. (ii) The cost of testing condition (ii) of Corollary 32 is in general much smaller than the cost of testing condition (ii) of Theorem 31: The former is proportional to k X d=1 S LW dL (An , . . . , An+m−1 ) d , while the latter is proportional to k Y S LW L (An , . . . , An+m−1 ) k = d=1 S LW dL (An , . . . , An+m−1 ) d . Figure 6 shows the comparisons that are made when testing one step of condition (ii) of Corollary 32. Note that a reordering of the rows of the matrix does not change the position at which the lexicographic comparison stops; this property is necessary for the stabiliser chain approach to work. 4 4.1 Implementation and results Implementation notes This section sketches some ideas that are crucial for an efficient implementation of the algorithm presented in the preceding sections. 4.1.1 Representing antichains using up-closed sets While the theoretical results of Section 3 are formulated in terms of antichains, it is easier and computationally more efficient to work with sets S that are upclosed, meaning that ↑ S = S holds. (For instance, testing whether A ⊆ L is a lattice-antichain for L only involves ↑L A.) Clearly, if A is an antichain, then ↑L A is up-closed and the set of minimal elements of ↑L A is equal to A. n n+m−1 n−1 2 Figure 6: Lexicographic comparison of a lattice LAn ,...,An+m−1 obtained from a sequence of partially constructed lattice-antichains and a relabelling; only the relevant part of the matrix is shown. Thick lines indicate the boundaries between levels. Parts of the lattice-antichains not yet constructed are shown in grey. Parts of the lattice-antichains known to coincide are shown in black. 13 Lemma 34. Let L be a levellised n-lattice with depL (n − 1) = k, let A and B be antichains in L, and let ℓ ∈ {1, . . . , k}. The following are equivalent:   (i) One has wtL A ∩ levd (L) = wtL B∩ levd (L) for d = k, . . . , ℓ − 1, and wtL A ∩ levℓ (L) < wtL B ∩ levℓ (L) .   for d = k, . . . , ℓ−1, (ii) One has wtL (↑L A)∩lev d(L) = wtL (↑L B)∩levd (L)  and wtL (↑L A) ∩ levℓ (L) < wtL (↑L B) ∩ levℓ (L) . Proof. If (i) holds, one has ((↑L A) \ A) ∩ levd (L) = ((↑L B) \ B) ∩ levd (L) for d = k, . . . , ℓ, since L is levellised. In particular, one has  (↑L B) \ (↑L A) ∩ levℓ (L) = (B \ A) ∩ levℓ (L) and  (↑L A) \ (↑L B) ∩ levℓ (L) = (A \ B) ∩ levℓ (L) , which together with (i) imply (ii). As A and B are the sets of minimal elements of ↑L A respectively ↑L B, the converse implication is obvious. Corollary 35. Let L be a levellised n-lattice with depL (n − 1) = k, let m ∈ N+ , and assume that Ai ⊆ L \ {0} for i = n, . . . , n + m − 1 satisfy condition (B) from Theorem 29. Then LAn ,...,An+m−1 is a canonical (n + m)-lattice if and only if: (i) L is canonical; and (ii) for d = k, . . . , 1, the sequence LW dL (↑L An , . . . , ↑L An+m−1 ) is lexicographically minimal under the action of Sd as in Corollary 32. Proof. The claim follows  from Corollary 32 with Lemma 34 and the observation that one has ↑L π(A) = π(↑L A) for any A ⊆ L and π ∈ Sd . 4.1.2 Packed representation of antichains and Beneš networks Let L be a canonical n-lattice with depL (n − 1) = k, and let m ∈ N+ . To generate the descendants of L with m elements on level k +1, we use a backtrack search to construct the sets (↑L Ai ) ∩ levd (L) for d = k, . . . , 1 (outer loop) and i = n, . . . , n + m − 1 (inner loop). Every time a candidate set (↑L Ai ) ∩ levd (L) has been chosen, we use condition (B) from Theorem 29 to check for possible contradictions (backtracking if there are any), and to keep track of any elements whose presence in (↑L Ai ) ∩ levd′ (L) for some d > d′ ≥ 1 is forced by the choices made so far (restricting the possible choices later in the backtrack search if there are any). Once all candidate sets on the current level have been chosen, we check for minimality under the action of the appropriate stabiliser Sd (cf. Corollary 35), backtracking if necessary. Given the large number of configurations that have to be generated and tested for canonicity, it is critical to use an efficient data structure to store a configuration of antichains. The sets (↑L An )∩levd (L), . . . , (↑L An+m−1 )∩levd (L) are encoded as a single (m · |levd (L)|)-bit integer. That way, a lexicographic comparison of two configurations reduces to a single comparison of two (m · |levd (L)|)-bit integers. 14 When constructing lattices with up to 18 elements, m·|levd (L)| is at most 64; when constructing lattices with up to 23 elements, m · |levd (L)| is at most 128. Thus, on a 64-bit CPU, a lexicographic comparison of two configurations costs only very few clock cycles. To be able to apply permutations to a packed representation as described above effectively, we pre-compute a Beneš network [Knu09, § 7.1.3] for each generator of the stabiliser Sd . That way, the application of the generator to the sequence (↑L An ) ∩ levd (L), . . . , (↑L An+m−1 ) ∩ levd (L) is realised by a sequence of bitwise operations (XOR and shift operations) on the (m · |levd (L)|)-bit integer representation.  If the sequence (↑L An ) ∩ levd (L), . . . , (↑L An+m−1 ) ∩ levd (L) is lexicographically minimal in its orbit under the action of Sd , then the computation of this orbit also yields generators of Sd−1 [Cam99, § 1.13]; we limit the number of generators by applying a technique known as Jerrum’s filter [Cam99, § 1.14]. 4.1.3 Vertically indecomposable lattices Restricting the construction to vertically indecomposable lattices is very easy: By the following lemma, the only modification required is to avoid adding a new level that contains a single element whose covering set equals the lowest non-trivial level of the given lattice in any step of the construction. Lemma 36. Let L be a levellised n-lattice, let m ∈ N+ , and let Ai ⊆ L \ {0} for i = n, . . . , n + m − 1 satisfy condition (B) from Theorem 29. (a) If L is vertically decomposable, then LAn ,...,An+m−1 is vertically decomposable. (b) If L is vertically indecomposable, then LAn ,...,An+m−1 is vertically decomposable if and only if m = 1 and ↑L An = L \ {0} hold. e = LAn ,...,An+m−1 . Proof. Let k = depL (n − 1) and let L (a) As L is vertically decomposable, there exists i ∈ L \ {0, 1} such that i is comparable to every element of L. In particular, j ⊑L i holds for any j ∈ levk (L), since L is levellised. For any a ∈ {n, . . . , n + m − 1}, one has e Aa ∩ levk (L) 6= ∅, and thus a ⊑Le i, so i is comparable to every element of L. e is vertically decomposable if and only if there exists (b) This is obvious, as L e \ {0}. a ∈ {n, . . . , n + m − 1}, such hat one has a ⊑Le i for all i ∈ L 4.1.4 Graded lattices Definition 37. A lattice L is called graded with rank function ρ : L → N if one has ρ(a) = ρ(b) + 1 for any a, b ∈ L satisfying a ≺ L b. Remark 38. It is clear from the definitions that a lattice L is graded if and only if depL is a rank function for L, that is, if and S only ifca ≺L b implies depL (a) = depL (b) + 1; the latter is equivalent to a∈levℓ (L) L a = levℓ−1 (L) for ℓ = 1, . . . , depL (0). 15 Lemma 39. Let L be a levellised n-lattice with depL (n − 1) = k, let m ∈ N+ , and assume that Ai ⊆ L \ {0} for i = n, . . . , n + m − 1 satisfy condition (B) from Theorem 29. Then LAn ,...,An+m−1 is graded if and only if (a) L is graded; and (b) n+m−1 S Ai = levk (L) holds. i=n e = LAn ,...,An+m−1 . By Theorem 29, one has dep e (a) = depL (a) for Proof. Let L L a ∈ L \ {0}, and depLe (i) = k + 1 = depL (0) for n ≤ i < n + m. By construction, e for ℓ = 0, . . . , k, as well depLe (0) = k + 2 holds. Thus, one has levℓ (L) = levℓ (L) e e as levk+1 (L) = {n, . . . , n + m − 1} and levk+1 (L) = {0} = levk+2 (L). As a ≺L b is equivalent to a ≺Le b for a, b ∈ L \ {0} and a ∈ L \ {0} (resp. e e if and only if a covers no other element of L a ∈ L\{0}) covers 0 in L (resp. in L) c e (resp. of L), the claim follows with Remark 38, noting that one has Le i = Ai for i = n, . . . , n + m − 1 by construction. In order to restrict the construction to graded lattices, it is thus sufficient to enforce condition (b) of Lemma 39 in every step of the construction. In the light of Lemma 36, the construction can be restricted to vertically indecomposable graded latticed by avoiding to add a new level that contains a single element in any step of the construction. Note, however, that the stabiliser chain approach does not yield any benefit for graded lattices compared to [HR02, JL15], since by condition (b) of Lemma 39, the construction of covering sets only involves a single level anyway. 4.2 Results and performance Table 1 shows the number in of isomorphism classes of vertically indecomposable unlabelled lattices on n elements, and the number un of isomorphism classes of unlabelled lattices on n elements for n ≤ 20; the values i20 and u20 are new. Table 2, Figure 7 and Figure 8 show the total CPU time and the real time taken by the computations for n ≥ 14 using the algorithm described in this paper and the algorithm from [JL15] for two hardware configurations: (A) 4 threads on a system with one 4-core Intel Xeon E5-1620 v2 CPU (clock frequency 3.70 GHz) with DDR3-1600 RAM (single thread bandwidth1 14.9 GB/s; total bandwidth 35.3 GB/s). The system load was just over 4 during the tests. (B) 20 threads on a system with two 10-core Intel Xeon E5-2640 v4 CPUs (clock frequency 2.60 GHz) with DDR4-2400 RAM (single thread bandwidth 9.9 GB/s; total bandwidth 89.2 GB/s). The system load was just over 20 during the tests. The algorithm described in this paper was implemented in C using pthreads by the first author. The authors of [JL15] kindly provided C-code using MPI implementing their algorithm. All code was compiled using GCC with maximal optimisations for the respective architecture. The compiler version was 4.8.1 for hardware configuration (A) and 4.8.5 for hardware configuration (B). 1 http://zsmith.co/bandwidth.html 16 n 1 2 3 4 5 6 7 8 9 10 11 in un 1 1 1 1 0 1 1 2 2 5 7 15 27 53 126 222 664 1 078 3 954 5 994 26 190 37 622 n 12 13 14 15 16 in un 190 754 262 776 1 514 332 2 018 305 12 998 035 16 873 364 119 803 771 152 233 518 1 178 740 932 1 471 613 387 n 17 18 19 20 in un 12 316 480 222 15 150 569 446 136 060 611 189 165 269 824 761 1 582 930 919 092 1 901 910 625 578 19 328 253 734 491 23 003 059 864 006 15 16 17 18 19 20 § 3-4 (B) 14 CPU real 8.69 4.07 58.1 16.8 550 142 5 840 1 490 66 600 16 900 — — — — § 3-4 [JL] (A) n CPU real 73.4 18.4 728 183 7 830 1 990 90 900 23 000 — — — — — — CPU real 10.6 3.27 72.9 7.01 695 42.5 7 360 428 84 200 4 900 1 030 000 61 500 13 000 000 805 000 [JL] Table 1: Numbers in and un of isomorphism classes of vertically indecomposable unlabelled lattices, respectively arbitrary unlabelled lattices, on n elements. CPU real 165 8.49 1 610 80.8 17 100 854 196 000 9 780 2 400 000 120 000 — — — — Table 2: Total CPU time and real time for the longer computations from Table 1 using the algorithm presented in this paper (labelled § 3-4) and the algorithm from [JL15] (labelled [JL]). Times are given in seconds. 17 1012 105 in 1011 in 1010 103 109 § 3-4, CPU § 3-4, real [JL], CPU [JL], real 108 107 14 15 16 n 17 18 time [s] 104 102 101 100 Figure 7: Growth of the number in of vertically indecomposable lattices with n elements, as well as of the CPU time and real time taken for their enumeration on hardware configuration (A), in terms of n. 1014 107 in 1013 106 1012 in 1011 104 1010 103 109 § 3-4, CPU § 3-4, real [JL], CPU [JL], real 108 107 14 15 16 17 n 18 19 20 time [s] 105 102 101 100 Figure 8: Growth of the number in of vertically indecomposable lattices with n elements, as well as of the CPU time and real time taken for their enumeration on hardware configuration (B), in terms of n. 18 Remark 40. 1. For longer enumerations, the speedup compared to the algorithm from [JL15] is between a factor of 13 (hardware configuration (A), n = 15) and a factor of 30 (hardware configuration (B), n = 18). On both hardware configurations, the speedup increases with n. This is expected in the light of Remark 33, as the benefits of the stabiliser chain approach are the more pronounced, the more levels a lattice has. 2. Obtaining a meaningful complexity analysis seems out of reach, as estimating the average case complexity would require a detailed understanding of the tree of canonical lattices. Experimentally, the algorithm seems to be close to optimal in the sense that the computation time grows roughly linearly in the number of lattices constructed. 3. On hardware configuration (A), the L2 and L3 cache hit rates during the computations were on average around 55-60% respectively 80-85%; on hardware configuration (B), this information could not be obtained. The high L3 cache hit hate suggests that DRAM bandwidth is not a significant limiting factor for the overall performance of the algorithm. 4. The throughput corresponds to roughly 1 800 CPU clock cycles per lattice on hardware configuration (A) and to roughly 1 750 CPU clock cycles per lattice on hardware configuration (B); these figures include pre-computations and inter-thread communication. This similarity of these values in spite of the different DRAM speeds on hardware configuration (A) and hardware configuration (B) is consistent with DRAM bandwidth not being a significant limiting factor for the overall performance. 5. C source code implementing the described algorithm is available at the URL https://bitbucket.org/vgebhardt/unlabelled-lattices under the GNU GPL v. 3+. References [Cam99] Peter J. Cameron. Permutation groups, volume 45 of London Mathematical Society Student Texts. Cambridge University Press, Cambridge, 1999. MR1721031 [HR02] Jobst Heitzig and Jürgen Reinhold. Counting finite lattices. Algebra Universalis, 48(1):43–53, 2002. MR1930032 [JL15] Peter Jipsen and Nathan Lawless. Generating all finite modular lattices of a given size. Algebra Universalis, 74(3-4):253–264, 2015. MR3397437 [KL71] Walter Klotz and Lutz Lucht. Endliche Verbände. J. Reine Angew. Math., 247:58–68, 1971. MR0276146 [Knu09] Donald E. Knuth. The art of computer programming. Vol. 4, Fasc. 1. Addison-Wesley, Upper Saddle River, NJ, 2009. Bitwise tricks & techniques; Binary decision diagrams. 19 [KW80] D. J. Kleitman and K. J. Winston. The asymptotic number of lattices. Ann. Discrete Math., 6:243–249, 1980. Combinatorial mathematics, optimal designs and their applications (Proc. Sympos. Combin. Math. and Optimal Design, Colorado State Univ., Fort Collins, Colo., 1978). MR593536 [McK98] Brendan D. McKay. Isomorph-free exhaustive generation. gorithms, 26(2):306–324, 1998. MR1606516 [Slo] J. Al- Neil J. A. Sloane. The on-line encyclopedia of integer sequences. https://oeis.org. Sequences A006966 and A058800. Volker Gebhardt [email protected] Stephen Tawn [email protected] http://www.stephentawn.info Western Sydney University Centre for Research in Mathematics Locked Bag 1797, Penrith NSW 2751, Australia http://www.westernsydney.edu.au/crm 20
4math.GR
1 Wyner’s Common Information under Rényi Divergence Measures Lei Yu and Vincent Y. F. Tan, Senior Member, IEEE arXiv:1709.02168v2 [cs.IT] 25 Jan 2018 Abstract We study a generalized version of Wyner’s common information problem (also coined the distributed source simulation problem). The original common information problem consists in understanding the minimum rate of the common input to independent processors to generate an approximation of a joint distribution when the distance measure used to quantify the discrepancy between the synthesized and target distributions is the normalized relative entropy. Our generalization involves changing the distance measure to the unnormalized and normalized Rényi divergences of order α = 1 + s ∈ [0, 2]. We show that the minimum rate needed to ensure the Rényi divergences between the distribution induced by a code and the target distribution vanishes remains the same as the one in Wyner’s setting, except when the order α = 1+s = 0. This implies that Wyner’s common information is rather robust to the choice of distance measure employed. As a byproduct of the proofs used to the establish the above results, the exponential strong converse for the common information problem under the total variation distance measure is established. Index Terms Wyner’s common information, Distributed source simulation, Rényi divergence, Total variation distance, Exponential strong converse I. I NTRODUCTION How much common randomness is needed to simulate two correlated sources in a distributed fashion? This problem, termed distributed source simulation, was first studied by Wyner [1], who used the normalized relative entropy (Kullback-Leibler divergence or KL divergence) to measure the approximation level (discrepancy) between the simulated joint distribution and the joint distribution of the original correlated sources. He defined the minimum rate needed to ensure that the normalized relative entropy vanishes asymptotically as the common information between the sources. He also established a single-letter characterization for the common information, i.e., the common information between correlated sources X and Y (with target distribution πXY ) is CWyner (X; Y ) = min I(XY ; W ). (1) PXY W : PXY =πXY , X−W −Y The common information is also known to be one of many reasonable measures of the dependence between two random variables [2, Section 14.2.2] (other measures include the mutual information and the Gács-Körner-Witsenhausen common information). A related notion is that of the exact common information which was introduced by Kumar, Li, and El Gamal [3]. They assumed variable-length codes and exact generation of the correlated sources (X, Y ), instead of block codes and approximate simulation of πXY as assumed by Wyner [1]. The exact common information is not smaller than Wyner’s common information. However, it is still not known whether they are equal in general. Furthermore, the common information problem can be also be regarded as a distributed coordination problem. The concept of coordination was first introduced by Cuff, Permuter, and Cover [4], [5], who used the total variation (TV) distance to measure the level of approximation between the simulated and target distributions. Wyner’s common information problem is also closely related to the channel resolvability problem, which was first studied by Han and Verdú [6], and subsequently studied by Hayashi [7], [8], Liu, Cuff, and Verdú [9], and Yu and Tan [10] among others. For the achievability part, both problems rely on so-called soft-covering lemmas [5]. The channel resolvability or common information problems have several interesting applications—including secrecy, channel synthesis, and source coding. For example, in [11] it was used to study the performance of a wiretap channel system under different secrecy measures. In [12] it was used to study the reliability and secrecy exponents of a wiretap channel with cost constraints. In [13] it was used to study the exact secrecy and reliability exponents for a wiretap channel. A. Main Contributions Different from Wyner’s work, we use (normalized and unnormalized) Rényi divergences of order 1 + s ∈ [0, 2] to measure the level of approximation between the simulated and target distributions. This is motivated in part by our desire to understand the sensitivity of the divergence as approximation measure on Wyner’s common information. We prove that for the distributed The authors are with the Department of Electrical and Computer Engineering, National University of Singapore (NUS) (Emails: [email protected], [email protected]). The second author is also with the Department of Mathematics, NUS. 2 source simulation problem, the minimum rate needed to guarantee that the (normalized and unnormalized) Rényi divergences vanish asymptotically is equal to Wyner’s common information (except for the case when Rényi parameter is equal to 0). This implies that Wyner’s common information in (1) is rather robust to the distance measure. For the achievability part, by using the method of types and typicality arguments, we prove that the optimal Rényi divergences vanish (at least) exponentially fast if the code rate is larger than Wyner’s common information. However, for the converse part, the proof is not straightforward and we have to first consider an auxiliary problem. We first prove an exponential strong converse for the common information problem under the TV distance measure, i.e., when the code rate is smaller than Wyner’s common information, the TV distance between the induced distribution and the target distribution tends to one (at least) exponentially fast. Even though our proof technique mirrors that of Oohama [14] to establish the exponential strong converse for the Wyner-Ziv problem, it differs significantly in some aspects. To wit, some intricate continuity arguments are required to assert that the strong converse exponent is positive for all rates below CWyner (X; Y ) (see part (i) of Lemma 1). Furthermore and interestingly, by leveraging a key relationship between the Rényi divergence and the TV distance [15], this exponential strong converse implies the converse for the normalized Rényi divergence (which in turn also implies the strong converse for the unnormalized Rényi divergence). It is worth noting that it is quite natural to use various divergences to measure the discrepancy between two distributions. Wyner [1] used the KL divergence to measure the level of approximation in the distributed source synthesis problem; Hayashi [7], [8] and Yu and Tan [10] respectively used the KL divergence and the Rényi divergence to study the channel resolvability problem. The latter also applied their results to study the capacity region for the wiretap channel under these generalized measures. Furthermore, in probability theory, Barron [16] and Bobkov, Chistyakov and Götze [17] respectively used the KL divergence and the Rényi divergence to study the central limit theorem, i.e., they used them to measure the discrepancy between the induced distribution of sum of i.i.d. random variables and the normal distribution with the same mean and variance. Furthermore, special instances of Rényi entropies and divergences—including the KL divergence, the collision entropy (the Rényi divergence of order 2), and min-entropy (the Rényi divergence of order ∞)—were used to study various information-theoretic problems (including security, cryptography, and quantum information) in several works in the recent literature [10], [11], [18]–[22]. B. Notation We use PX (x) to denote the probability distribution of a random variable X. This will also be denoted as P (x) (when the random variable X is clear from the context). We also use PeX , PbX and QX to denote various probability distributions with alphabet X . All alphabets considered in the sequel are finite. The set of probability measureson X is denoted as P (X ), and the set of conditional probability measures on Y given a variable in X is denoted as P(Y|X ) := PY |X : PY |X (·|x) ∈ P(Y), x ∈ X . Furthermore, the support P of a distribution P ∈ P(X ) is denoted as supp(P ) = {x ∈ X : P (x) > 0}. n We use Txn (x) := n1 i=1 1 {xi = x} to denote the type (empirical distribution) of a sequence xn , TX and VY |X to respectively denote a type of sequences in X n and a conditional type of sequences in Y n (given a sequence xn ∈ X n ). For a type TX , the type class (set of sequences having the same type TX ) is denoted by TTX . For a conditional type VY |X and a sequence xn , the VY |X -shell of xn (the set of y n sequences having the same conditional type VY |X given xn ) is denoted by TVY |X (xn ). For brevity, sometimes we use T (x, y) to denote the joint distributions T (x)V (y|x) or T (y)V (x|y). The ǫ-typical set of QX is denoted as Tǫn (QX ) := {xn ∈ X n : |Txn (x) − QX (x)| ≤ ǫQX (x), ∀x ∈ X } . (2) The conditionally ǫ-typical set of QXY is denoted as Tǫn (QY X |xn ) := {y n ∈ Y n : (xn , y n ) ∈ Tǫn (QXY )} . For brevity, sometimes we write and as and respectively. The TV distance between two probability mass functions P and Q with a common alphabet X is defined as 1 X |P (x) − Q(x)|. |P − Q| := 2 Tǫn (QX ) Tǫn (QY X |xn ) Tǫn (3) Tǫn (xn ) (4) x∈X By the definition of ǫ-typical set, we have that for any xn ∈ Tǫn (QX ), ǫ (5) |Txn − QX | ≤ . 2 Fix distributions PX , QX ∈ P(X ). The relative entropy and the Rényi divergence of order 1 + s are respectively defined as X PX (x) (6) PX (x) log D(PX kQX ) := QX (x) x∈supp(PX ) D1+s (PX kQX ) := 1 log s X x∈supp(PX ) PX (x)1+s QX (x)−s , (7) 3 ✒ PX n |Mn X n✲ Mn ❅ ❅ ❘ PY n |Mn ❅ Y n✲ Fig. 1. Distributed source synthesis problem, where the random variable Mn ∈ Mn := {1, . . . , enR }. and the conditional versions are respectively defined as D(PY |X kQY |X |PX ) := D(PX PY |X kPX QY |X ) D1+s (PY |X kQY |X |PX ) := D1+s (PX PY |X kPX QY |X ), (8) (9) where the summations in (6) and (7) are taken over the elements in supp(PX ). Throughout, log is to the natural base e and s ≥ −1. It is known that lims→0 D1+s (PX kQX ) = D(PX kQX ) so a special case of the Rényi divergence (or the conditional version) is the usual relative entropy (or the conditional version). Given a number a ∈ [0, 1], we define ā = 1 − a. We also define [x]+ = max {x, 0}. C. Problem Formulation In this paper, we consider the distributed source simulation problem illustrated in Fig. 1. Given a target distribution πXY , we wish to minimize the alphabet size of a random variable Mn that is uniformly distributed over1 Mn := {1, . . . , enR } (R is a positive number known as the rate), such that the generated (or synthesized) distribution PX n Y n (xn , y n ) := X 1 PX n |Mn (xn |m)PY n |Mn (y n |m) |Mn | (10) m∈Mn n . The pair of random mappings (PX n |Mn , PY n |Mn ) forms a good approximation to the product distribution πX n Y n := πXY constitutes a synthesis code. Different from Wyner’s seminal work on the distributed source simulation problem [1], we employ the unnormalized Rényi divergence (11) D1+s (PX n Y n kπX n Y n ) and the normalized Rényi divergence 1 D1+s (PX n Y n kπX n Y n ) n (12) to measure the discrepancy between PX n Y n and πX n Y n . The minimum rates required to ensure these two measures vanish asymptotically are respectively termed the unnormalized and normalized Rényi common information, and denoted as o n (13) T1+s (πXY ) := inf R : lim D1+s (PX n Y n kπX n Y n ) = 0 , n→∞ o n 1 (14) Te1+s (πXY ) := inf R : lim D1+s (PX n Y n kπX n Y n ) = 0 . n→∞ n It is clear that Te1+s (πXY ) ≤ T1+s (πXY ). (15) We also denote the minimum rate required to ensure the TV distance is bounded above by some constant ε ∈ [0, 1] asymptotically as o n (16) TεTV (πXY ) := inf R : lim sup |PX n Y n − πX n Y n | ≤ ε . n→∞ We say that the strong converse property for the common information problem under the TV distance holds if TεTV (πXY ) does not depend on ε ∈ [0, 1). 1 For simplicity, we assume that enR and similar expressions are integers. 4 II. M AIN R ESULTS Our main result concerns Wyner’s common information problem when the discrepancy measure is the unnormalized or normalized Rényi divergence. It is stated as follows. Theorem 1 (Rényi Common Informations). The unnormalized and normalized and Rényi common informations satisfy ( CWyner (X; Y ) s ∈ (−1, 1] T1+s (πXY ) = Te1+s (πXY ) = . (17) 0 s = −1 Furthermore, for s ∈ (−1, 1], the optimal Rényi divergence D1+s (PX n Y n kπX n Y n ) in the definitions of the Rényi common informations decays at least exponentially fast in n when R > CWyner (X; Y ). Remark 1. For the converse part, T1+s (πXY ) ≥ Te1+s (πXY ) ≥ CWyner (X; Y ) for s ∈ [0, 1] is implied by Wyner’s work [1] and the monotonicity of the Rényi divergence. For the achievability part, Te1+s (πXY ) ≤ CWyner (X; Y ) for s ∈ (−1, 0] is also implied by Wyner’s work [1] and the monotonicity of the Rényi divergence. Furthermore, since a channel resolvability code for the memoryless channel PX|W × PY |W can be used to form a common information code, the achievability part for the common information problem can be obtained from existing channel resolvability results. Specifically, T1+s (πXY ) ≤ CWyner (X; Y ) for s ∈ (−1, 0] can be obtained from Hayashi’s [7], [8] or Han, Endo, and Sasaki’s results [12]. In addition, Te1+s (πXY ) ≤ T1+s (πXY ) ≤ C1+s (X; Y ) for s ∈ (0, 1] with X  C1+s (X; Y ) := min (18) PW (w) D1+s PX|W (·|w)PY |W (·|w)kPXY PXY W : PXY =πXY , X−W −Y w can be obtained from the present authors’ results [10], but as shown in Theorem 1, this bound is not tight since C1+s (X; Y ) > CWyner (X; Y ) in general for s ∈ (0, 1]. This is because, on the one hand, for the channel resolvability problem, the discrete memoryless channel is fixed, and, by construction, imposes a product conditional distribution of the output given the input (which is a product distribution), but for the common information problem, the synthesizer has the freedom to choose PX n Y n |Mn = PX n |Mn × PY n |Mn , so that the Markov chain X n − Mn − Y n holds; on the other hand, for the common information problem, in the sequel, we will show that if we utilize a truncated channel (which is not memoryless) as the synthesizer. This results in a smaller achievable rate for the case s ∈ (0, 1]. Therefore, our converse for s ∈ [−1, 0) and achievability for s ∈ (0, 1] are new (and also tight). Remark 2. An exponential achievability result for s ∈ (−1, 0] can be obtained from Hayashi’s [7], [8] and Han, Endo, and Sasaki’s results [12], where i.i.d. codes were employed. For this theorem, the proof of the converse part for the unnormalized Rényi common information is provided in Appendix A, and the proof of the converse part for the normalized Rényi common information is provided in Section IV. Observe that the unnormalized Rényi divergence is stronger than the normalized one in the sense of (15), hence Te1+s (πXY ) ≤ T1+s (πXY ). This implies, on one hand, the achievability result for the normalized Rényi common information Te1+s (πXY ) can be obtained directly from the achievability result for the unnormalized version T1+s (πXY ), and on the other hand, the converse result for the normalized Rényi common information Te1+s (πXY ) implies the converse result for the unnormalized version T1+s (πXY ). The Rényi common informations are the same for all s ∈ (−1, 1], and also same as Wyner’s common information CWyner (X; Y ) (which corresponds to s = 0 for the normalized case). For the case s ∈ (−1, 1], to obtain the (unnormalized and normalized) Rényi common informations, we utilize a random code with (W n , X n , Y n ) (W is the auxiliary random variable in the definition of CWyner (X; Y )) distributed according to a truncated product distribution, i.e., a product distribution governed by QnW XY but whose mass is truncated to the typical set Tǫn (QW XY ).2 On one hand, the random sequences (W n , X n , Y n ) so generated are almost uniformly distributed over the typical set Tǫn (QW XY ); and on the other hand, the Rényi common informations can be expressed as some Rényi divergences. Moreover, these Rényi divergences evaluated at the truncated distribution are almost the same regardless of the parameter s ∈ (−1, 1]. Therefore, by using this truncated code, Wyner’s common information is achievable for any s ∈ (−1, 1]. However, the proof of the converse part for the normalized Rényi common information is not straightforward.3 We attempted to use the method of types to prove it, just as in [10] for the Rényi resovability problem, but failed since the code for the common information problem is arbitrary and does not need to be i.i.d. In particular, it is not i.i.d. In the following two 2 Interestingly, a truncated code is not necessary for s ∈ (−1, 0] case, since an i.i.d. code (without truncation) is optimal as well for this case. This point can be seen from the work of Yu and Tan [10]. 3 More precisely, the proof of the converse part for s ∈ (−1, 0) case is not easy. The converse part for s = 0 case was proven by Wyner [1], in which the 1 n ) is small then 1 H (X n Y n ) D(PX n Y n kπXY continuity of the normalized entropy under the normalized KL divergence measure was used, i.e., when n n P is arbitrarily close to Hπ (XY ). But it is not straightforward to apply Wyner’s proof to the case s ∈ (−1, 0), since we do not know whether a strong enough continuity condition for the normalized entropy holds under the normalized Rényi divergence measure with order α = 1 + s ∈ (0, 1). Even if a strong enough continuity condition holds, it is not straightforward to prove. Note that we were not able to directly utilize ideas in Wyner’s converse proof to demonstrate p (u) 1 this point. The definition of the “relative entropy typical set” A(n, ǫ1 ) := u : n log p1 (u) ≤ ǫ1 is crucial in Wyner’s proof. If we adopt this set with this 0 definition for the Rényi divergence setting (with Rényi parameter < 1), it is not clear to us whether P1 (Ac (n, ǫ1 )) vanishes (cf. Equation (A.7) in Wyner’s paper). 5 sections, we provide an indirect proof using the following strategy: We first prove an exponential strong converse for Wyner’s common information problem under the TV distance measure in Section III. Then by using a relationship between the Rényi divergence and the TV distance [15], we show this exponential strong converse implies the converse for normalized Rényi divergence in Section IV. As an intermediate result, the common information under the TV distance measure is characterized in the following theorem. Theorem 2 (Common Information under the TV Distance Measure). The following hold: (i) The common information under the TV distance measure satisfies ( CWyner (X; Y ) ε ∈ [0, 1) TV Tε (πXY ) = . 0 ε=1 (19) Hence, the strong converse property for the common information problem under the TV distance holds. (ii) Furthermore, there exists a sequence of synthesis codes with rate R > CWyner (X; Y ), such that |PX n Y n − πX n Y n | tends to zero exponentially fast as n tends to infinity. (iii) On the other hand, for any sequence of synthesis codes with rate R < CWyner (X; Y ), we have that |PX n Y n − πX n Y n | tends to one exponentially fast as n tends to infinity. Part (ii) is an exponential achievability result while the part (iii) is an exponential strong converse result. Combining parts (ii) and (iii) implies part (i). By Pinsker’s inequality for Rényi divergences [23], the achievability results (including the exponential achievability result) in Theorem 1 implies the achievability results (including the exponential achievability result) in Theorem 2. Conversely, the exponential strong converse result in part (iii) of Theorem 2 implies the converse results in Theorem 1 for both unnormalized and normalized Rényi divergences. To prove part (iii), we draw on several key ideas from Oohama’s work [14] on the exponential strong converse for the Wyner-Ziv problem. However, there are several key differences in our proofs, including the way we establish that the strong converse exponent is positive for all rates larger than CWyner (X; Y ) and the treatment of the cases when various probability mass functions take on the value zero. We note that conclusion in part (ii) (the exponential achievability result) in Theorem 2 can be also obtained by using the soft-covering lemma by Cuff [5, Lemma IV.1]. The proof of the conclusion in part (iii) is provided in the next section. As mentioned above, the other parts follow directly from Theorem 1. III. T HE P ROOF OF PART ( III ) IN T HEOREM 2 In this section, we provide an exponential strong converse theorem for the common information problem under the TV distance measure, which will be used to derive the converse for normalized Rényi divergence in next section. We define n o Q := QXY U ∈ P(X × Y × U) : |U| ≤ |X ||Y|, supp(QXY ) ⊆ supp(πXY ) . (20) Given α ∈ [0, 1] and an arbitrary distribution QXY U ∈ Q, define the linear combination of the likelihood ratios for (x, y, u) ∈ supp(QXY U ), ! QXY |U (x, y|u) QXY |U (x, y|u) QXY (x, y) (α) + α log + log . (21) ωQXY U (x, y|u) := ᾱ log πXY (x, y) QX|U (x|u)QY |U (y|u) πXY (x, y) This function is finite for all (x, y, u) ∈ supp(QXY U ). For QXY U ∈ Q and θ ∈ [0, ∞), define the negative cumulant generating functions as h i (α) (22) Ω(α,θ) (QXY U ) := − log EQXY U exp − θωQXY U (X, Y |U ) , Ω(α,θ) := min QXY U ∈Q Ω(α,θ) (QXY U ), (23) where EQXY U is only taken over the set supp(QXY U ) (this means we only sum over the elements (x, y, u) such that QXY U (x, y, u) > 0). Finally, we define the large deviations rate functions Ω(α,θ) − θαR , 1 + (5 − 3α)θ F (R) := sup F (α,θ) (R). F (α,θ) (R) := (24) (25) (α,θ)∈[0,1]×[0,∞) In view of the definitions above, we have the following theorem. The proof of this theorem is provided in Appendix B. 6 Theorem 3. For any synthesis code such that 1 log |Mn | ≤ R, n we have (26)  |PX n Y n − πX n Y n | ≥ 1 − 4 exp − nF (R) . (27) If we show F (R) > 0, then Theorem 3 implies the exponential strong converse for TV distance measure. To that end, we need the following lemma. Lemma 1. The following conclusions hold. (i) If R < CWyner (X; Y ), then F (R) > 0. (28) F (R) = 0. (29) (ii) If R ≥ CWyner (X; Y ), then The proof of Lemma 1 is provided in Appendix C. We remark that Lemma 1, especially part (i), plays an central role in claiming the exponential strong converse theorem for the common information problem with the TV distance measure. Its proof is completely different from that for the corresponding statement in [14] and requires some intricate continuity arguments (e.g., [25, Lemma 14]). As we have seen in Theorem 3, F (R) in (25) is a lower bound on the exponent of 1−|PX n Y n − πX n Y n |. This can be regarded as the strong converse exponent. Combining Lemma 1 and Theorem 3, we conclude that the exponent in the right hand side of (27) is strictly positive if the rate is smaller than CWyner (X; Y ). Hence, we obtain the exponential strong converse result given in the conclusion (iii) of Theorem 2. IV. C ONVERSE P ROOF OF T HEOREM 1 FOR THE N ORMALIZED R ÉNYI C OMMON I NFORMATION In this section, we provide a proof of the converse part of Theorem 1 for the normalized Rényi common information. To this end, we need the following relationships between the Rényi divergence and the TV distance. Lemma 2 (Relationship between the Rényi Divergence and the TV Distance (Sason [15])). For any s ∈ (−1, +∞), inf PX ,QX :|PX −QX |≥ǫ D1+s (PX kQX ) = = and for any s ∈ (0, 1), inf PX ,QX :|PX −QX |=ǫ inf q∈[0,1−ǫ] D1+s (PX kQX ) d1+s (q + ǫkq),   + 1−s 1 1 inf d1−s (q + ǫkq) ≥ min 1, log − log 2 , q∈[0,1−ǫ] s 1−ǫ s  where d1+s (pkq) := ( 1 s log(p1+s q −s + p̄1+s q̄ −s ), p log pq + p̄ log p̄q̄ , s ≥ −1, s 6= 0 s=0 (30) (31) (32) (33) denotes the binary Rényi divergence of order 1 + s.4 We also have inf PX ,QX :|PX −QX |≥ǫ D0 (PX kQX ) = inf PX ,QX :|PX −QX |=ǫ D0 (PX kQX ) = 0. (34) (35) Remark 3. Pinsker’s inequality provides a lower bound for inf PX ,QX :|PX −QX |≥ǫ D1+s (PX kQX ) or inf PX ,QX :|PX −QX |=ǫ D1+s (PX kQX ), i.e., inf PX ,QX :|PX −QX |=ǫ Hence 4 For (1+s)ǫ2 2 D1+s (PX kQX ) ≥ (1 + s)ǫ2 . 2 is also a lower bound of inf q∈[0,1−ǫ] d1+s (q + ǫkq). s = 0 case, the binary Rényi divergence is known as the binary relative entropy, and it is usually denoted as d(pkq). (36) 7 Remark 4. Using (31) and the lower bound in (32), it is easy to obtain the following improved lower bounds. For any s ∈ (0, 1), inf PX ,QX :|PX −QX |≥ǫ D1−s (PX kQX ) = inf sup D1−t (PX kQX ) (37) D1−t (PX kQX ) (38) PX ,QX :|PX −QX |≥ǫ t∈[s,1) ≥ sup inf t∈[s,1) PX ,QX :|PX −QX |≥ǫ ≥ sup inf t∈[s,1) q∈[0,1−ǫ] d1−t (q + ǫkq)  +   1 1 1−t log − log 2 ≥ sup min 1, t 1−ǫ t t∈[s,1) h i+  1  s ∈ (0, 12 ],  hlog 4(1−ǫ) i+ 1 1 1−s = s ∈ ( 12 , 1), ǫ > 12  s log 1−ǫ − s log 2   0 s ∈ ( 12 , 1), ǫ ≤ 12 and for any s ∈ [0, +∞), inf PX ,QX :|PX −QX |≥ǫ D1+s (PX kQX ) ≥ inf D1−t (PX kQX ) (43) inf inf t∈(0,1) q∈[0,1−ǫ] d1−t (q + ǫkq)   + 1−t 1 1 min 1, log − log 2 t 1−ǫ t t∈(0,1)  + 1 = log . 4 (1 − ǫ) ≥ sup (41) (42) t∈(0,1) PX ,QX :|PX −QX |≥ǫ ≥ sup (40) sup D1−t (PX kQX ) PX ,QX :|PX −QX |≥ǫ t∈(0,1) ≥ sup (39)  (44) (45) (46) Remark 5. The improved lower bounds (41) and (46) (or combining (31) and the lower bound in (32)) implies if |PX − QX | → 1, (47) D1+s (PX kQX ) → ∞. (48) then for any s ∈ (−1, +∞), Combining Lemma 2 with Theorem 3, we have the converse part for the normalized Rényi divergence, which implies the strong converse for the unnormalized Rényi divergence. Theorem 4. For any synthesis codes such that lim sup n→∞ 1 log |Mn | < CWyner (X; Y ), n (49) 1 D1+s (PX n Y n kπX n Y n ) > 0. n (50) we have for any s > −1, lim inf n→∞ Remark 6. This theorem establishes the converse part of Theorem 1 for the normalized Rényi common information. Remark 7. Since lim inf n→∞ n1 D1+s (PX n Y n kπX n Y n ) > 0 implies D1+s (PX n Y n kπX n Y n ) → ∞, the theorem above implies the strong converse for the Wyner’s common information problem under the unnormalized Rényi divergence. Proof: Theorem 3 states if 1 n log |Mn | < CWyner (X; Y ), then |PX n Y n − πX n Y n | → 1 exponentially fast. In other words, |PX n Y n − πX n Y n | ≥ 1 − e−nδn , for some sequence δn > 0 such that lim inf n→∞ δn > 0. Therefore, using Lemma 2 we have       1 1−s 1−s 1 n n n n δ − lim inf δn > 0. ) ≥ lim inf min 1, kπ D (P log 2 = min 1, lim inf n X Y 1+s X Y n→∞ n→∞ n→∞ n s ns s This completes the proof. (51) (52) 8 V. C ONCLUSION AND F UTURE W ORK In this paper, we studied a generalized version of Wyner’s common information problem (or the distributed source simulation problem), in which the unnormalized and normalized Rényi divergences were used to measure the level of approximation. We showed the minimum rate needed to ensure that the unnormalized or normalized Rényi divergence vanishes asymptotically remains the same as the one under Wyner’s setting where the relative entropy was used. In the future, we plan to investigate the second-order coding rate for Wyner’s common information under the unnormalized Rényi divergence or the TV distance. For the unnormalized Rényi divergence, the one-shot achievability bound given in Lemma 3 can be used to obtain an achievability bound for the second-order coding rate. In fact, it can easily be shown that the optimal second-order coding rate scales as O( √1n ). For the TV distance, the one-shot achievability bound given by Cuff [5] can be used to derive an achievability bound. However, the converse parts for both cases are not straightforward. One may leverage the perturbation approach [26] used to prove the second-order coding rate for the Gray-Wyner problem in [27], [28]. This is left as future work. Furthermore, we are also interested in various closely-related problems. Among them, the most interesting one is the distributed channel synthesis problem under the Rényi divergence measure: The coordination problem or distributed channel synthesis problem was first introduced by Cuff, Permuter, and Cover [4], [5]. In this problem, an observer (encoder) of a source sequence describes the sequence to a distant random number generator (decoder) that produces another sequence. What is the minimum description rate needed to produce achieve a joint distribution that is statistically indistinguishable, under the TV distance, from the distribution induced by a given channel? For this problem, Cuff [5] provided a complete characterization of the minimum rate. We can enhance the level of coordination by replacing the TV measure with the Rényi divergence. For this enhanced version of the problem, we are interested in characterizing the corresponding admissible rate region. P ROOF OF T HEOREM 1 FOR THE A PPENDIX A U NNORMALIZED R ÉNYI C OMMON I NFORMATION A. Achievability Next we focus on the achievability part. We first consider the case s ∈ (0, 1]. First we introduce the following one-shot achievability bound (i.e., achievability bound for blocklength n equal to 1). Lemma 3 (One-Shot Achievability Bound). [10] Consider a random mapping PX|W and a random codebook U = {W (i)}i∈M with W (i) ∼ PW , i ∈ M, where M = {1, . . . , eR }. We define 1 X PX|W (x|w(m)) (53) PX|U (x| {w(i)}i∈M ) := |M| m∈M Then we have for s ∈ (0, 1], esD1+s (PX|U kπX |PU ) ≤ esD1+s (PX|W kπX |PW )−sR + esD1+s (PX kπX ) ≤ 2e sΓ1+s (PW ,PX|W ,πX ,R) , (54) (55) where Γ1+s (PW , PX|W , πX , R)  := max D1+s (PX|W kπX |PW ) − R, D1+s (PX kπX ) . (56) Remark 8. This lemma provides a one-shot achievability bound for general source synthesis problems, not only for the distributed source synthesis or common information problem as studied in this paper. By setting πX , PX|W , PW , and R to πX n Y n , PX n Y n |W n = PX n |W n PY n |W n ,5 PW n , and nR respectively, Lemma 3 can be used to derive an achievability result for the common information problem. Applying Lemma 3 and taking limits appropriately, we obtain if there exists a sequence of distributions PW n PX n |W n PY n |W n such that limn→∞ D1+s (PX n Y n kπX n Y n ) → 0 and R > lim supn→∞ n1 D1+s (PX n Y n |W n kπX n Y n |PW n ), then there exists a sequence of codes such that lim sup D1+s (PX n Y n |Un kπX n Y n |PUn ) n→∞ o n 1 ≤ lim sup log esD1+s (PX n Y n |W n kπX n Y n |PW n )−nsR + esD1+s (PX n Y n kπX n Y n ) n→∞ s   1 ≤ log lim sup es(D1+s (PX n Y n |W n kπX n Y n |PW n )−nR) + 1 s  n→∞  1 s(n(R−ǫ)−nR) ≤ log lim sup e +1 s n→∞ = 0, 5 The pair (X n , Y n ) plays the role of X in Lemma 3. (57) (58) (59) (60) 9 where (59) follows since R > lim supn→∞ n1 D1+s (PX n Y n |W n kπX n Y n |PW n ) implies there exists a constant ǫ > 0 such that R − ǫ > n1 D1+s (PX n Y n |W n kπX n Y n |PW n ) holds for all sufficiently large n. Therefore, the minimum achievable rate satisfies  inf R : D1+s (PX n Y n |Un kπX n Y n |PUn ) → 0 1 (61) ≤ inf lim sup D1+s (PX n Y n |W n kπX n Y n |PW n ). n {PW n ,PX n |W n ,PY n |W n }∞ : n→∞ n=1 D1+s (PX n Y n kπX n Y n )→0 Let QW XY be a distribution such that QXY = πXY and X − W − Y . For the optimization in (61), to obtain an upper bound, we set the distributions PW n (wn ) ∝ QnW (wn ) 1 {wn ∈ Tǫn′ (QW )} , PX n |W n (xn |wn ) ∝ QnX|W (xn |wn ) 1 {xn ∈ Tǫn (QW X |wn )} , n n PY n |W n (x |w ) ∝ QnY |W n n n (x |w ) 1 {y ∈ Tǫn and n (QW Y |w )} , (62) (63) (64) where 0 < ǫ′ < ǫ ≤ 1. Then we have PX n Y n (xn , y n ) X = PW n (wn ) PX n |W n (xn |wn ) PY n |W n (xn |wn ) (65) wn = X Qn (wn ) 1 {wn ∈ T n′ (QW )} QnX|W (xn |wn ) 1 {xn ∈ Tǫn (QW X |wn )} W ǫ n (T n ) Q QnX|W (Tǫn (QW X |wn ) |wn ) ′ n ǫ W w × ≤ = QnW QnY |W (xn |wn ) 1 {y n ∈ Tǫn (QW Y |wn )} QnY |W (Tǫn (QW Y |wn ) |wn ) P n n n n w n QW XY (w , x , y ) n n n n n (Tǫ′ ) minwn ∈Tǫn′ QX|W (Tǫ (QW X |w ) |w ) minwn ∈Tǫn′ QnY |W (Tǫn (QW Y |wn ) |wn ) πX n Y n (xn , y n ) , 1 − δn (66) (67) (68) where in (68) δn is defined as 1 minus the denominator  This follows since  of (67). Here we claim that δn → 0 as n → ∞. QnW (Tǫn′ ) → 1, minwn ∈Tǫn′ QnX|W Tǫn (QW X |wn )|wn → 1, and minwn ∈Tǫn′ QnY |W Tǫn (QW Y |wn )|wn → 1, where the last two limits hold due to the following lemma.  Lemma 4. Assume 0 < ǫ′ < ǫ ≤ 1, then as n → ∞, QnX|W Tǫn (QW X |wn )|wn converges uniformly6 to 1 (in wn ∈ Tǫn′ (QW )). More explicitly,    1  ǫ−ǫ′ 2 (min) ′ 2 (min)   − nQX|W nQX|W − 1 ǫ−ǫ 1 − QnX|W Tǫn (QW X |wn )|wn ≤ |X | |W| e 3 1+ǫ′ + e 2 1−ǫ′ , (69) (min) for all wn ∈ Tǫn′ (QW ), where QX|W := min(x,w):QX|W (x|w)>0 QX|W (x|w). This lemma is a stronger version of the conditional typicality lemma in [2], since here the probability converges uniformly, instead of converging pointwise. However, the proof is merely a refinement of the conditional typicality lemma [2, Appendix 2A] (by applying the Chernoff bound, instead of the law of large numbers), and hence omitted here. Besides, a similar lemma can be found in [30, Lemma 2.12], which is established based on a slightly different definition of strong typicality. Using this upper bound of PX n Y n (xn , y n ) we have D1+s (PX n Y n kπX n Y n ) X 1 1+s n n −s n n = log PX n Y n (x , y )πX n Y n (x , y ) s n n x ,y X  πX n Y n (xn , y n ) 1+s 1 −s n n πX ≤ log n Y n (x , y ) s 1 − δ n n n x ,y 1+s  1 1 = log s 1 − δn → 0. 6 This (70) (71) (72) (73)  means that for any η > 0, there exists an integer N = Nη such that maxwn ∈T n′ (QW ) 1 − Qn Tǫn (QW X |w n )|w n ≤ η for all n > Nη . X|W ǫ Here the notion of “uniform convergence” is a slightly different from the conventional one [29, Definition 7.7]. In the conventional definition, the domain of the functions are fixed but here, the domain Tǫn′ (QW ) depends on n. 10 Let [TW VX|W ] denote the joint distribution of X and W induced by the type TW and conditional type VX|W . Now define the sets of tuples of types and conditional types:    (TW , VX|W , VY |W ) : ∀w, |TW (w) − QW (w)| ≤ ǫ′ QW (w)  ∀(w, x), [TW VX|W ](w, x) − QW X (w, x) ≤ ǫQW X (w, x) A := , and (74)   ∀(w, y), [TW VY |W ](w, y) − QW Y (w, y) ≤ ǫQW Y (w, y)   (TW , VX|W , VY |W ) : ∀(w, x, y),   [TW VX|W VY |W ](w, x, y) (1 + ǫ)2 (1 − ǫ)2 B := . (75) ≤ ≤   ′ ′ 1+ǫ QW XY (w, x, y) 1−ǫ   In (75), if QW XY (w, x, y) = 0, this imposes that TW VX|W VY |W (w, x, y) = 0. It is easy to verify that A ⊆ B. Let δ1,n and δ2,n be two arbitrary sequences tending to zero as n → ∞. Using these notations, we can write 1 D1+s (PW n X n Y n kPW n πX n Y n ) n X 1 = log P (wn ) (P (xn |wn ) P (y n |wn ))1+s π −s (xn , y n ) (76) ns w n ,xn ,y n X 1 log P (wn , xn , y n ) = ns w n ,xn ,y n !s QnX|W (xn |wn ) 1 {xn ∈ Tǫn (QW X |wn )} QnY |W (xn |wn ) 1 {y n ∈ Tǫn (QW Y |wn )} −s n n πX (77) × n Y n (x , y ) QnX|W (Tǫn (QW X |wn ) |wn ) QnY |W (Tǫn (QW Y |wn ) |wn ) X X 1 = log P (wn , xn , y n ) ns n n n TW ,VX|W ,VY |W w ∈TTW ,x ∈TVX|W (w ), y n ∈TVY |W (w n ) QnX|W (xn |wn ) 1 {xn ∈ Tǫn (QW X |wn )} QnY |W (xn |wn ) 1 {y n ∈ Tǫn (QW Y |wn )} × ≤ 1 log ns × ≤− QnY |W (Tǫn (QW Y |wn ) |wn ) −s n n πX n Y n (x , y ) (78) P (wn , xn , y n ) (TX ,VX|W ,VY |W )∈A w n ∈TTW ,xn ∈TVX|W (w n ), e y n ∈TVY |W (w n ) !s P P n w,x T (w,x) log Q(x|w) n w,y T (w,y) log Q(y|w) e 1 − δ1,n 1 1 log(1 − δ1,n )(1 − δ2,n ) + log n ns × ≤ QnX|W (Tǫn (QW X |wn ) |wn ) X X !s max (TX ,VX|W ,VY |W )∈B max (TX ,VX|W ,VY |W )∈B esn X P w,x 1 − δ2,n X e−ns P x,y T (x,y) log π(x,y) X n (79) P (wn , xn , y n ) n n (TX ,VX|W ,VY |W )∈B w ∈TTW ,x ∈TVX|W (w ), T (w,x) log Q(x|w)+sn T (w, x) log Q (x|w) + w,x X P y n ∈TVY |W (w n ) w,y T (w,y) log Q(y|w)−ns T (w, y) log Q (y|w) − w,y P x,y X T (x,y) log π(x,y) (80) ! T (x, y) log π (x, y) x,y 1 log(1 − δ1,n )(1 − δ2,n ) n ! X (1 + ǫ)2 X (1 − ǫ)2 X Q (w, x) log Q (x|w) + Q (x, y) log π (x, y) Q (w, y) log Q (y|w) − ≤ ′ 1+ǫ 1 − ǫ′ x,y w,x w,y − 1 log(1 − δ1,n )(1 − δ2,n ) n (1 − ǫ)2 (1 + ǫ)2 1 =− (H (X|W ) + H (Y |W )) + HQ (XY ) − log(1 − δ1,n )(1 − δ2,n ) Q Q 1 + ǫ′ 1 − ǫ′ n 4ǫ 1 (1 − ǫ)2 IQ (XY ; W ) + HQ (XY ) − log(1 − δ1,n )(1 − δ2,n ), = ′ ′ 1+ǫ 1−ǫ n where (79) follows from Lemma 4, (80) follows since A ⊆ B, (81) follows since X X P (wn , xn , y n ) ≤ 1, − n n n (TX ,VX|W ,VY |W )∈B w ∈TTW ,x ∈TVX|W (w ), y n ∈TVY |W (w n ) (81) (82) (83) (84) (85) 11 and (83) follows since QXY = πXY . Letting n → ∞ in (84), we have 1 4ǫ (1 − ǫ)2 IQ (XY ; W ) + HQ (XY ). D1+s (PW n X n Y n kPW n πX n Y n ) ≤ ′ 1+ǫ 1 − ǫ′ n→∞ n Combining (86) with (61) and (73), we obtain  4ǫ (1 − ǫ)2 IQ (XY ; W ) + HQ (XY ). inf R : D1+s (PX n Y n |Un kπX n Y n |PUn ) → 0 ≤ 1 + ǫ′ 1 − ǫ′ Since ǫ > ǫ′ > 0 are arbitrary, and HQ (XY ) = Hπ (XY ) ≤ log {|X ||Y|} is bounded, we have  inf R : D1+s (PX n Y n |Un kπX n Y n |PUn ) → 0 ≤ IQ (XY ; W ). (86) lim sup (87) (88) Since the distribution QW XY is arbitrary, we can minimize IQ (XY ; W ) over all distributions satisfying QXY = πXY and X − W − Y . Hence  min IQ (XY ; W ) = CWyner (X; Y ). (89) inf R : D1+s (PX n Y n |Un kπX n Y n |PUn ) → 0 ≤ QXY W : QXY =πXY , X−W −Y Observe that 1 D1+s (PX n Y n |Un kπX n Y n |PUn ) = log EUn s  X xn ,y n n n PX n Y n |Un (x , y |Un )  PX n Y n |Un (xn , y n |Un ) πX n Y n (xn , y n ) s  , (90) where EUn is the expectation taken with respect to the distribution PUn . Hence D1+s (PX n Y n |Un kπX n Y n |PUn ) → 0 implies that there must exist at least one sequence of codebooks indexed by {un }∞ n=1 such that D1+s (PX n Y n |Un =un kπX n Y n ) → 0. Therefore, the Rényi common information for s ∈ (0, 1] is not larger than CWyner (X; Y ). This completes the proof for the case s ∈ (0, 1]. Now we prove the case s ∈ (−1, 0). Since D1+s (PX n Y n kπX n Y n ) is non-decreasing in s, the result for s ∈ (0, 1] implies the achievability result for s ∈ (−1, 0). B. Exponential Achievability Since D1+s (PX n Y n Un kπX n Y n × PUn ) is non-decreasing in s, to prove the exponential result for s ∈ (−1, 1], we only need to show the result holds for s ∈ (0, 1]. To this end, we use the random code given in Appendix A-A. For this code, by Lemma 3, we obtain esD1+s (PX n Y n Un kπX n Y n ×PUn ) ≤ esD1+s (PW n X n Y n kPW n πX n Y n )−nsR + esD1+s (PX n Y n kπX n Y n )   = esD1+s (PX n Y n kπX n Y n ) esD1+s (PW n X n Y n kPW n πX n Y n )−nsR−sD1+s (PX n Y n kπX n Y n ) + 1 . (91) (92) Taking log’s and normalizing by s, D1+s (PX n Y n Un kπX n Y n × PUn )   1 = D1+s (PX n Y n kπX n Y n ) + log esD1+s (PW n X n Y n kPW n πX n Y n )−nsR−sD1+s (PX n Y n kπX n Y n ) + 1 (93) s 1 ≤ D1+s (PX n Y n kπX n Y n ) + esD1+s (PW n X n Y n kPW n πX n Y n )−nsR−sD1+s (PX n Y n kπX n Y n ) . (94) s Note that in (68), δn tends to zero exponentially fast as n → ∞, since QnW (Tǫn′ ), We first consider the first term of (94).   n n n n n n n n n n minwn ∈Tǫ′ QX|W Tǫ (QW X |w )|w , and minwn ∈Tǫ′ QY |W Tǫ (QW Y |w )|w all tend to one exponentially fast as n → ∞. Combining this with (73), we obtain that D1+s (PX n Y n kπX n Y n ) → 0 exponentially fast. Furthermore, by (84) we can write the exponent of the second term of (94) as 1 1 lim inf sR − sD1+s (PW n X n Y n kPW n πX n Y n ) + sD1+s (PX n Y n kπX n Y n ) n→∞ n  n (1 − ǫ)2 4ǫ = sR − s IQ (XY ; W ) + HQ (XY ) . (95) 1 + ǫ′ 1 − ǫ′ Since HQ (XY ) = Hπ (XY ) ≤ log {|X ||Y|} is bounded and R > IQ (XY ; W ), by choosing sufficiently small ǫ > ǫ′ > 0, we can ensure this exponent is positive. Combining the two points above, we conclude that the optimal D1+s (PX n Y n Un kπX n Y n × PUn ) tends to zero exponentially fast as long as R > CWyner (X; Y ). On the other hand, by a similar argument in Appendix A-A, D1+s (PX n Y n |Un kπX n Y n |PUn ) → 0 exponentially fast implies that there must exist at least one sequence of codebooks indexed by {un }∞ n=1 such that D1+s (PX n Y n |Un =un kπX n Y n ) → 0 exponentially fast. Hence the proof is completed. 12 A PPENDIX B P ROOF OF T HEOREM 3 A. Proof of Theorem 3 In this section, we present the proof of Theorem 3. In the proof, we adapt the information spectrum method proposed by Oohama [14] to first establish a non-asymptotic lower bound on |PX n Y n − πX n Y n |. Invoking the lower bound (cf. Lemma 6) and applying Cramér’s bound in the theory of large deviations [31], we can obtain a further lower bound on |PX n Y n − πX n Y n | leading to (27). Let PMn X n Y n be the joint distribution of (Mn , X n , Y n ), induced by the synthesis code, i.e., PMn X n Y n (m, xn , y n ) = 1 PX n |Mn (xn |m)PY n |Mn (y n |m). |Mn | (96) In the following, for brevity sometimes we omit the subscript, and write PMn X n Y n as P . Let QX n Y n and QX n Y n |Mn be arbitrary distributions. Given any η > 0, define the following information-spectrum sets and support sets: o n πX n Y n (xn , y n ) 1 ≥ −η × Mn , (97) A1 := (xn , y n ) : log n QX n Y n (xn , y n ) o n PX n |Mn (xn |w)PY n |Mn (y n |m) 1 ≥ −η , (98) A2 := (xn , y n , m) : log n QX n Y n |Mn (xn , y n |m) n o QX n Y n |Mn (xn , y n |m) 1 A3 := (xn , y n , m) : log ≤ R + η , (99) n πX n Y n (xn , y n ) (100) Ae1 := supp(πX n Y n ) × Mn , Ae2 := supp(PX n Y n Mn ), Ae := Ae1 ∩ Ae2 . (101) (102) n n ). For i = 1,Q . . . , n, let QXi Yi Ui Vi be any Qndistribution and let QX Y = QnChoose Ui = Mn andQVni = (X , Y n n n i−1 i−1 i−1 i−1 , where Q = Q = Q and Q Q = Q Xi Yi |Vi X Y |Mn Y Y i=1 Xi Yi |Mn X i=1 Xi Yi |Ui Vi i=1 Xi Yi |X i=1 Xi Yi |Vi and QXi Yi |Ui Vi are conditional distributions induced by QXi Yi Ui Vi . Paralleling (97) to (99), given any η > 0, we define the following memoryless version of information-spectrum sets: ( ) n QXi Yi |Vi (xi , yi |vi ) 1X n n n B1 := (x , y , v ) : log ≤ η × Mnn , (103) n i=1 πXY (xi , yi ) ( ) n QXi Yi |Ui Vi (xi , yi |ui , vi ) 1X n n n n B2 := (x , y , u , v ) : log ≤η , (104) n i=1 PXi |Ui Vi (xi |ui , vi )PYi |Ui Vi (yi |ui , vi ) ( ) n QXi Yi |Ui Vi (xi , yi |ui , vi ) 1X n n n n B3 := (x , y , u , v ) : log ≤R+η . (105) n i=1 πXY (xi , yi ) i−1 i−1 We first present a non-asymptotic lower bound on |PX n Y n − πX n Y n |, i.e., a non-asymptotic converse bound for the problem. Lemma 5. For any synthesis code such that 1 log |Mn | ≤ R, n we have |PX n Y n − πX n Y n | ≥ 1 − P 3 \ i=1  Ai Ae − 3e−nη , (106) (107) n n n n e =P n n where P (·|A) e denotes the conditional distribution of (X , Y , Mn ) ∼ PMn X n Y n given that (X , Y , Mn ) ∈ X Y Mn |A e with PMn X n Y n denoting the distribution induced by the synthesis code. A, The proof of Lemma 5 is given in Appendix B-B. Invoking Lemma 5 and choosing the distributions QX n Y n and QX n Y n |Mn as in the paragraph above (103), we obtain the following lemma. Lemma 6. Given the conditions in Lemma 5, we have |PX n Y n − πX n Y n | ≥ 1 − P 3 \ i=1  Bi Ae − 3e−nη . (108) 13 The proof of Lemma 6 is given in Appendix B-C. In the following, for simplicity, we will use Qi to denote QXi Yi Ui Vi and use Pi to denote PXi Yi Ui Vi . Let α ∈ [0, 1]. Then (α) we need the following definitions to further lower bound (108). Similar to the definition of ωQXY U (x, y|u) in (21), we define ! QXi Yi |Vi (xi , yi |vi ) QXi Yi |Ui Vi (xi , yi |ui , vi ) (α) ωQi ,Pi (xi , yi |ui , vi ) := ᾱ log + log πXY (xi , yi ) PXi |Ui Vi (xi |ui , vi )PYi |Ui Vi (yi |ui , vi ) + α log QXi Yi |Ui Vi (xi , yi |ui , vi ) . πXY (xi , yi ) (109) Then, similar to the definition of Ω(α,θ) (QXY U ) in (22), we define Ω(α,λ) ({Qi }ni=1 ) := − log X xn ,y n ,m n i h  X (α) ωQi ,Pi (xi , yi |ui , vi ) . PX n Y n Mn |Ae(xn , y n , m) exp − λ (110) i=1 e where ui = m, vi = (xi−1 , y i−1 ), and PX n Y n Mn |Ae is the conditional distribution of (X n , Y n , Mn ) given (X n , Y n , Mn ) ∈ A. Applying Cramér’s bound [31, Section 2.2] and utilizing Lemma 6, we obtain the following lemma. The proof of this lemma is similar to that of [14, Proposition 1], and hence we omit it for the sake of brevity. Lemma 7. For any (α, λ) ∈ [0, 1] × [0, ∞), given the condition in Lemma 5, we have |PX n Y n − πX n Y n | ≥ 1 − 4 exp −n 1 (α,λ) ({Qi }ni=1 ) nΩ − λαR 1 + (1 + ᾱ)λ ! . (111) Let Ω(α,λ) := inf 1 (α,λ) Ω ({Qi }ni=1 ). n inf n≥1 {Qi }n i=1 (112) Define θ := λ . 1 − 2ᾱλ (113) λ= θ . 1 + 2ᾱθ (114) Hence, we have The next lemma is essential in the proof. Lemma 8. For α ∈ [0, 1] and λ ∈ [0, 21ᾱ ), we have Ω(α,λ) ≥ Ω(α,θ) . 1 + 2ᾱθ (115) The proof of Lemma 8 is similar to that of [14, Proposition 2] and given in Appendix B-D. In the proof of Lemma 8, we adopt ideas from [14] and choose appropriate distributions QXi Yi Ui Vi via the recursive method. Combining Lemmas 7 and 8 yields |PX n Y n − πX n Y n | Ω(α,λ) − λαR ≥ 1 − 4 exp −n 1 + (1 + ᾱ)λ ≥ 1 − 4 exp −n Ω(α,θ) 1+2ᾱθ 1+ ! θαR 1+2ᾱθ (1+ᾱ)θ 1+2ᾱθ − Ω(α,θ) − θαR = 1 − 4 exp − n 1 + (5 − 3α)θ  ≥ 1 − 4 exp − nF (R) , ! ! (116) (117) (118) (119) where (119) follows from the definition of F (R) in (25) and the fact that (118) holds for any (α, θ) ∈ [0, 1] × (0, +∞). The proof of Theorem 3 is now complete. 14 B. Proof of Lemma 5 Define πX n Y n Mn := πX n Y n PMn |X n Y n . Then |PX n Y n − πX n Y n | = |PX n Y n Mn − πX n Y n Mn | ≥ π(Ae1 ∩ A1 ∩ A3 ) − P (Ae1 ∩ A1 ∩ A3 ) =1− π(Aec1 ∪ Ac1 ∪ 3 \  = 1 − P Ae ∩  ≥ 1 − P Ae ∩  = 1 − P Ae ∩ i=1 3 \ i=1 3 \ (120) (121) − P (Ae1 ∩ A1 ∩ A3 )  Ai − P (Ae1 ∩ A1 ∩ A3 ∩ (Ac2 ∪ Aec2 )) − π(Aec1 ∪ Ac1 ∪ Ac3 ) Ac3 ) Ai Ai i=1   − P (Ac2 ) − P (Aec2 ) − π(Aec1 ) − π(Ac1 ) − π(Ac3 ) − P (Ac2 ) − π(Ac1 ) − π(Ac3 ) (122) (123) (124) (125) The last three terms above can each be bounded above by e−nη because X P (Ac2 ) = P (xn , y n , m) (126) P (w)Q(xn , y n |m)e−nη (127) (xn ,y n ,m)∈Ac2 ≤ n (x ,y X n ,m)∈Ac2 ≤ e−nη , (128) and π(Ac3 ) = n (x ,y ≤ n (x ,y ≤ n X n X n (x ,y −nη ≤e (129) Q(xn , y n |m)e−n(R+η) P (m|xn , y n ) (130) Q(xn , y n |m)e−n(R+η) (131) ,m)∈Ac3 X n π(xn , y n )P (m|xn , y n ) ,m)∈Ac3 ,m)∈Ac3 (132) , and π(Ac1 ) = n X (x ,y ≤ n π(xn , y n ) (133) Q(xn , y n )e−nη (134) )∈Ac1 X (xn ,y n )∈Ac1 −nη ≤e (135) . Therefore, we have 3   \ Ai − 3e−nη |PX n Y n − πX n Y n | ≥ 1 − P Ae ∩ (136) i=1 ≥1−P 3 \ i=1  Ai Ae − 3e−nη . (137) 15 C. Proof of Lemma 6 Recall that in Appendix B-A, we choose Ui = Mn and Vi = (X i−1 , Y i−1 ). Then QX n Y n and QX n Y n |Mn can be written as follows: n Y QXi Yi |X i−1 Y i−1 (xi , yi |xi−1 , y i−1 ) (138) QX n Y n (xn , y n ) = = QX n Y n |Mn (xn , y n |m) = = i=1 n Y i=1 n Y i=1 n Y i=1 QXi Yi |Vi (xi , yi |vi ), (139) QXi Yi |Mn X i−1 Y i−1 (xi , yi |m, xi−1 , y i−1 ) (140) QXi Yi |Ui Vi (xi , yi |ui , vi ). (141) Now recall from Appendix B-A that the joint distribution of (X n , Y n , Mn ) induced by the code is PX n Y n Mn . The marginal distributions of PX n Y n Mn are as follows: πX n Y n (xn , y n ) = PX n |Mn (xn |m) = = = PY n |Mn (y n |m) = = = n Y i=1 n Y i=1 n Y i=1 n Y i=1 n Y i=1 n Y i=1 n Y i=1 πXY (xi , yi ), (142) PXi |Mn X i−1 (xi |m, xi−1 ) (143) PXi |Mn X i−1 Y i−1 (xi |m, xi−1 , y i−1 ) (144) PXi |Ui Vi (xi |ui , vi ), (145) PYi |Mn Y i−1 (yi |m, y i−1 ) (146) PYi |Mn X i−1 Y i−1 (yi |m, xi−1 , y i−1 ) (147) PYi |Ui Vi (yi |ui , vi ), (148) where (144) and (147) follow from the Markov chains Xi − Mn X i−1 − Y i−1 and Yi − Mn Y i−1 − X i−1 under distribution PX n Y n Mn (these two Markov chains can be easily obtained by observing that PX i Y i Mn = PMn PXi |Mn X i−1 PYi |Mn Y i−1 ). Using Lemma 5 and (138)–(148), we obtain |PX n Y n − πX n Y n | ≥ 1 − P 3 \ i=1  Bi Ae − 3e−nη . (149) D. Proof of Lemma 8 1) Removing Dependence on the Indices: Recall from Appendix B-A that the joint distribution of (X n , Y n , Mn ) is PX n Y n Mn and PXi Yi Ui Vi is induced by PX n Y n Mn . Further, recall that Qi denotes QXi Yi Ui Vi and Pi denotes PXi Yi Ui Vi . Define   (α) (α,λ) (150) gQi ,Pi (xi , yi |ui , vi ) := exp − λωQi ,Pi (xi , yi |ui , vi ) , (α) where ωQi ,Pi (xi , yi |ui , vi ) is defined in (109). Recall the definition of Ω(α,λ) ({Qi }ni=1 ) in (110), then we obtain that   exp − Ω(α,λ) ({Qi }ni=1 ) = X xn ,y n ,m PX n Y n Mn |Ae(xn , y n , m) n Y (α,λ) gQi ,Pi (xi , yi |ui , vi ), (151) i=1 e where ui = m, vi = (xi−1 , y i−1 ), and PX n Y n Mn |Ae is the conditional distribution of (X n , Y n , Mn ) given (X n , Y n , Mn ) ∈ A. 16 For i = 1, . . . , n, define C̃i := X xn ,y n ,m (α,λ)|i PX n Y n M |Ae(xn , y n , m) n (α,λ) Λi (α,λ)|i Obviously, PX n Y n M := ({Qj }ij=1 ) := n n e(x , y , m) n |A PX n Y n Mn |Ae(xn , y n , m) PX n Y n Mn |Ae(xn , y n , m) Qi j=1 i Y (α,λ) (152) gQj ,Pj (xj , yj |uj , vj ), j=1 (α,λ) gQj ,Pj (xj , yj |uj , vj ) C̃i (153) , C̃i . C̃i−1 (154) is a distribution induced by normalizing all the terms of the summation in the definition of C̃i . (α,λ) ({Qj }ij=1 ), defined in Similarly to [14, Lemma 7], we obtain the following lemma, which will be used to simplify Λi (154), in Appendix B-D2. Lemma 9. For i = 1, . . . , n, we have (α,λ) Λi ({Qj }ij=1 ) = X xn ,y n ,m (α,λ)|i−1 PX n Y n M (α,λ) n n e(x , y , m)gQi ,Pi (xi , yi |ui , vi ). (155) n |A Furthermore, combining (151), (152) and (154) gives us n   Y (α,λ) ({Qj }ij=1 ). Λi exp − Ω(α,λ) ({Qi }ni=1 ) = (156) i=1 2) Completion of the Proof of Lemma 8: Assume U and V are two countable sets. Paralleling (20) to (23), for (α, θ) ∈ (0, 1] × (0, ∞), we define the following quantities: n o e := QXY UV ∈ P(X × Y × U × V) : supp(QXY ) ⊆ supp(πXY ) , Q (157) ! QXY |UV (x, y|u, v) QXY |UV (x, y|u, v) QXY |V (x, y|v) (α) + α log + log , (158) ω eQXY U V (x, y|u, v) := ᾱ log πXY (x, y) QX|UV (x|u, v)QY |UV (y|u, v) πXY (x, y) i h  e (α,λ) (QXY UV ) := − log EQXY U V exp − θω (α) (159) Ω QXY U V (X, Y |U, V ) , e (α,λ) := Ω inf e QXY U V ∈Q e (α,λ) (QXY UV ), Ω (160) where EQXY U V in (159) is only taken over the set supp(QXY UV ). Recall that ui = m and vi = (xi−1 , y i−1 ). For each i = 1, . . . , n, define X (α,λ)|i−1 PX n Y n M |Ae(xn , y n , m), P (α,λ) (xi , yi , ui , vi ) := n xn i+1 ,yi+1 (161) n (α,λ)|i−1 where PX n Y n Mn (xn , y n , m) was defined in (153). Combining Lemma 9 and (161) yields (α,λ) Λi ({Qj }ij=1 ) = X (α,λ) P (α,λ) (xi , yi , ui , vi )gQi ,Pi (xi , yi |ui , vi ). (162) xi ,yi ,ui ,vi Note that Qi = QXi Yi Ui Vi can be chosen arbitrarily for all i = 1, . . . , n. Here we apply the recursive method. For each i = 1, . . . , n, we choose QXi Yi Ui Vi such that QXi Yi Ui Vi (xi , yi , ui , vi ) = P (α,λ) (xi , yi , ui , vi ). (163) Then, let QXi Yi |Vi , QXi Yi |Ui Vi be induced by QXi Yi Ui Vi . Define (α,λ) hQi (xi , yi |ui , vi ) := (α,λ) gQi ,Pi (xi , yi |ui , vi ) λᾱ (xi |ui , vi )PYλiᾱ|Ui Vi (yi |ui , vi ) PX i |Ui Vi λᾱ (yi |ui , vi ) (xi |ui , vi )QYλᾱ QX i |Ui Vi i |Ui Vi !−1 , (164) 17 (α,λ) where gQi ,Pi was defined in (150). In the following, for brevity, we drop the subscripts of the distributions. From (162), we obtain (α,λ) Λi (α,λ) ({Qj }ij=1 ) = EQi [gQi ,Pi (Xi , Yi |Ui , Vi )] " λᾱ λᾱ PX |U V (Xi |Ui , Vi )PYi |Ui Vi (Yi |Ui , Vi ) (α,λ) hQi (Xi , Yi |Ui , Vi ) λᾱi i i (Yi |Ui , Vi ) QXi |Ui Vi (Xi |Ui , Vi )QYλᾱ i |Ui Vi = EQi  ≤ EQi hn 1 i1−2λᾱ o 1−2λ ᾱ (α,λ) hQi (Xi , Yi |Ui , Vi ) #!λᾱ PYi |Ui Vi (Yi |Ui , Vi ) × EQi QYi |Ui Vi (Yi |Ui , Vi )    (α, λ ) e 1−2λᾱ (Qi ) ≤ exp − 1 − 2λᾱ Ω ! e (α,θ) (Qi ) Ω = exp − 1 + 2ᾱθ ! e (α,θ) Ω ≤ exp − 1 + 2ᾱθ   Ω(α,θ) , = exp − 1 + 2ᾱθ " EQi " # PXi |Ui Vi (Xi |Ui , Vi ) QXi |Ui Vi (Xi |Ui , Vi ) (165) (166) #!λᾱ (167) (168) (169) (170) (171) where (166) follows from (164); (167) follows from Hölder’s inequality; (168) follows from the definitions of Ω(α,θ) (·) and (α,λ) e (α,θ) hQi (·) in (22) and (164) respectively; (169) follows from (113) and (114); (170) follows since Ω(α,θ) (QXY UV ) ≥ Ω for any QXY UV such that supp(QXY ) ⊆ supp(πXY ) (The fact that Qi satisfies this point will be shown in the following paragraph); and (171) follows since by the support lemma [2], the cardinality bounds |V| ≤ 1, |U| ≤ |X ||Y| are sufficient to e (α,θ) . exhaust Ω Now we show that according to the choice of QXi Yi Ui Vi , we have supp(QXi Yi ) ⊆ supp(πXY ), which was used in (170). e and hence the marginal distributions Note that PX n Y n Mn (xn , y n , m) > 0 and πX n Y n (xn , y n ) > 0 for any (xn , y n , m) ∈ A, n n e PXi Yi Ui Vi , PXi |Ui Vi and PYi |Ui Vi when evaluated at any (x , y , m) ∈ A is positive as well. According to the choice of QXi Yi Ui Vi in (163), we have that QXi Yi Ui Vi is also positive when evaluated at (xn , y n , m) ∈ Ae (this point can be shown via mathematical induction), i.e., o n (172) supp(QXi Yi Ui Vi ) ⊇ (x, y, u, v) : ∃(xn , y n , m) ∈ Ae : xi = x, yi = y, m = u, (xi−1 , y i−1 ) = v . On the other hand, also according to the choice of QXi Yi Ui Vi , we have n o supp(QXi Yi Ui Vi ) ⊆ (x, y, u, v) : ∃(xn , y n , m) ∈ Ae : xi = x, yi = y, m = u, (xi−1 , y i−1 ) = v . Therefore, n o supp(QXi Yi Ui Vi ) = (x, y, u, v) : ∃(xn , y n , m) ∈ Ae : xi = x, yi = y, m = u, (xi−1 , y i−1 ) = v . Further, we have n o supp(QXi Yi ) = (x, y) : ∃(xn , y n , m) ∈ Ae : xi = x, yi = y (173) (174) (175) ⊆ {(x, y) : ∃(xn , y n , m) ∈ supp(πX n Y n ) × Mn : xi = x, yi = y} (176) = supp(πXY ). (177) Combining (156) and (171), we obtain that n 1 (α,λ) 1X (α,λ) ({Qj }ij=1 ) log Λi Ω ({Qi }ni=1 ) = − n n i=1 ≥ Ω(α,θ) . 1 + 2ᾱθ (178) (179) Finally, combining (112) and (179), we have that Ω(α,λ) ≥ The proof of Lemma 8 is now complete. Ω(α,θ) . 1 + 2ᾱθ (180) 18 A PPENDIX C P ROOF OF L EMMA 1 Let U be a random variable taking values in a finite alphabet U. Define a set of joint distributions on X × Y × U as n o P ∗ := PXY U : |U| ≤ |X ||Y|, PXY = πXY , X − U − Y . (181) and let R∗ := min PXY U ∈P ∗ I(XY ; U ). (182) A. Preliminary Lemmata for the Proof of Lemma 1 By the support lemma [2, Appendix C], we have the following lemma [1]. Lemma 10. Wyner’s common information CWyner (X; Y ) satisfies CWyner (X; Y ) = R∗ . (183) Before proceeding the proof of Lemma 1, we present an alternative expression for Wyner’s common information. Recall that given a number a ∈ [0, 1], we define ā = 1 − a. Then for any α ∈ [0, 1] and QXY U ∈ Q, define  R(α) (QXY U ) := ᾱ D(QXY kπXY ) + D(QXY |U kQX|U QY |U |QU ) R (α) := min + αD(QXY |U kπXY |QU ), QXY U ∈Q R (α) (QXY U ) 1 (α) R . α α∈(0,1] Rsh := sup (184) (185) (186) By observing that both P ∗ and Q are compact, and by utilizing the fact that a continuous function defined on a compact set attains its minimum, we obtain the following. Fact 1. Both the minima in the definitions of R∗ in (182) and R(α) in (185) are attained. We then have the following lemma. Lemma 11. The following conclusions hold. (i) For any α ∈ (0, 1], we have 1 (α) R ≤ R∗ . α (187) ∞ Moreover, there exists some decreasing sequence {αk }k=1 ⊂ R such that limk→∞ αk = 0 and 1 (αk ) R ≥ R∗ − c(αk ), αk (188) where {c(αk )}∞ k=1 ⊂ R is another sequence such that limk→∞ c(αk ) = 0. (ii) We have Rsh = R∗ = CWyner (X; Y ). (189) Lemma 11 is similar to [14, Property 3], but the proofs are different. Essentially, in both the proof of [14, Property 3] and e XY U is used to establish the inequality our proof, an intermediate distribution Q R∗ − c(αk ) ≤ 1 (αk ) e 1 (αk ) R (QXY U ) ≤ R . αk αk (190) However, the construction of such an intermediate distribution is different for these two proofs. The construction in [14] does e XY U to satisfy the Markov chain X − U − Y , but also requires not apply to our case, since our case does not only require Q e that QXY = πXY . Proof of Lemma 11: It is easy to show (187). Hence, by the definition of Rsh in (186), Rsh ≤ R∗ . ∞ {αm }m=1 (191) In the following we prove (188). Let be an arbitrary sequence of decreasing positive real numbers such that (m) limm→∞ αm = 0, and let QXY U be a minimizing distribution of (185) with α = αm . The existence of this minimizing 19 distribution is guaranteed by Fact 1. Since P(X × Y × U) is compact (following from the definition of Q), there must exist (mk ) e some sequence of increasing integers {mk }∞ k=1 such that QXY U converges to some distribution QXY U . Consider, 1 (α) R α α∈(0,1] 1 ≥ lim sup R(αmk ) α mk k→∞ (  ᾱmk  (m ) (m ) (m ) (m ) (m ) D(QXYk kπXY ) + D(QXYk|U kQX|Uk QY |Uk |QU k ) = lim sup αmk k→∞ ) Rsh = sup (mk ) (m ) + D(QXYk|U ||πXY |QU ≥ lim sup k→∞ +  ᾱmk αmk  ) o n (m ) (m ) (m ) (m ) (m ) lim inf D(QXYk kπXY ) + D(QXYk|U kQX|Uk QY |Uk |QU k ) (192) (193) (194) k→∞ (m ) lim inf D(QXYk|U ||πXY k→∞ (mk ) |QU )  e XY kπXY ) + D(Q e XY |U kQ e X|U Q e Y |U |Q e U ) + D(Q e XY |U ||πXY |Q e U ). = ∞ D(Q (195) (196) Observe that Rsh is finite due to (191). Hence it holds that That is, Therefore, under (199) and (200), we have e XY kπXY ) = 0, D(Q e XY |U kQ e X|U Q e Y |U |Q e U ) = 0. D(Q (197) (198) e XY = πXY , Q e XY |U = Q e X|U Q e Y |U . Q (199) e XY |U kπXY |Q eU ) (196) ≥ D(Q e XY |U , Q e XY ) = I(Q (201) (200) (202) (203) ∗ ≥R . Combining (191), (196) and (203) yields us Rsh = R∗ = lim k→∞ 1 R(αmk ) . αmk (204) ∗ Therefore, there exists some sequence {c(αmk )}∞ k=1 ⊂ R (e.g., the sequence {R − limk→∞ c(αmk ) = 0 and 1 R∗ − c(αmk ) ≤ R(αmk ) ≤ R∗ . αmk 1 αmk R(αmk ) }∞ k=1 ⊂ R) such that (205) This concludes the proof. We also have the following crucial lemma. Lemma 12. Let α ∈ (0, 1] and QXY U ∈ Q. Then we have or equivalently, 1 lim Ω(α,θ) = R(α) , θ↓0 θ (206) 1 (α,θ) Ω = R(α) + ǫ(α,θ) , θ (207) where Ω(α,θ) and R(α) were defined in (23) and (185) respectively, and ǫ(α,θ) is a term that vanishes as θ ↓ 0, the rate being dependent on α. Proof of Lemma 12: To show this lemma, we first need to show that ( 1 (α,θ) Ω (QXY U ), (α,θ) b R (QXY U ) := θ (α) R (QXY U ), θ>0 θ=0 (208) 20 is continuous in (θ, QXY U ) ∈ [0, 1+1 ᾱ ) × Q. It is easy to observe that h i (α) Ω(α,θ) (QXY U ) = − log EQXY U exp − θωQXY U (X, Y |U ) X 1−θ(1+ᾱ) θᾱ θ = − log (x, y, u) QU|XY (u|x, y)QX|U (x|u)QY |U (y|u) (QU (u)πXY (x, y)) QXY U (209) (210) x,y,u b (α,θ) (QXY U ) is jointly continuous on (0, 1 ) × Q. Therefore, is jointly continuous in (θ, QXY U ) ∈ [0, 1+1 ᾱ ) × Q, hence R 1+ᾱ (α,θ) b to show the continuity of R (QXY U ) in (θ, QXY U ) ∈ [0, 1+1 ᾱ ) × Q, it suffices to show it is continuous at any point in {0} × Q, i.e., 1 (α,θ) Ω (QXY U ) = R(α) (Q′XY U ) (211) lim (θ,QXY U )→(0,Q′XY U ) θ for any Q′XY U ∈ Q. Let   (α) QXY U (x, y, u) exp − θωQXY U (x, y|u) (α,θ) .  QXY U (x, y, u) := P (α) (x, y|u) Q (x, y, u) exp − θω XY U x,y,u QXY U (212) Invoking the definition of Ω(α,θ) (QXY U ) in (22), we obtain i h ∂Ω(α,θ) (QXY U ) (α) = EQ(α,θ) ωQXY U (X, Y |U ) , XY U ∂θ and Hence for fixed QXY U h i ∂ 2 Ω(α,θ) (QXY U ) (α) = −VarQ(α,θ) ωQXY U (X, Y |U ) . 2 XY U ∂θ ∈ Q, we have ∂Ω(α,θ) (QXY U ) ∂θ which implies that θ (213) (214) = R(α) (QXY U ) > 0, (215) ∂ 2 Ω(α,θ) (QXY U ) ≤ 0, ∂θ2 (216) θ=0 ∂Ω(α,θ) (QXY U ) ≤ Ω(α,θ) (QXY U ) ≤ θR(α) (QXY U ). ∂θ (217) Furthermore, observe that R(α) (QXY U ) is continuous in QXY U ∈ Q, hence lim QXY U →Q′XY U On the other hand, observe that ∂Ω(α,θ) (QXY U ) ∂θ lim (θ,QXY U )→(0,Q′XY U ) R(α) (QXY U ) = R(α) (Q′XY U ). (218) given in (213) is continuous in (θ, QXY U ) ∈ [0, 1+1 ᾱ ) × Q. Hence X ∂Ω(α,θ) (QXY U ) (α) = (X, Y |U ) Q′XY U (x, y, u)ωQ′ XY U ∂θ x,y,u = R(α) (Q′XY U ). (219) (220) Therefore, combining (217), (218), and (220), we observe that the limit lim(θ,QXY U )→(0,Q′XY U ) θ1 Ω(α,θ) (QXY U ) exists, and b(α,θ) (QXY U ) moreover, lim(θ,QXY U )→(0,Q′XY U ) θ1 Ω(α,θ) (QXY U ) = R(α) (Q′XY U ). Hence, we obtain (211). In other words, R 1 is jointly continuous in (θ, QXY U ) ∈ [0, 1+ᾱ ) × Q. In addition, observe that Q is a compact set. By using the following lemma b (α,θ) (QXY U ) is continuous in θ ∈ [0, 1 ). we can assert that minQXY U ∈Q R 1+ᾱ Lemma 13 (Lemma 14 in [25]). Let X and Y be two metric spaces and let K ⊂ X be a compact set. Let f : X × Y → R be a (jointly) continuous real-valued function. Then the function g : Y → R, defined as g(y) := min f (x, y), x∈K (221) ∀ y ∈ Y, is continuous on Y. Considering the point θ = 0, we obtain lim min θ→0 QXY U ∈Q b (α,θ) (QXY U ) = R min QXY U ∈Q b (α,0) (QXY U ) = R min QXY U ∈Q R(α) (QXY U ) = R(α) , (222) 21 where the first equality follows from Lemma 13 which essentially says that the limit and minimum operations can be swapped. On the other hand, observe that lim min θ→0 QXY U ∈Q b (α,θ) (QXY U ) = lim R min θ→0 QXY U ∈Q 1 1 (α,θ) Ω (QXY U ) = lim Ω(α,θ) . θ→0 θ θ (223) Combining (222) and (223), we obtain (206) as desired. B. Proof of Part (i) in Lemma 1 Using Lemma 10, we obtain that if R < CWyner (X; Y ), then R + τ ≤ R∗ (224) for some τ > 0. Further, invoking (188) and (224), we obtain that there exists k0 such that for any k ≥ k0 , R+τ ≤ 1 (αk ) R + c(αk ), αk (225) τ . 2 (226) and c(αk ) ≤ Referring to (225) and (226), we obtain that for any k ≥ k0 , R+ τ 1 (αk ) R . ≤ 2 αk (227) Therefore, invoking (25), we conclude that for any k ≥ k0 , F (R) ≥ sup F (αk ,θ) (R) (228) θ≥0 ≥ F (αk ,θ) (R) sup (229) θ∈[0, 1+1ᾱ ) k Ω(αk ,θ) − θαk R θ∈[0, 1+1ᾱ ) 1 + (5 − 3αk )θ k o 1 n (αk ) ≥ sup θR + θǫ(αk ,θ) − θαk R θ∈[0, 1+1ᾱ ) 1 + 5θ k θ n (αk ,θ) αk τ o ≥ sup ǫ + 2 θ∈[0, 1 ) 1 + 5θ = sup (230) (231) (232) 1+ᾱk ≥ sup e θ∈[0,θ] ≥ αk τ θ 4(1 + 5θ) (233) αk τ θe , e 4(1 + 5θ) (234) where (231) follows from Lemma 12 and the inequality 1 + (5 − 3αk )θ ≤ 1 + 5θ, (232) follows from (227), and (233) follows e Since the expression in (234) is since there exists a sufficiently small θe ∈ (0, 1+1ᾱk ) such that |ǫ(αk ,θ) | ≤ 14 αk τ for all θ ≤ θ. positive, we have F (R) > 0 as desired. C. Proof of Part (ii) in Lemma 1 Because exp(·) is convex, applying Jensen’s inequality, we obtain i h (α) Ω(α,θ) (QXY U ) ≤ θEQXY U ωQXY U (X, Y |U ) = θR(α) (QXY U ). (235) Hence we have Ω(α,θ) ≤ min QXY U ∈Q = θR(α) . θR(α) (QXY U ) (236) (237) 22 Thus, recalling the definition of F (α,θ) (R) in (24), we obtain that Ω(α,θ) − θαR 1 + (5 − 3α)θ θα( α1 R(α) − R) ≤ 1 + (5 − 3α)θ θα(Rsh − R) ≤ 1 + (5 − 3α)θ ≤ 0, F (α,θ) (R) = (238) (239) (240) (241) where (241) follows from the assumption R ≥ CWyner (X; Y ) = Rsh . On the other hand, note that lim F (α,θ) = 0. θ→0 (242) Hence, combining (241) and (242), we conclude that F = sup F (α,θ) (R) = 0. (243) (α,θ)∈[0,1]×[0,∞) ACKNOWLEDGEMENTS The authors are supported by a Singapore National Research Foundation (NRF) National Cybersecurity R&D Grant (R-263000-C74-281 and NRF2015NCR-NCR003-006). R EFERENCES [1] A. Wyner. The common information of two dependent random variables. IEEE Trans. on Inform. Theory, 21(2):163–179, 1975. [2] A. El Gamal and Y.-H. Kim. Network Information Theory. Cambridge university press, 2011. [3] G. R. Kumar, C. T. Li, and A. El Gamal. Exact common information. In IEEE International Symposium on Information Theory (ISIT), pages 161–165. IEEE, 2014. [4] P. Cuff, H. Permuter, and T. Cover. Coordination capacity. IEEE Trans. on Inform. Theory, 56(9):4181–4206, 2010. [5] P. Cuff. Distributed channel synthesis. IEEE Trans. on Inform. Theory, 59(11):7071–7096, 2013. [6] T. Han and S. Verdú. Approximation theory of output statistics. IEEE Trans. on Inform. Theory, 39(3):752–772, 1993. [7] M. Hayashi. General nonasymptotic and asymptotic formulas in channel resolvability and identification capacity and their application to the wiretap channel. IEEE Trans. on Inform. Theory, 52(4):1562–1575, 2006. [8] M. Hayashi. Exponential decreasing rate of leaked information in universal random privacy amplification. IEEE Trans. on Inform. Theory, 57(6):3989– 4001, 2011. [9] J. Liu, P. Cuff, and S. Verdú. Eγ -resolvability. IEEE Trans. on Inform. Theory, 63(5):2629–2658, 2017. [10] L. Yu and V. Y. F. Tan. Rényi resolvability and its applications to the wiretap channel. arXiv preprint 1707.00810, 2017. [11] M. R. Bloch and J. N. Laneman. Strong secrecy from channel resolvability. IEEE Trans. on Inform. Theory, 59(12):8077–8098, 2013. [12] T. S. Han, H. Endo, and M. Sasaki. Reliability and secrecy functions of the wiretap channel under cost constraint. IEEE Trans. on Inform. Theory, 60(11):6819–6843, 2014. [13] M. B. Parizi, E. Telatar, and N. Merhav. Exact random coding secrecy exponents for the wiretap channel. IEEE Trans. on Inform. Theory, 63(1):509–531, 2017. [14] Y. Oohama. Exponent function for source coding with side information at the decoder at rates below the rate distortion function. arXiv preprint arXiv:1601.05650, 2016. [15] I. Sason. On the Rényi divergence, joint range of relative entropies, and a channel coding theorem. IEEE Trans. on Inform. Theory, 62(1):23–34, 2016. [16] A. R. Barron. Entropy and the central limit theorem. The Annals of Probability, pages 336–342, 1986. [17] S. G. Bobkov, G. P. Chistyakov, and F. Götze. Rényi divergence and the central limit theorem. arXiv preprint arXiv:1608.01805, 2016. [18] J. Hou and G. Kramer. Effective secrecy: Reliability, confusion and stealth. In IEEE International Symposium on Information Theory (ISIT), pages 601–605. IEEE, 2014. [19] S. Beigi and A. Gohari. Quantum achievability proof via collision relative entropy. IEEE Trans. on Inform. Theory, 60(12):7980–7986, 2014. [20] Y. Dodis and Y. Yu. Overcoming weak expectations. In Theory of Cryptography, pages 1–22. Springer, 2013. [21] M. Hayashi and V. Y. F. Tan. Equivocations, exponents, and second-order coding rates under various Rényi information measures. IEEE Trans. on Inform. Theory, 63(2):975–1005, 2017. [22] V. Y. F. Tan and M. Hayashi. Analysis of remaining uncertainties and exponents under various conditional Rényi entropies. Accepted to IEEE Trans. on Inform. Theory. Available at arXiv preprint arXiv:1605.09551, 2018. [23] T. Van Erven and P. Harremoës. Rényi divergence and Kullback-Leibler divergence. IEEE Trans. on Inform. Theory, 60(7):3797–3820, 2014. [24] L. Yu and V. Y. F. Tan. Wyner’s common information under Rényi divergence measures. arXiv preprint arXiv:1709.02168, 2017. [25] V. Y. F. Tan, A. Anandkumar, L. Tong, and A. S. Willsky. A large-deviation analysis of the maximum-likelihood learning of Markov tree structures. IEEE Trans. on Inform. Theory, 57(3):1714–1735, 2011. [26] W.-H. Gu and M. Effros. A strong converse for a collection of network source coding problems. In IEEE International Symposium on Information Theory (ISIT), pages 2316–2320. IEEE, 2009. [27] S. Watanabe. Second-order region for Gray–Wyner network. IEEE Trans. on Inform. Theory, 63(2):1006–1018, 2017. [28] L. Zhou, V. Y. F. Tan, and M. Motani. Discrete lossy Gray-Wyner revisited: Second-order asymptotics, large and moderate deviations. IEEE Trans. on Inform. Theory, 63(3):1766–1791, 2017. [29] W. Rudin. Principles of Mathematical Analysis. McGraw-Hill, 1976. [30] I. Csiszar and J. Körner. Information Theory: Coding Theorems for Discrete Memoryless Systems. Cambridge University Press, 2011. [31] A. Dembo and O. Zeitouni. Large Deviations Techniques and Applications. Springer-Verlag, 2nd edition, 1998.
7cs.IT
Using Self-Organizing Maps for Sentiment Analysis Anuj Sharma Indian Institute of Management Indore – 453331, INDIA Email: [email protected] Shubhamoy Dey Indian Institute of Management Indore – 453331, INDIA Email: [email protected] Abstract— Web 2.0 services have enabled people to express their opinions, experience and feelings in the form of user-generated content. Sentiment analysis or opinion mining involves identifying, classifying and aggregating opinions as per their positive or negative polarity. This paper investigates the efficacy of different implementations of Self-Organizing Maps (SOM) for sentiment based visualization and classification of online reviews. Specifically, this paper implements the SOM algorithm for both supervised and unsupervised learning from text documents. The unsupervised SOM algorithm is implemented for sentiment based visualization and classification tasks. For supervised sentiment analysis, a competitive learning algorithm known as Learning Vector Quantization is used. Both algorithms are also compared with their respective multi-pass implementations where a quick rough ordering pass is followed by a fine tuning pass. The experimental results on the online movie review data set show that SOMs are well suited for sentiment based classification and sentiment polarity visualization. Keywords — Sentiment Analysis, Self-Organizing Map, Machine Learning, Text Mining. INTRODUCTION With the ever-growing popularity of Web 2.0 platforms and services, the Internet has emerged as a common medium to write reviews and share opinions for any product consumed or service received by users. Sentiment analysis, also known as opinion mining, attempts to address the problem of classifying opinionated texts (e.g. product reviews) according to their opinion polarity (positive or negative) (Pang et al, 2008). The sentiment based classification of online opinionated texts imposes several challenges which distinguishes it from topic based text classification. The variability and complexity of sentiment expressions, overlapping vocabularies, feature selection and domain dependency of sentiment features are the key issues in sentiment analysis (He et al, 2011; Bai, 2011). This paper investigates the efficacy of Self-Organizing Maps (SOM) for sentiment based visualization and classification of online product reviews. SOM can transform high-dimensional data into two-dimensional representation and can enable automatic clustering of text documents, while preserving higher order topology. Specifically, this paper implements the unsupervised SOM algorithm to support sentiment based visualization and classification tasks. For supervised sentiment analysis, a competitive learning algorithm known as Learning Vector Quantization (LVQ) is used in this work. Both algorithms are also compared with their respective multi-pass implementations. The movie reviews dataset (Pang and Lee 2004) is used to evaluate performance of these different SOM implementations. The rest of this paper is organized as follows: next Section presents related work on sentiment analysis, the methodology based on supervised and unsupervised SOM is described in the Section named methodology, the experimental result Section discusses our results, and the last Section concludes the paper. RELATED WORK There have been a considerable number of studies on sentiment analysis and opinion mining, most of which utilize machine learning, dictionary, statistical, and semantic based approaches. Supervised machine learning classifiers like Naïve Bayes, Support Vector Machine, Maximum Entropy, Decision Tree, Centroid classifier, K-nearest neighbor, Neural Networks, winnow classifier etc., are quite popular for sentiment based classification of online reviews (Sharma et al, 2012a, 2012b; Pang et al, 2002, 2004; Tan et al, 2008). Dictionaries of subjective words with some weighting schemes are also proposed in some studies to quantify polarities of sentiment expressing words in texts for sentiment analysis (Amati et al, 2008). Statistical approached apply statistics like Pointwise Mutual Information (PMI) based on co-occurrence of words to derive sentiment polarity of features (Turney, 2002). The Semantic approach exploits various semantic relationships between sentiment words like synonyms and antonyms, to calculate sentiment polarities for mining and summarizing customer reviews (Hu, 2004). The machine learning approaches that belong to supervised text classification techniques have shown good accuracy. But, the performance of machine learning approaches is heavily dependent on the selected features, the quality and quantity of training data and the domain of the data. The additional learning time required by the machine learning techniques is also a prominent issue since lexicon or semantic approaches take less time. The Naïve Bayes classifier is a probabilistic classifier based on Bayes’ theorem with strong feature independence assumptions. In real life, the sentiment features are seldom independent (Pang, 2002), severely limiting its applicability. The complexity of SVM due to its convoluted training and categorizing algorithms is a major drawback. The over-fitting issues with small training sets also limits the use of SVMs in many real world applications. The k-nearest neighbor (k-NN) classifier is easy to implement but becomes computationally intensive with growing size of the training set (Han et al, 1999). The decision tree and rule induction classifiers are simple to understand and explain but their performance suffer when the number of distinguishing features is large (Quinlan, 1993). To address the above issues, this paper implements both supervised and unsupervised SOM based approaches for sentiment based classification and compares their performances in the same domain. The findings will help gaining insight into the strengths and limitations of different SOM implementations for sentiment analysis. METHODOLOGY Data Pre-processing The data pre-processing involves collecting opinionated text documents, performing initial data cleaning and transforming relevant text data into the input variables form as needed by self-organizing maps. In this study we have performed tokenization, stemming and stop word removal in data pre-processing. We have used Vector Space Model (VSM) to generate the bag of words representation for each document. In VSM, the weights represent the tf·idf score which is computed from term frequency (tf) and inverse document frequency (idf). The tf·idf weight for term j in document Di is computed as tij  tfij  idf j  log(1  fij )  log( n ) df j (1) Where fij is frequency of term j in document I and dfj is number of document having term j. Information Gain Feature Selection Information gain (IG) as a feature goodness criterion has shown consistent performance for sentiment-feature selection (Tan et al, 2008). A feature from large feature space is selected based on its impact on decreasing the overall entropy. Entropy is the expected information needed to classify an instance document. The attributes ranked as per high to low IG score will minimize the overall information necessary to classify instances into predefined classes. The information gain of a feature j over all classes is given by: |C| |C| |C| i 1 i 1 i 1 IG ( j )   P(ci ) log 2 P(ci )  P( j ) P(ci | j ) log 2 P(ci | j )  P( j ) P(ci | j ) log 2 P(ci | j ) (2) Where, P(ci) is the probability that a random instance document belongs to class ci. P(j) is the probability of the occurrence of the feature j in a randomly selected document. P(ci |j) is the probability that a randomly selected document belongs to class ci if document has feature j. P( j ) is the probability of that j does not appear in a random document. P(ci | j ) is the probability that a random document belongs to class c if j does not occur in it. The top ranked features are selected based upon information gain feature selection in this study. Self-Organizing Maps (SOM) The self-organizing map (SOM) is an unsupervised competitive ANN based on the winner-takes-all (WTA) principle (Kohonen, 1995). A typical SOM is made up of a vector of nodes for input (input neurons layer), an array of nodes as output map, and a matrix of connections between each output layer unit and all the input layer units. The input nodes receive vectorized input data patterns and propagate them to a set of output nodes, which are arranged according to map topology (generally two-dimensional grid) forming so called Kohonen’s layer. Kohonen’s principle of topographic map formation, determines how the spatial location of an output node in the topographic map corresponds to a particular feature of the input data pattern (Merkl, 1998). Each of the output nodes j is associated with a weight vector wj of the same dimension as the input data vectors and a position in the map space. The popular arrangement of output nodes is a regular spacing in a hexagonal or rectangular grid. The procedure for placing a vector from input data space onto the map is to first find the node with the spatially closest weight vector to the input data space vector. Once the closest node is located it is assigned the values from the vector taken from the data space. For Sentiment based classification, each vector represents an online review document, while the output node represents the sentiment category that the review is classified to. The SOM learns through self organization of random nodes whose weights are associated to the layers of neurons. These weights are updated at every epoch during the training process. The change in weights depends upon the similarity or spatial closeness between the input data pattern and the map pattern (Michalski et al, 1999). Let xi ∈ ℝM, 1≤ i ≤ N, be the randomly chosen training feature vector representing document i in the corpus, where M is the number of indexed features and N is the number of documents in input dataset. These vectors are used to train the map which consists of a regular spacing grid of processing units called neurons. Each neuron in the map has M connections (synapses). The synaptic weight vector wj is associated with the jth neuron in the map having J neurons where, wj = {wjm|1≤ m ≤ M}, 1≤ j ≤ J. The SOM learning algorithm used in this paper includes the following steps: Step 1. Initialize weights of neurons in map. Step 2. Randomly select a training vector xi from the corpus. Step 3. Determine the neuron c, having the highest activity level with respect to xi. The activity level of a neuron is represented by the Euclidean distance between the input pattern and that neuron’s weight vector. The neuron having the highest activity level is referred as best matching unit (BMU). Hence, the selection of BMU may be written as: c:  min ( xi (t ) - w j (t ) ) xi (t ) - wc (t ) (3) j The operator  denotes Euclidean vector norm as the neuron j with synaptic weights wj is closest to xi. The discrete time notation t denotes the current training iteration. Step 4. Update the synaptic weights of the BMU and its neighbors in order to move the BMU closer to the input vector. The map adapts to the input pattern in all iterations of learning, and adaption is performed as a gradual reduction of the spatial distance between the input vector and BMU’s weight vector. The learning rate parameter α(t) controls the amount of adaptation that gradually decreases during the course of training. A neighborhood kernel function is used to determine the spatial range of neurons around the BMU that are subject to adaptation. Based upon the spatial distance between BMU and neighboring neurons, the neighborhood function determines the width of the adaptation kernel. This ensures that neurons spatially close to the BMU are adapted more strongly than neurons further away. A Gaussian function (Eq. (4)) can serve as the neighborhood function with ri representing the two-dimensional vector pointing to the location of neuron j within the map, and rj  rc denoting the distance between neuron j and BMU (i.e., c) in terms of the output space. hc ( x ), j  e  rj  rc 2 2 / 2 (t ) i (4) The time-varying parameter σ guides spatial width of adaptation in such a way that a wide area of the output space is subject to adaptation at the beginning of training. The number of neighboring neurons affected by adaptation is reduced gradually during the training process. The learning rule for any output neuron j can be represented as w j (t  1)  w j (t )   (t )hc ( x ), j (t )   xi (t ) - w j (t )  i (5) Step 5. Increase time stamp t representing training iteration. If t reaches the preset maximum training time T, stop the training process; otherwise decrease α(t) (0< α(t) <1), and the neighborhood size, and go to Step 2. The current study has implemented SOM algorithm as a baseline model and its multi-pass implementation for sentiment analysis. The multi-pass SOM is recommended for better results where two passes are executed on the same underlying model. The first pass is executed as a rough ordering pass with large neighborhood, learning rate and small training time. The second pass is performed as the fine tuning pass that has a longer training time, small initial neighborhood and smaller initial learning rate (Kohonen et al, 1996). SOM have two advantages over other clustering methods: non-linear projection of the input space and cluster topology preservation. Thus, SOM can extract inherent non-linear relationships amongst documents, and similar clusters of documents are mapped close to each other revealing higher level (i.e. emergent structures) (Ultsch et al, 2005). In Emergent Self Organizing Maps (ESOM) the cluster boundaries are ‘indistinct’, the degree of separation between ‘regions’ of the map (i.e. clusters) being depicted by ‘gradients’ (Ultsch et al, 2005). These attractive features of SOM have prompted researchers to use SOM for text clustering and visualization (Yen et al, 2008; Feng et al, 2010; Dey, 2010). Learning Vector Quantization (LVQ) The Learning Vector Quantization (LVQ) algorithm is the supervised version of the Kohonen’s SOM model designed for pattern classification tasks (Kohonen, 1995). The application of the LVQ algorithm to sentiment classification of user generated text has not been explored so far. The LVQ algorithm is based on neural competitive learning, which enables defining a group of class labels on the input data space using reinforced learning. LVQ also transforms high dimensional input data into a two-dimensional map like SOM, but without taking into consideration the topology of input data. For this transformation, LVQ utilizes pre assigned class labels to documents, thus minimizing the average expected misclassification probability. Hence, unlike the SOM, where clusters are generated by unsupervised manner based on feature-vector similarities, the LVQ categories are predefined. LVQ is a single-layer network where, the number of output neurons is equal to number of predefined classes (e.g., J=2 for binary sentiment classification). The weight vectors associated with each output neuron are called codebook vectors. Each class of input space is represented by its own set of codebook vectors and one codebook vector per class is used for classification. Any output neuron j, whose weight vector matches the input pattern xi as per a spatial distance definition is the winning unit (BMU), and its weights are updated during training (Goren-Bar et al, 2005). The LVQ algorithm is a competitive one, as for each input training vector output neurons compete among themselves in order to find the BMU as per Euclidean distance. The LVQ learning rule modifies weights of only the BMU, as it has the smallest Euclidean distance with regard to the input vector. The LVQ algorithm implemented in this paper includes the following steps: Step 1. Initialize the codebook vectors wi. Step 2. Randomly select a training vector xi from the corpus. Step 3. Determine the BMU (winner unit) closest to the input vector. The codebook vector wc associated with BMU has the smallest Euclidean distance with regard to the input vector xi. c : xi (t ) - wc (t )  min ( xi (t ) - w j (t ) ) (6) j Step 4. Update the synaptic weights of the BMU.  For correction classification (i.e. if codebook vector wc and input vector xi belong to the same class) w j (t  1)  w j (t )   (t )  xi (t ) - w j (t )  (7)  For incorrect classification (i.e. if codebook vector wc and input vector xi belong to different class) w j (t  1)  w j (t )   (t )  xi (t ) - w j (t )  (8) Step 5. Increase time stamp t representing training iteration and reduce the learning rate α. Repeat from step 2 until the neural network is stabilized or until a fixed number of iterations have been carried out. The current study has implemented LVQ1 with single BMU selection, Optimized LVQ1 (OLVQ1) with separate learning rate for each codebook vector and Multi-Pass LVQ. Multi-Pass LVQ implements two passes where, a quick rough pass is made on the model using OLVQ1, then a long fine tuning pass is made on the model with LVQ1 (Kohonen, 1990). EXPERIMENTAL EVALUATION Datasets and Performance Evaluations To evaluate the performance of SOM and LVQ for sentiment classification, this paper experiments with a data set of classified movie reviews (Pang and Lee, 2004). The data set contains 1,000 positive and 1,000 negative reviews and is a benchmark dataset for sentiment analysis tasks. This work has adopted classification accuracy, precision, recall and F1 Score to evaluate the performance of sentiment classification. Experimental Results We attempted to test the effectiveness of ESOM in visualizing sentiment polarity in classified movie reviews data set. Input features were selected using Information Gain criteria, and experiments were conducted with 50 to 1000 selected features. Fig. 1. ESOM of Vectors with 100 Features Fig, 2. ESOM of Vectors with 700 Features The resulting visualization maps for 100 and 700 features are shown in Figures 1 and 2 respectively. In the maps, the Blue dots represent positive polarity and Yellow dots represent negative polarity reviews (as classed in the movie reviews dataset). It can be observed from Figures 1 and 2 that SOMs are able to distinguish between reviews with positive and negative sentiment in the form of clusters. It is also seen that more distinct and well separated clusters are formed when vectors with 700 features are used to represent the reviews. Java based implementations on Microsoft Windows platform were used to implement all the versions of SOM and LVQ algorithms. The number of indexed terms N to be used in input document vector was selected from 50 to 1000. All the algorithms were experimented with taking random training data, proportional initialization mode with linear decay learning function. For SOM, this paper has used Gaussian function as the neighborhood function with neighborhood size 8 for a hexagonal map topology as suggested in (Kohonen, 1995). The learning rate, α was selected 0.3 for SOM, LVQ1, OLVQ1 and for the first pass of multi-pass SOM and multi-pass LVQ. For second pass, the learning rate was kept as small as 0.05. The number of iteration was varied from 1000 to 10000 to obtain a training map. All experiments were validated using 10-fold cross validation. Figures 3, 4, 5, and 6 shows comparison of both SOM and LVQ implementations on accuracy, precision, recall, and F1 score of sentiment based classification on movie reviews dataset. The Multi-Pass LVQ performed best among all algorithms with 89.1% overall accuracy. The (classic Kohonen’s) SOM was the worst performer among all. Fig. 3. Performance Comparison on Accuracy Almost all implementations of SOM and LVQ the best performances were obtained for 700-800 selected features. These results are consistent with previous studies related to sentiment analysis, as a few sentiment bearing features can express enough polarity for the whole document to be classified (Turney, 2002). Degradation in the performance of sentiment based classification can be seen beyond 900 features as the feature matrix becomes sparse, affecting the organization of maps. The same algorithms with larger number of features (i.e. 2000 to 10000) were also executed in this study and it was observed that performance of the classifier degraded significantly due to high dimensionality of the input space. Fig. 4. Performance Comparison on Precision Fig. 5. Performance Comparison on Recall As such, with respect to learning methods, LVQ and SOM are all suitable for sentiment analysis as their best performance results are comparable with other studies using other machine learning classifiers (Pang et al, 2004) and other unsupervised learning methods (Turney, 2002). The time and computation resources sacrificed to execute the two passes in multi-pass implementations are justified as they have shown better performance as compared to their respective single-pass implementations. The OLVQ1 has worked well with 84.9% best overall accuracy, which has ranked it as the second best algorithm among all in this study. Though this study did not compare the execution time of all the algorithms but it is worth noting that OLVQ1 was the quickest among all in convergence. Fig. 6. Performance Comparison on F1 Score CONCLUSION This paper proposes a sentiment classification model using SOM and LVQ algorithms. Specifically, this paper investigates performance of different implementations of SOM for sentiment based visualization and classification of online product reviews. The supervised learning algorithms (LVQ, OLVQ1, and Multi-pass LVQ) have been found to perform better than unsupervised learning algorithm for sentiment analysis purpose. Experiments have been performed to support sentiment based visualization and classification tasks. Both SOM and LVQ algorithms have also been compared with multi-pass implementations, which performed better than their respective single-pass implementations. The experimental results on the online movie review data set show that SOMs are well suited for sentiment based classification and sentiment polarity visualization. REFERENCES Amati, G., Ambrosi, E., Bianchi, M., Gaibisso, C., and Gambosi, G. 2008. Automatic Construction of an Opinion-Term Vocabulary for Ad Hoc Retrieval. In Proceedings of ECIR 2008, LNCS (Vol. 4956, pp. 89–100). Berlin, Heidelberg: Springer-Verlag. Bai, X. 2011. Predicting consumer sentiments from online text. Decision Support Systems, 50, 4, 732-742. Dey, S. 2010. Enhancing Text Cluster Visualization in Emergent Self Organizing Maps by Incorporating Bigram Vectors. Proceedings of the Third International Conference on Modeling, Simulation and Optimization, Beijing, China (CMSO 2010), 21-26. Feng Z., Bao J., Shen J. 2010, Dynamic and Adaptive Self Organizing Maps applied to High Dimensional Large Scale Text Clustering, IEEE International Conference on Software Engineering and Service Sciences (ICSESS). Goren-Bar, D. and Kuflik, T. 2005. Supporting user-subjective categorization with self-organizing maps and learning vector quantization. J. Am. Soc. Inf. Sci. Technol., 56, 4, 345-355. Han, E. H., Karypis, G., and Kumar, V. 1999. Text Categorization Using Weight Adjusted K-Nearest Neighbour Classification. Technical Report. Department of Computer Science and Engineering, Army HPC Research Center, University of Minnesota. He, Y. and Zhou, D. 2011. Self-training from labeled features for sentiment analysis. Information Processing &amp; Management, 47, 4, 606-616. Hu, M. and Liu, B. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Seattle, WA, USA). ACM. Kohonen, T. 1995. Self-organizing maps. Springer-Verlag, Berlin. Kohonen, T. 1990. Improved versions of learning vector quantization. In Proc. Int. Joint Conf. on Neural Networks, (San Diego, CA). 545–550. Kohonen, T., Hynninen, J., Kangas, J., and Laaksonen, J. 1996. SOM_PAK: The Self-Organizing Map Program Package, Technical Report A31, Helsinki University of Technology, Laboratory of Computer and Information Science. Merkl, D. 1998. Text classification with self-organizing maps: Some lessons learned. Neurocomputing, 21, 1–3, 61-77. Michalski, R. S., Bratko, I., and Kubat, M. 1999. Machine learning and data mining methods and applications. Wiley. Pang, B. and Lee, L. 2008. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1-2), 1-135. Pang, B. and Lee, L. 2004. A sentimental education: sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd annual meeting of the Association for Computational Linguistics (ACL) (Barcelona, Spain, July 21–26, 2004). 271–278. Pang, B., Lee, L., and Vaithyanathan, S. 2002. Thumbs up? sentiment classification using machine learning techniques. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing. ACL. Volume 10. Quinlan, J. R. 1993. C4.5: Program for Machine Learning. San Mateo, CA: Morgan Kaufmann. Sharma, A., & Dey, S. 2012a. A comparative study of feature selection and machine learning techniques for sentiment analysis. Proceedings of the 2012 ACM-SIGAPP Research in Applied Computation Symposium, San Antonio, Texas, 1-7. Sharma, A., & Dey, S. 2012b. A document-level sentiment analysis approach using artificial neural network and sentiment lexicons. SIGAPP Applied Computing Review, 12, 4, 67-75. Tan, S. and Zhang, J. 2008. An empirical study of sentiment analysis for Chinese documents. Expert Systems with Applications, 34, 4, 2622-2629. Turney, P. D. 2002. Thumbs up or thumbs down?: semantic orientation applied to unsupervised classification of reviews. In Proceedings of the 40th annual meeting on ACL (Morristown, NJ, USA) ACL, 417–424. Ultsch A., Moerchen F. 2005, ESOM-Maps: tools for clustering, visualization, and classification with Emergent SOM, Technical Report No. 46, Dept. of Mathematics and Computer Science, University of Marburg, Marburg, Germany. Yen G. G. and Wu Z. 2008, Ranked Centroid Projection: A Data Visualization Approach With Self-Organizing Maps, IEEE Transactions on Neural Networks, 19, 2.
9cs.NE
arXiv:1801.04658v1 [cs.IT] 15 Jan 2018 Information Geometric Approach to Bayesian Lower Error Bounds M. Ashok Kumar Kumar Vijay Mishra Discipline of Mathematics Indian Institute of Technology Indore, Madhya Pradesh, India Email: [email protected] IIHR - Hydroscience & Engineering The University of Iowa Iowa City, IA, USA Email: [email protected] Abstract—Information geometry describes a framework where probability densities can be viewed as differential geometry structures. This approach has shown that the geometry in the space of probability distributions that are parameterized by their covariance matrix is linked to the fundamentals concepts of estimation theory. In particular, prior work proposes a Riemannian metric - the distance between the parameterized probability distributions - that is equivalent to the Fisher Information Matrix, and helpful in obtaining the deterministic Cramér-Rao lower bound (CRLB). Recent work in this framework has led to establishing links with several practical applications. However, classical CRLB is useful only for unbiased estimators and inaccurately predicts the mean square error in low signal-to-noise (SNR) scenarios. In this paper, we propose a general Riemannian metric that, at once, is used to obtain both Bayesian CRLB and deterministic CRLB along with their vector parameter extensions. We also extend our results to the Barankin bound, thereby enhancing their applicability to low SNR situations. I. I NTRODUCTION Information geometry is a study of statistical models from a Riemannian geometric perspective. The differential geometric modeling methods were introduced to statistics by C. R. Rao in his seminal paper [1] and later formally developed by Cencov [2]. The information geometric concept of a manifold that describes the parameterized probability distributions has garnered considerable interest in recent years. The main advantages include structures in the space of probability distributions that are invariant to non-singular transformation of parameters [3], robust estimation of covariance matrices [4] and usage of Fisher Information Matrix (FIM) as a metric [5]. Information geometry has now transcended its initial statistical scope and expanded to several novel research areas including, but not limited to, Fisher-Rao Riemannian geometry [6], Finsler information geometry [7], optimal transport geometry [8], and quantum information geometry [9]. Many problems in science and engineering use probability distributions and, therefore, information geometry has been used as a useful and rigorous tool for analyses in applications such as neural networks [10, 11], optimization [12, 13], radar systems [14, 15], communications [16], computer vision [6], and machine learning [17, 18]. More recently, several developments in deep learning [19, 20] that employ various approximations to the FIM to calculate the gradient descent have incorporated information geometric concepts. The information geometry bases the distance between the parameterized probability distributions on the FIM [3]. In estimation theory, the well-known deterministic Cramér-Rao lower bound (CRLB) is the inverse of FIM. Therefore, the results derived from information geometry are directly connected with the fundamentals of estimation theory. Nearly all prior works exploited this information geometric link to CRLB in their analyses because the CRLB is most widely used benchmark for evaluating the mean square error (MSE) performance of an estimator. However, the classical CRLB holds only if the estimator is unbiased. In general, the estimators are biased in many practical problems such as nonparametric regression [21], communication [22], and radar [23]. The above-mentioned information geometric framework ceases its utility in these cases. Moreover, the classical CRLB is a tight bound only when the errors are small. It is well known that in case of the nonlinear estimation problems with finite support parameters, for example the time delay estimation in radar [23], the performance of the estimator is characterized by the presence of three distinct signal-to-noise-ratio (SNR) regions [24]. When the observations are large or the SNR is high (asymptotic region), the CRLB describes the MSE accurately. In case of few observations or low SNR regions, the information from signal observations is insufficient and the estimator criterion is hugely corrupted by the noise. Here, the MSE is close to that obtained via only a priori information, that is, a quasi-uniform random variable on the parameter support. In between these two limiting cases lies the threshold region where the signal observations are subjected to ambiguities and the estimator MSE increases sharply due to the outlier effect. The CRLB is used only in the asymptotic area and is not an accurate predictor of MSE when the performance breaks down due to increase in noise. In this paper, we propose the information geometric framework that addresses these drawbacks of the classical CRLB. The Bayesian CRLB [21] is typically used for assessing the quality of biased estimators. It is similar to the deterministic CRLB except that it assumes the parameters to be random with an a priori probability density function. We develop a general Riemannian metric that can be modified to link to both Bayesian and deterministic CRLB. To address the problem of the threshold effect, other bounds that are tighter than the CRLB have been developed (see e.g. [25] for an overview) to accurately identify the SNR thresholds that define the ambiguity region. In particular, Barankin bound [26] is a fundamental statistical tool to understand the threshold effect for unbiased estimators. In simple terms, the threshold effect can be understood as the region where Barankin deviates from the CRLB [27]. In this paper, we show that our metric is also applicable for the Barankin bound. Hence, compared to previous works [3], our information geometric approach to minimum bounds on MSE holds good for both the Bayesian CRLB and deterministic CRLB, their vector equivalents and the threshold effect through the Barankin bound. The paper is organized as follows: In the next section, we provide a brief background to the information geometry and describe the notation used in the later sections. Further, we explain the dual structure of the manifolds because, in most applications, the underlying manifolds are dually flat. Here, we define a divergence function between two points in a manifold. In Section III, we establish the connection between the Riemannian metric and the Kullback-Leibler divergence for the Bayesian case. The manifold of all discrete probability distributions is dually flat and the Kullback-Leibler divergence plays a key role here. We also show that, under certain conditions, our approach yields the previous results from [3]. In Section IV, we state and prove our main result applicable to several other bounds before providing concluding remarks in Section V. space of all probability distributions on X . Let S ⊂ P be a sub-manifold. Let θ = (θ1 , . . . , θk ) be a parameterization of S. Given a divergence function 1 on S, Eguchi [31] defines a Riemannian metric on S by the matrix i h (D) G(D) (θ) = gi,j (θ) , where (D) gi,j (θ) := −D[∂i , ∂j ] ∂ ∂ := − D(pθ , pθ′ ) ∂θi ∂θj′ where gi,j is the elements in the ith row and jth column of the matrix G, θ = (θ1 , . . . , θn ), θ′ = (θ1′ , . . . , θn′ ), and dual affine ∗ connections ∇(D) and ∇(D ) , with connection coefficients described by following Christoffel symbols (D) Γij,k (θ) := −D[∂i ∂j , ∂k ] := − ∂ ∂ ∂ D(pθ , pθ′ ) ∂θi ∂θj ∂θk′ θ=θ ′ and (D∗ ) Γij,k (θ) := −D[∂k , ∂i ∂j ] := − ∂ ∂ ∂ D(pθ , pθ′ ) ∂θk ∂θi′ ∂θj′ such that, ∇(D) and ∇(D sense that ∗ (D) ) , θ=θ ′ form a dualistic structure in the (D) (D∗ ) ∂k gi,j = Γki,j + Γkj,i , II. I NFORMATION G EOMETRY: A B RIEF BACKGROUND A n-dimensional manifold is a Hausdorff and second countable topological space which is locally homeomorphic to Euclidean space of dimension n [28–30]. A Riemannian manifold is a real differentiable manifold in which the tangent space at each point is a finite dimensional Hilbert space and, therefore, equipped with an inner product. The collection of all these inner products is called a Riemannian metric. In the information geometry framework, the statistical models play the role of a manifold and the Fisher information matrix and its various generalizations play the role of a Riemannian metric. Formally, by a statistical manifold, we mean a parametric family of probability distributions S = {pθ : θ ∈ Θ} with a “continuously varying" parameter space Θ (statistical model). The dimension of a statistical manifold is the dimension of the parameter space. For example, S = {N (µ, σ 2 ) : µ ∈ R, σ 2 > 0} is a two dimensional statistical manifold. The tangent space at a point of S is a linear space that corresponds to a “local linearization” at that point. The tangent space at a point p of S is denoted by Tp (S). The elements of Tp (S) are called tangent vectors of S at p. A Riemannian metric at a point p of S is an inner product defined for any pair of tangent vectors of S at p. In this paper, let us restrict to statistical manifolds defined on a finite set X = {a1 , . . . , ad }. Let P := P(X ) denote the θ=θ ′ (1) where D∗ (p, q) = D(q, p). III. F ISHER I NFORMATION M ATRIX FOR C ASE THE BAYESIAN Eguchi’s theory in section II can also be extended to the space P̃(X ) of all measures on X . That is, P̃ = {p̃ : X → (0, ∞)}. Let S = {pθ : θ = (θ1 , . . . , θk ) ∈ Θ} be a k-dimensional sub-manifold of P and let S̃ := {p̃θ (x) = pθ (x)λ(θ) : pθ ∈ S}, (2) where λ is a probability distribution on Θ. Then S̃ is a k + 1-dimensional sub-manifold of P̃. For p̃θ , p̃θ′ ∈ P̃, the Kullback-Leibler divergence (KL-divergence) between p̃θ and p̃θ′ is given by I(p̃θ kp̃θ′ ) = X x = X x p̃θ (x) log X p̃θ (x) X − p̃θ (x) + p̃θ′ (x) p̃θ′ (x) x x pθ (x)λ(θ) log pθ (x)λ(θ) − λ(θ) + λ(θ′ ). pθ′ (x)λ(θ′ ) (3) 1 By a divergence, we mean a non-negative function D defined on S × S such that D(p, q) = 0 iff p = q. (I) We define a Riemannian metric G(I) (θ) = [gi,j (θ)] on S̃ by and (e) kωX kp̃ := kXkp̃ = (I) gi,j (θ) := −I[∂i k∂j ] = = = pθ (x)λ(θ) ∂ ∂ X pθ (x)λ(θ) log − ∂θi ∂θj′ x pθ′ (x)λ(θ′ ) X ∂i (pθ (x)λ(θ)) · ∂j log(pθ (x)λ(θ)) x X = Now, for a (smooth) real function f on S̃, the differential of f at p̃, (df )p̃ , is a member of Tp̃∗ (S) which maps X to X(f ). The gradient of f at p̃ is the tangent vector corresponding to (df )p̃ , hence satisfies θ ′ =θ (e) (df )p̃ (X) = X(f ) = h(gradf )p̃ , Xip̃ , pθ (x)λ(θ)∂i (log pθ (x)λ(θ)) · ∂j (log(pθ (x)λ(θ))) x λ(θ) X (e) ·[∂j (log pθ (x)) + ∂j (log λ(θ))]  λ(θ) Eθ [∂i log pθ (X) · ∂j log pθ (X)] · + ∂i (log λ(θ)) · ∂j (log λ(θ)) = where  (e) λ (θ) , λ(θ) gi,j (θ) + Ji,j k(df )p̃ k2p̃ = h(gradf )p̃ , (gradf )p̃ ip̃ . (e) (10) Since gradf is a tangent vector, we can write gradf = (4) (5) gi,j (θ) := Eθ [∂i log pθ (X) · ∂j log pθ (X)], (9) and pθ (x)[∂i (log pθ (x)) + ∂i (log λ(θ))] x = q (e) hX, Xip̃ . (6) k X hi ∂i (11) i=1 for some scalars hi . Applying (9) with X = ∂j , for each j = 1, . . . , k, and using (11), we get +(e) * k X hi ∂i , ∂j (∂j )(f ) = i=1 and λ Ji,j (θ) := ∂i (log λ(θ)) · ∂j (log λ(θ)). (7) k X = i=1 (e) λ (θ)]. Then Let G(e) (θ) := [gi,j (θ)] and J λ (θ) := [Ji,j   G(I) (θ) = λ(θ) G(e) (θ) + J λ (θ) . (e) Tp̃ (P̃) = {X (e) : X ∈ Tp̃ (P̃)} = {A ∈ RX̃ : Ep̃ [A] = 0}. For the natural basis ∂i of a coordinate system θ = (θi ), (m) (e) (∂i )ξ = ∂i p̃θ and (∂i )ξ = ∂i log p̃θ . With these notations, for any two tangent vectors X, Y ∈ Tp̃ (P̃), the Fisher metric in (5) can be written as (e) hX, Y ip̃ = Ep̃ [X (e) Y (e) ]. Let S̃ be a sub-manifold of P̃ of the form as in (2), together with the metric G(I) as in (5). Let Tp̃∗ (S̃) be the dual space (cotangent space) of the tangent space Tp̃ (S̃) and let us consider for each Y ∈ Tp̃ (S̃), the element ωY ∈ Tp̃∗ (S̃) which maps X to hX, Y i(e) . The correspondence Y 7→ ωY is a linear map between Tp̃ (S̃) and Tp̃∗ (S̃). An inner product and a norm on Tp̃∗ (S̃) are naturally inherited from Tp̃ (S̃) by hωX , ωY ip̃ := hX, Y k X = (8) Notice that G(e) (θ) is the usual Fisher information matrix. Also observe that P̃ is a subset of RX̃ , where X˜ := X ∪ {ad+1 }. The tangent space at every point of P̃ is A0 := {A ∈ P RX̃ : x∈X̃ A(x) = 0}. That is, Tp (P̃) = A0 . We denote a tangent vector (that is, elements of A0 ) by X (m) . The manifold P̃ can be recognized by its homeomorphic image {log p̃ : p̃ ∈ P̃} under the mapping p̃ 7→ log p̃. Under this mapping the tangent vector X ∈ Tp̃ (P̃) can be represented X (e) which is defined by X (e) (x) = X (m) (x)/p̃(x) and we define (e) ip̃ hi h∂i , ∂j i(e) (e) hi gi,j , j = 1, . . . , k. i=1 From this, we have and so i−1 h [∂1 (f ), . . . , ∂k (f )]T , [h1 , . . . , hk ]T = G(e) X (g i,j )(e) ∂j (f )∂i . (12) From (9), (10), and (12), we get X (g i,j )(e) ∂j (f )∂i (f ) k(df )p̃ k2p̃ = (13) gradf = i,j i,j where (g i,j )(I) is the (i, j)th entry of the inverse of G(I) . The above and the following results are indeed an extension of [3, Sec.2.5.] to P̃. Theorem 1 ( [3]): Let A : X → R be any mapping (that is, a vector in RX . Let E[A] : P̃ → R be the mapping p̃ 7→ Ep̃ [A]. We then have Var(A) = k(dEp̃ [A])p̃ k2p̃ . Proof: For any tangent vector X ∈ Tp̃ (P̃), X X(Ep̃ [A]) = X(x)A(x) (14) x (e) = Ep̃ [Xp̃ · A] = (e) Ep̃ [Xp̃ (A − Ep̃ [A])]. (15) (16) (e) Since A − Ep̃ [A] ∈ Tp̃ (P̃), there exists Y ∈ Tp̃ (P̃) such that (e) A − Ep̃ [A] = Yp̃ , and grad(E[A]) = Y . Hence we see that k(dE[A])p̃ k2p̃ (e) (e) = Ep [Yp̃ Yp̃ ] = Ep̃ [(A − Ep̃ [A]) ]. This implies that X X ci cj (g (I) )i,j (θ). (21) ci cj Covθ (θbi , θbj ) ≥ λ(θ) Corollary 2 ( [3]): If S is a submanifold of P̃, then Varp̃ [A] ≥ k(dE[A]|S )p̃ k2p̃ (17) A − Ep̃ [A] ∈ {Xp̃ : X ∈ Tp̃ (S)} =: Tp̃ (S). E RROR B OUNDS We state our main result in the following theorem. Theorem 3: Let S and S̃ be as in (2). Let θb be an estimator of θ. Then (a) Bayesian Cramér-Rao:    b ≥ Eλ [G(I) (θ)] −1 , (18) Eλ Varθ (θ) b where Varθ (θ) = [Covθ (θbi (X), θbj (X))] is the covariance matrix and G(I) (θ) is as in (8). (In (18), we use the usual convention that, for two matrices A and B, A ≥ B means that A − B is positive semi-definite.) (b) Deterministic Cramér-Rao (unbiased): If θb is an unbiased estimator of θ, then (19) (c) Deterministic Cramér-Rao (biased): For any estimator θb of θ, b ≥ (1 + B ′ (θ))[G(e) (θ)]−1 (1 + B ′ (θ)) MSEθ [θ] +b(θ)b(θ)T , b is the bias where b(θ) = (b1 (θ), . . . , bk (θ))T := Eθ [θ] and 1 + B ′ (θ) is the matrix whose (i, j)th entry is 0 if i 6= j and is (1 + ∂i bi (θ)) if i = j. (d) Barankin Bound: (Scalar case) If θb be an unbiased estimator of θ, then hP i2 n al (θ(l) − θ) l=1 b ≥ sup , (20) Varθ [θ] i2 n Ph P n,al ,θ (l) al Lθ(l) (x) pθ (x) x i,j That is, Proof: Since (grad E[A]|S )p̃ is the orthogonal projection of (grad E[A])p̃ onto Tp̃ (S), the result follows from the theorem. b ≥ [G(e) (θ)]−1 . Varθ [θ] Hence, taking expectation on both sides with respect to λ, we get X     X ci cj Eλ (g (I) )i,j (θ) . ci cj Eλ Covθ (θbi , θbj ) ≥ i,j (e) (e) i,j i,j with equality iff OF i,j i,j 2 IV. D ERIVATION Pk b b b b (a) Let A = i=1 ci θi , where θ = (θ1 , . . . , θk ) is an unbiased estimator of θ, in corollary 2. Then, from (17), we have X X ci cj (g (I) )i,j (θ). ci cj Covθ̃ (θbi , θbj ) ≥ l=1 where Lθ(l) (x) := pθ(l) (x)pθ (x) and the supremum is over all a1 , . . . , an ∈ R, n ∈ N, and θ(1) , . . . , θ(n) ∈ Θ. Proof:   b ≥ Eλ [G(I) (θ)−1 ]. Eλ Varθ (θ)  −1 But Eλ [G(I) (θ)−1 ] ≥ Eλ [G(I) (θ)] by [32]. This proves the result. (b) This follows from (21) by taking λ(θ) = 1. (c) Let us first observe P that θb is an unbiased estimator of k θ + b(θ). Let A = i=1 ci θbi as before. Then E[A] = Pk i=1 ci (1 + bi (θ)). Then, from corollary 2 and (17), we have b ≥ (1 + B ′ (θ))[G(e) (θ)]−1 (1 + B ′ (θ)) Varθ [θ] b = Varθ [θ] b + b(θ)b(θ)T . This proves the But MSEθ [θ] assertion. (d) For fixed a1 , . . . , an ∈ R and θ(1) , . . . , θ(n) ∈ Θ, let us define a metric by the following formula n i2 XhX g(θ) := al Lθ(l) (x) pθ (x). (22) x l=1 b − θ, Let f be the mapping p 7→ Ep [A]. Let A(·) = θ(·) b where θ is an unbiased estimator of θ. Then the partial derivatives of f in (13) equals to n X  X b − θ) (θ(x) al Lθ(l) (x) pθ (x) x = l=1 n X al l=1 = n X X  b − θ) pθ(l) (x) pθ (x) (θ(x) pθ (x) x al (θ(l) − θ). l=1 Hence from (13) and corollary 2, we have hP i2 n al (θ(l) − θ) l=1 b ≥ Varθ [θ] i2 n Ph P al Lθ(l) (x) pθ (x) x (l) l=1 Since al and θ are arbitrary, taking supremum over all a1 , . . . , an ∈ R, n ∈ N, and θ(1) , . . . , θ(n) ∈ Θ, we get (20). V. C ONCLUSION We have shown that our Theorem 3 provides a general information geometric characterization of the statistical manifolds linking them to the Bayesian CRLB for vector parameters; the extension to estimators of measurable functions of the parameter θ is trivial. We exploited the general definition of Kullback-Leibler divergence when the probability densities are not normalized. This is an improvement over Amari’s work [3] on information geometry which only dealt with the notion of deterministic CRLB of scalar parameters. Further, we proposed an approach to arrive at the Barankin bound thereby shedding light on the relation between the threshold effect and information geometry. Both of our improvements enable usage of information geometric approaches in critical scenarios of biased estimators and low SNRs. This is especially useful in the analyses of many practical problems such as radar and communication. In future investigations, we intend to explore these methods further especially in the context of the threshold effect. R EFERENCES [1] C. R. Rao, “Information and the accuracy attainable in the estimation of statistical parameters,” Bulletin of Calcutta Mathematical Society, vol. 37, pp. 81–91, 1945. [2] N. N. Cencov, Statistical decision rules and optimal inference (Translations of mathematical monographs). American Mathematical Society, 2000, no. 53. [3] S. Amari and H. Nagaoka, Methods of information geometry. American Mathematical Society, Oxford University Press, 2000, vol. 191. [4] B. Balaji, F. Barbaresco, and A. Decurninge, “Information geometry and estimation of toeplitz covariance matrices,” in IEEE Radar Conference, 2014, pp. 1–4. [5] F. Nielsen, “Cramér-Rao lower bound and information geometry,” arXiv preprint arXiv:1301.3578, 2013. [6] S. J. Maybank, S. Ieng, and R. Benosman, “A Fisher-Rao metric for paracatadioptric images of lines,” International Journal of Computer Vision, vol. 99, no. 2, pp. 147–165, 2012. [7] Z. Shen, “Riemann-Finsler geometry with applications to information geometry,” Chinese Annals of Mathematics-Series B, vol. 27, no. 1, pp. 73–94, 2006. [8] W. Gangbo and R. J. McCann, “The geometry of optimal transportation,” Acta Mathematica, vol. 177, no. 2, pp. 113–161, 1996. [9] M. R. Grasselli and R. F. Streater, “On the uniqueness of the Chentsov metric in quantum information geometry,” Infinite Dimensional Analysis, Quantum Probability and Related Topics, vol. 4, no. 02, pp. 173–182, 2001. [10] S. Amari, “Information geometry of neural networks: An overview,” in Mathematics of Neural Networks, ser. Operations Research/Computer Science Interfaces Series, S. W. Ellacott, J. C. Mason, and I. J. Anderson, Eds. Springer US, 1997, vol. 8. [11] ——, “Information geometry of neural learning and belief propagation,” in IEEE International Conference on Neural Information Processing, vol. 2, 2002, pp. 886–vol. [12] S. Amari and M. Yukawa, “Minkovskian gradient for sparse optimization,” IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 4, pp. 576–585, 2013. [13] H.-G. Beyer, “Convergence analysis of evolutionary algorithms that are based on the paradigm of information geometry,” Evolutionary Computation, vol. 22, no. 4, pp. 679–709, 2014. [14] E. de Jong and R. Pribić, “Design of radar grid cells with constant information distance,” in IEEE Radar Conference, 2014, pp. 1–5. [15] F. Barbaresco, “Innovative tools for radar signal processing based on Cartan’s geometry of SPD matrices & information geometry,” in IEEE Radar Conference, 2008, pp. 1–6. [16] M. Coutino, R. Pribić, and G. Leus, “Direction of arrival estimation based on information geometry,” in IEEE International Conference on Acoustics, Speech and Signal Processing, 2016, pp. 3066–3070. [17] K. Sun and S. Marchand-Maillet, “An information geometry of statistical manifold learning,” in International Conference on Machine Learning, 2014, pp. 1–9. [18] S. Amari, “Natural gradient works efficiently in learning,” Neural computation, vol. 10, no. 2, pp. 251–276, 1998. [19] G. Desjardins, K. Simonyan, R. Pascanu et al., “Natural neural networks,” in Advances in Neural Information Processing Systems, 2015, pp. 2071–2079. [20] N. L. Roux, P.-A. Manzagol, and Y. Bengio, “Topmoumoute online natural gradient algorithm,” in Advances in neural information processing systems, 2008, pp. 849–856. [21] R. D. Gill and B. Y. Levit, “Applications of the van Trees inequality: A Bayesian Cramér-Rao bound,” Bernoulli, pp. 59–79, 1995. [22] R. Prasad and C. R. Murthy, “Cramér-rao-type bounds for sparse bayesian learning,” IEEE Transactions on Signal Processing, vol. 61, no. 3, pp. 622–632, 2013. [23] K. V. Mishra and Y. C. Eldar, “Performance of time delay estimation in a cognitive radar,” in IEEE Int. Conf. Acoustics, Speech and Signal Process., 2017, pp. 3141–3145. [24] H. L. Van Trees, K. L. Bell, and Z. Tian, Detection Estimation and Modulation Theory, Part I: Detection, Estimation, and Filtering Theory, 2nd ed. Wiley, 2013. [25] A. Renaux, P. Forster, P. Larzabal, C. D. Richmond, and A. Nehorai, “A fresh look at the Bayesian bounds of the Weiss-Weinstein family,” IEEE Transactions on Signal Processing, vol. 56, no. 11, pp. 5334–5352, 2008. [26] E. W. Barankin, “Locally best unbiased estimates,” The Annals of Mathematical Statistics, pp. 477–501, 1949. [27] L. Knockaert, “The Barankin bound and threshold behavior in frequency estimation,” IEEE Transactions on Signal Processing, vol. 45, no. 9, pp. 2398–2401, 1997. [28] S. Gallot, D. Hulin, and J. Lafontaine, Methods of information geometry. Riemannian Geometry, 2004. [29] J. Jost, Riemannian Geometry and Geometric Analysis. Springer, 2005. [30] M. Spivak, A Comprehensive Introduction to Differential Geometry Volume I. Publish or Perish Inc., 2005. [31] S. Eguchi, “Geometry of minimum contrast,” Hiroshima Mathematical Journal, vol. 22, no. 3, pp. 631–647, 1992. [32] T. Groves and T. Rothenberg, “A note on the expected value of an inverse matrix,” Biometrika, vol. 56, pp. 690–691, 1969.
7cs.IT
arXiv:cs/0309028v1 [cs.PL] 16 Sep 2003 Under consideration for publication in Theory and Practice of Logic Programming 1 cTI: A Constraint-Based Termination Inference Tool for ISO-Prolog FRED MESNARD Iremia, Université de La Réunion, Saint Denis, France (e-mail: [email protected]) ROBERTO BAGNARA Department of Mathematics, University of Parma, Italy (e-mail: [email protected]) Abstract We present cTI, the first system for universal left-termination inference of logic programs. Termination inference generalizes termination analysis and checking. Traditionally, a termination analyzer tries to prove that a given class of queries terminates. This class must be provided to the system, for instance by means of user annotations. Moreover, the analysis must be redone every time the class of queries of interest is updated. Termination inference, in contrast, requires neither user annotations nor recomputation. In this approach, terminating classes for all predicates are inferred at once. We describe the architecture of cTI and report an extensive experimental evaluation of the system covering many classical examples from the logic programming termination literature and several Prolog programs of respectable size and complexity. KEYWORDS: Termination Inference; Termination Analysis; Logic Programming; Abstract Interpretation. 1 Introduction Termination is a crucial aspect of program verification. It is of particular importance for logic programs (Lloyd 1987; Apt 1997), since there are no a priori syntactic restrictions to queries and, as a matter of fact, most predicates programmers tend to write do not terminate for their most general queries. In the last fifteen years, termination has been the subject of several research works in the field of logic programming (see, for instance, (Francez et al. 1985; Apt and Pedreschi 1990; Ruggieri 1999)). In contrast to what happens for other programming paradigms, there are two notions of termination for logic programs (Vasak and Potter 1986): existential and universal termination. To illustrate them, assume we are using a standard Prolog engine. Existential termination of a query means that either the computation finitely fails or it produces one solution in finite time. This does not exclude the possibility that the engine, when asked for further solutions, will loop. On the other hand, universal termination means that the computation yields all 2 F. Mesnard and R. Bagnara solutions and eventually fails in finite time (if we repeatedly ask for further solutions). Although the concept of existential termination plays an important role in connection with normal logic programs, it has severe drawbacks that make it not appropriate in other contexts: existential termination is not instantiation-closed (i.e., a goal may existentially terminate, yet some of its instances may not terminate), hence it is not and-compositional (i.e., two goals may existentially terminate while their conjunction does not); finally, existential termination depends on the textual order of clauses in the program. Universal termination is a stronger and much more robust concept: it implies existential termination and it is both and-compositional and instantiation-closed. Existential termination has been the subject of only a few works (Vasak and Potter 1986; Levi and Scozzari 1995; Marchiori 1996) whereas most research focused on universal termination. There are two main directions (see (De Schreye and Decorte 1994) for a survey): characterizing termination (Apt and Pedreschi 1990; Apt and Pedreschi 1993; Ruggieri 1999) and finding weaker but decidable sufficient conditions that lead to actual algorithms, e.g., (Ullman and Van Gelder 1988; Plümer 1990; Verschaetse 1992). Even though our research belongs to both streams, in this paper we focus on an intuitive presentation of the implementation of our approach. A companion paper presents a complete formalization of our work in the theoretical setting of acceptability for constraint logic programs (Mesnard and Ruggieri 2003), where we refine a necessary and sufficient condition for termination to the sufficient condition implemented in cTI. Our main contribution compared to other automated termination analyzers (Lindenstrauss and Sagiv 1997; Decorte 1997; Speirs et al. 1997; Codish and Taboch 1999) is that our tool infers sufficient universal termination conditions from the text of any Prolog program, adopting a bottom-up approach to termination. An important feature of this approach first presented in (Mesnard 1996) is that there is no need to define in advance a class of queries of interest. (If required, these classes can be provided after the analysis has finished in order to specialize the obtained results.) Our system, called cTI from constraint-based Termination Inference, is written in SICStus Prolog. A preliminary account of the work described in this paper appeared in (Mesnard and Neumerkel 2001), where we showed that numeric computations took most of the execution times. Now cTI relies on the specialized Parma Polyhedra Library (Bagnara et al. 2002), a modern C++ library for the manipulation of convex polyhedra that significantly speeds up the analysis. Moreover, cTI has been extended so that it can analyze any ISO-Prolog program (ISO/IEC 1995; Deransart et al. 1996). The only correctness requirement we currently impose on programs is that they must not create infinite rational terms. Hence we assume execution with occurs-check or, equivalently, NSTO programs (i.e., programs that are Not Subject to Occur-Check (Deransart et al. 1991) and thus are safely executed with any standard conforming system). We point out that simple, sufficient syntactic methods for ensuring occurs-check freedom are presented in (Apt and Pellegrini 1994) while (Søndergaard 1986; Crnogorac et al. 1996) describe abstract-interpretation based solutions. Recently, finite-tree analysis (Bagnara et al. 2001a; Bagnara et al. 2001b) cTI: A Termination Inference Tool for Prolog 3 has been proposed to confine infinite rational terms in programs that are not occurscheck free. Both the approach described in (Mesnard and Ruggieri 2003) and the cTI system can be extended, with the help of finite-tree analysis, to deal also with such programs. Throughout the paper we assume a basic knowledge of logic programming (see, e.g., (Apt 1997)), constraint logic programming (see, e.g., (Marriott and Stuckey 1998)), abstract interpretation (see, e.g., (Cousot and Cousot 1992)), and propositional µcalculus (see, e.g., (Clarke et al. 2000)). In Section 2 we present cTI informally with an example analysis. How to use cTI is described in Section 3. An experimental evaluation of the system is the subject of Section 4. Related work is discussed in Section 5 while Section 6 concludes. 2 An Overview of cTI Our aim is to compute classes of queries for which universal left termination is guaranteed. We call such classes termination conditions. More precisely, let P be a Prolog program and q a predicate symbol of P . A termination condition for q is a set TCq of goals of the form ← c, q(x̃) where c is a CLP(H) constraint such that, for any goal G ∈ TCq , each derivation of P and G using the left-to-right selection rule is finite. Our analyzer uses three main constraint structures: Herbrand terms for the initial program P (seen as a CLP(H) program), non-negative integers, and booleans (P is abstracted into both a CLP(N ) and a CLP(B) program). We illustrate our method to infer termination conditions by means of an example. The method consists of six distinct steps, which will be illustrated on the following definition for the predicates app/3, nrev/2 and app3/4. app([], X, X). app([E|X], Y, [E|Z]) :app(X, Y, Z). nrev([], []). app3(X, Y, Z, U) :nrev([E|X], Y) :app(X, Y, V), nrev(X, Z), app(V, Z, U). app(Z, [E], Y). Step 1: From Prolog to CLP(N ). From the Prolog program P , a CLP(N ) program P N is obtained by applying a symbolic norm. In our example, we use the term-size norm, which is the one cTI applies by default. All ISO-predefined predicates have been manually pre-analyzed for this norm. Notice that, as explained in (Mesnard and Ruggieri 2003), termination inference for pure Prolog programs can be based on any linear norm. The symbolic term-size norm is inductively defined as follows:  Pn   1 + i=1 kti kterm-size , if t = f (t1 , . . . , tn ) with n > 0; def ktkterm-size = 0, if t is a constant;   t, if t is a variable. For example, kf (0, 0)kterm-size = 1. All non-monotonic elements of the program are approximated by monotone constructs. For instance, Prolog’s unsound nega- 4 F. Mesnard and R. Bagnara tion \+ G is approximated by ((G, false) ; true). More generally, extra-logical predicates are mapped to their first-order counterparts so that the termination property is preserved. For our running example, we obtain the following CLP(N ) clauses: nrevN (0, 0). app3N (x, y, z, u) ← appN (0, x, x). appN (x, y, v), appN (1 + e + x, y, 1 + e + z) ← nrevN (1 + e + x, y) ← nrevN (x, z), appN (v, z, u). appN (x, y, z). appN (z, 1 + e, y). Step 2: Computing a numeric model. A model of the CLP(N ) program is now computed. For each predicate p, the model describes, with a finite conjunction of linear equalities and inequalities denoted by postN p , the linear inter-argument relations that hold for every solution of p. In our example we obtain the following model: postN app (x, y, z) ⇐⇒ x + y = z, postN nrev (x, y) ⇐⇒ x = y, postN app3 (x, y, z, u) ⇐⇒ x + y + z = u. The actual computation is performed on the set of nonnegative, infinite precision rational numbers, using a fixpoint calculator based on PPL, the Parma Polyhedra Library (Bagnara et al. 2002), and the standard widening (Cousot and Halbwachs 1978; Halbwachs 1979). In our example the least model is found. In general, however, only a less precise model can be determined. Step 3: Computing a numeric level mapping. The information provided by the numerical model is crucial to compute a level mapping | · |N . Let p be an n-ary predicate symbol in the CLP(N ) program. The level mapping associates to p a function fp : Nn → N that is guaranteed to decrease when going from the head of the clause to each recursive call(s), if any, for each clause defining p. For example, a level mapping |·|N such that |nrevN (x, y)|N = x intuitively means: for each ground instance1 of each recursive clause defining nrevN , the first argument decreases when going from the head of the clause to the recursive call (since 1 + e + x > x for each e, x ∈ N). Since no clause defining app3N is recursive, the level mapping can be defined so that |app3(x, y, z, t)|N = 0. The level mapping computed for our example is defined by: |app(x, y, z)|N = min(x, z), |nrev(x, y)|N = x, |app3(x, y, z, u)|N = 0. This is obtained by means of an improvement of the technique by K. Sohn and A. Van Gelder for the automatic generation of linear level mappings. Their algorithm, which is based on linear programming, is complete in the sense that it will 1 That is, where natural numbers have replaced variable symbols. cTI: A Termination Inference Tool for Prolog 5 always provide a linear level mapping if one exists (Sohn and Van Gelder 1991). Our extension, which is described in (Mesnard and Neumerkel 2001), consists in first computing a constraint over the coefficients of a generic linear level mapping (step 3a). Then we generate a concrete level mapping (step 3b). Notice that for a multi-directional predicate (such as app/3) we may get multiple linear level mappings. These are combined, with the min operator, into one non-linear level mapping. In contrast with the well-known standard framework of acceptability, the decrease of the level mapping has to be shown only for predicates belonging to the same strongly connected component (SCC) of the call graph. Step 5 below will ensure that the other calls to predicates from lower SCC’s do left terminate. The advantage of this approach is twofold: first, the computation of a level mapping, being SCCbased, is modular. Secondly, the expressive power of linear level mappings with respect to termination is much higher than in the acceptability case. Step 4: From CLP(N ) to CLP(B). From the CLP(N ) program P N a CLP(B) program, P B , is obtained by mapping each natural number to 1 (true), each variable symbol to itself, and addition to logical conjunction. nrevB (1, 1). app3B (x, y, z, u) ← appB (1, x, x). appB (x, y, v), appB (1 ∧ e ∧ x, y, 1 ∧ e ∧ z) ← nrevB (1 ∧ e ∧ x, y) ← nrevB (x, z), appB (v, z, u). appB (x, y, z). appB (z, 1 ∧ e, y). The purpose of P B is the one of capturing boundedness dependencies within P N or, equivalently, rigidity dependencies within the original program.2 A model for P B is then computed and a boolean level mapping | · |B is obtained from the numerical level mapping computed in Step 3. In order to do that, the translation scheme outlined above is augmented with the association of the logical disjunction x ∨ y to min(x, y): this means that min(x, y) is a bounded quantity if x or y or both are bounded. Here is what we obtain for the example program: postB app (x, y, z) ⇐⇒ (x ∧ y) ↔ z, postB nrev (x, y) ⇐⇒ x ↔ y, postB app3 (x, y, z, u) ⇐⇒ (x ∧ y ∧ z) ↔ u, |app(x, y, z)|B = x ∨ z, |nrev(x, y)|B = x, |app3(x, y, z, u)|B = 1. For instance, as we use the term-size norm, this model tells us that for any computed answer θ to a call nrev(x, y), xθ is ground if and only if yθ is ground. Step 5: Computing boolean termination conditions. The information obtained from P B for each program point is combined with the level mapping by means of the following boolean µ-calculus formulæ, whose solution gives the desired boolean termination conditions. 2 A term t is rigid with respect to a symbolic norm k · k if and only if its measure is invariant by instantiation, i.e., ktk = ktθk for any substitution θ. 6 F. Mesnard and R. Bagnara preapp = νT . λ(x, y, z) .   app(x, y, z) B    ∀e, x′ , z ′ : x ↔ (1 ∧ e ∧ x′ ) ∧ z ↔ (1 ∧ e ∧ z ′ ) → T (x′ , y, z ′ ) prenrev = νT . λ(x, y) .   |nrev(x, y)|B     ∀e, x′ , z : x ↔ (1 ∧ e ∧ x′ ) → T (x′ , z)     ∀e, x′ , z : x ↔ (1 ∧ e ∧ x′ ) ∧ postB (x′ , z) → pre nrev preapp3 = νT . λ(x, y, z, u) .  B    app3(x, y, z, u) ∀v : 1 → preapp (x, y, v)   ∀v : postB (x, y, v) → pre app app (z, 1 (1) (2) ∧ e, y) (3) app (v, z, u) Here is the intuition behind such boolean µ-calculus formulæ. Consider the nrev/2 predicate. Its unit clause is taken into account in the computation of the numeric and the boolean model. For computing the boolean termination condition prenrev , we consider the clause nrevB (x, y) ← [x ↔ (1 ∧ e ∧ x′ )], nrevB (x′ , z), appB (z, 1 ∧ e, y). We are looking for a boolean relation T (x, y) satisfying the following conditions: • for each (x, y) in T , the level mapping has to be bounded, which leads to condition (1) above; • the recursive call to nrev/2 has to terminate, hence condition (2); • for any state resulting from the evaluation of the first call, the subsequent call to app/3 has to terminate, giving condition (3); • finally, we are interested in the weakest solution for T , hence the boolean termination condition is defined as a greatest fixpoint:  prenrev = νT . λ(x, y) . (1) ∧ (2) ∧ (3) . Solving the equations for our example gives: preapp (x, y, z) = x ∨ z, prenrev (x, y) = x, preapp3 (x, y, z, u) = (x ∧ y) ∨ (x ∧ u). The greatest fixpoint is evaluated with the boolean µ-solver described in (Colin et al. 1997), which computes on the domain Pos of positive boolean formulæ (Armstrong et al. 1998) and is based on the boolean solver of SICStus Prolog. Step 6: Back to Prolog. In the final step of the analysis, the boolean termination conditions are lifted to termination conditions with the following interpretation, where the c’s are CLP(H) constraints: cTI: A Termination Inference Tool for Prolog 7 • each goal ‘?- c, app(X,Y,Z).’ left-terminates if X or Z are ground in c; • each goal ‘?- c, nrev(X,Y).’ left-terminates if X is ground in c; • each goal ‘?- c, app3(X,Y,Z,U).’ left-terminates if X and Y are ground in c or X and U are ground in c. 3 Using cTI Once compiled and installed, cTI is invoked with the command ‘cti source’, where the program in ‘source’ is assumed to be an ISO-Prolog program. The user may then control the behavior of cTI with some options. We describe the main ones. ‘-p file’ By default, undefined predicates are assumed to fail. The user may enrich or redefine the set of built-ins recognized by the system, by specifying ‘-p file’ on the command line. This has the effect of importing the predicates whose numerical model, boolean model, and termination condition are given in ‘file’. As predicates imported that way cannot be redefined in the analyzed program, this scheme provides a way to overcome potential weaknesses of the analysis. ‘-t timeout in ms’ The analysis steps 2, 3a, 3b, 4, and 5 described in Section 2 all include potentially expensive computations. Because of this, for each such step, the computation concerning each SCC is subject to a timeout, whose default value is 2 seconds. The ‘-t’ option allows the user to modify this value. ‘-n N’ For the computation of the numeric model (step 2), a widening is used after n iterations of the approximate fixpoint iteration. The default value for n is 1. The user may also modify a program to give specific information for selected program points. We illustrate this facility by means of examples; the precise syntax is given in the cTI’s documentation. One may specify that particular program variables will only be bound to non-negative integers and that the analyzer should take into account some constraints involving them. For instance, cTI does not detect that the following program terminates: p1(N) :- N > 0, M is N-1, p1(M). p1(N) :- N > 1, A is N>>1, Z is N-A, p1(A), p1(Z). where the predefined arithmetic functor ‘>>/2’ is the bitwise arithmetic right shift. On the other hand, cTI is able to show that p2(N) terminates: p2(N) :- cti:{N > 0, M = N-1}, p2(M). p2(N) :- cti:{N > 1, 2*A =< N, N =< 2*A+1, Z = N-A}, p2(A), p2(Z). Finally, at any program point, the user can add linear inter-argument relations or groundness relations that the analyzer will take for granted. The system can thus prove the termination of the goal ‘?- top.’ where the predicate top/0 is defined by the program given in Section 2 augmented with the following clause, where the term-size of L1 is declared to be less than 10 and L2 is declared to be ground: top :- cti:{n(L1) < 10},app(L1,Zs,L2),cti:{b(L2)},app(Xs,Ys,L2). 8 F. Mesnard and R. Bagnara While such programs are no longer ISO-Prolog programs, the annotations can be automatically removed so as to obtain the original programs back. The assertion language currently used in cTI is only experimental, and future versions of the system may be based on the language defined in (Hermenegildo et al. 2000). 4 Experimental Evaluation Unless otherwise specified, the experiments we present here were all conducted with the option (see Section 3) -p predef_for_compatibility.pl, which ensures that non-ISO built-ins used in the benchmarks (several of which are written in a nonISO dialect of Prolog) are predefined. This experimental evaluation was done on a GNU/Linux system with an Intel i686 CPU clocked at 2.4 GHz, 512 Mb of RAM, running the Linux kernel version 2.4, SICStus Prolog 3.10.0 (28.3 MLips), PPL version 0.5, and cTI version 1.0. Standard programs from the termination literature. Table 1 presents timings and results of cTI on some standard LP termination benchmarks. The columns are labeled as follows: program: the name of the analyzed program (the asterisk near a name means that we had to use one of the options that allow to tune the behavior of cTI); top-level predicate: the predicate of interest; checked: the class of queries checked by the analyzers of (Decorte et al. 1999; Lindenstrauss and Sagiv 1997; Speirs et al. 1997); result: the best result among those reported in (Decorte et al. 1999; Lindenstrauss and Sagiv 1997; Speirs et al. 1997) (where, of course, ‘yes, the program terminates’ is better than ‘no, don’t know ’); inferred: the termination condition inferred by cTI (1 means that any call to the predicate terminates, 0 means that cTI could not find a terminating mode for that predicate); time: the running time, in seconds, for cTI to infer the termination conditions. For all the examples presented in Table 1, our analyzer is able to infer a class of terminating queries at least as large than the one checked by the analyzers of (Decorte et al. 1999; Lindenstrauss and Sagiv 1997; Speirs et al. 1997) (although we manually tuned cTI three times). We point out that TermiLog (Lindenstrauss and Sagiv 1997) and TerminWeb (Codish and Taboch 1999) are sometimes able to prove termination whereas cTI is not and vice versa. Standard programs from the abstract interpretation literature. Table 2 presents timings of cTI using some standard benchmarks3 from the LP program analysis community. We have chosen eleven middle-sized, well-known logic programs. All the programs are taken from (Bueno et al. 1994) except credit and plan. The first column of Table 2 gives the name of the analyzed program and the second one 3 These have been collected by N. Lindenstrauss, see www.cs.huji.ac.il/~naomil. cTI: A Termination Inference Tool for Prolog 9 Table 1. De Schreye’s, Apt’s, and Plümer’s programs. Others program top-level predicate permute duplicate sum merge dis-con reverse append list fold lte map member mergesort mergesort* mergesort ap mergesort ap* naive rev ordered overlap permutation quicksort select subset sum pl2.3.1 pl3.5.6 pl3.5.6a pl4.0.1 pl4.5.2 pl4.5.3a pl5.2.2 pl7.2.9 pl7.6.2a pl7.6.2b pl7.6.2c pl8.3.1 pl8.3.1a pl8.4.1 pl8.4.2 permute(x, y) duplicate(x, y) sum(x, y, z) merge(x, y, z) dis(x) reverse(x, y, z) append(x, y, z) list(x) fold(x, y, z) goal map(x, y) member(x, y) mergesort(x, y) mergesort(x, y) mergesort ap(x, y, z) mergesort ap(x, y, z) naive rev(x, y) ordered(x) overlap(x, y) permutation(x, y) quicksort(x, y) select(x, y, z) subset(x, y) sum(x, y, z) p(x, y) p(x) p(x) append3(x,y,z,v) s(x, y) p(x) turing(x, y, z, t) mult(x, y, z) reach(x, y, z) reach(x, y, z, t) reach(x, y, z, t) minsort(x, y) minsort(x, y) even(x) e(x, y) cTI checked result inferred x x x∧y x∧y x x∧z x∧y x x∧y 1 x y x x x x x x x∧y x x y x∧y z x 1 1 x∧y∧z x x x∧y∧z x∧y x∧y∧z x∧y∧z∧t x∧y∧z∧t x x x x yes yes yes yes yes yes yes yes yes yes yes yes no no yes yes yes yes yes yes yes yes yes yes no no yes yes no no no yes no no yes no yes yes yes x x∨y x∨y∨z (x ∧ y) ∨ z x x x∨z x y 1 x∨y y 0 x z x∨z x x x∧y x x y∨z x∧y y∨z 0 x x (x ∧ y) ∨ (x ∧ v) 0 0 0 x∧y 0 0 z∧t x∧y x x x time (s) 0.03 0.02 0.03 0.03 0.03 0.02 0.02 0.01 0.02 0.02 0.02 0.01 0.06 0.07 0.11 0.11 0.03 0.01 0.01 0.03 0.06 0.01 0.02 0.02 0.01 0.01 0.01 0.02 0.03 0.01 0.11 0.02 0.02 0.03 0.04 0.04 0.04 0.02 0.07 gives the number of its clauses (before any program transformation takes place). The following six columns indicate the running times (minimum execution times over ten runs), in seconds, for computing: MPN : a numeric model (step 2); Cµ : the constraint over the coefficients of a generic linear level mapping (step 3a); µ: the concrete level mapping (step 3b); MPB : a boolean model (step 4); 10 F. Mesnard and R. Bagnara Table 2. Running times for middle-sized programs. analysis times (s) program ann bid boyer browse credit peephole plan qplan rdtok read warplan clauses 177 50 136 30 57 134 29 148 55 88 101 MPN MPB Cµ µ 0.17 0.03 0.07 0.05 0.02 0.18 0.02 0.20 0.13 0.26 0.10 0.48 0.04 0.06 0.12 0.03 0.56 0.03 0.52 0.39 1.00 0.25 0.08 0.02 0.02 0.03 0.02 0.03 0.01 0.12 0.03 0.04 0.01 0.17 0.02 0.08 0.04 0.02 0.20 0.02 0.18 0.07 0.31 0.08 18% 50% 6% 17% TC 0.06 0.02 0.02 0.01 0.01 0.06 0.02 0.07 0.02 0.08 0.02 total Q% 1.00 49% 0.14 100% 0.30 85% 0.26 60% 0.11 100% 1.08 94% 0.11 100% 1.13 68% 0.65 44% 1.72 52% 0.49 33% 6% 100% average % of time for each analysis phase TC: the boolean termination conditions (step 5). The next column reports the total runtime in seconds while the last column, labeled ‘Q%’, expresses the quality of the analysis, computed as the ratio of the number of user-defined predicates that have a non-empty termination condition over the total number of user-defined predicates (the result of an analysis presents all the user-defined predicates together with their corresponding termination conditions). We note that cTI can prove that bid, credit, and plan are left-terminating: every ground atom left-terminates. For any such program P , TP has only one fixpoint (Apt 1997, Theorem 8.13), which may help proving its partial correctness. Moreover, as the ground semantics of such a program is decidable, Prolog is its own decision procedure, which does help testing and validating the program. On the other hand, when the quality of the analysis is less than 100%, it means that there exists at least one SCC where the inferred termination condition is 0. Let us call such SCC’s failed SCC’s. They are clearly identified, which may help the programmer. Here are some reasons why cTI may fail: potential non-termination, poor numeric model, non-existence of a linear level mapping for a predicate with respect to the model, inadequate norm. Also, the analysis of the SCC’s which depend on a failed SCC is likely to fail, but this does not prevent cTI from analyzing other parts of the call graph. Some larger programs. Finally, we have tested cTI on the following programs: • chat is a parser written by F.C.N. Pereira and D.H.D. Warren; • lptp is an interactive theorem prover for Prolog written by R. Stärk (Stärk 1998); • pl2wam is the compiler from Prolog to WAM of GNU-Prolog 1.1.2 developed by D. Diaz (Diaz and Codognet 2000); cTI: A Termination Inference Tool for Prolog 11 Table 3. Running times for larger programs. analysis times (s) program chat lptp pl2wam slice symbolic1 clauses 515 1298 1190 952 923 MPN 3.89 3.99 2.12 2.20 1.49 Cµ 2.78 14.10 2.37 10.46 0.67 µ 2.84 1.84 0.99 0.13 0.04 MPB TC total timeouts Q% 0.35 1.65 1.29 0.93 0.29 12.80 25.10 9.22 16.20 3.47 1/1/1/0/0 0/1/0/0/0 0/0/0/0/0 0/3/0/0/0 0/0/0/0/0 71% 67% 64% 55% 58% 2.85 2.88 2.00 2.08 0.61 • slice is a multi-language interpreter developed by R. Bagnara and A. Riaudo; • symbolic1 seems to be a simulator for a Prolog machine. We do not know the origin of this file. The results of the analysis are given in Table 3. As explained in the previous section, we set up a timeout of 2 seconds per SCC for computing a CLP(N ) model, the constraints defining level mappings, a CLP(B) model, and the termination condition. So we have a limit of 10 seconds of CPU time per SCC. The last but one column in Table 3 summarizes the number of timeouts for steps 2/3a/3b/4/5, respectively. 5 Related Work The compiler of the Mercury programming language (Somogyi et al. 1996) includes a termination checker, described in (Speirs et al. 1997). The speed of the analyzer is quite impressive. We see two reasons for this. First, the termination checker is written in Mercury itself. Second, and most importantly, the analyzer takes high profit of the mode informations that are part of the text of the program being checked. On the other hand, while the running times of cTI are bigger, termination inference is a more general problem than termination checking: in the worst case, an exponential number of termination checks are needed to simulate termination inference. TALP (Arts and Zantema 1996) is an automatic tool that transforms a wellmoded logic program (see, e.g., (Apt 1997)) into a term rewriting system such that termination of the latter implies termination of the former. The generated term rewriting system is then proved terminating by the CiME tool (http://cime.lri.fr/). The system seems quite powerful for this class of logic programs. (Genaim and Codish 2001) made recently a link between backward analysis (King and Lu 2002) and termination analysis, which leads to termination inference. Although they used a completely different scheme for computing level mappings, the results of the analysis on the programs described in Tables 1 and 2 were rather similar, both in time and quality, to previous versions of cTI that rely on the rational linear solver of SICStus Prolog. Thanks to the PPL, cTI is now significantly faster (speed-ups from a factor of two to more than an order of magnitude have been observed). The latest version of TerminWeb emphasizes termination analysis of typed logic programs. Termination of logic programming where numerical computations are taken into account are studied in (Serebrenik and De Schreye 2001; Serebrenik and De Schreye 2002). 12 F. Mesnard and R. Bagnara The authors present some advanced techniques for explicitly dealing with integers and floating point numbers computations. The size-change termination principle has been proposed in (Lee et al. 2001) for deciding termination of first-order functional programs. The resulting analysis is close to the TermiLog approach (Lindenstrauss and Sagiv 1997) and the authors establish its intrinsic complexity. Finally, we point out that the system Ciao-Prolog (Bueno et al. 1997) adopts another approach for termination, based on complexity analysis (Debray et al. 1994). 6 Conclusion We have presented cTI, the first bottom-up left-termination inference tool for ISOProlog, and its experimental evaluation over standard termination benchmarks as well as middle-sized and larger logic programs. Running cTI on large programs shows that the approach scales up satisfactorily. We believe that, thanks to the Parma Polyhedra Library, cTI is today the fastest and most robust termination inference tool for logic programs. When a SCC is too large, computations relying on projection may become too expensive. So we have added for each computation which may be too costly a timeout and if necessary we are able to return a value which does not destroy the correctness of the analysis, although the quality of the inference is obviously weaker. It allows cTI to keep on analyzing the program. As a side effect, the running time of cTI is linear with respect to the number of SCC’s in the call graph. Finally, one can observe that the termination conditions computed in Section 2 are actually optimal with respect to the language used for describing classes of queries. Can one prove such properties automatically? (Mesnard et al. 2002) presents a first step in this direction. Acknowledgments. We would like to thank Ulrich Neumerkel for numerous discussions we had on termination inference and for the help he provided while debugging cTI. Thanks also to the readers of a previous version of this paper for their comments. Availability. cTI is distributed under the GNU General Public License. The analyzer, together with the programs analyzed for benchmarking, are available from cTI’s web site: http://www.cs.unipr.it/cTI. References Apt, K. and Pellegrini, A. 1994. On the occur-check-free Prolog programs. ACM Transactions on Programming Languages and Systems 16, 3, 687–726. Apt, K. R. 1997. From Logic Programming to Prolog. International Series in Computer Science. Prentice Hall. Apt, K. R. and Pedreschi, D. 1990. Studies in pure Prolog: Termination. In Computational Logic: Symposium Proceedings, J. W. Lloyd, Ed. ESPRIT Basic Research Series. Springer-Verlag, Berlin, Brussels, Belgium, 150–176. cTI: A Termination Inference Tool for Prolog 13 Apt, K. R. and Pedreschi, D. 1993. Reasoning about termination of pure Prolog programs. Information and Computation 106, 1, 109–157. Armstrong, T., Marriott, K., Schachte, P., and Søndergaard, H. 1998. Two classes of Boolean functions for dependency analysis. Science of Computer Programming 31, 1, 3–45. Arts, T. and Zantema, H. 1996. Termination of logic programs using semantic unification. In Logic Program Synthesis and Transformation: Proceedings of the 5th International Workshop, M. Proietti, Ed. Lecture Notes in Computer Science, vol. 1048. Springer-Verlag, Berlin, Utrecht, The Netherlands, 219–233. Bagnara, R., Gori, R., Hill, P. M., and Zaffanella, E. 2001a. Finite-tree analysis for constraint logic-based languages. See Cousot (2001), 165–184. Bagnara, R., Ricci, E., Zaffanella, E., and Hill, P. M. 2002. Possibly not closed convex polyhedra and the Parma Polyhedra Library. In Static Analysis: Proceedings of the 9th International Symposium, M. V. Hermenegildo and G. Puebla, Eds. Lecture Notes in Computer Science, vol. 2477. Springer-Verlag, Berlin, Madrid, Spain, 213–229. Bagnara, R., Zaffanella, E., Gori, R., and Hill, P. M. 2001b. Boolean functions for finite-tree dependencies. See Nieuwenhuis and Voronkov (2001), 579–594. Bueno, F., Cabeza, D., Carro, M., Hermenegildo, M., López-Garcı́a, P., and Puebla, G. 1997. The Ciao Prolog system. Reference manual. Tech. Rep. CLIP3/97.1, School of Computer Science, Technical University of Madrid (UPM). Available from http://www.clip.dia.fi.upm.es/. Bueno, F., de la Banda, M. G., and Hermenegildo, M. V. 1994. Effectiveness of global analysis in strict independence-based automatic program parallelization. In Logic Programming: Proceedings of the 1994 International Symposium, M. Bruynooghe, Ed. MIT Press Series in Logic Programming. The MIT Press, Ithaca, NY, USA, 253–268. Clarke, E. M., Grumberg, O., and Peled, D. A. 2000. Model Checking. The MIT Press. Codish, M. and Taboch, C. 1999. A semantic basis for the termination analysis of logic programs. Journal of Logic Programming 41, 1, 103–123. Colin, S., Mesnard, F., and Rauzy, A. 1997. Constraint logic programming and mucalculus. In Proceedings of the ERCIM/COMPULOG Workshop on Constraints. Schloss Hagenberg, Austria. Cousot, P., Ed. 2001. Static Analysis: 8th International Symposium, SAS 2001. Lecture Notes in Computer Science, vol. 2126. Springer-Verlag, Berlin, Paris, France. Cousot, P. and Cousot, R. 1992. Abstract interpretation and applications to logic programs. Journal of Logic Programming 13, 2&3, 103–179. Cousot, P. and Halbwachs, N. 1978. Automatic discovery of linear restraints among variables of a program. In Conference Record of the Fifth Annual ACM Symposium on Principles of Programming Languages. ACM Press, Tucson, Arizona, 84–96. Crnogorac, L., Kelly, A. D., and Søndergaard, H. 1996. A comparison of three occur-check analysers. In Static Analysis: Proceedings of the 3rd International Symposium, R. Cousot and D. A. Schmidt, Eds. Lecture Notes in Computer Science, vol. 1145. Springer-Verlag, Berlin, Aachen, Germany, 159–173. De Schreye, D. and Decorte, S. 1994. Termination of logic programs: The neverending story. Journal of Logic Programming 19 & 20, 199–260. Debray, S. K., López-Garcı́a, P., Hermenegildo, M. V., and Lin, N.-W. 1994. Estimating the computational cost of logic programs. In Static Analysis: Proceedings of the 1st International Symposium, B. Le Charlier, Ed. Lecture Notes in Computer Science, vol. 864. Springer-Verlag, Berlin, Namur, Belgium, 255–265. 14 F. Mesnard and R. Bagnara Decorte, S. 1997. Enhancing the power of termination analysis of logic programs through types and constraints. Ph.D. thesis, Department of Computer Science, K. U. Leuven, Leuven, Belgium. Decorte, S., De Schreye, D., and Vandecasteele, H. 1999. Constraint-based termination analysis of logic programs. ACM Transactions on Programming Languages and Systems 21, 6, 1137–1195. Deransart, P., Ed-Dbali, A., and Cervoni, L. 1996. Prolog: The Standard. SpringerVerlag, Berlin. Deransart, P., Ferrand, G., and Téguia:, M. 1991. NSTO programs (Not Subject to Occur-Check). In Logic Programming: Proceedings of the 1991 International Symposium, V. A. Saraswat and K. Ueda, Eds. MIT Press Series in Logic Programming. The MIT Press, San Diego, USA, 533–547. Diaz, D. and Codognet, P. 2000. The GNU Prolog system and its implementation. In Proceedings of the 2000 ACM Symposium on Applied Computing, J. Carrol, E. Damiani, H. Haddad, and D. Oppenheim, Eds. Vol. 2. Como, Italy. Francez, N., Grumberg, O., Katz, S., and Pnueli, A. 1985. Proving termination of Prolog programs. In Proceedings of Third Workshop on Logics of Programs, R. Parikh, Ed. Lecture Notes in Computer Science, vol. 193. Springer-Verlag, Berlin, Brooklyn, New York, 89–105. Genaim, S. and Codish, M. 2001. Inferring termination conditions for logic programs using backwards analysis. See Nieuwenhuis and Voronkov (2001), 685–694. Halbwachs, N. 1979. Thèse de 3ème cycle d’informatique. Ph.D. thesis, Université scientifique et médicale de Grenoble, Grenoble, France. Hermenegildo, M. V. and Puebla, G., Eds. 2002. Static Analysis: 9th International Symposium, SAS 2002. Lecture Notes in Computer Science, vol. 2477. Springer-Verlag, Berlin, Madrid, Spain. Hermenegildo, M. V., Puebla, G., and Bueno, F. 2000. An assertion language for constraint logic programs. In Analysis and Visualization Tools for Constraint Programming, P. Deransart, M. V. Hermenegildo, and J. Maluszyński, Eds. Lecture Notes in Computer Science, vol. 1870. Springer-Verlag, Berlin, 23–61. ISO/IEC. 1995. ISO/IEC 13211-1: 1995 Information technology — Programming languages — Prolog — Part 1: General core. International Standard Organization. King, A. and Lu, L. 2002. A backward analysis for constraint logic programs. Theory and Practice of Logic Programming 2, 4–5, 517–547. Lee, C. S., Jones, N. D., and Ben-Amram, A. M. 2001. The size-change principle for program termination. In Proceedings of the 28th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL-01), C. Norris and J. J. B. Fenwick, Eds. ACM SIGPLAN Notices, vol. 36. Association for Computing Machinery, London, UK, 81–92. Levi, G. and Scozzari, F. 1995. Contributions to a theory of txistential termination for definite logic programs. In Proceedings of the “1995 Joint Conference on Declarative Programming (GULP-PRODE’95)”, M. Alpuente and M. I. Sessa, Eds. Marina di Vietri, Italy, 631–642. Lindenstrauss, N. and Sagiv, Y. 1997. Automatic termination analysis of logic programs. In Logic Programming: Proceedings of the Fourteenth International Conference on Logic Programming, L. Naish, Ed. MIT Press Series in Logic Programming. The MIT Press, Leuven, Belgium, 63–77. Lloyd, L. W. 1987. Foundations of Logic Programming, second ed. Springer-Verlag, Berlin. cTI: A Termination Inference Tool for Prolog 15 Marchiori, M. 1996. Proving existential termination of normal logic programs. In Proceedings of the Fifth International Conference on Algebraic Methodology and Software Technology (AMAST’96), M. Wirsing and M. Nivat, Eds. Lecture Notes in Computer Science, vol. 1101. Springer-Verlag, Berlin, Munich, Germany, 375–390. Marriott, K. and Stuckey, P. J. 1998. Programming with Constraints: An Introduction. The MIT Press. Mesnard, F. 1996. Inferring left-terminating classes of queries for constraint logic programs by means of approximations. In Logic Programming: Proceedings of the Joint International Conference and Symposium on Logic Programming, M. J. Maher, Ed. MIT Press Series in Logic Programming. The MIT Press, Bonn, Germany, 7–21. Mesnard, F. and Neumerkel, U. 2001. Applying static analysis techniques for inferring termination conditions of logic programs. See Cousot (2001), 93–110. Mesnard, F., Payet, E., and Neumerkel, U. 2002. Detecting optimal termination conditions of logic programs. See Hermenegildo and Puebla (2002), 509–525. Mesnard, F. and Ruggieri, S. 2003. On proving left termination of constraint logic programs. ACM Transactions on Computational Logic 4, 2, 1–26. Nieuwenhuis, R. and Voronkov, A., Eds. 2001. Proceedings of the 8th International Conference on Logic for Programming, Artificial Intelligence and Reasoning (LPAR 2001). Lecture Notes in Artificial Intelligence, vol. 2250. Springer-Verlag, Berlin, Havana, Cuba. Plümer, L. 1990. Terminations Proofs for Logic Programs. Number 446 in Lecture Notes in Computer Science. Springer-Verlag, Berlin. Ruggieri, S. 1999. Verification and validation of logic programs. Ph.D. thesis, Dipartimento di Informatica, Università di Pisa, Pisa, Italy. Printed as Report TD-6/99. Serebrenik, A. and De Schreye, D. 2001. Inference of termination conditions for numerical loops in prolog. See Nieuwenhuis and Voronkov (2001), 654–668. Serebrenik, A. and De Schreye, D. 2002. On termination of logic programs with floating point computations. See Hermenegildo and Puebla (2002), 151–164. Sohn, K. and Van Gelder, A. 1991. Termination detection in logic programs using argument sizes. In Proceedings of the Tenth ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems. Association for Computing Machinery, Denver, Colorado, 216–226. Extended abstract. Somogyi, Z., Henderson, F., and Conway, T. 1996. The execution algorithm of Mercury: An efficient purely declarative logic programming language. Journal of Logic Programming 29, 1–3, 17–64. Søndergaard, H. 1986. An application of abstract interpretation of logic programs: Occur check reduction. In Proceedings of the 1986 European Symposium on Programming, B. Robinet and R. Wilhelm, Eds. Lecture Notes in Computer Science, vol. 213. Springer-Verlag, Berlin, Saarbrücken, Federal Republic of Germany, 327–338. Speirs, C., Somogyi, Z., and Søndergaard, H. 1997. Termination analysis for Mercury. In Static Analysis: Proceedings of the 4th International Symposium, P. Van Hentenryck, Ed. Lecture Notes in Computer Science, vol. 1302. Springer-Verlag, Berlin, Paris, France, 157–171. Stärk, R. F. 1998. The theoretical foundations of LPTP (a Logic Program Theorem Prover). Journal of Logic Programming 36, 3, 241–269. Ullman, J. D. and Van Gelder, A. 1988. Efficient tests for top-down termination of logical rules. Journal of the ACM 35, 2, 345–373. Vasak, T. and Potter, J. 1986. Characterisation of terminating logic programs. In Proceedings of the 1986 Symposium on Logic Programming, R. M. Keller, Ed. IEEE Computer Society Press, Salt Lake City, Utah, 140–147. 16 F. Mesnard and R. Bagnara Verschaetse, K. 1992. Static termination analysis for definite Horn clause programs. Ph.D. thesis, Department of Computer Science, K. U. Leuven, Leuven, Belgium.
2cs.AI
COMMENSURABILITY FOR CERTAIN RIGHT-ANGLED COXETER GROUPS AND GEOMETRIC AMALGAMS OF FREE GROUPS arXiv:1610.06245v2 [math.GR] 5 Oct 2017 PALLAVI DANI, EMILY STARK AND ANNE THOMAS Abstract. We give explicit necessary and sufficient conditions for the abstract commensurability of certain families of 1-ended, hyperbolic groups, namely right-angled Coxeter groups defined by generalized Θ-graphs and cycles of generalized Θ-graphs, and geometric amalgams of free groups whose JSJ graphs are trees of diameter ≤ 4. We also show that if a geometric amalgam of free groups has JSJ graph a tree, then it is commensurable to a right-angled Coxeter group, and give an example of a geometric amalgam of free groups which is not quasi-isometric (hence not commensurable) to any group which is finitely generated by torsion elements. Our proofs involve a new geometric realization of the right-angled Coxeter groups we consider, such that covers corresponding to torsionfree, finite-index subgroups are surface amalgams. 1. Introduction Two groups G and H are (abstractly) commensurable if they contain finite-index subgroups G0 < G and H 0 < H so that G0 and H 0 are abstractly isomorphic. There are related but stronger notions of commensurability for subgroups of a given group. Commensurability in the sense we consider is an equivalence relation on abstract groups which implies quasi-isometric equivalence (for finitely generated groups). In this paper, we give explicit necessary and sufficient conditions for commensurability within certain families of Coxeter groups and amalgams of free groups. Let Γ be a finite simplicial graph with vertex set S. The right-angled Coxeter group WΓ with defining graph Γ has generating set S and relations s2 = 1 for all s ∈ S, and st = ts whenever s, t ∈ S are adjacent vertices. We assume throughout that WΓ is infinite, or equivalently, that Γ is not a complete graph. The graph Γ is 3-convex if every path between a pair of vertices of valence at least three in Γ has at least three edges. For each induced subgraph Λ of Γ, with vertex set A, the corresponding special subgroup of WΓ is the right-angled Coxeter group WΛ = WA = hAi. See Section 2.1 for additional terminology for graphs, and Figure 1.1 for some examples of 3-convex defining graphs. Γ Γ0 Γ00 X X0 Figure 1.1. The graphs Γ, Γ0 , and Γ00 are examples of 3-convex defining graphs, and the spaces X and X 0 are examples of surface amalgams. 1 Geometric amalgams of free groups were introduced by Lafont in [16]. These are fundamental groups of spaces called surface amalgams, which, roughly speaking, consist of surfaces with boundary glued together along their boundary curves. See Section 2.3 for details, and Figure 1.1 for some examples. We now outline our main results and the ideas of their proofs, including some new constructions which may be of independent interest, and discuss previous work on related questions. We defer precise definitions and theorem statements to Section 1.1. In Theorems 1.8 and 1.12, we give explicit necessary and sufficient conditions for commensurability of right-angled Coxeter groups defined by two families of graphs. Theorem 1.8 classifies those defined by 3-convex generalized Θ-graphs (see Definition 1.6 and the graph Γ in Figure 1.1), and Theorem 1.12 classifies those defined by 3-convex cycles of generalized Θ-graphs (see Definition 1.10 and the graphs Γ0 and Γ00 in Figure 1.1). We prove Theorem 1.8 in Section 5, and the necessary and sufficient conditions of Theorem 1.12 in Sections 6 and 7, respectively. Our commensurability criteria are families of equations involving the Euler characteristics of certain special subgroups, and we express these criteria using commensurability of vectors with entries determined by these Euler characteristics. In [6], Crisp and Paoluzzi classified up to commensurability the right-angled Coxeter groups defined by a certain family of three-branch generalized Θ-graphs, and in Remark 1.9 we recover their result using Theorem 1.8. If Γ is a 3-convex generalized Θ-graph and Γ0 a 3-convex cycle of generalized Θ-graphs, we can also determine the commensurability of WΓ and WΓ0 , as explained in Remark 1.14. The results described in the previous two paragraphs fit into a larger program of classifying 1-ended, hyperbolic right-angled Coxeter groups up to commensurability. Since groups that are commensurable are quasi-isometric, a step in this program is provided by Dani and Thomas [7], who considered Bowditch’s JSJ tree [3], a quasi-isometry invariant for 1-ended hyperbolic groups which are not cocompact Fuchsian. If G is such a group then G acts cocompactly on its Bowditch JSJ tree TG with edge stabilizers maximal 2-ended subgroups over which G splits. The quotient graph for the action of G on TG is called the JSJ graph of G, and the induced graph of groups is the JSJ decomposition for G. In Section 2.2, we recall results from [7] that give an explicit “visual” construction of the JSJ decomposition for right-angled Coxeter groups WΓ satisfying the following. Assumption 1.1. The graph Γ has no triangles (WΓ is 2-dimensional); Γ is connected and no vertex or (closed) edge separates Γ into two or more components (WΓ is 1-ended); Γ has no squares (WΓ is hyperbolic); Γ is not a cycle of length ≥ 5 (WΓ is not cocompact Fuchsian); and Γ has a cut pair of vertices {a, b} (WΓ splits over a 2-ended subgroup). Moreover, Theorem 1.3 of [7] says that Bowditch’s tree is a complete quasi-isometry invariant for the family of groups satisfying, in addition, Assumption 1.2. Γ has no induced subgraph which is a subdivided copy of K4 , the complete graph on four vertices. We denote by G the family of graphs satisfying Assumptions 1.1 and 1.2. Generalized Θ-graphs and cycles of generalized Θ-graphs are two infinite families of graphs in G. Thus Theorems 1.8 and 1.12 and Remarks 1.9 and 1.14 provide a finer classification (up to commensurability) within some quasi-isometry classes determined by graphs in G. In Section 9 we discuss the obstructions to extending our results to other families of graphs in G. Our proofs of the necessary conditions in Theorems 1.8 and 1.12 follow the same general strategy used by Crisp–Paoluzzi [6] and Stark [23] on commensurability of certain geometric amalgams of 2 free groups (discussed further below). Given two groups which are commensurable, the first step in both these papers is to consider covers corresponding to isomorphic (torsion-free) finite-index subgroups. In both cases such covers are surface amalgams, and a crucial ingredient is Lafont’s topological rigidity result from [16], which says that any isomorphism between a pair of geometric amalgams of free groups is induced by a homeomorphism of the corresponding surface amalgams. This homeomorphism between the covers is then analyzed to obtain the necessary conditions. The natural spaces to apply this strategy to in our setting are quotients ΣΓ /G where ΣΓ is the Davis complex for WΓ , and G is a torsion-free, finite-index subgroup of WΓ . However, Stark proves in [22] that topological rigidity fails for such quotients, by constructing an example where G and G0 are isomorphic torsion-free, finite-index subgroups of WΓ , but ΣΓ /G and ΣΓ /G0 are not homeomorphic. The graph Γ in this example is a 3-convex cycle of generalized Θ-graphs. In light of the result of [22], in Section 3 we introduce a new geometric realization for right-angled Coxeter groups WΓ with 3-convex Γ ∈ G, by constructing a piecewise hyperbolic orbicomplex OΓ with fundamental group WΓ . The orbicomplex OΓ has underlying space obtained by gluing together right-angled hyperbolic polygons, and each edge of OΓ which is contained in only one such polygon is a reflection edge. It follows that any cover of OΓ corresponding to a torsion-free, finite-index subgroup of WΓ is a surface amalgam (tiled by right-angled polygons). Thus we can apply Lafont’s result to these spaces. With a view to generalizing the commensurability classification, we give the construction of the orbicomplex OΓ for all 3-convex graphs in G, not just for generalized Θ-graphs and cycles of generalized Θ-graphs. Our construction of OΓ makes heavy use of the JSJ decomposition from [7]. We restrict to 3-convex defining graphs in this paper so that the correspondence between Γ and the JSJ decomposition of WΓ is more straightforward than the general case in [7]. A reference for orbifolds is Kapovich [15]; we view orbicomplexes as complexes of groups, and use the theory of these from Bridson–Haefliger [4]. The proofs of the necessary conditions in Theorems 1.8 and 1.12 then involve a careful analysis of the homeomorphic finite covers guaranteed by topological rigidity. For generalized Θ-graphs, we adapt Stark’s proof in [23] to the setting where the orbicomplexes considered do not have the same Euler characteristic. The proof of the necessary conditions for cycles of generalized Θgraphs is considerably more difficult. Here, the groups WΓ and WΓ0 are fundamental groups of orbicomplexes OΓ and OΓ0 with “central” orbifolds A and A0 that have many branching edges along which other orbifolds are attached (see the lower right of Figure 4.4). A key ingredient in the proof of Theorem 1.12 is a careful argument to show that the homeomorphism f : X → X 0 between finite covers π : X → OΓ and π 0 : X 0 → OΓ0 guaranteed by Lafont’s topological rigidity result can be modified so that either f (π −1 (A)) = π 0−1 (A0 ), or f (π −1 (A)) has no surfaces in common with π 0−1 (A0 ). To prove the sufficient conditions in Theorem 1.8 and 1.12, given any pair of groups WΓ and WΓ0 which satisfy the putative sufficient conditions, we construct finite-sheeted covers of OΓ and OΓ0 with isomorphic fundamental groups. It follows that WΓ and WΓ0 have isomorphic finite-index subgroups. In the case of generalized Θ-graphs, these finite-sheeted covers are orbicomplexes whose construction is an immediate generalization of Crisp and Paoluzzi’s. The finite covers in the case of cycles of generalized Θ-graphs are homeomorphic surface amalgams, and their construction is quite delicate. In order to explain our constructions of surface amalgams covering OΓ , we introduce the notion of a half-covering of graphs (see Section 4.1). The idea is that if a surface amalgam Y is a finitesheeted cover of another surface amalgam X , or of an orbicomplex OΓ , then the induced map of JSJ graphs is a half-covering. For the proofs of the sufficient conditions in Theorem 1.12, we first 3 construct the JSJ graphs for the homeomorphic finite-sheeted covers, together with the associated half-coverings. We then construct suitable surfaces and glue them together along their boundaries, guided by adjacencies in the JSJ graphs, to obtain a surface amalgam covering OΓ . One construction in our proof of sufficient conditions for Theorem 1.12 may be of independent interest. In Section 4, given any orbicomplex OΓ (with Γ ∈ G and Γ being 3-convex), we construct a particularly nice degree 16 cover X which is a surface amalgam. Hence WΓ has an index 16 subgroup which is a geometric amalgam of free groups (this result is stated below as Theorem 1.15). Our construction of X uses tilings of surfaces similar to those in Futer and Thomas [10, Section 6.1]. Given two groups WΓ and WΓ0 which satisfy the sufficient conditions from Theorem 1.12, we first pass to our degree 16 covers X and X 0 of OΓ and OΓ0 , then construct homeomorphic finite-sheeted covers of X and X 0 . Theorem 1.16, proved in Section 8, says that for all geometric amalgams of free groups G, if the JSJ graph of G is a tree, then G is commensurable to some WΓ (with Γ ∈ G). In Section 8 we also give an example of a geometric amalgam of free groups which is not quasi-isometric (and hence not commensurable) to any WΓ , or indeed to any group finitely generated by torsion elements. This leads to the question: Question 1.3. Which geometric amalgams of free groups are commensurable to right-angled Coxeter groups (with defining graphs in G)? The proof of Theorem 1.16 uses the torsion-free degree 16 cover of the orbicomplex OΓ that we construct in Section 4. We show that any surface amalgam whose JSJ graph is a tree admits a finite-sheeted cover which “looks like” our torsion-free cover. Then we follow our construction in Section 4 backwards to obtain an orbicomplex OΓ as a finite quotient, with fundamental group the right-angled Coxeter group WΓ . As a corollary to Theorems 1.8, 1.12, and 1.16, we obtain the commensurability classification of geometric amalgams of free groups whose JSJ graph is a tree with diameter at most 4 (see Corollary 1.17). This recovers part of Theorem 3.31 of Stark [23], which gives the commensurability classification of fundamental groups of surface amalgams obtained by identifying two closed surfaces along an essential simple closed curve in each. We remark that Malone [18] provides a complete quasi-isometry classification within the class of geometric amalgams of free groups; in particular, he proves that the isomorphism type of Bowditch’s JSJ tree for such a group is a complete quasiisometry invariant, and supplies an algorithm to compute this tree. We conclude this part of the introduction by mentioning some earlier work on commensurability and on the relationship between commensurability and quasi-isometry for groups closely related to the ones we study. We refer to the surveys by Paoluzzi [21] and Walsh [24] for more comprehensive accounts. In [14] and [11], the related notion of wide commensurability is studied for Coxeter groups generated by reflections in the faces of polytopes in n-dimensional real hyperbolic space. Apart from these two papers and [6], we do not know of any other work on commensurability for (infinite non-affine) Coxeter groups. In [5], Crisp investigated commensurability in certain 2-dimensional Artin groups, while Huang has studied quasi-isometry and commensurability in right-angled Artin groups with finite outer automorphism groups in [13] and [12]. Huang’s combined results show that within a class of right-angled Artin groups defined by a few additional conditions on the defining graph, the quasiisometry, commensurability and isomorphism classes are the same. We note that none of the groups we consider is quasi-isometric to a right-angled Artin group, since the groups we consider are all 1-ended and hyperbolic, and a right-angled Artin group is hyperbolic if and only if it is free. 4 The above results of Huang are in contrast to our settings of right-angled Coxeter groups and geometric amalgams of free groups: Theorems 1.8, 1.12, and 1.16 above, together with the descriptions of Bowditch’s JSJ tree from [7] and [18], show that each quasi-isometry class containing one of the groups considered in our theorems contains infinitely many commensurability classes. For geometric amalgams of free groups, Malone [18] and Stark [23] had both given examples to show that commensurability and quasi-isometry are different. 1.1. Definitions and statements of results. We now give precise definitions and statements for our main results. First, we recall the definition of the Euler characteristic of a Coxeter group in the case we will need. A reference for the general definition is pp. 309–310 of [8]. Definition 1.4 (Euler characteristic of WΓ ). Let WΓ be a right-angled Coxeter group with defining graph Γ having vertex set V (Γ) and edge set E(Γ). Assume that Γ is triangle-free. Then the Euler characteristic of WΓ is the rational number χ(WΓ ) given by: card V (Γ) card E(Γ) + . 2 4 We remark that χ(WΓ ) is equal to the Euler characteristic of the orbicomplex induced by the action of WΓ on its Davis complex (we will not need this interpretation). To each group we consider, we associate a collection of vectors involving the Euler characteristics of certain special subgroups. We will use the following notions concerning commensurability of vectors. χ(WΓ ) = 1 − Definition 1.5 (Commensurability of vectors). Let k, ` ≥ 1. Vectors v ∈ Qk and w ∈ Q` are commensurable if k = ` and there exist non-zero integers K and L so that Kv = Lw. Given a nontrivial commensurability class V of vectors in Qk , the minimal integral element of V is the unique vector v0 ∈ V ∩ Zk so that the entries of v0 have greatest common divisor 1 and the first nonzero entry of v0 is a positive integer. Then for each v ∈ V, there is a unique rational R = R(v) so that v = Rv0 . Our first main result, Theorem 1.8, classifies the commensurability classes among right-angled Coxeter groups with defining graphs generalized Θ-graphs, defined as follows. Definition 1.6 (Generalized Θ-graph). For k ≥ 1, let Ψk be the graph with two vertices a and b, each of valence k, and k edges e1 , . . . , ek connecting a and b. For integers 0 ≤ n1 ≤ · · · ≤ nk , the generalized Θ-graph Θ(n1 , n2 , . . . , nk ) is obtained by subdividing the edge ei of Ψk into ni + 1 edges by inserting ni new vertices along ei , for 1 ≤ i ≤ k. For example, the graph Γ in Figure 1.1 is Θ(2, 2, 3, 4). We write βi for the induced subgraph of Θ = Θ(n1 , n2 , . . . , nk ) which was obtained by subdividing the edge ei of Ψk , and call βi the ith branch of Θ. A generalized Θ-graph is 3-convex if and only if ni ≥ 2 for all i, equivalently n1 ≥ 2. Note that if Θ is 3-convex and has at least 3 branches then Θ satisfies Assumptions 1.1 and 1.2. To each generalized Θ-graph, we associate the following Euler characteristic vector. Definition 1.7 (Euler characteristic vector for generalized Θ-graphs). Let WΘ be the right-angled Coxeter group with defining graph Θ = Θ(n1 , . . . , nk ). The Euler characteristic vector of WΘ is v = (χ(Wβ1 ), . . . , χ(Wβk )). Note that, by definition, χ(Wβ1 ) ≥ . . . ≥ χ(Wβk ). For example, if Γ = Θ(2, 2, 3, 4) as in Figure 1.1, then the Euler characteristic vector of WΓ is v = (− 41 , − 14 , − 12 , − 43 ). The commensurability classification for the corresponding right-angled Coxeter groups is then as follows. 5 Theorem 1.8. Let Θ and Θ0 be 3-convex generalized Θ-graphs with at least 3 branches, and let v and v 0 be the Euler characteristic vectors of the right-angled Coxeter groups WΘ and WΘ0 , respectively. Then WΘ and WΘ0 are abstractly commensurable if and only if v and v 0 are commensurable. Remark 1.9. We now explain how Theorem 1.8 can be used to classify all right-angled Coxeter groups defined by generalized Θ-graphs satisfying Assumptions 1.1 (so that the corresponding groups are 1-ended and hyperbolic), with Θ not required to be 3-convex. If Θ = Θ(n1 , . . . , nk ) satisfies every condition in Assumptions 1.1, it is easy to check that k ≥ 3, n1 ≥ 1, and ni ≥ 2 for all i ≥ 2. Theorem 1.8 covers the case n1 ≥ 2, so we just need to discuss the case n1 = 1. Let c be the unique vertex of valence 2 on the first branch of Θ. Then we may form the double of Θ over c, defined by Dc (Θ) = Θ(n2 , n2 , n3 , n3 , . . . , nk , nk ). This is a 3-convex generalized Θ-graph with 2(k − 1) ≥ 4 branches, obtained from Θ by deleting the open star of c, and then identifying two copies of the resulting graph along a and b. The group WDc (Θ) is isomorphic to the kernel of the map WΘ → Z/2Z which sends c to 1 and all other generators to 0. In particular, WDc (Θ) is commensurable to WΘ . This, together with Theorem 1.8, can be used to extend our classification result above to all generalized Θ-graphs satisfying Assumptions 1.1. In this way we recover Crisp and Paoluzzi’s result from [6], as they considered the family of graphs Θ(1, m + 1, n + 1) with m, n ≥ 1. Next, in Theorem 1.12 we consider groups WΓ where Γ is a cycle of generalized Θ-graphs, defined as follows. Definition 1.10 (Cycle of generalized Θ-graphs). Let N ≥ 3 and let r1 , . . . , rN be positive integers so that for each i, at most one of ri and ri+1 (mod N ) is equal to 1. Now for 1 ≤ i ≤ N , let Ψri be the graph from Definition 1.6, with ri edges between ai and bi . Let Ψ be the graph obtained by identifying bi with ai+1 for all i (mod N ). A cycle of N generalized Θ-graphs is a graph obtained from Ψ by (possibly) subdividing edges of Ψ. For example, the graph Γ0 (respectively, Γ00 ) in Figure 1.1 is a cycle of three (respectively, four) generalized Θ-graphs. If Γ is a cycle of generalized Θ-graphs, we denote by Θi the ith generalized Θ-graph of Γ, that is, the subdivided copy of Ψri inside Γ, and we say that Θi is nontrivial if ri > 1. Observe that the condition on the ri guarantees that if Θi is trivial, then Θi−1 and Θi+1 are not, hence the vertices a1 , . . . , aN have valence at least three in Γ. It follows that a cycle of generalized Θ-graphs Γ is 3-convex if and only if each edge of Ψ is subdivided into at least three edges, by inserting at least two vertices. We now use this notation to define Euler characteristic vectors associated to cycles of generalized Θ-graphs. Definition 1.11. [Euler characteristic vectors for cycles of generalized Θ-graphs] Let Γ be a 3convex cycle of N generalized Θ-graphs and let I ⊂ {1, . . . , N } be the set of indices with ri > 1 (so I records the indices i for which Θi is nontrivial). (1) For each i ∈ I, we define the vector vi ∈ Qri to be the Euler characteristic vector of WΘi (from Definition 1.7). Thus if Θi has branches βi1 , . . . , βiri , with ri > 1, then vi = (χ(Wβi1 ), . . . , χ(Wβiri )) =: (χi1 , . . . , χiri ). (2) If there is some r ≥ 2 so that each nontrivial Θi has exactly r branches, we define another vector w associated to Γ as follows. Let A be the union of {a1 , . . . , aN } with the vertex sets of all trivial Θi . Recall from the beginning of this section that WA is the special subgroup 6 of WΓ defined by the subgraph of Γ induced by A. Then w ∈ Qr+1 is the reordering of the vector ! X X X χi1 , χi2 , . . . , χir , χ(WA ) i∈I i∈I i∈I obtained by putting its entries in non-increasing order. For example, the graph Γ0 in Figure 1.1 has I = {1, 2, 3}, with say v1 = v2 = (− 41 , − 21 ) and v3 = (− 41 , − 14 , − 21 ). If Γ00 is as in Figure 1.1, then r = 2, we can choose I = {1, 2, 3}, and then v1 = v2 = (− 14 , − 21 ), v3 = (− 34 , − 32 ), χ(WA ) = − 54 , and w = (− 54 , − 54 , − 52 ). We can now state the commensurability classification of right-angled Coxeter groups defined by 3-convex cycles of generalized Θ-graphs. Theorem 1.12. Let Γ and Γ0 be 3-convex cycles of N and N 0 generalized Θ-graphs, respectively (with N, N 0 ≥ 3). Let ri be the number of branches of the ith generalized Θ-graph Θi in Γ, and let I be the set of indices with ri > 1. Let {vi | i ∈ I} and w be the vectors from Definition 1.11, and let WA be the special subgroup from Definition 1.11(2). Here, {vi | i ∈ I} denotes a multiset of vectors (since the vi may not all be distinct). Let rk0 , Θ0k , I 0 , {vk0 | k ∈ I 0 }, w0 , and WA0 be the corresponding objects for Γ0 . The right-angled Coxeter groups W = WΓ and W 0 = WΓ0 are abstractly commensurable if and only if at least one of (1) or (2) below holds. (1) (a) The set of commensurability classes of the vectors {vi | i ∈ I} coincides with the set of commensurability classes of the vectors {vk0 | k ∈ I 0 }; and (b) given a commensurability class of vectors V, if IV ⊂ I is the set of indices of the vectors in {vi | i ∈ I} ∩ V, and IV0 ⊂ I 0 is the set of indices of the vectors in {vk0 | k ∈ I 0 } ∩ V, then     X X χ(WΘ0k ) . χ(WA0 ) ·  χ(WΘi ) = χ(WA ) ·  0 k∈IV i∈IV (2) There exists r ≥ 2 such that each nontrivial generalized Θ-graph in Γ and in Γ0 has r branches, and: (a) the vectors vi for i ∈ I are contained in a single commensurability class; likewise, the vectors vk0 for k ∈ I 0 are contained in a single commensurability class; and (b) the vectors w and w0 are commensurable. Remark 1.13. For a pair of graphs Γ and Γ0 as in the statement of Theorem 1.12, with WΓ and WΓ0 abstractly commensurable, it is possible that (1) holds but not (2), that (2) holds but not (1), or that both (1) and (2) hold. See Figure 1.2 for examples of this. Remark 1.14. If Γ is a generalized Θ-graph as in the statement of Theorem 1.8, then the result of doubling Γ along a vertex of valence 2 which is adjacent to a vertex of valence at least 3 is a cycle of three generalized Θ-graphs, and by doubling again if necessary, we can obtain a 3-convex cycle of generalized Θ-graphs Γ0 such that the groups WΓ and WΓ0 are commensurable (compare Remark 1.9). Hence we can determine which groups from Theorems 1.8 and 1.12 are commensurable to each other. We then turn our attention to the relationship between right-angled Coxeter groups and geometric amalgams of free groups. We prove the following two theorems. Recall that G is the class of graphs satisfying Assumptions 1.1 and 1.2. 7 Γ1 Γ01 a u u b 4u c v a0 Γ02 Γ2 2v u β0 u v 2u Γ03 Γ3 2v u u 3u 3u u Figure 1.2. We illustrate Remark 1.13. Let u = (− 14 , − 14 , − 12 ) and v = (− 14 , − 12 , − 12 ). By Definition 1.4, the Euler characteristic vectors associated to the nontrivial generalized Θ-subgraphs of the graphs above are the multiples of u and v depicted. Consider first the pair Γ1 , Γ01 . Condition (1a) in Theorem 1.12 clearly holds. Now WA = ha, b, ci < WΓ1 satisfies χ(WA ) = − 12 , and WA0 < WΓ01 is the group defined by the branch β 0 together with the vertex a0 , so χ(WA0 ) = −1. It is easy to check that for the commensurability class of u, both sides of the equation in (1b) are 2 and for that of v, both sides of the equation in (1b) are 54 . Thus condition (1) is satisfied for Γ1 , Γ01 . However, since u and v are not commensurable, (2a) fails for this pair. Now consider the graphs Γ2 , Γ02 . It is clear that condition (2a) holds, but since u and v are not commensurable, (1a) fails. For this pair, we have w = (− 21 , −1, −1, −2) and w0 = (− 34 , − 23 , − 32 , −3), so (2b) holds as well. Thus condition (2) but not condition (1) is satisfied for Γ2 , Γ02 . Finally, we leave it to the reader to check that both (1) and (2) hold for the pair Γ3 , Γ03 . Theorem 1.15. If Γ ∈ G and Γ is 3-convex, then WΓ has an index 16 subgroup which is a geometric amalgam of free groups. Theorem 1.16. If a geometric amalgam of free groups has JSJ graph which is a tree, then it is abstractly commensurable to a right-angled Coxeter group (with defining graph in G). In addition, we show that if the diameter of the JSJ graph of the geometric amalgam of free groups is at most 4, then the defining graph of the corresponding right-angled Coxeter group guaranteed by Theorem 1.16 is either a generalized Θ-graph or a cycle of generalized Θ-graphs. Thus we have the following corollary. Corollary 1.17. Geometric amalgams of free groups whose JSJ graphs are trees of diameter at most 4 can be classified up to abstract commensurability. Acknowledgements. We thank the Forschungsinstitut für Mathematik at ETH Zürich, for hosting Thomas in Spring Semester 2016 and supporting a visit by Dani in June 2016, during which part of this research was carried out. This research was also supported by the National Science Foundation under Grant No. DMS-1440140 while Dani and Thomas were in residence at the Mathematical Sciences Research Institute (MSRI) in Berkeley, California during the Fall 2016 semester. We also thank MSRI for supporting a visit by Stark in Fall 2016. Dani was partially supported by NSF Grant No. DMS-1207868 and this work was supported by a grant from the Simons Foundation (#426932, Pallavi Dani). Thomas was partially supported by an Australian Postdoctoral Fellowship and this research was supported in part by Australian Research Council Grant No. DP110100440. Finally, we thank the anonymous referee for careful and swift reading, which included spotting a mistake in the proof of Theorem 1.16 in an earlier version, and numerous other helpful suggestions. 2. Preliminaries We recall relevant graph theory in Section 2.1. Section 2.2 states the results on JSJ decompositions from [7] that we will need, and establishes some technical lemmas. In Section 2.3 we recall from [16] the definitions of geometric amalgams of free groups and surface amalgams, and 8 state Lafont’s topological rigidity result. Section 2.4 contains some well-known results on coverings of surfaces with boundary, and Section 2.5 recalls the definition of Euler characteristic for orbicomplexes. 2.1. Graph theory. In this paper, we work with several different kinds of graphs. We now recall some graph-theoretic terminology and establish notation. We will mostly consider unoriented graphs, and so refer to these as just graphs. As in [1], a graph Λ consists of a vertex set V (Λ), an edge set E(Λ), and a map  : E(Λ) → V (Λ)2 /C2 from the edge set to the set of unordered pairs of elements of V (Λ). For an edge e, we write (e) = [x, y] = [y, x], where x, y ∈ V (Λ). An edge e is a loop if (e) = [x, x] for some x ∈ V (Λ). If (e) = [x, y] we say that e is incident to x and y, and when x 6= y, that x and y are adjacent vertices. The valence of a vertex x is the number of edges incident to x, counting 2 for each loop e with (e) = [x, x]. A vertex is essential if it has valence at least 3. Identifying Λ with its realization as a 1-dimensional cell complex, a cut pair in Λ is a pair of vertices {x, y} so that Λ \ {x, y} has at least two components, where a component by definition contains at least one vertex. A cut pair {x, y} is essential if x and y are both essential vertices. A reduced path in Λ is a path which does not self-intersect. A branch of Λ is a subgraph of Λ consisting of a (closed) reduced path between a pair of essential vertices, which does not contain any essential vertices in its interior. A graph Λ is bipartite if V (Λ) is the disjoint union of two nonempty subsets V1 (Λ) and V2 (Λ), such that every edge of Λ is incident to exactly one element of V1 (Λ) and exactly one element of V2 (Λ). In this case, we sometimes refer to the vertices in V1 (Λ) as the Type 1 vertices and those in V2 (Λ) as the Type 2 vertices. An oriented graph Λ consists of a vertex set V (Λ), an edge set E(Λ), and maps i : E(Λ) → V (Λ) and t : E(Λ) → V (Λ). For each edge e ∈ E(Λ), we refer to i(e) as the initial vertex of e and t(e) as the terminal vertex of e. Other definitions are similar to the unoriented case. Throughout this paper, we reserve the notation Γ and Γ0 for defining graphs of right-angled Coxeter groups, and we assume throughout that Γ and Γ0 are finite, simplicial graphs in G, that is, they satisfy Assumptions 1.1 and 1.2. (A graph Γ is simplicial if it has no loops and the map  is injective, that is, Γ does not have any multiple edges.) 2.2. JSJ decomposition of right-angled Coxeter groups. In this section we recall the results we will need from [7]. We also establish some technical lemmas needed for our constructions in Section 3, which use similar arguments to those in [7]. 2.2.1. JSJ decomposition. Let W = WΓ . The main result of [7] gives an explicit description of the W -orbits in Bowditch’s JSJ tree T = TWΓ and the stabilizers for this action. From this, we can obtain an explicit description of the canonical graph of groups induced by the action of W on T . Recall from the introduction that the quotient graph Λ = W \T is the JSJ graph of W and the canonical graph of groups over Λ is the JSJ decomposition of W . The JSJ graph Λ is in fact a tree, since W is not an HNN extension (like any Coxeter group, W is generated by torsion elements hence does not surject to Z). The next result follows from Theorem 3.36 of [7], and is illustrated by the examples in Figure 2.1. Corollary 2.1. Let Γ be a finite, simplicial graph satisfying Assumptions 1.1 and 1.2, so that Γ is 3-convex. The JSJ decomposition for W = WΓ is as follows: (1) For each pair of essential vertices {a, b} of Γ such that Γ \ {a, b} has k ≥ 3 components, the JSJ graph Λ has a vertex of valence k, with local group ha, bi. (2) For each set A of vertices of Γ satisfying the following conditions: 9 a s1 Γ= s2 ha, b, si i v1 t1 u1 u2 t 2 u3 a1 v2 v3 v4 b ha, b, ui i Γ0 ha, bi ha, b, ti i a2 a6 = a 5 hai i a4 ha, b, vi i ha1 , a2 i ha2 , a3 i a3 Figure 2.1. Examples of JSJ decompositions given by Corollary 2.1. From left to right, we have a defining graph Γ, the JSJ decomposition of WΓ , a defining graph Γ0 , and the JSJ decomposition of WΓ0 . In the JSJ decompositions, the Type 1 vertices are white, all unlabelled Type 2 vertex groups are the special subgroups corresponding to the branches indicated by color, and all edge groups are equal to the groups on the adjacent Type 1 vertices. (α1 ) elements of A pairwise separate the geometric realization |Γ| of Γ, that is, given a, b ∈ A with a 6= b, the space |Γ| \ {a, b} has at least two components; (α2 ) the set A is maximal among all sets satisfying (α1 ); and (α3 ) hAi is infinite but not 2-ended; the JSJ graph Λ has a vertex of valence ` where ` ≥ 1 is the number of distinct pairs of essential vertices in A which are as in (1). The local group at this vertex is hAi. (3) Every edge of Λ connects some vertex as in (1) above to some vertex as in (2) above. Moreover, a vertex v1 as in (1) and a vertex v2 as in (2) are adjacent if and only if their local groups intersect, with this intersection necessarily ha, bi where {a, b} are as in (1). There will then be an edge with local group ha, bi between these vertices. All maps from edge groups to vertex groups are inclusions. For i = 1, 2, we refer to the vertices of the JSJ graph Λ as in part (i) of Corollary 2.1 as the Type i vertices, and denote these by Vi (Λ). The JSJ graph Λ is a bipartite graph, with vertex set V (Λ) = V1 (Λ) t V2 (Λ). 2.2.2. Technical lemmas. We will use in Section 3 the following technical lemmas related to the JSJ decomposition, where Γ is as in the statement of Corollary 2.1. Lemma 2.2. Suppose a Type 2 vertex in the JSJ decomposition of WΓ has stabilizer hAi. Let L = LA be the number of essential vertices in A. Then L ≥ 2, and L = 2 if and only if A is equal to the set of vertices of a branch of Γ (including its end vertices). Proof. The fact that L ≥ 2 follows from the vertex v having valence ` ≥ 1, and the description given in Corollary 2.1 of the vertices adjacent to v. If A equals the set of vertices of a branch of Γ, then it is clear that L = 2. Conversely, if L = 2 let the two essential vertices of A be a and b. Since hAi is not 2-ended by (α3 ), the set A must also contain a non-essential vertex, say a0 . Let β be the branch of Γ containing a0 . Using Lemma 3.20 of [7], we obtain that A contains all vertices of β, including its end vertices. Since L = 2, these end vertices must be a and b. If A contains a vertex c of Γ which is not in the branch β, then since L = 2, the vertex c must be non-essential and lie on another branch, say β 0 , between a and b. Now the graph Γ is not a cycle, so the pair {a0 , c} cannot separate |Γ|. This contradicts (α1 ). So if L = 2, the set A is equal to the set of vertices of a branch β in Γ.  Lemma 2.3. Suppose a Type 2 vertex in the JSJ decomposition of WΓ has stabilizer hAi. Assume A contains the set of vertices of a branch β of Γ, including its end vertices b and b0 . Let k be the 10 number of components of Γ \ {b, b0 }. Then k ≥ 2, and k ≥ 3 if and only if A equals the set of vertices of β. Proof. First observe that k ≥ 2 since the branch β is not all of Γ. Now assume k ≥ 3, and let Λ1 , . . . , Λk be the components of Γ \ {b, b0 }, with Λ1 = β \ {b, b0 }. Suppose A contains a vertex a which is not on β. Without loss of generality, a ∈ Λ2 . Let c be an interior vertex of β. We will show that Γ \ {a, c} is connected, hence the (non-adjacent) pair {a, c} does not separate |Γ|. This contradicts condition (α1 ) for A. Define Λ to be the induced subgraph of Γ which is the union of b, b0 , and the graphs Λi for 3 ≤ i ≤ k. Since k ≥ 3, the graph Λ is connected, and since c ∈ Λ1 and a ∈ Λ2 , we have that Λ is contained in one component of Γ \ {a, c}. It now suffices to show that for any vertex d ∈ Λ1 ∪ Λ2 with d 6∈ {a, c}, there is a path in Γ from d to either b or b0 which misses both a and c. If d ∈ Λ1 \ {a, c} = Λ1 \ {c}, then d is an interior vertex of the branch β and the result is clear. If d ∈ Λ2 \ {a, c} = Λ2 \ {a}, consider the induced graph Λ2 with vertices b, b0 , and the vertex set of Λ2 . We claim that d must be in a component of Λ2 \ {a} containing either b or b0 , hence there is a path in Λ2 from d to either b or b0 which misses a (and also c), and we are done. Suppose for a contradiction that d is in a component of Λ2 \ {a} containing neither b nor b0 . By Assumption 1.1, the graph Γ has no separating vertices, so there is a reduced path η in Γ from d to say b which does not pass through a. Without loss of generality, we may assume that η does not pass through b0 . Then the entire path η is contained in a component of Λ2 \ {a}. So in fact d is in a component of Λ2 \ {a} containing b, and the claim holds. Thus Γ \ {a, c} is connected. We conclude that if k ≥ 3, the set A equals the vertex set of β. Now suppose that A equals the vertex set of β, and let ` be valence of the vertex of the JSJ graph which has stabilizer hAi. Then by (2) of Corollary 2.1, we have that ` ≥ 1 and that A must contain at least one pair of vertices as in (1) of Corollary 2.1. Since A equals the vertex set of the branch β, the set A contains at most one pair of vertices as in (1) of Corollary 2.1, namely the pair {b, b0 }. By the description given in (1) of Corollary 2.1, it follows that k ≥ 3.  Lemma 2.4. Suppose a Type 2 vertex in the JSJ decomposition of WΓ has stabilizer hAi. (1) The set A has a well-defined cyclic ordering a1 , . . . , an . (2) Assume A contains L ≥ 3 essential vertices, and let ai1 , . . . , aiL be the essential vertices of A in the induced cyclic ordering. For 1 ≤ j ≤ L let kj be the number of components of Γ \ {aij , aij+1 }. (a) For all 1 ≤ j ≤ L, kj ≥ 2, hence {aij , aij+1 } is an essential cut pair. (b) For each 1 ≤ j ≤ L, there is at least one branch β of Γ between aij and aij+1 . (c) Suppose a ∈ A is a non-essential vertex, lying between aij and aij+1 in the cyclic ordering. Then kj = 2, there is a unique branch β = βj of Γ between aij and aij+1 , all vertices of β are contained in A, and β contains the vertex a. (d) If there are no non-essential vertices of A lying between aij and aij+1 in the cyclic ordering on A, then kj ≥ 3. (e) Suppose a, b ∈ A are such that ha, bi is the stabilizer of a Type 1 vertex in the JSJ decomposition of WΓ . Then a and b are adjacent in the cyclic ordering on A. Proof. Part (1) is Lemma 3.14(1) of [7]. For (2)(a), we have kj ≥ 2 for all j by (α1 ) in Corollary 2.1 and the graph Γ being 3-convex. For (2)(b), let σ be an induced cycle in Γ containing all vertices of A, as guaranteed by Lemma 3.12 of [7]. Then since aij and aij+1 are adjacent essential vertices in the cyclic ordering on A, there is an arc of σ \ {aij , aij+1 } which contains no essential vertices of A. To see that there is a branch β between aij and aij+1 , it suffices to show that this arc contains no essential 11 vertices of Γ. Assume there is an essential vertex b of Γ which lies on σ between aij and aij+1 . Then it is not hard to see that either every reduced path from aij to aij+1 in the component of Γ \ {aij , aij+1 } containing b passes through b, or Γ contains a subdivided K4 subgraph. In the first case, by the maximality condition (α2 ) of Corollary 2.1, the vertex b is in A, which is a contradiction. The second case contradicts Assumption 1.2. Hence there is at least one branch β of Γ between aij and aij+1 . To prove (2)(c), let a be a non-essential vertex of A lying between aij and aij+1 in the cyclic ordering, and let β be the branch of Γ containing a. (Note that every non-essential vertex of Γ lies on a unique branch of Γ.) Then all vertices of β are in A, by Lemma 3.20 of [7]. Since the cyclic ordering on A is well-defined, and β contains a, it follows that the branch β has endpoints aij and aij+1 . Now as L ≥ 3, the set A is not equal to the vertex set of β. Thus by Lemma 2.3 we get that kj = 2. So there is at most one branch of Γ between aij and aij+1 . Hence β = βj is the unique branch of Γ between aij and aij+1 . We have proved all claims in (2)(c). For (2)(d), since there are no non-essential vertices of A lying between aij and aij+1 in the cyclic ordering on A, without loss of generality aij = a1 and aij+1 = a2 , and Γ \ {a1 , a2 } has kj = k ≥ 2 components. By (2)(b), there is a branch β of Γ between a1 and a2 . Assume k = 2. We will obtain a contradiction by showing that every interior vertex of β is in A. Let b be an interior vertex of β. First notice that the pairs {a1 , b} and {a2 , b} both separate |Γ|, and that if b0 is any other interior vertex of β then {b, b0 } also separates |Γ|. Now since k = 2, for every a ∈ A which is not in β, the pair {a, b} separates |Γ| if and only if both {a, a1 } and {a, a2 } separates |Γ|. There is at least one such a since L ≥ 3. It follows that for all a ∈ A and for all vertices b in the interior of β, the pair {a, b} separates |Γ|. Hence by the maximality condition (α2 ), every interior vertex of β is in A. This contradicts a1 and a2 being adjacent in the cyclic ordering on A, so kj ≥ 3 as required. To prove (2)(e), we have by Corollary 2.1 that Γ \ {a, b} has k ≥ 3 components. Let σ be an induced cycle in Γ containing all vertices of A, as guaranteed by Lemma 3.12 of [7], and suppose a and b are not consecutive in the cyclic ordering on A. Then there are vertices c, d ∈ A so that one of the arcs of σ from a to b contains c and the other arc contains d. Since k ≥ 3, there is also a reduced path from a to b which misses both c and d. But this contradicts Lemma 3.14(2) of [7]. Hence a and b are consecutive in the cyclic ordering on A, as required.  Lemma 2.5. If a Type 2 vertex in the JSJ decomposition of WΓ has stabilizer hAi, then the set A is not equal to the vertex set of an induced cycle in Γ. Proof. If A is equal to the vertex set of an induced cycle in Γ, then the group hAi is cocompact Fuchsian, which contradicts the characterization of stabilizers of Type 2 vertices in [3].  Remark 2.6. Suppose Λ is the JSJ graph of a right-angled Coxeter group WΓ as in the statement of Corollary 2.1. Then Λ is a finite tree (containing at least one edge). By Corollary 2.1(1), all Type 1 vertices of Λ have valence ≥ 3. Hence all valence one vertices of Λ are of Type 2, so Λ has even diameter. Using Corollary 2.1 and the above technical lemmas, it is not hard to check that Λ has diameter 2 if and only if Γ is a generalized Θ-graph, and that Λ has diameter 4 if and only if Γ is a cycle of generalized Θ-graphs (with Γ being 3-convex and satisfying Assumptions 1.1 and 1.2 in both cases). 2.3. Surface amalgams and topological rigidity. We now recall some definitions and a topological rigidity result from Lafont [16]. Definition 2.7 (Surface amalgams and geometric amalgams of free groups). Consider a graph of spaces over an oriented graph Λ with the following properties. 12 (1) The underlying graph Λ is bipartite with vertex set V (Λ) = V1 t V2 and edge set E(Λ) such that each edge e ∈ E(Λ) has i(e) ∈ V1 and t(e) ∈ V2 . (2) The vertex space Cx associated to a vertex x ∈ V1 is a copy of the circle S 1 . The vertex space Sy associated to a vertex y ∈ V2 is a connected surface with negative Euler characteristic and nontrivial boundary. (3) Given an edge e ∈ E(Λ), the edge space Be is a copy of S 1 . The map φe,i(e) : Be → Ci(e) is a homeomorphism, and the map φe,t(e) : Be → St(e) is a homeomorphism onto a boundary component of St(e) . (4) Each vertex x ∈ V1 has valence at least 3. Given any vertex y ∈ V2 , for each boundary component B of Sy , there exists an edge e with t(e) = y, such that the associated edge map identifies Be with B. The valence of y is the number of boundary components of Sy . A surface amalgam X = X (Λ) is the space   G G G Be  . X = Cx t Sy t x∈V1 y∈V2 e∈E(Λ) ∼ where for each e ∈ E(Λ) and each b ∈ Be , we have b ∼ φe,i(e) (b) and b ∼ φe,t(e) (b). If X is a surface amalgam, the surfaces in X are the surfaces Sy for y ∈ V2 (Λ). The fundamental group of a surface amalgam is a geometric amalgam of free groups. Note that in [16], surface amalgams are called simple, thick, 2-dimensional hyperbolic P -manifolds. For an oriented graph Λ as in Definition 2.7, we may by abuse of notation write Λ for the unoriented graph with the same vertex and edge sets and  : E(Λ) → V (Λ)/C2 given by (e) = [i(e), t(e)]. We note that: Remark 2.8. If X = X (Λ) is a surface amalgam, then the JSJ graph of the geometric amalgam of free groups π1 (X ) is the (unoriented) graph Λ. For details, see [18, Section 4.1]. We will use the following topological rigidity result of Lafont. Theorem 2.9. [16, Theorem 1.2] Let X and X 0 be surface amalgams. Then any isomorphism φ : π1 (X ) → π1 (X 0 ) is induced by a homeomorphism f : X → X 0 . 2.4. Coverings of surfaces. We now recall some results on coverings of surfaces. We write Sg,b for the connected, oriented surface of genus g with b boundary components. This surface has Euler characteristic χ(Sg,b ) = 2 − 2g − b. The next lemma allows us to obtain positive genus covers of any Sg,b with negative Euler characteristic. Lemma 2.10. Suppose χ(Sg,b ) < 0. Then Sg,b has a connected 3-fold covering Sg0 ,b , where g 0 = 3g + b − 2 and so g 0 > 0. Proof. By Proposition 5.2 of Edmonds, Kulkarni, and Stong [9], it is enough to check that χ(Sg0 ,b ) = 3χ(Sg,b ). This is an easy calculation. Alternatively, as the referee suggested to us, one may explicitly construct a degree 3 covering map from S1,3 to S0,3 , and then use a pants decomposition to deduce the result for other surfaces. To construct such a map, note that a regular hexagon H with opposite sides attached forms a torus, and the vertices of H project to two points on the torus. Obtain S1,3 by removing open balls around these two points as well as around the center of H. Then the order 3 rotation of H induces an isometry of S1,3 and the corresponding quotient is S0,3 .  We have the following easy corollary. 13 Corollary 2.11. If X = X (Λ) is a surface amalgam, then X has a degree 3 cover X 0 which is a surface amalgam whose underlying graph is also Λ, so that each surface in X 0 has positive genus.  We will make repeated use of the following lemma concerning coverings of positive genus surfaces with boundary, from Neumann [19]. As discussed in [19], the result appears to be well-known. Lemma 2.12. [19, Lemma 3.2] Let S = Sg,b where g > 0 and b > 0. Let D be a positive integer. Suppose that for each boundary component of S, a collection of degrees summing to D is specified. Then S has a connected D-fold covering S 0 with b0 ≥ b boundary components and these specified degrees on the collection of boundary components of S 0 lying over each boundary component of S if and only if b0 has the same parity as D · χ(S). 2.5. Euler characteristic for orbicomplexes. We now recall the definition of orbicomplex Euler characteristic, in the special case that we will need. All of the orbicomplexes that we construct will be 2-dimensional and have (possibly disconnected) underlying spaces obtained by gluing together some collection of right-angled hyperbolic polygons. When the underlying space is a single polygon, the orbifold will be a reflection polygon. We will then identify certain of these reflection polygons along non-reflection edges to obtain other orbicomplexes. The local groups are as follows. All edge groups will be either trivial or C2 , corresponding to non-reflection and reflection edges, respectively. All vertex groups will be the direct product of the adjacent edge groups, and will be either C2 or C2 × C2 . If O is such an orbicomplex, write F for the number of faces (i.e. polygons) in O, E1 (respectively, E2 ) for the number of edges in O with trivial (respectively, C2 ) local groups, and V2 (respectively, V4 ) for the number of vertices in O with C2 (respectively, C2 ×C2 ) local groups. Then     E2 V2 V4 χ(O) = F − E1 + + + . 2 2 4 The fundamental groups of the connected orbicomplexes that we construct will be right-angled Coxeter groups with triangle-free defining graphs, such that if WΓ = π1 (O) then the Euler characteristic of WΓ from Definition 1.4 is equal to χ(O). 3. Orbicomplex construction From now on, Γ is a finite, simplicial, 3-convex graph satisfying Assumptions 1.1 and 1.2. In this section we construct a piecewise-hyperbolic orbicomplex OΓ with fundamental group WΓ , such that covers of OΓ corresponding to torsion-free, finite-index subgroups of WΓ are surface amalgams. The bottom right of Figure 4.3 gives an example of the orbicomplex OΓ when Γ is a generalized Θ-graph, and Figure 4.4 contains an example of OΓ when Γ is a cycle of generalized Θ-graphs. We begin by constructing hyperbolic orbifolds in Sections 3.1 and 3.2 which have fundamental groups the stabilizers of Type 2 vertices in the JSJ decomposition of WΓ (see Corollary 2.1). The underlying spaces of these orbifolds are right-angled hyperbolic polygons. We then in Section 3.3 glue these orbifolds together along their non-reflection edges to obtain OΓ . Some features of the orbifolds constructed in this section are summarized in the first two columns of Table 1 at the end of Section 4. 3.1. Branch orbifolds. For each branch β in Γ, we construct an orbifold Pβ with fundamental group the special subgroup Wβ . We call Pβ a branch orbifold. We will also assign types to some of the edges and vertices of Pβ , which will later be used to glue Pβ to other orbifolds. 14 Let β be a branch with n = nβ vertices (including its endpoints). Since Γ is 3-convex, we have n ≥ 4. Let P = Pβ be a right-angled hyperbolic p-gon where p = n + 1 ≥ 5. We construct P to have one edge of length 1. Now we construct the orbifold Pβ over P . The distinguished edge of P with length 1 is a nonreflection edge of Pβ . The other n = p − 1 edges of P are reflection edges of Pβ , with local groups hb1 i, . . . , hbn i in that order, where b1 , . . . , bn are the vertices in β going in order along the branch. For the vertex groups of Pβ , the endpoints of the unique non-reflection edge of Pβ have groups hb1 i and hbn i, so that for i = 1 and i = n the vertex group hbi i is adjacent to the edge group hbi i. The other n − 1 = p − 2 vertex groups of Pβ are hbi , bi+1 i ∼ = C2 × C2 for 1 ≤ i < n, with hbi , bi+1 i the local group at the vertex of Pβ whose adjacent edges have local groups hbi i and hbi+1 i. 4−p An easy calculation shows that the Euler characteristic of Pβ is χ(Pβ ) = 3−n 4 = 4 = χ(Wβ ). In fact, we have: Lemma 3.1. The fundamental group of Pβ is Wβ . Proof. We regard Pβ as a simple polygon of groups over the underlying polygon P , with trivial face group, trivial group on the non-reflection edge, and the other edge and vertex groups as described above (see [4, Example 12.17(6)] for the general definition of a simple polygon of groups). Notice that each vertex group is generated by its adjacent edge groups. Since the face group is trivial, it follows that the fundamental group of Pβ is generated by its edge groups, subject to the relations imposed within its vertex groups (compare [4, Definition 12.12]). Hence by construction, π1 (Pβ ) is generated by the vertices of the branch β, which are b1 , . . . , bn going in order along the branch, with relations b2i = 1 for 1 ≤ i ≤ n, and [bi , bi+1 ] = 1 for 1 ≤ i < n. That is, π1 (Pβ ) is the special subgroup Wβ generated by the vertex set of β. Alternatively, as suggested to us by the referee, consider the hyperbolic orbifold whose underlying space is S0,1 , the sphere with one boundary component, with n−1 = p−2 ≥ 3 cone points of order 2. Its fundamental group is the free product of n − 1 copies of C2 . Then Pβ is obtained by quotienting this orbifold by a reflection in a properly embedded segment containing all the cone points. An easy computation shows that the orbifold fundamental group of Pβ is Wβ , as desired.  We assign the non-reflection edge of Pβ to have type {b1 , bn }. Note that since b1 and bn are the end vertices of a branch in Γ, the pair {b1 , bn } is an essential cut pair in Γ. For i = 1 and i = n, we assign type {bi } to the vertex of Pβ which has group hbi i. 3.2. Essential vertex orbifolds and non-branch orbifolds. Now let A be a subset of vertices of Γ so that WA = hAi is the stabilizer of a Type 2 vertex in the JSJ decomposition, and let L be the number of essential vertices of A. Recall from Lemma 2.2 that L ≥ 2. In this section, we assume that L ≥ 3 and construct two hyperbolic orbifolds, QA and A. By Lemma 2.2, since L ≥ 3 the set A is not equal to the set of vertices of a branch of Γ. However the set A may still contain the vertex sets of branches. The orbifold A will be constructed in two stages: we first construct the orbifold QA over a 2L-gon, and then obtain A by gluing on the branch orbifold Pβ for each branch β whose vertex set is contained in A. The fundamental group of QA will be the special subgroup generated by the essential vertices in A, and A will have fundamental group WA . We refer to QA as an essential vertex orbifold and to A as a non-branch orbifold. We now construct QA . Let Q = QA be a right-angled hyperbolic 2L-gon (L ≥ 3). Using the following lemma, we can specify that alternate edges of Q have length 1. Lemma 3.2. Given L ≥ 3, there exists a right-angled hyperbolic 2L-gon in which alternate edges have length 1. 15 Proof. Given any three positive numbers, there exists a right-angled hyperbolic hexagon with alternate edges having lengths equal to these three numbers (cf. Proposition B.4.13 of [2].) The result then follows by induction on L, since if a right-angled hyperbolic hexagon with alternate edges of length 1 is glued along an edge of length 1 to a right-angled hyperbolic 2L-gon with alternate edges of length 1, the result is a right-angled hyperbolic 2(L + 1)-gon with alternate edges of length 1.  The essential vertex orbifold QA is constructed over Q. The alternate edges of Q of length 1 will be non-reflection edges of QA . The remaining edges of QA will be reflection edges, as follows. By Lemma 2.4(1), the set A has a well-defined cyclic ordering a1 , . . . , an . Let ai1 , . . . , aiL be the essential vertices of A in this induced cyclic order. The reflection edges of QA will have groups haij i ∼ = C2 for 1 ≤ j ≤ L, going in order around Q. Now each vertex of QA is adjacent to one non-reflection edge and one reflection edge with group haij i, and this vertex will also have group haij i. An easy calculation shows that χ(QA ) = 2−L 2 . This is equal to Euler characteristic of the special subgroup generated by the L essential vertices in A, since as Γ is 3-convex, there are no edges between any two essential vertices. We next assign types to certain edges and all vertices of QA . The non-reflection edges going around QA are assigned type {aij , aij+1 } (where j ∈ Z/LZ), and we assign type {aij } to the vertices of QA with group haij i, so that the endpoints of the non-reflection edge of QA with type {aij , aij+1 } have types {aij } and {aij+1 }. Notice that each pair {aij , aij+1 } is an essential cut pair, by Lemma 2.4(2)(a). We now construct the non-branch orbifold A. If A consists entirely of essential vertices, then we put A = QA . Otherwise, by Lemma 2.4(2)(c), there is at least one pair {aij , aij+1 } of essential vertices in A so that Γ \ {aij , aij+1 } has kj = 2 components, and for all such j, there is a unique branch βj of Γ between aij and aij+1 . Denote by Pj the right-angled polygon underlying the branch orbifold Pβj constructed in Section 3.1 above. Recall that all non-reflection edges in QA have length 1, and that the unique non-reflection edge in Pβj has length 1 as well. Also, the non-reflection edge of Pβj has type {aij , aij+1 }. To obtain A, for each j such that A contains the vertex set of βj , we glue the non-reflection edge of QA of type {aij , aij+1 } to the unique non-reflection edge of Pβj , so that the types of the end-vertices match up. In both QA and Pβj , the end-vertices of the edge which has just been glued have groups haij i and haij+1 i. Also, the edge groups adjacent to the vertex group haij i (respectively, haij+1 i) are both haij i (respectively, haij+1 i). So we may erase all of the non-reflection edges along which we just glued branch orbifolds to QA , and combine the edges of QA and Pβj with group haij i (respectively, haij+1 i) into a single edge with group haij i (respectively, haij+1 i). We now define A to be the resulting orbifold over the right-angled hyperbolic polygon obtained by gluing together the polygon QA which underlies QA and the polygons Pj which underly the Pβj . Lemma 3.3. The fundamental group of A is WA . Proof. By construction, the reflection edges of A have groups ha1 i, . . . , han i, where a1 , . . . , an is the cyclic ordering on A given by Lemma 2.4(1), and the reflection edges with groups hai i and hai+1 i are adjacent in A if and only if ai and ai+1 are adjacent in Γ (for i ∈ Z/nZ). The proof then uses similar arguments to Lemma 3.1.  For Euler characteristics, a somewhat involved calculation shows that χ(A) = WA . We remark that there is no simple formula for this Euler characteristic in terms of `, the valence of the Type 2 vertex stabilized by WA in the JSJ decomposition, since χ(WA ) depends on the number of edges between vertices in A, and this varies independently of `. 16 Observe that since A is not the vertex set of an induced cycle in Γ (by Lemma 2.5), it follows from our construction that A has at least one non-reflection edge. The non-reflection edges of A retain their types {aij , aij+1 } from QA , as do the endpoints of such edges. 3.3. Construction of orbicomplex. We now construct the orbicomplex OΓ by gluing together certain branch orbifolds Pβ from Section 3.1 and all of the non-branch orbifolds A from Section 3.2. Consider a Type 2 vertex in the JSJ decomposition with stabilizer hAi, where A is a set of vertices of Γ. Let L be the number of essential vertices of A. Then by Lemma 2.2, we have that L ≥ 2, and L = 2 exactly when A is the set of vertices of a branch β of Γ. So if L = 2 then WA = Wβ is the fundamental group of the branch orbifold Pβ , and if L ≥ 3 then WA is the fundamental group of the non-branch orbifold A. Let C be the collection of all branch orbifolds Pβ such that Wβ = π1 (Pβ ) is a Type 2 vertex stabilizer, together with all non-branch orbifolds A. By the discussion in the previous paragraph, we have: Corollary 3.4. The set of orbifolds C is in bijection with the set of Type 2 vertices in the JSJ decomposition of WΓ . Moreover, for each orbifold O in the collection C, we have that π1 (O) is equal to the stabilizer of the corresponding Type 2 vertex. We now consider the relationship between Type 1 vertices in the JSJ decomposition and nonreflection edges of orbifolds in the collection C. Recall that each Pβ has a unique non-reflection edge, each A has at least one non-reflection edge, and each non-reflection edge in either a Pβ or an A has type {a, b} where {a, b} is an essential cut pair of Γ. Lemma 3.5. Each non-reflection edge in the collection of orbifolds C has type {a, b} where ha, bi is the stabilizer of a Type 1 vertex in the JSJ decomposition of WΓ . Proof. First suppose that Pβ is in C, let the endpoints of the branch β be b and b0 , and let A be the vertex set of the branch β. Then since Wβ = WA is the stabilizer of a Type 2 vertex, Lemma 2.3 implies that Γ \ {b, b0 } has k ≥ 3 components. The result then follows from Corollary 2.1(1). Now let A be a non-branch orbifold and let {aij , aij+1 } be the type of a non-reflection edge of A. Then by construction of A and Lemma 2.4(d), we have that Γ \ {aij , aij+1 } has kj ≥ 3 components (otherwise, we would have glued a branch orbifold on at this edge of QA ). The result then also follows from Corollary 2.1(1).  We note that, by similar arguments to those in Lemma 3.5, the number of non-reflection edges of A is equal to ` ≥ 1, the valence of the Type 2 vertex in the JSJ decomposition stabilized by WA . Lemma 3.6. Let a and b be vertices of Γ so that ha, bi is the stabilizer of a Type 1 vertex v of valence k ≥ 3 in the JSJ decomposition of WΓ . Then {a, b} is the type of a non-reflection edge in exactly k orbifolds in the collection C. Proof. Since v has valence k, by the first statement in Corollary 3.4 there are exactly k orbifolds in the collection C whose fundamental groups are stabilizers of Type 2 vertices adjacent to v. By Corollary 2.1(3) and the second statement in Corollary 3.4, these k orbifolds are exactly the elements of C whose fundamental groups contain the generators a and b. If a branch orbifold Pβ is one of these k orbifolds, then a and b are the endpoints of the branch β, so by construction the unique non-reflection edge of Pβ is of type {a, b}. If a non-branch orbifold A is one of these k orbifolds, let A be the set of vertices of Γ so that WA = π1 (A). By the construction of A, it suffices to show that a and b are consecutive in the cyclic ordering on A given by Lemma 2.4(1), and this is Lemma 2.4(2)(e).  17 Recall that each non-reflection edge in the orbifolds we have constructed has length 1. Now for each {a, b} which is the type of some non-reflection edges in the collection C, we glue together all non-reflection edges of type {a, b} in this collection, so that the types of their end-vertices are preserved. The resulting orbicomplex is OΓ . Note that in the resulting space, each non-reflection edge still has a well-defined type {a, b}, and is the unique non-reflection edge of this type. Also, if a non-reflection edge e of OΓ has type {a, b}, and v is its vertex of type {a} (respectively, {b}), then the vertex group of OΓ at v is hai (respectively, {b}), and each reflection edge of OΓ which is adjacent to v has group hai (respectively, {b}). Lemma 3.7. The orbicomplex OΓ has fundamental group WΓ . Proof. The underlying space of OΓ is obtained by gluing together the polygons underlying the orbifolds in the collection C along non-reflection edges. After gluing, there is a polygon in the resulting space for each Type 2 vertex in the JSJ decomposition, and a single non-reflection edge contained in at least 3 distinct polygons for each Type 1 vertex in the JSJ decomposition. Since the JSJ decomposition is over a connected graph, it follows that the underlying space after gluing is connected. Moreover, since the JSJ decomposition is over a tree, it follows that the underlying space after gluing is contractible. Hence the fundamental group of OΓ is generated by the fundamental groups of the orbifolds in C, subject to the identifications of generators induced by gluing nonreflection edges. This gives fundamental group WΓ .  We leave it to the reader to verify that χ(OΓ ) = χ(WΓ ). 4. Half-coverings and torsion-free covers Let OΓ be the orbicomplex with fundamental group WΓ constructed in Section 3. In this section we construct a covering space X of OΓ so that π1 (X ) is an index 16 torsion-free subgroup of π1 (OΓ ) = WΓ . Examples appear in Figures 4.3 and 4.4. The space X will be a surface amalgam with each connected surface in X having positive genus (so that we can obtain further covers by applying Lemma 2.12). We describe the construction using the terminology of half-coverings, which we define in Section 4.1. The surfaces Sβ , which cover the branch orbifolds Pβ , and SA , which cover the non-branch orbifolds A, are constructed in Sections 4.2 and 4.3 respectively. In Section 4.4, we explain how to glue the Sβ and SA together to obtain X . Section 4.5 contains a table summarizing the surfaces we construct, and an explanation of why we use degree 16 covers. 4.1. Half-coverings. In this section we define half-coverings of general bipartite graphs, and a particular half-covering H(T ) where T is a bipartite tree. Given a graph Λ which contains no loops, for each vertex x ∈ V (Λ), let Λ(x) the set of edges of Λ which are incident to x. Now let Λ and Λ0 be (unoriented) graphs, with associated maps  : E(Λ) → V (Λ)2 /C2 and 0 : E(Λ0 ) → V (Λ0 )2 /C2 . (See Section 2.1 for graph-theoretic definitions.) Recall that a graph morphism θ : Λ → Λ0 is a map taking V (Λ) to V (Λ0 ) and E(Λ) to E(Λ0 ), so that for all e ∈ E(Λ), if (e) = [x, y] then 0 (θ(e)) = [θ(x), θ(y)]. Definition 4.1. Let Λ and Λ0 be bipartite graphs with vertex sets V (Λ) = V1 t V2 and V (Λ0 ) = V10 t V20 , respectively. A graph morphism θ : Λ → Λ0 is a half-covering if: (1) For i = 1, 2, the map θ takes Vi to Vi0 . (2) For all x ∈ V1 , the restriction of θ to Λ(x) is a bijection onto Λ0 (θ(x)). (3) For all y ∈ V2 , and for every edge e0 ∈ Λ0 (θ(y)), there is an e ∈ Λ(y) so that θ(e) = e0 . If there is a half-covering θ : Λ → Λ0 then we say that Λ half-covers Λ0 . 18 In short, a half-covering is a morphism of bipartite graphs which preserves the bipartition, is locally bijective at vertices of Type 1, and is locally surjective at vertices of Type 2. (In Section 9, we compare half-coverings with the weak coverings in [1].) In several of our constructions we will use half-coverings of the following particular form. Examples appear in Figures 4.3 and 4.4. Definition 4.2 (The graph H(T ) half-covering a tree T ). Let T be a bipartite tree with vertex set V (T ) = V1 t V2 . Let H(T ) be the bipartite graph defined as follows. (1) The vertex set V (H(T )) equals V10 tV20 , where V10 consists of two disjoint copies of V1 , and V20 is a copy of V2 . (2) Each edge of H(T ) connects a vertex in V10 to one in V20 . Suppose u ∈ V1 corresponds to u0 and u00 in V10 and v ∈ V2 corresponds to v 0 in V20 . Then u0 and u00 are adjacent to v 0 in H(T ) if and only if u is adjacent to v in T . In other words, H(T ) is obtained by taking two copies of the tree T and identifying them along their vertices of Type 2. By construction, the morphism H(T ) → T induced by sending each vertex in V (H(T )) to the corresponding vertex in V1 or V2 (according to the identification of V10 with two copies of V1 , and of V20 with V2 ) is a half-covering. 4.2. Covering the branch orbifolds. Let β be a branch in Γ with nβ vertices in total. The orbifold Pβ constructed in Section 3.1 has underlying space P a right-angled p-gon with p = nβ + 1 ≥ 5. We now construct a connected surface Sβ with genus 2(p − 4) ≥ 2 and 2 boundary components so that Sβ is a 16-fold cover of Pβ . The surface Sβ we construct will be tessellated by 16 right-angled p-gons, so that: • each p-gon has exactly one edge in a boundary component of Sβ ; • the two boundary components of Sβ each contain 8 edges; and • types can be assigned to the edges of this tessellation and to the edges of the p-gon P , such that there is a type-preserving map from Sβ to P which takes each edge in a boundary component of Sβ to the (unique) non-reflection edge of Pβ . It follows that the type-preserving map Sβ → P induces a degree 16 covering map from the surface Sβ to the orbifold Pβ . (More precisely, we can consider the trivial complex of groups G1 (Sβ ) over this tessellation of Sβ , that is, the complex of groups in which each local group is trivial. The type-preserving map Sβ → P then induces a covering of complexes of groups from G1 (Sβ ) to the simple complex of groups Pβ ; see [4] for the general definition of a covering of complexes of groups. This induced covering is 16-sheeted since each face group in G1 (Sβ ) and in Pβ is trivial, and Sβ contains 16 p-gons while P is one p-gon.) For the construction of Sβ , we consider two cases, p ≥ 5 odd and p ≥ 6 even. Case 1: p ≥ 5 is odd. We obtain a tessellated surface Sβ by gluing together, in a type-preserving manner, 16 copies of the right-angled p-gon “jigsaw puzzle piece” shown on the top left in Figure 4.1. Each piece is straight on the bottom and sides, while along the top we have p−5 2 full scoops and p−5 p−5 2 +1 horizontal edges, and there is one half-scoop in the top right corner. (So there are 3+2 2 + 1 + 1 = p sides in total.) Glue 16 of these pieces together into an 8-by-2 block as in Figure 4.1. Now glue the left and right boundaries of this block as indicated by the arrows. The resulting surface will be a sphere with 2 boundary components corresponding to non-reflection edges (these are the outer horizontal edges drawn in thick lines; they form 2 cycles of length 8 after the gluing), 4 boundary components corresponding to the half-scoops, and p−5 2 × 8 boundary components corresponding to the full scoops. These boundary components contain 8, 4, and 2 edges respectively. The final step 19 p=5 16 p=7 p=9 p=6 16 p=8 Figure 4.1. On the left are the “jigsaw puzzle pieces” in the odd and even cases. On the right are the degree 16 covers constructed by gluing the pieces together. The left and right blue boundaries are glued and the center circles are glued in pairs to construct a surface with two boundary components, drawn in green. to obtain Sβ is to glue together the boundary components coming from half-scoops and full-scoops in type-preserving pairs. Case 2: p ≥ 6 is even. In this case we use similar jigsaw puzzle pieces to Case 1, except that now we have p−4 2 full scoops along the top, as shown on the bottom left in Figure 4.1. Glue 16 of these pieces together into an 8-by-2 block as shown in Figure 4.1, and identify the left and right boundaries as indicated by the arrows. The resulting surface will be a sphere with 2 boundary components corresponding to non-reflection edges and p−4 2 × 8 boundary components corresponding to the scoops along the top edge of the piece. The final step to obtain Sβ is to glue together these latter boundary components in type-preserving pairs. 4.3. Covering the essential vertex and non-branch orbifolds. Let A be a subset of vertices of Γ so that hAi is the stabilizer of a Type 2 vertex in the JSJ decomposition of WΓ , and let L ≥ 2 be the number of essential vertices of A. In this section, we assume that L ≥ 3 and construct a connected surface SA which is a 16-fold cover of the orbifold A. (See Section 3.2 for the construction of A.) The surface SA will have genus at least 2 and 2` boundary components, where ` ≥ 1 is the valence in the JSJ decomposition of the vertex with stabilizer hAi. An illustration of the construction appears in Figure 4.2 (which should be viewed in color). As in Section 3.2, we will first consider the essential vertex orbifold QA . We will cover QA by a connected surface SA with genus 3L − 7 ≥ 2 and 2L boundary components. Recall that the underlying space of QA is a right-angled 2L-gon Q. The surface SA we construct will be tessellated by 16 right-angled 2L-gons, so that: • each 2L-gon has its alternate edges in boundary components of SA ; • each boundary component of SA contains 8 edges; and • types can be assigned to the edges of this tessellation of SA and to the edges of the 2L-gon Q, such that there is a type-preserving map from SA to Q which takes each edge in a boundary 20 L=4 L=5 2 1 L=3 1 1' 2' 3 2 1 1' 1'' 2'' 2' 2 3 4 3' 4' 4 1 4 3 3' 3'' 4'' 4' 4 1 2' 3' 3 2 4 1 2' 1'' 4'' 3' 4 2 4' 1' 16 3 2 4' 2'' 3'' 1' 3 16 16 Figure 4.2. On the right are orbifolds with 2L sides which alternate between reflection edges (non-horizontal, and drawn in thin lines) and non-reflection edges (horizontal, and drawn in thick lines). On the left are the initial 8 by 2 blocks of jigsaw puzzle pieces for their degree 16 covers. The blue vertical sides of these blocks are glued as indicated by the arrows, and the center circles are glued as indicated by the labels and arrows, with the top half of a circle in the left half glued to the top (respectively, bottom) half of a circle in the right half as shown by the arrow on the right-hand circle being on the top (respectively, bottom). After gluing, the center horizontal arcs are cut to obtain a surface with 2L boundary components. component of SA to a non-reflection edge of QA . In particular, each boundary component of SA has the same type as a non-reflection edge of QA , and SA has two boundary components of each type. As for branch orbifolds in Section 4.2 above, it follows that the type-preserving map SA → Q induces a degree 16 covering map from the surface SA to the orbifold QA . Start with the 8 by 2 block of jigsaw puzzle pieces from the case p = 2L ≥ 6 even above. The non-reflection edges will be the horizontal edges of the puzzle piece. We will first glue this into a surface with just 2 boundary components of the same type, this type being one of the non-reflection edges of QA . We will then cut along L − 1 curves which have types the other non-reflection edges of QA , to obtain a connected surface with 2 + 2(L − 1) = 2L boundary components, and with 2 boundary components for every type of non-reflection edge of QA . The first stage is to glue together the left and right sides of the block, as indicated by arrows. We then glue pairs of boundary components which consist of 2 edges; these are the “center circles” in Figure 4.2. This is done is such a way that the inner horizontal edges of the 8 by 2 block, those coming from the top edges of the jigsaw piece, form cycles of a single type of length 8. For each non-reflection type, there is one such cycle of 8 edges. 21 In more detail, when L = 3, the gluing pattern for the center circles is as indicated by the arrows and labels in the top row of Figure 4.2. The orange horizontal edges, which occur in pairs in the initial block, then form a cycle of 8 edges passing through the center circles with labels 1, 2, 3, 4 in that order, and similarly the pink horizontal edges form a cycle of 8 edges pass through the center circles with labels 1, 3, 2, 4 in that order. The construction is similar for the pink edges for all L ≥ 4, as indicated in Figure 4.2, with the pink edges passing through light green center circles labeled 1, 3, 2, 4 for all L ≥ 4. For the other horizontal edges, a suitable labeling is constructed by induction on L ≥ 4. When L = 4 we use the labeling from the middle row of Figure 4.2, so that the orange edges form a cycle of 8 edges passing through the center circles with labels 10 , 20 , 30 , 40 in that order, and the purple edges form a cycle of 8 edges passing through alternately black and light green center circles with labels 1, 10 , 3, 30 , 4, 40 , 2, 20 in that order. Write n(k) for the integer n followed by k primes. The labeling of center circles in the left half of the initial block is then, from left to right, 1, 10 , . . . , 1(L−3) , 2(L−3) , . . . , 20 , 2, 3, 30 , . . . , 3(L−3) , 4(L−3) , . . . , 40 , 4. On the right half of the initial block, going from left to right, the 5th (respectively, 6th, 7th, and 8th) puzzle pieces have center circles labeled 1, 20 , 100 , . . . (respectively, 4(L−3) , 3(L−4) , 4(L−5) , . . . ; 2, 40 , 200 , . . . ; and 3(L−3) , 1(L−4) , 3(L−5) , . . . ) when L is odd, and center circles labeled 1, 20 , 100 , . . . (respectively, 3(L−3) , 4(L−4) , 3(L−5) , . . . ; 2, 40 , 200 , . . . ; and 1(L−3) , 3(L−4) , 1(L−5) , . . . ) when L is even. By induction and the alternating pattern of labels on the right half, the only color to be checked is orange. The orange horizontal edges come in pairs in the initial block, and it is not hard to verify that for all L ≥ 4, they form a cycle of length 8 passing through black center circles with labels 1(L−3) , 2(L−3) , 3(L−3) , 4(L−3) in that order. Now cut along each of these cycles of length 8. There are then 2 boundary components of each non-reflection type, each consisting of 8 edges. Using the gluing described above, the resulting surface is connected: for this, we just need to glue at least one top half of a scoop to the bottom half of a scoop, and this follows from our construction. We now describe the surface SA which 16-fold covers the orbifold A. If A consists entirely of essential vertices, then we put SA = SA . Otherwise, for each branch β so that the vertex set of β is in the set A, we glue Sβ to SA as follows. Recall that all edges in the tessellations of Sβ and SA have types, that these types are preserved by the covering maps Sβ → Pβ and SA → QA , and that each boundary component of Sβ and SA has 8 edges. Recall also that A is obtained by gluing Pβ to QA along a non-reflection edge in a type-preserving manner, for each branch β with vertex set in A. For all such β, we now glue one boundary component of Sβ to one boundary component of SA with its same type, and the other boundary component of Sβ to the other boundary component of SA with its same type, so that these gluings match up edges and vertices of the existing tessellations, and so that at each vertex of the resulting tessellation, the incident edges have exactly two types. Finally, we erase the edges in the resulting tessellation which were in boundary components of SA , and denote the resulting tessellated surface by SA . Note that SA is still tessellated by right-angled polygons whose edges have well-defined type. By construction, the 16-fold coverings Sβ → Pβ and SA → QA induce a 16-fold covering SA → A. We observe that since SA has genus 3L − 7 ≥ 2, the surface SA has genus at least 2. Also, the surface SA has two boundary components for each non-reflection edge of A. Now by construction of the orbicomplex OΓ in Section 3.3, the number of non-reflection edges of A equals the valence ` ≥ 1 of the vertex stabilized by hAi in the JSJ decomposition of WΓ . Hence SA has 2` boundary components. 22 θ 16 Γ Λ Figure 4.3. The graph Γ is a generalized Θ-graph, and the graph Λ is the JSJ graph of WΓ . On the right is the degree 16 cover XΓ → OΓ , where OΓ is the orbicomplex described in Section 3.3. In the center is the half-covering θ from the JSJ graph of XΓ to the JSJ graph of OΓ . 4.4. The surface amalgam X . We now construct a surface amalgam X which 16-fold covers the orbicomplex OΓ , by gluing together certain surfaces Sβ constructed in Section 4.2 and all of the surfaces SA constructed in Section 4.3. Examples appear in Figures 4.3 and 4.4. Let Λ be the JSJ graph for WΓ and recall that Λ is a bipartite tree. Let H(Λ) be the graph from Definition 4.2 which half-covers Λ. For each v ∈ V2 (Λ), we write v 0 for the corresponding vertex in V2 (H(Λ)). Now for all v ∈ V2 (Λ), let Sv be the surface Sβ , if v has stabilizer Wβ for β a branch of Γ, and let Sv be the surface SA , if v has stabilizer hAi with A not equal to the vertex set of a branch of Γ. Then the collection of surfaces {Sv | v ∈ V2 (Λ)} = {Sv | v 0 ∈ V2 (H(Λ))} is in bijection with the Type 2 vertices of H(Λ). We now obtain X by gluing together the surfaces {Sv } along boundary components, according to the adjacencies in the graph H(Λ). (In fact, the JSJ graph of X will be H(Λ).) By construction, each Sv has 2` boundary components, where ` ≥ 1 is the valence of v in Λ. Since v 0 has valence 2` in H(Λ), each Sv has number of boundary components equal to the valence of v 0 . We now relabel the boundary components of Sv using the vertices of H(Λ) which are adjacent to v 0 , as follows. Recall that the boundary components of Sv come in pairs which cover the same non-reflection edge (in either Pβ or A), and so have type {a, b} where ha, bi is the stabilizer of a Type 1 vertex in Λ. For each u ∈ V1 (Λ) with stabilizer ha, bi, with corresponding Type 1 vertices u0 and u00 in H(Λ), and for each v adjacent to u in Λ, we label one boundary component of Sv of type {a, b} by u0 and the other by u00 . Now every boundary component in the collection {Sv | v 0 ∈ V2 (H(Λ))} has been assigned a type in V1 (H(Λ)), and for each v ∈ V1 (Λ), there is a bijection between the types of boundary components of Sv and the vertices adjacent to v 0 in H(Λ). We then glue together all boundary components in the collection {Sv } which have the same type, so that these gluings match up edges and vertices of the existing tessellations. The resulting surface amalgam is X . By construction, the 16-fold covers Sβ → Pβ and SA → A induce a 16-fold cover X → OΓ . 4.5. Summary and remarks. We now provide in Table 1 a summary of the surfaces constructed in this section, and explain why we constructed covers of degree 16 (rather than some smaller degree). To cover a branch orbifold Pβ with underlying space a right-angled p-gon, we wanted a construction for all p ≥ 5 of a surface with positive genus and a small positive number of boundary components. The Euler characteristic of Pβ is 4−p 4 so a degree D cover has Euler characteristic 23 θ 16 Γ Λ Figure 4.4. The graph Γ is a cycle of generalized Θ-graphs, and the graph Λ is the JSJ graph of WΓ . On the right is the degree 16 cover XΓ → OΓ , where OΓ is the orbicomplex described in Section 3.3. In the center is the half-covering θ from the JSJ graph of XΓ to the JSJ graph of OΓ . Orbifold Orbifold Connected Genus of Euler covering covering characteristic surface surface No. of boundary components of covering surface Branch orbifold Pβ χ(Wβ ) = 4−p Sβ 2(p − 4) ≥ 2 2 4 over p-gon, p ≥ 5, with p − 1 reflection edges, 1 non-reflection edge 2−L Essential vertex orbifold QA SA 3L − 7 ≥ 2 2L 2 over 2L-gon, L ≥ 3, with L reflection edges, L non-reflection edges Non-branch orbifold A χ(WA ) SA ≥2 2` with ` ≥ 1 non-reflection edges Table 1. Summary of orbifolds constructed in Section 3 and their 16-fold connected covering surfaces constructed in Section 4. D(4−p) . 4 For a surface, the Euler characteristic must be an integer, hence D = 4k for some k ≥ 1. Now if the covering surface Sβ has genus g and b boundary component, then χ(Sβ ) = 2 − 2g − b. Equating Euler characteristics when b = 1, we get k(4 − p) = 1 − 2g. This has no solution when p is even. Thus we consider b = 2 boundary components. Equating Euler characteristics when b = 2, we get k(4 − p) = −2g. This has a positive genus solution for all p ≥ 5, provided k is even. So we work with covers of degree D = 4k where k ≥ 2 is even. Now, an essential vertex orbifold QA with L reflection edges and L non-reflection edges has Euler characteristic 2−L 2 , where L ≥ 3. We wanted to cover each such QA by a surface SA of positive genus g with 2L boundary components (since each surface Sβ covering a branch orbifold Pβ will 24 have 2 boundary components, and certain Sβ will be glued onto the surface covering QA in order to cover the non-branch orbifold A). If the degree of the cover is D = 8, by equating Euler characteristics we get 4(2 − L) = 2 − 2g − 2L hence g = L − 3. For L = 3 this gives g = 0, but we would like SA to have positive genus. The next possible degree is D = 16, and the Euler characteristic equation 8(2 − L) = 2 − 2g − 2L has positive genus solution g = 3L − 7 for all L ≥ 3. Hence we work with degree 16 covers. 5. Generalized Θ-graphs In this section we prove Theorem 1.8, which gives the commensurability classification of rightangled Coxeter groups with defining graph a 3-convex generalized Θ-graph. We remark that by Corollary 2.1, the JSJ graph of such a group is a k-valent star, so in particular, the JSJ graph is a tree of diameter 2. Let WΘ and WΘ0 be right-angled Coxeter groups whose defining graphs are 3-convex generalized Θ-graphs Θ = Θ(n1 , . . . , nk ) and Θ0 = Θ(n01 , . . . , n0k0 ) respectively, with k, k 0 ≥ 3. If Θ has branches βi for 1 ≤ i ≤ k, then WΘ is the fundamental group of the orbicomplex O = OΘ constructed in Section 3, obtained by gluing together the branch orbifolds Pi = Pβi , for 1 ≤ i ≤ k, along their non-reflection edge. Since Wi = Wβi is the fundamental group of Pi , the Euler characteristic vector of WΘ from Definition 1.7 is v = (χ(W1 ), . . . , χ(Wk )). Similarly, WΘ0 is the fundamental group of O0 = OΘ0 obtained by gluing together the branch orbifolds Pi0 = Pβi0 , for 1 ≤ i ≤ k 0 , and WΘ0 has Euler characteristic vector v 0 = (χ(W10 ), . . . , χ(Wk0 0 )) where Wi0 = Wβi0 . 5.1. Sufficient conditions for commensurability. Suppose that the vectors v and v 0 are commensurable. Then by definition, k = k 0 and there exist integers K, K 0 ≥ 1 so that Kv = K 0 v 0 . Let a and b be the two essential vertices of Θ. Let ra be the subcomplex of O consisting of all reflection edges with local group hai (so ra is a star of valence k), and let rb ⊂ O be the corresponding subcomplex for b. An immediate generalization of [6, Section 3.1] is that, for any positive integer R, there is a degree R orbicomplex covering RO → O given by unfolding R times along copies of ra and rb so that the central branching edge of RO forms a geodesic path. An easy counting argument then proves that the orbicomplexes KO and K 0 O0 are homeomorphic, hence have isomorphic fundamental groups which are finite-index subgroups of WΘ and WΘ0 , respectively. Therefore WΘ and WΘ0 are commensurable. 5.2. Necessary conditions for commensurability. The proof of the necessary conditions in Theorem 1.8 follows from a slight generalization of [23, Proposition 3.3.2]. In that setting, it was assumed that the analogs of O and O0 had the same Euler characteristic, while here, χ(WΘ ) = χ(O) and χ(WΘ0 ) = χ(O0 ) could be unequal. The proof from [23] can thus be altered as follows. Suppose that the groups WΘ and WΘ0 are commensurable. Then they are quasi-isometric, and by [7, Theorem 3.36], we have k = k 0 . Since WΘ and WΘ0 are virtually torsion-free, they have isomorphic torsion-free, finite-index subgroups H ≤ WΘ and H 0 ≤ WΘ0 . Then the corresponding covers X and X 0 of O and O0 are surface amalgams. Let D and D0 be the degrees with which X and X 0 respectively cover O and O0 . By Theorem 2.9, there is a homeomorphism f : X → X 0 that induces the isomorphism between H and H 0 . Suppose χ(W1 ) = χ(W2 ) = · · · = χ(Ws ) > χ(Ws+1 ) ≥ · · · ≥ χ(Wk ), and 0 χ(W10 ) = χ(W20 ) = · · · = χ(Wt0 ) > χ(Wt+1 ) ≥ · · · ≥ χ(Wk0 ) for some s, t ≤ k. (Note that the ordering in the above display is reversed with respect to the ordering in the proof of Proposition 3.3.2 of [23].) Without loss of generality, D·χ(W1 ) ≥ D0 ·χ(W10 ), 25 and if D · χ(W1 ) = D0 · χ(W10 ), then s ≥ t. The remainder of the proof [23, Proposition 3.3.2] may now be applied with the first and last line of the main equation in the proof changed to account for the fact that the degrees of the covering maps are different. 6. Necessary conditions for cycles of generalized Θ-graphs In this section, we establish our necessary conditions for the commensurability of right-angled Coxeter groups defined by cycles of generalized Θ-graphs, from Theorem 1.12. We show: Proposition 6.1. Let W and W 0 be as in Theorem 1.12. If W and W 0 are commensurable, then at least one of (1) and (2) from Theorem 1.12 holds. In Section 6.1 we fix notation, and we explain the three cases we will consider in Section 6.2. Case 1 is proved in Section 6.3. We establish some results for both Cases 2 and 3 in Section 6.4, then complete the proof of Case 2 in Section 6.5 and that of Case 3 in Section 6.6. 6.1. Notation. Suppose W and W 0 as in Theorem 1.12 are commensurable. We continue all notation from Section 1.1. In addition, let OΓ and OΓ0 be the orbicomplexes with fundamental groups W = WΓ and W 0 = WΓ0 , respectively, constructed in Section 3, and let A and A0 be the (unique) non-branch orbifolds in OΓ and OΓ0 , with fundamental groups WA and WA0 , respectively. So OΓ (respectively, OΓ0 ) is obtained by gluing certain branch orbifolds to A (respectively, A0 ) along non-reflection edges. For each i ∈ I and 1 ≤ j ≤ ri , let Pij be the branch orbifold in OΓ with fundamental group Wij := Wβij , and let Oi be the sub-orbicomplex of OΓ obtained by gluing together the ri orbifolds Pij along their non-reflection edge. Then π1 (Oi ) = WΘi for each i ∈ I. 0 , W 0 , and O 0 with π (O 0 ) = W 0 for each k ∈ I 0 and 1 ≤ l ≤ r 0 . We write E Similarly define Pkl 1 i Θk kl k k k for the ith branching edge of OΓ , that is, the non-reflection edge along which Oi is glued to A, and similarly write Ek0 for the kth branching edge of OΓ0 . If ri = r = rk0 for each i ∈ I and each k ∈ I 0 , then for all 1 ≤ j ≤ r we define the subcomplex Rj of OΓ to be the disjoint union of the “jth ring” of branch orbifolds {Pij | i ∈ I} in OΓ , and define χj to be its Euler characteristic. (Note that, by construction, if i1 6= Pi2 ∈ I then the P underlying spaces of Pi1 j and Pi2 j are disjoint polygons.) Thus χj = χ(Rj ) = i∈I χ(Pij ) = i∈I χij . For each 1 ≤ j ≤ r, we define R0j ⊂ OΓ0 analogously, and put χ0j = χ(R0j ). Let ρ : X → OΓ and ρ0 : X 0 → OΓ0 be the degree 16 torsion-free covers from Section 4. Then π1 (X ) and π1 (X 0 ) are commensurable as well. Let η : Y → X and η 0 : Y 0 → X 0 be covers corresponding to isomorphic finite-index subgroups of π1 (X ) and π1 (X 0 ), and set π = η ◦ ρ and π 0 = η 0 ◦ ρ0 . Let D and D0 be the degrees of the covering maps π : Y → OΓ and π 0 : Y 0 → OΓ , respectively. Finally, let f : Y → Y 0 be the homeomorphism guaranteed by Theorem 2.9. 6.2. Cases. To prove Proposition 6.1, we consider three cases, by comparing the subsets f (π −1 (A)) and π 0−1 (A0 ) of the surface amalgam Y 0 . Note first that since A (respectively, A0 ) contains every branching edge of OΓ (respectively, OΓ0 ), the sets f (π −1 (A)) and π 0−1 (A0 ) will always have non◦ empty intersection containing all branching curves in Y 0 . We write f (π −1 (A)) ∩ π 0−1 (A0 ) = ∅ if the interiors of f (π −1 (A)) and π 0−1 (A0 ) are disjoint, that is, these subsets of Y 0 have no surfaces ◦ in common, and we write f (π −1 (A)) ∩ π 0−1 (A0 ) 6= ∅ if the interiors of f (π −1 (A)) and π 0−1 (A0 ) are non-disjoint, that is, these subsets of Y 0 have at least one surface in common. The cases we consider, and their consequences, are as follows: Case 1. If f (π −1 (A)) = π 0−1 (A0 ), we show that condition (1) holds. ◦ Case 2. If f (π −1 (A)) ∩ π 0−1 (A0 ) = ∅, we show that condition (2) holds. 26 ◦ Case 3. If f (π −1 (A)) 6= π 0−1 (A0 ) and f (π −1 (A)) ∩ π 0−1 (A0 ) 6= ∅, we construct new homeomorphic finite-sheeted covers of OΓ and OΓ0 which satisfy Case 1. It follows that condition (1) holds. 6.3. Case 1. In this case we prove: Proposition 6.2. If f (π −1 (A)) = π 0−1 (A0 ) then condition (1) in Theorem 1.12 holds. Proof. Given i ∈ I, consider one component S of the preimage of Oi in Y. The assumption that f (π −1 (A)) = π 0−1 (A0 ) implies that π 0 (f (S)) cannot intersect the interior of A0 , and so f (S) must cover some orbicomplex Ok0 in OΓ0 with rk0 = ri . If ri ≥ 3, then the vectors vi and vk0 are commensurable by Theorem 1.8 for generalized Θ-graphs. If ri = 2 and Θi has branches βi1 and βi2 , then form a new Θ-graph Θ with branches βi1 and two copies of βi2 , and do the same for Θ0k to get Θ0 . Now WΘ and WΘ0 are commensurable, since we can construct homeomorphic covers of OΘ and OΘ0 from S and f (S) by adding extra copies of 0 , respectively. It follows from Theorem 1.8 that the vectors the surfaces mapping to Pi2 and Pk2 0 0 0 (χi1 , χi2 , χi2 ) and (χk1 , χk2 , χk2 ) are commensurable, and so vi = (χi1 , χi2 ) and vk0 = (χ0k1 , χ0k2 ) are as well. Applying the same argument with f −1 , we conclude that the sets of commensurability classes of the vectors {vi | ri ≥ 2} and {vk0 | rk0 ≥ 2} coincide, proving condition (1)(a). We now prove (1)(b). It follows from the proof of (1)(a) that ! !! [ [ f π −1 Oi = (π 0 )−1 Ok0 . k∈I 0 i∈I π1 (Ok0 ) Now π1 (Oi ) = WΘi for each i ∈ I and are D and D0 respectively, we deduce that = WΘ0k for each k ∈ I 0 . As the degrees of π and π 0 ! (6.1) D X χ(WΘi ) ! = D0 X χ(WΘ0k ) . k∈I 0 i∈I Now f (π −1 (A)) = (π 0 )−1 (A0 ) by assumption, and π1 (A) = WA and π1 (A0 ) = WA0 , so we also have D · χ(WA ) = D0 · χ(WA0 ). This together with Equation (6.1) implies (1)(b).  6.4. Results for both Cases 2 and 3. In this section we establish some results which are relevant to both of the remaining cases. From now on, we suppose that f (π −1 (A)) 6= π 0−1 (A0 ). We start by showing in Lemma 6.3 that all nontrivial generalized Θ-graphs in both Γ and Γ0 have the same number of branches. Next, in Proposition 6.4 we prove that the vectors w and w0 are commensurable, hence condition (2)(b) in Theorem 1.12 holds. We then color certain suborbicomplexes of OΓ and OΓ0 , using the commensurability of w and w0 , and make some observations about this coloring in Remark 6.5 and Lemma 6.6. In Remark 6.7 we discuss the structure of the subset π −1 (A) ∪ f −1 (π 0−1 (A0 )) of Y, and then in Lemma 6.9 we consider pre-images of branching edges in OΓ and OΓ0 . Lemma 6.3. If f (π −1 (A)) 6= π 0−1 (A0 ) then there exists r ≥ 2 such that ri = rk0 = r for all i ∈ I and all k ∈ I 0 . Proof. We prove the contrapositive. If the conclusion fails, then one of the following holds: (i) ri1 6= ri2 for some ri1 , ri2 ∈ I; (ii) rk0 1 6= rk0 2 for some rk0 1 , rk0 2 ∈ I 0 ; or (iii) ri = r for all i ∈ I and rk0 = s for all k ∈ I 0 , but r 6= s. 27 Suppose (i) holds. If S is any component of π −1 (A), then S is incident to branching curves in Y with different branching degrees ri1 + 1 and ri2 + 1. Then π 0 (f (S)) is an orbifold of OΓ0 which is incident to branching edges of at least two different degrees. It follows that π 0 (f (S)) = A0 and A0 has branching edges of at least two degrees. Applying the same argument with f −1 , we see that π(f −1 (T )) = A for each component T of π 0 −1 (A0 ). Thus f (π −1 (A)) = π 0−1 (A0 ). The proof is identical if (ii) occurs. Finally, (iii) cannot occur, since the degree of branching is preserved by homeomorphisms.  We next extend the techniques used to show that the Euler characteristic vectors of generalized Θ-graphs are commensurable (in the proof of Theorem 1.8) to prove the following. Proposition 6.4. The vectors w and w0 are commensurable. 0 Proof. Let w = (w1 , . . . , wr+1 ) and w0 = (w10 , . . . , wr+1 ), and suppose w1 = · · · = ws > ws+1 ≥ · · · ≥ wr+1 , and 0 0 w10 = · · · = wt0 > wt+1 ≥ · · · ≥ wr+1 . Without loss of generality, we may assume that Dw1 ≥ D0 w10 , and if Dw1 = D0 w10 then t ≤ s. For each 1 ≤ u ≤ r + 1, let Tu ⊂ OΓ be the sub-orbicomplex corresponding to the entry wu of w. That is, if wu = χ(WA ) then Tu = A, and if wu = χj then Tu = Rj . Similarly, let Tu0 ⊂ OΓ0 be the sub-orbicomplex corresponding to the entry wu0 of w0 . Note that f (π −1 (Tu )) ⊂ Y 0 is a disjoint collection of connected surfaces with boundary, and the set of boundary curves of these surfaces is exactly the set of branching curves of Y 0 . We now partition f (π −1 (T1 )) ⊂ Y 0 as follows: • SA0 is the union of the connected surfaces in f (π −1 (T1 )) which cover A0 . • For each k ∈ I 0 , Sk is the union of the connected surfaces in f (π −1 (T1 )) which cover a branch orbifold in Ok0 . Suppose SA0 forms a cover of degree D00 ≤ D0 of A0 . This includes the possibility that SA0 is empty, in which case we put D00 = 0. Then (π 0 )−1 (Ek0 ) ∩ SA0 covers the branching edge Ek0 by degree D00 as well, for each k ∈ I 0 . For each component surface S of a collection Sk , let d(S) 0 0 −1 0 0 0 be Xthe degree of π restricted to S. Since (π ) (Ek ) covers Ek by degree D , it follows that d(S) = D0 − D00 for all k ∈ I 0 . So, S⊂Sk Dw1 = χ(π −1 (T1 )) = χ(f (π −1 (T1 ))) = χ(SA0 ) + X χ(Sk ) k∈I 0  = D00 χ(A0 ) + X  X 0 0 d(S) χ(Pkl ) where S covers Pkl S S  k∈I 0 S⊂Sk  ≤ D00 w10 + X k∈I 0 0 χ(Pk1 )  X d(S) ≤ D00 w10 + (D0 − D00 )w10 = D0 w10 . S⊂Sk By our assumption, Dw1 ≥ D0 w10 . Thus we conclude that Dw1 = D0 w10 . Now, each branching curve in Y 0 is incident to exactly s connected surfaces in f (π −1 (T1 )) ∪ · · · ∪ f (π −1 (Ts )). It follows 0 }, that π 0 (f (π −1 (T1 )) ∪ . . . ∪ f (π −1 (Ts ))) must have in its image at least s orbifolds in {T10 , . . . , Tr+1 Ss 0 0 −1 and therefore t ≥ s. Thus we have Dwi = D wi for 1 ≤ i ≤ s = t. Moreover, i=1 π (Ti ) = S s 0−1 (T 0 ). Now the above argument can be repeated (at most finitely many times) with the i i=1 π remaining sub-orbicomplexes of OΓ and OΓ0 which are of strictly smaller Euler characteristic, proving the claim.  28 The next step in both Cases 2 and 3 is to color certain sub-orbicomplexes of OΓ and OΓ0 , as follows. Let w0 ∈ Zr+1 be the minimal integral element in the commensurability class of w and w0 , so that w = Rw0 and w0 = R0 w0 . Let C = {c1 , . . . , cn } be the set of distinct values occurring in w0 , and assume that c1 > · · · > cn , so that w0 = (c1 , . . . , c1 , c2 , . . . , c2 , . . . , cn , . . . , cn ). We call the elements of the set C the colors. Let Tu and Tu0 be defined as in the proof of Proposition 6.4 for 1 ≤ u ≤ r. We now color OΓ and OΓ0 so that Tu has color c ∈ C if wu = Rc and Tu0 has color c ∈ C if wu0 = R0 c. Note that for each color c, there is an m = mc so that each branching edge Ei 0 ) of color c. (respectively, Ek0 ) is incident to m branch orbifolds Pij (respectively, Pkl Remark 6.5. (The map f preserves colors.) The following equation is an easy consequence of the proof of Proposition 6.4:   f π −1 ({Tu | Tu ⊂ OΓ has color c}) = π 0−1 {Tu0 | Tu0 ⊂ OΓ0 has color c} . It also follows that orbifolds of the same color which are attached to a single edge in OΓ (respectively, OΓ0 ) are identical: Lemma 6.6. For i ∈ I, if Pij1 and Pij2 have the same color c, then χij1 = χij2 , and Pij1 and 0 0 Pij2 are therefore identical orbifolds. The analogous statement holds for Pkl and Pkl of the same 1 2 0 color, where k ∈ I . Proof. Let i ∈ I. If the conclusion fails, we may assume that χP ij1 > χij2 . Then P by our assumption 0 0 that χuj ≥ χuj whenever j < j , we see that j1 < j2 , and u∈I χuj1 > u∈I χuj2 . This is a contradiction as both of these sums are equal to Rc. Thus χij1 = χij2 , and Pij1 and Pij2 are identical orbifolds. The proof of the second sentence is identical.  For the proofs of Cases 2 and 3, it will be important to understand the structure of the subset π −1 (A) ∪ f −1 (π 0−1 (A0 )) of Y, which is described in the following remark. Remark 6.7. (Structure of π −1 (A) ∪ f −1 (π 0−1 (A0 ))). Each of π −1 (A) and f −1 (π 0−1 (A0 )) is a disjoint union of (connected) surfaces in Y. Let W be the collection of surfaces in Y which are in both these sets, that is, W is the collection of surfaces in π −1 (A) ∩ f −1 (π 0−1 (A0 )). Now if S is a surface in π −1 (A) \ W, then since π 0 (f (S)) 6= A0 , for each boundary component of S there is necessarily a surface in f −1 (π 0−1 (A0 )) incident to this boundary component. If S 0 is one of these, then π(S 0 ) 6= A, since S and S 0 share a boundary curve and S is in π −1 (A). Hence every other boundary component of S 0 also borders a component of π −1 (A). Continuing in this way, we see that the connected component of π −1 (A) ∪ f −1 (π 0−1 (A0 )) which contains S has no boundary. Thus there is a decomposition π −1 (A) ∪ f −1 (π 0−1 (A0 )) = W t Z, where Z is a disjoint union of closed surfaces. Moreover, given a component Z of Z, the branching curves on Z partition it as 0 are the subsets of Z consisting of surfaces in Y such that Z = ZA ∪ ZA0 , where ZA and ZA π(ZA ) = A, π(ZA0 ) ⊂ ∪i∈I Oi , π 0 (f (ZA0 )) = A0 , and π 0 (f (ZA )) ⊂ ∪k∈I 0 Ok0 . Observe that ZA0 contains at least one surface mapping into Oi for each i ∈ I, so ZA0 , and similarly ZA , is necessarily disconnected. We denote the set of branching curves of a component Z of Z by the (slightly counterintuitive) notation ∂Z. Then by the description above, ZA and ZA0 intersect in exactly in ∂Z. So we have that ∂Z = ∂ZA = ∂ZA0 . It follows that the degree of π restricted to ZA , the degree of π restricted to ZA0 , and the degree of π restricted to ∂Z are all equal. Observe that for any component Z of Z, the collection ∂Z intersects π −1 (Ei ) for each branching edge Ei of OΓ , as well as f −1 (π 0 −1 (Ek0 )) for each branching edge Ek0 of OΓ0 . In what follows it will be necessary to consider the following subset of ∂Z. 29 Definition 6.8. Let Z be a component of Z. For i ∈ I and k ∈ I 0 , let ik = ik (Z) be the collection of branching curves in ∂Z which map to Ei under π and to Ek0 under π 0 ◦ f . That is, ik = π −1 (Ei ) ∩ f −1 (π 0−1 (Ek0 )) ∩ ∂Z. The degrees of π and π 0 restricted to ik satisfy the following useful property. Lemma 6.9. Let Z be a component of Z, and let ik be the curves from Definition 6.8. Then for each i ∈ I and k ∈ I 0 , the degree of the map π : ik → Ei is a number δk = δk (Z) which depends on k but not on i, and the degree of the map π 0 : f (ik ) → Ek0 is a number δi0 = δi0 (Z) which depends on i but not on k. Moreover, if d and d0 are the degrees of π and π 0 restricted to ∂Z and f (∂Z) respectively, then we have the following equations: X X (6.2) δk = d and δi0 = d0 . k∈I 0 i∈I Proof. We write Z = ZA ∪ ZA0 as in Remark 6.7. By definition, no component of ZA maps to A0 , so we partition ZA into a disjoint collection of possibly disconnected surfaces Sk so that π 0 (f (Sk )) is contained in Ok0 for each k ∈ I 0 . Observe that ∂Sk contains all curves of ∂Z = ∂ZA which map to Ek0 under π 0 ◦ f . Thus ∂Sk = ∪i∈I ik . On the other hand, π maps Sk (and hence ∂Sk ) to A with some degree, say δk , which depends on k. Then for any i, the edge Ei has exactly δk lifts in Sk , or equivalently, the degree of π restricted to ik is δk , which is independent of i. The left equation in (6.2) follows from the fact that ∂Z = ∪k∈I 0 ∂Sk , so the degree d of π restricted to ∂Z is the sum of the degrees of π restricted to ∂Sk over all k ∈ I 0 . The corresponding statements about δi0 are proved similarly.  6.5. Case 2. We now complete the proof of our necessary conditions in the case that ◦ f (π −1 (A)) ∩ π 0−1 (A0 ) = ∅. Lemma 6.3 establishes the first sentence of (2) in Theorem 1.12 and Proposition 6.4 proves (2)(b). Thus it remains to show: ◦ Proposition 6.10. If f (π −1 (A)) ∩ π 0−1 (A0 ) = ∅ then condition (2)(a) holds. Proof. Recall from the discussion preceding Remark 6.5 that w = Rw0 , where the distinct entries of w0 correspond to distinct colors. Let ŵ0 denote the vector obtained from w0 by deleting the pth entry, where p is the index of the entry χ(WA ) of w. We show below that for each i ∈ I, the vector vi = (χi1 , . . . , χir ) is commensurable to ŵ0 . Hence the vectors {vi | i ∈ I} belong to a single commensurability class. Fix i ∈ I. We begin by showing that there exists a number Di0 such that for 1 ≤ j ≤ r, if Pij has color c, then (6.3) χ(f (π −1 (Pij ))) = Di0 R0 c. Suppose the color of A0 is c0 (with c0 possibly equal to c). Partition f (π −1 (Pij )) into (possibly disconnected) subsurfaces TA0 and Tk such that π 0 (TA0 ) = A0 and π 0 (Tk ) is contained in Ok0 , for each k ∈ I 0 . Note that π −1 (Pij ) contains the full preimage of the branching edge Ei . Let Eik be the collection of all branching curves in Y which map to Ei under π and to Ek0 under π 0 ◦ f . Then f (Eik ) consists of ∂Tk together with some curves from ∂TA0 . ◦ Let W and Z be as in Remark 6.7. Then W is empty, since f (π −1 (A)) ∩ π 0−1 (A0 ) = ∅. Thus Eik is the union of the curves ik = ik (Z) corresponding to all components Z of Z. Then it follows 30 P from Lemma 6.9 that the degree of π 0 restricted to f (Eik ) is a number Di0 = Z⊂Z δi0 (Z), which is independent of k. Suppose the degree of π 0 restricted to TA0 is Di00 < Di0 . Then Remark 6.5 implies that for 0 consisting of orbifolds of color c by total each k, the map π 0 sends Tk into the subset of ∪rl=1 Pkl 0 of color c, the Euler characteristic of W 0 = π (P 0 ) is degree Di0 − Di00 . By Lemma 6.6, for all Pkl 1 kl kl 0 independent of l, and is equal to χ(Wkl0 ) for some l0 . Now R0l0 consists of one orbifold of color c for P 0 ). Thus we obtain Equation (6.3) as follows: each k ∈ I 0 , and therefore χ0l0 = R0 c = k∈I 0 χ(Wkl 0 X 0 ) = Di00 R0 c + (Di0 − Di00 )R0 c = Di0 R0 c. χ(f (π −1 (Pij ))) = Di00 R0 c + (Di0 − Di00 )χ(Wkl 0 k∈I 0 To complete the proof, note that χ(f (π −1 (Pij ))) = Dχ(Wij ) = Dχij , since π is a degree D map. Thus Dvi = D(χi1 , . . . , χir ) = (Di0 R0 c1 , . . . , Di0 R0 c1 , . . . , . . . , Di0 R0 cn , . . . , Di0 R0 cn ) = Di0 R0 ŵ0 . This proves the first part of (2)(a), and the second part has a similar proof.  ◦ 6.6. Case 3. We are now in the case that f (π −1 (A)) ∩ π 0−1 (A0 ) 6= ∅. We show that condition (1) of Theorem 1.12 holds in this case by constructing new homeomorphic finite-sheeted covers of OΓ and OΓ0 such that the hypothesis of Case 1 is satisfied. More precisely, we show: ◦ Proposition 6.11. Suppose f (π −1 (A)) ∩ π 0−1 (A0 ) 6= ∅. Then there exist finite-sheeted, connected torsion-free covers µ : M → OΓ and µ0 : M0 → OΓ0 and a homeomorphism f˜ : M → M0 such that f˜(µ−1 (A)) = µ0−1 (A0 ). Then by applying Case 1 to the maps µ, µ0 , and f˜, we get: ◦ Corollary 6.12. If f (π −1 (A)) ∩ π 0−1 (A0 ) 6= ∅, then condition (1) holds.  To prove Proposition 6.11, we assume that Case 1 does not already hold for Y, Y 0 , and f . We observe: ◦ Observation 6.13. Since f (π −1 (A)) ∩ π 0−1 (A0 ) 6= ∅, Remark 6.5 implies that A and A0 have the same color, say a. As Case 1 does not hold, each of OΓ and OΓ0 also has orbifolds besides A and A0 of color a. Idea of the construction of M and M0 . As described in Remark 6.7, we have a decomposition π −1 (A) ∩ f −1 (π 0−1 (A0 )) = W t Z. The hypothesis of Case 1 holds on W and fails on Z. Our assumption that Case 1 does not hold for Y, Y 0 , and f implies that Z is nonempty. As in Remark 6.7, each component Z of Z has a decomposition Z = ZA ∪ ZA0 such that π(ZA ) = A, π(ZA0 ) ⊂ ∪i∈I Oi , π 0 (f (ZA0 )) = A0 , and π 0 (f (ZA )) ⊂ ∪k∈I 0 Ok0 . The key to constructing the new covers is the fact, which we establish below, that χ(ZA ) = χ(ZA0 ). 0 were connected, then by redefining f to interchange its images on these two If ZA and ZA surfaces, we would get a homeomorphism such that the hypothesis of Case 1 is satisfied. However, as observed in Remark 6.7, the surfaces ZA and ZA0 are never connected, so it is not clear that there is a homeomorphism from ZA to f (ZA0 ). Trying to modify π 0 poses similar difficulties, as it is not clear that π 0 can be consistently redefined so that the individual pieces of f (ZA ) each map 0 . Thus a more subtle approach is to A0 , and the individual pieces of f (ZA0 ) each map to some Pkl required. 31 Our strategy is to cut out the interiors of ZA and ZA0 from Y (and correspondingly cut out f (ZA ) and f (ZA0 ) from Y 0 ) and replace them with different surfaces to get new (homeomorphic) surface amalgams M and M0 covering OΓ and OΓ0 respectively. We will replace ZA and f (ZA ) with homeomorphic connected surfaces with boundary which cover both A and A0 . The Euler characteristic of these surfaces will be χ(ZA ) = χ(ZA0 ). We will replace ZA0 and f (ZA0 ) by a collection of homeomorphic connected surfaces with boundary, so that each such homeomorphic 0 of color a in O 0 . The crucial pair covers some orbifold Pij of color a in OΓ and also some orbifold Pkl Γ ingredient for this step is Lemma 6.9, which allows us to derive the necessary Euler characteristic equations required to construct such pairs. After all these replacements, the new covering maps need to extend the already existing covering maps on ∂Z = ∂ZA = ∂ZA0 and f (∂Z). We construct the replacement surfaces in Lemmas 6.14 and 6.15 by passing to our torsion-free covers from Section 4 and then using Lemma 2.12 for constructing further covers. We use Lemma 6.9 to prove the replacement surfaces are homeomorphic. The parity condition in Lemma 2.12 forces the covering maps on the replacement surfaces to have twice the degrees of the restrictions of π and π 0 to Z and f (Z). Thus to glue in these replacement surfaces, we need to pass to degree 2 covers of Y and Y 0 . A further difficulty is posed by the fact that each connected replacement surface of the second 0 , but it may be replacing a collection of surfaces in Z 0 whose type maps to a single Pij and Pkl A image in either OΓ or OΓ0 contains all orbifolds of color a at either Ei or Ek0 . This would make it impossible to extend at least one of π or π 0 to the new surface. To fix this we pass to a further degree m2 cover, where m is the number of orbifolds of color a in any Oi , i ∈ I. Thus we first construct spaces Ȳ and Ȳ 0 consisting of 2m2 disjoint copies of Y and Y 0 . We define new covering maps π̄ and π̄ 0 to OΓ and OΓ0 respectively, designed to make the eventual extensions to M and M0 possible. Then we cut out all 2m2 copies of Z and replace them with m2 copies of each of the replacement surfaces in such a way that π̄ and π̄ 0 extend. The fact that Case 1 holds follows from the construction. We now give the details of the construction. Construction of the replacement surfaces. Let Z be a component of Z. The following lemma defines the surfaces which will be used to replace copies of ZA . Lemma 6.14. Let d and d0 be the degrees of π and π 0 restricted to ∂Z = ∂ZA and f (∂Z) respectively. Then there exist connected surfaces T and T 0 and orbifold covering maps α : T → A and α0 : T 0 → A0 of degree 2d and 2d0 respectively such that the following properties hold. (1) There is a partition ∂T = ∂T + t ∂T − and homeomorphisms β ± : ∂T ± → ∂Z such that α = π ◦ β±. (2) There is a partition ∂T 0 = ∂T 0+ t ∂T 0− and homeomorphisms β 0± : ∂T 0± → f (∂Z) such that α0 = π 0 ◦ β 0± . (3) There is a homeomorphism ψ : T → T 0 whose restriction to ∂T consists of the maps (β 0 ± )−1 ◦ f ◦ β ± . Proof. Recall that π = ρ ◦ η, where ρ : X → OΓ is the degree 16 torsion-free covering space from Section 4. If SA = ρ−1 (A), then η maps the possibly disconnected surface ZA to the connected surface SA by degree d/16 ∈ Z. We will use Lemma 2.12 to construct a connected cover T of SA of degree d/16. Let ∂T + and ∂T − be sets of circles, each equal to a copy of ∂Z, and let β + and β − be homeomorphisms from ∂T + and ∂T − respectively to ∂Z. Then η ◦ β ± a covering map from each circle in ∂T ± to a circle in ∂SA . Moreover, the sum of the degrees of η ◦ β + and η ◦ β − over all circles that 32 cover a given one in ∂SA is 2(d/16) ∈ 2Z by construction (since η is a covering map of degree d/16 on ∂Z). Now by Lemma 2.12, there exists a connected surface T and a covering map ᾱ : T → SA of degree 2d/16 which extends the maps η ◦ β ± on ∂T + t ∂T − , because both the prescribed number of boundary components, namely 2 card (∂Z), and the number 2(d/16)χ(SA ) are even. Hence α = ρ◦ ᾱ is a degree 2d covering map from T to A. This proves (1), and the map α0 : T 0 → A0 of degree 2d0 satisfying (2) is constructed similarly. To prove (3) we first show that T and T 0 have the same Euler characteristic. By construction χ(T ) = 2dχ(WA ) and χ(T 0 ) = 2d0 χ(WA0 ), and we now show that these are equal. As in the proof of Lemma 6.9, we partition ZA into unions of connected surfaces Sk for k ∈ I 0 , where π 0 (f (Sk )) is contained in Ok0 . Observe that the set of boundary components of Sk is exactly ∪i∈I ik , and by Lemma 6.9, the degree of π 0 restricted to f (ik ) is δi0 . Now the degree of π 0 restricted 0 to P f (Sk0 ) is 0exactly the sum of the degrees of π restricted to f (ik ) over all i ∈ I, and this sum is i∈I δi = d (again by Lemma 6.9). Thus we have 0 χ(f (Sk )) = d0 χ(Wkt ), (6.4) 0 = π (P 0 ) for P 0 a branch orbifold of color a. By Lemma 6.6 all branch orbifolds of where Wkt 1 kt kt 0 with total degree d0 for each color a are identical. So f (Sk ) covers an orbifold identical to Pkt 0 1 ≤ k ≤ N . Thus we have X X 0 ) = d0 χ0t = d0 χ(WA0 ) (6.5) d χ(WA ) = χ(ZA ) = χ(f (Sk )) = d0 χ(Wkt k∈I 0 χ0t k∈I 0 χ(R0t ), where the last equality comes from = χ(WA0 ) = χ(π1 (A0 )), and the fact that A0 and R0t have the same color a. Now T and T 0 have the same Euler characteristic and the same number of boundary components, so they are homeomorphic. Moreover, we can choose a homeomorphism ψ between them which extends the homeomorphism (β 0 ± )−1 ◦ f ◦ β ± on their boundaries.  The following lemma constructs pairs of homeomorphic surfaces which map to fixed pairs of 0 . Collections of surfaces of this form will be used to replace copies of Z 0 . orbifolds Pis and Pkt A Lemma 6.15. Let Z be a component of Z, and let ik , δk and δi0 be as in Lemma 6.9. For each i 0 have color a, there and k such that ik is nonempty, and for each s and t such that Pis and Pkt 0 0 of exist connected surfaces S and S and orbifold covering maps ζ : S → Pis and ζ 0 : S 0 → Pkt 0 degree 2δk and 2δi respectively, such that the following properties hold. (1) There is a partition ∂S = ∂S + t ∂S − and homeomorphisms ξ ± : ∂S ± → ∂Z such that ζ = π ◦ ξ±. (2) There is a partition ∂S 0 = ∂S 0+ t ∂S 0− and homeomorphisms ξ 0± : ∂S 0± → f (∂Z) such that ζ 0 = π ◦ ξ 0± . (3) There is a homeomorphism ϕ : S → S 0 whose restriction to ∂S consists of the maps (ξ 0 ± )−1 ◦ f ◦ ξ ± . 0 of the required Proof. As in the proof of Lemma 6.14, we construct covers S → Pis and S 0 → Pkt 0 degrees using Lemma 2.12. By construction, S and S have homeomorphic boundaries. We show below that χ(S) = χ(S 0 ). It follows that S and S 0 are homeomorphic, and we choose a homeomorphism ϕ between them which extends the homeomorphism (ξ 0 ± )−1 ◦ f ◦ ξ ± on their boundaries. 0 ) = δ 0 χ0 , it is enough to show that Since χ(S) = δk χ(Wis ) = δk χis and χ(S 0 ) = δi0 χ(Wkt i kt δk χis = δi0 χ0kt . This follows from the two facts below, where d and d0 are as in Lemma 6.9. (i) dχs = d0 χ0t . 33 (ii) d0 χ0kt = δk χs and dχis = δi0 χ0t . More precisely, facts (i) and (ii) give  0 0  0   0 0 0  δi χkt d χkt δi d χs d0 = . = 0 = d χt δk dχis d δk χis This yields δk χis = δi0 χ0kt as desired. Equation (6.5) in the proof of Lemma 6.14 shows that dχ(WA ) = d0 χ(WA0 ). Now (i) follows from the fact that χs = χ(WA ) and χ0t = χ(WA0 ). To prove (ii), we recall from the proof of Lemma 6.9 that Sk covers A with degree δk . This, together with equation (6.4) in the proof of Lemma 6.14 and the fact that χ(WA ) = χs gives: δk χs = δk χ(WA ) = χ(Sk ) = χ(f (Sk )) = d0 χ0kt . The proof of the second equation in (ii) is similar.  We are now ready to construct the new covers M and M0 . Proof of Proposition 6.11. Suppose that each Oi and Ok0 has m branch orbifolds of color a. Then we will construct homeomorphic surface amalgam covers M and M0 of OΓ and OΓ0 of degrees 2m2 D and 2m2 D0 respectively. New homeomorphic covering spaces Ȳ and Ȳ 0 . Take 2m2 disjoint copies of Y and Y 0 to obtain Ȳ ± : Y → Y ± and τ 0 ± : Y → Y 0 ± be homeomorphisms, and Ȳ 0 . More precisely, for 1 ≤ p, q ≤ m, let τpq pq pq pq and set:         m m m m G G G G G G − + + −   Y 0 pq  . Ȳ =  Ypq Ypq and Ȳ 0 =  Y 0 pq  p,q=1 p,q=1 p,q=1 p,q=1 We define the map π̄ : Ȳ → OΓ as follows. First define an isometry σ : OΓ → OΓ so that if 1 ≤ j1 < · · · < jm ≤ r are the indices of the color a orbifolds Rj in OΓ , then σ(Rju ) = Rju+1 , where ± by π̄ = σ p ◦π◦(τ ± )−1 . u+1 is taken mod m, and σ is the identity elsewhere. Then π̄ is defined on Ypq pq Clearly the resulting map π̄ : Ȳ → OΓ is an orbifold covering map of degree 2m2 D. Similarly we choose the analogous isometry σ 0 : OΓ0 → OΓ0 and define the corresponding covering −1 on Y 0 ± . Finally, we define a map π̄ 0 : Y 0 → OΓ0 of degree 2m2 D0 , by setting π̄ 0 = σ 0 q ◦ π 0 ◦ (τ 0 ± pq pq ) ± )−1 on Y ± . ◦ f ◦ (τ homeomorphism f¯ : Ȳ → Ȳ 0 by setting f¯ = τ 0 ± pq pq pq Construction of M and M0 . Let Z ⊂ Y be as in Remark 6.7. Given a component Z of Z ⊂ Y, ± = τ ± (Z) and ∂Z ± = τ ± (∂Z) denote the let ∂Z denote the branching curves it contains. Let Zpq pq pq pq ± corresponding objects in Ypq . Let MZ be the subspace of Ȳ obtained by cutting out the interiors of all 2m2 copies of Z. More precisely,   m G G  ± ±  MZ = Ȳ \  Zpq \ ∂Zpq . Z∈Z p,q=1 The space M is obtained by gluing a new collection of surfaces to MZ in such a way that each branching curve in M has a neighborhood homeomorphic to a neighborhood of the corresponding curve in Y, and moreover there is an extension of π̄|MZ to a map µ on M, such that π̄ and µ agree on such neighborhoods of branching curves. The space M0 is obtained similarly. The surfaces are glued as follows. Fix a component Z of Z. We glue two types of surfaces to MZ and f (MZ ) along the curves ± m ± ¯ tm p,q=1 ∂Zpq and tp,q=1 f (∂Zpq ) respectively. Let Z = ZA t ZA0 as in Remark 6.7. 34 (1) (Replacements for ZA .) Construct m2 copies each of the surfaces T and T 0 from Lemma 6.14, ± : ∂T 0± 0 indexed by 1 ≤ p, q ≤ m. Let βpq pq → ∂Z and β pq : ∂Tpq → f (∂Z) denote the maps + to ∂Z + via τ + ◦ β + , from the lemma. Then Tpq is attached to MZ by identifying ∂Tpq pq pq pq − to ∂Z − via τ − ◦ β − . Similarly T 0 is attached to f¯(∂Z ± ) in f¯(M ) via the maps and ∂Tpq Z pq pq pq pq pq 0 ± ◦ β ± + . Note that by Lemma 6.14 there are covering maps α τpq pq pq : Tpq → A and 0 0 0 αpq : Tpq → A . (2) (Replacements for ZA0 .) For each non-empty collection of curves ik in Z, we glue in m2 copies of a single surface. Thus ZA0 is replaced by the union of such surfaces for all pairs i, k such that ik is nonempty. Fix such an ik . Construct m2 copies of the surfaces S and S 0 described in Lemma 6.15, 0 , with 1 ≤ p, q ≤ m. Corresponding to p and q, choose s = j and denoted by Spq and Spq p 0 : S0 → P 0 t = lq in the statement of Lemma 6.15, and let ζpq : Spq → Pijp and ζpq be the pq klq ± : ∂S 0± 0 covering maps given by the lemma. Let ξpq pq → ∂Z and ξ pq : ∂Spq → f (∂Z) be the associated maps on the boundaries. ± as in the previous case, because we want to define We cannot simply attach Spq to ∂Zpq a covering map µ on M which agrees with π̄ on a neighborhood of each branching curve in Ȳ. Now µ(Spq ) will have to be a single orbifold Pijs for some 1 ≤ s ≤ m. On the + ∪ Y − (namely all surfaces other hand, Spq is replacing a disjoint collection of surfaces in Ypq pq ± ± in τpq (ZA0 ) with boundary in τpq (ik )). A priori the image of the union of these surfaces under π̄ could be all the Pij of color a. Then for any choice of s there would be some branching curve incident to two surfaces mapping to Pijs , and µ would not be an orbifold covering. To take care of this issue, we define the attaching maps as follows. + . Then ξ + (γ) is a branching curve of Y, which is Let γ be a boundary curve in ∂Spq pq incident to two subsurfaces in Z, say S1 in ZA0 and S2 in ZA . Then π(S1 ) = Piju for some 0 for some 1 ≤ v ≤ m. We then attach the γ boundary 1 ≤ u ≤ m and π 0 (f (S2 )) = Pkl v + + + . Note that u and v depend on γ. of Spq to Y(p−u)(q−v) ∩ MZ via the map τ(p−u)(q−v) ◦ ξpq 0 given by Lemma 6.15, then we attach Now if ϕ is the homeomorphism from Spq to Spq −1 + 0 to Y ¯ + the ϕ(γ) boundary of Spq (p−u)(q−v) ∩ f (MZ ) via the map f ◦ τ(p−u)(q−v) ◦ ξpq ◦ ϕ . ± We repeat this procedure for each γ in ∂Spq to attach all the boundary components of Spq 0 to f (M ). to MZ and the corresponding boundary components of Spq Z We repeat the above procedure for each 1 ≤ p ≤ m and 1 ≤ q ≤ m. This completes the replacement of surfaces corresponding to the copies of the curves ik . We now do the same for each i and k such that ik is nonempty. ± We claim that at the end of steps (1) and (2), each curve in tm p,q=1 ∂Zpq is incident to exactly one surface from (1) and one surface from (2). The first claim is easy. Now suppose (without loss + of generality) that γ ∈ Zst is incident to two surfaces, say along γ1 in Sp1 q1 and γ2 in Sp2 q2 . Then + + ξp1 q1 (γ1 ) = ξp2 q2 (γ2 ) = (τ + st )−1 (γ) in ∂Z, and if u and v are defined as in (2), then s = p1 − u = p2 − u. Thus p1 = p2 , and similarly q1 = q2 . Finally, we repeat steps (1) and (2) above for each component Z of Z to get M and M0 . Note ± that since the ∂Z’s for different Z’s partition ∂Z, we see that each branching curve in tZ∈Z tp,q ∂Zpq ± for some Z, p and q, is incident to exactly one surface from (1) and one from (2). In fact, if γ ∈ ∂Zpq then there exist neighborhoods NY (γ) and NM (γ) in Y and M respectively, and a homeomorphism hγ : NY (γ) → NM (γ), which is the identity on MZ ∩ NY (γ), maps NY (γ) ∩ ZA to the part of NM coming from the surface from (1), and maps NY (γ) ∩ ZA0 to the part of NM coming from the surface from (2). 35 The covering maps µ and µ0 , and the homeomorphism f˜. Define 1 ≤ p, q ≤ m:   0   on MZ π̄ π̄ 0 0 µ = αpq on each Tpq and µ = αpq     ζuv on each Spq ζpq maps µ and µ0 as follows, where on f (MZ ) 0 on each Tpq 0 . on each Spq The overlap between the domains of the various maps is exactly the set of branching curves ± . Thus it follows from the definitions of the gluing maps, the definition of π̄, and tZ∈Z tp,q ∂Zpq items (i) and (ii) in Lemmas 6.14 and 6.15 that µ and µ0 are well-defined. ± . If γ We already know that µ is a covering map away from the branching curves in tZ∈Z tp,q ∂Zpq is such a branching curve, then the above description of the surfaces incident to γ, together with the fact that µ and π̄ agree on γ, shows that there is a homeomorphism h from some neighborhood of γ in M to a neighborhood of γ in Y such that M = π̄ ◦ h. It follows that µ is an orbifold covering map in a neighborhood of each branching curve. Thus µ, and similarly µ0 , are orbifold covering maps. Finally, define f˜ by setting it equal to f¯ on MZ , to ψpq on each surface of the form Tpq as in (1) above, and to ϕpq on each surface of the form Spq as in (2) above, where ψpq and ϕpq are the homeomorphisms from Lemmas 6.14 and 6.15. Then f˜ : M → M0 is a well-defined homeomorphism, and we have that f˜(µ−1 (A)) = µ0−1 (A)0 . If the covers M and M0 constructed above are not connected, we consider the restriction of f˜ to a single connected component of M to get the desired homeomorphic connected covers satisfying the hypothesis of Case 1.  This completes the proof of Proposition 6.1. 7. Sufficient conditions for cycles of generalized Θ-graphs In this section we establish sufficient conditions for the commensurability of right-angled Coxeter groups defined by cycles of generalized Θ-graphs. We show that conditions (1) and (2) from Theorem 1.12 are sufficient in Sections 7.1 and 7.2, respectively. 7.1. Sufficiency of condition (1). We now prove the following result. Proposition 7.1. Let W = WΓ and W 0 = WΓ0 be as in Theorem 1.12. If condition (1) holds, then W and W 0 are commensurable. By the construction given in Section 4, we have surface amalgams X and X 0 which are finitesheeted covers of OΓ and OΓ0 , respectively. To prove Proposition 7.1, we will construct homeomorphic surface amalgams Y and Y 0 so that Y is a finite-sheeted cover of X and Y 0 is a finite-sheeted cover of X 0 . It follows that π1 (Y) ∼ = π1 (Y 0 ) has finite index in both W = π1 (OΓ ) and W 0 = π1 (OΓ0 ), 0 hence W and W are abstractly commensurable. Let Λ and Λ0 be the JSJ graphs of W and W 0 , respectively, and recall from Section 4 that the JSJ graphs of X and X 0 are the half-covers H(Λ) and H(Λ0 ). The surface amalgams Y and Y 0 will have isomorphic JSJ graphs Ψ and Ψ0 , with Ψ and Ψ0 being half-covers of H(Λ) and H(Λ0 ), respectively. After introducing some notation and constants for use in the proof, in Sections 7.1.1 and 7.1.2 respectively, we construct Ψ and Ψ0 in Section 7.1.3. We then construct common covers of the surfaces in X and X 0 in Sections 7.1.4 and 7.1.5. Finally, in Section 7.1.6, we describe how these surfaces are glued together to form homeomorphic surface amalgams Y and Y 0 so that there are finite-sheeted coverings Y → X and Y 0 → X 0 . 36 7.1.1. Notation. We first establish some notation to be used throughout the proof of Proposition 7.1. This includes labeling the vertices of H(Λ) and H(Λ0 ). Using (1)(a) of Theorem 1.12, let V1 , . . . , VM be the commensurability classes of the Euler characteristic vectors {vi | ri ≥ 2} and {vk0 | rk ≥ 2}. For each 1 ≤ p ≤ M , let rp ≥ 2 be the number of entries of each vector in Vp , let wp be the minimal integral element of the commensurability class Vp , 0 | 1 ≤ q0 ≤ N 0 } let {vpq | 1 ≤ q ≤ Np } be the vectors from {vi | ri ≥ 2} which lie in Vp , and let {vpq 0 p 0 0 0 0 be the vectors from {vk | rk ≥ 2} which lie in Vp . For each p, q, and q , let Rpq and Rpq 0 be the 0 = R0 w . (unique) rationals so that vpq = Rpq wp and vpq 0 pq 0 p 1 y11 2 y11 x+ 11 x− 11 y0 x− 12 x+ 21 1 y12 1 y21 2 y21 − x+ 12 x21 2 y12 Γ H(Λ) 3 y21 Figure 7.1. The graph Γ on the left is an example of a cycle of generalized Θ-graphs in which there are two commensurability classes of Euler characteristic vectors for WΓ . On the right, the graph H(Λ) is the JSJ graph of the degree 16 torsion-free cover X of OΓ constructed in Section 4. The vertices of H(Λ) are labeled as described in Section 7.1.1. We use the notation from the previous paragraph to label the vertices of the graphs H(Λ) and H(Λ0 ), as follows. An example of the labeling is given in Figure 7.1. The graph H(Λ) has two P Type 1 vertices for every ri ≥ 2, hence has 2 M p=1 Np Type 1 vertices. We label these as: − V1 (H(Λ)) = {x+ pq , xpq | 1 ≤ p ≤ M, 1 ≤ q ≤ Np }. The central vertex of H(Λ), which is of Type 2, we denote by y0 . This vertex is adjacent to every P Type 1 vertex, so has valence 2 M p=1 Np . For each p, the graph H(Λ) has rp Np additional Type 2 vertices, all of valence 2, which we label so that j V2 (H(Λ)) = {y0 } ∪ {ypq | 1 ≤ p ≤ M, 1 ≤ q ≤ Np , 1 ≤ j ≤ rp }. j − Each of the vertices ypq is adjacent to both x+ pq and xpq . P 0 Similarly, H(Λ0 ) has 2 M p=1 Np vertices of Type 1, given by − 0 0 V1 (H(Λ0 )) = {x+ pq 0 , xpq 0 | 1 ≤ p ≤ M, 1 ≤ q ≤ Np }, and 1 + PM 0 p=1 rp Np vertices of Type 2, given by j 0 0 V2 (H(Λ0 )) = {y00 } ∪ {ypq 0 | 1 ≤ p ≤ M, 1 ≤ q ≤ Np , 1 ≤ j ≤ rp }, j + − so that y00 is adjacent to every Type 1 vertex, and ypq 0 is adjacent to xpq 0 and xpq 0 . 37 7.1.2. Constants. We next define quite a few constants for use in the proof of Proposition 7.1, continuing notation from Section 7.1.1. We will need the following lemma. Lemma 7.2. Assume W = WΓ and W 0 = WΓ0 are as in Theorem 1.12, and that condition (1) holds. Then for each 1 ≤ p ≤ M ,    0  Np Np X X 0  χ(WA0 ) ·  Rpq  = χ(WA ) ·  Rpq . 0 q 0 =1 q=1 Proof. For any vector v, write v for the sum of the entries of v. Using this and the notation of Theorem 1.12, an easy calculation shows that for each i ∈ I and each k ∈ I 0 , χ(WΘi ) = ri X 0 χ(Wβij ) = vi and χ(WΘ0k ) = j=1 rk X χ(Wβkl ) = vk0 . l=1 Now combining this with the notation established in Section 7.1.1, we have by condition (1)(b) of Theorem 1.12 that for each 1 ≤ p ≤ M ,     0   0   Np Np Np Np X X X X 0 0 . Rpq wp  = χ(WA ) ·  Rpq vpq  = χ(WA0 ) ·  vpq χ(WA0 ) ·  0 wp  = χ(WA ) ·  0 q=1 q 0 =1 q=1 q 0 =1 The result follows.  Next, for each 1 ≤ p ≤ M , each 1 ≤ q ≤ Np , and each 1 ≤ q 0 ≤ Np0 , let (7.1) kpq = 16|Rpq | 0 0 and kpq 0 = 16|Rpq 0 |. Then by Lemma 7.2, we may define constants Bp for 1 ≤ p ≤ M by    0  Np Np X X 0  Bp = |χ(WA0 )| ·  kpq  = |χ(WA )| ·  kpq . 0 q 0 =1 q=1 We also define (7.2) B= M X Bp . p=1 Lemma 7.3. With the same hypotheses as Lemma 7.2: 0 (1) each kpq and kpq 0 is a positive integer divisible by 4; (2) each Bp is a positive integer, so B is a positive integer; and (3) for each p, we have Bp ≥ Np . 0 and the rationals Proof. By Definition 1.4, the entries in the Euler characteristic vectors vpq and vpq 0 χ(WA ) and χ(WA0 ) all have denominator at most 4. Since wp is the minimal integral element of 0 the commensurability class Vp , which contains the vectors vpq and vpq0 , the rationals Rpq and Rpq 0 have denominator at most 4 as well. The lemma follows easily from this.  0 , that is, We now define K to be the product of all of the kpq and kpq 0   0  Np Np M Y Y Y 0   (7.3) K= kpq   kpq . 0 p=1 q 0 =1 q=1 38 We also define (7.4) dpq = K , kpq d0pq0 = K 0 , kpq 0 D = K|χ(WA0 )|, and D0 = K|χ(WA )|. Lemma 7.3 now implies: Lemma 7.4. The constants K, dpq , d0pq0 , D, and D0 are all positive integers.  7.1.3. Construction of the half-covers Ψ and Ψ0 . We now construct the graphs Ψ and Ψ0 , and show that they are isomorphic. Let B be as defined at (7.2) above, and recall from Lemma 7.3 that B is a positive integer. The PM PNp 0 graphs Ψ and Ψ will both have 2B Type 1 vertices. In Ψ, we use that B = p=1 q=1 |χ(WA0 )|kpq to put i,− V1 (Ψ) = {xi,+ pq , xpq | 1 ≤ p ≤ M, 1 ≤ q ≤ Np , 1 ≤ i ≤ |χ(WA0 )|kpq }. PN 0  P p 0 In Ψ0 , we use that B = M |χ(W )|k to put 0 0 A p=1 pq q =1 0 0 i ,− 0 0 0 0 V1 (Ψ0 ) = {xipq,+ 0 , xpq 0 | 1 ≤ p ≤ M, 1 ≤ q ≤ Np , 1 ≤ i ≤ |χ(WA )|kpq 0 }. We now describe the Type 2 vertices of Ψ and Ψ0 , and the edges. The graph Ψ has a distinguished P Type 2 vertex y1 of valence 2B, which is adjacent to every Type 1 vertex. There are M p=1 Bp rp additional Type 2 vertices in Ψ, each of valence 2, so that: ij V2 (Ψ) = {y1 } ∪ {ypq | 1 ≤ p ≤ M, 1 ≤ q ≤ Np , 1 ≤ i ≤ |χ(WA0 )|kpq , 1 ≤ j ≤ rp }. PM ij i,− 0 For 1 ≤ j ≤ rp , the vertex ypq is adjacent to xi,+ pq and xpq 0 . Similarly, the graph Ψ has 1+ p=1 Bp rp Type 2 vertices, given by 0 ij 0 0 0 0 V2 (Ψ0 ) = {y10 } ∪ {ypq 0 | 1 ≤ p ≤ M, 1 ≤ q ≤ Np , 1 ≤ i ≤ |χ(WA )|kpq 0 , 1 ≤ j ≤ rp }, 0 0 0 i ,− i ,+ ij with y10 adjacent to every Type 1 vertex, and ypq 0 adjacent to xpq 0 and xpq 0 . Lemma 7.5. The graphs Ψ and Ψ0 are isomorphic. Proof. By construction, it suffices to show that for each p, the Bp vertices {xi,+ pq | 1 ≤ q ≤ Np , 1 ≤ i0 ,+ 0 0 0 0 } in Ψ0 have i ≤ |χ(WA0 )|kpq } in Ψ and the Bp vertices {xpq0 | 1 ≤ q ≤ Np , 1 ≤ i ≤ |χ(WA )|kpq 0 the same collection of valences. But each of these vertices in fact has the same valence, namely i0 ,+ 0 1 + rp , since each xi,+ pq (respectively, xpq 0 ) is adjacent to the central vertex y1 (respectively, y1 ), 0 ij ij and to the rp vertices {ypq | 1 ≤ j ≤ rp } (respectively, {ypq 0 | 1 ≤ j ≤ rp }). The result follows.  The next lemma is easily verified. ij j ± Lemma 7.6. There is a half-covering Ψ → H(Λ) induced by y1 7→ y0 , xi,± pq 7→ xpq , and ypq 7→ ypq , and similarly for Ψ0 → H(Λ0 ). 7.1.4. Covering the central surfaces. Denote by SA the surface in X that covers the orbifold A with degree 16, and by SA0 the corresponding surface in X 0 (see Section 4.3 for the constructions of these surfaces). In this section, we construct a common cover T of the surfaces SA and SA0 . Recall that the boundary components of SA (respectively, SA0 ) are in bijection with the Type 1 vertices of H(Λ) (respectively, H(Λ0 )). Now label the boundary components of SA and SA0 using the notation for the Type 1 vertices of these graphs from Section 7.1.1 above. Let dpq , dpq0 , D, and D0 be the positive integers defined at (7.4) above. 39 Lemma 7.7. There exists a connected surface T with 2B boundary components which covers SA with degree D and SA0 with degree D0 . Proof. Notice first that D · χ(SA ) = (K|χ(WA0 )|) · (16χ(WA )) = (K|χ(WA )|) · (16χ(WA0 )) = D0 · χ(SA0 ), so these covering degrees are compatible with the existence of T . Next we consider the number of boundary components. Recall from Section 4.3 that SA is obtained by gluing together the connected surface SA , which has 2N boundary components, and a connected surface Sβ with two boundary components for every branch β in Γ whose vertices are contained in A. There is one such Sβ for every ri = 1, thus SA has two boundary components P PM 0 for every ri ≥ 2. Hence SA has 2 M p=1 Np boundary components. Similarly, SA0 has 2 p=1 Np boundary components. Now by Lemma 7.3, we have Bp ≥ Np (respectively, Bp ≥ Np0 ) for 1 ≤ p ≤ M . Since B = PM p=1 Bp , it follows that the number of boundary components of SA (respectively, SA0 ) has the same parity as 2B, and is less than or equal to 2B. Now let T be a connected surface with Euler characteristic χ(T ) = D·χ(SA ) = D0 ·χ(SA0 ) and 2B boundary components. To complete the proof, by Lemma 2.12 it suffices to specify the degrees by which the boundary components of T will cover the boundary components of SA and SA0 . For this, we label the 2B boundary components of T in two different ways, first by the Type 1 vertices of Ψ, and second by the Type 1 vertices of Ψ0 (see Section 7.1.3 above). We specify that + for 1 ≤ i ≤ |χ(WA0 )|kpq , the boundary component xi,+ pq of T covers the boundary component xpq 0 0 , the boundary component xi ,+ of SA with degree dpq = K/kpq , and that for 1 ≤ i0 ≤ |χ(WA )|kpq 0 pq 0 i,+ 0 0 0 of T covers the boundary component x+ pq 0 of SA with degree dpq 0 = K/kpq 0 . Hence the union ∪i xpq 0 i ,+ + covers x+ pq with total degree |χ(WA0 )|kpq dpq = |χ(WA0 )|K = D and the union ∪i0 xpq 0 covers xpq 0 i,− − 0 d0 0 with total degree |χ(WA )|kpq 0 pq 0 = |χ(WA )|K = D . Similarly, the union ∪i xpq covers xpq ⊂ SA 0 − with degree D in total and degree dpq on each component, and ∪i0 xipq,− 0 ⊂ T covers xpq 0 ⊂ SA0 with  degree D0 in total and degree d0pq0 on each component. This completes the proof. j j (respectively, Spq 7.1.5. Covering the other surfaces. For each p, q, q 0 , and j, let Spq 0 ) be the surface j j 0 ∈ V2 (H(Λ)) (respectively, ypq in X (respectively, X 0 ) corresponding to the vertex ypq 0 ∈ V2 (H(Λ ))). j j j Recall that each Spq and Spq 0 has 2 boundary components. In this section we construct a surface Tp j j which is a common cover of the surfaces Spq and Spq 0. Lemma 7.8. For each 1 ≤ p ≤ M and each 1 ≤ j ≤ rp , there exists a connected surface Tpj with 2 boundary components so that: j (1) for all 1 ≤ q ≤ Np , the surface Tpj covers Spq with degree dpq in total, and degree dpq on each boundary component; and j 0 0 (2) for all 1 ≤ q 0 ≤ Np0 , the surface Tpj covers Spq 0 with degree dpq 0 in total, and degree dpq 0 on each boundary component. The constants dpq and d0pq0 in this statement are as defined at (7.4) above. j Proof. We prove (1); the proof of (2) is similar. Write wpj and vpq for the jth entry of the vectors wp j and vpq , respectively. Let Tp be a connected surface with 2 boundary components and Euler j characteristic χ(Tpj ) = −Kwpj , where K is the positive integer defined at (7.3) above. Since each Spq 40 j also has 2 boundary components, by Lemma 2.12 it suffices to show that χ(Tpj ) = dpq χ(Spq ). j j Now Spq is a 16-fold cover of an orbifold with Euler characteristic vpq , so j j dpq χ(Spq ) = 16dpq vpq = 16dpq Rpq wpj = −kpq dpq wpj = −Kwpj = χ(Tpj ) where the constant kpq is as defined at (7.1) above. This completes the proof.  7.1.6. Constructions of surface amalgams Y and Y 0 . We finish the proof of Proposition 7.1 by constructing homeomorphic surface amalgams Y and Y 0 which cover X and X 0 , respectively. Recall that the surface amalgam Y will have JSJ graph Ψ, and Y 0 will have JSJ graph Ψ0 , where Ψ and Ψ0 are the isomorphic graphs from Section 7.1.3 above. The surfaces in Y and Y 0 will be those constructed in Sections 7.1.4 and 7.1.5 above, as we now explain. In Y, put Sy1 = T , and in Y 0 , put Sy10 = T , where T is the surface with 2B boundary components from Lemma 7.7. Now ij ij for each p, q, q 0 , i, i0 , and j, if y = ypq ∈ V2 (Ψ) write Spq for the surface Sy in Y, and similarly, if 0 0 ij ij ij i0 j j 0 0 0 y = ypq0 ∈ V2 (Ψ ) write Spq0 for the surface Sy0 in Y . We then put Spq = Tpj in Y and Spq 0 = Tp in Y 0 , where Tpj is the surface with 2 boundary components constructed in Lemma 7.8 above. To glue these surfaces together to form a surface amalgam Y which covers X , we label the ij ij two boundary components of Spq by the two vertices of Ψ which are adjacent to ypq , namely xi,+ pq i,− and xpq . Now recall from the proof of Lemma 7.7 that if we label the boundary components of T = Sy1 by the Type 1 vertices of Ψ, then the boundary component xi,± pq of Sy1 covers the boundary ij ± to the component xpq of SA ⊂ X with degree dpq . Also, by Lemma 7.8, the covering from Tpj = Spq j surface Spq ⊂ X has degree dpq on each boundary component. Since the degrees of these coverings j ij are equal on each boundary component which has the same label, we → Spq Sy1 → SA and Spq ij may glue together boundary components of the collection {Sy1 } ∪ {Spq } which have the same labels to obtain a surface amalgam Y which covers X . The construction of Y 0 which covers X 0 is similar. Now all of the surfaces in Y and Y 0 are compact, hence the coverings Y → X and Y 0 → X 0 are finite-sheeted (in fact, they have degrees D and D0 respectively, but we will not need this). The following lemma then completes the proof of Proposition 7.1. Lemma 7.9. The surface amalgams Y and Y 0 are homeomorphic. Proof. We note first that Y and Y 0 have isomorphic JSJ graphs Ψ and Ψ0 , by Lemma 7.5, so it suffices to consider the surfaces in Y and Y 0 . By construction, Y and Y 0 have the same central ij surface Sy1 = T = Sy10 . Now for each 1 ≤ p ≤ M , consider the Bp rp surfaces {Spq } in Y and the 0 0 ij j ij ij 0 0 0 Bp rp surfaces {Spq 0 } in Y . Since Spq = Tp = Spq 0 for each q, q , i, and i , we have by construction 0 0 i,− i ,+ i ,− 0 that each of the Bp pairs of branching curves {xi,+ pq , xpq } in Y (respectively, {xpq 0 , xpq 0 } in Y ) is incident to the same collection of 1+rp surfaces {T }∪{Tpj }. Hence Y and Y 0 are homeomorphic.  This completes the proof of Proposition 7.1. 7.2. Sufficiency of condition (2). We now prove the following result. Proposition 7.10. Let W = WΓ and W 0 = WΓ0 be as in Theorem 1.12. If condition (2) holds, then W and W 0 are commensurable. By the construction described in Section 4, there are surface amalgams X and X 0 which form degree 16 covers of the orbicomplexes OΓ and OΓ0 , respectively. To prove that W and W 0 are abstractly commensurable, we prove there are finite-sheeted covers Z → X and Z 0 → X 0 and a 41 -36 Z -72 -144 -36 -432 -72 -216 -1440 -108 -144 -72 -288 9 X -4 -8 -144 -16 -72 -288 -144 18 -4 -48 -18 -24 W -8 -160 -12 -80 -16 -8 -32 -16 -8 -32 -36 -72 -16 Figure 7.2. Illustrated above is an example of the covers described in Section 7.2, on the level of the JSJ graphs. The white vertices represent surfaces with specified Euler characteristic; the valence of the vertex is the number of boundary components of the surface. The black vertices represent branching curves that are glued to boundary components of adjacent surfaces. The covering maps between surface amalgams restrict to half-coverings between the JSJ graphs. In this example, the space X does not cover the space W, and there is no minimal surface amalgam within this commensurability class. surface amalgam W so that Z and Z 0 finitely cover W. An example of the construction given in this section appears in Figure 7.2. We will describe these covering spaces and maps in terms of their JSJ graphs and half-covers as defined in Section 4.1. We set notation in Section 7.2.1. The surface amalgam W is described in Section 7.2.2. The surface amalgam Z and the finite cover Z → X is given in Section 7.2.3. Finally, the finite cover Z → W is given in Section 7.2.4. The construction for X 0 is analogous. 7.2.1. Notation. For ease of notation, suppose the elements of the multiset of Euler characteristic vectors {vi | i ∈ I} in the statement of Theorem 1.12 are labeled v1 , . . . , vn . By assumption, there exists r ≥ 2 so that each vector vi for 1 ≤ i ≤ n has r entries. Let V denote the commensurability class of the vectors {vi | 1 ≤ i ≤ n}. The orbicomplex OΓ contains a central orbifold A with n non-reflection edges e1 , . . . , en . Attached to the non-reflection edge ei is a collection of r branch orbifolds Pi1 , . . . , Pir , with Pij = Pβij and π1 (Pij ) = Wij = Wβij for each 1 ≤ i ≤ n and 1 ≤ j ≤ r. By definition, for 1 ≤ i ≤ n the Euler characteristic vector vi is given by vi = (χi1 , . . . , χir ) = (χ(Wi1 ), . . . , χ(Wir )), and w is the reordering of   n n n X X X  χj1 , χj2 , . . . , χjr , χ(WA ) j=1 j=1 j=1 so that its entries are in non-increasing order. 42 The surface amalgam X has a central surface SA with Euler characteristic 16χ(WA ) and 2n boundary components. Attached to each of the n pairs of boundary components of SA are surfaces Si1 , . . . , Sir for 1 ≤ i ≤ n where each Sij has two boundary components and Euler characteristic 16χ(Wij ), and forms a degree 16 cover of Pij . Let Λ be the JSJ graph of W . As described in Section 4.4, the JSJ graph of X is the halfcover H(Λ). The graph H(Λ) has 2n vertices of Type 1. Label these vertices as − V1 (H(Λ)) = {x+ i , xi | 1 ≤ i ≤ n}. There is a central vertex of Type 2 associated to the surface SA ; label this vertex as y0 . This vertex is adjacent to every Type 1 vertex, hence has valence 2n. For each i, the graph H(Λ) has r additional Type 2 vertices, associated to the surfaces Sij for 1 ≤ j ≤ r. Label these vertices accordingly, so that V2 (H(Λ)) = {y0 } ∪ {yij | 1 ≤ i ≤ n, 1 ≤ j ≤ r}. − Each vertex yij is adjacent to x+ i and xi . For 1 ≤ i ≤ n, define vei = 16vi = (χ(Si1 ), . . . , χ(Sir )) ∈ 2Zr<0 . Let w e = 16w, so that w e is the reordering of   n n n X X X  χ(Sj1 ), χ(Sj2 ), . . . , χ(Sjr ), χ(SA ) ∈ 2Zr+1 <0 j=1 j=1 j=1 so that its entries are in non-increasing order. Suppose that in this ordering, χ(SA ) is the kth entry of w. e Let w b ∈ Zr be the vector w e with the kth entry deleted. Then w b = ve1 + · · · + ven ∈ 2Zr<0 . By condition (2)(b) in Theorem 1.12, the vectors w and w0 are commensurable. Let w0 = (w1 , . . . , wr+1 ) ∈ Zr+1 be the minimal integral element in the commensurability class of w and w0 . Since all entries of w and w0 are negative, all entries of w0 are positive. Finally, let w c0 be the vector w0 minus its kth entry, so w c0 = (w1 , . . . , wk−1 , wk+1 , . . . , wr+1 ) ∈ Zr . 7.2.2. The surface amalgam W. The surface amalgam W has JSJ graph Φ with two Type 1 vertices labeled V1 (Φ) = {x+ , x− }. The graph Φ has r + 1 Type 2 vertices labeled V2 (Φ) = {y1 , . . . , yr+1 }. Each Type 2 vertex is adjacent to both x+ and x− . To construct the surface amalgam W, for 1 ≤ i ≤ r + 1, let Wi be the surface with Euler characteristic −2wi and 2 boundary components Ci1 and Ci2 ; the surface Wi is associated to the vertex yi . Identify the curves Ci1 for 1 ≤ i ≤ r + 1 to create a single curve C1 associated to the vertex x+ , and identify the curves Ci2 for 1 ≤ i ≤ r + 1 to create a single curve C2 associated to the vertex x− . This forms a surface amalgam W which has two singular curves C1 and C2 and JSJ graph Φ. Lemma 7.11. The graph H(Λ) forms a half-cover of the graph Φ. 43 − + − Proof. The cover is induced by the maps x+ i 7→ x , xi 7→ x , and yij 7→ yj for 1 ≤ i ≤ n and 1 ≤ j ≤ r, and y0 7→ yr+1 .  Remark 7.12. In view of Lemma 7.11, note that X does not necessarily cover W; an example of this is given in Figure 7.2. To resolve this, we construct a finite cover Z → X so that Z has JSJ graph H(Λ) and finitely covers W. 7.2.3. The surface amalgam Z and the finite covering map Z → X . We will need the following lemma. Lemma 7.13. Let u ∈ Zr be the minimal integral element in the commensurability class V. There exists a positive integer K so that w c0 = Ku. Proof. Since vei = 16vi , the set of integer vectors {e vi , w, b w c0 | 1 ≤ i ≤ n} is also contained in V. Thus there exists an integer K so that w c0 = Ku. Since all entries of w c0 and u are positive, K is a positive integer.  Let Z be the following surface amalgam with JSJ graph H(Λ). Associated to the Type 2 vertex y0 of H(Λ), the space Z contains one central surface Sf A with 2n boundary components f and Euler characteristic χ(SA ) = Kχ(SA ). Attached to the ith pair of boundary curves of Sf A, + − f which are associated to the vertices xi , xi ∈ H(Λ) for 1 ≤ i ≤ n, there are r surfaces Sij for 1 ≤ j ≤ r, where Sf ij has 2 boundary components and Euler characteristic Kχ(Sij ); these surfaces are associated to the vertices yij in H(Λ). Lemma 7.14. The surface amalgam Z forms a degree K cover of the surface amalgam X , where K is the positive integer guaranteed by Lemma 7.13. Proof. By Lemma 2.12, the surface Sf A forms a degree K cover of the surface SA so that the degree restricted to each boundary component of Sf A over the corresponding boundary component of SA is equal to K. Similarly, by Lemma 2.12, each surface Sf ij forms a degree K cover of Sij so that the f degree restricted to each boundary component of Sij over the corresponding boundary component of Sij is equal to K. Thus, since the covering maps agree along the gluings, Z forms a degree K cover of X .  7.2.4. The finite cover Z → W. For 1 ≤ i ≤ n, define f zi = Ke vi = (χ(Sf e i1 ), . . . , χ(Sij )) and z = K w, where K is the positive integer from Lemma 7.13. Note that the entries of z are in non-increasing order. If χ(Sf b be the vector z with the kth entry deleted. Then A ) is the kth entry of z, let z zb = K w b = z1 + · · · + zn ∈ 2Zr<0 . Since w0 is the minimal integral element in the commensurability class of w e ∈ 2Zr+1 <0 , there exists a positive integer M so that w e = −2M w0 . Hence, z = M K(−2w0 ). Let D = M K ∈ N. Then z = D(−2w0 ) and zb = D(−2c w0 ). We compute D a second way, and use this to prove that Z finitely covers W. Lemma 7.15. Let di be positive integers so that vei P = di (−2u), where u is the minimal integral element in the commensurability class V. Then D = ni=1 di . 44 Proof. By construction of the covering map and the definition of K, we have zi = Ke vi = K(di (−2u)) = di (−2c w0 ). Hence zb = z1 + · · · + zn = n X ! di (−2c w0 ). i=1 So, combining the two equations for zb given above, we have that Pn i=1 di = D.  Lemma 7.16. The surface amalgam Z forms a degree D cover of the surface amalgam W. Proof. Suppose for ease of notation that χ(Sf A ) is the (r + 1)st entry of the vector z. Then for all f 1 ≤ i ≤ n and all 1 ≤ j ≤ r, we have that χ(Sij ) = di χ(Wj ), where the surface Wj in W has Euler characteristic wj by definition, and di is defined in Lemma 7.15. By Lemma 2.12, Sf ij covers Wj by degree di so that the degree restricted to each of the boundary components is equal to di . So, Sn f f i=1 Sij forms a degree D cover of Wj for all 1 ≤ j ≤ r. In addition, χ(SA ) = Dχ(Wr+1 ), and the n f pairs of boundary components of SA can be partitioned so that the ith pair of boundary curves Pn of Sf A covers the pair of boundary curves of Wr+1 by degree di , for 1 ≤ i ≤ n, with i=1 di = D. Thus by Lemma 2.12, the space Z forms a D-fold cover of W, concluding the proof.  8. Geometric amalgams and RACGs In this section we further investigate the relationship between geometric amalgams of free groups (see Definition 2.7) and right-angled Coxeter groups. In Section 8.1 we prove Theorem 1.16 from the introduction, which states that the fundamental groups of surface amalgams whose JSJ graphs are trees are commensurable to right-angled Coxeter groups (with defining graphs in G). As a consequence, we obtain the commensurability classification of geometric amalgams of free groups whose JSJ graphs are trees of diameter at most four (Corollary 1.17). Then in Section 8.2 we give an example of a geometric amalgam of free groups which is not quasi-isometric (hence not commensurable) to any Coxeter group, or indeed to any group which is finitely generated by torsion elements. 8.1. Geometric amalgams over trees. Let X (T ) be a surface amalgam whose JSJ graph is a tree T . To prove Theorem 1.16, we construct a surface amalgam X (Λ) which finitely covers X (T ), and which looks like one of the covers constructed in Section 4. We then follow the construction from Section 4 backwards to construct an orbicomplex O with fundamental group a right-angled Coxeter group whose JSJ graph is T . It follows that the fundamental group of X (T ) is commensurable to a right-angled Coxeter group with JSJ graph T . Proof of Theorem 1.16. Let X (T ) be as above. By Corollary 2.11 we may assume that every surface associated to a Type 2 vertex of X (T ) has positive genus. Let Λ = H(T ) be the graph from Definition 4.2, with associated half-covering map θ : Λ → T . We construct a surface amalgam over Λ which covers X (T ) with degree 8. Let y and y 0 be Type 2 vertices in T and Λ respectively, with θ(y 0 ) = y. Recall that for each edge e incident to y in T , there are two edges e0 , e00 incident to y 0 in Λ, such that θ(e0 ) = θ(e00 ) = e. In particular, the valence of y 0 is twice that of y. Suppose Sy is the surface associated to y, with genus g and b boundary components. The surface Sy0 associated to y 0 is defined to be the degree 8 cover of Sy which has 2b boundary components, such that each boundary component of Sy is covered by two boundary curves of Sy0 , each mapping with degree 4. Such a cover exists by Lemma 2.12, since 8 and 2b are both even. This induces a bijection between the boundary components of Sy0 45 and the edges of Λ incident to y 0 . Now form X (Λ) by gluing the surfaces associated to the Type 2 vertices of Λ according to the adjacencies in Λ. The covering maps on the individual surfaces induce a degree 8 cover X (Λ) → X (T ). If y, y 0 , Sy , g, b and Sy0 are as in the previous paragraph, and g 0 is the genus of Sy0 , then we have 8χ(Sy ) = χ(Sy0 ), or equivalently, 8(2 − 2g − b) = 2 − 2g 0 − 2b. Simplifying, we get: (8.1) g 0 = 3b + 8g − 7. Observe that g 0 is even if and only if b is odd, and that g 0 > 3b, since g was assumed to be positive. Next we construct an orbicomplex O as a graph of spaces with underlying graph T such that X (Λ) is a degree 16 cover of O. If y is a terminal Type 2 vertex of T , then the vertex space associated to y is the orbifold Py defined as follows. Let Sy0 ⊂ X (Λ) be as above. It follows from Equation (8.1), since b = 1 in this case, that g 0 is even. Moreover, g 0 is positive. Then Sy0 forms a degree 16 cover 0 of Py , an orbifold with underlying space a right-angled hyperbolic polygon with g2 + 4 ≥ 5 sides, 0 as described in Section 4. The orbifold Py has g2 + 3 ≥ 4 reflection edges and one non-reflection edge of length 1. Now let y be a non-terminal Type 2 vertex of T with valence b. Then the surface Sy0 of X (Λ) has 2b boundary components and genus g 0 > 3b such that g 0 and b have opposite parity. We will realize Sy0 as a cover of an orbifold Ay which is the union of orbifolds QA and P, identified to each other along a non-reflection edge of each. To do this, decompose Sy0 into two surfaces S1 and S2 by cutting along a pair of non-separating curves whose union separates Sy0 , and such that S1 has 2b + 2 boundary components and genus 3(b + 1) − 7 < g 0 , and S2 has 2 boundary components and genus h satisfying g 0 = 1 + 3(b + 1) − 7 + h. Thus h = g 0 − 3b + 3. Since g 0 and b have opposite parity, h is even, and since g 0 > 3b, we know that h is positive. By the construction in Section 4, the surface S1 can be realized as a degree 16 cover of an orbifold QA with underlying space a right-angled 2(b + 1)-gon, so that QA has b + 1 non-reflection edges, each of length 1, and b + 1 reflection edges. Since h is positive and even, the surface S2 forms a degree 16 cover of an orbifold P with h2 + 4 ≥ 5 sides, as in the terminal vertex case. The orbifold P has h2 + 3 reflection edges and one non-reflection edge of length 1. Form the orbifold Ay by gluing the non-reflection edge of P to one of the non-reflection edges of QA . Now let Ay be the vertex space associated to y. We now identify the orbifold vertex spaces described above according to the incidence relations of T along non-reflection edges to get an orbicomplex O. Arguments similar to those from Section 3 show that the fundamental group of O is a right-angled Coxeter group. Since the covering maps from the surfaces in X to the orbifolds in O agree along their intersections, the space X finitely covers O. The theorem follows.  Proof of Corollary 1.17. If T is a tree, then by Theorem 1.16, any geometric amalgam of free groups with JSJ graph T is commensurable to a right-angled Coxeter group W with JSJ graph T . If, moreover, T has diameter at most 4, then by Remark 2.6, the defining graph of W is either a generalized Θ-graph or a cycle of generalized Θ-graphs. Corollary 1.17 follows.  8.2. Geometric amalgams not over trees. In the case of surface amalgams whose JSJ graphs are not trees, the situation is more complicated. Since H(T ) is not a tree, clearly some geometric amalgams of free groups whose JSJ graphs are not trees are commensurable to right-angled Coxeter groups. On the other hand, the following example shows that there exist geometric amalgams of free groups which are not quasi-isometric (and hence not commensurable) to any Coxeter group, or indeed to any group which is finitely generated by torsion elements. 46 Example 8.1. Let G be a geometric amalgam of free groups with JSJ graph Λ as in Figure 8.1, where the Type 1 vertices are white and the Type 2 vertices black. a0 a5 a1 a4 a2 Λ= a3 Figure 8.1. Suppose G is quasi-isometric to a group H which is finitely generated by torsion elements. Then H is also 1-ended and hyperbolic, and G and H have the same JSJ tree T . Moreover, the JSJ graph of H is a finite tree, say Ψ. (The graph Ψ is a tree since any group with a nontrivial loop in its JSJ graph admits a nontrivial homomorphism to Z, but no such homomorphism from H to Z exists.) Now the JSJ tree T half-covers Ψ. We show that this is a contradiction, by showing that if θ : T → T is a half-covering, where T is any tree, then T must be infinite. In the following argument, all indices are taken mod 6. For each vertex ai of Λ = G\T , by abuse of notation write Gai for the G-orbit of some lift of ai in T . The following are straightforward consequences of the definition of a half-cover. (1) Each vertex in θ(Gai ) ⊂ V (T ) is adjacent to at least one vertex from each of θ(Gai−1 ) and θ(Gai+1 ). (2) For all i 6= j, we have θ(Gai ) ∩ θ(Gaj ) = ∅. To show that T must be infinite, we define a map ν : [0, ∞) → T as follows. Choose ν(0) to be an element of θ(Ga0 ), and then for each i > 0, inductively choose ν(i) to be an element of θ(Gai ) adjacent to θ(Gai−1 ), which exists by (1) above. The adjacency condition can now be used to extend ν to [0, ∞). By (2), we conclude that for each i > 0, the vertices ν(i − 1) and ν(i + 1) are distinct (i.e. there is no “folding”). This, together with the fact that T is a tree, implies that ν is injective, and it follows that T is infinite. 9. Discussion In this section we discuss some obstructions to extending our commensurability results. A natural strategy that we considered is that used by Behrstock–Neumann in [1] for quasiisometry classification. In Theorem 3.2 of [1], it is shown that two non-geometric graph manifolds M and M 0 which have isomorphic Bass–Serre trees T and T 0 are quasi-isometric, by first proving that there is a minimal graph Λ00 which is weakly covered by both of the quotient graphs Λ = π1 (M )\T and Λ0 = π1 (M 0 )\T 0 . The notion of weak covering just requires the graph morphism to be locally surjective, unlike the half-coverings we introduce in Section 4.1, which must be locally bijective at vertices of Type 1. The minimal graph Λ00 determines a 3-manifold M 00 which is covered by both M and M 0 , and a bilipschitz map M → M 0 is constructed by showing that there are bilipschitz maps M → M 00 and M 0 → M 00 . We now give an example to explain why such a strategy cannot be pursued in our setting. Let G = G(Λ) and G0 = G(Λ0 ) be geometric amalgams of free groups with JSJ graphs Λ and Λ0 as shown in Figure 9.1, with Type 1 vertices white and Type 2 vertices black. Since all Type 1 vertices in Λ and Λ0 are of valence 3, the groups G and G0 have isomorphic JSJ trees. Now the graph Λ0 47 y3 Λ= y4 Λ0 = y2 y5 y1 Figure 9.1. does not half-cover any other graph, so if any graph is half-covered by both Λ and Λ0 , then there is a half-covering θ : Λ → Λ0 . The map θ must send each of the triples of Type 2 vertices {y1 , y2 , y3 }, {y2 , y3 , y4 }, and {y1 , y4 , y5 } of Λ to the three distinct Type 2 vertices of Λ0 , but this is impossible. Hence there is no (minimal) graph which is half-covered by both Λ and Λ0 . Instead, we may approach commensurability by considering common finite half-covers of Λ and Λ0 . More precisely, we consider common finite half-covers of finite, connected, bipartite graphs Λ and Λ0 in which all Type 1 vertices have valence ≥ 3, such that Λ and Λ0 are both half-covered by the same (infinite) bipartite tree in which all Type 2 vertices have countably infinite valence. With these assumptions, any corresponding geometric amalgams of free groups G = G(Λ) and G0 = G(Λ0 ) will have isomorphic JSJ trees (a necessary condition for commensurability). Using similar methods to those of Leighton [17] (see also Neumann [20]), we can construct a common finite half-cover Λ00 of any two such graphs Λ and Λ0 . However, a half-covering of JSJ graphs does not necessarily induce a topological covering map of associated surface amalgams. For example, in Figure 7.2, the JSJ graph of π1 (X ) does half-cover the JSJ graph of π1 (W), but by considering Euler characteristics, one sees that the surface amalgam X does not cover the surface amalgam W. Hence half-coverings of graphs may not induce finite-index embeddings of associated geometric amalgams of free groups. Indeed, our results distinguishing commensurability classes can be used to construct examples where for certain surface amalgams X over Λ and X 0 over Λ0 , there is no surface amalgam X 00 over a common finite half-cover Λ00 so that X 00 covers both X and X 0 . Such examples exist even when Λ and Λ0 are isomorphic graphs. A priori, we see no reason why there would be a bound on the size of common half-covers so that in determining whether two groups are commensurable, it is enough to look at the set of common half-covers of their JSJ graphs up to a given size and determine whether suitable surface amalgams over these graphs exist. Finally we remark that the proofs of our necessary and sufficient conditions for cycles of generalized Θ-graphs make substantial use of Λ and Λ0 both having a distinguishable central vertex, so it is not clear how to generalize the commensurability classification to arbitrary pairs of JSJ graphs. References [1] Jason A. Behrstock and Walter D. Neumann. Quasi-isometric classification of graph manifold groups. Duke Math. J., 141(2):217–240, 2008. [2] Riccardo Benedetti and Carlo Petronio. Lectures on hyperbolic geometry. Universitext. Springer-Verlag, Berlin, 1992. [3] Brian H. Bowditch. Cut points and canonical splittings of hyperbolic groups. Acta Math., 180(2):145–186, 1998. [4] Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature, volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999. [5] John Crisp. Automorphisms and abstract commensurators of 2-dimensional Artin groups. Geom. Topol., 9:1381– 1441, 2005. 48 [6] John Crisp and Luisa Paoluzzi. Commensurability classification of a family of right-angled Coxeter groups. Proc. Amer. Math. Soc., 136(7):2343–2349, 2008. [7] Pallavi Dani and Anne Thomas. Bowditch’s JSJ tree and the quasi-isometry classification of certain Coxeter groups. arxiv:1402.6224. [8] Michael W. Davis. The geometry and topology of Coxeter groups, volume 32 of London Mathematical Society Monographs Series. Princeton University Press, Princeton, NJ, 2008. [9] Allan L. Edmonds, Ravi S. Kulkarni, and Robert E. Stong. Realizability of branched coverings of surfaces. Trans. Amer. Math. Soc., 282(2):773–790, 1984. [10] David Futer and Anne Thomas. Surface quotients of hyperbolic buildings. Int. Math. Res. Not. IMRN, (2):437– 477, 2012. [11] Rafael Guglielmetti, Matthieu Jacquemet, and Ruth Kellerhals. On commensurable hyperbolic Coxeter groups. Geom. Dedicata, 183:143–167, 2016. [12] Jingyin Huang. Commensurability of groups quasi-isometric to RAAG’s. arxiv:1603.08586. [13] Jingyin Huang. Quasi-isometry rigidity of right-angled Artin groups I: the finite out case. arxiv:1410.8512. [14] N. W. Johnson, R. Kellerhals, J. G. Ratcliffe, and S. T. Tschantz. Commensurability classes of hyperbolic Coxeter groups. Linear Algebra Appl., 345:119–147, 2002. [15] Michael Kapovich. Hyperbolic manifolds and discrete groups. Modern Birkhäuser Classics. Birkhäuser Boston, Inc., Boston, MA, 2009. Reprint of the 2001 edition. [16] Jean-François Lafont. Diagram rigidity for geometric amalgamations of free groups. J. Pure Appl. Algebra, 209(3):771–780, 2007. [17] Frank Thomson Leighton. Finite common coverings of graphs. J. Combin. Theory Ser. B, 33(3):231–238, 1982. [18] William Malone. Topics in geometric group theory. ProQuest LLC, Ann Arbor, MI, 2010. Thesis (Ph.D.)–The University of Utah. [19] Walter D. Neumann. Immersed and virtually embedded π1 -injective surfaces in graph manifolds. Algebr. Geom. Topol., 1:411–426 (electronic), 2001. [20] Walter D. Neumann. On Leighton’s graph covering theorem. Groups Geom. Dyn., 4(4):863–872, 2010. [21] Luisa Paoluzzi. The notion of commensurability in group theory and geometry. RIMS Kôkyûroku, 1836:124–137, 2013. [22] Emily Stark. Topological rigidity fails for quotients of the Davis complex. arxiv:1610.08699. [23] Emily Stark. Abstract commensurability and quasi-isometry classification of hyperbolic surface group amalgams. Geometriae Dedicata, pages 1–36, 2016. [24] Genevieve S. Walsh. Orbifolds and commensurability. In Interactions between hyperbolic geometry, quantum topology and number theory, volume 541 of Contemp. Math., pages 221–231. Amer. Math. Soc., Providence, RI, 2011. 49
4math.GR
Adapting Local Sequential Algorithms to the Distributed Setting Ken-Ichi Kawarabayashi∗ Gregory Schwartzman∗ arXiv:1711.10155v2 [cs.DC] 19 Feb 2018 Abstract It is a well known fact that sequential algorithms which exhibit a strong ”local” nature can be adapted to the distributed setting given a legal graph coloring. The running time of the distributed algorithm will then be at least the number of colors. Surprisingly, this well known idea was never formally stated as a unified framework. In this paper we aim to define a robust family of local sequential algorithms which can be easily adapted to the distributed setting. We then develop new tools to further enhance these algorithms, achieving state of the art results for fundamental problems. We define a simple class of greedy-like algorithms which we call orderless-local algorithms. We show that given a legal c-coloring of the graph, every algorithm in this family can be converted into a distributed algorithm running in O(c) communication rounds in the CONGEST model. We show that this family is indeed robust as both the method of conditional expectations and the unconstrained submodular maximization algorithm of Buchbinder et al. [BFNS15] can be expressed as orderless-local algorithms for local utility functions — Utility functions which have a strong local nature to them. We use the above algorithms as a base for new distributed approximation algorithms for the weighted variants of some fundamental problems: Max k-Cut, Max-DiCut, Max 2-SAT and correlation clustering. We develop algorithms which have the same approximation guarantees as their sequential counterparts, up to a constant additive ǫ factor, while achieving an O(log ∗ n) running time for deterministic algorithms and O(ǫ−1 ) running time for randomized ones. This improves exponentially upon the currently best known algorithms. National Institute of Informatics, Tokyo, Japan. {k keniti, greg}@nii.ac.jp This work was supported by JST ERATO Grant Number JPMJER1201, Japan. ∗ 1 1.1 Introduction Definitions and motivation A large part of research in the distributed environment aims to develop fast distributed algorithms for problems which have already been studied in the sequential setting. Ideally, we would like to use the power of the distributed environment to achieve a substantial improvement in the running time over the sequential algorithm, and indeed, for many problems distributed algorithms achieve an exponential improvement over the sequential case. One approach to designing distributed algorithms is using the sequential algorithm as natural staring point [CLS17, BCS17, BCGS17, GHS83, BS07], then certain adjustments are made for the distributed environment in order to achieve a faster running time. In this paper we focus our attention on approximation algorithms for unconstrained optimization problems on graphs. We are given some graph G(V, E), where each vertex v is assigned a variable Xv taking values in some set A. We aim to maximize some utility function f over these variables (For a formal definition see Section 2). Our distributed model is the CONGEST model of distributed computation, where the network is represented by a graph, s.t nodes are computational units and edges are communication links. Nodes communicate in synchronous communication rounds, where at each round a node sends and receives messages from all of its neighbors. In the CONGEST model the size of messages sent between nodes is limited to O(log n) bits. This is more restrictive than the LOCAL model, where message size is unbounded. Our complexity measure is the number of communication rounds of the algorithm. Adapting a sequential algorithm of the type we describe above to the distributed setting, means we wish each node v in the communication graph to output an assignment to Xv such that the approximation guarantee is close to that of the sequential algorithm, while minimizing the number of communication rounds of the distributed algorithm. Our goal is to formally define a family of sequential algorithms which can be easily converted to distributed algorithms, and then develop tools to allow these algorithms to run exponentially faster, while achieving almost the same approximation ratio up. To achieve this we focus our attention on a family of sequential algorithms which exhibit a very strong local nature. Finally, we apply our techniques to the weighted variants of the classical problems of Max k-Cut, Max-DiCut, Max 2-SAT and correlation clustering, achieving an exponential improvement over the current state of the art distributed algorithms [CLS17], while losing only an additive ǫ-factor in the approximation ratio. 1.2 Tools and results We define a family of utility functions, which we call local utility functions (Formally defined in Section 2). We say that a utility function f is a local utility function, if the change to the value of the function upon setting one variable Xv can be computed locally. Intuitively, while optimizing a general utility function in the distributed setting might be difficult for global functions, the local nature of the family of local utility functions makes it a perfect candidate. We focus on adapting a large family of, potentially randomized, local algorithms to the distributed setting. We consider orderless-local algorithms - algorithms that can traverse the variables in any order and in each iteration apply some local function to decide the value of the variable. By local we mean that the decision only depends on the local environment of the node in the graph, the variables of nodes adjacent to that variable and some randomness only used by that node. This 1 is similar to the family of Priority algorithms first defined in [BNR02]. The goal of [BNR02] was to formally define the notion of a greedy algorithm, and then to explore the limits of these algorithms. Our definition is similar (and can be expressed as a special case of priority algorithms), but the goal is different. While [BNR02] aims to prove lower bounds, we provide some sufficient conditions that allow us to easily transform local sequential algorithms into fast distributed algorithms. Our definitions are also similar to the S-LOCAL model [GKM17], which also shows that sequential algorithms which traverse the graph vertices in any order and make local decisions can be adapted to the distributed LOCAL model in poly logarithmic rounds using randomization. While the results of [GKM17] are much more broad, our transformation does not require any randomization and works in the CONGEST model. Finally, we should also mention the field of local computation algorithms [RTVX11] whose aim is developing efficient local sequential algorithms. We refer the reader to an excellent survey by Levi and Medina [LM17]. One might expect that due to the locality of this family of algorithms it can be distributed if the graph is provided with a legal coloring. The distributed algorithm goes over the color classes one after another and executes all nodes in the color class simultaneously. This solves any conflicts that may occur form executing two neighboring nodes, while the orderless property guarantees that this execution is valid. In a sense this argument was already used for specific algorithm (Coloring to MIS [Lin92], MaxIS of [BCGS17], Max-Cut of [CLS17]). We provide a more general result, using this classical argument. Specifically, we show that given a legal c-coloring, any orderless-local algorithm can be distributed in O(c)-communication rounds in the CONGEST model. To show that this definition is indeed robust, we show two general applications. The first is adapting the method of conditional expectations (Formally defined in Section 2) to the distributed setting. This method is inherently sequential, but we show that if the utility function optimized is a local utility function, then the algorithm is an orderless-local algorithm. A classical application of this technique is for Max k-cut, where an (1 − 1/k)-approximation is achieved when every node chooses a cut side at random. This can be derandomized using the method of conditional expectations, and adapted to the distributed setting, as the cut function is a local utility function. We note that the same exact approach results in a (1/2−ǫ)-approximation for max-agree correlation clustering on general graphs (see Section 2 for a definition). Because the tools used for Max-Cut directly translate to correlation clustering, we focus on Max-Cut for the rest of the paper, and only mention correlation clustering at the very end. The second application is the unconstrained submodular maximization algorithms of [BFNS15], where a deterministic 1/3-approximation and a randomized expected 1/2-approximation algorithms are presented. We show that both are orderless-local algorithms when provided with a local utility function. This can be applied to the problem of Max-DiCut, as it is an unconstrained submodular function, and also a local utility function. The algorithms of [BFNS15] were already adapted to the distributed setting for the specific problem of Max-DiCut by [CLS17] using similar ideas. The main benefit of our definition is the convenience and generality of adapting these algorithms without the need to consider their analysis or reprove correctness. We conclude that the family of orderless-local algorithms indeed contains robust algorithms for fundamental problems, and especially the method of conditional expectations, which has no analogue in the distributed environment to the best of our knowledge. Next, we wish to consider the running time of these algorithms. Recall that we expressed the running time of orderless local algorithms in terms of the colors of some legal coloring for the graph. For a general graph, we cannot hope for a legal coloring using less than ∆ + 1, where ∆ is the maximum degree in the graph. This means that using the distributed version of an orderless-local 2 algorithm as is will have a running time linear in ∆. We show how to overcome this obstacle for Max k-Cut and Max-DiCut. The general idea is to compute a defective coloring of the graph which uses few colors, drop all monochromatic edges, and call the algorithm for the new graph which now has a legal coloring. A key tool in our algorithms is a new type of defective coloring we call a weighted ǫ-defective coloring. The classical defective coloring allows each vertex to have at most d monochromatic edges, for some defect parameter d. We consider positively edge weighted graphs and require a weighted fractional defect - for every vertex the total weight of monochromatic edges is at most an ǫ-fraction of the total weight of all edges. We show that a weighted ǫ-defective coloring using O(ǫ−2 ) colors can be computed deterministically in O(log∗ n) rounds using the defective coloring algorithm of [Kuh09]. The classical algorithm of Kuhn was found useful in the adaptation of sequential algorithms to the distributed setting [FGK17, GKM17], thus its effectiveness for weighted ǫ-defective coloring might be of further use. Although we cannot guarantee a legal coloring with a small number of colors for any graph G(V, E, w), we may remove some subset of E which will result in a new graph G′ with a low chromatic number. We wish to do so while not decreasing the total sum of edge weights in G′ , which we prove guarantees the approximation will only be mildly affected for our cut problems. Formally, we show that if we only decrease the total edge weight by an ǫ-fraction, we will incur an additive ǫ-loss in the approximation ratio of the cut algorithms for G. For the randomized algorithm this is easy, simply color each vertex randomly with a color in [⌈ǫ−1 ⌉] and drop all monochromatic edges. For the deterministic case, we execute our weighted ǫ-defective coloring algorithm, and then remove all monochromatic edges. We then execute the relevant cut algorithm on the resulting graph G′ which now has a legal coloring, using a small number of colors. The above results in extremely fast approximation algorithms for weighted Max k-Cut and weighted Max-DiCut, while having almost the same approximation ratio as their sequential counterpart. Finally, our techniques can also be applied to the problem of weighted Max 2-SAT. To do so we may use the randomized expected 3/4-approximation algorithm presented in [PSWvZ17]. It is based on the algorithm of [BFNS15], and thus is almost identical to the unconstrained submodular maximization algorithm. Because the techniques we use are very similar to the above, we defer the entire proof to the appendix (Section C). We state our results in Table 1. Problem Weighted Weighted Weighted Weighted Weighted Our Approx. Our Time Prev Approx. Correlation-Clustering* Max k-Cut Max-Dicut Max-Dicut Max 2-SAT Prev Time Notes ∗ 1/2 − ǫ 1 − 1/k − ǫ 1/3 − ǫ 1/2 − ǫ 3/4 − ǫ O(log n) det. ∗ ∗ † O(log n) 1/2 [CLS17] Õ(∆ + log n) det. O(log∗ n) 1/3 [CLS17]† Õ(∆ + log∗ n) det. O(ǫ−1 ) 1/2 [CLS17]† Õ(∆ + log∗ n) rand. −1 O(ǫ ) rand. Table 1: Summary of our results for the CONGEST model (Õ hides factors polylogarithmic in ∆). (*) General graphs, max-agree (†) Unweighted graphs, only Max-Cut (k = 2). 1.3 Previous research Cut problems: An excellent overview of the Max-Cut and Max-DiCut problems appears in [CLS17], which we follow in this section. Computing Max-Cut exactly is NP-hard as shown by Karp 3 [Kar72] for the weighted version, and by [GJS76] for the unweighted case. As for approximations, it is impossible to improve upon a 16/17-approximation for Max-Cut and a 12/13-approximation for Max-DiCut unless P = N P [TSSW00, Hås01]. If every node chooses a cut side randomly, an expected 1/2-approximation for Max-Cut, a 1/4-approximation for Max-DiCut and a (1 − 1/k)approximation is achieved. This can be derandomized using the method of conditional expectations. In the breakthrough paper of Goemans and Williamson [GW95] a 0.878-approximation is achieved using semidefinite programming. This is optimal under the unique games conjecture [KKMO07]. In the same paper a 0.796-approximation for Max-DiCut was presented. This was later improved to 0.863 in [MatuuraM01]. Other results using different techniques are presented in [KS11, Tre12]. In the distributed setting the problem has not received much attention. A node may choose a cut side at random, achieving the same guarantees as above in constant time. In [HRSS14] a distributed √ algorithm for d-regular triangle free graphs which achieves a (1/2 + 0.28125/ d)-approximation ratio in a single communication round is presented. The only results for general graphs in the distributed setting is due to [CLS17]. In the CONGEST model they present a deterministic 1/2approximation for Max-Cut, a deterministic 1/3-approximation for Max-DiCut, and a randomized expected 1/2 approximation for Max-DiCut running in Õ(∆ + log∗ n) communication rounds. The results for Max-DiCut follow from adapting the unconstrained submodular maximization algorithm of [BFNS15] to the distributed setting. Better results are presented for the LOCAL model; we refer the reader to [CLS17] for the full details. Max 2-SAT: The decision version of Max 2-SAT is NP-complete [GJS76], and there exist several approximation algorithms [GW95,FG95,LLZ02,MM01], of which currently the best known approximation ratio is 0.9401 [LLZ02]. In [Aus07] it is shown that assuming the unique games conjecture, the approximation factor of [LLZ02] cannot be improved. Assuming only that P 6= N P it cannot be approximated to within a 21/22-factor [Hås01]. To the best of our knowledge the problem of Max 2-SAT (or Max-SAT) was not studied in the distributed model. Correlation clustering: An excellent overview of correlation clustering (see Section 2 for a definition) appears in [ACG+ 15], which we follow in this section. Correlation clustering was first defined by [BBC02]. Solving the problem exactly is NP-Hard, thus we are left with designing approximation algorithms for the problem, here one can try to approximate max-agree or min-disagree. If the graph is a clique, there exists a PTAS for max-agree [BBC02, GG06], and a 2.06-approximation for max-disagree [CMSY15]. For general (even weighted) graphs there exists a 0.7666-approximation for max-agree [CGW05,Swa04], and a O(log n)-approximation for min-disagree [DEFI06]. A trivial 1/2-approximation for max-agree on general graphs can be achieved by considering putting every node in a separate cluster, then considering putting all nodes in a single cluster, and taking the more profitable of the two. In the distributed setting little is known about correlation clustering. In [CHK16] a dynamic distributed MIS algorithm is provided, it is stated that this achieves a 3-approximation for mindisagree correlation clustering as it simulates the classical algorithm of Ailon et al. [ACN08]. We note that the algorithm of Ailon et al. assumes the graph to be a clique, thus the above result is limited to complete graphs where the edges of the communication graph are taken to be the positive edges, and the non-edges are taken as the negative edges (as indeed for general graphs, the problem is APX-Hard, and difficult to approximate better than Θ(log n) [DEFI06]). We also note that using only two clusters, where each node chooses a cluster at random, guarantees an expected 1/2-approximation for max-agree on weighted general graphs. We derandomize this approach in this paper. 4 2 Preliminaries Sequential algorithms: The main goal of this paper is converting (local) sequential graph algorithms for unconstrained maximization (or minimization) to distributed graph algorithms. Let us first define formally this family of algorithms. The sequential algorithm receives as input a graph G = (V, E), we associate each vertex v ∈ V with a variable Xv taking values in some finite set A. The algorithm outputs a set of assignments X = {Xv = αv }. The goal of the algorithms is to maximize some utility function f (G, X) taking in a graph and the set of assignments and outputting some value in R. For simplicity we assume that the order of the variables in X does not affect f , so we use a set notation instead of a vector notation. We somewhat abuse notation, and when assigning a variable we write X ∪ {Xv = α}, meaning that any other assignment to Xv is removed from the set X. We also omit G as a parameter when it is clear from context. When considering randomized algorithms we assume the algorithm takes in a vector of random bits denoted by ~r. This way of representing random algorithms is identical to having the algorithm generate the random coins, and we use these two definitions interchangeably. The randomized algorithm aims to maximize the expectation of f , where the expectation is taken over the random bits of the algorithm. Max k-Cut, Max-DiCut: In this paper we provide fast distributed approximation algorithms to some fundamental problems, which we now define formally. In the Max k-Cut problem we wish to divide the vertices into k disjoint sets, such that the weight of edges between different sets is maximized. In the Max-DiCut problem the edges are directed and we wish to divide the edges into two disjoint sets, denoted A, B, such that the weight of edges directed from A to B is maximized. Max 2-SAT: In the Max 2-SAT problem we are given a set of unique weighted clauses over some set of variables, where each clause contains at most two literals. Our goal is to maximize the weight of satisfied clauses. This problem is more general than the cut problems, so we must define what it means in the distributed context. First, the variables will be node variables as defined before. Second, each node knows all of the clauses it appears in as a literal. Correlation clustering: We are given an edge weighted graph G(V, E, w), such that each edge is also assigned a value from {+, −} (referred to positive and negative edges). Given some partition, C, of the graph into disjoint clusters, we say that an edge agrees with C if it is positive and both endpoints are in the same cluster, or it is negative, and its endpoints are in different clusters. Otherwise we say it disagrees with C. We aim to find a partition C, using any number of clusters, such that the weight of edges that agree with C (agreements) is maximized (max-agree), or equivalently the weight of edges that disagree with C is minimized (min-disagree). The problem is usually expressed as an LP using edge variables, where each variable indicates whether the nodes are in the same cluster. This allows a solution to use any number of clusters. In this paper we only aim to achieve a (1/2 − ǫ)-approximation for the problem. This can be done rather simply without employing the full power of correlation clustering. Specifically, two clusters are enough for our case as we show that we can deterministically achieve (1/2 − ǫ) |E| agreements which results in the desired approximation ratio. Local utility functions: We are interested in a type of utility function which we call a local utility function. Before we continue with the definition let us define an operator on assignments X, we define Lv [X] = {Xu = αu } ∈ X | u ∈ N (v) . For convenience, when we pass Lv [X] as parameter to a function, we assume that the function also receives the local topology of v which we do not write explicitly. We say that a utility function f , as defined above, is a local utility function if for every v there exists a function gv s.t f (X ∪ {Xv = α}) − f (X ∪ {Xv = α′ }) = gv (Lv [X], α, α′ ). 5 That is, to compute the change in the utility function which is caused by changing Xv from α′ to α, we only need to know the local topology of v, and the assignment to neighboring node variables. We note that for the cut problems considered in this paper the utility functions are indeed local utility functions. This is proven in the following Lemma (proof deferred to the appendix): Lemma 1. The utility functions for Max k-Cut, Max-DiCut and max-agree correlation clustering with 2 clusters are local utility functions. Submodular functions: A family of functions that will be of interest in this paper is the family of submodular functions. A function f : 0, 1Ω → R is called set function, with ground set Ω. It is said to be submodular if for every S, T ⊆ Ω it holds that f (S) + f (T ) ≥ f (S ∪ T ) + f (S ∩ T ). The functions we are interested in have V as their ground set, thus we remain with our original notation, setting A = {0, 1} and having f take in a set of binary assignments X as a parameter. The method of conditional expectations: Next, we consider the method of conditional n expectations. Let f : A → R, and let X = (X1 , ..., Xn ) be a vector of random variables taking values in A. We wish to be consistent with the previous notation, thus we treat X as a set of assignments. If E[f (X)] ≥ β, then there is an assignment of values Z = {Xi = αi }ni=1 such that f (Z) ≥ β. We describe how to find P the vector Z. We first note that from the law of total expectation it holds that E[f (X)] = α∈A E[f (X) | X1 = α]P r[X1 = α], and therefore for at least some α ∈ A it holds that E[f (X) | X1 = α] ≥ β. We set this value to be α1 . We then repeat this process for the rest of the values in X, which results in the set Z. In order for this method to work we need it to be possible to compute the conditional expectation of f (X). Graph coloring: A c-coloring for G(V, E) is defined as a function ϕ : V → C. For simplicity we treat any set C of size c with some ordering as the set of integers [c]. This simplifies things as we can always consider ϕ(v) ± 1, which is very convenient. We say that a coloring is a legal coloring if ∀v, u s.t (v, u) ∈ E it holds that ϕ(v) 6= ϕ(u). An important tool in this paper is defective coloring. Let us fix some c-coloring function ϕ : V → [c]. We define the defect of a vertex to be the number of monochromatic edges it has. Formally, def ect(v) = size{u ∈ N (v) | ϕ(v) = ϕ(u)}. We call ϕ a ccoloring with defect d if it holds that ∀v ∈ V, def ect(v) ≤ d. A classic result by Kuhn [Kuh09] states that for all d ∈ {1, 2, ..., ∆} an O(∆2 /d2 )-coloring with defect d can be computed deterministically in O(log∗ n) rounds in the CONGEST model. In this paper we define a new kind of defective coloring which we call a weighted ǫ-defective coloring. Given a positively edge weighted graph and any coloring, for every P vertex we denote by Em (v) its monochromatic edges. Define its weighted defect as def ectw (v) = e=(u,v)∈Em (v) w(e). P We aim to find a coloring s.t the defect for every v is below ǫw(v) = ǫ v∈e w(e). We show that the algorithm of Kuhn actually computes a weighted ǫ-defective O(ǫ−2 )-coloring. We state the following theorem (As the analysis is rather similar to the original analysis of Kuhn, the proof is deferred to Section A in the Appendix): Theorem 2. For any constant ǫ ∈ (0, 1/e) a weighted ǫ-defective O(ǫ−2 )-coloring can be computed deterministically in O(log∗ n) rounds in the CONGEST model. 3 Orderless-local algorithms Next we turn our attention to a large family of (potentially randomized) greedy algorithms. We limit ourselves to graph algorithms s.t every node v has a variable Xv taking values in some set A. We aim to maximize some global utility function f (X). We focus on a class of algorithms we call 6 orderless-local algorithms. These are greedy-like algorithms which may traverse the vertices in any order, and at each step decide upon a value for Xv . This decision is local, meaning that it only depends on the 1-hop topology of v and the values of neighboring variables. The decision may be random, but each variable has its own random bits, keeping the decision process local. The code for a generic algorithm of this family is given in Algorithm 1. The algorithm first initiates the vertex variables. Next it traverses the variables in some order π : V → [n]. Each Xvi is assigned a value according to some function decide, which only depends on Lv [X] and some random bits ~ri which are only used to set the value for that variable. Finally the assignment to the variables is returned. We are guaranteed that the expected value of f is at least β(G) for any ordering π of the variables. Formally, E~r [f (OL(G, ~r, π))] ≥ β(G). Algorithm 1: OL(G, ~r, π) 1 2 3 4 ∀v ∈ V, Xv = init(Lv [X]) Order the variables according to π: v1 , v2 ..., vn for i from 1 to n do Xvi = decide(Lv [X], ri ) Return X We show that this family of algorithms can be easily distributed using coloring, s.t the running time of the distributed version depends on the number of colors. The distributed version, OLDist, is presented as Algorithm 4 in the appendix. The variables are all initiated as in the sequential version, and then the color classes are executed sequentially, while in each color class the nodes execute decide simultaneously, and send the newly assigned value to all neighbors. Decide does not communicate with the neighbors, so the algorithm finishes in O(c) rounds. It is easy to see that given the same randomness both the sequential and distributed algorithms output the same result, this is because all decisions of the distributed algorithm only depend on the 1-hop environment of a vertex, and we are provided with a legal coloring. Thus, one round of the distributed algorithm is equivalent to many steps of the sequential algorithm. We prove the following lemma (the proof is deferred to the appendix): Lemma 3. For any graph G with a legal coloring ϕ, there exists an order π on the variables s.t it holds that OL(G, ~r, π) = OLDist(G, ~r, π) for any ~r. Finally we show that for any graph G with a legal coloring ϕ, it holds that E~r [f (OLDist(G, ~r, ϕ))] ≥ β(G). We know from Lemma 3 that for any coloring ϕ there exists an ordering π s.t OL(G, ~r, π) = OLDist(G, ~r, ϕ) for any ~r. TheP proof is direct from here: E~r [f (OLDist(G, ~r, ϕ))] = P P r[~ r ]f (OLDist(G, ~ r , ϕ)) = r ]f (OL(G, ~r, π)) = E~r [f (OL(G, ~r, π))] ≥ β(G). ~ r ~ r P r[~ We conclude that any orderless-local algorithm can be distributed, achieving the same performance guarantee on f , and requiring O(c) communication rounds to finish, given a legal c-coloring. We state the following theorem: Theorem 4. Given some utility function f , any sequential orderless-local algorithm for which it holds that E~r [f (OL(G, ~r, π))] ≥ β(G), can be converted into a distributed algorithm for which it holds that E~r [f (OLDist(G, ~r, ϕ))] ≥ β(G), where ϕ is a legal c-coloring of the graph. The running time of the distributed algorithm is O(c) communication rounds. 7 3.1 Distributed derandomization We consider the method of conditional expectations in the distributed case for some local utility function f (G, X), as defined in the preliminaries. Assume that the value of every Xv is set independently at random according to some distribution on A which depends only on the local topology of v. We are guaranteed that ∀G, E[f (G, X)] ≥ β(G). Thus in the sequential setting we may use the method of conditional expectations to compute a deterministic assignment to the variables with the same guarantee. We show that because f is a local utility function, the method of conditional expectations applied on f is an orderless-local algorithm, and thus can be distributed. Initially all variables are initiated to some value ∅ ∈ / A, meaning the variable is unassigned. Let Y = {Xu = αu | u ∈ U ⊆ V } be some partial assignment to the variables. The method of conditional expectations goes over the variables in any order,  and in each iteration sets Xvi = argmaxα E[f (X) | Y, Xvi = α]. This is equivalent to argmaxα E[f (X) | Y, Xvi = α] − E[f (X) | Y ] , as the subtracted term is just a constant. With this in mind, we present the pseudo code for the method of conditional expectations in Algorithm 2. Algorithm 2: CondExpSeq(G) 1 2 3 4 ∀v ∈ V, Xv = ∅ Order the variables according to any order: v1 , v2 ..., vn for i from 1 to n do Xvi = argmaxα E[f (X) | Y, Xvi = αv ] − E[f (X) | Y ] Return X To show that Algorithm 2 is an orderless-local algorithm we only need to show that argmaxα E[f (X) | Y, Xv = αv ] − E[f (X) | Y ] can be computed locally for any v. We state the following lemma (proof deferred to the appendix), followed by the main theorem for this section. Lemma 5. The value argmaxα E[f (X) | Y, Xv = αv ] − E[f (X) | Y ] can be computed locally. Theorem 6. Let G be any graph and f a local utility function for which it holds that E[f (X)] ≥ β, where the random assignments to the variables are independent of each other, and depend only on the local topology of the node. There exists a distributed algorithm achieving the same expected value for f , running in O(c) communication rounds in the CONGEST model, given a legal c-coloring. 3.2 Submodular Maximization In this section we consider both the deterministic and randomized algorithms of [BFNS15], achieving 1/3 and 1/2 approximation ratios for unconstrained submodular maximization. We show that both can be expressed as orderless-local algorithms for any local utility function. As the deterministic and randomized algorithms of [BFNS15] are almost identical, we focus on the randomized algorithm achieving a 1/2-approximation in expectation (Algorithm 3), as it is a bit more involved (The deterministic algorithm appears as Algorithm 5 in the appendix). The algorithms of [BFNS15] are defined for any submodular function, but as we are interested only in the case where the ground set is V , we will present it as such. The algorithm maintains two variable assignment Zi , Yi , initially Z0 = {Xv = 0 | v ∈ V }, Y0 = {Xv = 1 | v ∈ V }. It iterates over the variables in any order, at each iteration it considers two nonnegative quantities ai , bi . These quantities represent the gain of either setting Xvi = 1 in Zi−1 or setting Xvi = 0 in Yi−1 . Next a coin is flipped with probability p = ai /(ai + bi ), if ai = bi = 0 8 we set p = 1. If we get heads we set Xvi = 1 in Zi and otherwise we set it to 0 in Yi . When the algorithm ends it holds that Zn = Yn , and this is our solution. The deterministic algorithm is almost identical, only that it allows ai , bi to take negative values, and instead of flipping a coin it makes the decision greedily by comparing ai , bi . Algorithm 3: rand-usm(f ) 1 2 3 4 5 6 7 8 9 Z0 = {Xv = 0 | v ∈ V }, Y0 = {Xv = 1 | v ∈ V } Order the variables in any order v1 , ..., vn for i from 1 to n do ai = max {f (Zi−1 ∪ {Xvi = 1}) − f (Zi−1 ), 0} bi = max {f (Yi−1 ∪ {Xvi = 0}) − f (Yi−1 ), 0} if ai + bi = 0 then p = 1 else p = ai /(ai + bi ) Yi = Yi−1 , Zi = Zi−1 Flip a coin with probability p, if heads Zi = Zi ∪ {Xvi = 1}, else Yi = Yi ∪ {Xvi = 0} return Zn We first note that the algorithm does not directly fit into our mold, as each vertex has two variables. We can overcome this, by taking Xv to be a binary tuple, the first coordinate stores its value for Zi , and the other for Yi . Initially it holds that ∀v ∈ V, Xv = (0, 1), and our final goal function will only take the first coordinate of the variable. We note that because f is a local utility function the values ai , bi can be computed locally, this results directly from the definition of a local utility function, as we are interested in the change in f caused by flipping a single variable. Now we may rewrite the algorithm as an orderless-local algorithm, the pseudocode appears in the appendix as Algorithm 6. Using Theorem 4 we state our main result: Theorem 7. For any graph G and a local unconstrained submodular function f with V as its ground set, there exists a randomized distributed 1/2-approximation, and a deterministic 1/3approximation algorithms running in O(c) communication rounds in the CONGEST model, given a legal c-coloring. 3.3 Fast approximations for cut functions Using the results of the previous sections we can provide fast and simple approximation algorithms for Max-DiCut and Max k-Cut. Lemma 1 guarantees that the utility functions for these problems are indeed local utility functions. For Max-DiCut we use the algorithms of Buchbinder et al., as this is an unconstrained submodular function. For Max k-Cut each node choosing a side uniformly at random achieves a (1 − 1/k) approximation, thus we use the results of Section 3.1. Theorem 7 and Theorem 6 immediately guarantee distributed algorithms, running in O(c) communication rounds given a legal c-coloring. Denote by Cut(G, ϕ) one of the cut algorithms guaranteed by Theorem 7 or Theorem 6. We present two algorithms, approxCutDet, a deterministic algorithm to be used when Cut(G, ϕ) is deterministic (Algorithm 8 in the appendix), and, approxCutRand, a randomized algorithm (Algorithm 9 in the appendix) for the case when Cut(G, ϕ) is randomized. approxCutDet works by coloring the graph G using a weighted ǫ-defective coloring and then defining a new graph G′ by dropping all of the monochromatic edges. This means that the coloring is a legal coloring for G′ . 9 Finally we call one of the deterministic cut functions. approxCutRand is identical, apart from the fact that nodes choose a color uniformly at random from [⌈ǫ−1 ⌉]. For approxCutDet, the running time of the coloring is O(log∗ n) rounds, returning a weighted ǫ-defective O(ǫ−2 )-coloring. The running time of the cut algorithms is the number of colors, thus the total running time of the algorithm is O(ǫ−2 + log∗ n) rounds. Using the same reasoning, the running time of approxCutRand is O(ǫ−1 ). It is only left to prove the approximation ratio. We prove the following lemma: Lemma 8. Let G(V, E, w) be any graph, and let G′ (V,PE ′ , w) be a graph resulting from removing any subset of edges from G of total weight at most ǫ e∈E w(e). Then for any constant p, any p-approximation for Max-DiCut or Max k-Cut for G′ is a p(1 − 4ǫ)-approximation for G. Proof. Let OP T, OP T ′ be the size of optimal solutions for G, G′ . It holds that OP T ′ P ≥ OP T − P ǫ e∈E w(e), as any solution for G is also a solution for G′ whose value differs by at most ǫ e∈E w(e) (the weight of discarded P edges). Assigning every node a cut side uniformly at random the expected cut weight is at least e∈E P w(e)/4 for Max-DiCut and Max k-Cut. Using the probabilistic method this implies that OP T ≥ e∈E w(e)/4. Using all of the above P we can say that given a p-approximate solution for OP T ′ it holds that: p · OP T ′ ≥ p(OP T − ǫ e∈E w(e)) ≥ p(OP T − 4ǫOP T ) = p(1 − 4ǫ)OP T Lemma 8 immediately guarantees the approximation ratio for the deterministic algorithm. As for the randomized algorithm, let the random variable δ be the fraction of edges removed, let p be the approximation ratio guaranteed by one of the cut algorithms and let ρ be the approximation ratio achieved by approxCutRand. We know that Eρ [ρ | δ] = p(1 − 4δ). Applying the law of total expectations we get that E[ρ] = Eδ [Eρ [ρ | δ]] = Eδ [p(1 − 4δ)] = p(1 − 4ǫ). We state our main theorems for this section. Theorem 9. There exists a deterministic (1 − 1/k − ǫ)-approximation algorithms for Weighted Max k-Cut running in O(log∗ n) communication rounds in the CONGEST model. Theorem 10. There exists a deterministic (1/3 − ǫ)-approximation algorithm for Weighted MaxDiCut running in O(log∗ n) communication rounds in the CONGEST model. Theorem 11. There exists a randomized distributed expected (1/2−ǫ)-approximation for Weighted Max-DiCut running in O(ǫ−1 ) communication rounds in the CONGEST model. Correlation clustering We note the the same techniques used for Max-Cut work directly for max-agree correlation clustering on general graphs. Specifically, if we divide the nodes into two clusters, s.t each node selectes a cluster uniformly at random, each edge hasPexactly probability 1/2 to agree with the clustering, thus the expected value of the clustering is e∈E w(e)/2, which is a 1/2-approximation. The above can be derandomized exactly in the same manner as Max-Cut, meaning this is an orderless local algorithm. Finally, we apply the weighted ǫ-defective coloring algorithm twice (note that we ignore the sign of the edge), discard all monochromatic edges and execute the deterministic algorithm guaranteed from Theorem 6 with a legal coloring. Because there P must exists a clustering which has a value at least e∈E w(e)/2, a lemma identical to Lemma 8 can be proved and hence we are done. We state the following theorem: Theorem 12. There exists a deterministic (1/2 − ǫ)-approximation algorithms for weighted maxagree correlation clustering on general graphs, running in O(log∗ n) communication rounds in the CONGEST model. 10 Acknowledgments: We would like to thank Ami Paz and Seri Khoury for many fruitful discussions and useful advice. References [ACG+ 15] Kook Jin Ahn, Graham Cormode, Sudipto Guha, Andrew McGregor, and Anthony Wirth. Correlation clustering in data streams. In ICML, volume 37 of JMLR Workshop and Conference Proceedings, pages 2237–2246. JMLR.org, 2015. [ACN08] Nir Ailon, Moses Charikar, and Alantha Newman. Aggregating inconsistent information: Ranking and clustering. J. ACM, 55(5):23:1–23:27, 2008. [Aus07] Per Austrin. Balanced max 2-sat might not be the hardest. In STOC, pages 189–197. ACM, 2007. [BBC02] Nikhil Bansal, Avrim Blum, and Shuchi Chawla. Correlation clustering. In FOCS, page 238. IEEE Computer Society, 2002. [BCGS17] Reuven Bar-Yehuda, Keren Censor-Hillel, Mohsen Ghaffari, and Gregory Schwartzman. Distributed approximation of maximum independent set and maximum matching. In PODC, pages 165–174. ACM, 2017. [BCS17] Reuven Bar-Yehuda, Keren Censor-Hillel, and Gregory Schwartzman. A distributed (2 + ǫ)-approximation for vertex cover in o(log ∆ / ǫ log log ∆) rounds. J. ACM, 64(3):23:1–23:11, 2017. [BFNS15] Niv Buchbinder, Moran Feldman, Joseph Naor, and Roy Schwartz. A tight linear time (1/2)-approximation for unconstrained submodular maximization. SIAM J. Comput., 44(5):1384–1402, 2015. [BNR02] Allan Borodin, Morten N. Nielsen, and Charles Rackoff. (incremental) priority algorithms. In SODA, pages 752–761. ACM/SIAM, 2002. [BS07] Surender Baswana and Sandeep Sen. A simple and linear time randomized algorithm for computing sparse spanners in weighted graphs. Random Struct. Algorithms, 30(4):532–563, 2007. [CGW05] Moses Charikar, Venkatesan Guruswami, and Anthony Wirth. Clustering with qualitative information. J. Comput. Syst. Sci., 71(3):360–383, 2005. [CHK16] Keren Censor-Hillel, Elad Haramaty, and Zohar S. Karnin. Optimal dynamic distributed MIS. In PODC, pages 217–226. ACM, 2016. [CLS17] Keren Censor-Hillel, Rina Levy, and Hadas Shachnai. Fast distributed approximation for max-cut. 10718:41–56, 2017. [CMSY15] Shuchi Chawla, Konstantin Makarychev, Tselil Schramm, and Grigory Yaroslavtsev. Near optimal LP rounding algorithm for correlationclustering on complete and complete k-partite graphs. In STOC, pages 219–228. ACM, 2015. 11 [DEFI06] Erik D. Demaine, Dotan Emanuel, Amos Fiat, and Nicole Immorlica. Correlation clustering in general weighted graphs. Theor. Comput. Sci., 361(2-3):172–187, 2006. [FG95] Uriel Feige and Michel X. Goemans. Aproximating the value of two prover proof systems, with applications to MAX 2sat and MAX DICUT. In ISTCS, pages 182–189. IEEE Computer Society, 1995. [FGK17] Manuela Fischer, Mohsen Ghaffari, and Fabian Kuhn. Deterministic distributed edgecoloring via hypergraph maximal matching. In FOCS, pages 180–191. IEEE Computer Society, 2017. [GG06] Ioannis Giotis and Venkatesan Guruswami. Correlation clustering with a fixed number of clusters. Theory of Computing, 2(13):249–266, 2006. [GHS83] Robert G. Gallager, Pierre A. Humblet, and Philip M. Spira. A distributed algorithm for minimum-weight spanning trees. ACM Trans. Program. Lang. Syst., 5(1):66–77, 1983. [GJS76] M. R. Garey, David S. Johnson, and Larry J. Stockmeyer. Some simplified np-complete graph problems. Theor. Comput. Sci., 1(3):237–267, 1976. [GKM17] Mohsen Ghaffari, Fabian Kuhn, and Yannic Maus. On the complexity of local distributed graph problems. In STOC, pages 784–797. ACM, 2017. [GW95] Michel X. Goemans and David P. Williamson. Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming. J. ACM, 42(6):1115–1145, 1995. [Hås01] Johan Håstad. Some optimal inapproximability results. J. ACM, 48(4):798–859, 2001. [HRSS14] Juho Hirvonen, Joel Rybicki, Stefan Schmid, and Jukka Suomela. Large cuts with local algorithms on triangle-free graphs. CoRR, abs/1402.2543, 2014. [Kar72] Richard M. Karp. Reducibility among combinatorial problems. In Complexity of Computer Computations, The IBM Research Symposia Series, pages 85–103. Plenum Press, New York, 1972. [KKMO07] Subhash Khot, Guy Kindler, Elchanan Mossel, and Ryan O’Donnell. Optimal inapproximability results for MAX-CUT and other 2-variable csps? SIAM J. Comput., 37(1):319–357, 2007. [KS11] Satyen Kale and C. Seshadhri. Combinatorial approximation algorithms for maxcut using random walks. In ICS, pages 367–388. Tsinghua University Press, 2011. [Kuh09] Fabian Kuhn. Weak graph colorings: distributed algorithms and applications. In SPAA, pages 138–144. ACM, 2009. [Lin92] Nathan Linial. Locality in distributed graph algorithms. SIAM J. Comput., 21(1):193– 201, 1992. 12 [LLZ02] Michael Lewin, Dror Livnat, and Uri Zwick. Improved rounding techniques for the MAX 2-sat and MAX DI-CUT problems. In IPCO, volume 2337 of Lecture Notes in Computer Science, pages 67–82. Springer, 2002. [LM17] Reut Levi and Moti Medina. A (centralized) local guide. Bulletin of EATCS, 2(122), 2017. [MM01] Shiro Matuura and Tomomi Matsui. 0.863-approximation algorithm for MAX DICUT. In RANDOM-APPROX, volume 2129 of Lecture Notes in Computer Science, pages 138–146. Springer, 2001. [PSWvZ17] Matthias Poloczek, Georg Schnitger, David P. Williamson, and Anke van Zuylen. Greedy algorithms for the maximum satisfiability problem: Simple algorithms and inapproximability bounds. SIAM J. Comput., 46(3):1029–1061, 2017. [RTVX11] Ronitt Rubinfeld, Gil Tamir, Shai Vardi, and Ning Xie. Fast local computation algorithms. In ICS, pages 223–238. Tsinghua University Press, 2011. [Swa04] Chaitanya Swamy. Correlation clustering: maximizing agreements via semidefinite programming. In SODA, pages 526–527. SIAM, 2004. [Tre12] Luca Trevisan. Max cut and the smallest eigenvalue. SIAM J. Comput., 41(6):1769– 1786, 2012. [TSSW00] Luca Trevisan, Gregory B. Sorkin, Madhu Sudan, and David P. Williamson. Gadgets, approximation, and linear programming. SIAM J. Comput., 29(6):2074–2097, 2000. 13 A Extending Kuhn’s algorithm for ǫ-defective coloring We aim to extend the algorithm of Kuhn [Kuh09] to the case of a weighted ǫ-defective coloring. Given a positively edge weighted graph and any coloring, for every vertex we denote by Em (v), Eb (v) its Pmonochromatic and bi-chromatic edges respectively. Define its weighted defect as def ectw (v) = e=(u,v)∈Em (v) w(e). We aim to find a coloring s.t the defect for every v is below P ǫw(v) = ǫ v∈e w(e). We show that Kuhn’s algorithm can be adapted to this problem. The algorithm (Algorithm 1) : Given some coloring [M ] of the graph, each iteration of Kuhn’s algorithm consists of assigning some unique function φx : A → B to every color x ∈ [M ] (We abuse notation and also Prefer by φv the function assigned to v). Then every node iterates over all α ∈ A and picks α s.t v∈e,φv (α)=φu (α) w(e) is minimized. We note that the family of functions used is the family of polynomials degree at most k over some field. Finally the vertex is assigned the color (α, φv (α)). Analysis: We state a lemma analogous to Lemma 4.1 in [Kuh09]. We P show that it is possible to assign a color to v s.t the defect of the vertex increases by at most ǫ e∈Eb (v) w(e). Lemma 13. Assume we are given an M -coloring of G. For a value k > 0 let the functions assigned to the colors be such that for any two colors x, y ∈ [M ], the functions φx , φy intersect for at most k values and s.t that |A| > kǫ−1 then the increase P in the weighted defect of every vertex v of the new coloring computed by the algorithm is at most ǫ e∈Eb (v) w(e). Proof. We note that in the proof of Lemma 4.1 in [Kuh09], it is shown that if there two neighboring nodes u, v with different colors that choose values αu , αv s.t φv (αv ) 6= φu (αu ) then after applying Algorithm 1, the edge will be bi-chromatic. We show that P for every node v there exists some value α ∈ A s.t the increase in the weighted defect is at most ǫ e∈Eb (v) w(e). Fix some node v and assume by contradiction that this is not the case, thus P for every α ∈ A it holds that the increase in weight of monochromatic edges is greater than ǫ e∈Eb (v) w(e). Let us count the total weight of monochromatic edges for all values of α in two different ways: X X X k w(e) ≥ w(e) · χφv (α)=φu (α) e∈Eb (v) = X X α∈A e=(u,v)∈Eb (v) e=(u,v)∈Eb (v) α∈A w(e) · χφv (α)=φu (α) > |A| ǫ X w(e) e∈Eb (v) Thus, we get that |A| < ǫ−1 k which is a contradiction. Next the construction of such a family of functions remains the same using polynomials. Specifically, restating Theorem 4.6 in [Kuh09] using our parameters we get that: Theorem 14. Assume we are given an M coloring of the graph. There are an explicit function φx for x ∈ [M ] and a constant CD > 0 s.t Algorithm 1 computes a CD ǫ−2 log21/ǫ M coloring s.t the weighted defect of every vertex v increases by at most ǫw(v). 14 Finally we prove a theorem equivalent to Theorem 4.9 in [Kuh09] using our parameters. The proof is almost identical to the original one, and we present it here for completeness. Theorem 2 For any constant ǫ ∈ (0, 1/e) a weighted ǫ-defective O(ǫ−2 )-coloring can be computed deterministically in O(log∗ n) rounds in the CONGEST model. Proof. Let M = [n] be the coloring induced by the unique identifiers of the√vertices. Define ǫi = ǫ0 /2T −i , ǫ0 = ǫ/2, where T is the smallest positive integer s.t ln(T ) M ≤ 8 CD /ǫ0 . We run Algorithm 1 T times with parameters ǫi in the i-th iteration. For all values of i this implies that: p p T −i 8 CD /ǫi = 8 CD ǫ−1 ≤ 2T −i ln(T −1) M ≤ ln(i−1) M, (1) 0 2 where the first inequality is due to the definition of T (that is, for T − 1 the inequality holds in the other direction), and the second is due to the fact that for all x > 0 it holds that 2 ln x < x. Using induction we show that for all i ≤ T we have that (i) 2 Mi ≤ 16CD (ǫ−1 i ln M ) . For i = 1 the claim holds by Theorem 14 (and because ǫ−1 0 > e). For the remaining values of i, 2 due to Theorem 14 it holds that Mi ≤ CD (ǫ−1 · log M i−1 ) . Thus, it is enough to show that 1/ǫi i log1/ǫi Mi−1 ≤ 4 ln(i) M . It holds that log1/ǫi Mi−1 ≤(1) ln Mi−1 (i−1) (i−1) ≤ ln 16CD (ǫ−1 M )2 = ln 16CD (2ǫ−1 M )2 i−1 ln i ln ≤(2) ln (ln(i−1) M )4 ≤ 4 ln(i) M, where the first inequality is due to the fact that ǫ−1 to inequality (1). 0 ≥ e and the third is due √ −1 (T ) Finally plugging in i = T we get that: MT ≤ 16CD ǫ0 ln M ≤ 16CD /ǫ0 · 8 CD /ǫ0 = O(ǫ−2 ). P As the the weighted ǫ-defect for every vertex v is bounded by w(v) Ti=0 ǫi ≤ ǫw(v), this completes the proof. B Omitted Pseudocodes and Proofs Algorithm 4 is the pseudo code for the distributed version of an orderless-local algorithm. Algorithm 5 is the deterministic 1/3-approximation for unconstrained submodular maximization of [BFNS15]. Algorithm 6 and Algorithm 7 are the psedocode the randomized 1/2-approximation algorithm of [BFNS15] expressed as an orderless local algorithm. Algorithm 8 and Algorithm 9 are the deterministic and randomized approximation algorithms for the cut functions. 15 Algorithm 4: OLDist(G, ~r, ϕ) 1 2 3 4 5 6 ∀v ∈ V, Xv = init(Lv [X]) for i from 1 to c do foreach v s.t ϕ(v) = i simultaneously do Xvi = decide(Lv [X], ri ) Send Xvi to neighbors return X Algorithm 5: det-usm(f ) 1 2 3 4 5 6 7 8 9 10 11 Z0 = {Xv = 0 | v ∈ V }, Y0 = {Xv = 1 | v ∈ V } for i from 1 to n do ai = f (Zi−1 ∪ {Xvi = 1}) − f (Zi−1 ) bi = f (Yi−1 ∪ {Xvi = 0}) − f (Yi−1 ) if ai ≥ bi then Zi = Zi−1 ∪ {Xvi = 1} Yi = Yi−1 else Zi = Zi−1 Yi = Yi−1 ∪ {Xvi = 0} return Zn Algorithm 6: rand-usm(G, ~r, π) 4 ∀v ∈ V, Xv = (0, 1) Order the vertices according to π for i from 1 to n do Xu = decide(Lv [X], ~ri ) 5 return Xn 1 2 3 16 Algorithm 7: decide(Lv [X], r)  1 Z = Xu = αu,1 | {Xu = (αu,1 , αu,2 )} ∈ Lv [X]  2 Y = Xu = αu,2 | {Xu = (αu,1 , αu,2 )} ∈ Lv [X] 3 a = max {gv (Z, 0, 1), 0} 4 b = max {gv (Z, 1, 0), 0} 5 if a + b = 0 then 6 p=1 7 8 9 10 11 12 13 else p = a/(a + b) Flip coin with probability p if heads then return (1,1) else return (0,0) Algorithm 8: approxCutDet(G, ǫ) 1 2 3 ϕ = epsilonColor(G, ǫ) Let G′ = (V, E ′ = {(v, u) ∈ E | ϕ(v) 6= ϕ(u)}) Cut(G′ , ϕ) Algorithm 9: approxCutRand(G, ǫ) 1 2 3 Each vertex v chooses ϕ(v) uniformly at random from [⌈ǫ−1 ⌉] Let G′ = (V, E ′ = {(v, u) ∈ E | ϕ(v) 6= ϕ(u)}) Cut(G′ , ϕ) 17 Lemma 1 The utility functions for Max k-Cut, Max-DiCut and max-agree correlation clustering with 2 clusters are local utility functions. P Proof. The utility functions for Max k-Cut is given by f (X) = e=(w,u)∈E w(e) · Xw ⊕ Xu where Xw ⊕ Xu = 0 if Xw = Xu and 1 otherwise. Thus, if we fix some v it holds that  f (X ∪ Xv = α′ ) − f (X ∪ {Xv = α}) X X = w(e) · α′ ⊕ Xu − w(e) · α ⊕ Xu e=(v,u)∈E = X e=(v,u)∈E e=(v,u)∈E ′ w(e) · (α ⊕ Xu − α ⊕ Xu ) = gv (Lv [X], α′ , α) Note that the sum now only depends on vertices u ∈ N (v), and we are P done. For the problem of Max-DiCut the utility functions is given by f (X) = e=(v→u)∈E w(e) · Xv ∧ (1 − Xu ),Pand for max-agree correlation clustering with 2 clusters the utility function is given by P f (X) = e=(v,u)∈E + w(e) · (1 − Xv ⊕ Xu ) + e=(v,u)∈E − w(e) · Xv ⊕ Xu (E + , E − are the positive and negative edges, respectively), and the proof is exactly the same. Lemma 3 For any graph G with a legal coloring ϕ, there exists an order π on the variables s.t it holds that OL(G, ~r, π) = OLDist(G, ~r, π) for any ~r. Proof. We prove the claim by induction on the executions of color classes by the distributed algorithm. We note that the execution of the distributed algorithm defines an order on the variables.  k Let us consider the i-th color class. Let us denote these variables as Xvj j=1 , assigning some arbitrary order within the class. The ordering we analyze for the sequential algorithm would be π(vj ) = (ϕ(v), j). Now both the distributed and sequential algorithms follow the same order of color classes, thus we allow ourselves to talk about the sequential algorithm finishing an execution of a color class. Let Yi be the assignments to all variables of the distributed algorithm after the i-th color class finishes execution. And let Yi′ be the assignments made by the sequential algorithm following π until all variable in the i-th color class are assigned. Both algorithms initiate the variables identically, ′ . The coloring is legal, so for any so it holds that Y0 = Y0′ . Assume that it holds that Yi−1 = Yi−1 Xu , Xv , s.t ϕ(u) = ϕ(v) = i it holds that N (v) ∩ u = ∅. Thus, when assigning v, its neighborhood is not affected by any other assignments done in the color class, so the randomness is identical for both algorithms, and using the induction hypothesis all assignments up until this color class were identical. Thus, for all variables in this color class decide will be executed with the same parameters for both the distributed and sequential algorithms, and all assignments will be identical. Lemma 5 The value argmaxα E[f (X) | Y, Xv = αv ] − E[f (X) | Y ] can be computed locally. Proof. It holds that: E[f (X) | Y, Xv = αv ] − E[f (X) | Y ] X X E[f (X) | Y, Xv = α]P r[Xv = α] = E[f (X) | Y, Xv = αv ]P r[Xv = α] − α∈A α∈A = X α∈A P r[Xv = α](E[f (X) | Y, Xv = αv ] − E[f (X) | Y, Xv = α]) 18 P Where the first equality is due to the law of total expectation and the fact that α∈A P r[Xv = α] = 1. The probability of assigning Xv to some value can be computed locally, so we are only left with the difference between the expectations. To show that this is indeed a local quantity we use the definition of expectation as a weighted summation over all possible assignments to unassigned variables. Let Uv be the set of all possible assignments to unassigned variables in N (v) and let U be the set of all possible assignments to the rest of the unassigned variables. It holds that: E[f (X) | Y, Xv = αv ] − E[f (X) | Y, Xv = α] X X P r[Zv ]P r[Z]f (X ∪ Zv ∪ Z ∪ {Xv = αv }) − f (X ∪ Zv ∪ Z ∪ {Xv = α}) = Zv ∈Uv Z∈U = X X Zv ∈Uv Z∈U = X Zv ∈Uv P r[Zv ]P r[Z]gv (Lv [X ∪ Zv ∪ Z], α, αv ) = X X Zv ∈Uv Z∈U P r[Zv ]P r[Z]gv (Lv [X ∪ Zv ], α, αv ) P r[Zv ]gv (Lv [X ∪ Zv ], α, αv ), where in the first equality we use the definition of expectations and the fact that the variables are set independently of each other. Then we use the definition of a local utility function, and finally the dependence on U disappears due to the law of total probability. The final sum can be computed locally, as the probabilities for assigning variables in Zv are known and gv is local. C Max 2-SAT In this section we consider the problem of Max 2-SAT. We are given a set of weighted clauses C over the set of node variables, such that each clause contains at most two literals. We wish to find an assignment maximizing the weight of satisfied clauses. As before, we are interested in adapting a sequential algorithm to the distributed setting. The algorithm we shall adapt is the sequential algorithm presented in [PSWvZ17] which achieves a 3/4-approximation in expectation for the problem of weighted Max-SAT. It is based on the results of [BFNS15], thus it is almost identical to Algorithm 3. Before presenting the algorithm we need some preliminary definitions. We allow the node variables to take on values in {0, 1, ∅} where Xv = ∅ means that the value to Xv has not yet been assigned. We define two utility functions, the first, fT counts the weight of clauses satisfied given the assignment, and the second fF counts the weight of clauses that are unsatisfied given the assignment (all literals are false). We note that until now our functions depended on the node variables and the topology of the graph (which we omitted as a parameter), and now they also depend on the clauses. As with the topology, when we pass X as a parameter we assume we also pass all of the clauses, while when we pass Lv [X] as a parameter, we assume we also pass all of the clauses that contain Xv as a literal. For every C ∈ C we define two values, CT (X), CF (X), where CT (X) = w(C) if the clause is satisfied by X (and 0 otherwise) while CP is falsified by the assignment F (X) = w(C) if the clause P (and 0 otherwise). It holds that fT (X) = C∈C CT (X) and fF (X) = C∈C CF (X). Both functions are local utility functions. We prove the following lemma. Lemma 15. Both fT and fF are local utility functions. Proof. Let us assign each clause C to one of the variables which appears in it as a literal (perhaps there is only a single one). Denote for each variable by Cv the set of clauses assigned to it. Because this is a 2-SAT instance, it holds that if C ∈ Cv then CT (X) = CT (Lv [X]) and CF (X) = CF (Lv [X]). 19 It holds that fT (X) = P v∈V P C∈Cv CT (Lv [X]). Thus it holds that:  fT (X ∪ Xv = α′ ) − fT (X ∪ {Xv = α}) X X  CT (Lu [X] ∪ Xv = α′ ) − CT (Lu [X] ∪ {Xv = α}) = u∈N (v) C∈Cu = X X u∈N (v) C∈Cu  CT (Lv [X] ∪ Xv = α′ ) − CT (Lv [X] ∪ {Xv = α}) = gv (Lv [X, α′ , α]), where the first equality is because all clauses that do not contain Xv as a literal immediately get removed from the sum, and the final equality is because the clauses are of size at most two, thus every clause that is affected by an assignment to Xv only depends on Xv and Xu s.t u ∈ N (v). This finishes the proof for fT . The proof for fF is identical. Now we describe the algorithm of [PSWvZ17] (Algorithm 10) for the special case of weighted Max 2-SAT on graphs. Initially all variables are unassigned. The algorithm iterates over the variables in any order. Our aim in each iteration is to maximize the difference between the weight of the satisfied clauses and that of the unsatisfied clauses given the partial assignment. The algorithm calculates two quantities ti , fi , which correspond to this difference if we set the current variable to 0 or 1, given the current partial assignment. Finally, we flip a coin and set the variable to each value with a probability proportional to its benefit. If the benefit is negative, this probability is 0. Algorithm 10: rand-MaxSAT(G, C) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 X = {Xv = ∅ | v ∈ V } Define any order over the variables: Xv1 , ..., Xvn for i from 1 to n do ti = (fT (X ∪ {Xvi = 1}) − fT (X)) − (fF (X ∪ {Xvi = 1}) − fF (X)) fi = (fT (X ∪ {Xvi = 0}) − fT (X)) − (fF (X ∪ {Xvi = 0}) − fF (X)) if fi ≤ 0 then p=1 else if ti ≤ 0 then p=0 else p = ti /(ti + fi ) Flip a coin with probability p if heads then Xvi = 1 else Xvi = 0 return X Due to Lemma 15 we see that calculating ti , fi can be done locally, and thus Algorithm 10 is an orderless-local algorithm. Thus we can immidiatly state the following theorem: Theorem 16. For any graph G and a Weighted Max 2-SAT instance, there exists a randomized distributed expected 3/4-approximation algorithm running in O(c) communication rounds in the 20 CONGEST model, given a legal c-coloring. We now show that the same technique used for the Max-DiCut problem can be applied here. We consider the distributed algorithm guaranteed from Theorem 16. We note that in order to compute ti , fi for some node v we only need to know the variables in the 1-hop environment that share a clause with v. Thus, if there is an edge (u, v) ∈ E but no clause depends on both Xv , Xu we may drop that edge. Formally, if we define by H(V, EH ) the graph resulting from G by dropping all said edges, the execution of the algorithm for H is identical to that of G. And now we may use the same method as before of coloring the vertices at random with colors from [⌈ǫ−1 ⌉]. Let us ignore nodes in H which have no edges. These nodes may only contain unit clauses (clauses with a single literal), and the solution achieved by the greedy algorithm for that subset of clauses is optimal. This is because if a node only contains a single literal, the algorithm will satisfy that literal. If it contains two literals which contradict each other, any assignment satisfies one of them, which is optimal. Let us now assign every clause containing two literals to the edge between the two nodes, and any unit clause to one of the node’s edges, denote this set for every edge eP∈ EH . Next P of clauses by Ce P we define for every e ∈ EH its weight to be w(e) = C∈Ce w(C), thus e∈EH w(e) = C∈C w(c) = w(C). From here the proof follows the lines of Section 3.3. Let H ′ be the graph resulting from H after removing all monochromatic edges and all clauses associated with those edges. This results in a an expected decrease of ǫw(C) in total clause weight. We prove a lemma equivalent to Lemma 8 in Section 3.3. ′ ) be a graph resulting from removing any subset of edges from H of Lemma 17. Let H ′ (V, EH weight at most ǫw(C). Then for any constant p, any p-approximation for Max 2-SAT for H ′ is a p(1 − 2ǫ)-approximation for H. Proof. Let OP T, OP T ′ be optimal solutions for H, H ′ , respectively. We can immediately see that OP T ′ ≥ OP T − ǫw(C), as any solution for H is also a solution for H ′ whose value differs by at most ǫw(C). Using the probabilistic method it is easy to see that for Weighted Max 2-SAT it holds that OP T ≥ w(C)/2. Using all of the above we can say that given a p-approximate solution for OP T ′ it holds that: p · OP T ′ ≥ p(OP T − ǫw(C)) ≥ p(OP T − 2ǫOP T ) = p(1 − 2ǫ)OP T Finally the proof of the expected approximation value follows exactly as in Section 3.3. We state the main theorem for this section. Theorem 18. There exists a randomized distributed (3/4 − ǫ)-approximation for Weighted Max 2-SAT running in O(ǫ−1 ) communication rounds in the CONGEST model. 21
8cs.DS
1 Distribution System Outage Detection using Consumer Load and Line Flow Measurements Raffi Sevlian∗ , Yue Zhao∗‡ , Ram Rajagopal† , Andrea Goldsmith∗ , H. Vincent Poor‡ ∗ Department of Electrical Engineering, Stanford University † Department of Civil and Environmental Engineering, Stanford University ‡ Department of Electrical Engineering, Princeton University arXiv:1503.08244v2 [cs.SY] 17 Nov 2016 Abstract—An outage detection framework for power distribution networks is proposed. Given the tree structure of the distribution system, a method is developed combining the use of real-time power flow measurements on edges of the tree with load forecasts at the nodes of the tree. A maximum a posteriori detector (MAP) is formulated for arbitrary number and location of outages on trees which is shown to have an efficient detector. A framework relying on the maximum missed detection probability is used for optimal sensor placement and is solved for tree networks. Finally, a set of case studies is considered using feeder data from the Pacific Northwest National Laboratories. We show that a 10% loss in mean detection reliability network wide reduces the required sensor density by 60 % for a typical feeder if efficient use of measurements is performed. I. I NTRODUCTION Outage detection and management has been a long-standing problem in power distribution networks. Outages are caused by protective devices closing off a part of the network to automatically isolate faults. Usually, a short circuit fault will trigger this protective operation. We employ the term outage detection to denote the task of finding the status of the protective devices, and the term fault detection to denote finding the faults that caused the resulting outage situation. Many methods for outage and fault detection based on artificial intelligence have been developed. Outage detection is often performed prior to fault detection and can greatly improve the accuracy of fault diagnosis. For outage detection, fuzzy set approaches have been proposed based on customer calls and human inspection [14], and based on real-time measurement with a single sensor at the substation [1]. In networks where supervisory control and data acquisition (SCADA) systems are available, a subset of the protective devices’ status can be obtained via direct monitoring. When two-way communications from the operator and the smart meters are available, AMI polling has been proposed to enhance outage detection [8]. There have also been knowledge based systems that combine different kinds of information (customer calls, SCADA, AMI polling) [6]. For fault detection, using only a single digital transient recording device at the substation, fault location and diagnosis systems have been developed based on fault distance computation using impedance information in the distribution system [15]. Using only the outage detection results, i.e., the status of the protective relays, expert systems have been applied to locate the underlying faults [4]. Incorporating voltage measurements in the distribution system This research was supported in part by the DTRA under Grant HDTRA1-081-0010, in part by the Tomkat Center, in part by Powell Foundation Fellowship, in part by the National Science Foundation award CIF1116377, in part by the Air Force Office of Scientific Research under MURI Grant FA9550-09-1-0643, and in part by the Office of Naval Research, under Grant N00014-12-1-0767. with the outage detection results, fault detection methods based on knowledge based systems have been proposed [2]. Fault detection that uses fault voltage-sag measurements and matching has been proposed in [9], [5]. Fault diagnosis based on fuzzy systems and neural networks have also been proposed that can resolve multiple fault detection decisions [13]. Existing outage and fault detection methods based on artificial intelligence do not provide an analytical performance metric, so it is in general hard to examine their optimality. Their performance can however be evaluated numerically and in simulation studies. Moreover, because of this lack of an analytical metric, while some of the existing approaches depend on near real-time sensing (e.g. SCADA), they do not provide guidance on where to deploy the limited sensing resources within the distribution system. A major alternative to these mechanism is the so called last gasp, where area’s in outage will notify, via distress signal that they are out of power. These provide a duplicate method of outage detection which can be combined with the proposed methods here. In fact, combining both of these methods can further reduce the time to outage in practical scenarios. The proposed sensing and feedback framework exploits the combination of real-time sensing and feedback from a limited number of power flow sensors and the infrequent load updates from AMI or forecasting mechanisms. This can is practically possible since there is a growing number of deployments of distribution system line measurements [7], which can measure line current with high precision. II. P ROBLEM F ORMULATION AND M AIN C ONTRIBUTIONS Consider a power distribution network that has a tree structure. Power is supplied from the feeder at the root, and is drawn by all the downstream loads. An outage is a protective device isolating a faulted area. When this occurs, the loads downstream of the faulted area will be in outage. We investigate the optimal design and performance of automatic outage detection systems with the use of the following two types of measurements: Noisy Nodal Consumption typically in the form of forecasts which have forecast errors that must be taken into account. Error Free Edge Flows which typically come from realtime SCADA measurements of the power flows on a fraction of the lines. The issue of noisy and error free measurement comes from the fact that loads come from delayed information which needs to be forecasted, while SCADA systems have real time communication potential. This work assumes lossless power flow, but can also be applied in the case of current measurements on the line and load level. This can be done, since practical distribution line sensing 2 is accurate in terms of current measurements, and smart meter interval data provides power and voltage information, making current inference possible. The main contributions of this work are the following: Outage Detection We formulate the problem of detecting any number of possible outages via nodal and edge measurements as a general hypothesis testing problem where the number and locations of outages are unknown. We show that this general formulation results in a computationally efficient decentralized hypothesis detector. Sensor Placement We use the decentralized nature of the detector provide a optimal sensor placement with respect to the maximum missed detection error for all hypotheses. III. S YSTEM M ODEL AND N OTATION 0 1 s0 3 7 where the set Vi (H) indicate the set of vertices to be summed over under any particular hypothesis. A general representation of the observed flow is the following. The set of full flow observations s, given a particular hypothesis H ∈ H k , we can represent the observations as: 9 10 s1 (4) where ΓH ∈ {0, 1}|M|×|V | . Here ΓH is generated for each hypothesis and we assume the forecast error covariance Σ can be estimated from the load forecasting process. 8 5 6 v∈Vi (H) s = ΓH x ∀H ∈ Hk , 2 4 Measurement Model: For any edge e, denote by s the power flow on it towards all active downstream loads. The measured flow depends on the network topology, outage situation and the true loads. The sensor placement is denoted as M with M ⊂ E. The vector of all measurements is s ∈ R|M| . Given a tree T assume hypothesis H corresponds to the outage of any number of disconnected edges. The measured power consumption of the ith sensor measurement under hypothesis H ∈ Hk is X si (H) = x(v), (3) 11 IV. G ENERAL O UTAGE D ETECTION 12 13 14 16 17 18 15 Fig. 1. Example tree T1 used to illustrate various properties. Each node in the network is numbered. Node vn is connected to it’s parent via edge en consuming x(v) power at each node. Two flow measurement sensors s0 , s1 along with load pseudo measurements x̂(v). Topology of the Distribution System: The vertices in the distribution network are indexed by V = {v0 , v1 , . . . , vN }, with bus v0 denoting the root of the tree. We index by en the line that connects bus vn and its parent node. Outage Hypothesis Model: Outages are modeled as disconnected edges corresponding to protective devices disconnecting loads on a network. For example, consider single line outages in a tree with N edges: In this situation, there will exist N single edge outage hypotheses and a single non-outage situation. Let H1 = {e1 , . . . , eN ∪ ∅} be the set of all single outage hypotheses for a tree T . We consider a more general case of an unknown number and location of potential outages. We define the set of up to k edge outages Hk as the set of k edge hypothesis. This set follows: Hk = H1 × H1 . . . H1 . | {z } (1) k times Load Model: Each node v in the graph has a consumption load x(v). The forecast of each load is x̂(v) with error (v) = x(v) − x̂(v). We assume errors are mutually independent random variables that follow (v) ∼ N (0, σ(v)2 ). Given the forecasts we treat the true load, which is unknown to us, as a random variable x(v) ∼ N (x̂(v), σ 2 (v)). In the vector case, we have x̂ ∼ N (x, Σ) where we can assume Σ is a diagonal covariance matrix. (2) Consider the general outage detector. Given the vector of load forecasts, x̂, nominal forecast error Σ and real time load flows s along a set of branches, the detector must determine the correct ? number and location of each edge in outage H ∈ Hk . These are single snapshot values of load forecast and line flow. A multi period detection framework can be analyzed in a similar fashion. We first construct a simple but naive multiple hypothesis detector relying on a maximum likelihood estimator. Consider the flow model in eq. (4), relating the true load at each node to the observed flow on the network. Given the forecast model in eq. (2) and the hypothesis model in eq. (4), the Maximum a Posteriori detector is the following: {k ? , Ĥ} ∈ arg max Pr (s | x̂, H) (5) H∈Hk See Appendix B, for details. The flow likelihood can be computed as follows: s|{x̂, H} = ΓH x = ΓH (x̂ + ) ∼ N ΓH x̂, ΓH ΣΓTH  ∀H ∈ Hk (6) Eq. (6) allows us to evaluate a likelihood under each possible hypothesis. Therefore, a naive detector will enumerate every possible outage, evaluate it’s likelihood, and choose the maximum. This is difficult for the following reasons:  k 1) The set Hk is of size |E| k , so computing the set H can be very expensive. Enumerating the  entire maximum P likelihood detector requires k=1 |E| evaluations. k 2) Many of the potential hypotheses map to the same observed flows, therefore the detector output is not unique. This occurs when one edge is a descendant of another. 3) Missed detection errors in a multivariate hypothesis testing framework can only be evaluated via monte carlo testing. There is no insight in optimizing placement. 3 V. D ECOUPLED M AXIMUM L IKELIHOOD D ETECTION 0 We show that the issues regarding the general maximum likelihood detector can be overcome by decoupling the hypotheses and the observations, given the tree structure of the outage detection problem. This leads to a simple decentralized detector which is equivalent to eq. (5), where each decision is a scalar hypothesis test. This leads to an efficient hypothesis enumeration, detection and error evaluation. In the following sections, we show the following: 1) The original search space Hk can be replaced by a set Hu of uniquely detected outages, due to the tree structure of the network. 2) Processing zero/positive flow information reduces the search space from Hu to Hu+ . which decouples a V into product set of local hypotheses: Hu+ = Hu+ (A). A∈A+ Where, A indicates a local area with it’s own hypothesis set. 3) The joint likelihood function Pr(s | x̂, H) decouples along each area. Combining these results leads to a decentralized detector which can be solved easily. A. Unique Outages Maximizing the likelihood of observations over the k outage set can lead to a non-unique solution. An alternative is to only consider uniquely detectable outages, where no possible outage event is downstream of any other. Define the set of outage hypotheses Hu (u is for unique) as follows: Hu ={H ∈ H for some k, s.t no two edges (7) This definition is not constructive, but useful. Consider tree T2 shown in Figure 2(c). Here Hu can be enumerated by simple observation. Hu = {∅, e1 , e2 , e3 , e4 , e5 , (e3 × e5 ), (e4 × e5 )} 2 b3 b2 b4 b3 b2 3 b5 b5 b4 (a) (8) There is a single non-outage hypothesis ∅, and 5 single outage hypotheses, and 2 double outage hypotheses. For a more general case (Figure 1), enumerating this set for a tree can be performed recursively. The set Hu can be enumerated using a ’branch-network’ based on the original tree. The tree in Figure 1 is depicted in Figure 2(a) with nodes and edges removed which highlights the various branches of the graph. Each set of branches are aggregated as a node to be traversed, in the hypothesis enumeration procedure. Consider a set function E(b) = {e ∈ E : e is along branch bi } to enumerate the set of edges on a branch. Given the two examples, we have the following branch-edges: • T1 : E(b1 ) = {e1 , e2 , e3 }, E(b2 ) = {e4 , e5 , e6 , e7 }, E(b3 ) = {e8 , e9 , e10 , e11 , e12 }, E(b4 ) = {e13 , e14 , e15 } and E(b5 ) = {e16 , e17 , e18 }. • T2 : E(b1 ) = {e1 , e2 }, E(b2 ) = {e3 , e4 }, E(b3 ) = {e5 }. Using this definition, we propose the following recursive definition of the set Hu (b), which indicates the set of hypotheses formed from branch b and all descendants. From this definition, s 1 b1 b1 5 4 (b) b3 b2 (c) (d) Fig. 2. 2(a) Branches of tree T1 . 2(b) Branch network for tree T1 . 2(c) Simple tree network T2 . 2(d) Branch graph for T2 . it is clear that Hu ≡ Hu (b1 ), since this is the root branch of the tree.  ! [ ^ Hu (b) = E(b) ∪  (9) Hu (b)  b∈b b∈P(child(b)) Here b is the current node, and child(b) is the set of children of b in the branch tree. The set P(child(b)) is the power set of all the child branches, where any element of the power set is b. Note that if a branch has no descendants, we merely evaluate E(b), since the remaining terms are null. Eq. (9) is quite unwieldy, but can be interpreted easily. For tree T2 in Figure 2(c) the child branches are child(b1 ) = {b2 , b3 } while the power set is: P({b2 , b3 }) = { {∅} , {b2 }, {b3 }, {b2 , b3 }}. |{z} |{z} |{z} | {z } b0 b1 b2 (10) b3 Evaluating (9), we arrive at the following:     ^ ^ Hu (b1 ) = E(b1 ) ∪  Hu (b) ∪  Hu (b) b∈{b2 } b∈{∅}  k are descendant of each other.} b1  ^ ∪  Hu (b) ∪  b∈{b3 }  ^ Hu (b) (11) b∈{b2 ,b3 } We rely on the following definitions: V D1 Base case: Hu (b) = ∅. b∈{∅} D2 Hypotheses double counting: ei ∪ ei = ei , including the empty set ∅ ∪ ∅ = ∅. D3 Cross product reduction: ∅ × ei = ei , which implies E(b) × ∅ = E(b). For example, eq. (9) is applied to T2 as follows: For b1 , we have that Hu (b1 ) = {E(b1 ) ∪ Hu (b2 ) ∪ Hu (b3 ) ∪ Hu (b2 ) × Hu (b3 )}. (12) For the branches b2 and b3 we use eq. (9) and D1 resulting in Hu (b2 ) = {E(b2 ) ∪ ∅} and Hu (b3 ) = {E(b3 ) ∪ ∅}. Using D2 and D3, we have the following: This results in: Hu (b1 ) = {E(b1 ) ∪ {∅} ∪ {E(b2 ) ∪ ∅} ∪ {E(b3 ) ∪ ∅} ∪ {E(b2 ) ∪ ∅} × {E(b3 ) ∪ ∅}} = {∅ ∪ E(b1 ) ∪ {E(b2 ) ∪ E(b2 ) ∪ E(b3 ) ∪ (E(b2 ) × E(b3 ))} = {∅, e1 , e2 , e3 , e4 , e5 , (e3 × e5 ), (e4 × e5 )} Which is identical to simple enumeration. At this point, we have reduced the maximum likelihood detector to the following form: Ĥ = arg max Pr (s | x̂, H) H∈Hu (13) 4 Note the equality in the maximization since the restriction of the search space allows us to find a unique solution. Notice that this procedure will automatically enumerate all possible number of outages as well as their positions in the graph. Under the general model of any number of edge outages, this is the complete enumeration of the hypothesis set which has a high computational burden. In the following sections we show that (1) the search space decouples to a product space of local hypotheses given binary flow indicators and (2) the likelihood function decouples along these local ’areas’, thereby reducing to a decoupled scalar hypothesis test for each local area. B. Reducing Hypotheses from binary flow information A1, H+u(A1) A1 b 1 s1 s2 b2 A2 H+u(A2) s1 s1 b u3 A2 b4 (a) bl s2 H1 s2 3 H2 b5 (b) (c) Fig. 3. 3(a) The branch network for network T1 with associated sensors s1 , + + s2 and area hypotheses Hu (A1 ) and Hu (A2 ). 3(b) Each area will keep the local branches. Branch b3 is split between two areas and processed as different u vertices in branch-network (bu 3 and b3 . 3(c) The network can be modeled by a directed graphical model indicating observations (shaded) and variables that must be maximized over. The detection problem will encounter a set of positive flows as well as zero’s when a sensor is downstream of an edge outage. The ML detection observations and search space can be reduced as: Ĥ = arg max Pr ({sz s+ } | x̂, H) (14) ∀H∈Hu = arg max Pr ({s+ } | x̂, H) (15) ∀H∈H+ u By processing the flow information we reduce the set of hypothesis in the detector from the set Hu to the reduced set Hu+ . More importantly, we show that Hu+ does not require a recursive enumeration, but actually decouples as a product space of local hypotheses. This is first shown in an example, then the general form is stated. Consider T1 with branch network with flow measurements in Figure 3(b). In this example, all branches are unchanged except b3 , which is split into bu3 and bl3 and is separated by flow measurement s1 . In the case of splitting branch nodes, the upper branch will take the edge with the measurements so E(bu3 ) = {e8 , e9 , e10 } and E(bl3 ) = {e11 , e12 }. The following illustrative example highlights a general decoupling principle of the hypothesis set Hu+ . Let’s consider the two cases separately (s1 , s2 > 0 and s1 > 0, s2 = 0) which provides the intuition for the general case. 1) Case 1: (s1 > 0, s2 > 0) This implies that there cannot be any outage with edges in b1 or b3 , else s2 = 0. A brute force enumeration of Hu+ (b1 ) is done by enumerating H(b1 ) according to Figure 2(b), then removing any terms with E(b1 ) and E(b3 ) outages. It can be shown that: Hu (b1 ) = {∅ ∪ E(b1 ) ∪ E(b2 ) ∪ (E(bu3 ) ∪ Hu (bl3 )) ∪ E(b2 ) × (E(bu3 ) ∪ Hu (bl3 ))} Removing the possible outages due to the positive flow information (i.e. any tuple with edges in E(bu3 ), or E(b1 )) , we have:  Hu+ (b1 ) = {∅ ∪ E(b2 ) ∪ Hu+ (bl3 ) ∪ E(b2 ) × Hu+ (bl3 ) } = {∅1 ∪ E(b2 )} × {∅2 ∪ Hu+ (bl3 )} = Hu+ (A1 ) × Hu+ (A2 ) We use A1 and A2 to define a local area. An area is a partition of the original tree, which will decouple the set of all hypotheses into a product space of ’area hypotheses’. Each area will contain a root sensor (ex. A1 contains s1 ) and a set of descendent sensors (ex. child(s1 ) = {s2 }). Within a local area, a set of unique outage hypotheses are evaluated Hu+ (A1 ) which satisfy the binary observations of whether any flow is observed along the downstream sensors child(s1 ) > 0. The local hypotheses of each area are later combined to form any possible hypothesis from the original enumeration of unique outages. Appendix C contains a more detailed discussion. 2) Case 2: (s1 > 0, s2 = 0) This implies that all possible outages must contain b1 or bu3 , else s2 > 0. First we enumerate Hu (b1 ) then keep only those elements which lead to s2 = 0 (i.e. every tuple with edges in E(bu3 ), or E(b1 )) , we have: Hu+ (b1 ) = {∅ ∪ E(b1 ) ∪ E(bu3 ) ∪ E(b2 ) × E(bu3 )} = {Hu+ (A1 )} Note that in this case, Hu+ (A2 ) is never evaluated since s2 = 0. Different positive and zero flow patterns lead to changes in the local hypothesis set Hu+ (A). 3) General Hypothesis Decoupling: A complete description of the set should be Hu+ (A, f ), where f = I{child(si )>0} , since this set depends on the binary flow information of the child sensors. For example Case 2 is Hu+ = Hu+ (A1 , {1 0}) and Case 2 is Hu+ = Hu+ (A1 , {1 1}) × Hu+ (A2 , {1 1}). For any arbitrary tree and flow sensors, the unique hypothesis set conditional on binary flows Hu+ will decouple according to: ^ Hu+ = Hu+ (A). (16) A∈A+ + Where the set A indicates the areas which have a root measurements si > 0. Finally, Hu+ (A) is the local conditional hypothesis, which can be computed in the general case. Any given Hi ∈ Hu+ can be represented by a product of area hypotheses Hi = H1,i(1) × . . . × HM,i(M ) , where for the k th area Hk,i(k) ∈ Hu+ (Ak ). Index i(k) is the particular index into the k th hypothesis set corresponding to global hypothesis Hi . Appendix D presents a general algorithm generate the set Hu+ (A, f ), for arbitrary binary flow information. C. Decoupling the Joint Likelihood: Due to the model of noiseless flow and forecasted nodal measurements joint likelihood of all observations, given x̂, and H ∈ Hu+ , decouple across the areas. In general we have: Y Pr (s+ | x̂, H) = Pr (si | child(si ), x̂, Hi ) (17) i:Ai ∈A+ 5 The tree T1 is presented here as an example, while the general case can be easily shown as an extension of the example. Consider again the sub-graph in Figure 3(a), where we have already reduced the hypotheses using binary flow information (both s0 , s1 > 0). As discussed in Section V-B3, a unique hypothesis Hi ∈ Hu+ can be represented as Hi = H1,i(1) × H2,i(2) . Where H1,i(1) ∈ H(A1 ) and H2,i(2) ∈ H(A2 ). From eq. (3) the observed flow s1 , can be computed as follows: X s1 |{x̂, Hi } = x(v) (18) D. Decoupled Maximum Likelihood Function: We can combine the results shown so far to a decoupled likelihood function. {k ? , Ĥ} ∈ arg max Pr (s|x̂, H) = arg max Pr ({s+ sz }|x̂, H) X x(v) + v∈V (H1,i(1) ×H2,∅ )\V (H2,∅ ) X = arg max Pr (s+ |x̂, H) X = x(v) = Pr (s+ |x̂, H1 . . . HM ) Y arg max ∀A∈A+ ∀Hi ∈H+ (A) Pr (si |child(si ), x̂, Hi ) (26) ^ = Where: • Eq. (18) is the summation of true loads. By decoupling of the hypotheses across areas, we represent Hi as the product of the two local hypotheses H1,i(1) and H2,i(2) . • In (19), this can be separated as the sum of parts 1) V (H1,i(1) × H2,∅ ) \ V (H2,∅ ) are the vertices in the summation in A1 independent of what is happening of downstream areas. H2,∅ indicates the non-outage hypothesis in area 2. 2) V (H2,i(2) ) is the set of vertices in the summation for P area Area 2. Although H2,i(2) is unknown, s2 = v∈V (H2,i(2) ) x(v), the the unknown hypothesis can be eliminated. • Eq. (20), replaces the second summation with the observed flow, since they are equivalent. Therefore, if we condition on the remaining flow observation, child(s1 ), the flow decouple between different areas. Finally, since x(v), is not known, a likelihood function for the net flow in the area can be constructed given the load forecasts. The first term in eq. (20) is modeled with the following: (21) Evaluating µ(x̂, H), and σ(Σ, H) can be computed easily and is described in Section VI-B. Similarly, s2 is decoupled from any hypothesis in Hu+ (A1 ), since the measurement depends only on downstream hypotheses (assuming s2 has some positive flow to begin with). In this example we can decouple the likelihood function as follows. Pr (s1 s2 | x̂, H1 , H2 ) = Pr (s2 | x̂, H2 ) Pr (s1 |s2 x̂, H1 ) In this example, this decoupling can be represented as a simple graphical model as shown in Figure 3(c), where each local hypothesis is an unknown variable that must be determined via likelihood maximization. Conditioning on the the only observations s1 , s2 , the two hypotheses variables are independent. This graphical model formulation is used to show the general case where there may be noise in the flow measurements. In such a case, the decoupling would not work, but a message passing algorithm can be used. The graphical model formulation can also be applied in different sensor types. To show the general case, eq (18)-(20), can be extended to a general area network with multiple downstream sensors. (25) i:Ai ∈A+ (20) v∈V (H1,i(1) )\V (H2,∅ ) s1 − s2 |{s2 x̂, H} ∼ N (µ(x̂, H), σ(x̂, H)). arg max ∀A∈A+ ∀Hi ∈H+ (A) v∈V (H2,i(2) ) x(v) + s2 (24) ∀H∈H+ u (19) = (23) ∀H∈Hu v∈V (H1,i(1) ×Hi,i(2) ) = (22) ∀k,∀H∈Hk i:Ai ∈A+ arg max Pr (si |child(si ), x̂, Hi ) (27) ∀Hi ∈H+ (Ai ) Here eq. (22) is the original maximum likelihood hypothesis test over all Hk outages. This is reduced to a search space over Hu in eq. (23). This further reduces to an even smaller search space Hu+ due to processing of binary flow information in eq. (24). Eq. (25) decouples Hu+ to a product space of local search hypotheses. Eq. (26) decouples the likelihood functions by conditioning on the set of observations. This likelihood function is a product of terms which only depend on the local hypotheses. Therefore maximizing this product is equivalent to maximizing each term separately, as in eq. (27). The decoupling of the centralized likeAlgorithm 1: Maximum Likelihood Hypothesis Detector. Result: Maximum Likelihood Hypothesis Detector Input: [1] Load Forecast/Nominal Statistics: x̂, Σ [2] Real Time Load: s = {s+ , sz } + 1 A ← prune − areas(A, s) + 2 for Ai ∈ A do 3 // Generate Local Hypothesis set. 4 Hu+ (Ai ) ← local − hypotheses(T , si , d(si )) 5 // Local MAP Detector 6 Ĥi ← arg max Pr (si |d(si ), x̂, Hi ) ∀Hi ∈H+ (Ai ) 7 8 9 end // Combine Local Hypotheses V Ĥ ← Ĥi Ai ∈A+ lihood function in (26), leads to a simple decentralized detector in Algorithm 1. The input is (1) the set of load forecasts x̂ with their nominal statistics Σ, (2) and the real time load information s. The function prune-areas simply discards areas with zero flow. The function local-hypotheses performs the generation of local hypothesis set Hu+ (Ai ) as described in section D. Finally the local MAP detector is simple to evaluate as a multihypothesis test involving scalar gaussian of known means and variances as follows: For a local area, since H+ (A) is enumerated, we determine: !2 P si − s∈child(si ) s − µ(x̂, H) Ĥ = arg max σ(Σ, H) H∈H+ (A) 6 Algorithm 2: Solution to optimization (OPT-3) for tree network. Result: Placement for a Tree Network Input: [1] Tree network T s∈child(si ) [2] Nominal loads statistics x̂, Σ [3] Subproblem Ordering Vprocess for each local area. Computation of the means and variance are [4] Target error P target discussed in Section VI-B. 1 // Generate node process ordering 2 Eprocess ← generate-edge-order(T ) VI. S ENSOR P LACEMENT P ROBLEM 3 // initialize sensor placement as empty In evaluating a placement, we use the maximum missed de- 4 Mg ← ∅ tection probability over all hypotheses. This is a useful metric 5 for et ∈ Eprocess do A ← construct-area(et , Mg ) since we are considering a very large number of alternative 6 // Evaluate the current subtree maximum missed hypotheses. First consider the maximum error PE (H, M) = 7 if PEmax (A) ≤ P target then Pr(Ĥ 6= H; placement M), where Ĥ is the optimal solution 8 9 // continue to next node of the outage detection. 10 else This can be used for sensor placement evaluation. 11 if |child(vt )| == 1 then M? = arg min max PE (H, M) (OPT-1) 12 Mg ← line-action(A, Mg ) |M|=M H∈Hu 13 else if |child(vt )| > 1 then This is difficult to do outside of a combinatorial enumeration of 14 Mg ← tree-action(A, Mg ) sensor locations and hypotheses. A suitable proxy we optimize 15 instead is the following: 16 end 17 end M? = arg min max PE (A) (OPT-2) g 18 return M |M|=M A∈A Therefore each decision only depends on an effective measurement X ∆si , si − s (28) Where PEmax (A) = maxf,H∈Hu+ (A,f ) PE (H, M), that is we only search over the hypothesis in the local area. This second optimization very closely approximates an upper bound to (OPT-1) (see appendix F for details). Optimization OPT-2 is solved via a bisection method on the following feasibility problem: M? = find M |M|≤M s.t. PEmax (A) (OPT-3) ≤ P target So the minimum P target is determined which yields a solution of size |M| = M . This can be solved very efficiently, with the algorithm that follows. A. Feasibility Placement Algorithm The intuition for the greedy placement algorithm is the following. Starting from the bottom of a tree, we successively maintain a temporary local area with root sensor in et . The root sensor is iteratively moved closer to the root edge e1 , while maintaining that the maximum error of all areas is less than P target . This is done by maintaining that the local area has error less than P target . If this is true we move closer to the root, if not, we place a sensor and start a new area. Since the objective function decouples across areas, we can maintain that the feasibility problem is always satisfied. Finally, if the number of sensors are less than M , M? is returned. This framework can be realized in Algorithm 2. The inputs to the method are the tree network T and the set of nominal load forecasts x̂ and forecast variance Σ. To have the effect of starting at the leaf of the network and move our way up to the root, the algorithm will process a sequence of edges Eprocess . For example in Figure 1, Eprocess = {E(b4 ) E(b5 ) E(b3 ) E(b2 ) E(b1 )}. We generate the list Eprocess in line 2 with function generate-edge-order. The function takes the tree T and traverses via breadth first search keeping track of the depth of each vertex/edge. Reversing this list of depths yields a list of nodes to process, the parent edge being e ∈ Eprocess . The greedy solution Mg must first be initialized as empty. In line 5 we iterate over the current root node et and current sensor placement M. In line 6 we construct the current area network A. We then evaluate PEmax (A) in line 8; If PEmax (A) > P target we perform a placement action line-action or tree-action depending on the number of child nodes of vt (downstream of et ). Each sub function is described in more detail as follows. 1) construct-area: For each iteration, the temporary node et is visited and the area network A is constructed with et and the previous solution Mg . The current edge et is the temporary root sensors of the area st . The terminal sensors of the area are any sensors in Mg that are children of st . Note that this may be empty at the start of the algorithm. 2) line-action: Given that our current subproblem satisfies PEmax (A) < P target , if the next subproblem does not satisfy the condition, our only option is place a sensor at child (vt ). So Mg ← Mg ∪ et . 3) tree-action: Given that the algorithm up to now has placed sensors on the two disjoint trees with roots with v1 and v2 We must move as far up to the root as possible before we are forced to place a measurement. This leads to two different types of actions: a greedy strategy that is easy to implement and the optimal strategy. The greedy strategy is implemented in practice and is almost always equal to the optimal strategy, which is discussed in Appendix H. The greedy strategy chooses the area network with the smallest PEmax (A). For example, in Figure 4, first assume PEmax (A0 ) and PEmax (A00 ) < P target and PEmax (A1 ) > P target . In moving et closer to the root, we must choose either A2 or A3 based on the placement which has the smallest error. 7 outage detector with greedy placement is extremely efficient (only O(|E|2 ) complexity). … For the remaining case studies only single area outages are considered. … A3 … … … … A2 … … … A1 … … … A0 … … … A '0 … VII. D ISTRIBUTION S YSTEM C ASE S TUDY A. PNNL Case Study Fig. 4. (greedy-tree-action) Correct node traversal in the tree network assumes max (A ) < P target and P max (A0 ) < P target . If P max (A ) < P target we that PE 0 1 0 E E do nothing. Else, we generate and evaluate the error one of the 2|child(vt )|−1 0 00 area networks that can be constructed for example A1 and A1 . R1−12.47−1 R2−12.47−3 R5−12.47−1 R5−12.47−4 R5−25.00−1 0.6 is computed as follows. For each local hypothesis H ∈ H (Ai , f ), the conditional distribution ∆s| x̂ H ∼ N (µ(x̂, H), Σ(x̂, H)) is computed. µ PTo compute this, we introduce the following W (k) = load forecast of v∈desc(v) x̂(v) which computes the cumulative P all descendent vertices. Similarly W σ (v) = v∈desc(v) σ 2 (v) for the forecast variances. For a given area, we evaluate the ∆si under hypothesis H, assuming the root si and child sensor locations child(si ) for a particular observed flow. X X µ(x̂, H) = W µ (si ) − W µ (e) − W µ (e) (29) e∈V (H) e∈child(si ) Note that this is computed for a particular binary flow pattern. Therefore, no outage edge upstream from the child sensors that we consider. Likewise σ(x̂, H) is computed via W σ . Finally, given the scalar distribution for ∆si |{x̂ H}, the probability of missed detection for a scalar maximum likelihood detector can be computed. C. Optimality and Complexity This discussion concerns only optimal-tree-action since it guarantees optimality, although the greedy strategy output is almost always identical. Theorem 1. The bottom up placement solution Mg relying on optimal-tree-action traversal solves OPT-3. Complexity of evaluating the detector and the objective as well as the placement algorithm is discussed in Appendix E. The main results are summarized as follows: • The worst case complexity of evaluating the detector for an area and it’s missed detection PEmax (A) for an area of size |E| is O(4|E| ). Evaluating only outages of size k is O(|E|k ). max • Given a fixed size to evaluating PE (A) the greedy placement algorithm is of O(|E|) complexity while the optimal strategy is of O(4log |E| ) complexity. The detector and placement complexity in the worst case is quite poor. However in any practice, this cost is averted using a detector fixed hypothesis size. The outage model was of each edge having some finite prior likelihood of outage. Therefore a multi-edge outage of large size is much less likely. For this reason, all k outages do not need to be enumerated. In practical a single edge outage per area may be sufficient. Using a single 0.5 Sensor Density B. Evaluating PEmax (A) The objective PEmax (Ai ) + 0.4 0.3 0.2 0.1 0 0.1 0.3 0.5 0.7 Misdetection Error 0.9 Fig. 5. Outage detection performance for selected PNNL feeders. Square marker denotes specified error target for optimization. Circular marker denotes empirical mean missed detection error for hypothesis set. We perform outage detection using a subset of the Pacific Northwest National Laboratory test feeders [11]. Table I gives overview of the feeders chosen for the simulation study. The primary applications of the feeders, are heavy to light urban networks, as well as suburban and rural networks. The climate zones refer to (1) temperate (2) hot/arid (3) cold (4) hot/cold (5) hot/humid according to [11]. TABLE I PNNL T EST FEEDERS USED IN CASE STUDY Network Voltage Climate Zone Type Size R1-12.47-1 R2-12.47-3 R5-12.47-1 R5-12.47-4 R5-25.00-1 12.5 kV 12.47 kV 13.8 kV 12.47 kV 22.9 kV 1 2 5 5 5 suburban urban urban commercial suburban 613 52 265 643 946 1) Outage Model: For the PNNL feeders, outages are simulated by fuses and switches disconnected the downstream loads from the substation feeders. For each network, all fuses and switches are reduced to edges in the general tree network. The set of loads which are disconnected by a fuse or switch disconnecting are lumped to aggregate loads. For these loads, the mean load of each group of fuses can vary. 8 B. General Line and Tree Network Sensitivity κ = 0.01 κ = 0.1 κ = 0.2 κ = 0.3 Sensor Density 0.8 0.6 0.4 0.2 0 1 Sensor Density 1 κ = 0.01 κ = 0.1 κ = 0.2 κ = 0.3 0.8 0.3 0.5 0.7 Error Target Probability 0.9 (a) Err−Mean: κ = 0.01 ErrTarget: κ = 0.01 Err−Mean: κ = 0.25 ErrTarget: κ = 0.25 0.8 0.4 Sensor Density 0.5 Mean Misdetection Error 0.3 0.2 0.4 0.2 0.1 0 0 0.6 0.05 0.1 0.15 Empirical Misdetection Error 0.2 0.1 (a) 0.3 0.5 0.7 Misdetection Error 0.9 (b) Fig. 7. Hypothesis missed detection analysis: 7(a) error histogram for P target = 0.2; 7(b) tree network reduction to mean error. record the value of each hypothesis error. The empirical maximum error is close to the target 0.2. This makes sense because in successively solving the feasibility problem, we will expand the area network until the maximum error surpasses P target . However, we see that in fact almost all of the missed detection probabilities are less than the target. For this example in particular 34% of the hypothesis are less than 1e−3 therefore essentially zero. In comparing the mean and maximum errors for the range of achievable values of κ and P target . The mean error is in on average 25% of the P target . The maximum error in the network and P target very closely, therefore the optimization yields a very tight result. 0.6 VIII. C ONCLUSION 0.4 0.2 0.1 0.6 Empirical Count 2) Forecast Error Model: In [12] the authors present a rule of thumb model for day ahead load forecasting at various aggregation levels based on smart meter data. The day ahead forecast coefficient of variation, κ = σ/µ is shown to be dependent on the mean load of the group. Many studies make simplified assumptions on the relative forecast error. However, at the level of small aggregates, the forecast κ can vary greatly on the size of the aggregate and must beqtaken into account. A Reasonable fit shown in [12] is κ(W ) = 3562 W + 41.9. This formula is used to show that each set of islanded loads will have a different value of κ(W ). Figure 5 shows the application of the sensor placement algorithm for each network. Even though each networks represents a different applications, they show somewhat similar performance in terms of placement density. Averaged over each of the network configurations, attaining 10% mean missed detection error is possible by having 30% sensor density. Seen another way, we can reduce the realtime monitoring of each fuse by 70% by tolerating a small amount of error in the outage decision. 0 0.1 0.3 0.5 0.7 Error Target Probability 0.9 (b) Fig. 6. The effect of optimization error target P target and relative forecast error κ on both line (6(a)) and tree (6(b)) networks. Figure 6 shows the sensitivity of the line and tree networks under different simulation parameters. Both networks are of length 100 nodes, the tree was generated using the method in [10]. In an ideal line network with extremely high forecast accuracy (κ = 1%), 1 or 2 sensors are required for extremely low missed detection errors. This extreme situation does not occur in practice, but serves as a baseline for realistic networks. From Figure 6(a) we see that the required sensor density decreases quite quickly vs. P target . The relation between sensor density and P target is smoothly decaying.In comparison, randomly generated tree networks require on average 2 − 3 times as many sensors to achieve the same error target. C. Missed Detection Error Optimization (OPT-2) is meant to minimize the maximum missed detection error among all possible hypothesis. This is clearly can be too conservative of a requirement. Therefore it is useful to understand the nature of the actual hypothesis missed detection values that arise from a given sensor placement. Figure 7(a) shows the distribution of missed detection probabilities for a tree network. Setting P target = 0.2, and κ = 0.3 we We propose an outage detection framework combining power flow measurements on edges of the distribution system along with consumption forecasts at nodes of the network. We formulate the detection problem and provide an optimal placement for the maximum missed detection error metric. Finally, relying on feeder information from the pacific northwest national labs as well as a forecast error scaling law derived from Pacific Gas and Electric smart meter data, we demonstrate our formulation. R EFERENCES [1] J.R. Aguero and A. Vargas. Inference of operative configuration of distribution networks using fuzzy logic techniques - part i: real-time model. IEEE Transactions on Power Systems, 20(3):1551– 1561, 2005. [2] R. Balakrishnan and A. Pahwa. A computer assisted intelligent storm outage evaluator for power distribution systems. IEEE Transactions on Power Delivery, 5(3):1591–1597, 1990. [3] Christopher M Bishop et al. Pattern recognition and machine learning, volume 1. Springer New York, 2006. [4] C. Fukui and J. Kawakami. An expert system for fault section estimation using information from protective relays and circuit breakers. IEEE Transactions on Power Delivery, 1(4):83–90, 1986. [5] M. Kezunovic. Smart fault location for smart grids. IEEE Transactions on Smart Grid, 2(1):11–22, 2011. [6] Y. Liu and N.N. Schulz. Knowledge-based system for distribution system outage locating using comprehensive information. IEEE Transactions on Power Systems, 17(2):451–456, 2002. [7] Sentient Energy LLC. Mm3 intelligent sensor, 2016. [8] S.T. Mak and N. Farah. Synchronizing scada and smart meters operation for advanced smart distribution grid applications. Proc. IEEE PES Innovative Smart Grid Technologies (ISGT), pages 1–7, 2012. [9] R.A.F. Pereira, L.G.W. da Silva, M. Kezunovic, and J.R.S. Mantovani. Improved fault location on distribution feeders based on matching duringfault voltage sags. IEEE Transactions on Power Delivery, 24(2):852–862, 2009. [10] Alexey Rodionov and Hyunseung Choo. On generating random network structures: Trees. In Computational Science?ICCS 2003, pages 879–887. Springer, 2003. 9 [11] Kevin Schneider, Yousu Chen, David Chassin, Robert Pratt, David Engel, and Sandra Thompson. Modern grid initiative distribution taxonomy final report. arXiv preprint arXiv:1404.0058, 2008. [12] Raffi Sevlian, Patel Siddarth, and Ram Rajagopal. Day ahead electricity load forecasting on varying levels of aggregation. Power Engineering Letters, 2014. [13] D. Srinivasan, R.L. Cheu, Y.P. Poh, and A.K.C. Ng. Automated fault detection in power distribution networks using a hybrid fuzzy–genetic algorithm approach. Elsevier Engineering Applications of Artificial Intelligence, 13(4):407–418, 2000. [14] Z. Sumic and R. Vidyanand. Fuzzy set theory based outage determination. Proc. of 1996 International Conference on Intelligent Systems Applications to Power Systems, pages 204–208, 1996. [15] J. Zhu, D.L. Lubkeman, and A.A. Girgis. Automated fault location and diagnosis on electric power distribution feeders. IEEE Transactions on Power Delivery, 12(2):801–809, 1997. A PPENDIX A. Nomenclature Table T Tree network representation of a distribution feeder (V, E) Vertex and edge set of tree T V (H) Set of vertices that are connected to root under outage hypothesis H. desc(v) Descendants of vertex v. child(v) Children of vertex v. xn , x Scalar load and forecast value at vertex v n ,  Forecast residual for load l(v) σn2 , Σ Forecast residual variance and covariance matrix. H1 Single outage hypothesis and element and set. Hk k-outage hypothesis. Hu Unique hypothesis where no edges are downstream of any others. A, A, A+ Local area network A ∈ A; Pruned areas under binary flow processing. Set of local area hypotheses post binary flow Hu+ (Ak , f ) processing. Hk,i(k) i(k)th hypothesis in area Ak . Used to reconstruct global hypothesis Hi . M Sensor placement (M ⊂ E) s Set of observations on edges of network rij Acceptance region for pairwise test of hypotheses: Hi and Hj RH (H) Acceptance region of hypothesis H over all alternatives in H. PE,C (H, M), Probability of error (E) and correct detection (C) for hypothesis H, under placement M. PEmax (A) Maximum probability of incorrect detection over all hypothesis error in area A. PCmin (A) Minimum probability of correct detection over all hypothesis error in area A. B. MAP Detection for Outage Hyptheses Here we show how the general MAP detector rule can be evaluated for where we combine edge flows s, load forecasts x̂ and candidate outages H. Ĥ = arg max Pr (H | s, x̂) (30) H∈Hk Pr (s, x̂ | H) Pr (H) Pr (s, x̂) (31) = arg max Pr (s, x̂ | H) Pr (H) (32) = arg max H∈Hk H∈Hk = arg max Pr (s | x̂, H) Pr (x̂ | H) Pr (H) (33) H∈Hk = arg max Pr (s | x̂, H) Pr (x̂) Pr (H) (34) H∈Hk = arg max Pr (s | x̂, H) Pr (H) (35) H∈Hk = arg max Pr (s | x̂, H) (36) H∈Hk Lines (30) - (32) convert the MAP detector to a likelihood detector with prior weights. Line (33) conditions on the load forecast x̂. Since x̂ does not depend on the outage hypothesis, (only s does), the term can be removed leading to (35). In (36), 10 we assume a uniform prior over all hypotheses, however this does not have to be the case. Given the assumption of each edge going into outage with some fixed prior probability, a single edge outage hypothesis should have Pr (H) = ρ, while a k outage condition should have a prior of Pr (H) = ρk . This motivates enumerating fewer outage hypotheses when evaluating Hu in practice. we have that: H+ ({1 0}) = {E(b1 )}. We see that splitting the hypotheses based on flow information, conserves the search space, since Hu+ ({1 1}) ∪ Hu+ ({1 0}) = Hu . 4) Tree T1 : For tree T1 , we have: C. Extended Discussion of Recursive Evaluation of Hu D. General Hypothesis Decoupling s1 s1 b1 (a) Fig. 8. 3. b1 b1 s1 b2 b3 b2 (b) s2 b3 (c) 8(a) Worked example 1. 8(b) Worked example 2. 8(c) Worked example We present here additional worked out examples, that show how the recursive definition can enumerate all hypotheses and focus on some corner cases that must be defined. Recall the recursive definition:  ! [ ^ Hu (b) = E(b) ∪  Hu (b)  b∈P(child(b)) b∈b 1) Example 1 (Figure 8(a)): This is the simplest case to evaluate and is: Hu = E(b1 ) ∪ ∅ The null hypothesis set arises from evaluating P(d(b1 )) = ∅, since b1 has no children. 2) Example 2 (Figure 8(b)): This is the simplest case to evaluate and is: Hu = {E(b1 ) ∪ ∅ ∪ Hu (b2 ) ∪ Hu (b3 ) ∪ Hu (b2 ) × Hu (b3 )} = {E(b1 ) ∪ ∅ ∪ {∅ ∪ E(b2 )} ∪ {∅ ∪ E(b3 )}∪ {∅ ∪ E(b2 )} × {∅ ∪ E(b3 )}} = {∅ ∪ E(b1 ) ∪ E(b2 ) ∪ E(b3 ) ∪ (E(b2 ) × E(b3 ))} This example is reduced to it’s final form in eq. (13), in Section V-A. However, the following equalities are omitted in the enumeration: ∅∪∅=∅ (37) ∅×∅=∅ (38) ∅ × ei = ei (39) Hu (b1 ) = {∅ ∪ E(b1 ) ∪ E(b2 ) ∪ Hu (b3 ) ∪ E(b2 ) × Hu (b3 )} Hu (b3 ) = {∅ ∪ E(b3 ) ∪ E(b4 ) ∪ E(b5 ) ∪ {E(b4 ) × E(b5 )}} These two cases provide the intuition for a general procedure which is as follows: Given binary information from flows, all areas Ak with rooted sensor with sk = 0, are discarded in generating a local hypothesis. Each node in the branch graph is assigned a label, l ∈ L L = {P, Z, U } for (P ) positive, (Z) zero, and (U ) undetermined branches. These are defined as follows: Positive Branch: Branch is upstream from a sensor measuring positive flow, therefore can never be evaluated in any outage hypothesis. Also, it’s immediate parent branch cannot be enumerated either. Zero Branch: This branch is directly upstream from a zero measurement therefore it’s edges must always be enumerated in any outage hypothesis. Undetermined Branch: This branch has no information, so is enumerated without any restriction. This definition leads to the following procedure to enumerate H+ (Ai , I{child(si )>0} ). First each branch-node is labeled with the following procedure: Initialization Branch with descendent sensor (1) s > 0 assigned label P and (2) s = 0 assigned label Z, and (3) no descendants assigned label U . Update Given a current branch node b and the set of children, the node is assigned as follows: (1) If any descendent node is labelled P then it must be labelled P . (2) If descendants are U and Z, then it must be labeled U . Once the the branch-nodes are labeled, enumerating H+ (Ai , I{child(si )>0} ) can be done recursively using the following rules: Positive Rule Never enumerate a branch (E(b)) with positive flow label (P ). Zero Rule When evaluating the recursive definition H(b) on an element of the power set. If any descendent is labelled Z, only evaluate product set elements that contain this branch. d+1 b1 The final relation leads to ∅ × E(b) = E(b). 3) Example 3 (Figure 8(c)): Consider the two binary flow indicators I{s1 ,s2 >0} = {1 1} and {1 0}. In the first case, we have the following product set: d b2 b3 s0 H+ ({1 1}) = {E(b2 ) ∪ ∅1 } × {E(b3 ) ∪ ∅2 } b4 We use the fact that ∅ × E(b) = E(b) and that the product ∅1 × ∅2 = ∅, which is the global null hypothesis from the naive enumeration in Example 2. Similarly, enumerating the {1 0} case, b5 s1 = {E(b2 ) × E(b3 ) ∪ ∅1 × E(b3 ) ∪ E(b2 ) × ∅2 ∪ ∅2 × ∅2 } = {E(b2 ) × E(b3 ) ∪ E(b2 ) ∪ E(b3 ) ∪ ∅} 1 (a) 2… K d-1 … … … … (b) Fig. 9. 9(a) General network reduced to individual branches. 9(a) Worst case tree network of depth D and K children for each vertex. 11 TABLE II F. Proxy Function Optimization Binary Flow Id(si )>0} Branch Labels Hypotheses Branches 0, 0 U , Z, U , U , Z 0, 1 1, 0 1, 1 P , Z, P , U , P P , P , U, U, Z P , P , P , U, P b1 , b2 , b2 × b3 , b2 × b5 , b2 × b4 × b5 b2 , b2 × b4 b3 , b5 , b4 × b5 b4 The placement problem OPT-1 will output the optimal sensor locations and minimizing maximum missed detection error α? (M? ), as: α? (M? ) = min max PE (H, M) |M|=M H∈Hu An example local area is provided in Figure 9(a). The general method is applied under each of the binary flow cases, where the results are shown in Table II. The method is applied to each binary flow, and the branches to be enumerated are given. This can be approximated by using the decoupling of hypotheses in different areas. Recall the solution to OPT-1 repeated above, is the distributed detector in Algorithm 1 where each area performs a local hypothesis Ĥ, . . . , ĤM . The complete hypothesis is only correct if every local detection output is correct. So for any hypothesis, we have following lower bound: min P r(Ĥ = H;M) H∈Hu E. Complexity Analysis  In analyzing the complexity of various algorithms we assume the tree in Figure 9(b). Each vertex has K children and is of depth D. It can be shown that the number of edges is related to D −1 . these quantities by E = KK−1 max 1) Evaluating PE (A): For simplicity, we focus on a binary tree, so K = 2 and |E| = 2D − 1. The number of possible hypotheses at each depth C(d) using this network is related recursively as the following: K   X K C(d + 1) = C n (d + 1) (40) n n=1 K = (C(d) + 1) − 1 (41) This can be derived directly from (9). d In the binary tree case, it’s simple to show that C(d) = 42 −1. Therefore the number of hypotheses are double exponential in the depth of the tree. For the entire tree, this leads to the root D node value of C(D) = 42 − 1 which is 4(|E|+1) − 1. Therefore  |Hu | = O 4|E| , exponential in the size of the graph. This cost may be averted due to the following: • Fixed multi-hypothesis size. The MAP detector requires a prior probability of hypotheses. Since multiple edge outages are less likely, they don’t always have to be enumerated. For example, considering only single edge outages leads to |Hu | = O (|E|) complexity for an area. • Small area sizes, and Binary Flow segmentation. The number of hypotheses are exponential in |E| for an area. Given many sensors, this can divide the number of edges for an area considerably. Additionally, the binary flow information from downstream sensors will on average divide each H+ (A) by a factor of 2|d(s)| . The following analysis is in terms of the evaluation of PEmax (A) since we assume appropriate approximation of this function has been performed. 2) Evaluation of Algorithm 2 using greedy strategy: The greedy strategy will have to evaluate all 2K subproblems at each vertex, and choose the minimum PEmax (A). The worst case complexity is therefore O(|E| 2K ), which reduces to O(|E|), since K is a constant. 3) Evaluation of Algorithm 2 using optimal strategy: The optimal strategy will expand the problem size by a factor 2K at each vertex. The number of sub-problems to consider after a depth of D will be (2K )D which for a binary tree becomes O(4log |E| ). = min  H∈Hu ≥ = Y P r(Ĥi = Hi ; M) (42) ∀Ai , Hi Y ∀A∈A  Y min P r(Ĥ = H; A) H∈H(A) PCmin (A). (43) (44) ∀A∈A Line (42) follows from the decoupling of the decentralized detector. The overall MAP decision can be correct only if each local MAP decision is correct. For any H ∈ Hu the probability of each area making a correct decision is always greater than the worst case probability of correct decision for each area. We interchange the sensor placement and area notation since local areas are constructed from sensor placements. Here PCmin (A) is the minimum probability of correct detection within a local area A. This lower bound can be used to first upper bound the optimal α? . Finally, only an approximate solution to the upper bound is formulated. α(M? ) = min max PE (H, M) (45) max (1 − PC (H, M)) (46) |M|=M H∈Hu = min |M|=M H∈Hu = 1 − max min PC (H, M) Y PCmin (A) |M|=M H∈Hu ≤ 1 − max |M|=M (47) (48) ∀A∈A ≈ 1 − max min PCmin (A) (49) = 1 − max max (1 − PEmax (A)) (50) |M|=M ∀A∈A = |M|=M ∀A∈A min max PEmax (A) |M|=M ∀A∈A (51) Optimization OPT-1 is identical to (46), since the probability of a single hypothesis error can be exchanged for it’s compliment. The min-max to max-min change is due to the negative sign in 47. In line 47, instead of maximizing the minimum correct probability over all hypotheses we maximize Q a computationally tractable lower bound which is the product ∀A∈A PCmin (A). In 49 we introduce a close approximate solution which is the following: Instead of maximizing the product of PCmin (A) for each A, it is sufficient to maximizing the minimum of each PCmin (A) Experimentally the two solutions have been shown identical for a large number instances, and only sub-optimal in a small number of cases where the gap is small. 1 1 0.8 0.8 Lemma 1. An equivalent definition is \ R(Hi ) = ri,j . (55) j:Hj ∈H 0.6 0.6 0.4 0.4 0.2 0.2 Proof: Using the definition of the right hand side, we have \ ri,j = {s : s ∈ ri,1 ∩ . . . ∩ s ∈ ri,N } ! A P Cm i n(A, M) 12 j∈H = {s : Pr(s|Hi ) ≥ Pr(s|H1 ) ∩ . . . 0 0 0.5 maxA P Cmin(A, M) (a) 1 0 0 0.2 0.4 0.6 0.8 maxA P Cmin(A, M) ∩ Pr(s|Hi ) ≥ Pr(s|HN )} 1 = {s : Pr(s|Hi ) ≥ Pr(s|Hj ) ∀Hj ∈ H} (b) = RH (Hi ). Fig. 10. 10(a) Brute force placement evaluation where the two objectives are identical. 10(b) Brute force placement evaluation where the two objectives differ. Experimentally the two solutions have been shown identical for a large number instances, and only sub-optimal in a small number of cases where the gap is small. Figure 10(a), 10(b), shows a pair of randomly generated trees with N = 15 nodes with random loads and a forecast coefficient of variation of 0.02. In both cases, the bottom up placement was used to determine Mg , where |Mg | = 5.  15 A brute force enumeration of all 5 placements is evaluated Q for max∀A∈A PCmin (A, M) and ∀A∈A PCmin (A, M). In both cases, there is a strong correlation between the two solutions. The two solutions are not however equal Figure 10(a), the solutions are identical, while in 10(b), the two solutions differ by 7.2%. Intuitively they should intuitively be very close to each other. Decreasing one area error will increase the other area’s error due to the monotonic growth of PEmax (A) and the finite tree size. Therefore, maximizing the product of all the terms tends to a solution where each area error is as close to each other as possible. Minimizing the maximum error often leads to such a solution, since we must trade off one area error for another. Note, that this statement is for the maximum likelihood classifier. Under an arbitrary classifier, a procedure of constructing one-to-many classifier will lead to ambiguity. See [3] (pg. 183) for discussion. Conjecture 1. Given a ML detection problem with a set of hypothesis of the form: sk ∼ N (µk , σk2 + ∆) for k = 1, . . . , K. The missed detection error for each hypothesis will monotonically increase w.r.t ∆. This is seen to hold with 200,000 random problem instantiations. We now state Lemma 2. Definition 2. Two area networks are nested A ⊂ A0 if the vertices of each area V , V 0 are such that V ⊂ V 0 . Lemma 2. Given two area networks A and A0 where A ⊂ A0 , PEmax (A) ≤ PEmax (A0 ). G. Proof of Theorem 1 We prove Theorem 1 by showing the following: 1) The objective function PEmax (A) monotonically increases for nested areas. 2) Algorithm 2, will recover the solution to OPT-3 First we prove propositions 1 and state a conjecture shown to hold in large scale simulation experiments. These are needed to prove Lemma 2 which is needed to prove Theorem 1. First consider the following: Definition 1. For a single pairwise test, Hi vs Hj we have the following decision region: ri,j = {s ∈ RM : Pr(s|Hi ) ≥ Pr(s|Hj )}. (52) The observation space is therefore partitioned into two regions. So the detector is the following: ( Hi , s ∈ rij Ĥ = (53) Hj , s ∈ / rij . For the one-to-many ML test: Hi vs ∀Hj ∈ H, we have an acceptance region defined as: RH (Hi ) = {s ∈ RM : Pr(s|Hi ) ≥ Pr(s|Hj ) ∀Hj ∈ H}. (54) (a) (b) Fig. 11. 11(a) Case 1 showing growth by adding new nodes by moving terminal sensor down. The conditional pdf ∆s|H ∀h ∈ H does not change. However the acceptance region shrinks from RH to RH∪x . 11(b) Case 2 showing growth by adding new nodes by moving root sensor up. The conditional pdf ∆s0 |H ∀h ∈ H 0 will change. Again, the acceptance region shrinks from RH to RH∪x . max Proof: Given some α∗ = Perr (A), which is the maximum missed detection probability of the local hypothesis in area A which is evaluated at some hypothesis H ∗ . We aim to show that for an enlarged area A0 , the error probability for H ? will always be larger, which is quite intuitive. Therefore the maximum error PEmax (A0 ) regardless if H ∗ maximizes the missed detection in the enlarged area. Expansion of A is analyzed in two cases: Case 1 Terminal sensors expand downstream (away from v0 ) shown in Figure 11(a). Case 2 Root sensor of A moves upstream (closer to v0 ) shown in Figure 11(b). 13 We use the following shorthand: Hypotheses for area A and A0 : H , H(A), H ∪ xi , H(A0 ). The set xi differ in how the area is enlarged, where i = 1, 2 for case 1 and case 2. Therefore under case 1, and 2 we have RH∪x1 (Hi ) and RH∪x2 (Hi ). Now we show how the effective measurement distribution (as defined in (28)) and acceptance regions changes under each case. µ0i = µi Case 1 For all hypothesis Hi ∈ H we have that 2 and σ 0 i = σi2 . The distribution ∆s|H is unchanged. Given i T that RH (Hi ) =T j:Hj ∈H rij , the new acceptance region is RH∪x1 (Hi ) = j:Hj ∈H∪x1 rij , with rij from the original alternatives unchanged. Case 2 For all hypothesis Hi ∈ H we have that µ0i = µe +µi 2 and σ 0 i = σe2 + σi2 . The distribution and acceptance region change ∆s|Hi → ∆s0 |Hi . The T new acceptance T T region is the 0 following: RH∪x2 (Hi ) = { j:Hj ∈H rij } { j:Hj ∈H rij }, where the acceptance regions under the previous area alternatives are now different. To see why the distribution ∆s|Hi changes, first recall that:  ∆si |Hk ∼ N   X µ(v) − µT , v∈Vi \Vk µk , σk2 ∼N X  ∀H ∈ H (57) The terms µT and σT2 are the sum of loads forecasts and variances of all terminal sensors. Now moving the root node upstream leads to:  X s0i |Hk ∼ N  X µ(v) − µT , v∈Vi0 \Vk σ 2 (v) − σT2  (58) v∈Vi0 \Vk  ∼ N µe + µi , σe2 + σi2 ∀H ∈ H. (59) Therefore changing the position of si so as to add additional vertices will increase every original hypothesis mean and variance by the same amount. Now consider Case 1 first, where we merely add new alternatives, keeping the distributions of ∆s|H unchanged. Here we have: Pr(∆s ∈ RH∪x (H ∗ )|H ∗ true) \ = Pr(∆s ∈ ri,j | H ∗ true) (60) (61) j:Hj ∈H∪x = Pr({∆s ∈ \ ri,j } ∩ { j:Hj ∈H \ ∆s ∈ ri,j }| H ∗ true) j:Hj ∈x (62) ≤ Pr(∆s ∈ \ ∈ ri,j |H ∗ true) (63) Hj ∈H = Pr(∆s ∈ RH (H ∗ )|H ∗ true). Pr(∆s0 ∈RH∪x |H ∗ true) ≤ Pr(∆s0 ∈ RH (H ∗ )|H ∗ true) (65) = Pr(∆s0 − µe ∈ RH (H ∗ ) − µe |H ∗ true) (66) ∗ ≤ Pr(∆s ∈ RH |H true). (67) The inequality in line 65 uses the identical procedure in 61 63. In line 66 we are merely shifting the gaussian ∆s and the acceptance region by µe using a shorthand notation. This can be done since the MAP test is scalar. Finally the inequality in line 67, follows from conjecture 1. We can now prove Theorem 1. Proof: The bottom up solution Mg moved to the root node enlarging each area network so that each PEmax (A) < P target but any further up will violate the target area. Consider some other method produces a solution M0 which minimizes the number of sensors the error constraint, where |M0 | < |Mg |. This implies that some area must increase in size, as compared to the Mg solution, and from Lemma 2, some area will violate the error constraint. σ 2 (v) − σT2  (56) v∈Vi \Vk  σe2 in mean and variance. This implies that: (64) Line 61 defines the area RH (H ∗ ) using proposition 1. This is split into the the intersection of two separate events using our definitions of H and H∪x. Next we use the fact that Pr(A∩B) ≤ Pr(A). Therefore if Pr(∆s ∈ RH (H ∗ )|H ∗ true) ≤ Pr(∆s ∈ RH∪x2 (H ∗ )|H ∗ true) then PEmax (A) ≤ PEmax (A0 ). We next prove Case 2 where not only are more alternatives considered for H ∗ , but ∆s|H ∗ is translated by fixed amount µe , H. Greedy and Optimal Tree Action optimal-tree-action: The correct action at a node-junction is to enumerate each 2|child(vr )|−1 possible trees and process them until only one remains closest to the root. In the example in Figure 4, we must process to the root node processing both A2 and A3 in parallel as separate problem instances, with its own vt and M g . This is in contrast to the greedy strategy that chooses one placement and moves on. Each problem instance is then processed, until the area objective function violates P target . All but one problem instance is kept; the one where vt was closest to the root vertex. It turns out that the greedy-tree-action and the optimal-treeaction procedures are in practice extremely close as discussed in Appendix H. Algorithm 2 can only implement this technique, since we do not grow the search space with multiple bottom up scenarios. The bottom up placement using greedy-tree-action relies on moving the current node as close to root as possible while keeping the error < P target . At a juncture, recall that the current network is chosen as the one which minimizes the maximum error, and continues with that choice. To see why this is sub optimal, consider Figure 12 shows a typical subtree where we have the following events: 1) (Figure 12(a)) The bottom up algorithm will evaluate target target PEmax (AL and PEmax (AR . 1) < P 1)<P 2) (Figure 12(b)) Move onto evaluating the combined area network with the parent node. Evaluating PEmax (A2 ) > P target , we must choose in a greedy manner via greedy-tree-action. max 3) (Figure 12(c)) Evaluating both PEmax (AL (AR 3 ), 3 ) and PE L max R target max . The greedy where PE (A3 ) < PE (A3 ) < P R choice will keep AL 3 and discard the A3 . 4) (Figure 12(d)) The optimal choice is in fact AR 4 since target max R PEmax (AR ) < P < P (A ). 4 4 E The greedy choice will choose AL 3 and be forced to place a sensor in AL 4 , while the optimal can continue upstream. The following numerical example will lead to this: µi = 1, ∀i and σ 2 = {0.0599, 0.0125, 0.0835, 0.0945, 0.0906, 0.0607} with P target = 0.1923. 4 4 6 4 6 6 14 L PEmax(A 1 4) 1PE 3 2 3 4 5 3 5 6 3 4 2 5 2 4 1 2 2 3 5 3 5 6 4 6 4 2 4 3 4 4 6 6 PEmax(AL3) L PEmax 1 (A 4) 5 3 1 6 PEmax (AL3) PEmax(AR3) 4 5 6 PEmax4 (AL1) PEmax(AR3) 3 2 3 5 5 4 6 R PEmax 4 (A 1)6 6 3 5 6 4 6 σ2L + ∆ σ2R + ∆ 0.1 (d) 5 3 5 3 Ptarget 0.02 1 1 0.04 ∆ 0.06 0.08 0.1 (a) 2 5 PEmax(AR1) TABLE III max L (A 3)STUDYPEmax(AR3) PNNL T EST FEEDERS USEDPEIN CASE H+ (A, b) µk σk2 max (A) PE s2 > 0, s5 > 0 s2 > 0, s3 > 0 s1 > 0, s3 > 0 s1 > 0, s3 > 0 e3 , e4 , ∅ 1, 2, 3 0.0125, 0.1031, 0.1637 0.1083 e5 , e6 , , ∅ 1, 2, 3 0.0125, 0.0960, 0.1905 0.0961 e3 , e4 , ∅ 2, 3, 4 0.0724, 0.1558, 0.2504 0.1885 e5 , e6 , , ∅ 2, 3, 4 0.0724, 0.1630, 0.2236 0.1960 Notice that in the hypothesis means and variances, we have the following: L µk (AL 4 ) = µ(v1 ) + µk (A3 ) R µk (AR 4 ) = µ(v1 ) + µk (A3 ) L σ 2 (AL 4 ) = ∆ + σk (A3 ) R σ 2 (AR 4 ) = ∆ + σk (A3 ) Where µ(v1 ) = 1, and ∆ = σ 2 (v1 ) = 0.0599. The translation of mean and variance causes the maximum error over an area to switch from AL to AR . Recall that conjecture 1 stated that the maximum error in such a case will monotonically increase with respect to some translation in variances. Regardless, there is no domination between any pair of triplets, whereby one will always be greater than another under the same translation. 6 4 6 L ) in Figure R ) where the maximum hypotheThis can be (A seen 13(a) PEmax PEmax(A 3 3 sis error between the tuples is shown with respect to translation ∆. As indicated, any ∆ > 0.02 will cause their ordering to change, with the counterexample shown to be beyond this point. However, we should not that although the transition does occur, the gap is quiet small so any realistic gap will be very small. Since R P target is between P (AL 4 ) and P (A4 ) it is very unlikely to occur. In finding a counterexample 10000 monte carlo runs produced 5 examples. For this reason, the greedy and optimal placement strategies will have identical outcomes of random instances. Flow AR 4 4 4 AL 3 AL 4 5 Fig. 13. The maximum probability of error over three hypotheses as a function of 5 5 3 3 R a translation of each hypothesis variance. At the point ∆ = 0, P (AL 3 ) < P (A3 ), R and in point ∆ = 0.0599, P (AL 4 ) > P (A4 ) 5 Area AR 3 2 2 The triplet of scalar hypotheses which cause this are shown in 4III. 6 6 Table 4 6 4 6 PEmax(AL1) 2 0.05 0 PEmax(AR4) Fig. 12. 12(a) Both areas have maximum error smaller than target error: max (AL )1 < P target , P max (AR ) < P target . 12(b) The area’s are PE 1 1 1 E 1 max (A ) > P target combined, and tested where PE . 12(c) The greedy choice 2 L max L max R will choose A3 as the candidate: PE (A3 ) < PE (A3 ) < P target . 12(d) 2 2 max max (AR ) < P target 2 (AL ). 2 < PE The correct choice is AR 4 since PE 3 3 3 0.2 0.15 6 1 5 1 2 1 (c) 3 1 (b) 1 3 5 3 PEmax(AR4) 0.25 3 2 1 1 2 5 5 (a) PEmax(A2) 2 4 P4 Emax(AR61) PE4max(AL1)6 2 3 3 5 6 1 1 1 PEmax(A2) 21 21 max(AL ) P0.3 E 4 max(AR ) 4
3cs.SY
Testing Properties of Functions on Finite Groups arXiv:1509.00930v1 [cs.DS] 3 Sep 2015 Kenta Oono Preferred Networks, Inc. [email protected] Yuichi Yoshida∗ National Institute of Informatics, and Preferred Infrastructure, Inc. [email protected] September 4, 2015 Abstract We study testing properties of functions on finite groups. First we consider functions of the form f : G → C, where G is a finite group. We show that conjugate invariance, homomorphism, and the property of being proportional to an irreducible character is testable with a constant number of queries to f , where a character is a crucial notion in representation theory. Our proof relies on representation theory and harmonic analysis on finite groups. Next we consider functions of the form f : G → Md (C), where d is a fixed constant and Md (C) is the family of d by d matrices with each element in C. For a function g : G → Md (C), we show that the unitary isomorphism to g is testable with a constant number of queries to f , where we say that f and g are unitary isomorphic if there exists a unitary matrix U such that f (x) = U g(x)U −1 for any x ∈ G. 1 Introduction In property testing [27, 14], we want to decide whether the input function f satisfies a predetermined property P or “far” from it. More specifically, an algorithm is called a tester for a property P if, given a query access to the input function f and a parameter ǫ > 0, it accepts with probability at least 2/3 when f satisfies P, and rejects with probability at least 2/3 when f is ǫ-far from P. If a tester accepts with probability one when f satisfies P, then it is called a one-sided error tester. The definition of ǫ-farness depends on the model, but for the case of Boolean functions, we say that a function f : Fn2 → {−1, 1} is ǫ-far from P if the distance dist(f, g) := Prx [f (x) 6= g(x)] of f and g is at least ǫ for any function g : Fn2 → {−1, 1} satisfying P. Here −1 and 1 are corresponding to true and false, respectively. The efficiency of the tester is measured by the number of queries to f , called the query complexity. We say that a property P is constant-query testable if there is a tester with query complexity depending only on ǫ (and not on n at all). The study of testing properties of Boolean functions, or more generally, functions on finite fields was initiated by Rubinfeld and Sudan [27], and then subsequently many properties have been shown to be constant-query testable [14, 11, 16, 13]. To incorporate the algebraic structure of the finite field, Kaufman and Sudan [20] asked to study affine-invariant properties, that is, properties P such that, if f : Fnp → {0, 1} satisfies P, then f ◦ A also satisfies P for any bijective affine transformation ∗ Supported by JSPS Grant-in-Aid for Young Scientists (B) (No. 26730009), MEXT Grant-in-Aid for Scientific Research on Innovative Areas (No. 24106001), and JST, ERATO, Kawarabayashi Large Graph Project. 1 A. A lot of progress has been made on the study of the constant-query testability of affine-invariant properties [10, 17], and finally (almost) complete characterizations of constant-query testability were achieved [9, 33]. For further details on function property testing, refer to [8, 25] for surveys. Besides finite fields, functions over finite groups such as the cyclic group and the permutation group are also objects that naturally appear in various contexts, e.g., circuit complexity [1], computational learning [32], and machine learning [18]. Despite its importance, there are only a few works on testing properties on functions over finite groups [5, 14], and extending this line of research is the main focus of the present paper. More specifically, we consider testing properties of functions f of the form f : G → D, and more generally, f : G → D(d), where G is a finite group, D = {z ∈ C | |z| ≤ 1} is the unit disk, and D(d) = {A ∈ Md (C) | kAkF ≤ 1} is the set of d by d matrices with Frobenius norm at most one. Note that D(1) = D. We regard d as a constant. The reason that we use D and D(d) is that they are maximal sets closed under multiplication. Below, we get into the details of these two settings. Testing properties of functions of the form f : G → D: We define the distance between two p functions f, g : G → D as dist(f, g) = 12 kf − gk2 , where kf k2 := Ex∈G |f (x)|2 is the L2 norm. Note that dist(f, g) is always in [0, 1]. We say that a function f : G → D is ǫ-far from a property P if dist(f, g) ≥ ǫ for any function g : G → D satisfying P. We note that, for {−1, 1}-valued functions √ f, g : G → {−1, 1}, we have Pr[f (x) 6= g(x)] = ǫ if and only if dist(f, g) = ǫ holds. Hence, we have a quadratic gap between our definition and the standard definition using the Hamming distance. However, we adopt the L2 norm as it is more friendly with our analysis. We first show the following: • Conjugate invariance, that is, f (yxy −1 ) = f (x) for any x, y ∈ G, is one-sided error testable with O(1/ǫ2 ) queries. • Homomorphism, that is, f (x)f (y) = f (xy) for any x, y ∈ G, is one-sided error testable with O(1/ǫ2 log(1/ǫ)) queries. We show the constant-query testability of conjugate invariance by a simple combinatorial argument. When G = Fn2 , then homomorphism is often called linearity and intensively studied in the area of property testing [4, 6, 14, 28]. PIndeed in this case, a function f : Fn2 → D is homomorphism if and only if f (x) = χS (x) := (−1) i∈S xi for some S ⊆ {1, . . . , n}. The case G = Sn , the permutation group of order n, is easy to understand. Let f : Sn → D be a function on the permutation group. If f is conjugate invariant, then the value of f (π) only depends on the cycle pattern of π. If f is homomorphism, then f is the all-zero function, the all-one function, or the function that returns the sign of the input permutation. We note that Ben-Or et al. [5] studied the constant-query testability of homomorphism from a finite group to another finite group. The query complexity of their algorithm is O(1/ǫ), where the distance is measured by the Hamming distance. Their algorithm and analysis by a combinatorial argument extends to our setting, in which the range is D and the distance is measured by the L2 norm, with query complexity O(1/ǫ2 ). In this sense, our result on homomorphism is not new. Nevertheless, we prove it again using harmonic analysis over finite groups. By doing so, we can generalize it for testing other properties. To describe our next result, we need to introduce the basic of representation theory. A representation of a group G is a homomorphism ϕ of the form ϕ : G → Md (C) for some integer d. In particular, we study the family of irreducible representations, where any representation can be 2 described as the direct sum of irreducible representations. We mention that irreducible representations ϕ : G → Md (C) can be chosen as unitary, that is, ϕ(x) is unitary for all x ∈ G. The character of a representation ϕ is the function χϕ (x) = tr(ϕ(x)), where tr(·) denotes the trace of a matrix. The character carries the essential information about the representation in a more condensed form and is intensively studied in character theory. The character of an irreducible representation is called an irreducible character. For example, every irreducible character of Fn2 is of the form P χS (x) = (−1) i∈S xi for some S ⊆ {1, . . . , n}, which is often called a character in Fourier analysis of Boolean functions. If G is abelian, that is, commutative, then representations of G always map to one-dimensional matrices, and hence a representation and its corresponding character coincide. However, this is not the case when G is not abelian, which makes the analysis more involved. We show the following: • The property of being proportional to an irreducible character, that is, f = cχϕ for some c ∈ C and irreducible representation ϕ, is testable with O(1/ǫ8 log2 (1/ǫ)) queries. The reason that we do not consider irreducible characters themselves is that irreducible characters may take values outside of D. In particular, χϕ (1) = d holds for the identity element 1 ∈ G and the dimension d of ϕ. When G = Fn2 , then irreducible characters coincide with linear functions, that is, χS for some S ⊆ {1, . . . , n}. Hence, testing the property of being proportional to an irreducible character can be seen as another generalization of linearity testing. The form of irreducible characters is quite complicated in general. When G = Sn , however, its combinatorial interpretation via Young Tableau is well studied [12, 21, 29] (though still complicated to state here). Testing properties of functions of the form f : G → D(d): Since the representations of a group G are matrix-valued, it is natural to consider testing properties of q functions of the form f : G → D(d). For two functions f, g : G → D(d), we define dist(f, g) = 21 Ex kf (x) − g(x)k2F . Note that this is indeed a metric and matches the previous definition of distance when d = 1. Let U (d) be the set of d by d unitary matrices with each element in C. We show the following: • Unitary equivalence to g : G → D(d), that is, f = U gU −1 for some unitary matrix U ∈ U (d), 2 is testable with (d3/2 /ǫ)O(d ) queries. Here g is a parameter of the problem and not a part of the input. Unitary equivalence is an important notion when studying representations since the irreducibility of a representation ϕ is preserved by the transformation ϕ 7→ U ϕU −1 for a unitary matrix U . Our tester samples unitary matrices from the Haar measure, a fundamental tool in the representation theory of Lie Groups, and then checks whether g becomes close to f by applying these unitary matrices. Arguably the simplest property of matrix-valued functions is again homomorphism. However, homomorphism is known to be constant-query testable by a combinatorial argument [5], and the harmonic analysis does not facilitate the analysis. Therefore we do not study homomorphism of matrix-valued functions in this paper. 3 Related work: There are a number of works on testing whether a function on a finite group is a homomorphism. Blum et al. [14] gave a tester (the BLR tester ) for homomorphism of functions on a finite group. However, the number of queries depends on the number of generators of the group, which may depend on the size of the group in general. Ben-Or et al. [5] gave another algorithm without this dependency. Bellare et al. [4] gave a Fourier-analytic proof of the BLR tester when the domain and the range are Fn2 and F2 , respectively. Our tester for homomorphism can be seen as a generalization of their analysis to general groups. There has been an interest in improving various parameters of homomorphism testing results, due to their applications in the construction of probabilistically checkable proof (PCP) [3]. Bellare et al. [4] gave an almost tight connection between the distance to homomorphism and the rejection probability of the BLR tester. BenSasson et al. [6] and Shpilka and Wigderson [28] reduced the number of random bits required by the test as it affects the efficiency of the proof system and in turn the hardness of approximation results that one can achieve using the proof system. Rubinfeld [26] studied properties of a function on a finite group that are defined by functional equations and gave a sufficient condition of constantquery testability. Using the Lp norm for p ≥ 1 as a distance measure in property testing is recently systematically studied by Berman et al. [7]. One of their motivations is exploring the connection of property testing with learning theory and approximation theory. For this purpose, the Lp norm is more favorable than the Hamming distance because, in learning theory and approximation theory, we typically measure errors in the Lp norm for p = 1 or 2. Indeed, several lower bounds and upper bounds for property testing in the Lp norm were shown using this connection. See [7] for more details. Representation theory is one of the most important areas in modern mathematics. Representation theory itself is intensively studied and it is also used as an analytical tool in harmonic analysis, invariant theories, and modular theory. Representation theory have been used in various problems of theoretical computer science such as constructing pseudorandom objects [2, 19], circuit complexity [1], communication complexity [24], computational learning [32], machine learning [18], and quantum property testing [23]. Organization: In Section 2, we introduce representation theory, harmonic analysis on finite groups, and the Haar measure in more detail. We discuss the testability of conjugacy invariance in Section 3. We show that homomorphism and the property of a being proportional to an irreducible character are constant-query testable in Sections 4 and 5, respectively. Section 6 is devoted to testing unitary equivalence. 2 Preliminaries For an integer n ≥ 1, [n] denotes the set {1, 2, . . . , n}. Let δij be Kronecker’s delta, that is, δij = 1 if i = j and δij = 0 if i 6= j. For a matrix M , we denote its (i, j)-th element by Mij . We √ write the real and imginary part of a complex number z as ℜz and ℑz, respectively (hence z = ℜz + −1ℑz). For a complex number z, z denotes its conjugate. We frequently use the following lemma. Lemma 2.1. Let f : G → D be a function for some finite group G. For any ǫ > 0, with probability at least 1 − δ, we can compute an estimate z of Ex∈G [f (x)] such that |z − Ex∈G [f (x)]| ≤ ǫ . The number of queries to f is O(1/ǫ2 log 1/δ). 4 x1 , . . . , xs uniformly at random from G, Proof. Let c = Ex∈G [f (x)]. To estimate c, we sample 1P 2 where s = Θ(1/ǫ log 1/δ). Then, we output e c := s i∈[s] f (xx ). Clearly, the query complexity is O(1/ǫ2 log 1/δ). We now show that e c is indeed a good approximation to c. Since |ℜ(f (x))| ≤ 1, from Hoeffding’s inequality, we have |ℜ(c) − ℜ(e c)| ≤ ǫ/2 with probability at least 1 − δ/2 by choosing the hidden constant in s large enough. Similarly, we have |ℑ(c) − ℑ(e c)| ≤ ǫ/2 with probability at least 1 − δ/2. By the union bound, we have |c − e c| ≤ |ℜ(c) − ℜ(e c)| + |ℑ(c) − ℑ(e c)| ≤ ǫ with probability at least 1 − δ. 2.1 Representation theory We introduce basic notions and facts in representation theory. See, e.g., [30] for more details. For a vector space V over a field F, GLF (V ) denotes the set of invertible linear transformations. We only consider the case F = C in this paper, and hence we omit the subscript for simplicity. A representation of G is a pair (ϕ, V ) of a finite-dimensional vector space V and a homomorphism ϕ : G → GL(V ), that is, ϕ(xy) = ϕ(x)ϕ(y) for every x, y ∈ G and ϕ(1) is the identity transformation for the identity element 1 ∈ G. For a representation (ϕ, V ), V is called the representation space of it. When V is clear from the context, we simply call ϕ a representation. The dimension of a representation (ϕ, V ) is the dimension of V . When V is a finite-dimensional vector space, then we say that (ϕ, V ) is finite-dimensional representation. In our argument, we only need finite-dimensional representations. We describe the decomposition of a representation into irreducible representations, which is a fundamental tool used in representation theory. For a representation (ϕ, V ) and a subspace W of V , we say that W is G-invariant if ϕ(G)W ⊆ W . If W is a G-invariant space, then we can regard the range of ϕ as GL(W ), and hence we obtain a representation (ϕ, W ). Note that {0} and V are G-invariant from the definition. A representation (ϕ, V ) is called irreducible if {0} and V are the only G-invariant spaces. Note that a one-dimensional representation is always irreducible. When G is abelian, then we have the converse from Schur’s Lemma, that is, any irreducible representation is one-dimensional. When G is non-abelian, however, an irreducible representation might have dimension more than one. This fact makes the analysis of algorithms for functions on a non-abelian group more involved. Two representations (ϕ, V ) and (ψ, W ) of G are equivalent if there exists an invertible linear transformation T : V → W such that, for every x ∈ G, it holds that ψ(x)◦T = T ◦ϕ(x) . We identify b the family of equivalence classes of irreducible equivalent representations, and we denote by G representations. It is known that there is a one-to-one correspondence between conjugacy classes b of G and G. A representation (ϕ, V ) is unitary if, for all x ∈ G, ϕ(x) is a unitary transformation. For any representation of G, there is an equivalent unitary representation. Hence, we can take unitary b and we identify it with G. b Since representations as a complete system of representatives of G, b b G is finite, so is G. For ϕ ∈ G, we denote the dimension of its representation space by dϕ . In what follows, we fix a basis of the vector space of each representation (ϕ, V ), and we regard it as a homomorphism from G to Mdϕ (C), where dϕ is the dimension of V . 5 2.2 Fourier analysis on non-abelian finite groups We regard the space of C-valued functions of G as an inner product space by defining hf, gi = Ex∈G [f (x)g(x)] for f, g : G → C. The following fact is known. p b i, j ∈ [dϕ ] forms an orthonordϕ ϕij | ϕ ∈ G, Lemma 2.2 ([30]). For a finite group G, the set mal basis of the space of C-valued functions of G. Hence, we can decompose f : G → C as X X X X fb(ϕ)ij ϕij (x), hf, ϕij iϕij (x) = dϕ f (x) = dϕ i,j∈[dϕ ] b ϕ∈G i,j∈[dϕ ] b ϕ∈G where fb(ϕ) ∈ Mdϕ (C) is defined as fb(ϕ) = Ex∈G [f (x)ϕ(x)] and called the Fourier coefficient of ϕ. This decomposition is called the Fourier expansion of f . Note that Fourier coefficients are matrix-valued functions. The following is well known. Lemma 2.3 ([30]). Let f, g : G → C be functions. Then, we have X X fb(ϕ)ij gb(ϕ)ij , hf, gi = dϕ ϕ kf k22 = 2.3 X ϕ (Plancherel’s identity) i,j∈[dϕ ] dϕ X i,j∈[dϕ ] |fb(ϕ)ij |2 . (Parseval’s identity) Class functions and characters For a representation ϕ : G → Mdϕ (C), the character χϕ : G → C of ϕ is defined as χϕ (x) = tr(ϕ(x)) for x ∈ G. We say that a function f : G → C is conjugate invariant if f (x) = f (yxy −1 ) for all x, y ∈ G. A conjugate invariant function is sometimes called a class function. It is not hard to check that characters are conjugate invariant. Indeed, the following fact is known.  b forms an orthonormal basis of the space Lemma 2.4. For a finite group G, the set χϕ | ϕ ∈ G of C-valued class functions of G. Note that if a representation is one-dimensional, its character is identical to the original representation, hence is a homomorphism. This is not the case in general. The following lemma says that Fourier coefficients of a class function are always diagonal. Lemma 2.5. For any class function f : G → C, it holds that fb(ϕ) = hf,χϕ i dϕ Idϕ . In order to prove Lemma 2.5, we need the following two auxiliary lemmas: Lemma 2.6. For a function f : G → C and an irreducible representation ϕ, we have Proof. hf, χϕ i = E x = X X k ρ dρ X i,j hf, χϕ i = tr(fb(ϕ)). fb(ρ)ij ρij (x) fb(ϕ)kk = tr(fb(ϕ)). X ϕkk (x) = X ρ k 6 dρ X i,j fb(ρ)ij X hρij , ϕkk i k Lemma 2.7. Let f : G → C be a function and g : G → C be a class function. Then, X g (ϕ)). hf, gi = tr(fb(ϕ))tr(b ϕ Proof. Since g is a class function, we can represent g(x) = X X fb(ϕ)ij hg, χ′ϕ i E[ϕij (x)χϕ′ (x)] hf, gi = dϕ = X X dϕ i,j ϕ,ϕ′ = X dϕ ϕ i X ϕ,ϕ′ = = XX X ϕ Proof of Lemma 2.5. ϕ ≤ i,j ϕ fb(ϕ)ij fb(ϕ)ij fb(ϕ)ii X ϕ X X X Now we have X X X i′ i′ i′ gb(ϕ′ )i′ i′ gb(ϕ′ )i′ i′ X j′ X j′ E[ϕij (x)ϕ′j ′ j ′ (x)] x (from Lemma 2.6) hϕij , ϕ′j ′ j ′ i gb(ϕ)i′ i′ g (ϕ)). tr(fb(ϕ))tr(b kf k22 = = ϕ hg, χϕ iχϕ (x). x i,j ϕ,ϕ′ P i∈[dϕ ] dϕ tr(fb(ϕ)) fb(ϕ)ii X i,j∈[dϕ ] 2 ≤ fb(ϕ)ij 2 (from Lemma 2.7) X ϕ 2 dϕ X i∈[dϕ ] = kf k22 . fb(ϕ)ii 2 (by Cauchy-Schwarz) (by Parseval’s identity) Therefore, the equality holds for both inequalities in the formula above. In particular, fb(ϕ) is proportional to the identity matrix cIdϕ for some c ∈ C. By Lemma 2.6, hf, χϕ i = tr(fb(ϕ)) = cdϕ . Hence, c = hf, χϕ i/dϕ , and we have the lemma. 2.4 Introduction to the Haar measure In this section, we introduce Haar measure briefly. See, e.g., a textbook [22] for more details. A topological group is a group equipped with a topology and whose group operations are continuous in its topology. A finite group is a topological group if it is endowed with a discrete topology. 2 Any subgroup of Md (C) is a topological group, in which we identify Md (C) with Cd and introduce 2 the topology induced by Cd . We call a measure µ on a topological group G left invariant (resp., right invariant) if µ(xS) = µ(S) (resp., µ(S) = µ(Sx)) for any x ∈ G and Borel set S ⊆ G. Similarly, We call µ invariant under taking inverse if µ(S −1 ) = µ(S) for any Borel set S ⊆ G where S −1 = {x−1 | x ∈ S}. For any compact topological group G, there exists a measure on G which is left invariant, right invariant, and invariant under taking inverse. Such a measure is unique up to scalar multiplication and called the Haar measure on G. For example, the Haar measure µ of a finite group G is (a scalar multiplication of) the counting measure, that is, µ(S) = |S|/|G| for subset S ⊆ G. We regard U (d) as a closed subgroup of Md (C) and regard it as a compact topological group. Hence, the Haar measure of U (d) exists. 7 Algorithm 1 (Tester for conjugate invariance) 1: for s := O(1/ǫ2 ) times do 2: Sample x and y ∈ G uniformly at random. 3: if f (x) 6= f (yxy −1 ) then reject. 4: Accept. 3 Conjugate Invariance In this section, we first show that conjugate-invariance is constant-query testable. Theorem 3.1. Conjugate invariance is one-sided error testable with O(1/ǫ2 ) queries. Then, we show the following lemma, which simplifies testing properties that imply conjugate invariance. Lemma 3.2. Let P be a property such that every f satisfying P is a class function. Suppose that there is a tester A for P with query complexity q(ǫ) if the input is restricted to be a class function. Then there is a tester A′ for P with query complexity O(1/ǫ2 + q(ǫ/2) log q(ǫ/2)). Moreover, if A is a one-sided error tester, then A′ is also a one-sided error tester. 3.1 Proof of Theorem 3.1 Our algorithm for testing conjugate invariance is described in Algorithm 1. It is easy to see that the query complexity of Algorithm 1 is O(1/ǫ2 ) and the tester always accepts when f is conjugate invariant. Thus, it suffices to show that Algorithm 1 rejects with probability at least 2/3 when f is ǫ-far from class functions. It is well known that conjugacy classes of G form a partition of G. We define G♯ as the set of conjugacy classes of G. Also, for an element x ∈ G, we define x♯ as the unique conjugacy class x belongs to. ′ ♯ −1 ′ For y ∈ x♯ , we define Nx,y = F {z ∈ G | zxz = y}. Since we have Nx,y ∩ Nx,y = ∅ for y, y ∈ x ′ with y 6= y , it holds that G = y∈x♯ Nx,y . Therefore, the following lemma guarantees that uniform sampling from a conjugacy class is executed by uniformly sampling from the whole group. Lemma 3.3. The number of elements in Nx,y depends only on the conjugacy class to which y belongs. Proof. For y, y ′ ∈ x♯ , fix z0 , z0′ ∈ G so that z0 xz0−1 = y and z0′ xz0′ −1 = y ′ hold. We construct mappings Φy,y′ : Nx,y → Nx,y′ by z 7→ z0′ z −1 z0 and Φy′ ,y : Nx,y′ → Nx,y by z 7→ z0 z −1 z0′ . By a direct calculation, we can check that Φy,y′ ◦ Φy′ ,y = idNx,y , and Φy′ ,y ◦ Φy,y′ = idNx,y′ . Therefore |Nx,y | = |Nx,y′ | holds. Fix a function f : G → D. For a conjugacy class C ∈ G♯ and z ∈ C, define pC (z) := ♯{x ∈ C | f (x) = z}/|C| as the probability that f (x) = z if we sample x ∈ C uniformly at random. We define pC := maxz∈C pC (z) and zC := arg maxz∈C pC (z). Then, we define fe as fe(x) = zx♯ . Note that fe is a class function such that fe(x) ∈ D for any x ∈ G. Lemma 3.4. Pr [f (x) 6= f (yxy −1 )] ≥ dist(f, fe)2 . x,y∈G 8 Proof. Since |x − y| ≤ 2 for any x, y ∈ D, we have 1 X 1 |C|(1 − pC ). dist(f, fe)2 = E |f (x) − fe(x)|2 ≤ Pr[f (x) 6= fe(x)] = x 4 x |G| ♯ C∈G By Lemma 3.3, if we fix x ∈ G and sample y ∈ G uniformly at random, then yxy −1 forms a uniform distribution over elements in x♯ . Thus, Z 1 X Pr [f (x) 6= f (yxy −1 )] ≥ |C| pC (z)(1 − pC (z))dz x,y∈G |G| C∈G♯ Z X 1 X 1 |C| pC (z)(1 − pC )dz = |C|(1 − pC ) ≥ dist(f, fe)2 . ≥ |G| |G| ♯ ♯ C∈G C∈G Lemma 3.5. If f is ǫ-far from being conjugate invariant, then Algorithm 1 rejects with probability at least 2/3. Proof. Since fe : G → D is a class function, we have dist(f, fe) ≥ ǫ. Hence, the probability we reject at Line 3 in each trial is at least ǫ2 by Lemma 3.4. Hence the tester rejects with probability 2/3 by choosing the hidden constant in s large enough. We establish Theorem 3.1 by Lemma 3.5. 3.2 Proof of Lemma 3.2 The following lemma shows that we can obtain a query access to a class function that is close to f . Lemma 3.6. Let f : G → D be a function that is ǫ-close to a class function. There exists a class function f ′ : G → D with the following property. • For any x ∈ G, with O(log 1/δ) queries to f , we can correctly compute f ′ (x) or find a witness that f is not a class function with probability at least 1 − δ. Moreover, if f itself is a class function, then we can always compute f ′ (x) correctly. • dist(f ′ , f ) ≤ 3ǫ. In particular, f ′ = f when f itself is a class function. P ∗ ∈ D be the unique value that minimizes ∗ 2 Proof. For a conjugacy class C, let zC x∈C |f (x) − zC | . ∗ ∗ ∗ ∗ We define f : G → D as f (x) = zx♯ . Note that f is the class function closest to f . We define f ′ : G → D as follows: ( zx∗♯ if px♯ ≤ 21 , f ′ (x) = zx♯ otherwise. We first show the first claim. Our algorithm for computing f ′ (x) is as follows. Given x ∈ G, we pick y1 , . . . , ys ∈ G for s := O(log 1/δ) uniformly at random, and compute f (yi xyi−1 ) for each i ∈ [s]. If f (yi xyi−1 ) 6= f (yj xyj−1 ) for some i 6= j, then we reject f and output the pair as the witness that f is not a class function. If all of them are the same, we output the value as f ′ (x). 9 Now we analyze the correctness of the algorithm above. If px♯ ≤ 1/2, then with probability at least 1 − δ, we have f (yi xyi−1 ) 6= f (yj xyj−1 ) for some i 6= j, and we reject. If px♯ > 1/2, with probability at least 1 − δ, the majority of {f (yi xyi−1 )}i∈[s] is equal to zC . Hence, with probability at least 1 − δ, either we output zC or reject. Moreover, if f itself is a class function, then we have f ′ (x) = f (x) for any x ∈ G, and our algorithm always outputs f (x) as f ′ (x). We turn to the second pP claim. For two functions g, h : G → C and a conjugacy class C, ♯ 2 define distC (g, h) := x∈C |g(x) − h(x)| . We will show that, for each conjugacy class C ∈ G , distC (f, f ′ ) ≤ 3distC (f, f ∗ ), which implies dist(f, f ′ ) ≤ 3ǫ. If C satisfies pC ≤ 1/2, we have nothing to show. Thus suppose pC > 1/2. Then, we have sX q ′ ∗ ∗ ′ ∗ − z |2 . |f (x) − f ∗ (x)|2 + |C||zC distC (f, f ) ≤ distC (f, f ) + distC (f , f ) = C x∈C ∗ when constructing f ∗ from f , we have Since a pC -fractionP of values has moved from zC to zC P ∗ 2 ∗ − z |2 ≤ 2 ∗ 2 pC |C||zC − zC | ≤ x∈C |f (x) − f (x)| . By pC > 1/2, we have |C||zC C x∈C |f (x) − √ ∗ 2 ′ f (x)| . Combining this with the previous inequality, we have distC (f, f ) ≤ (1 + 2)distC (f, f ∗ ) ≤ 3distC (f, f ∗ ). Proof of Lemma 3.2. We first apply the ǫ/6-tester for conjugate invariance (Algorithm 1). If the tester rejects, we immediately reject f as it implies that f does not satisfy P . Otherwise, using Lemma 3.6, we construct a query access to a class function f ′ with δ = O(1/q(ǫ/2)). Then we apply the tester A to f ′ with the error parameter ǫ/2. The query complexity is clearly as stated. Suppose that f satisfies the property P . Then, we never reject when testing conjugate invariance. Also f ′ (x) = f (x) holds for every x ∈ G and it follows that f ′ satisfies the property P . Hence, the tester A accepts f ′ with probability at least 2/3. Moreover if A is a one-sided error tester, then A accepts f ′ with probability one. Suppose that f is ǫ-far from the property P . If f is ǫ/6-far from conjugate invariance, then we reject f with probability at least 2/3. Thus assume that f is ǫ/6-close to conjugate invariance. In this case f ′ is a class function that is ǫ/2-close to f . Hence f ′ is still ǫ/2-far from the property P . Then the tester A on f ′ should reject with probability 2/3. 4 Testing Homomorphism In this section, we show the following: Theorem 4.1. Homomorphism is one-sided error testable with O(1/ǫ2 log(1/ǫ)) queries. We note that, if f : G → C is a homomorphism, then it is a one-dimensional representation and hence an irreducible representation. First we observe that homomorphism implies conjugate invariance. Lemma 4.2. If f : G → C is a homomorphism, then f is conjugate invariant. Proof. Since f is a homomorphism, we have for any x, y ∈ G, f (yxy −1 ) = f (y)f (x)f (y −1 ) = f (y)f (y −1 )f (x) = f (yy −1 )f (x) = f (1)f (x). By setting y = 1, we have f (x) = f (1)f (x), which means f (x) = 0 or f (1) = 1. 10 Algorithm 2 (Tester for homomorphism) Input: A class function f : G → D. 1: for s = O(1/ǫ2 ) times do 2: Sample x, y ∈ G uniformly at random. 3: if f (x)f (y) 6= f (xy) then reject. 4: Accept. If f (x) = 0 for all x ∈ G, then f is clearly conjugate invariant. If f (x) 6= 0 for some x ∈ G, then f (1) = 1. In this case, we have f (yxy −1 ) = f (1)f (x) = f (x) and f is again conjugate invariant. From Lemmas 3.2 and 4.2, to test homomorphism, it suffices to show that homomorphism is one-sided error testable with O(1/ǫ2 ) queries when the input function is a class function. Our tester is given in Algorithm 2. It is clear that the query complexity is O(1/ǫ2 ). We next see that Algorithm 2 always accepts homomorphisms: Lemma 4.3. If a class function f : G → D is a homomorphism, then Algorithm 2 always accepts. Proof. We always accept because f (x)f (y) = f (xy) for any x, y ∈ G Now we turn to the case that f is ǫ-far from homomorphisms. To show that Pr[f (x)f (y) 6= f (xy)] is much smaller than 1, we analyze the term f (x)f (y)f (xy). Lemma 4.4. For any function f : G → C, we have E [f (x)f (y)f (xy)] = x,y X dϕ ϕ X i,j,k∈[dϕ ] Proof. The left hand side is equal to X dϕ dϕ′ dϕ′′ ϕ,ϕ′ ,ϕ′′ X X X fb(ϕ)ij fb(ϕ′ )i′ j ′ fb(ϕ′′ )i′′ j ′′ E [ϕij (x)ϕ′i′ j ′ (y)ϕ′′i′′ j ′′ (xy)]. x,y i,j∈[dϕ ] i′ ,j ′ ∈[dϕ′ ] i′′ ,j ′′ ∈[dϕ′′ ] Now we analyze the expectation in (1). h ′ ′ E [ϕij (x)ϕi′ j ′ (y)ϕ′′i′′ j ′′ (xy)] = E ϕij (x)ϕi′ j ′ (y) x,y x,y = X k ′′ ∈[dϕ′′ ] Hence (1) = hϕij , ϕ′′i′′ k′′ ihϕ′i′ j ′ ϕ′′k′′ j ′′ i P ϕ dϕ P b = b fb(ϕ)ij fb(ϕ)jk fb(ϕ)ik . ( X ϕ′′i′′ k′′ (x)ϕ′′k′′ j ′′ (y) k ′′ ∈[dϕ′′ ] 1 d2ϕ if ϕ = ϕ′ = ϕ′′ , i = i′′ , j = i′ = k′′ , and j ′ = j ′′ , 0 otherwise. b i,j,j ′ ∈[dϕ ] f (ϕ)ij f (ϕ)jj ′ f (ϕ)ij ′ . Corollary 4.5. For any class function f : G → C, we have X X fb(ϕ)ii |fb(ϕ)ii |2 . dϕ E [f (x)f (y)f (xy)] = x,y i ϕ 11 i∈[dϕ ] (1) b and i 6= j ∈ [dϕ ] by Lemma 2.5. Proof. If f is a class function, then fb(ϕ)ij = 0 for any ϕ ∈ G Hence, we have the corollary from Lemma 4.4. The following lemma completes the proof of Theorem 4.1. Lemma 4.6. If a class function f : G → D is ǫ-far from homomorphism, then Algorithm 2 rejects with probability at least 2/3. Proof. From Corollary 4.5, we have ℜ E [f (x)f (y)f (x + y)] = ℜ x,y ≤ = max b ϕ∈G,i∈[d ϕ] max b ϕ∈G,i∈[d ϕ] ℜfb(ϕ)ii · X X ϕ dϕ ϕ dϕ X i∈[dϕ ] ℜfb(ϕ)ii · kf k22 ≤ X i∈[dϕ ] fb(ϕ)ii |fb(ϕ)ii |2 |fb(ϕ)ii |2 max b ϕ∈G,i∈[d ϕ] ℜfb(ϕ)ii . For any ϕ with dimension more than one, |fb(ϕ)ii | ≤ 1/2 and hence ℜfb(ϕ)ii ≤ 1/2 (see Lemma 2.5). Now consider a one-dimensional irreducible representation ϕ. Since f is ǫ-far from homomorphism, we have q q 1 1 2 2 ǫ ≤ dist(f, ϕ) = kf k2 + kϕk2 − 2ℜhf, ϕi ≤ 2 − 2ℜfb(ϕ). 2 2 Note that kf k22 ≤ 1 and kϕk22 = 1 as ϕ is a (non-zero) homomorphism. Hence, ℜfb(ϕ) ≤ 1 − 2ǫ2 . We have shown that ℜ Ex,y [f (x)f (y)f (x + y)] ≤ 1 − 2ǫ2 . Since |f (x)f (y)f (x + y)| ≤ 1, at least an Ω(ǫ2 )-fraction of pairs (x, y) satisfy f (x)f (y)f (x + y) 6= 1. Hence we have Prx,y [f (x)f (y) = f (x+ y)] ≤ 1− Ω(ǫ2 ). By choosing the hidden constant in s large enough, we reject with probability at least 2/3. 5 Testing the Property of Being Proportional to an Irreducible Character In this section, we show the following: Theorem 5.1. The property of being proportional to an irreducible character is testable with O(1/ǫ8 log2 (1/ǫ)) queries. As any character is a class function, by Lemma 3.2, it suffices to give a tester with query complexity O(1/ǫ8 log(1/ǫ)) that works when the input function is a class function. The following fact is crucial for our algorithm. Lemma 5.2 ([31]). For a function f : G → C, the following are equivalent. 1. f (x) = f (1)e χϕ (x) for some irreducible representation ϕ, where χ eϕ = χϕ /dϕ . 2. f (x)f (y) = f (1) Ez∈G [f (yzxz −1 )] for any x, y ∈ G. 12 Algorithm 3 (Tester for being proportional to an irreducible character) Input: A function f : G → D. 1: Let e1 be the estimation to kf k22 obtained by applying Lemma 2.1 with the error parameter ǫ2 /100 and the confidence parameter 1/100. 2: if e1 < ǫ2 /2 then accept. 3: for each i = 1 to s := O(1/ǫ4 ) do 4: Sample x, y ∈ G uniformly at random. 5: Let ei5 be the estimation to Ez [f (yzxz −1 )] obtained by applying Lemma 2.1 with the error parameter ǫ2 /10 and the confidence parameter 1/100s. 6: Let ei6 = |f (x)f (y) − f (1)e5i |2 . 7: if ei6 > ǫ4 /100 then reject. 8: accept. As we can freely change the value of f (1) by multiplying a constant, the second condition is a necesary and sufficient condition of being proportional to an irreducible character. The most simple test based on Lemma 5.2 is checking whether f (x)f (y) ≈ f (1) Ez [f (yzxz −1 )] (by estimating the latter by sampling z ∈ G a constant number of times). However, we were unable to handle the term Ex,y,z [f (x)f (y)f (yzxz −1 )f (1)] that naturally arises when analyzing this test. Instead, we estimate |f (x)f (y) − f (1) Ez [f (xyzxz −1 )]| and check whether it is small. The detail is given in Algorithm 3. It is clear that the query complexity of Algorithm 3 is O(1/ǫ8 log 1/ǫ). Lemma 5.3. If a function f : G → C is proportional to an irreducible character, then Algorithm 3 accepts with probability at least 2/3. Proof. By the union bound, all the estimations succeed with probability at least 2/3. Below we assume this indeed happens. Recall that f (x)f (y) − f (1) Ez [f (yzxz −1 )] = 0 for any x, y ∈ G by Lemma 5.2. Then for each i i, e6 = |f (x)f (y) − f (1)ei5 |2 = |f (1) Ez [f (yzxz −1 )] − f (1)ei5 |2 ≤ ǫ4 /100 holds for every i ∈ [s]. Hence, we accept with probability at least 2/3. Now we turn to the case that f is ǫ-far from being proportional to an irreducible character. We need the following auxiliary lemma. Lemma 5.4. For any function f : G → C, we have h i 2 ≥ kf k22 min kf − f (1)e χϕ k22 . E f (x)f (y) − f (1) E[f (yzxz −1 )] x,y ϕ z 13 Proof. We have i hX i X −1 −1 b f (ϕ) ϕ (yzxz ) d f (yzxz ) = E E ij ij ϕ z =E z = h z hX dϕ ϕ = = dϕ ϕ X X X i,j∈[dϕ ] X i,j∈[dϕ ] dϕ X ϕ i,j∈[dϕ ] X X dϕ ϕ i,j∈[dϕ ] ϕ fb(ϕ)ij fb(ϕ)ij fb(ϕ)ij i,j∈[dϕ ] X k,l,m∈[dϕ ] X k,l,m∈[dϕ ] X i ϕik (y)ϕkl (z)ϕlm (x)ϕmj (z −1 ) h i ϕik (y)ϕlm (x) E ϕkl (z)ϕjm (z) z ϕik (y)ϕlm (x) k,l,m∈[dϕ ] fb(ϕ)ij ϕij (y) δkj δlm dϕ X ϕkk (x) X X fb(ϕ)ij ϕij (y) = dϕ χ eϕ (x) dϕ ϕ k∈[dϕ ] i,j∈[dϕ ] In the third equality, we used the fact that ϕmj (z −1 ) = ϕjm (z). This follows from the fact that ϕ(z −1 )ϕ(z) = ϕ(1) = I and ϕ(z) is unitary. Therefore, X X X X fb(ϕ)ij ϕij (y) fb(ϕ)ij ϕij (y) − f (1) dϕ χ eϕ dϕ f (x)f (y) − f (1) E[f (yzxz −1 )] = f (x) z = X ϕ dϕ (f (x) − f (1)e χϕ (x)) X i,j∈[dϕ ] ϕ ϕ i,j∈[dϕ ] i,j∈[dϕ ] fb(ϕ)ij ϕij (y) It follows that i hX i h X 2 −1 2 fb(ϕ)ij ϕ(y)ij dϕ (f (x) − f (1)e χϕ (x)) = E E f (x)f (y) − f (1) E[f (yzxz )] = X ϕ,ϕ′ = X ϕ x,y z x,y  ϕ dϕ dϕ′ E (f (x) − f (1)e χϕ′ (x)) χϕ (x))(f (x) − f (1)e x dϕ kf − f (1)e χϕ k22 ≥kf k22 min kf ϕ − X i,j∈[dϕ ] f (1)e χϕ k22 fb(ϕ)ij 2  X i,j∈[dϕ ] i,j∈[dϕ ] i′ ,j ′ ∈[dϕ′ ] ≥ min kf − f (1)e χϕ k22 ϕ X X ϕ dϕ   fb(ϕ)ij fb(ϕ′ )i′ j ′ E ϕij (y)ϕ′i′ j ′ (y) X i,j∈[dϕ ] y fb(ϕ)ij 2   In the third equality, we used the fact that Ey ϕij (y)ϕ′i′ j ′ (y) is equal to 1/dϕ if ϕ = ϕ′ , i = i′ , and j = j ′ , and is equal to zero otherwise. Lemma 5.5. If a function f : G → C with f (1) = 1 is ǫ-far from being proportional to an irreducible character, then Algorithm 3 rejects with probability at least 2/3. Proof. By the union bound, with probability at least 5/6, all the estimations succeed. Below we assume it indeed happens. From Lemma 5.4, when f is ǫ-far, the expectation of e6i is at least i h 2ǫ2 ǫ4 ǫ2 ǫ2 ǫ4 ǫ4 i 2 ≥ (1 − )kf k22 min kf − f (1)e χϕ k22 − ≥ (1 − ) · kf k22 − ≥ . E f (x)f (y) − f (1)e5 ϕ 10 100 5 4 100 25 x,y 14 Algorithm 4 (Tester for unitary equivalence) Input: Functions f, g : G → D(d) Θ(d2 ) times do 1: for s := d3/2 /ǫ 2: Sample U ∈ U (d) with respect to the (normalized) Haar measure of U (d). 3: Let e be the estimation of dist(f, U gU ∗ ) obtained by applying Lemma 2.1 with the error parameter ǫ2 /100 and the confidence parameter 1/6s. 4: if e < ǫ2 /10 then accept. 5: Reject. We also note that ei6 ≤ (1 + ǫ2 /100)2 ≤ 2. Let p = Prx,y [e6i > ǫ4 /100]. Then, we have 2 · p + ǫ4 /100 · (1 − p) ≥ ǫ4 /25, and it follows that p ≥ ǫ4 /100. By choosing the hidden constant in s large enough, we reject f with probability at least 2/3. 6 Testing Unitary Equivalence In this section, we prove the following: Theorem 6.1. The unitary equivalence to g : G → D(d) is testable with d3/2 /ǫ O(d2 ) queries. Our algorithm is described in Algorithm 4. We use the Haar measure on U (d) to sample unitary matrices. We do not need the detailed definition of the Haar measure, and we only have to understand that it defines a probability distribution on U (d). See Section 2.4 for a brief introduction to the Haar measure. The basic idea of our algorithm and analysis is the following. Suppose that functions f, g : G → D(d) are unitary equivalent, that is, f = U0 gU0∗ for some unitary matrix U0 ∈ U (d). Then, by sampling a sufficient number of unitary matrices from the Haar measure, we get a unitary matrix U that is sufficiently close to U0 in the sense that the Frobenius norm of U − U0 is small (Lemma 6.3). Then, we can show that the Frobenius norm of f (x) − U g(x)U ∗ is also small for any x ∈ G (Lemma 6.5). On the other hand, if f and g are ǫ-far from being unitary equivalent, then the average Frobenius norm of f (x) − U g(x)U ∗ over x ∈ G is large for any unitary matrix U . Hence, by checking whether there is a unitary matrix U (in the sample) such that the average Frobenius norm is small, we can distinguish the case that f and g are unitary equivalent from the case that f and g are ǫ-far from being unitary equivalent. Let U ∈ U (d) be a random matrix sampled with respect to the Haar measure. We diagonalize U as U = W ΛW ∗ where W ∈ U (d) and Λ = diag(λ1 , . . . λd ). By the √ unitarity of U , the absolute value of each eigenvalue of U is 1. Therefore we can write λi = exp( −1θi ) for some θi ∈ [−π, π), which we call the phase of λi . We use the following proposition, which is Weyl’s integral formula applied to U (d). Q Proposition 6.2 ([15]). The distribution µ of the phases θ = (θ1 , . . . , θd ) is dµ(θ) = Z1d i>j |λi − λj |2 dθ, where Zd := (2π)d d! is a normalization constant and dθ is a standard Euclid measure. For ǫ > 0, we write BU (d) (ǫ) = {U ∈ U (d) | kU − Id kF ≤ ǫ}. We need the following auxiliary lemma, which says that the set of a sufficiently large number of randomly chosen unitary matrices forms an “ǫ-net” of unitary matrices with respect to the Frobenius norm. 15 Lemma 6.3. Let U0 ∈ U (d) and U be a random matrix sampled with respect to the Haar measure of U (d). For sufficiently small ǫ > 0, the probability Pr[kU − U0 kF ≤ ǫ] ≥ δ6.3 (ǫ, d), where O(d2 ) ǫ δ6.3 (ǫ, d) = d3/2 . Proof. Since the Haar measure is invariant under left multiplication, we can assume U0 = Id without loss of generality. Hence, we want to bound Pr[U ∈ BU (d) (ǫ)]. Let λ1 , . . . , λd be eigenvalues of U and θ1 , . . . , θd be corresponding phases. By the conjugate invariance of Frobenius norm, U ∈ BU (d) (ǫ) P iff di=1 |λi − 1|2 ≤ ǫ2 . √ Suppose kθk2 ≤ 3ǫ/4. Since θi ’s are sufficiently small, we can expand as λi = 1+ −1θi +O(θi2 ). P P Then di=1 |λi − 1|2 = kθk22 + O( di=1 θRi3 ) = kθk22 + kθk22 maxi |θi | < ǫ2 . It implies the probability is bounded below by Pr[U ∈ BU (d) (ǫ)] ≥ Bd (3ǫ/4) dµ where Bd (r) is the ball in Rd of radius r centered at the origin. ǫ , θ0 = [0, ǫ̃, . . . , (d − 1)ǫ̃], and θ1 = [ǫ̃/3, 4ǫ̃/3, . . . , (d − 2/3)ǫ̃]. Note that kθ0 k22 ≤ Let ǫ̃ = d3/2 ǫ2 /3 and kθ1 k22 = ǫ2 (6d2 − d + 2)/18d3 < ǫ2 /2. Therefore, Bd (3ǫ/4) contains the d-dimensional hypercube V = {θ ∈ Rd | θ0 ≤ θ ≤ θ1 }, where we write x ≤ y if xi ≤ yi for each i. Note that if θ ∈ V , |θi − θj | ≥ ǫ̃/3 for any i 6= j. Therefore, we have Z Z Z Y Z Y 1 1 dµ ≥ dµ ≥ |λi − λj |2 dθ ∼ |θi − θj |2 dθ Z Z d V d V V Bd (3ǫ/4) i>j i>j (By the fact that ǫ is small enough) Z Z Y 2   1  ǫ̃ d2 ǫ̃ 1 ǫ̃ d(d−1) 1 dθ = . dθ = ≥ Zd V 3 Zd 3 Zd 3 V i>j Let Skew(d) be the set of d-dimensional skew-Hermitian matrices, i.e., Skew(d) = {X ∈ Md (C) | = −X}. Although the following lemma is an almost immediate consequence of the fact that Skew(d) is the Lie algebra of U (d), we prove it for completeness. X∗ Lemma 6.4. If ǫ > 0 is small enough, then the image of the exponential map exp : Skew(d) → U (d) contains BU (d) (ǫ). Furthermore, if U ∈ BU (d) (ǫ), then we can choose X ∈ Skew(d) such that exp(X) = U and kXkF ≤ 2ǫ. ∗ Proof. 1 , . . . , λd ). Let λi = √ We diagonalize U as W ΛW , where W ∈ U (d) and Λ = diag(λ P∞ n+1 z n /n. This exp( −1θi ) for θi ∈ R and µi = 1 − λi . For z ∈ C, we define √ log(1 + z) = i=1 (−1) Taylor expansion converges for |z| ≤ 1. Note that log λi = −1(θi + 2πmi ) for some mi ∈ Z. Since kU − Id kF ≤ ǫ, we can define X = log U . Note that U = exp(X) as z = exp(log z). We have the following formulas: X = log U = log W ΛW ∗ = W (log Λ)W ∗ , log Λ = diag(log λ1 , . . . , log λd ), √ √ log λi = −1(θi + 2πmi ) = − −1(θi + 2πmi ) = − log λi , (log Λ)∗ = diag(log λ1 , . . . , log λd ) = −diag(log λ1 , . . . , log λd ) = − log Λ. Combining these formulas, X ∗ = W (log Λ)∗ W ∗ = −W (log Λ)W ∗ = −X, that is, X ∈ Skew(d). 16 P Next, we estimate the Frobenius norm of X. The condition kU − Id kF ≤ ǫ implies di=1 |µi |2 ≤ ǫ2 . It implies µi is small enough. So we can expand log λi = log(1 + µi ) = µi + O(µ2i ). Therefore | log λi |2 = |µi |2 + O(µ3i ) and we can conclude kXk2F = kW (log Λ)W ∗ k2F = k log Λk2F = Pd Pd 2 2 3 2 i=1 | log λi | = i=1 |µi | + O(µi ) ≤ 2ǫ . Lemma 6.5. Let A ∈ D(d) be a matrix and U, V ∈ U (d) be unitary matrices. If kU − V kF ≤ ǫ for sufficiently small ǫ > 0, then kU AU ∗ − V AV ∗ kF ≤ 3ǫ. Proof. Let U0 = V ∗ U . By a direct calculation, kU − V kF = kU0 − Id kF and kU AU ∗ − V AV ∗ kF = kU0 AU0∗ − AkF . Therefore, we can assume that V = Id . By Lemma 6.4, there exists X ∈ Skew(d) such that kXkF ≤ 1 and U = exp(2ǫX). Note that U ∗ = exp(−2ǫX). Then, U ∗ AU − A = exp(−2ǫX)A exp(2ǫX) − A = (Id − 2ǫX + O(ǫ2 )J)A(Id + 2ǫX + O(ǫ2 )J) = 2ǫ(AX − XA) + O(ǫ2 )J, where O(ǫ2 )J denotes a matrix with each entry having absolute value O(ǫ2 ). We evaluate the Frobenius norm of the commutator as kAX − XAkF ≤ kAXkF + kXAkF = 2kXkF ≤ 2, where we use the triangle inequality of the Frobenius norm, and the assumption kAkF ≤ 1. Therefore, kU AU ∗ − Ak2F ≤ 4ǫ2 kAX − XAk2F + O(ǫ3 ) ≤ 8ǫ2 + O(ǫ3 ) < 9ǫ2 . Now, we prove Theorem 6.1. Proof of Theorem 6.1. We assume that the all the estimations have succeeded, which happens with probability at least 5/6 by the union bound. Suppose that f and g are unitary equivalent and U0 ∈ U (d) be the unitary matrix with f (x) = U0 g(x)U0∗ for any x ∈ G. From Lemma 6.3, we sample U such that kU0 − U kF ≤ ǫ/10 with probability at least 5/6. From Lemma 6.5, we have Ex∈G [kf (x) − U g(x)U ∗ k2F ] ≤ 9ǫ2 /100. For such U , we obtain the estimation e satisfies e < ǫ2 /10. By the union bound, we accept with probability at least 2/3. Suppose that f and g are ǫ-far from being unitary equivalent. For every U we sample, we have Ex∈G [kf (x) − U g(x)U ∗ k2F ] ≥ 4ǫ2 . Hence, the estimation e satisfies e > ǫ2 /10 and we accept. To summarize, we reject with probability at least 5/6 > 2/3. Acknowledgements We thank Mitsuru Kusumoto and anonymous referees for comments that greatly improved the manuscript. 17 References [1] E. Allender, J. Jiao, M. Mahajan, and V. Vinay. Non-commutative arithmetic circuits: depth reduction and size lower bounds. Theoretical Computer Science, 209(1-2):47–86, 1998. [2] N. Alon and S. Lovett. Almost k-wise vs. k-wise independent permutations, and uniformity for general group actions. Theory of Computing, 9(1):559–577, 2013. [3] S. Arora, C. Lund, R. Motwani, M. Sudan, and M. Szegedy. Proof verification and the hardness of approximation problems. Journal of the ACM, 45(3):501–555, 1998. [4] M. Bellare, D. Coppersmith, J. Hastad, M. Kiwi, and M. Sudan. Linearity testing in characteristic two. IEEE Transactions on Information Theory, 42(6):1781–1795, 1996. [5] M. Ben-Or, D. Coppersmith, M. Luby, and R. Rubinfeld. Non-abelian homomorphism testing, and distributions close to their self-convolutions. Random Structures & Algorithms, 32(1):49– 70, 2007. [6] E. Ben-Sasson, M. Sudan, S. Vadhan, and A. Wigderson. Randomness-efficient low degree tests and short PCPs via epsilon-biased sets. In Proceedings of the 35th annual ACM symposium on Theory of computing (STOC), pages 612–621, 2003. [7] P. Berman, S. Raskhodnikova, and G. Yaroslavtsev. Lp -testing. In Proceedings of the 46th Annual ACM Symposium on Theory of Computing (STOC), pages 164–173, 2014. [8] A. Bhattacharyya. Guest column: on testing affine-invariant properties over finite fields. ACM SIGACT News, 44(4):53–72, 2013. [9] A. Bhattacharyya, E. Fischer, H. Hatami, P. Hatami, and S. Lovett. Every locally characterized affine-invariant property is testable. In Proceedings of the 45th Annual ACM Symposium on Theory of Computing (STOC), pages 429–436, 2013. [10] A. Bhattacharyya, E. Fischer, and S. Lovett. Testing low complexity affine-invariant properties. Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1337–1355, 2012. [11] A. Bhattacharyya, S. Kopparty, G. Schoenebeck, M. Sudan, and D. Zuckerman. Optimal testing of Reed-Muller codes. In Proceedings of the 51st Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 488–497, 2010. [12] P. Biane. Characters of symmetric groups and free cumulants. Asymptotic Combinatorics with Applications to Mathematical Physics, 1815:185–200, 2003. [13] E. Blais, A. Weinstein, and Y. Yoshida. Partially symmetric functions are efficiently isomorphism-testable. In Proceedings of the 53rd Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 551–560, 2012. [14] M. Blum, M. Luby, and R. Rubinfeld. Self-testing/correcting with applications to numerical problems. Journal of Computer and System Sciences, 47(3):549–595, 1993. 18 [15] T. Bröcker and T. Dieck. Representations of Compact Lie Groups. Graduate Texts in Mathematics. Springer, 1985. [16] B. Green. A Szemerédi-type regularity lemma in abelian groups, with applications. Geometric and Functional Analysis, 15(2):340–376, 2005. [17] H. Hatami and S. Lovett. Estimating the distance from testable affine-invariant properties. In Proceedings of the 54th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 237–242, 2013. [18] J. Huang, C. Guestrin, and L. Guibas. Fourier theoretic probabilistic inference over permutations. The Journal of Machine Learning Research, 10:997–1070, 2009. [19] M. Kassabov. Symmetric groups and expander graphs. Inventiones Mathematicae, 170(2):327– 354, 2007. [20] T. Kaufman and M. Sudan. Algebraic property testing: the role of invariance. In Proceedings of the 40th Annual ACM Symposium on Theory of Computing (STOC), pages 403–412, 2008. [21] S. V. Kerov. Talk at IHP conference. 2000. [22] A. Kirillov. Introduction to Lie groups and Lie algebras. Cambridge University Press, 2008. [23] R. O’Donnell and J. Wright. Quantum spectrum testing. In Proceedings of the 47th ACM Symposium on Theory of Computing (STOC), 2015. [24] R. Raz and B. Spieker. On the “log rank”-conjecture in communication complexity. Combinatorica, 15(4):567–588, 1995. [25] D. Ron. Algorithmic and analysis techniques in property testing. Foundations and Trends R in Theoretical Computer Science, 5(2):73–205, 2009. [26] R. Rubinfeld. On the robustness of functional equations. SIAM Journal on Computing, 28(6):1972–1997, 2006. [27] R. Rubinfeld and M. Sudan. Robust characterizations of polynomials with applications to program testing. SIAM Journal on Computing, 25(2):252–271, 1996. [28] A. Shpilka and A. Wigderson. Derandomizing homomorphism testing in general groups. SIAM Journal on Computing, 36(4):1215–1230, 2006. [29] R. P. Stanley. Irreducible symmetric group characters of rectangular shape. Lotharingien de Combinatoire, 50, 2003. Séminaire [30] B. Steinberg. Representation Theory of Finite Groups: An Introductory Approach. Springer, 2011. [31] H. Weyl. Quantenmechanik und gruppentheorie. Zeitschrift für Physik, 46(1-2):1–46, 1927. [32] K. Wimmer. Agnostically learning under permutation invariant distributions. In Proceedings of the 51st Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 113–122, 2010. 19 [33] Y. Yoshida. A characterization of locally testable affine-invariant properties via decomposition theorems. In Proceedings of the 46th Annual ACM Symposium on Theory of Computing (STOC), pages 154–163, 2014. 20
8cs.DS
Proceedings 15th International Conference on Information Fusion, Singapore 2012 Online Learnability of Statistical Relational Learning in Anomaly Detection Magnus Jändel, Pontus Svenson and Niclas Wadströmer Swedish Defence Research Agency Sweden Contact: [email protected] Abstract— Statistical Relational Learning (SRL) methods for anomaly detection are introduced via a security-related application. Operational requirements for online learning stability are outlined and compared to mathematical definitions as applied to the learning process of a representative SRL method – Bayesian Logic Programs (BLP). Since a formal proof of online stability appears to be impossible, tentative common sense requirements are formulated and tested by theoretical and experimental analysis of a simple and analytically tractable BLP model. It is found that learning algorithms in initial stages of online learning can lock on unstable false predictors that nevertheless comply with our tentative stability requirements and thus masquerade as bona fide solutions. The very expressiveness of SRL seems to cause significant stability issues in settings with many variables and scarce data. We conclude that reliable anomaly detection with SRL-methods requires monitoring by an overarching framework that may involve a comprehensive context knowledge base or human supervision. Keywords: Relational detection; Online learning; Learning stability; Statistical Learning, Bayesian Logic Programs; Anomaly I. INTRODUCTION A. Application scenario Real-time abnormal behavior detection is an important component of many security applications today. A simple example is a facility surveillance system, where one wants to detect anomalies and take appropriate action to either prevent them or mitigate their consequences. Generally, it will be necessary to combine information from many different sources in order to detect the anomalies; hence fusion is a necessary component of the surveillance systems. The anomalies of interest could range from terrorists planning an attack to more pedestrian situations such as theft by employees or customers in a port facility. Even though it is the terrorist cases that presently get the most attention, it is important not to forget the possibilities to detect everyday crime using the same systems. The research described in this paper is motivated by applications in EU FP7 projects on supply chain security [1] [2] [3].There are several different aspects of supply chain security [4] [5]. Perhaps most obvious is the scanning and screening of cargo and bags that take place in airports. Similar scanning is done on containers in ports and on postal items in sorting facilities. Ideally all cargo/containers should be scanned but economic realities favor a risk-based approach, where a risk 1150 index is computed for each cargo item, and only those with a high risk index are subjected to the most expensive scanning. One way of computing risk indicators is to use abnormal behavior detection on all available data regarding the item. This can include information from sensors and other sources. For instance, relevant information for containers might include senders and receivers of all goods packaged inside the container, the owner and crew of the ship transporting it, and the journey details of the ship. This set of information is highly heterogeneous, which complicates the task of fusing and analyzing it. Furthermore, the analysis of what constitutes a threat and not is not static, but will change dynamically. Hence, there is a need for methods for fusion of heterogeneous information and anomaly detection that can adapt to changing threat profiles. An equally important application of anomaly detection for supply chain security is in facility protection and surveillance. While abnormal behavior detection in crowds is more common in mass transport security applications, several similar issues arise also in facility surveillance. Consider a port where both passengers and employees are at large. Anomaly detection can be used to detect individuals who appear to be preparing criminal acts. Here, too, the threat patterns are evolving quickly and there is a need for online learning to adapt the anomaly detection system to the changing conditions. A full description of an anomaly detection system is out of the scope of this paper. We will however assume that the system consists of a set of sensors that give information about the current state of the world. This information is fused to produce higher-level situation descriptions which are fed into the anomaly detection system. In this paper we consider anomaly detection to be a form of classification where the system learns from a training set including both normal and anomalous situations. The trained system classifies situations as either normal or anomalous and operators are alerted to detected anomalies. Note that the availability of labeled data including both normal and anomalous events often is an issue in the supervised classification approach to anomaly detection employed here. For simplicity we assume that there is only one type of anomaly although practical applications typically include many anomaly types. Alternative paradigms, not considered here, include unsupervised learning where lowprobability events are considered anomalous. is expressed as a set of Horn clauses 1 . In the following we will use the implication format of a Horn clause and it is understood that “clause” means “Horn clause”. The core of the anomaly detection system is assumed to be based on Statistical Relational Learning (SRL) [6]. A key requirement is that the classification performance of the continuously learning anomaly detection system is stable. In this paper we therefore investigate the online stability of SRL in the context of our application scenario. B. Statistical Relational Learning Extending machine learning to take account of the rich relational structure of real-world problems is the goal of the emerging science of SRL (see [6] for a comprehensive overview). Many different methods have been proposed including Bayesian Logic Programs (BLP) [6] which is the SRL method selected for the present investigation. SRL methods describe a very broad class of probability density functions (PDFs) using a two-tier representation consisting of a structural part and a parametric part. The structural part is usually written in the language of first order logic while the parametric part expresses the statistical significance of the relational structure. An explicit representation of a PDF, typically in the form of a Bayesian network or a Markov random field, is generated from the SRL models in response to a query. This explicit PDF representation is then used to compute the answer to the query, e.g., the probability that an observed event is anomalous. SRL enables potentially a wide range of powerful applications but the profuse expressiveness of SRL is also the source of the stability issues that are discussed in the following. C. Bayesian Logic Programs A BLP consists of a set of ground terms, a set of logical atoms and a set of so-called Bayesian clauses that represent relations. The ground terms represent known objects. Logical atoms are well-founded facts and relations including, in particular, universally quantified relations. The ground terms together with the logical atoms form a base-line logical knowledge base KB0. A Bayesian clause describes a dependency relation in the form of a logical clause and is associated with probabilistic parameters describing the statistical significance of the relation. The statistical parameters take the form of a conditional probability table (CPT). The set of Bayesian clauses constitute an extended knowledge base KBB. A total knowledge base KB is formed by merging KB0 and KBB. Consider a new object e.g. Q23 that has been discovered by the surveillance system. First the new ground term and the observed facts are added to the knowledge base. Then the BLP system estimates the degree of anomalousness of the object in three steps. A) The knowledge base KB is first used to construct a logical proof of the predicate Anomalous(Q23). For simplicity we assume that precisely one such proof is found. Ref. [6] describes how combination rules are used for handling multiple proofs. The logical proof B) The set of clauses that constitute the proof can be viewed as a network with predicates as nodes and inference relations as edges. By transforming predicates to random variables and inference relations to statistical dependency relations, the proof is converted to a Bayesian network that is called the support network of the query. Each node in the support network inherits a CPT from the Bayesian clause that is the template of the node. The CPT codifies the statistical dependency between the dependent variable and its parents. The specific statistical parameters of the support network are hence instantiated from the generic parameters of the BLP. C) The probability of Q23 being anomalous is finally estimated by inference in the support network. In summary, queries to the BLP system are requests for estimates of the marginal probability distribution of random variables. The system interprets random variables as predicates, attempts to prove the predicates, converts the proof to a support network and estimates the requested marginal probability distribution from the support network. The main challenge is how to learn Bayesian clauses from data. Presently this is done in a two-stage loop where the space of relations first is explored for the purpose of identifying candidate sets of clauses. The second step strives to find optimal statistical parameters for each candidate set of clauses. By selecting the best set of Bayesian clauses as the starting point of the next iteration, the learning process eventually converges. BLP learning is further discussed in section II A. D. Introduction to online learning in SRL SRL learning processes including the BLP learning algorithm that is outlined in the previous section are primarily designed for batch processing where the learning process is executed once using all available data. In many applications, including our surveillance scenario, it is more natural to consider online learning where the system learns continuously from a stream of incoming data while it simultaneously must respond to queries. Disregarding computational efficiency and memory issues it is possible to run a batch learning process on all the accumulated data as soon as new data arrives. It is, however, vital that online learning is stable in security-critical systems. A single new training example could otherwise cause 1151 1 Recall that a Horn clause is a disjunction of literals, where all except at most one literal is negated. It can also be viewed as an implication, where the literal that was negated is implied from the conjunction of the others. Universally stable systems can be trusted to learn online indefinitely. a sudden performance degradation. Definitions of stability are discussed from an operational point of view in section I E and in the context of learning theory in section II B. Discussion of SRL online learning in the literature seems to focus on computational complexity aspects. Online learning of statistical parameters in Markov Logic Networks (MLN) is considered in [7], [8] and [9]. Huynh and Mooney address online learning of both structural and statistical aspects of MLN and introduce a regularization method that suppresses the creation of new clauses while learning online [10]. For reasons that are spelled out in the following, universal stability in SRL online learning is a chimera at the present state of the art. We can strive for enhancing online stability but never reach the epitome of universal stability. E. Operative requirements for online stability in anomaly detection Generic critical properties of anomaly detection systems are the false alarm rate Pfalse and the missed alarm rate Pmissed. The former quantity is the probability of erroneously detecting an anomaly in a normal situation while the latter quantity is the probability of not detecting an anomaly in an anomalous situation. Pfalse is the same as False Positive Rate in machine learning, while Pmissed is the False Negative Rate. The anomaly detection system is useful only if Pfalse and Pmissed fall below thresholds Tfalse and Tmissed respectively where threshold values depend on contingent operative requirements. Furthermore, the system should be able to run with reliable good performance between maintenance check-points. We assume here that the system runs unattended except for regularly scheduled maintenance occasions when an operator monitors the performance of the system and decides if it can be relied on. This means that the online learning algorithm should comply with performance thresholds for at least m* new training examples where m* is given by the maximum rate of incoming training examples and the maintenance period. Assuming that m training examples have been processed initially practical requirements for online stability hence are that , ∀m´: 0 < m´≤ m * Pfalse (S m + m´ ) < T false , (1) ∀m´: 0 < m´≤ m * Pmissed (S m + m´ ) < Tmissed , (2) and where Sm means a training set consisting of m samples. An anomaly detection system such that, ∀m : Pfalse (S m +1 ) ≤ Pfalse (S m ) , (3) ∀m : Pmissed (S m +1 ) ≤ Pmissed (S m ) , (4) and is guaranteed to be stable for any value of the operational parameters Tfalse, Tmissed and m*. A system with this property is universally stable since adding a new training example to a system with satisfactory performance will not, under any circumstances, impair the performance of the system. 1152 II. ANALYSIS OF ONLINE STABILITY This section provides a more detailed description of BLP learning, a brief review of online stability in machine learning and discussion of a simplified BLP model where sources of online instability are identified and discussed. A. BLP Learning In the following we review BLP learning as described in [6] and [11]. The goal of BLP learning is to infer a BLP from a set of training examples S. Each training example Si ∈ S consists of a list of random variables where values may be assigned to some of the variables e.g. S56 = {flies(Donald) = ?, colour(Donald) = white, father(Quackmore, Donald) = true, mother(Hortense, Donald) = true }. Inductive Logic Programming (ILP) [12] is first applied to create the initial set of clauses from the data. For this purpose each training example is viewed as list of logical assertions where random variables are transformed to predicates and the assigned values are ignored. The training example S56 becomes for example the assertion {flies(Donald), colour(Donald), father(Quackmore, Donald), mother(Hortense, Donald)}. Using the transformed data set and optionally background knowledge in KB0, the ILP engine generates a set of universally quantified Horn clauses codifying generic relations that can be learnt from S and KB0. Note this set of clauses typically is not unique. ILP delivers hence the structure of an initial BLP where each Horn clause becomes a Bayesian clause with yet unknown statistical parameters. Bayesian support networks are generated by instantiating the universally quantified variables with ground terms from the training set. The statistical parameters of the Bayesian clauses are computed by maximizing the combined likelihood of the support networks with respect to S (now taking variable values into account). The value of the combined likelihood for optimal statistical parameters becomes the utility of the initial BLP. Refinement operators are now applied to the initial BLP so that a set of candidate BLPs are generated. In the presently studied implementation of BLP [11] two refinement operators are employed that, applied to the body of a clause, either adds or removes a universally quantified predicate. Viewed as a set of Horn clauses, each candidate is required to be consistent with KB0. Viewed as a BLP, each candidate is scored by computing statistical parameters as before and evaluating utility by summing up the combined maximum likelihood of the support networks. The candidate with the highest utility is selected as the next preferred solution. The learning process consists hence of a series of iterations where refinement operators are applied and the best candidate BLP is selected. When no further significant improvement in utility is achieved, the learning process is deemed to have converged. Note that this process must be constrained so that only acyclic support networks are generated and that the maximum likelihood computation must handle combining rules. This section is only intended as a brief overview of BLP learning and rests completely on Ref. [6] and [11] that should be consulted for further details. B. Formal stability conditions for online learning Ross and Bagnell [13] provide an up-to-date review on what is known about stability conditions for online learnability in the context of the General Setting of Learning (GSL) [14]. In the online version of GSL an algorithm A observes a sequence of data points z1 , z 2 ...z m ∈ Z . Each time the algorithm has received a subsequence z1 , z 2 ...zi −1 it must select a hypothesis hi ∈ H and suffer a penalty given by the loss function f : H × Z → R applied to the next data point zi. Symbols Z and H denote the space of data and hypotheses respectively. The problem of online stability in anomaly detection, as discussed in section I E, is strictly not a GSL problem since there are several different objectives and the problem therefore belongs to the class of multiobjective optimization problems [15]. To transform our stability requirements to a GSL problem we must define a computationally feasible loss function that captures the combined operational constraints expressed in terms of available data. classes of algorithms notably those with a convex loss function. Convexity means that there are no local optima so that a greedy gradient ascent algorithm is guaranteed to find the global optimum. SRL problems typically feature a jagged fitness landscape that violates the convexity condition. The BLP learning process that is outlined in the previous section involves greedy gradient search with respect to both structure and parameters. Since it is unknown if BLP is online learnable and the convergence rate anyway is unknown it appears that it presently is difficult to use the formal online stability conditions of Ross and Bagnell as a basis for testing the practical online stability of BLP applications. C. A tentative rule of thumb for SRL online stability Because of the great potential advantages of using SRL methods it is tempting to ignore the lack of formal stability assurance and embrace a more cavalier approach according to the following sketch. Since real-valued statistical parameters always will vary under the impact of accumulating training samples it appears that the permanence of the “digital” clause structure is the best signature of stability. The clause structure should hence be monitored according to the following guidelines, Ross and Bagnell define the regret of the sequence of hypotheses in GSL as, m Rm = m ∑ f ( h , z ) − min ∑ f ( h, z ) . i i =1 i h∈H i (5) i =1 An algorithm has no regret at convergence rate ε regret ( m ) if for any sequence of data and for all m, Rm / m ≤ ε regret ( m) , 1) Run the online learning process until the structure of the solution is stable for a time that is significantly longer than the maintenance interval. 2) Make sure that only minor refinements of the structure have occurred over the last few modifications of the solution. It is reasonable to assume that only small simpifications or specializations of the logical structure occurr as the SRL system converges to a bona fide solution. 3) Monitor the life time (the number of new online training samples that a structure survives). Certify that the life time of consecutive structures increases since this can be assumed to be a signature of approaching a valid solution. Our tentative rule of thumb amounts to accepting the solution for operational use if all of the conditions 1-3 are satisfied. Note that the requirement 2) of small structural changes only is used for verifying stability. Early learning phases may include large structural changes where the (unstable) system breaks out of local minima. (6) where the convergence rate is a monotonously non-increasing function of m and ε regret → 0 as m → ∞ . If an algorithm has no regret the problem is online learnable. III. Ross and Bagnell discuss online learnability for the major classes of machine learning algorithms. It is not known if any SRL method is online learnable according to the definition of Ross and Bagnell. Even if we assume that an SRL algorithm has no regret and hence is online learnable we would only be able to say that the system will comply with operational stability requirements for sufficiently large m without knowing how much data that actually is needed for a given system and situation. To assess if the system is online stable for any specific finite m we also need an estimate of ε regret ( m ) . The convergence rate is, however, known only for some special 1153 ANALYSIS OF ONLINE STABILTY IN A SIMPLIFIED MODEL For the purpose of better understanding online stability in BLP and similar SRL concepts we will discuss a highly simplified model where stability issues are more tractable than in generic SRL learning. This toy model will also be the basis of numerical experiments where we test the rule of thumb in the previous section. A. A toy model Consider a set of n+2 random variables XA, X0, X1, … Xn where each variable has a binary domain X i ∈ {0,1} . The ground truth PDF is given by the following description. All variables with the exception of XA are independent with a uniform probability distribution. The anomalousness variable XA depends on X0 so that XA with 80% probability has the same value as X0. Applying BLP6 to a test set obviously gives much worse performance than BLPT thus illustrating that BLP learning, at least in some situations, might reduce performance by overfitting to statistical fluctuations in the training sample. A training set Sm consists of m complete samples of all variables drawn from the ground truth PDF. We are interested in learning a BLP from the training set and then predict the value of XA from the values of the other variables in a test sample. B. False predictors The set {X17, X34, X47, X61} together with the associated CPT is an example of a false predictor where a set of variables accidentally codes for another variable in a training set although the variables really are independent. In this section we will analyze how false predictors impact the BLP online learning process. To prepare for BLP learning we consider each training example as a set of values {xA(yi), x0(yi), x1(yi), … xn(yi)} where xj(yi) is the value of random variable Xj(yi) , yi is the value of a dummy random variable Y and i is the index of the example. The Bayesian clause that fully models the ground truth probability distribution of relevance for the anomalousness indicator is, X A (Y ) : − X 0 (Y ) , To facilitate this analysis we shall further simplify the toy model by assuming that the CPTs that are associated with each Bayesian clause consist only of binary-valued conditional probabilities p ∈ {0,1} . The structure of the single Bayesian clause of our toy model is represented by a list of body variables such as {X17, X34, X47, X61}. We further define a pattern as a bit pattern of body variable values e.g. {X17=1, X34=0, X47=1, X61=0}. The binary CPT of the Bayesian clause with s variables in the body is hence represented by 2s rows each row consisting of a pattern and an associated conditional probability p ∈ {0,1} where p=1 means that XA=1 with certainty. The total number of different CPTs for a given structure of size s is countable in this model and amounts (7) where Y, in the structural perspective, is a universally quantified logical variable representing a training example label and a CPT describes the relevant probability distribution. In addition to (7) there are n redundant independent binary variables. This will in the following be called BLPT, the ground truth BLP. Consider now a BLP learning process according to section II A where the initial model BLP0, for good measure, is taken to be identical to BLPT. Note that the learning process outlined in section II A normally would generate multiple initial clauses. For the purpose of creating an analytically tractable problem we assume here that the initial structure learning algorithm has found the ideal starting point and that hence just one clause is used in the learning process. A sufficiently small initial training set includes redundant variables that accidentally are strongly correlated with XA and the total likelihood of the training set is improved by applying the expanding refinement operator to include such variables in the body of the clause as for example in the following sequence, BLP1 : X A ( Y ) : − X 0 ( Y ) , X 17 ( Y ) to 2 2s . The variables X1, … Xn that are decoupled from XA are called redundant variables. A redundant structure is a structure that consists of redundant variables. A false predictor is a Bayesian clause consisting of a redundant structure and a CPT where the clause classifies the values of XA in the training set with 100 % accuracy. Under the assumptions of the toy model, the average number of false predictors containing s variables drawn from a set of n variables {X1, … Xn} and applied to a training set consisting of m examples is, N nsm = ⎛ n ⎞ 2 2s − m . ⎜ ⎟ ⎝s⎠ (8) The binomial coefficient in the equation corresponds to the number of ways that a structure of s variables can be selected BLP2 : X A ( Y ) : − X 0 ( Y ) , X 17 ( Y ) , X 47 ( Y ) …. 2 BLP5 : X A ( Y ) : − X 0 ( Y ) , X 17 ( Y ) , X 34 ( Y ) , X 47 ( Y ) , X 61 ( Y ) Each step increases the combined likelihood of the training set and is therefore a valid learning iteration according to section II A. In the present batch of training examples, a set of redundant variables in the body of BLP5 {X17, X34, X47, X61} just happens to code for the values of XA in the training set with 100 % accuracy since each combination of values for the set either maps to a given value of XA or does not occur in the training set. Applying the reductive refinement operator in the next learning iteration hence produces the 100% accurate BLP, s from a total of n variables. The 2 factor corresponds to the total number of CPTs for a given structure. Each new training example has an even chance of falsifying each of the false predictors thus explaining the negative factor m in the exponent. By approximating the binomial coefficient, it is easy to show that for large n and under certain other s n+2 −m conditions N nsm 2 . False predictors are hence ubiquitous e.g. for n = 20, s = 10 and m = 1000. Note that the 2s term in the exponent of (8) is characteristic for the rich hypothesis space of relational algorithms and would be absent for non-relational machine learning methods. If a false predictor has been established as the preferred hypothesis it will be challenged by further online learning. Each new training example will, under the BLP6 : X A ( Y ) : − X 17 ( Y ) , X 34 ( Y ) , X 47 ( Y ) , X 61 ( Y ) . 1154 assumptions of the toy model, have a 50 % chance of falsifying the false predictor. A false predictor survives hence in average for two online learning cycles under the assumptions of the toy model. collapse when the true solution is found. In general we expect a sharp shift in structural size as the online learning process exits from the false predictor phase and transits from a highly specialized false predictor to a relevant solution. During initial online learning there will be a phase where the fitness landscape of the loss function includes many false minima each associated with a false predictor. This phase will be called the false predictor phase. As the online learning process proceeds, the population of false predictors is decimated. The life time of each selected hypothesis is short and the learning process will rapidly switch between false optima. Relevant solutions are found only after almost all false predictors have been falsified. In the toy model there is a 50 % probability that all false predictors are falsified after approximately (to leading order in n), The life time of the selected redundant structure will also increase as the training set expands. Larger structures correspond to a larger supply of possible CPTs that will require more training examples to falsify. Increased structural life time is therefore not per se an indicator of convergence to a stable and relevant solution. m1/2 = 2 n ( +1) 2 , (9) training examples. In generic SRL problems it will in general be impossible to estimate how much data that is required to suppress false predictors since prior knowledge about the true joint probability distribution typically not is available. From the toy model we shall endeavor to extrapolate some generic insights about SRL online learning. In particular, we evaluate the rule of thumb in section II C by studying the evolution of structure in online learning. In preparation for this investigation quantities that characterize the structure life cycle are defined. We expect the average hop size to be fairly constant while the learning process is in the false predictor phase. The learning process favors minimal hops and the distribution of false predictors over the fitness landscape should be uniform. Let us now relax the toy model restriction to binary CPTs and allow real-valued probabilistic parameters. We expect the same general behavior but with a smoother fitness landscape since the expanded model is less constrained. One single training example will now not be able to falsify a false predictor but will only weaken the likelihood of the hypothesis thus extending the life span. A more complex relational structure should not fundamentally change the picture. The size and life time of preferred redundant structures should still increase with increasing size of the training set. Hence we expect that the behavior of false predictors that we discussed in the context of the toy model qualitatively will remain the same in generic online learning of BLP and other similar SRL models. The life time of a structure is the number of training examples in online learning from the point where the structure is selected by the learning algorithm to the point where it is discarded and replaced by a different structure. The minimum life time is one. A redundant structure will be more long-lived than the individual Bayesian clauses that incorporate the structure since each structure correspond to many Bayesian clauses with different CPTs each having some probability of being a false predictor. The structural size is a measure of the logical complexity of a BLP. In our toy example the structural size is simply the number of variables in the body of the clause. The structural distance between two BLP models is the minimum number of refinement operations that is required to transform one of the models to be identical to the other model. For two Bayesian clauses with the same head, the structural distance is just the number of predicates that differ between the clauses. The hop size is the difference in structural distance between two consecutive solutions. As the number of training examples increases during online learning, false predictors will be depleted at the same rate independently of the structure but smaller structures are associated with fewer CPTs and will hence run out of options sooner. Starting from a structurally simple hypothesis, we expect the average structural size of the preferred solution to increase gradually during online learning in the false predictor phase. In our example, the structural size would suddenly 1155 IV. EXPERIMENTAL SETUP AND RESULTS To confirm the theoretical arguments in section III we have performed numerical experiments where the toy model described in section III A is used for generating training samples. We apply a BLP learning algorithm as described in section II A modified with the constraint to binary CPTs. The loss function is the inverse of the likelihood of the training data according to the hypothesis. For each training set a best fitting hypothesis is found. The training set is incremented by one sample and the learning process is repeated on the expanded training set. The structural size, structural life time and hop size (see definitions in section III B) are measured and averaged over many online learning histories. Table 1 quantifies the evolution of structural size and life time for a progression of training set sizes. Each row of the table presents the stability indicators for a given range of training set sizes averaged over 1000 online training histories with data drawn from the ground truth distribution. We assume that all simulated observations are accurate so that there is no uncertainty in the data. For this experiment we used a model with 12 redundant variables. Training set range Average structural size Average structural life time 0 - 19 3,69 ± 1,83 3,2 ± 2,08 20 -39 7,24 ± 1,06 7,52 ± 5,01 40 - 59 8,79 ± 0,89 12,84 ± 8,31 60 - 79 9,73 ± 0,85 18,94 ± 11,70 80 - 99 10,38 ± 0,80 26,72 ± 15,84 100 -119 10,86 ± 0,73 34,82 ± 19,62 120 -139 11,17 ± 0,64 41,52 ± 22,44 140 - 159 11,44 ± 0,58 53,42 ± 27,96 160 179 11,64 ± 0,50 63,38 ± 28,28 180 -199 11,84 ± 0,37 84,18 ± 37,81 200 - 219 11,87 ± 0,34 112 ± 29,80 220 - 239 11,90 ± 0,30 126 ± 10,65 Table 1. Average structural size and structural life time for the preferred solution according to the learning algorithm described in section II A. The standard deviation d of the tabulated values is indicated by (± d). The data is batched over a range of training set sizes in order to make trends visible in spite of the large variance of the structural life time. Note that only false predictors contribute to the data shown in Table 1. The learning process is terminated as soon as the system finds the ground truth hypothesis. Equation (9) gives an idea of how many training samples that are be needed to suppress the false predictors so that the algorithm captures the relevant solution. In the experiments we found that the average number of training samples needed for exiting the false predictor phase is 116 ± 61 where again the standard deviation is indicated. In the experiments we also measured the hop size and found that the average hop size always is very close to unity independently of the training set size. Our experiments illustrate and confirm the effects that was presaged in section III namely that both the size and the life time of redundant structures increases with increasing training size in the false predictor phase. V. CONCLUSIONS Comparing the theoretical and experimental results on the toy model with the rule of thumb suggested in section II C we note that the evolution of solution structure in the zone of false predictors is dangerously close to satisfying all the suggested stability requirements. False predictors can masquerade as relevant solutions since, 1) The life time of redundant structures increases with the number of redundant variables and might be longer than the maintenance interval. In a real-life situation it is impossible to assess if this is the case since the true joint probability distribution is unknown and the relational structure is much more complex than in the toy model. 2) Redundant structures suffer only minor refinements for the later part of the false predictor phase. 3) The life time of consecutive redundant structures increases as training examples accumulate. Redundant structures imitate our expectations for valid solutions mainly because larger structures have an exponentially larger number of associated CPTs that take a progressivly longer time to falsify. Obviously it is only in the case of binary CPTs that we can talk about the number of CPTs but the survivablity of redundant structures is actually expected to increase for real-valued CPTs since the parameter space is less constrained. The masquerading effect is hence expected to preavail for generic SRL models and the effect is in fact unique for relational models. More specialized redundant hypotheses are in general easier to falsify with new data in classical machine learning while more specialized redundant SRL structures are harder to falsify. The BLP learning algorithm that is briefly reviewed in section II A and comprehensibly described in [11] could be amended by more sophisticated search and regularization methods although the scope for such improvements is severely limited by computational efficiency issues. Replacing greedy hill climbing with search techniques designed for a rough fitness landscape, such as simulated annealing, would reduce the risk of ending up in a false optimum. Note, however, that false predictors typically are global optima in the false predictor phase and hence prevail even if robust search techniques are applied. The learning algorithm of the KReator toolkit [16] includes simple regularization since the default setting limits the number of predicates in the body of a Bayesian clause to three. A wide range of alternative methods for discriminating against too complex solutions can be tried including replacing the hard cut-off with a soft weighting and also entropy-based methods. It is, however, often difficult to estimate the complexity of the wanted solution. Too strict regularization could emasculate the expressive power of SRL. No foreseeable modification of the learning algorithm is hence a panacea for the online stability issue. Better search methods are not a certain remedy since false predictors can have a lower loss function value than valid solutions. Regularization favors simple solutions but the main argument for SRL is the ability to model complex real-world relational structures so simplicity cannot be driven too far. Presently it appears that online learning stability can be guaranteed only by selecting non-relational machine learning methods with convex loss functions where the form of the convergence rate (6) is known. Relational algorithms such as BLP are not provable online stable and our analysis indicates that false predictor structures can behave as bona fide solutions according to our rules of thumb in section II C. The online learning stability of SRL algorithms is a therefore in doubt since real-life SRL scenarios often include a large number of variables, a limited supply of training examples and little a priori knowledge about the ground truth probability distribution. Since the underlying probability distribution is unknown it is difficult to estimate how much initial training data that is needed to eradicate false predictors and generate a relevant hypothesis. Because the life time of false predictors is unknown it is difficult to know when the learning process has converged to a hypothesis that can be trusted to be sufficiently stable according to the operational 1156 requirements that were discussed in section I E. These problems remain even in an ideal case where the algorithm fortuitously is supplied with the correct model for example from human expert. As the algorithm continues to learn online it can still get lost in the jagged fitness landscape ACKNOWLEDGMENT Should one in spite of these issues select to use a SRLbased solution in e.g. a surveillance application we can presently just offer the following advice for enhancing online stability. Helpful advice on BLP learning and on the KReator toolkit from Matthias Thimm is gratefully acknowledged. This research was supported by the SecurityLink strategic research centre as well as by the European Commission under Grant No 261679 (CONTAIN). REFERENCES 1) Design a loss function that reflects operational requirements (rather than some default loss function of the SRL package). 2) Approve the system for operational use only when all of the following conditions are satisfied, a) The rules of thumb in section II C are satisfied. [1] [2] [3] [4] [5] b) The solution gives satisfactory performance from an operational point of view. c) The selected structure is consistent with an appropriate domain knowledge base (see discussion in section I C) [6] d) The selected structure makes sense according to human expert evaluation that is repeated for each maintenance interval. The first requirement ensures that online training is directed to improving operational performance while the second requirement reduces the risk for launching a live system while the learning process still is thrashing in the false predictor phase. [7] Although it appears that human monitoring may help us to establish at least some partial control of online stability practical applications often include other pressing stability issues. In this paper we have tacitly assumed that there is a stable underlying joint probability distribution that is explored by a growing data set. Many applications include probability distributions that change over multiple and initially unknown time scales. The very concept of stochastic modeling is also challenged in scenarios including intelligent adversarial opposition. It appears that there are good reasons for further investigating the stability of SRL online learning under conditions that are relevant for real-life applications. 1157 [8] [9] [10] [11] [12] [13] [14] [15] [16] “Secure Upgrade of Ports” project http://www.support-project.eu/, EU FP7 Grant Agreement no 242112 “Container Security Advanced Information Networking” project, http://www.containproject.com/, EU FP7 Grant Agreement no 261679 “Reuse and development of Security Knowledge assets for International Postal supply chains” project, EU FP7 Grant Agreement no 285104 A.R. Thomas, Supply Chain Security: International Practices and Innovations in Moving Goods Safely and Efficiently, Praeger, 2010. M. McNicholas, Maritime Security: An Introduction, ButterworthHeinemann, 2007. L. Getoor, and B. Taskar. (Eds), Introduction to Statistical Relational Learning, Boston: The MIT Press, 2007. T.N. Huynh, and R.J. Mooney, “Online Max-Margin Weight Learning for Markov Logic Networks,” in Proc. 11th Int. Conf. on Data Mining, 2011, pp. 642-651. L. Mihalkova, and R.J. Mooney, “Learning to disambiguate search queries from short sessions,” In: Proc. of the European Conf. on Machine Learning and Knowledge Discovery in Databases, 2009, pp.111-127. S. Riedel, and I. Meza-Ruiz, “Collective semantic role labelling with Markov logic,” In Proc. of the 12th Conf. on Computational Natural Language Learning, 2008, pp. 193-197. T.N. Huynh, and R.J. Mooney, “On online structure learning for Markov Logic Networks,” in Proc. European Conf. on Machine Learning an Principles and Practice of Knowledge Discovery in Databases, 2011, pp. 81-96. K. Kersting, and L. De Raedt, “Basic Principles of Learning Bayesian Logic Programs”, Univ. Freiburg Inst. of Computer Science Technical Report No. 174, 2002. N. Lavrac, and S. Dzeroski, Inductive Logic Programming Techniques and Applications, New York: Ellis Horwood, 1994. S. Ross, J.A. and Bagnell, “Stability conditions for online learnability”, arXiv:1108.3154v2 [cs.LG] 17 Aug. 2011. V. Vapnik, The Nature of Statistical Learning Theory. Springer, 1995. M. Ehrgott, “Multiobjective optimization”, AI Magazine, vol. 29, pp. 47-57, 2008 M. Finthammer, S. Loh, and M. Thimm, ”Towards a Toolbox for Relational Probabilistic Knowledge Representation”, In: Proc. Workshop on Relational Approaches to Knowledge Representation and Learning, 2009, pp. 34-48.
2cs.AI
Monolithic simulation of convection-coupled phase-change - verification and reproducibility arXiv:1801.03429v1 [physics.flu-dyn] 10 Jan 2018 Alexander G. Zimmerman and Julia Kowalski AICES Graduate School, RWTH Aachen University, Schinkelstr. 2, 52062 Aachen, Germany [email protected] Abstract. Phase interfaces in melting and solidification processes are strongly affected by the presence of convection in the liquid. One way of modeling their transient evolution is to couple an incompressible flow model to an energy balance in enthalpy formulation. Two strong nonlinearities arise, which account for the viscosity variation between phases and the latent heat of fusion at the phase interface. The resulting coupled system of PDE’s can be solved by a single-domain semi-phase-field, variable viscosity, finite element method with monolithic system coupling and global Newton linearization [9]. A robust computational model for realistic phase-change regimes furthermore requires a flexible implementation based on sophisticated mesh adaptivity. In this article, we present first steps towards implementing such a computational model into a simulation tool which we call Phaseflow [1]. Phaseflow utilizes the finite element software FEniCS [3], which includes a dual-weighted residual method for goal-oriented adaptive mesh refinement. Phaseflow is an open-source, dimension-independent implementation that, upon an appropriate parameter choice, reduces to classical benchmark situations including the lid-driven cavity and the Stefan problem. We present and discuss numerical results for these, an octadecane PCM convection-coupled melting benchmark, and a preliminary 3D convection-coupled melting example, demonstrating the flexible implementation. Though being preliminary, the latter is, to our knowledge, the first published 3D result for this method. In our work, we especially emphasize reproducibility and provide an easy-to-use portable software container using Docker [7]. Keywords: incompressible flow, finite element method, Newton method, phase-change, reproducibility 2 A. G. Zimmerman, J. Kowalski Nomenclature t tf x x∗ p u T fB r Tr Pr Ra Ste µ φ ∆t ()n Time Final time  Spatial coordinates x = x y z phase-change interface’s position Pressure field Velocity field Temperature field Buoyancy force Regularization smoothing factor Regularization central temperature Prandtl number Rayleigh number Stefan number Dynamic viscosity Semi-phase-field Time step size Values from discrete time n 1 Introduction γ Coefficient for penalty stabilization h Finite element cell diameter F The vector-valued strong form F Functional for the variational form Ω The spatial domain ∂Ω Boundary of the spatial domain ∂ΩT Dirichlet boundary for T V The scalar solution function space V The vector solution function space W Mixed finite element function  space w System solution w = p u T ψ Finite element basis functions M Adaptive goal functional M Adaptive solver tolerance δw Residual of linearized system ω Newton method relaxation factor ()k Values from Newton iteration k The melting and solidification of so-called phase-change materials (PCM’s) are relevant to many applications ranging from the design of latent heat based energy storage devices [12], to ice-ocean coupling and its effects on Earth’s climate [10], to the evolution of ocean worlds on the icy moons of our solar system [14] and the design of robotic melting probes for their exploration [16]. The predictive modeling of phase-change systems is, however, challenging due to 1. strong nonlinearities at the phase-change interface (PCI), 2. the coupling of several physical processes (i.e multi-physics), and 3. a large range of relevant scales both in space and time (i.e. multi-scale). Any mathematical model of a complex phase-change process hence manifests as a multi-scale and multiparameter, nonlinear PDE system. We aim to develop a robust and flexible model to simulate these systems. In this work, we will focus on phase-change in the presence of liquid convection. Convection can have a tremendous effect on the evolution of phaseinterfaces, as shown in Figure 1. A comprehensive introduction to melting and freezing without convection is given in [2]. A mathematical model that accounts for convection-coupled phase-change is presented in [5]. Therein mushy layer theory is introduced, which provides a model for understanding the PCI at macroscale. The physical system is mathematically modeled by considering balance laws for mass, momentum, and energy, with the momentum and energy balances coupled via buoyancy (which forces natural convection). This constitutes a difficult multi-physics problem. In a PCM domain, certain physics dominate in the solid and liquid subdomains. To solve the coupled 3 Hot plate Monolithic simulation of convection-coupled phase-change 𝒈 Ice Fig. 1. Sketch of the physical model superimposed on a melting experiment [17]. Heating from the left advances the phase-change interface (PCI) rightward. Buoyancy forces a convection cell, causing the upper PCI to advance more rapidly. In the absence of convection, the PCI would propagate as a planar front, indicating the significant impact of convection on the PCI’s evolution. problem, out of many approaches in the literature, we employ the single-domain semi-phase-field enthalpy method. With this approach, the convection itself is handled by an incompressible Navier-Stokes flow model, while the energy balance is modeled as the convection and diffusion of an enthalpy field. The key idea is to solve the same equations on the entire domain. This is commonly referred to as the fixed grid approach, for which various techniques are reviewed in [19]. Using an enthalpy method, phase-change latent heat effects are isolated to a source term in the energy balance [18]. Single-domain methods require a velocity correction scheme, for which we select the variable viscosity method, thereby treating the solid as a highly viscous fluid. Efficiently applying this approach requires local mesh refinement to resolve the moving PCI. This can be accomplished either with mesh refinement or fronttracking [12]. Adopting the former approach, we leverage the existing work on goal-oriented adaptive mesh refinement (AMR) methods, particularly the dualweighted residual method [4]. These days, goal-oriented AMR is widely practiced, and multiple open-source software libraries provide this capability. In the following, Section 2 describes the mathematical model. Section 3 presents numerical methods for the model’s discretization and linearization. Section 4 presents our implementation, Phaseflow [1], based on the finite element software FEniCS [3]. Finally, Section 5 presents verification via comparison to benchmark problems, a convergence study for the 1D Stefan problem, and a preliminary result for convection-coupled melting in a 3D domain. We close with conclusions and an outlook. 4 2 2.1 A. G. Zimmerman, J. Kowalski Mathematical Phase-Change Model The Governing Equations For the coupled phase-change system, out of many approaches from the literature, we adopt an enthalpy formulated [18], single-domain semi-phase-field [6], variable viscosity model. The mass and momentum balances (1) and (2) are given by the incompressible Navier-Stokes equations, with velocity field u = u(x, t) and pressure field p = p(x, T ). Invoking the Boussinesq approximation, we extend the momentum equation with a temperature-dependent buoyancy forcing term fB (T ) which couples it to the energy equation. This approach is well established in the context of natural convection [20]. Furthermore, we consider the phase-dependent viscosity µ(φ) in the momentum equation, where φ = φ(T ) is the temperature-dependent phase. For constant heat capacity, the energy balance in enthalpy form reduces to (3), which is an extended form of the convectiondiffusion equation for the temperature field T = T (x, t). The diffusion term in (3) 1 ∂ is scaled by the Prandtl number Pr. The nonlinear source term Ste ∂t φ accounts for the phase-change latent heat, where Ste is the Stefan number. Altogether the system of governing equations is ∇·u=0 (1) ∂ u + (u · ∇) u + ∇p − ∇ · (2µ(φ)D(u)) + fB (T ) = 0 (2) ∂t ∂ 1 ∂ 1 T− φ + ∇ · (T u) − ∆T = 0 (3) ∂t Ste ∂t Pr   T where the symmetric part of the rate-of-strain tensor is D(u) = 21 ∇u + (∇u) . The equations are unitless per the normalization in [9], shifted always such that T = 0 corresponds to the temperature of fusion. We refer to φ as a semi -phase-field [6], because we do not treat φ as an additional unknown. Rather, φ maps the temperature field to values between zero and one. For this we use   1 Tr − T φ(T ) = 1 + tanh , (4) 2 r where Tr is the central temperature and r is a smoothing parameter. The Newton method requires the differentiability of (4). Subject to this requirement, there are many other useful regularizations from which to choose. Figure 2 plots φ and φ0 . For the Stefan problem in Section 5.3, we simply set Tr equal to the physical temperature of fusion (i.e. Tr = 0 for the normalized system). Following [9], for the coupled convection problem, we set Tr such that the strongest variation is localized near the newly appearing phase, e.g. liquid for melting. Physically, the region where 0 < φ < 1 is analogous to a mushy layer. In this sense, we can view φ as the solid volume-fraction. Given φ, we define the phase-dependent viscosity as functions of the liquid and solid values (assuming constants µL and Monolithic simulation of convection-coupled phase-change 5 1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 -0.1 -0.05 0 0.05 0.1 Fig. 2. The regularized semi-phase-field (4) with Tr = 0 and two values of r. For smaller r, φ(T ) steepens and its derivative φ0 (T ) approaches the Dirac delta function. Here the derivative is scaled by 2r for convenience. µS respectively) with µ(φ) = µL + (µS − µL )φ, (5) With the variable viscosity method, the solid is treated as a highly viscous fluid, forcing its flow velocity to zero. In general, not only viscosity varies in the PCM domain. For example, the thermal material properties of water-ice vary significantly between the solid and liquid phases. In the case of octadecane PCM’s, such as the benchmark in Section 5.4, we may treat these material properties as constants [20]. 2.2 The Initial Boundary Value Problem To simulate the time evolution of the PCM system, we will solve (8) as an initial boundary value problem subject to the initial values T T p(x, t), u(x, t), T (x, t) = 0, u0 (x), T0 (x) ∀(x, t) ∈ Ω × 0 (6) and boundary conditions u(x, t) = uD (x, t) ∀(x, t) ∈ ∂Ω × (0, tf ], T (x, t) = TD (x, t) ∀x ∈ ∂ΩT × (0, tf ], (7) (n̂ · ∇)T (x, t) = 0 ∀x ∈ (∂Ω \ ∂ΩT ) × (0, tf ] For the applications in Section 5, we apply for all boundaries the velocity no-slip condition. This implies homogeneous Dirichlet boundary conditions, uD = 0, for 6 A. G. Zimmerman, J. Kowalski all cases except the lid-driven cavity in Section 5.1, where the ”moving” lid makes uD non-homogeneous. The notation ∂ΩT refers to the boundary subdomain with non-homogeneous Dirichlet boundary conditions on the temperature. Physically this models a thermal reservoir beyond the wall, keeping the wall at a constant temperature. All other walls are prescribed homogeneous Neumann boundary conditions on the temperature, which physically models thermally insulated (i.e. adiabatic) walls. The generality of this initial boundary value problem allows us to solve a variety of interesting benchmarks by setting the appropriate parameters. In Section 5 we will demonstrate the 2D lid-driven cavity, 2D heat-driven cavity, 1D Stefan problem, 2D convection-coupled melting of an octadecane PCM, and 3D convection-coupled melting. Our implementation which we present in Section 4 is as versatile as the model. Testing these benchmarks is accomplished with short Pythons scripts specifying the parameters. 3 Numerical Methods We base our work on the numerical approach from [9]. Therefore, we discretize in time via finite differences, discretize in space via the finite element method (FEM) with a penalty formulation for stabilization, couple the system monolithically, and solve the nonlinear system globally via Newton’s method. 3.1 The Nonlinear Variational Problem  We denote the system’s solution as w = p u T and write the strong form (1), (2), and (3) as the vector-valued functional F(w) = 0 ∀(x, t) ∈ Ω × (0, tf ], (8) with the three components respectively being the mass, momentum, and energy equations. To stabilize the finite element method, following the penalty formulation in [9], we add a pressure stabilizing term γp to the mass component. The coefficient γ is generally small, and in this case is γ = 10−7 . We consider 1D, 2D, and 3D spatial domains, respectively allowing for Ω ⊂ R1 , Ω ⊂ R2 , or Ω ⊂ R3 . Per the standard Galerkin finite element method, we write the variational (i.e. weak) problem whose solution approximates (8). To obtain the monolithic system, we use mixed finite elements  [8]. Therefore, we multiply (8) from the left by test functions ψ = ψp ψ u ψT and integrate by parts over the domain. Finally, we employ the fully implicit Euler method by substituting the time discretizations ∂t u = ∆t−1 (un+1 − un ), ∂t T = ∆t−1 (Tn+1 − Tn ), ∂t φ = ∆t−1 (φn+1 − φn ). Monolithic simulation of convection-coupled phase-change 7 Altogether, for homogeneous Neumann boundary conditions, this yields the time-discrete nonlinear variational form Z  ψp ψ u ψT F(p, u, T )dx = F (ψ, w) = Ω b(u, ψp ) − (ψp , γp)  1 + ψu , (un+1 − un ) + fB (T ) + c(u; u, ψ u ) + b(ψ u , p) + a(µ(φ); u, ψ u ) ∆t     1 1 1 + ψT , Tn+1 − Tn − (φ(Tn+1 ) − φ(Tn )) + ∇ψT , ∇T − T u ∆t Ste Pr (9) R R where we use the short-hand (u, v) = Ω uvdx, (u, v) = Ω u · vdx for integrating inner products. Additionally, we use a common notation [11] for the linear, bilinear, and trilinear forms of the variational Navier-Stokes equations Z a : V × V → R, a(µ; u, v) = 2 µD(u) : D(v)dx Z Ω b : V × V → R, b(u, p) = − p∇ · udx (10) Ω Z c : V × V × V → R, c(u; z, v) = vT (∇z) udx  Ω Given (9), we write the variational problem as Find w ∈ W such that F (ψ, w) = 0 ∀ψ ∈ Ŵ (11) where W = V ×V×V and Ŵ = V̂ ×V̂×V̂ . This distinction comes from how nonhomogeneous boundary conditions are often handled in finite element method implementations. The solution is split into homogeneous and non-homogeneous parts, the former is found, and then the  latter is reconstructed. Therefore, for example, T belongs to the space V = v ∈ H01 (Ω) : v = TD on ∂Ω , while ψT  belongs to the space V̂ = v ∈ H01 (Ω) : v = 0 on ∂Ω , where H01 (Ω) is the classical Hilbert space fulfilling requirements for continuity and compact support. The vector-valued function space V depends on the spatial domain’s dimensionality, e.g. V = V × V in 2D. For the incompressible Navier-Stokes solution, we use the Taylor-Hood element [11], i.e. we restrict the pressure solution to piece-wise continuous linear Lagrange polynomials, and we restrict the velocity solution to piece-wise continuous quadratic Lagrange polynomials. We use the same polynomial space for temperature and pressure. 3.2 Linearization via Newton’s Method We apply Newton’s method by solving a sequence (indexed by superscript k) of linear problems 8 A. G. Zimmerman, J. Kowalski Find δw ∈ W such that Dw F (ψ, wk ; δw) = F (ψ, wk ) ∀ψ Ŵ (12) for the residual δw which updates the solution wk , converging wk to the approximate solution of the nonlinear problem. The Gâteaux derivative of F (ψ, wk ), d F (ψ, wk + δw)|=0 , is given by defined as Dw F (ψ, wk ; δw) ≡ d Dw F (ψ, wk ; δw) = b(δu, ψp ) − (ψp , γδp)  1 0 k + ψu , δu + δT fB (T ) + c(uk ; δu, ψ u ) + c(δu; uk , ψ u ) ∆t   +b(ψ u , δp) + a δT µ0 (T k ); uk , ψ u + a µ(T k ); δu, ψ u     1 1 0 k 1 k k + ψT δT, 1 − φ (T ) + ∇ψT , ∇δT − T δu − δT u , ∆t Ste Pr  (13) For a given wk (13) and hence (12) are linear with respect to the unknown Newton residual δw. In typical fashion, we define the test functions, solution and Newton residual as linear combinations of the same basis. Upon selecting a mesh and concrete basis, this allows (12) to be re-written as a discrete linear system of the form Ax = b, which can be efficiently solved on a computer by standard methods and software. In this work, we directly solve each linear system with LU decomposition (using the interface of FEniCS, discussed in Section 4). We use the latest discrete time solution as the initial guess for the Newton solver, i.e. we initialize w0 = wn . After each iteration solving for δw, we update the solution with the relaxed residual wk+1 := wk + ωδw, where 0 < ω ≤ 1 is a relaxation factor. Relaxing Newton’s method is often useful for highly nonlinear problems. The 1D and 2D results in Section 5 use the full Newton (ω = 1) method. So far only the preliminary 3D result requires relaxation, where we will set ω = 0.8. 3.3 Adaptive Mesh Refinement The single-domain approach requires local mesh refinement, or else the computational cost would quickly become impractical, especially in 3D. Furthermore, we cannot a priori prescribe where to locally refine the mesh, since our goal is to predict the position of the PCI. This means we must employ adaptive mesh refinement, now commonly referred to as AMR. The theory of AMR fundamentally requires an error estimator, which exist for many interesting problems [15]. In the context of phase-change simulations, hierarchical error estimators have been derived for the Stefan problem [6]; but no such rigorous work has been completed for the problem with coupled convection. Promising results are reported in [9], wherein they used a mesh adaptivity procedure by metric control which was particular to the software library and its Delaunay mesh generation procedure. Unfortunately, it is unclear exactly which metrics were used for adaptivity in those results. Monolithic simulation of convection-coupled phase-change 9 We instead employ goal-oriented AMR [4]. To discuss AMR, let us briefly write the spatially discrete problem which is dependent on a mesh Wh ⊂ W(Ω): Find wh ∈ Wh ⊂ W(Ω) such that F (ψ h , wh ) = 0 ∀ψ h ∈ Ŵh ⊂ Ŵ (14) Goal-oriented AMR requires some goal functional M (w) to be integrated over the domain Ω. The goal-oriented adaptive solution of (14) can then be written: Find Wh ⊂ W(Ω) and wh ∈ Wh such that |M (w) − M (wh )| < M (15) where M is some prescribed tolerance. Since M (w) is unknown, we still require an error estimator. To this end, we use the dual-weighted residual method, as implemented in FEniCS [3]. Computing cell-wise error estimates requires solving a linearized adjoint (with respect to the goal) problem. The primal and adjoint problems are solved on a hierarchy of meshes. Computing the linearized adjoint solution on each mesh is relatively cheap compared to solving the nonlinear primal problem. See [4] for a full explanation of the method. For the adaptive solutions in this work, we set the goal functional Z M (w) = φ(T )dx (16) Ω which represents the volume of solid material remaining in the domain. 4 Implementation The method presented in this paper, along with its application to a series of test problems, were implemented by the author into an open-source Python module named Phaseflow [1], using the finite element library FEniCS [3]. 4.1 FEniCS The abstract interface of FEniCS [3] makes it an ideal library to quickly prototype models and methods which use FEM. Additionally, standard algorithms for Newton linearization and goal-oriented AMR are already implemented. FEniCS is an umbrella project. A major back-end component is the C++ library DOLFIN, which implements most of the classes and methods seen by the user. Python interfaces are generated mostly automatically, and indeed FEniCS is primarily used as a Python module. From a conceptual perspective, we are most interested in two components of FEniCS: the Unified Form Lanugage (UFL) and the FEniCS Form Compiler (FFC). UFL allows us to write down the abstract discrete variational (i.e. weak) form in a way that is understood by FEniCS. This means that to implement variational forms, we write them as source code almost character-for-character. The FEniCS Form Compiler (FFC) then automatically implements the FEM matrix assembly routine in optimized C++ code using just-in-time (JIT) compiling. 10 4.2 A. G. Zimmerman, J. Kowalski Phaseflow Phaseflow [1], created by the author, is a Python module maintained openly on GitHub. Phaseflow implements the methods of this paper using the open-source finite element library FEniCS [3]. The interface allows users to run these methods for the variational form (9) with any set of similarity parameters Ste, Pr, buoyancy model fB (T ), regularization φ(T ), liquid and solid viscosities µL and µS , time step size ∆t, and stabilization coefficient γ. The initial values (6) are set with a function in the solution space. Combined with the built-in FEniCS method, this allows the user to interpolate general mathematical expressions written in the C syntax, or to use an existing solution. Dirichlet boundary conditions (7) are also set with general C mathematical expressions. Additionally the interface allows users to control algorithm parameters, such as tolerances for the solvers, the Newton relaxation parameter ω, and other options. The interface accepts any FEniCS mesh object, which can either be generated from scratch or can be converted from a standard format via the FEniCS library. Leveraging the abstract interfaces of FEniCS, Phaseflow is dimensionindependent, and has been applied to 1D, 2D, and 3D spatial domains. In this work, unit square and rectangular prism domains have been used. Phaseflow writes all solutions to disk in the efficient XDMF+HDF5 format using the builtin FEniCS interface. Furthermore, this is combined with the H5Py library to write checkpoint files for easily restarting simulations at later times. Phaseflow has other ease-of-use features, such as a stopping criterion for steady state solutions. This was useful for the heat-driven cavity test in Section 5.2, where the Newton method required a small time step size. 4.3 Reproducibility with Docker Phaseflow leverages open technology, and it is meant to contribute further to open scientific and engineering research. Open research in the computational sciences is often hindered by the difficulty of reproducing results, primarily because of the ever-increasing complexity of computing environments [7]. Recently, the software Docker has emerged as a technology which greatly facilitates reproducibility in the computational sciences. Built upon an official FEniCS Docker image, we provide a Phaseflow Docker with all dependencies. Enabled by the Docker image and Phaseflow’s test suite, changes to the master branch on GitHub are continuously tested using the Travis-CI continuous integration service [7]. Docker has been central to the development of Phaseflow. Initially, a primary motivation for its use was the existence of Docker images where most of the dependencies of Phaseflow are already installed and maintained by the FEniCS developers. This allows us to spend more time implementing models, and less time building tool chains. From there, maintaining a Docker image which runs Phaseflow takes little effort. With the Phaseflow Docker image in hand, we were then able to set up a continuous integration process with remarkably little effort. Monolithic simulation of convection-coupled phase-change 11 The vast majority of Phaseflow’s source code is covered with a suite of unit tests and integration tests written with PyTest. Most results in this paper are included in Phaseflow’s test suite, while the others are covered by the example scripts and notebooks in its repository [1]. Such results are easy to reproduce, and we highly encourage the reader to do this. To further aid reproducibility, a specific version of Phaseflow has been archived (see the release versions at [1]) to coincide with this publication. 5 Verification To verify our implementation, we consider a series of examples, including the liddriven cavity benchmark from [13] extended with a solid subdomain, the heatdriven cavity benchmark from [20], an approximation of the analytical 1D Stefan problem from [2] with and without AMR, a preliminary 2D convection-coupled melting demonstration with AMR, and a preliminary 3D convection-coupled melting demonstration without AMR. Furthermore, for the Stefan problem, we verify the convergence orders of the temporal and spatial discretizations, and of the Newton method. Phaseflow’s flexible interface described in Section 4 allows us to implement each of these applications with minimal effort, and to test the same lines of source code, increasing our confidence with the implementation. 5.1 Lid-Driven Cavity with Solid Subdomain To test phase-dependent viscosity, we consider an extension of the lid-driven cavity benchmark based on data published in [13]. The standard benchmark uses a unit square geometry. We extend the geometry below the bottom wall, and set the temperature in the new region below the freezing temperature, such that the liquid subdomain per (4) covers the original unit square geometry. Therefore the initial temperature values are ( Tc for y ≤ 0, (17) T0 (x) = Th otherwise where we chose Th = 1 and Tc = −1. For the semi-phase-field (4) we set central temperature Tr = −0.01 and smoothing parameter r = 0.01. We set variable viscosity (5) with µL = 1 and µS = 108 . To capture this strong variation, we refined all cells of the initial mesh which touched coordinate y = 0 for four refinement cycles. Note that in this case, the local refinement is not adaptive. This problem at steady state fits our general model (9) with null buoyancy fB = 0 and with an arbitrarily large Prandtl number to nullify thermal conduction. As boundary conditions uD , we set positive horizontal velocity on the lid, zero horizontal velocity away from the lid, and zero vertical velocity everywhere. We set initial velocity u0 similarly, which serves as a suitable initialization for solving until steady state with Newton’s method. Because Phaseflow implements only the unsteady problem, we solve a single large time step to obtain the approximate 12 A. G. Zimmerman, J. Kowalski 1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 Fig. 3. Left: Result from lid-driven cavity test (Re = 100). The mesh (which is locally refined near the PCI) is shown in translucent gray. Velocity streamlines are shown in black. As expected, the moving lid causes circulation in the cavity, and we see the expected recirculation zone in the bottom-right corner of the liquid subdomain. Right: Horizontal velocity sampled from the vertical centerline, compared to benchmark data published in [13]. steady state solution. It is advantageous to model the steady problem this way, when the goal is to verify components of our unsteady implementation. Figure 3 shows and discusses the result. The solution agrees very well with the published benchmark data. 5.2 Heat-Driven Cavity To verify the coupled energy equation without phase-change, we compare to the heat-driven cavity benchmark data published in [20]. This problem fits our general model with constant viscosity µS = µL = 1 and arbitrarily large Stefan number. We handle buoyancy with an idealized linear 1 Boussinesq model   Ra 0 T , (18) fB (T ) = −1 PrRe2 The momentum equation (2) is scaled such that the Reynolds number Re is always unity. For this benchmark, the Rayleigh number is Ra = 106 and the Prandtl number is Pr = 0.71. This Rayleigh number is considered to be high, and demonstrates substantial natural convection. We set homogeneous Dirichlet (i.e. no slip) boundary conditions on the velocity uD = 0, and non-homogeneous Dirichlet boundary conditions for the temperature with hot and cool temperatures, Th = 0.5 and Tc = −0.5, respectively on the left and right vertical walls, 1 The method handles nonlinear bouyancy with (13). Phaseflow’s test suite includes a benchmark with the nonlinear density anomaly of water. Monolithic simulation of convection-coupled phase-change 13 1 Wang 2010 Phaseflow 0.8 0.6 0.4 0.2 0 -0.1 -0.05 0 0.05 0.1 Fig. 4. Left: Result from heat-driven cavity test (Ra = 106 ). The mesh is shown in translucent gray. Velocity streamlines are shown in black. As expected, we see the flow circulating, being forced upward near the hot wall and downward near the cold wall. Right: The horizontal component of velocity is plotted along the vertical centerline. The solution agrees very well with the data published in [20]. i.e. ( Th TD (x) = Tc for x = 0, for x = 1 (19) Again we solve the unsteady problem until it reaches steady state. For this we initialize the velocity field to zero (i.e. u0 = 0) and the temperature field to vary linearly between the hot and cold walls (i.e. T0 (x) = Th +x(Tc −Th )). Unlike with the lid-driven cavity, here we cannot obtain the steady state solution in a single time step, because the initial guess is not sufficient for the Newton method to converge. Therefore we solve a sequence of time steps using ∆t = 0.001. Figure 4 shows the successful result with further discussion. 5.3 The Stefan Problem To verify the energy equation with phase-change, we compare to the analytical 1D Stefan problem as written in [2], with parameters comparable to the octadecane PCM melting benchmark in Section 5.4, including the Stefan number Ste = 0.045. This problem fits our general model with nullified buoyancy fB = 0, unity Prandtl number, and zero initial velocity u0 = 0 (which remains zero). Solving the Stefan problem as a special case of the coupled problem results in an unnecessarily large system. In this case, the vast majority of degrees of freedom are trivial. From an implementation perspective, this exercise is quite valuable, because we test the exact lines of code which are used for the coupled problem. 14 A. G. Zimmerman, J. Kowalski We again set temperature boundary conditions (19), this time with Th = 1 and Tc = −0.01. The initial temperature field is set such that a thin layer of melt exists near the hot wall, with the rest of the domain at the cold wall temperature, i.e. ( Th for x ≤ x∗0 , T0 (x) = (20) Tc otherwise We parameterize this initial PCI position as x∗0 = L 1−q 2 N0 (21) where q is the number of initial hot wall refinement cycles. This ensures that for the given initial mesh, the thinnest possible layer of melt exists. For the semi-phase-field regularization (4) we set parameters Tr = 0, r = 0.01. We simulate until time tf = 0.1, using the time step size ∆t = 0.001. In testing, this size was needed to bound the point-wise error between the numerical and analytical solutions from above by T (x) − Texact (x) < 0.01 ∀x ∈ {0, 0.025, 0.05, 0.075, 0.1, 0.5, 1} at time t = tf (shown in Figure 5). This discrete set of verification points is a good proxy for the global solution. Figure 5 shows the successful result and further discussion. We also take this opportunity to demonstrate AMR. We set the adaptive goal (16) with tolerance M = 10−6 . During the time-dependent simulation, new cells are only added, and never removed, i.e. the mesh is never coarsened. This is an unfortunate limitation of the current AMR algorithm. We see that cells have been clustered near the PCI, and these clusters remain everywhere the PCI has been during the time-dependent simulation. Rightward of the PCI, the cells grow much larger, e.g. the right half of the domain is covered by only two cells. Even without coarsening, already in 1D this is a large improvement. The gains in 2D and 3D will be even greater. 5.4 2D Convection-Coupled Melting of Octadecane PCM To demonstrate the entire coupled system, we present a preliminary result for the convection-coupled melting of the octadecane PCM benchmark from [20] and [9]. This problem uses all aspects of our general model (9). We set variable viscosity (5) (with µL = 1 and µS = 108 ), the buoyancy model (18) (with Ra = 3.27×105 , Pr = 56.2, and Re = 1), no-slip velocity boundary conditions uD = 0, and again the temperature boundary conditions (19) with Th = 1 and Tc = −0.01. Again we initialize the temperature field with (20) such that the simulation begins with a thin layer of melt, and initialize a stationary velocity field u0 = 0. For the semiphase-field (4), we set regularization parameters Tr = 0.01 and r = 0.025. For the time-dependent simulation, we set ∆t = 1. For AMR, we again set the adaptive 3 For the uniform case, this is the minimum number of cells for which the Newton method would converge. Similarly for the AMR case, starting with a coarser initial mesh disrupts the Newton method. Monolithic simulation of convection-coupled phase-change 15 0.1 1 Analytical Phaseflow, Uniform Phaseflow, Adaptive 0.8 0.08 0.6 0.06 0.4 0.04 0.2 Analytical Phaseflow, Uniform Phaseflow, Adaptive 0.02 0 0 -0.2 0 0.5 1 0 0.05 0.1 Fig. 5. Left: Comparison at t = 0.1 between the analytical Stefan problem solution, Phaseflow using a uniform mesh of 311 cells3 , and Phaseflow using AMR (with 133 cells in the final adapted mesh). The AMR solution is shown with a marker at every mesh vertex. This emphasizes the much smaller cell sizes in the wake of the PCI. Right: Comparison of the PCI positions at three times between the analytical, uniform, and AMR solutions. We observe a bias in the uniform mesh solution which shrinks over time. This is primarily because of the mesh-dependent initial PCI position (21). goal (16) from our Stefan problem example, but with tolerance M = 10−5 until t = 36 and M = 0.5 × 10−5 after that time. The result in Figure 6 is promising; but compared to the benchmark, we see that the PCI has not advanced far enough by simulated time t = 80. We still need to investigate the effects of our choices for ∆t, r, x∗0 , M (w), M , and the initial mesh refinement. Of these, we know that a much smaller smoothing parameter r was used in [9]. 5.5 3D Convection-Coupled Melting The dimension-independent implementation, facilitated by the abstractions from FEniCS, allow us to quickly demonstrate a 3D example. We consider the a problem similar to the 2D convection-coupled melting in Section 5.4, but with the domain extruded in the z direction, adiabatic no-slip boundary conditions on the walls parallel to the z plane, similarity parameters Ste = 1, Ra = 106 , Pr = 0.71, and numerical parameters µS = 104 , r = 0.05. For the previous examples, we employed the ”full” Newton method with ω = 1. For this problem, we relaxed the Newton method with a factor of ω = 0.8. Figure 7 shows the preliminary result. 16 A. G. Zimmerman, J. Kowalski Fig. 6. Preliminary result for the 2D convection-coupled octadecane PCM melting benchmark. Solutions are shown at simulated times t = 36 and t = 80 (on the left and right). The mesh is shown in translucent gray to highlight the adaptive mesh refinement (AMR). The PCI is shown as a thick black temperature isoline. Velocity streamlines are shown as thin black lines. This result looks promising when compared to octadecane PCM melting results from [20] and [9]. We see the flow circulating, and the PCI advanced more quickly at the top than at the bottom. We also see that the lack of grid coarsening is becoming expensive. By the final time, many refined cells from earlier times are likely not needed. Monolithic simulation of convection-coupled phase-change 17 Fig. 7. Preliminary result for 3D convection-coupled melting, with front and isometric views (on the left and right). The mesh is shown in translucent gray to highlight the local mesh refinement. The PCI is a darker translucent gray iso-surface. The streamlines are colored by a grayscale from white to black as velocity magnitude increases. We use a coarse mesh and omit AMR, because the refined 3D problem quickly exceeded the capabilities of the desktop-scale computers used for this work. Despite the coarse mesh, we still observe the primary features we expect in the solution. The flow is circulating, and we see that the top of the PCI advancing more quickly than the bottom. 5.6 Convergence To verify the accuracies of the finite difference time discretization, finite element space discretization, and Newton linearization methods, we consider the 1D Stefan problem from Section 5.3. The mixed finite element formulation was shown to be second order accurate for the incompressible Navier-Stokes equations in [8]. We are not aware of such a result for the energy-coupled problem. Here, as a first step, we focus only on the energy equation with phase-change. Based on the choices of discretizations, with fully implicit Euler for time and finite elements for space, we expect first order convergence in time and second order in space. Figure 8 compares this with the actual convergence orders of Phaseflow’s solution from Section 5.3. With respect to the time step sizes δt, the observed convergence order is only slightly higher than expected. With respect to the grid spacing h, for a sufficiently smooth solution with well-behaved data, we should expect second order accuracy. Though there is some deviation from second order near h = 0.001, the results in Figure 8 show good agreement. Next, we verify the accuracy of the Newton method. Figure 8 shows its convergence behavior for multiple time steps, highlighting the difference in performance between the first and last time steps. Ideally, Newton’s method will converge quadratically for well-posed problems with a suitable initial guess. At earlier times, a few Newton iterations pass before reaching approximately quadratic convergence. Later times converge slightly better than quadratically. 18 A. G. Zimmerman, J. Kowalski 100 100 10-1 10-1 10-2 10-3 Actual First Order 10-2 10-3 10-2 Actual Second Order 10-1 10 0 10-4 10-4 10-2 100 10 -5 10-5 10 -10 10-10 10 -15 Actual Quadratic 10 -20 0 2 4 Actual Quadratic 6 8 10 10-15 1 2 3 4 5 Fig. 8. Convergence study results for Phaseflow applied to the Stefan problem from Section 5.3. Top: Temporal (on the left) and spatial (on the right) convergence results based on the solution at the final time tf . Bottom: Newton method convergence shown for multiple time steps. Each sub-plot has a series of lines which represent ideal (quadratic) iterations. Monolithic simulation of convection-coupled phase-change 6 19 Conclusions and Outlook In this work, we presented the computational tool Phaseflow for simulating the convection-coupled melting and solidification of PCM’s. Our work is based on the numerical method proposed in [9], and so we used an enthalpy formulated, singledomain semi-phase-field, variable viscosity, finite element method, with monolithic system coupling and global Newton linearization. The primary difference between our numerical approach and that in [9] is the mesh adaptivity algorithm, where we employ the dual-weighted residual method for goal-oriented AMR. We implemented the method into our open source Python module Phaseflow, based on the finite element software library FEniCS, and verified this against a series of classical benchmarks. We obtained a promising result for the octadecane PCM convection-coupled melting benchmark. Leveraging our dimension-independent implementation, we contributed detailed convergence results for the 1D phasechange problem, and have applied the method to a preliminary 3D convectioncoupled melting example. Furthermore, we openly shared a Docker container which allows anyone to reproduce our results in the same software environment. It is our hope that this facilitates the further development of this and related methods. The method appears promising, and Phaseflow is ready for application to interesting problems. FEniCS was a good choice as the base of our implementation, allowing us to focus on the models and numerical methods rather than on implementing standard algorithms. Furthermore, the existing Docker software container allowed us to quickly leverage the library. Some difficulties do remain, primarily: 1. FEniCS lacks mesh coarsening capability, prohibiting the efficient application of AMR to the moving PCI problem. 2. The adaptive solver (using dual-weighted residual goal-oriented AMR) has not been implemented for distributed memory systems, therefore currently limiting the implementation’s applicability to a single compute node, prohibiting realistic 3D applications. On the other hand, there is yet potential to simplify Phaseflow with existing FEniCS features. Most interestingly, there is an automatic differentiation capability which could serve as an alternative to computing the Gâteaux derivative (13) directly. During the development of Phaseflow, we have successfully applied this feature to some simplified cases; and we would like to further explore this capability. Next steps of our work include the development of robustness features, including... 1. a priori bounds on the time step size ∆t, based on the properties of the linear system (12) and initial values (6), 2. an algorithm for obtaining the initial mesh on which to obtain the first solution to begin AMR. The latter is important, because we discovered a limitation with relying on AMR to resolve strong nonlinearities. Typically, goal-oriented AMR is used to provide either the most accurate solution (with respect to the goal) for a given cost, or the minimum-cost discretization for a given accuracy (with respect to the goal). In the case of the presented method, AMR is required for the Newton method to converge and to hence obtain the first solution in the hierarchy. Handling these issues is necessary before practically applying this method to a wider range of realistic problems. 20 A. G. Zimmerman, J. Kowalski References 1. Phaseflow. https://github.com/geo-fluid-dynamics/phaseflow-fenics. 2. V Alexiades. Mathematical modeling of melting and freezing processes. CRC Press, 1992. 3. M Alnæs, J Blechta, J Hake, et al. The fenics project version 1.5. Archive of Numerical Software, 3(100):9–23, 2015. 4. W Bangerth and R Rannacher. Adaptive Finite Element Methods for Differential Equations. Springer Science & Business Media, 2003. 5. GK Batchelor, HK Moffatt, MG Worster, and TR Osborn. Perspectives in Fluid Dynamics. Cambridge University Press Cambridge, 2000. 6. Y Belhamadia, A Fortin, and É Chamberland. Anisotropic mesh adaptation for the solution of the stefan problem. J. Comput. Phys., 194(1):233–255, 2004. 7. C Boettiger. An introduction to docker for reproducible research. ACM SIGOPS Operating Systems Review, 49(1):71–79, 2015. 8. F Brezzi and M Fortin. Mixed and hybrid finite element methods, volume 15. Springer Science & Business Media, 1991. 9. I Danaila, R Moglan, F Hecht, and S Le Masson. A newton method with adaptive finite elements for solving phase-change problems with natural convection. J. Comput. Phys., 274:826–840, 2014. 10. MS Dinniman, XS Asay-Davis, BK Galton-Fenzi, et al. Modeling ice shelf/ocean interaction in antarctica: A review. Oceanography, 29(4):144–153, 2016. 11. J Donea and A Huerta. Finite element methods for flow problems. John Wiley & Sons, 2003. 12. Y Dutil, DR Rousse, NB Salah, et al. A review on phase-change materials: mathematical modeling and simulations. Renewable and Sustainable Energy Reviews, 15(1):112–130, 2011. 13. UKNG Ghia, KN Ghia, and CT Shin. High-re solutions for incompressible flow using the navier-stokes equations and a multigrid method. J. Comput. Phys., 48(3):387–411, 1982. 14. HW Hsu, F Postberg, Y Sekine, et al. Ongoing hydrothermal activities within enceladus. Nature, 519(7542):207–210, 2015. 15. DW Kelly, De SR Gago, OC Zienkiewicz, et al. A posteriori error analysis and adaptive processes in the finite element method: Part i—error analysis. Int. J. Numer. Methods Eng., 19(11):1593–1619, 1983. 16. J Kowalski, P Linder, S Zierke, et al. Navigation technology for exploration of glacier ice with maneuverable melting probes. Cold Reg. Sci. Technol., 123:53–70, 2016. 17. K Schüller and J Kowalski. Integrated modeling and validation for phase change with natural convection. Pending, 2018. 18. VR Voller, M Cross, and NC Markatos. An enthalpy method for convection/diffusion phase change. Int. J. Numer. Methods Eng., 24(1):271–284, 1987. 19. VR Voller, CR Swaminathan, and BG Thomas. Fixed grid techniques for phase change problems: a review. Int. J. Numer. Methods Eng., 30(4):875–898, 1990. 20. S Wang, A Faghri, and TL Bergman. A comprehensive numerical model for melting with natural convection. International Journal of Heat and Mass Transfer, 53(9):1986–2000, 2010.
5cs.CE
Confidence sets with expected sizes for Multiclass Classification Christophe Denis∗ and Mohamed Hebiri† arXiv:1608.08783v2 [math.ST] 18 Dec 2017 LAMA, UMR-CNRS 8050, Université Paris Est – Marne-la-Vallée Abstract Multiclass classification problems such as image annotation can involve a large number of classes. In this context, confusion between classes can occur, and single label classification may be misleading. We provide in the present paper a general device that, given an unlabeled dataset and a score function defined as the minimizer of some empirical and convex risk, outputs a set of class labels, instead of a single one. Interestingly, this procedure does not require that the unlabeled dataset explores the whole classes. Even more, the method is calibrated to control the expected size of the output set while minimizing the classification risk. We show the statistical optimality of the procedure and establish rates of convergence under the Tsybakov margin condition. It turns out that these rates are linear on the number of labels. We apply our methodology to convex aggregation of confidence sets based on the V -fold cross validation principle also known as the superlearning principle [vdLPH07]. We illustrate the numerical performance of the procedure on real data and demonstrate in particular that with moderate expected size, w.r.t. the number of labels, the procedure provides significant improvement of the classification risk. Keywords : Multiclass classification, confidence sets, empirical risk minimization, cumulative distribution functions, convex loss, superlearning. 1 Introduction The advent of high-throughput technology has generated tremendous amounts of large and highdimensional classification data. This allows classification at unprecedented scales with hundreds or even more classes. The standard approach to classification in the multiclass setting is to use a classification rule for assigning a single label. More specifically, it consists in assigning a single label Y ∈ Y, with Y = {1, . . . , K}, to a given input example X ∈ X among a collection of labels. However, while a large number of classes can lead to precise characterizations of data points, similarities among classes also bear the risk of confusion and misclassification. Hence, assigning a single label can lead to wrong or ambiguous results. In this paper, we address this problem by introducing an approach that yields sets of labels as outputs, namely, confidence sets. A confidence set Γ is a function that maps X onto 2Y . A natural way to obtain a set of labels is to use ranked outputs of the classification rule. For example, one could take the classes that correspond to the top-level conditional probabilities P(Y = ·|X = x). Here, we provide a more general approach where we control the expected size of ∗ [email protected][email protected] 1 the confidence sets. For a confidence set Γ, the expected size of Γ is defined as E[|Γ(X)|], where | · | stands for the cardinality. For a sample X and given an expected set size β, we provide an algorithm that outputs a set Γ̂(X) such that E[|Γ̂(X)|] ≈ β. Furthermore, the procedure aims at minimizing the classification error given by   P Y ∈ / Γ̂(X) ≈ min P (Y ∈ / Γ(X)) = R∗β . Γ : E[|Γ(X)|]=β We establish a close formula of the oracle confidence set Γ∗ = argminΓ : E[|Γ(X)|]=β P (Y ∈ / Γ(X)) that involves the cumulative distribution functions of the conditional probabilities. Besides, we formulate a data-driven counterpart of Γ∗ based on the minimization of the empirical risk. However, the natural risk function in the multiclass setting is non convex, and then minimizing it is proving to be a computational issue. As a remedy, convex surrogates are often used in machine learning, and more specifically in classification as reflected by the popularity of methods such as Boosting [FS97], logistic regression [FHT00] and support vector machine [Vap98]. In our problem this translates by considering some convex surrogate of the 0/1-loss in P (Y ∈ / Γ(X)); we introduce a new convex risk function which enables us to emphasize specific aspects of confidence sets. That convex risk function is partly inspired by the works in [Zha04, BJM06, YW10] that deal with binary classification. Our approach displays two main features. First, our method can be implemented in a semisupervised way [Vap98]. More precisely, the method is a two-steps procedure that requires the estimation of score functions (as minimizers of some convex risk function) in a first step and the estimation of the cumulative distribution of these scores in a second one. Hence, the first step requires labeled data whereas the second one involves only unlabeled samples. Notably, the unlabeled sample does not necessary consists of examples coming from the whole classes. This aspects is fondamental when we deal with a large number of classes, some of which been sparsely represented. Second, from the theoretical point of view, we provide an oracle inequality satisfied by Γ̂, the empirical confidence set that results from the minimization of an empirical convex risk. The obtained oracle inequality shall enable us to derive rates of convergence on a particular class of confidence sets under the Tsybakov noise assumption on the data generating distribution. These rates are linear in the number of classes K and also depend on the regularity of the cumulative distribution functions of the conditional probabilities. An obvious benefit of considering convex risk minimization is when we deal with aggregation. However, aggregating confidence sets is no simple matter. Another contribution of the present paper is about applying the above methodology to aggregation of confidence sets. More specifically, we provide a generalization of the superlearning algorithm [vdLPH07] initially introduced in the context of regression and binary classification. This algorithm relies on the V -fold cross-validation principle. We prove the consistency of this aggregation procedure and illustrate its relevance on real datasets. Let us point out that any arbitrary library of machine learning algorithms may be used in this aggregation procedure; we propose in the present paper to exploit support vector machines, random forest procedures or softmax regression since these methods are popular in machine learning and that each of them is associated to a different shape. We end up this discussion by highlighting in two words what we perceive as being the main contributions of the present paper. We describe an optimal strategy for building confidence sets in multiclass classification setting, and we derive a new aggregation procedure for confidence sets that still allows controlling the expected size of the resulting confidence set. Related works: The closest learning task to the present work is classification with reject option which is a particular setting in binary classification. Several papers fall within the scope of this area [Cho70, HW06, YW10, WY11, Lei14, DH15] and differ from each other by the goal they 2 consider. Among these references, our procedure is partially inspired by the paper [DH15] that also considers a semi-supervised approach to build confidence sets invoking some cumulative distribution functions (of the conditional probabilities themselves in their case). The similarity is however limited to the definition of oracle confidence sets, the oracle confidence in the present paper being an extension of the one defined in [DH15] to the multiclass setting. On the other hand, all the data-driven considerations are completely different (in particular, [DH15] focuses on plug-in rules) and importantly, we develop here new probabilistic results on sums of cumulative distribution functions of random variables, that are of own interest. Assigning a set of labels instead of a single one for an input example is not new [VGS05, WLW04, dCDB09, LRW13, CCB16]. One of the most popular methods is based on Conformal Prediction approach [VGS99, Vov02, VGS05]. In the multiclass classification framework, the goal of this algorithm is to build the smallest set of labels such that its classification error is below a pre-specified level. Since our procedure aims at minimizing the classification error while keeping under control the size of the set, Conformal Prediction can be seen as a dual of our method. It is worth mentioning that conformal predictors approaches need two labeled datasets where we only need one labeled dataset, the second being unlabeled. We refer to the very interesting statistical study of Conformal Predictors in the binary case in the paper [Lei14]. Notation: First, we state general notation. Let Y = {1, . . . , K}, with K ≥ 2 being an integer. Let (X, Y ) be the generic data-structure taking its values in X × Y with distribution P. The goal in classification is to predict the label Y given an observation of X. This is performed based on a classifier (or classification rule) s which is a function mapping X onto Y. Let S be the set of all classifiers. The misclassification risk R associated with s ∈ S is defined as R(s) = P(s(X) 6= Y ). Moreover, the minimizer of R over S is the Bayes classifier, denoted by s∗ , and is characterized by s∗ (·) = argmax pk (·), k∈Y where pk (x) = P(Y = k|X = x) for x ∈ X and k ∈ Y. Let us now consider more specific notation related to the multiclass confidence set setting. Let a confidence set be any measurable function that maps X onto 2Y . Let Γ be a confidence set. This confidence set is characterized by two attributes. The first one is the risk associated to the confidence set R (Γ) = P (Y ∈ / Γ(X)) , (1) and is related to its accuracy. The second attribute is linked to the information given by the confidence set. It is defined as I(Γ) = E (|Γ(X)|) , (2) where | · | stands for the cardinality. Moreover, for some β ∈ [1, K], we say that, for two confidence sets Γ and Γ′ such that I (Γ) = I (Γ′ ) = β, the confidence set Γ is “better” than Γ′ if R (Γ) ≤ R (Γ′ ). Organization of the paper: The rest of the paper is organized as follows. Next section is devoted to the definition and the main properties of the oracle confidence set for multiclass classification. The empirical risk minimization procedure is provided in Section 3. Rates of convergence for the confidence set that results from this minimization can also be found in this section. We present an application of our procedure to aggregation of confidence sets in Section 4. We finally draw some conclusions and present perspectives of our work in Section 5. Proofs of our results are postponed to the Appendix. 3 2 Confidence set for multiclass classification In the present section, we define a class of confidence sets that are suitable for multiclass classification and referred as Oracle β-sets. For some β ∈ (0, K), these sets are shown to be optimal according to the risk (1) with an information (2) equal to β. Moreover, basic but fondamental properties of Oracle β-sets can be found in Proposition 1, while Proposition 3 provides another interpretation of these sets. 2.1 Notation and definition First of all, we introduce in this section a class of confidence sets that specifies oracle confidence sets. Let β ∈ (0, K) be a desired information level. The so-called Oracle β-sets are optimal according to the risk (1) among all the confidence sets Γ such that I(Γ) = β. Throughout the paper we make the following assumption (A1) For all k ∈ {1, . . . , K}, the cumulative distribution function Fpk of pk (X) is continuous. The definition of the Oracle β-set relies on the continuous and decreasing function G defined for any t ∈ [0, 1] by K X G(t) = F̄pk (t), k=1 where for any k ∈ {1, . . . , K}, we denote by F̄pk the tail distribution function of pk (X), that is, F̄pk = 1 − Fpk with Fpk being the cumulative distribution function (c.d.f.) of pk (X). The generalized inverse G−1 of G is given by (see [vdV98]): G−1 (β) = inf{t ∈ [0, 1] : G(t) ≤ β}, ∀β ∈ (0, K). The functions G and G−1 are central in the construction of the Oracle β-sets. We then provide some of their useful properties in the following proposition. Proposition 1. The following properties on G hold i) For every t ∈ (0, 1) and β ∈ (0, K), G−1 (β) ≤ t ⇔ β ≥ G(t). ii) For every β ∈ (0, K), G(G−1 (β)) = β. iii) Let ε be a random variable, independent of X, and distributed from a uniform distribution on {1, . . . , K} and let U be uniformly distributed on [0, K]. Define Z= K X k=1 pk (X)1{ε=k} . L L If the function G is continuous, then G(Z) = U and G−1 (U ) = Z. The proof of Proposition 1 relies on Lemma 1 in the Appendix. Now, we are able to defined the Oracle β-set: Definition 1. Let β ∈ (0, K), the Oracle β-set is given by Γ∗β (X) = {k ∈ {1, . . . , K} : G(pk (X)) ≤ β}  = k ∈ {1, . . . , K} : pk (X) ≥ G−1 (β) . 4 This definition of the Oracle β-set is intuitive and can be related to the binary classification with reject option setting [Cho70, HW06, DH15] in the following way: a label k is assigned to the Oracle β-set if the probability pk (X) is large enough. It is worth noting that the function G plays the same role as the c.d.f. of the score function in [DH15]. As emphasized by Proposition 1, their introduction allows to control exactly the information (2). Indeed, it follows from the definition of the Oracle β-set that for each β ∈ (0, K) |Γ∗β (X)| = K X k=1 1{pk (X)≥G−1 (β)} , h i and then I(Γ∗β ) = E |Γ∗β (X)| = G(G−1 (β)). Therefore, Proposition 1 ensures that I(Γ∗β ) = β. This last display points out that the Oracle β-sets are indeed β-level (that is, its information equals β). In the next section, we focus on the study of the risk of these oracle confidence sets. Remark 1. Naturally, the definition of Oracle β-sets can be extended to any β ∈ [0, K]. However, the limit cases β = 0 and β = K are of limited interest and are completely trivial. We then exclude these two limit cases from the present study. 2.2 Properties of the oracle confidence sets Let us first state the optimality of the Oracle β-set: Proposition 2. Let Assumption (A1) be satisfied. For any β ∈ (0, K), we have both: 1. The Oracle β-set Γ∗β satisfies the following property:  R Γ∗β = inf R (Γ) , Γ where the infimum is taken over all β-level confidence sets. 2. For any β-level confidence set Γ, the following holds  X  0 ≤ R (Γ) − R Γ∗β = E  k∈(Γ∗ (X) ∆ Γ(X)) β  pk (X) − G−1 (β)  , (3) where the symbol ∆ stands for the symmetric difference of two sets, that is, for two subsets A and B of {1, . . . , K}, we write A ∆ B = (A \ B) ∪ (B \ A). Several remarks can be made from Proposition 2. First, for β ∈ (0, K), the Oracle β-set is optimal for the misclassification risk, over the class of β-level confidence sets. Moreover, the excess risk of any β-level confidence set relies on the behavior of the score functions pk around the threshold G−1 (β). Finally, we can note that if K = 2 and β = 1, which implies that G−1 (β) = 1/2, Equation (3) reduces to the misclassification excess risk in binary classification. Remark 2. One way to build a confidence set Γ with information β is to set Γ as the β top levels conditional probabilities. In the sequel this method is referred as the max procedure. This strategy is natural but actually suboptimal as shown by the first point of Proposition 2. As an illustration, we consider a simulation scheme with K = 10 classes. We generate (X, Y ) according to a mixture model. More precisely, 5 i) the label Y is distributed from a uniform distribution on {1, . . . , K}; ii) conditional on Y = k, the feature X is generated according to a multivariate gaussian distribution with mean parameter µk ∈ R10 and identity covariance matrix. For each k = 1, . . . , K, the vectors µk are i.i.d realizations of uniform distribution on [0, 4]. For β = 2 we evaluate the risks of the Oracle β-set and the max procedure and obtain respectively 0.05 and 0.09 (with very small variance). Remark 3. An important motivation behind the introduction of confidence sets and in particular of Oracle β-sets is that they might outperform the Bayes rule which can be seen as the Oracle β-set associated to β = 1. This gap in performance is even larger when the number of classes K is large and there is a big confusion between classes. Such improvement will be illustrated in the numerical study (see Section 4.3). We end up this section by providing another characterization of the Oracle β-set. Proposition 3. For t ∈ [0, 1], and Γ a confidence set, we define Lt (Γ) = P (Y ∈ / Γ(X)) + tI(Γ). For β ∈ [0, K], the following equality holds: LG−1 (β) (Γ∗β ) = min LG−1 (β) (Γ). Γ The proof of this proposition relies on the same arguments as those of Proposition 2. It is then omitted. Proposition 3 states that the Oracle β-set is defined as the minimizer, over all confidence sets Γ, of the risk function Lt when the tuning parameter t is set equal to G−1 (β). Note moreover that the risk function Lt is a trade-off, controlled by the parameter t, between the risk of a confidence set on the one hand, and the information provided by this confidence set on the other hand. Hence, the risk function Lt can be viewed as a generalization to the multiclass case of the risk function provided in [Cho70, HW06] for binary classification with reject option setting. 3 Empirical risk minimization In this section we introduce and study confidence sets which rely on the minimization of convex risks. Their definitions and main properties are given in Sections 3.1-3.2. As a consequence, we deduce a data-driven procedure described in Section 3.3 with several theoretical properties, such as rates of convergence, that we demonstrate in Section 3.4. 3.1 φ-risk PK Let f = (f1 , . . . , fK ) : X → RK be a score function and Gf (.) = k=1 F̄fk (.). Assuming that the function Gf is continuous and given an information level β ∈ (0, K), there exists δ ∈ R, such that Gf (−δ) = β. Given this simple but important fact, we define the confidence set Γf,δ associated to f and δ as Γf,δ (X) = {k : fk (X) ≥ −δ}. (4) In this way, the confidence set Γf,δ consists of top scores, and the threshold δ is fixed so that I (Γf,δ ) = β. As a consequence, we naturally aim at solving the problem min R(Γf,δ ), f ∈F 6 where F is a class of functions. Due to computational issues, it is convenient to focus on a convex surrogate of the previous minimization problem. To this end, let φ : R → R be a convex function. We define the φ-risk of f by "K # X Rφ (f ) = E φ(Zk fk (X)) , (5) k=1 where Zk = 2 1{Y =k} − 1 for all k = 1, . . . , K. Therefore, our target score becomes f¯ ∈ argmin Rφ (f ) , f ∈F for the purpose of building the confidence Γf¯,δ . In the sequel, we also introduce f ∗ , the minimizer over the class of all measurable functions, of the φ-risk. The notation suppresses the dependence on φ. It is worth mentioning at this point that the definition of the risk function Rφ is dictated by Equation (3) and suits for confidence sets. Moreover, this function differs from the classical risk function used in the multiclass setting (see [TB07]). The reason behind this is that building a confidence set is closer to K binary classification problems. 3.2 Classification calibration for confidence sets Convexification of the risk in classification is a standard technique. In this section we adapt classical results and tools to confidence sets in the multiclass setting. We refer the reader to earlier papers as [Zha04, BJM06, YW10] for interesting developments. One of the main important concept when we deal with convexification of the risk is the notion of calibration of the loss function φ. This property permits to connect confidence sets deduced from the convex risk and from the classification risk. Definition 2. We say that the function φ is confidence set calibrated if for all β > 0, there exists δ ∗ ∈ R such that Γf ∗ ,δ∗ = Γ∗β , with f ∗ being the minimizer of the φ-risk f ∗ ∈ argmin Rφ (f ) , f where the infimum is taken over the class of all measurable functions. Hence, the confidence set based on f ∗ coincides with the Bayes confidence set. Given this, we can state the following proposition that gives a characterization of the confidence set calibration property in terms of the function G. Proposition 4. The function φ is confidence set calibrated if and only if for all β ∈ (0, K), there exists δ ∗ ∈ R such that φ′ (δ ∗ ) and φ′ (−δ ∗ ) both exists, φ′ (δ ∗ ) < 0, φ′ (−δ ∗ ) < 0 and G−1 (β) = φ′ (δ ∗ ) , + φ′ (−δ ∗ ) φ′ (δ ∗ ) where φ′ denotes the derivative of φ. The proof of the proposition follows the lines of Theorem 1 in [YW10] with minor modifications. These characterizations of calibration for confidence sets generalize the notion of 7 calibration in the classification setting as well as the necessary and sufficient condition for the minimizer of the φ-risk to be calibrated. Indeed, if we pick δ = 0 in Definition 2 and Proposition 4 we exactly come back to the calibration property in the classical classification setting (see [BJM06]). Note that commonly used loss functions as boosting (x 7→ exp(−x)), least squares (x 7→ (x − 1)2 ) and logistic (x 7→ log(1 + exp(−x))) are examples of calibrated losses (see for instance [BJM06, WY11]). Now, for some score function f and some real number δ such that Gf (−δ) = β, we define the excess risk ∆R(Γf,δ ) = R(Γf,δ ) − R(Γ∗β ), and the excess φ-risk ∆Rφ (f ) = Rφ (f ) − Rφ (f ∗ ) . We also introduce the marginal conditional excess φ-risk on f = (f1 , . . . , fK ) as ∆Rφk (f (X)) = pk (X)(φ(fk (X)) − φ(fk∗ (X))) + (1 − pk (X))(φ(−fk (X)) − φ(−fk∗ (X))), for k = 1, . . . , K. The following theorem shows that the consistency in terms of φ-risk implies the consistency in terms of classification risk R. Theorem 1. Assume that φ is confidence set calibrated and assume that there exists constants C > 0 and s ≥ 1 such that1 |pk (X) − G−1 (β)|s ≤ C∆Rφk (−δ ∗ ). (6) Let fˆn be a sequence of scores. We assume that for each n, the function Gfˆn is continuous. Let δn ∈ R be such that Gfˆn (−δn ) = β, then P ∆Rφ (fˆn ) → 0 P ∆R(Γfˆn ,δn ) → 0. ⇒ The theorem ensures that the convergence in terms of φ risk implies the convergence in terms of risk R. This convergence is made possible since we manage, in the proof, to bound the excess risk by (a power of) the excess φ-risk. The assumption needed in this theorem is also standard and is for instance satisfied for the boosting, least square and logistic losses with the parameter s being equal to 2 (see [BJM06]). 3.3 Data-driven procedure In this section we provide the steps of the construction of our empirical confidence set that is deduced from the empirical risk minimization. Before going into further details, let us first mention that our procedure is semi-supervised in the sense that it requires two datasets, one of which being unlabeled. Hence we introduce a first data set Dn = {(Xi , Yi ), i = 1, . . . , n}, which consists of independent copies of (X, Y ). We define the empirical φ-risk associated to a score function f (which is the empirical counterpart of Rφ given in (5)): n R̂φ (f ) = K 1 XX φ(Zki fk (Xi )), n i=1 (7) k=1 where Zki = 2 1{Yi =k} − 1 for all k = 1, . . . , K. We also define the empirical risk minimizer over F , a convex set of score functions, as fˆ = arg min R̂φ (f ). f ∈F 1 With abuse of notation, we write ∆Rkφ (−δ∗ ) instead of ∆Rkφ ((−δ∗ , . . . , −δ∗ )) since no confusion can occur. 8 At this point, we have in hands the optimal score function fˆ and need to build the corresponding confidence set with the right expected size. However, let us before introduce an intermediate confidence set that would help comprehension since it mimics the oracle β-set Γ∗ with regard to its construction using fˆ instead of f ∗ . For this purpose, we define   Ffˆk (t) = PX fˆk (X) ≤ t|Dn , for t ∈ R, where PX is the marginal distribution of X. As for the c.d.f. of pk and fk , we make the following assumption: (A2) The cumulative distribution functions Ffˆk with k = 1, . . . , K are continuous. At this point, we are able to define an empirical approximation of the Oracle β-set for β ∈ (0, K): n o e β (X) = k ∈ {1, . . . , K} : G( e fˆk (X)) ≤ β , Γ (8) where for t ∈ R e = G(t) K X F̄fˆk (t), (9) k=1 e depends on the unknown distribution of X, we consider with F̄fˆk = 1 − Ffˆk . Since the function G a second but interestingly unlabeled dataset DN = {Xi , i = 1, . . . , N }, independent of Dn in order to compute the empirical versions of the F̄fˆk ’s. By now, we can define the empirical β-set based on fˆ: Definition 3. Let fˆ be the minimizer of the empirical φ-risk given in (7) based on Dn , and consider the unlabeled dataset DN . Let β ∈ (0, K). The empirical β-set is given by n o Γ̂β (X) = k ∈ {1, . . . , K} : Ĝ(fˆk (X)) ≤ β , (10) where N K 1 XX Ĝ(.) = 1{fˆk (Xi )≥.} . N i=1 k=1 The most important remark about the construction of this data-driven confidence set is that it is made in a semi-supervised way. Indeed, the estimation of the tail distribution functions of fˆk requires only a set of unlabeled observations. This is particularly attractive in applications where the number of label observations is small (because labelling examples is time consuming or may be costly) and where one has in hand several unlabeled features that can be used to make prediction more accurate. As an important consequence is that the estimation of the tail distribution functions of fˆk does not depend on the number of observations in each class label. That is to say, this unlabeled dataset can be unbalanced with respect to the classes, that can often occur in multiclass classification settings where the number of classes K is quite large. Next section deals with the theoretical performance of the empirical β-sets. 3.4 Rates of convergence In this Section, we provide rates of convergence for the empirical confidence sets defined in Section 3.3. First, we state some additional notation. In the sequel, the symbols P and E 9 stand for generic probability and expectation, respectively. the empirical βMoreover, given    / Γ̂β (X) and the information set Γ̂β from Definition 3, we consider the risk R Γ̂β = P Y ∈ h i   √ I Γ̂β = E |Γ̂β (X)| . Our first result ensures that Γ̂β is of level β up to a term of order K/ N . Proposition 5. For each β ∈ (0, K), the following equality holds     K . I Γ̂β = β + O √ N In order to precise rates of convergence for the risk, we need to formulate several assumptions. First, we consider loss functions φ which have in common that the modulus of convexity of their underlying risk function Rφ , defined by ) (   X K   Rφ (f ) + Rφ (g) f +g 2 2 (11) δ(ε) = inf EX (fk − gk ) (X) ≥ ε , , − Rφ 2 2 k=1 satisfies δ(ε) ≥ c1 ǫ2 for some c1 > 0 (we refer to [BJM06, BM06] for more details on modulus of convexity for classification risk). Moreover, we assume that φ is classification calibrated and L-Lipschitz for L > 0. On the other hand, for α > 0, we assume a margin condition Mkα on each pk , k = 1, . . . , K.  Mkα : PX 0 < |pk (X) − G−1 (β)| ≤ t ≤ c2 tα , for some constant c2 > 0 and for all t > 0. It is important to note that since we assume that the distribution functions of pk (X) are contin uous for each k, we have PX 0 < |pk (X) − G−1 (β)| ≤ t → 0 with t → 0. Therefore, the margin condition only precise the rate of this decay to 0. Now, we provide the rate of convergence that we can expect for the empirical confidence sets defined by (10). Theorem 2. Assume that kf k∞ ≤ B for all f ∈ F. Let Mn = N (1/n, L∞ , F ) be the cardinality of the set of closed balls with radius 1/n in L∞ -norm needed to cover F . Under the assumptions of Theorem 1, with the modulos of convexity δ(ε) ≥ c1 ǫ2 with c1 > 0, if φ is L-Lipschitz and if the margin assumptions Mkα are satisfied with α > 0, then α/(α+s)  h i ′ K K log(Mn ) 1−α/(α+s) inf ∆Rφ (f ) + E |∆R(Γ̂β )| ≤ C(B, L, s, α)K +C √ , f ∈F n N ′ where C > 0 is an absolute constant and C(B, L, s, α) > 0 is a constant that depends only on L, B, s and α.  α/(α+s) n) From this theorem we obtain the following rate of convergence: K log(M + √KN . n This is the first bound for confidence sets in multiclass setting. The regularity parameter α plays a crucial role and governs the rate; larger values of α lead to faster rates. Moreover, this rate is linear in K, the number of classes, that seems to be the characteristic of multiclass classification as compared to binary classification. Note that the exponent α/(α + s) is not classical √ and is due to the estimation of quantiles. The second part of the rates which is of order K/ N relies on the estimation of the function G̃ given in (9). This part of the estimation is established under the mild Assumptions (A1) and (A2). For this term the proof of the linearity on K is tricky and is obtained thanks to a new technical probabilistic results on sums of cumulative distribution functions (see Lemma 1). Let us conclude this paragraph by mentioning that Theorem 2 applies for instance for F being the convex hull of a finite family of a score functions which is the scope of the next section. 10 4 Application to confidence sets aggregation This Section is devoted to an aggregation procedure which relies on the superlearning principle. The specifics of our superlearning procedure is given in Section 4.1. In Section 4.2, we show the consistency of our procedure and finally we provide a numerical illustration in Section 4.3 of our algorithm. 4.1 Description of the procedure: superlearning In this section, we describe the superlearning procedure for confidence sets. The superlearning procedure is based on the V -fold cross-validation procedure (see [vdLPH07]). Initially, this aggregation procedure has been introduced in the context of regression and binary classification. Let V ≥ 2 be an integer and let (Bv )1≤v≤V be a regular partition of {1, . . . , n}, i.e., a partition such that, for each v = 1, . . . , V , card(Bv ) ∈ {⌊n/V ⌋, ⌊n/V ⌋ + 1}, where we write ⌊x⌋ for the (v) floor of any real number x (that is, x − 1 ≤ ⌊x⌋ ≤ x). For each v ∈ {1, . . . , V }, we denote by Dn (−v) (respectively Dn ) the dataset {(Xi , Yi ), i ∈ Bv } (respectively {(Xi , Yi ), i 6∈ Bv }), and define the corresponding empirical measures Pn(v) = Pn(−v) = X 1 Dirac(Xi , Yi ), and card(Bv ) i∈Bv X 1 Dirac(Xi , Yi ). n − card(Bv ) i6∈Bv For a score algorithm f , that is to say a function which maps the empirical distribution to a score function. We define the cross-validated risk of f as V   X (P (v) ) bφn (f ) = 1 Rφ n f (Pn(−v) ) , R V v=1 (Pn(v) ) where for each v ∈ {1, . . . , V }, Rφ based on (v) Dn (Pn(v) )   f (Pn(−v) ) = Next, we consider F = f 1 , . . . , f validated score by (−v) )) is the empirical estimator of Rφ (f (Pn ), (−v) Dn : and conditionally on Rφ (−v) (f (Pn (12)  M K XX 1 φ(Zki fk (Pn(−v) )(Xi )). card(Bv ) i∈Iv k=1 a family of M score algorithms. We define the cross- bn (f ). fb ∈ argmin R φ f ∈conv(F ) Finally, we consider the resulting cross-validated confidence set defined by n o ˆ Γ̂CV β (X) = k ∈ {1, . . . , K} : Ĝ(fk (X)) ≤ β , that we analyse in the next section. 11 (13) 4.2 Consistency of the algorithm In this Section, we show some results that illustrate the consistency of the superlearning procedure described above. To this end, we introduce the oracle counterpart of the cross-validated risk defined in (12). For a score f , we define R̃φn (f ) = V 1 X Rφ (f (Pn(−v) ), V v=1 that yields to the oracle counterpart of the cross-validated score defined in (13) f˜ ∈ argmin R̃φn (f ). f ∈conv(F ) Here again, we assume that the loss function φ satisfies the properties described in Section 3.4 so that we can state the following result: (−v) Proposition 6. We assume that for each f ∈ conv(F ) and v ∈ {1, . . . , V }, kf (Pn Then the following holds )k∞ ≤ B. h i KM log(n) E R̃φn (fˆ) − R̃φn (f˜) ≤ C(B, L) , ⌊n/V ⌋ where C(B, L) > 0 is a constant that depends only on B and L. As usual when one deals with cross-validated estimators, the theorem compares fˆ to the oracle counterpart f˜ in terms of the oracle cross-validated φ-risk. The theorem teaches us that, for sufficiently large n, fˆ perform as well as f˜. However our main goal remains confidence sets. Therefore the next step consists in showing that the confidence set associated to the crossvalidated score fˆ has good properties in terms of the classification risk. Let Γf,δ be the confidence set that results from a choice of β ∈ (0, K) and a score function f ∈ conv(F ). We introduce the following excess risks ∆R̃n (Γf,δ ) = ∆R̃φn (f ) =  V 1 X R Γf V v=1 (−v) Pn  ,δ  − R∗ , V 1 X Rφ (f (Pn(−v) )) − Rφ (f ∗ ). V v=1 These quantities can be view as cross-validated counterparts of the excess risks introduced in Section 3.2. Hereafter, we provide a result which can be view as the cross-validation counterpart of Theorem 2. It is interpreted in a same way.     (−v) ≤ t|Dn is conProposition 7. We assume that for each v ∈ {1, . . . , V }, t 7→ PX fˆ Pn tinuous. Under the assumptions of Proposition 6 and if the margin assumptions Mkα are satisfied with α > 0,   h i   KM log(n) α/(α+s) ′ K n ˜ n CV 1−α/(α+s) E ∆R̃φ (f ) + E |∆R̃ (Γ̂β )| ≤ C(B, L, α, s)K +C √ . ⌊n/V ⌋ N The proof of Proposition 7 relies on Proposition 6 and similar arguments as in Theorem 2. 12 Forest (K = 4) β-set β 2 rforest R I 0.02 (0.02) 2.00 (0.09) softmax reg svm 0.06 (0.02) 2.00 (0.08) 0.02 (0.01) 2.00 (0.09) kknn 0.05 (0.03) 2.00 (0.08) CV 0.02 (0.01) 2.00 (0.08) Plant (K = 100) β-set β 2 10 rforest R I R I 0.18 2.00 0.02 9.95 (0.03) (0.09) (0.01) (0.38) softmax reg svm 0.77 (0.02) 2.02 (0.18) 0.42 (0.04) 10.06 (0.58) 0.32 1.99 0.03 9.98 (0.04) (0.10) (0.02) (0.22) kknn 0.20 2.00 0.08 9.98 (0.03) (0.08) (0.03) (0.23) CV 0.17 2.00 0.02 9.96 (0.03) (0.08) (0.01) (0.37) Table 1: For each of the B = 100 repetitions and for each dataset, we derive the estimated risks R and information I of the different β-sets w.r.t. β. We compute the means and standard deviations (between parentheses) over the B = 100 repetitions. For each β, the β-sets are based on, from left to right, rforest, softmax reg and svm, kknn and CV which are respectively the random forest, the softmax regression, support vector machines, k nearest neighbors and the superlearning procedure. Top: the dataset is the Forest – the dataset is the Plant. 4.3 Application to real datasets In this section, we provide an application of our aggregation procedure described in Section 4.1. For the numerical experiment we focus on the boosting loss and consider the library of algorithms constituted by the random forest, the softmax regression, the support vector machines and the k nearest neighbors (with k = 11) procedures. To be more specifics, we respectively exploit the R packages randomForest, polspline, e1071 and kknn. All the R functions are used with standard tuning parameters. Finally the parameter V of the aggregation procedure is fixed to 5. We evaluate the performance of the procedure on two real datasets: the Forest type mapping dataset and the one-hundred plant species leaves dataset coming from the UCI database. In the sequel we refer to these two datasets as Forest and Plant respectively. The Forest dataset consists of K = 4 classes and 523 labeled observations (we gather the train and test sets) with 27 features. Here the classes are unbalanced. In the Plant dataset, there are K = 100 classes and 1600 labeled observations. This dataset is balanced so that each class consists of 16 observations. The original dataset contains 3 covariates (each covariate consists of 64 features). In order to make the problem more challenging, we drop 2 covariates. To get an indication of the statistical significance, it makes sense to compare our aggregated confidence set (referred as CV) to the confidence sets that result from each component of the library. Roughly speaking, we evaluate risks (and informations) of these confidence sets on each dataset. To do so, we use the cross validation principle. In particular, we run B = 100 times the procedure where we split the data each time in three: a sample of size n to build the scores fˆ; a sample of size N to estimate the function G and to get the confidence sets; and a sample of size M to evaluate the risk and the information. For both datasets, we make sure that in the sample of size n, there is the same number of observations in each class. As a benchmark, we notify that the misclassification risks of the best classifier from the library in the Forest dataset is evaluated at 0.15 , whereas in the Plant dataset, it is evaluated at 0.40. As planned, the performance of the classical methods are quite bad in this last dataset. We set the sizes of the samples as n = 200, N = 100 and M = 223 for the Forest dataset, 13 and n = 1000, N = 200 and M = 400 for the Plant one. The results are reported in Table 1, and confirm our expectations. In particular, our main observation is that the aggregated confidence set (CV) outperforms all components of the library in the sense that it is at least as good as the best component in all of the experiments. Second, let us state some remarks that hold for all of the confidence sets and in particular our aggregated confidence set. First, we note that the information I(Γ) has the good level β which is supported by our theory. Moreover, we see that the risk gets drastically better with moderate β as compared to the best misclassification risk. For instance, for the Plant, the error rate of the confidence set with β = 2 based on random forests is 0.18 whereas the misclassification error rate of the best component is 0.40. 5 Conclusion In multiclass classification setting, the present paper propose a new procedure that assigns a set of labels instead of a single label to each instance. This set has a controlled expected size (or information) and its construction relies on cumulative distribution functions and on an empirical risk minimization procedure. Theoretical guarantees, especially rates of convergence are also provided and rely on the regularity of these cumulative distribution functions. The obtained rates of convergence highlight a linear dependence w.r.t the number of classes K. The procedure described in Section 3 is defined as a two steps algorithm whose second step consists in the estimation of the function G (which is a sum of tail distribution functions). Interestingly, this step does not require a set of labeled data and neither to explore the whole classes, that is suitable for semi-supervised learning. Moreover, we apply our methodology to derive an aggregation algorithm which is based on the V -fold cross-validation principle. Future works will focus on the optimality in the minimax sense with respect to the classification error. In particular, we will investigate whether the rates of convergence are optimal in terms of their dependence on K. We believe that this dependence (linearity on K) is the correct one. However we will instigate whether this dependence can be reduced under some sparsity assumption. 6 Appendix This section gathers the proofs of our results. Let us first add a notation that will be used throughout the Appendix: for any random variable (or vector) Z, we denote by PZ the probability w.r.t. Z and by EZ , the corresponding expectation. 6.1 Technical Lemmas We first lay out key lemmata, which are crucial to establish the main theory. We consider K ≥ 2 be an integer, and Z1 , . . . , ZK , K random variables. Moreover we define function H by: H(t) = K 1 X Fk (t), ∀t ∈ [0, 1], K k=1 where for all k = 1, . . . , K, Fk is the cumulative distribution function of Zk . Finally, let us define the generalized inverse H −1 of H: H −1 (p) = inf{t : H(t) ≥ p}, ∀p ∈ (0, 1). 14 Lemma 1. Let ε distributed from a uniform distribution on {1, . . . , K} and independent of Zk , k = 1, . . . , K. Let U distributed from a uniform distribution on [0, 1]. We consider Z= K X k=1 Zk 1{ε=k} . If H is continuous then L L H(Z) = U and H −1 (U ) = Z  Proof. First we note that for every t ∈ [0, 1], P (H(Z) ≤ t) = P Z ≤ H −1 (t) . Moreover, we have P (H(Z) ≤ t) = = K X k=1 P(Z ≤ H −1 (t), ε = k) K 1 X P(Zk ≤ H −1 (t)) with ε independent of X K k=1 = H(H −1 (t)) = t with H continuous. To conclude the proof, we observe that P H −1 (U ) ≤ t  = = P (U ≤ H(t)) K 1 X Fk (t) K k=1 = K X k=1 = P (Zk ≤ t, ε = k) P(Z ≤ t). Lemma 2. There exists an absolute constant C ′ > 0 such that   ′ K e fˆk (X))| ≥ |G( e fˆk (X)) − β| ≤ C √ . P |Ĝ(fˆk (X) − G( N k=1 K X Proof. We define, for γ > 0 and k ∈ {1, . . . K} n o e fˆk (X)) − β ≤ γ Ak0 = G( n o e fˆk (X)) − β| ≤ 2j γ , j ≥ 1. Akj = 2j−1 γ < |G( Since, for every k, the events (Akj )j≥0 are mutually exclusive, we deduce K X k=1   e fˆk (X))| ≥ |G( e fˆk (X)) − β| = P |Ĝ(fˆk (X) − G( K X X k=1 j≥0   e fˆk (X))| ≥ |G( e fˆk (X)) − β|, Akj . (14) P |Ĝ(fˆk (X) − G( 15 Now, we consider a random variable ε uniformly distributed on {1, . . . , K} and independent of Dn and X. Conditional on Dn and under Assumption (A2), we apply Lemma 1 with Zk = fˆk (X), PK e is uniformly distributed on [0, K]. Therefore, Z = k=1 Zk 1{ε=k} and then obtain that G(Z) for all j ≥ 0 and γ > 0, we deduce K   1 X e fˆk (X)) − β| ≤ 2j γ|Dn PX |G( K k=1 Hence, for all j ≥ 0, we obtain K X k=1 Next, we observe that for all j ≥ 1 K X k=1   e − β| ≤ 2j γ|Dn = PX |G(Z) ≤ 2j+1 γ . K P(Akj ) ≤ 2j+1 γ. (15)   e fˆk (X))| ≥ |G( e fˆk (X)) − β|, Akj ≤ P |Ĝ(fˆk (X) − G( K X k=1 i   h e fˆk (X))| ≥ 2j−1 γ|Dn , X 1Ak . (16) E(Dn ,X) PDN |Ĝ(fˆk (X)) − G( j Now, since conditional on (Dn , X), Ĝ(fˆk (X)) is an empirical mean of i.i.d random variables of e fˆk (X)) ∈ [0, K], we deduce from Hoeffding’s inequality that common mean G(     N γ 2 22j−1 j−1 ˆ ˆ e . PDN |Ĝ(fk (X)) − G(fk (X))| ≥ 2 γ|Dn , X ≤ 2 exp − K2 Therefore, from Inequalities (14), (15) and (16), we get K X k=1   e fˆk (X))| ≥ |G( e fˆk (X)) − β| P |Ĝ(fˆk (X) − G( !   X K  N γ 2 22j−1 k P + 2 exp − ≤ P Aj K2 j≥1 k=1 k=1   X N γ 2 22j−1 j+2 . ≤ 2γ + γ 2 exp − K2 K X Ak0  X j≥1 K Finally, choosing γ = √ in the above inequality, we finish the proof of the lemma. N 16 6.2 Proof of Proposition 2 Let β > 0 and Γ be a confidence set such that I(Γ) = β. First, we note that the following decomposition holds R(Γ) − R(Γ∗β ) = = K X K X K X K X E j=1 k=1 = E j=1 k=1 K X K X " E j=1 k=1 " K X l=1 K X l=1 (1{Y ∈Γ(X)} − 1{Y ∈Γ / / ∗ (X)} )1{Y =l} 1{|Γ(X)|=k} 1{|Γ∗ (X)|=j} β β (1{l∈Γ(X)} − 1{l∈Γ / / ∗ (X)} )pl (X)1{|Γ(X)|=k} 1{|Γ∗ (X)|=j} β β "K X l=1 # # # (1{l∈Γ∗β (X)\Γ(X)} − 1{l∈Γ(X)\Γ∗β (X)} )pl (X)1{|Γ(X)|=k,|Γ∗β (X)|=j} , where we conditioned by X to get the second equality. From the last decomposition and with   E [|Γ(X)|] = E |Γ∗β (X)|   K X K X = E k1{|Γ(X)|=k} 1{|Γ∗β (X)|=j}  j=1 k=1   K X K X E j1{|Γ(X)|=k} 1{|Γ∗ (X)|=j}  , = β j=1 k=1 we can express the excess risk as the sum of two terms: R(Γ) − R(Γ∗β ) + = K X K X E j=1 k=1 K X K X j=1 k=1 E " " −1 kG K X l=1 −1 1{l∈Γ∗β (X)\Γ(X)} pl (X) − jG (β) − K X l=1 ! ! (β) 1{|Γ(X)|=k} 1{|Γ∗β (X)|=j} 1{l∈Γ(X)\Γ∗β (X)} pl (X) 1{|Γ(X)|=k} 1{|Γ∗β (X)|=j} . (17) Now, for j, k ∈ {1, . . . , K} on the event {|Γ(X)| = k, |Γ∗β (X)| = j}, we have k= K X 1{l∈Γ(X)\Γ∗β (X)} + K X 1{l∈Γ(X)∩Γ∗β (X)} , K X 1{l∈Γ∗β (X)\Γ(X)} + K X 1{l∈Γ(X)∩Γ∗β (X)} . l=1 and j= l=1 Therefore, since # # l=1 l=1 l ∈ Γ∗β ⇔ pl (X) ≥ G−1 (β), Equality (17) yields the result. 17 6.3 Proof of Theorem 1 First we recall that fˆn = (fˆn,1 , . . . , fˆn,K ) is a sequence of score functions and δn ∈ R is such that Gfˆn (−δn ) = β. We suppress the dependence on n to simplify notation and write fˆ = (fˆ1 , . . . , fˆK ) and δ for fˆ and δn respectively. Moreover, since there is no doubt, we also suppress everywhere the dependence on X. We also define the events Bk = {fˆk ∈ (−δ, −δ ∗ ) or fˆk ∈ (−δ ∗ , −δ)}, (18) for k = 1, . . . , K. We aim at controlling the excess risk ∆R(Γfˆ,δ ). Since the risk R is decomposable, it is convenient to introduce “marginal excess risks”: ∆Rk (Γf,δ ) = 1{k∈(Γfˆ,δ ∆ Γ∗β )} |pk − G−1 (β)|. Recall also that by convexity of the loss function φ, we have that for all x, y ∈ R, φ(y) − φ(x) ≥ φ′ (x)(y − x). (19) Assume that fˆk ≤ −δ and fˆk ≤ −δ ∗ ≤ fk∗ , which translates as pk − G−1 (β) ≥ 0, we get thanks to (19) pk (φ(fˆk ) − φ(−δ)) − (1 − pk )(φ(−fˆk ) − φ(δ)) ≥ (φ′ (δ ∗ ) − φ′ (−δ ∗ ))(pk − G−1 (β))(fˆk + δ ∗ ) ≥ 0. Similarly, if fˆk ≥ −δ and fˆk ≥ −δ ∗ ≥ fk∗ , that is, pk − G−1 (β) ≤ 0 we have pk (φ(fˆk ) − φ(−δ)) − (1 − pk )(φ(−fˆk ) − φ(δ)) ≥ 0. Note that in the following two cases • fˆk ≤ −δ • fˆk ≥ −δ and fk∗ ≤ −δ ∗ ; and fk∗ ≥ −δ ∗ , we have ∆Rk (Γfˆ,δ ) = 0. Therefore, from the above inequalities, on Bkc and by assumption (6) we get (20) 1Bkc 1{k∈(Γfˆ,δ ∆ Γ∗β )} |pk − G−1 (β)|s ≤ C(∆Rφk (fˆ)). Therefore, since s ≥ 1, we have "K #!s X E 1Bkc ∩{k∈(Γfˆ,δ ∆ Γ∗β )} |pk − G−1 (β)| k=1 K i 1 X h E 1Bkc ∩{k∈Γfˆ,δ ∆Γ∗β } K s |pk − G−1 (β)|s K ≤ k=1 CK s−1 ∆Rφ (fˆ). ≤ Moreover E " K X k=1 −1 1Bk 1{k∈(Γfˆ,δ ∆ Γ∗β )} |pk − G # (β)| ≤ ≤ ≤ 18 K X k=1 P(Bk ) h i E |Gfˆ(−δ) − Gfˆ(−δ ∗ )| h i E |Gf ∗ (−δ ∗ ) − Gfˆ(−δ ∗ )| . Finally, we get the following bound ∆R(Γfˆ,δ ) ≤ K s−1 s Now we observe that i h ∆Rφ (fˆ)1/s + E |Gf ∗ (−δ ∗ ) − Gfˆ(−δ ∗ )| . |1{fˆk ≥−δ∗ } − 1{fk∗ ≥−δ∗ } | ≤ 1{|pk −G−1 (β)|s ≤C∆Rk (fˆ)} . φ (21) Therefore, for each γ > 0, we have i h EX |Gf ∗ (−δ ∗ ) − Gfˆ(−δ ∗ )| ≤ ≤ K X k=1 K X k=1   PX |pk (X) − G−1 (β)|s ≤ C∆Rφk (fˆ)     PX |pk (X) − G−1 (β)| ≤ γ 1/s + PX γ ≤ C∆Rφk (fˆ) . Now using Markov Inequality, we have that K X k=1   ≤ PX γ ≤ C∆Rφk (fˆ) K h i CX EX ∆Rφk (fˆ) γ k=1 C∆Rφ (fˆ) . γ ≤ The above inequality yields K   i C∆R (fˆ) X h φ PX |pk (X) − G−1 (β)| ≤ γ 1/s . + EX |Gf ∗ (−δ ∗ ) − Gfˆ(−δ ∗ )| ≤ γ (22) k=1 Hence, with Equation (21), we get ∆R(Γfˆ,δ ) ≤ K s−1 s ∆Rφ (fˆ)1/s + K   C∆Rφ (fˆ) X + PX |pk (X) − G−1 (β)| ≤ γ 1/s . γ k=1  PK The term k=1 PX |pk (X) − G−1 (β)| ≤ γ 1/s → 0 when γ → 0, given that the distribution function of the p′k s are continuous. Then using the convergence in distribution of ∆Rφ (fˆ) to zero, the last inequality ensures the desired result. 6.4 Proof of Proposition 5 For any β ∈ (0, K), and conditional on Dn we define e−1 (β) = inf{t ∈ R : G(t) e ≤ β}. G e We note that Assumption (A2) ensures that t 7→ G(t) is continuous and then e G e−1 (β)) = β. G( 19 (23) Now, we have e β (X)| = |Γ = K X k=1 K X k=1 Hence, the last equation implies that 1 {G e(fˆ (X))≤β} k 1{fˆ (X)≥G e−1 (β)} . k K   i X h e G e −1 (β)) = β. e−1 (β)|Dn = G( e β (X)||Dn = PX fˆk (X) ≥ G EX |Γ (24) k=1 h i e β (X)| = β. Also, we can write Therefore, we obtain E |Γ h i E Γ̂β (X) − β ≤ ≤ ≤ ≤ ≤ h i e β (X)| E |Γ̂β (X)| − |Γ "K #  X E 1{Ĝ(fˆk (X))≤β} − 1{G e(fˆk (X))≤β} k=1 h i e β (X)| E |Γ̂β (X) ∆ Γ K X k=1 K X k=1 h i E |1{Ĝ(fˆk (X))≤β} − 1{G | ˆ e(f (X))≤β} k   e fˆk (X))| ≥ |G( e fˆk (X)) − β| . P |Ĝ(fˆk (X) − G( Hence, applying Lemma 2 in the above inequality, we obtain the desired result. 6.5 Proof of Theorem 2 When there is no doubt, we suppress the dependence on X. First, let us state a intermediate result that is also needed to prove the theorem. Lemma 3. Consider Γfˆ,δ the confidence set based on the score function fˆ with information β (that is, Gfˆ(−δ) = β). Under assumptions Mαk , the following holds o n ∆R(Γfˆ,δ ) ≤ C(α, s) K 1−1/(s+λ−λs) ∆Rφ (fˆ)1/(s+λ−λs) + K 1−λ/(s+λ−λs) ∆Rφ (fˆ)λ/(s+λ−λs) , where λ = α α+1 and C(α, s) is non negative constant which depends only on α and s. Proof. For each k = 1, . . . , K, we define the following events Sk = Bkc ∩ {k ∈ (Γfˆ,δ ∆ Γ∗β )} and Tk = Bk ∩ {k ∈ (Γfˆ,δ ∆ Γ∗β )}, where the Bk ’s are the events given in Eq. (18). Now we observe that "K # "K # X X −1 −1 ∆R(Γfˆ,δ ) = E 1Sk |pk − G (β)| + E 1Tk |pk − G (β)| = ∆R1 (Γfˆ,δ ) + ∆R2 (Γfˆ,δ ), k=1 k=1 20 where ∆R1 (Γfˆ,δ ) ∆R2 (Γfˆ,δ ) = E = E " " K X k=1 K X # 1Sk |pk − G−1 (β)| , −1 k=1 1Tk |pk − G # (β)| . The end of the proof consists in controlling each of these two terms. Let us first consider ∆R1 (Γfˆ,δ ). For ε > 0, we have ∆R1 (Γfˆ,δ ) = E ≤ E " " K X k=1 K X −1 1Sk |pk − G 1Sk ε 1−s k=1 (β)| 1{|pk −G−1 (β)|≥ε} + 1{|pk −G−1 (β)|≤ε} −1 |pk − G ≤ Cε1−s ∆Rφ (fˆ) + ε K X s (β)| # +ε K X #  P(Sk ) k=1 P(Sk ), (25) k=1 where we used the assumption (6) and more precisely (20) to deduce the last inequality. To PK control k=1 P(Sk ), we require the following result that is a direct application of Lemma 5 in [BJM06]. Lemma 4. Under the assumptions Mαk we have K X k=1 α   α+1 , P (Sk ) ≤ C(α) K 1/α ∆R1 (Γfˆ,δ ) where C(α) > 0 is a constant that depends only on α. Proof. The proof of this result relies on the following simple fact: for all ε > 0     E 1Sk |pk − G−1 (β)| ≥ εE 1Sk 1{|pk −G−1 (β)|≥ε} Choosing ε =  1 c2 K(α+1) PK k=1 P (Sk )) 1/α ≥ ε [P (Sk ) − c2 εα ] . we get the lemma. We go back to the proof of Lemma 3. Applying Lemma 4 to (25), we get α   α+1 . ∆R1 (Γfˆ,δ ) ≤ Cε1−s ∆Rφ (fˆ) + εC(α) K 1/α ∆R1 (Γfˆ,δ ) Choosing ε = s−1 (λ−1) ∆R1 (Γfˆ,δ )(1−λ) , sC(α) K we obtain ∆R1 (Γfˆ,δ ) ≤ C1 (α, s)K 1−1/(s−λs+λ) ∆Rφ (fˆ)1/(s+λ−λs) , for a non negative constant C1 (α, s) that depends only on α and s. 21 (26) Let us now focus on the second term, ∆R2 (Γfˆ,δ ). Since the assumptions of Theorem 1 are satisfied, we can use Eq. (22) for any γ > 0. Combined with the Margin assumptions Mαk , we obtain i h ∆R2 (Γfˆ,δ ) ≤ EX |Gf ∗ (−δ ∗ ) − Gfˆ(−δ ∗ )| ≤ C∆Rφ (fˆ) + c2 Kγ α/s . γ Therefore, optimizing in γ, we have ∆R2 (Γfˆ,δ ) ≤ C2 (α, s)K 1−λ/(s+λ−λs) ∆Rφ (fˆ)λ/(s+λ−λs) , for a non negative constant C2 (α, s) that depends only on α and s. The result stated in the lemma is deduced by combining the last equation with Eq. (26) and by setting C(α, s) = max{C1 (α, s); C2 (α, s)}. We now state another important lemma that describes the behavior of the empirical minimizer of the φ-risk on the class F : Lemma 5. Let f¯ ∈ F be the minimizer of Rφ (f ) over F . Under the assumptions of Theorem 2, we have that h i 3KL KC(B, L) log(N ) n + , E Rφ (fˆn ) − Rφ (f¯) ≤ n n where C(B, L) > 0 is a constant that only depends on the constant B given in the Proposition 6 and on the Lipschitz constant L. Proof. First, according to Eq. (11), for each f ∈ F, we can write  v uK   uX ¯ ¯   f +f Rφ (f ) + Rφ (f ) ≥ δ t EX (fk − f¯k )2 (X)  , − Rφ 2 2 k=1 Hence, by assumption on the modulus of convexity, we deduce   K X   Rφ (f ) + Rφ (f¯) f + f¯ ≥ c1 EX (fk − f¯k )2 (X) . − Rφ 2 2 k=1 Since Rφ   ¯ f +f 2 ≥ Rφ (f¯), we obtain K X k=1   1 Rφ (f ) − Rφ (f¯). EX (fk − f¯k )2 (X) ≤ 2c1 Now, denoting by h(z, f (x)) = PK k=1 φ(zk fk (x)) − φ(zk f¯k (x)), we get the following bound   EX h2 (Zf (X)) ≤ KL2 ≤ K X k=1   EX (fk − f¯k )2 (X) KL2 EX [h(Zf (X))] , 2c1 22 (27) where L is the Lipschitz constant L for φ. On the other hand, we have the following decomposition Rφ (fˆ) − Rφ (f¯) = Rφ (fˆ) + 2(R̂φ (fˆ) − R̂φ (f¯)) − 2(R̂φ (fˆ) − R̂φ (f¯)) − Rφ (f¯). Also, since R̂φ (fˆ) − R̂φ (f¯) ≤ 0, we get Rφ (fˆ) − Rφ (f¯) ≤ (Rφ (fˆ) − Rφ (f¯)) − 2(R̂φ (fˆ) − R̂φ (f¯)) 3KL + sup (Rφ (f ) − Rφ (f¯)) − 2(R̂φ (f ) − R̂φ (f¯)), ≤ n f ∈Fn where Fn is the ǫ-net of F w.r.t the L∞ -norm and with ǫ = 1/n. Now, using Bernstein’s Inequality, we have that for all f ∈ Fn and t > 0   P (Rφ (f ) − Rφ (f¯)) − 2(R̂φ (f ) − R̂φ (f¯)) ≥ t ≤   P 2((Rφ (f ) − Rφ (f¯)) − (R̂φ (f ) − R̂φ (f¯)) ≥ t + Rφ (f ) − Rφ (f¯)   n(t + E [h(Z, f (X))])2 /8 ≤ exp − . E [h2 (Z, f (X))]) + (2KLB/3)(t + E [h(Z, f (X))])) Using Eq. (27), we get for all f ∈ Fn    P (Rφ (f ) − Rφ (f¯)) − 2(R̂φ (f ) − R̂φ (f¯)) ≥ t ≤ exp −  nt , 8(KL2 /(2c1 ) + KLB/3) Therefore, using a union bound argument, and then integrating we deduce that # " h i 3KL ˆ ¯ ¯ ¯ E Rφ (f ) − Rφ (f ) ≤ + E sup (Rφ (f ) − Rφ (f )) − 2(R̂φ (f ) − R̂φ (f )) n f ∈Fn ≤ 3KL KC(B, L) log(Mn ) + . n n We are now ready to conclude the proof of the theorem. We have the following inequality e β ) + |R(Γ̂β ) − R(Γ e β )|. |R(Γ̂β ) − R∗β | ≤ ∆R(Γ (28) We deal which each terms in the r.h.s separately. First, we have from Jensen’s Inequality that h i  h i s+λ−λs λ e β ) s+λ−λs eβ ) λ ≤ E ∆R(Γ . E ∆R(Γ Hence, from Lemma 3, we deduce  h i s+λ−λs h i s+λ−λs s+λ−λs λ eβ ) ≤ C(α, s) λ K λ −1 E ∆Rφ (fˆ) . E ∆R(Γ Moreover, from Lemma 5, we have that h i 3KL KC(B, L) log(Mn ) + . E ∆Rφ (fˆ) ≤ inf ∆Rφ (f ) + f ∈F n n 23 Therefore, we can write λ/(s+λ−λs)  h i e β ) ≤ C(α, s)K 1−λ/(s+λ−λs) inf ∆Rφ (f ) + 3KL + KC(B, L) log(Nn ) . E ∆R(Γ f ∈F n n (29) e β )| in (28), we observe that For the second term |R(Γ̂β ) − R(Γ 1{Y ∈/ Γ̂β (X)} − 1{Y ∈/e Γ β (X)} = K X k=1 1{Y =k} 1{k∈/ Γ̂β (X)} − K X k=1 Therefore, we can write i h = E 1Y ∈/ Γ̂β (X)} − 1{Y ∈/e Γβ (X)} = K X k=1 K X k=1 1{Y =k} 1k∈/e Γ β (X) . i h  E pk (X) 1{k∈/ Γ̂β (X)} − 1{k∈/e Γβ (X)} i h  E pk (X) 1{Ĝ(fˆk (X))>β } − 1{G̃(fˆk (X))>β} . Since 0 ≤ pk (X) ≤ 1 for all k ∈ {1, . . . , K}, the last equality implies h    R Γ̂β − R Γ̃β = E 1Y ∈/ Γ̂β (X)} − 1{Y ∈/e Γ β (X)} ≤ K X k=1 i h i E 1{Ĝ(fˆk (X))>β } − 1{G̃(fˆk (X))>β} ≤ K X k=1 Therefore, Lemma 2 implies   e fˆk (X))| ≥ |G( e fˆk (X)) − β| . P |Ĝ(fˆk (X) − G(    C ′K R Γ̂β − R Γ̃β ≤ √ . N Injecting Eqs. (29) and (30) to Eq. (28) we conclude the proof of the theorem. 6.6 (30) Proof of Proposition 6 We begin with the following decomposition R̃φn (fˆ) − R̃φn (f˜) = R̃φn (fˆ) + 2(R̂φn (fˆ) − R̂φn (f˜)) − 2(R̂φn (fˆ) − R̂φn (f˜)) − R̃φn (f˜), since R̂φn (fˆ) − R̂φn (f˜) ≤ 0, we get R̃φn (fˆ) − R̃φn (f˜) ≤ (R̃φn (fˆ) − R̃φn (f˜)) − 2(R̂φn (fˆ) − R̂φn (f˜)). (31) Now, we denote by Cn = {(i1 /n, . . . , iM /n), (i1 , . . . , iM ) ∈ {0, . . . , n}} ∩ CM , PM and Fn = {f = i=1 mi fi , m1 , . . . mM ) ∈ Cn }. For each f ∈ conv(F ), there exists fn ∈ Fn such that |R̃φn (f ) − R̃φn (fn )| ≤ |R̂φn (f ) − R̂φn (fn )| ≤ 24 2KLBM n 2KLBM . n Therefore, with Equation (31), we obtain 6LKB R̃φn (fˆ) − R̃φn (f˜) ≤ + sup (R̃φn (f ) − R̃φn (f¯)) − 2(R̂φn (f ) − R̂φn (f¯)). n f ∈Fn Now, Similar arguments as in [DvdL05] and Lemma 5 yield the proposition. References [BJM06] P. Bartlett, M. Jordan, and J. McAuliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138–156, 2006. [BM06] P. Bartlett and S. Mendelson. Empirical minimization. Probability Theory and Related Fields, 135(3):311–334, 2006. [CCB16] A. Choromanska, K. Choromanski, and M. Bojarski. On the boosting ability of topdown decision tree learning algorithm for multiclass classification. preprint, 2016. [Cho70] C.K. Chow. On optimum error and reject trade-off. IEEE Transactions on Information Theory, 16:41–46, 1970. [dCDB09] J. del Coz, J. Díez, and A. Bahamonde. Learning nondeterministic classifiers. Journal of Machine Learning Research, 10:2273–2293, 2009. [DH15] C. Denis and M. Hebiri. Consistency of plug-in confidence sets for classification in semi-supervised learning. preprint, 2015. [DvdL05] S. Dudoit and M. van der Laan. Asymptotics of cross-validated risk estimation in estimator selection and performance assessment. Statistical Methodology, 2(2):131– 154, 2005. [FHT00] J. Friedman, T. Hastie, and R. Tibshirani. Additive logistic regression: a statistical view of boosting. The Annals of Statistics, 28(2):337–407, 2000. [FS97] Y. Freund and R. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55(1):119–139, 1997. [HW06] R. Herbei and M. Wegkamp. Classification with reject option. The Canadian Journal of Statistics, 34(4):709–721, 2006. [Lei14] J. Lei. Classification with confidence. Biometrika, 101(4):755–769, 2014. [LRW13] J. Lei, J. Robins, and L. Wasserman. Distribution-free prediction sets. Journal of the American Statistical Association, 108(501):278–287, 2013. [TB07] A. Tewari and P. Bartlett. On the consistency of multiclass classification methods. Journal of Machine Learning Research, 8:1007–1025, 2007. [Vap98] V. Vapnik. Statistical learning theory. Adaptive and Learning Systems for Signal Processing, Communications, and Control. John Wiley & Sons Inc., New York, 1998. A Wiley-Interscience Publication. 25 [vdLPH07] M. van der Laan, E. Polley, and A. Hubbard. Super learner. Statistical Applications in Genetics and Molecular Biology, 6, 2007. [vdV98] A. van der Vaart. Asymptotic statistics, volume 3 of Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 1998. [VGS99] V. Vovk, A. Gammerman, and C. Saunders. Machine-learning applications of algorithmic randomness. In International Conference on Machine Learning, pages 444–453. 1999. [VGS05] V. Vovk, A. Gammerman, and G. Shafer. Algorithmic learning in a random world. Springer, New York, 2005. [Vov02] V. Vovk. Asymptotic optimality of transductive confidence machine. In Algorithmic learning theory, volume 2533 of Lecture Notes in Computer Science, pages 336–350. Springer, Berlin, 2002. [WLW04] T.-F. Wu, C.-J. Lin, and R. Weng. Probability estimates for multi-class classification by pairwise coupling. Journal of Machine Learning Research, 5:975–1005, 2004. [WY11] M. Wegkamp and M. Yuan. Support vector machines with a reject option. Bernoulli, 17(4):1368–1385, 2011. [YW10] M. Yuan and M. Wegkamp. Classification methods with reject option based on convex risk minimization. Journal of Machine Learning Research, 11:111–130, 2010. [Zha04] T. Zhang. Statistical behavior and consistency of classification methods based on convex risk minimization. The Annals of Statistics, 32(1):56–85, 02 2004. 26
10math.ST
0/1/all CSPs, Half-Integral A-path Packing, and Linear-Time FPT Algorithms arXiv:1704.02700v2 [cs.DS] 7 Nov 2017 Yoichi Iwata∗ National Institute of Informatics [email protected] Yutaro Yamaguchi† Osaka University yutaro [email protected] Yuichi Yoshida‡ National Institute of Informatics [email protected] Abstract A recent trend in the design of FPT algorithms is exploiting the half-integrality of LP relaxations. In other words, starting with a half-integral optimal solution to an LP relaxation, we assign integral values to variables one-by-one by branch and bound. This technique is general and the resulting time complexity has a low dependency on the parameter. However, the time complexity often becomes a large polynomial in the input size because we need to compute half-integral optimal LP solutions. In this paper, we address this issue by providing an O(km)-time algorithm for solving the LPs arising from various FPT problems, where k is the optimal value and m is the number of edges/constraints. Our algorithm is based on interesting connections among 0/1/all constraints, which has been studied in the field of constraints satisfaction, A-path packing, which has been studied in the field of combinatorial optimization, and the LPs used in FPT algorithms. With the aid of this algorithm, we obtain improved FPT algorithms for various problems, including Group Feedback Vertex Set, Subset Feedback Vertex Set, Node Multiway Cut, Node Unique Label Cover, and Non-monochromatic Cycle Transversal. The obtained running time for each of these problems is linear in the input size and has the current smallest dependency on the parameter. In particular, these algorithms are the first linear-time FPT algorithms for problems including Group Feedback Vertex Set and Nonmonochromatic Cycle Transversal. ∗ Supported by JSPS KAKENHI Grant Number JP17K12643. Supported by JSPS KAKENHI Grant Number JP16H06931 and JST ACT-I Grant Number JPMJPR16UR. ‡ Supported by JST ERATO Grant Number JPMJER1305 and JSPS KAKENHI Grant Number JP17H04676. † Contents 1 Introduction 1.1 FPT Algorithms using Half-Integral LP Relaxations 1.2 0/1/all Deletion and A-path Packing . . . . . . . 1.3 Related Work on Half-Integral A-path Packing . . . 1.4 Proof Sketch . . . . . . . . . . . . . . . . . . . . . . 1.5 Comparison to Babenko’s algorithm . . . . . . . . . 1.6 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 3 4 5 7 7 2 Definitions 8 2.1 Basic Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2 0/1/all Deletion and Half-Integral Relaxation . . . . . . . . . . . . . . . . . . . . 8 2.3 Single-Branching Pair and Incremental-Test Oracle . . . . . . . . . . . . . . . . . . . 10 3 Half-Integral Packing and Covering 3.1 Preliminaries . . . . . . . . . . . . . . . . . 3.1.1 Basic F-Packing . . . . . . . . . . . 3.1.2 Augmenting Path/Pair . . . . . . . . 3.2 Finding Augmenting Path/Pair . . . . . . . 3.3 Constructing Half-Integral F-Cover . . . . . 3.4 Augmentation . . . . . . . . . . . . . . . . . 3.4.1 Simplification of Alternating Path . 3.4.2 Augmentation by Augmenting Path 3.4.3 Augmentation by Augmenting Pair . 4 Farthest Cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 11 11 12 15 19 21 21 23 25 31 5 Linear-Time FPT Algorithms 35 5.1 Algorithm for 0/1/all Deletion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 5.2 Applications to Other Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 A Proof of Persistency 46 B Axiomatic Model 47 1 Introduction 1.1 FPT Algorithms using Half-Integral LP Relaxations Parameterized complexity is the subject of studying the complexity of parameterized problems. A parameterized problem with a parameter k is fixed parameter tractable (FPT) if we can solve the problem in f (k)poly(n) time, where n is the input size. Various parameterized problems are known to be FPT. See [8, 10] and references therein for a comprehensive list of FPT problems. One of the motivations of studying parameterized complexity is to understand tractable subclasses of (NP-)hard problems; hence, the primary interest has been which parameterized problems admit FPT algorithms. However, from a practical point of view, the running time with respect to the input size must also be small. Indeed, linear-time FPT algorithms (i.e., FPT algorithms whose running times are linear in the input size) are proposed for several problems including Treewidth [4], Almost 2-SAT [18, 31], Feedback Vertex Set (FVS) [3], Subset FVS [21], Directed FVS [22], and Node Unique Label Cover [23]. These works focused on reducing the running time with respect to the input size, and hence dependency on the parameter is often suboptimal. For example, Subset FVS admits an FPT algorithm running in O∗ (4k )1 time [19] whereas the current best linear-time FPT algorithm has time complexity O(25.6k m) [21], where m is the number of edges in the input graph. The half-integrality of the LP relaxations has recently been used to design FPT algorithms for a broad range of problems [9, 15, 17, 19, 20, 33]. To see the idea, let us consider a minimization problem whose goal is to find a solution of size k, and suppose that it admits a half-integral LP relaxation2 , that is, an LP relaxation with an optimal LP solution that only uses values in {0, 21 , 1}. The algorithm is based on the standard branch-and-bound framework. First, we compute a halfintegral (optimal) LP solution. We can stop if all the variables have integral values or the sum of the values in the LP solution exceeds k. Otherwise, we fix variables with values 1. We then pick an arbitrary variable with value 21 , and branch into the case that its value is fixed to 0 and the case that its value is fixed to 1. This approach has several big advantages. It can be applicable to various problems just by changing the LP, and has a small time complexity with respect to parameter k. For many problems including Almost 2-SAT [20], Node Multiway Cut [9], and Group FVS [19], the current smallest dependency on the parameter is indeed achieved by this approach. A drawback of the abovementioned approach is that it is not trivial how to efficiently compute half-integral LP solutions. Iwata, Wahlström, and Yoshida [19] viewed half-integrality as a discrete relaxation. They showed that one can compute half-integral LP solutions for Almost 2-SAT and (Edge) Unique Label Cover in time linear in the input size by reducing them to the s-t cut problem. Moreover, they showed that we can compute LP solutions with an extremal condition, which we call the farthest condition herein (see Section 2.2 for details), in the same running time. Using farthest solutions, the LP lower bound strictly increases for each branching. Consequently, they obtained linear-time FPT algorithms for these problems. O∗ (·) hides a polynomial dependency on the input size. When focusing on reducing the f (k) part, the poly(n) part is often ignored using this notation. 2 Most of the LPs used in the FPT algorithms are not natural LP relaxations of the original problems, but are LP relaxations of rooted problems. For example, the rooted version of FVS is a problem of finding a minimum vertex subset S, such that the graph obtained by removing S contains no cycles reachable from a prescribed vertex s. Note that the existence of a half-integral LP relaxation to the rooted problem does not imply a 2-approximability of the original problem. 1 1 Table 1: Summary of our FPT results. Here, d denotes the maximum domain size, Σ is the alphabet set; m denotes the number of edges/constraints in the input; and TΓ denotes the time complexity for performing group operations. All the algorithms are deterministic except for the O(25.6k m)-time algorithm for Subset FVS. See Section 5.2 for the problem definitions. Problem Smallest f (k) Existing linear-time FPT Our result ∗ 2k 0/1/all Deletion O (d ) [19] — O(d2k km) ∗ 2k O(k|Σ|) Node Unique Label Cover O (|Σ| ) [19] |Σ| m [23] O(|Σ|2k km) Two-fan Deletion O∗ (9k ) [19] — O(4k km) Monochromatically — — O(4k km) Orientable Deletion Subset Pseudoforest — — O(4k km) Deletion Node Multiway Cut O∗ (2k ) [9] O(4k m) [6, 18] O(2k km) Group FVS O∗ (4k TΓ ) [19] — O(4k kmTΓ ) k O(25.6 m) (randomized) [21] Subset FVS O∗ (4k ) [19] O(4k km) 2O(k log k) m (deterministic) [21] Non-monochromatic — O∗ (4k ) [33] O(4k km) Cycle Transversal The s-t cut approach unfortunately does not work well for other problems such as Group FVS, Subset FVS, Node Multiway Cut, and Node Unique Label Cover because it is essentially applicable only to edge-deletion problems3 and because the auxiliary network size is generally not linear in the input size (e.g., for Subset Feedback Edge Set, the size of the network becomes 2O(m) ). We need to resort to solving linear programs for these problems, and hence the resulting FPT algorithms have large dependencies on the input size. Among these problems, linear-time FPT algorithms have been obtained for Node Multiway Cut [6, 18], Subset FVS [21], and Node Unique Label Cover [23] by problem-specific arguments without using LP relaxations. However, they have larger dependencies on k (Table 1). The main contribution of this study is the development of an algorithm that computes farthest half-integral solutions to those LPs in time linear in the input size. Using the connection of computing half-integral LP solutions and FPT algorithms, we obtain linear-time FPT algorithms for various problems, which are summarized in Table 1. In particular, for Subset FVS, Node Multiway Cut, and Node Unique Label Cover, we substantially improve the dependency on the parameter. For the other problems including Group FVS and Non-monochromatic Cycle Transversal, we obtain the first linear-time FPT algorithms. We note that, for every problem in the table, the f (k) part in the running time of our algorithm matches or improves the smallest known. All of these results are obtained by the same approach, i.e., the branch-and-bound framework combined with the efficient computation of half-integral LPs, which demonstrates its generality. 3 Note that edge-deletion problems are easily reducible to the corresponding vertex-deletion problems in most cases by subdividing the edges and creating k copies of the original vertices. 2 1.2 0/1/all Deletion and A-path Packing We consider the following problem, called 0/1/all Deletion, to establish a unified framework and provide linear-time FPT algorithms for the abovementioned problems using half-integral LP relaxations. Let V be a set of variables. Each variable v ∈ V has an individual domain D(v). A function ϕ on V that maps each variable v ∈ V to a value ϕ(v) ∈ D(v) is called an assignment for V . We consider the following two types of binary constraints on (u, v) ∈ V × V , called 0/1/all constraints [7]4 . 1. Permutation constraint π(ϕ(u)) = ϕ(v), where π : D(u) → D(v) is a bijection. 2. Two-fan constraint (ϕ(u) = a) ∨ (ϕ(v) = b), where a ∈ D(u) and b ∈ D(v). Let C be a set of 0/1/all constraints on V . We assume that C contains at most one constraint for each pair of distinct variables u, v ∈ V . A constraint on (u, v) in C is denoted by Cuv . For a subset U ⊆ V , we denote the set {Cuv ∈ C | u, v ∈ U } by C[U ]. 0/1/all Deletion Parameter: k, d Input: A set of variables V each of which has a domain of size at most d, a set C of 0/1/all constraints each of which is given as a table of size O(d), a partial assignment ϕA for a subset A ⊆ V , and an integer k. Question: Is there a pair of a set X ⊆ V of at most k variables and a partial assignment ϕ for V \ X such that (1) ϕ(v) = ϕA (v) holds for every v ∈ A \ X and (2) ϕ satisfies all of C[V \ X]? The set X in the question is called a deletion set. The task of the optimization version of this problem is to compute the size of a minimum deletion set. Various FPT problems can be expressed as 0/1/all Deletion (by using a large d). Note that for several problems, we need exponentialsize domains, and hence a linear-time FPT algorithm for 0/1/all Deletion does not directly imply linear-time FPT algorithms for such problems. We show that, by giving constraints not as a table but as an oracle and by using a specialized branching strategy, we can obtain linear-time FPT algorithms even for such problems in a unified way. See Section 5.2 for details. The primal graph for C is a simple undirected graph G = (V, E) such that an edge uv ∈ E exists if and only if a constraint Cuv on (u, v) exists. An important property of the 0/1/all constraints is that, when fixing the value of a variable u ∈ V to p ∈ D(u), the set of values of v ∈ V satisfying the constraint Cuv is either D(v) (when the constraint is a two-fan with a = p) or a singleton {q} (when the constraint is a permutation with q = π(p) or a two-fan with a 6= p and b = q). The latter-type implication is called a unit propagation. When we are given a partial assignment ϕA for a subset A ⊆ V , unit propagations occur along walks in the primal graph G starting at the vertices in A. If the unit propagations along two different walks starting at (possibly the same) vertices in A lead to a contradiction (i.e., implicate distinct singletons for the same variable), then at least one variable on the two walks must be contained in the deletion set. The concatenation of such two walks is a walk between the vertices in A (called an A-walk) that is said to be conflicting (see Section 2.2 for a formal definition). This observation provides a lower bound on the minimum size of a deletion set as follows. Suppose that a deletion set X ⊆ V with |X| = k exists, and let F = FC,ϕA be the set of all 4 Precisely speaking, the permutation and two-fan constraints together with empty and complete constraints are obtained by enforcing arc-consistency on the 0/1/all constraints introduced in [7]. 3 conflicting A-walks in the primal graph G for C with respect to ϕA . Then, the remaining graph G − X cannot have a walk in F (i.e., X is a cover (or a hitting set) of F). Hence, the minimum size of such a cover is at most k. We now consider an LP relaxation of finding a minimum cover of the set F of all conflicting A-walks in G, called the F-covering a function x : V → R≥0 P problem: we are required to find P minimizing the total value |x| := x(v) under the constraint that v∈V v∈W x(v) ≥ 1 for every P W ∈ F, where “ v∈W ” means the summation over the occurrences of vertices v in W considering the multiplicity (e.g., x(v) is summed twice if W intersects v twice). The dual of this LP is written down as follows, as the F-packing P problem: we are required P to find a function y : F → R≥0 maximizing the total value |y| := W ∈F y(W ) subject to W ∈F : v∈W y(W ) ≤ 1 for every v ∈ V , where P we also consider the multiplicity of the occurrences of vertices in a walk in the summation “ W ∈F : v∈W ” (e.g., y(W ) is summed twice if W intersects v twice). We then propose an O(kmT )-time algorithm for finding a pair of a half-integral F-cover x and a half-integral F-packing y with |x| = |y| ≤ k2 (if exists, and otherwise correctly concluding it), where m = |C| and T is the running time of a certain oracle for simulating unit propagations (see Section 2.3 for the detail). Note that, by LP duality, these x and y are both optimal solutions to the F-covering and F-packing problems, respectively. Combining our algorithm with the result in [19] (see also Theorem 1 in Section 2.2), we can obtain an FPT algorithm for 0/1/all Deletion and, hence, FPT algorithms for other various problems. Section 5 provides the detailed discussion for each specific problem. 1.3 Related Work on Half-Integral A-path Packing When C contains only the permutation constraints, the integral version of the F-packing problem has been studied under the name of non-returning A-path packing [26, 28, 34]. This is the current most general case of tractable (integral) path packing problems. Our half-integral result suggests a conjecture that the integral packing of conflicting paths will also be tractable. We also believe that ideas behind our half-integral algorithm will be useful for obtaining a faster algorithm for integral path packing problems. For a further special case (of non-returning A-path packing), called internally disjoint5 A-path packing, the integral version of the dual covering problem coincides with Node Multiway Cut, and its LP relaxation is used in the branch-and-bound FPT algorithm [9] and a 2-approximation algorithm [14] for this problem. Several previous works in the field of combinatorial optimization can be found as regards the half-integral version of internally disjoint A-path packing. Garg et al. [14] and Pap [27, 29] found that both LPs always enjoy half-integral optimal solutions even if each non-terminal vertex has an individual integral capacity instead of 1. Hirai [16] and Pap [27, 29] developed algorithms for finding such half-integral optimal solutions, which both run in strongly polynomial time (i.e., the number of elementary operations performed through each algorithm does not depend on the capacity values). One makes use of a sophisticated algorithm for the submodular flow problem [13] whereas the other relies on the ellipsoid method to solve LPs whose coefficient matrices only have 0, ±1 entries [12]. Specializing on the uncapacitated case, Babenko [1] provided an O(knm)-time algorithm for finding a maximum half-integral packing, where n and m are the numbers of vertices and edges, respectively, and k denotes the optimal value, which is at most O(n). Our algorithm 5 They can share terminals, but not inner vertex or edge. 4 (1) connecting distinct vertices (2) path to an odd cycle (3) prefix-sharing pair Figure 1: Augmentations for half-integral matching. Thick solid lines denote edges of weight 1, dashed lines denote edges of weight 12 , and thin solid lines denote edges of weight 0 for the F-packing/covering problems improves the previous best running time even against this (internally disjoint) special case. 1.4 Proof Sketch Basically, we iteratively augment a half-integral F-packing y, and construct a half-integral F-cover x of the same size when no augmentation is possible. Since the F-covering and F-packing problems are the dual LPs to each other, this implies that x and y are optimal solutions to these problems. In the case of the maximum s-t flow, when we failed to find an augmenting path, we can construct a minimum s-t cut by taking the set of edges on the flow that separates the vertices visited by the failed search from the unvisited vertices. Similarly, we can construct a minimum half-integral F-cover by taking the set of vertices on the half-integral F-packing that separates visited vertices from the unvisited vertices. We describe the idea behind our algorithm for F-packing by showing a relation to computing half-integral (non-bipartite) matchings, which are often called 2-matchings in the field of combinatorial optimization (see [32, Chapter 30] for the basics). Although we can easily obtain a maximum half-integral matching by a reduction to maximum bipartite matching [24], we propose a different approach herein. The idea behind this approach can be used for the half-integral F-packing. We focus on special half-integral matchings that consist of vertex-disjoint edges of weight 1 and odd cycles6 with each edge having weight 12 . A maximum half-integral matching with this special structure is known to always exist [2]. In each step, we search for an alternating path (a path that alternately uses edges of weight 0 and 1 and never uses the edges of weight 21 ) from the vertices not used in the current matching. Augmentation can be categorized into three types (see Figure 1). The first case is when we found an alternating path P connecting two distinct vertices not used in the current matching. The current matching uses all the even edges in P . We can augment the matching size by 1 by taking all the odd edges and discarding all the even edges. The second case is when we found an alternating path P of length 2a + 1 ending at a vertex on an odd cycle C of length 2b + 1. The current matching uses all the even edges in P with weight 1 and all the edges in C with weight 12 . Thus, the size of the current matching induced by P and 6 The number of edges in the cycle is odd. 5 C is a + b + 12 . We can easily augment this matching to an integral matching of size a + b + 1 by alternately taking the edges in P and C. The third case is when we found a pair of prefix-sharing alternating paths P of odd length and Q of even length ending at the same vertex. This case corresponds to a blossom in Edmonds’ algorithm for the maximum (integral) matching problem [11]. In Edmonds’ algorithm, the alternation is applied to the common prefix and then the vertices on the cycles induced by P and Q are shrunk. In our approach, we can augment the matching size by 12 by applying the alternation to the common prefix, then by transforming the cycle induced by P and Q to an odd cycle with each edge having weight 21 . We use a similar approach in our algorithm for F-packing/covering. We focus on a special type of half-integral F-packings that consist of disjoint paths in F of weight 1, called integral paths, and wheels, which are the sums of an odd number of walks in F of weight 12 and correspond to odd cycles for matching. See Section 3.1.1 for the precise definition. Although a maximum half-integral packing with a similar special structure for a very special case of F-packing, internally disjoint Apath packing, is known to always exist [25], the existence of such a special solution was previously unknown for any other cases. The correctness of our algorithm provides a constructive proof of the existence. In each step, we search for an alternating path from the vertices in A not used in the current packing. As opposed to the case of computing half-integral matchings, the alternating paths may use edges contained in wheels. However, for simplicity, we ignore such a case in the explanation that follows. Roughly speaking, an alternating path is a sequence of paths P1 , . . . , P` , where Pi is a path internally disjoint from any integral path and wheel for odd i and is fully contained in an integral path for even i. In the alternation operation, we replace the integral path containing P2 with the path obtained by concatenating P1 and the prefix of the integral path containing P2 . We then replace the integral path containing P4 with . . . , and so on. The definition of alternating paths is rather complicated in ensuring that each introduced path is in F. However, it essentially plays the same role as that for the half-integral matching case. Augmentation can be categorized into three types. Each of which corresponds to the one for the half-integral matching case. When we find an alternating path connecting two distinct vertices in A, we can augment the F-packing by 1 by applying the alternation. When we find an alternating path ending at a vertex on a wheel that is the sum of 2a + 1 walks in F of weight 21 , we can augment the F-packing by 12 by decomposing the wheel into a integral paths and introducing a new integral path. We call these two types of alternating path as an augmenting path. When we find a pair of prefix-sharing alternating paths P and Q ending at the same vertex (with some additional conditions), we can augment the F-packing by 12 by applying the alternation to the common prefix and then by introducing a new wheel. We call this type of a pair of alternating paths as an augmenting pair. We need to test the membership in F in constant time to achieve the linear time complexity. To this end, we exploit an observation that the algorithm only tests membership against some special walks. To obtain linear-time FPT algorithms, we need to compute a farthest minimum half-integral F-cover, which is an F-cover satisfying some extremal condition, in linear time. In the existing work using the s-t cut approach [19], this is achieved by exploiting a structural property of all the minimum s-t cuts [30]. We use a different approach because we do not have the corresponding structural property for all the minimum half-integral F-covers. Naively, we can find a farthest 6 minimum half-integral F-cover by at most n computations of the minimum half-integral F-cover. We show that we can regard the whole sequence of computations as at most 2k computations of augmenting path/pair search by properly deciding the order of these computations using the maximum half-integral F-packing. Thus, it runs in linear time in total. Another obstacle for obtaining linear-time FPT algorithms is the existence of two-fan constraints. Because 2-SAT can be expressed as 0/1/all Deletion with k = 0 and d = 2, any linear-time FPT algorithm for 0/1/all Deletion must be able to solve 2-SAT in linear time. The standard linear-time algorithm for 2-SAT uses the strongly connected component decomposition of the implication graph, and the existing linear-time FPT algorithms for Almost 2-SAT, a parameterized version of 2-SAT, also use the strongly connected component decomposition of an auxiliary network [18, 19, 31]. We cannot use this approach because the size of the auxiliary network becomes super-linear for our problems. In our algorithm, we do not use the strongly connected component decomposition, but instead use a parallel unit-propagation, which is an alternative linear-time algorithm for 2-SAT. 1.5 Comparison to Babenko’s algorithm Because our algorithm for the F-packing/covering problems improves the previous best running time even against the internally-disjoint special case, we compare our algorithm with Babenko’s algorithm to clarify the reason that we obtain such an improvement. While both the algorithms iteratively augment a packing, the approaches are completely different. The main difference is the existence of augmenting pairs and the algorithm for computing augmenting paths/pairs. In our algorithm, we focus on packings with a special structure. Both the definition of alternating paths and the algorithm for searching augmenting paths/pairs strongly rely on this structure. While the existence of a maximum half-integral packing with the special structure was already known for internally disjoint A-path packing [25], Babenko’s algorithm does not directly exploit the structure, but uses a much weaker structure. This is because his augmentation strategy does not preserve the special structure because it does not consider a notion corresponding to augmenting pairs of our algorithm. In our algorithm, we directly compute an augmenting path/pair in O(m) time and hence we can compute a maximum packing of size k in O(km) time. In contrast, in Babenko’s algorithm, an auxiliary network and its s-t flow f are constructed from the current packing, then an f -augmenting path for the standard maximum flow problem is computed. From the obtained f -augmenting path, we can either augment the current packing or we can find a set of vertices that can be safely contracted to some vertex in A. Because contractions occur at most O(n) time per augmentation, the running time becomes O(knm). 1.6 Organization We introduce the notions used throughout the paper in Section 2. Section 3 shows a fast algorithm that computes a maximum half-integral F-packing and transforms it into a minimum half-integral F-cover. Section 4 presents a fast algorithm for computing a farthest minimum half-integral Fcover. We provide linear-time FPT algorithms in Section 5 using this algorithm. 7 2 2.1 Definitions Basic Notations The multiplicity function 1S : U → Z≥0 for a multiset S on the ground set U is defined such that 1S (a) is the number of times that a ∈ U appears in S. For two multisets A and B on the same ground set U , we denote by A \ B the multiset such that 1A\B (a) = max{1A (a) − 1B (a), 0} holds for any element a ∈ U . For a function f : U → R and a multiset S on the ground set U , we define P f (S) := a∈U 1S (a)f (a). For a value i ∈ R, we define f −1 (i) := {a ∈ U | f (a) = i}. All the graphs in this study are undirected. However, we sometimes need to take care of the direction of edges. For an undirected graph G = (V, E), we use the symbol Ê when we take care of the direction of the edges, i.e., uv = vu for uv ∈ E but uv 6= vu for uv ∈ Ê. For simplicity, we assume that the graphs are simple; if a graph contains multiple edges or self-loops, we can easily obtain an equivalent simple graph by subdividing the edges. For vertex v ∈ V , we denote the set of incident edges by δ(v). For a subset U ⊆ V , we denote the induced subgraph by G[U ] = (U, E[U ]). For an undirected graph G = (V, E), we define a walk in G as an ordered list W = (v0 , . . . , v` ) of vertices such that vi−1 vi ∈ E for all i = 1, . . . , `. The integer ` is called the length of the walk. We denote the first and last vertices of W by s(W ) = v0 and by t(W ) = v` , respectively, and we say that W starts from s(W ) and ends at t(W ). We denote the multisets of vertices, inner vertices, and (undirected) edges appeared in W by V (W ) = {v0 , . . . , v` }, by Vin (W ) = {v1 , . . . , v`−1 } = V (W ) \ {s(W ), t(W )}, and by E(W ) = {v0 v1 , . . . , v`−1 v` }, respectively. For an edge e = uv ∈ Ê, we simply use the same symbol e to denote the walk (u, v). A walk W is called a (simple) path if 1V (W ) (v) ≤ 1 for every v ∈ V . A walk W is called a closed walk if s(W ) = t(W ). A closed walk W is called a (simple) cycle if ` ≥ 3 and 1V (W )\{s(W )} (v) ≤ 1 for every v ∈ V (which implies 1V (W ) (s(W )) = 2). We may regard a walk W as a subgraph by ignoring the direction and the multiplicity. We say that a walk W is internally disjoint from a subgraph G0 = (V 0 , E 0 ) if none of the inner vertices of W are in V 0 and none of the edges of W are in E 0 . For a walk W = (v0 , . . . , v` ), we define the reversed walk as W −1 = (v` , . . . , v0 ). For a walk W1 = (v0 , . . . , v`0 ) and a walk W2 = (v`0 , . . . , v` ) (where 0 ≤ `0 ≤ `), we define the concatenation of the two walks as W1 ◦ W2 = (v0 , . . . , v` ). The notation W1 ◦ W2 implicitly implies t(W1 ) = s(W2 ). 2.2 0/1/all Deletion and Half-Integral Relaxation We first recall 0/1/all Deletion defined in Section 1.2. We are given a set V of variables v with individual domains D(v), a set C of 0/1/all constraints (permutation and two-fan constraints) that can be represented by a simple undirected graph G, called the primal graph, and a partial assignment ϕA for a subset A ⊆ V . The task is to determine whether a deletion set X ⊆ V of at most k variables exists, for which there exists an assignment ϕ for V \X such that (1) ϕ(v) = ϕA (v) holds for every v ∈ A \ X, and (2) ϕ satisfies every constraint Cuv ∈ C[V \ X]. As described in Section 1.2, an important property of the 0/1/all constraints is that, when fixing the value of a variable u ∈ V to p ∈ D(u), the set of values of v ∈ V satisfying the constraint Cuv is either D(v) or a singleton {q}. We define Cuv (p) := all in the former case and define Cuv (p) := q in the latter case. We extend this definition to walks in the primal graph as follows. For a walk (s) of length zero and an element p ∈ D(s), we define C(s) (p) := p. For a walk W = W 0 ◦ e starting from s ∈ V and an element p ∈ D(s), we define CW (p) := all when CW 0 (p) = all and CW (p) := Ce (CW 0 (p)) when CW 0 (p) 6= all. Suppose that an assignment ϕ satisfying all the 8 constraints exists. Then, either CW (ϕ(s(W ))) = all or ϕ(t(W )) = CW (ϕ(s(W ))) holds for any walk W in the primal graph. Let ϕA be a partial assignment for a subset A ⊆ V . For a walk W with s(W ) ∈ A, we define impϕA (W ) := CW (ϕA (s(W ))), which represents the set of assignments for t(W ) induced by W . A walk W is called a ϕA -implicational walk if s(W ) ∈ A and impϕA (W ) 6= all. A ϕA -implicational walk W is called ϕA -conflicting if t(W ) ∈ A and impϕA (W ) 6= ϕA (t(W )) hold. We omit the prefix/subscript ϕA if it is clear from the context. We use the following lemma. Lemma 1. For any two ϕA -implicational walks P and Q ending at the same vertex, P ◦ Q−1 is ϕA -conflicting if and only if impϕA (P ) 6= impϕA (Q). Proof. We prove the lemma by induction on the length of Q. When the length of Q is zero, P ◦ Q−1 = P is conflicting if and only if imp(P ) 6= ϕA (t(P )) = imp(Q). When Q = Q0 ◦ vu, we consider three cases. (Case 1) If P ◦ uv is not implicational, Cuv must be a two-fan (ϕ(u) = a) ∨ (ϕ(v) = b) and imp(P ) = a. Because Q is implicational, we have imp(Q) = a. Therefore, we have imp(P ) = imp(Q). Because any prefix of a conflicting walk is implicational, P ◦ Q−1 is not conflicting. Thus, the lemma holds for this case. (Case 2) If P ◦ uv is implicational and Cuv is a permutation π(ϕ(u)) = ϕ(v), we have imp(P ◦ uv) = π(imp(P )) and imp(Q0 ) = π(imp(Q)). Therefore, imp(P ◦ uv) = imp(Q0 ) if and only if imp(P ) = imp(Q) holds. Thus, from the induction hypothesis, the lemma holds for this case. (Case 3) If P ◦ uv is implicational and Cuv is a two-fan (ϕ(u) = a) ∨ (ϕ(v) = b), we have imp(P ) 6= a, imp(Q) = a, imp(P ◦ uv) = b, and imp(Q0 ) 6= b. Therefore, from the induction hypothesis, P ◦ Q−1 is conflicting. Thus, the lemma holds for this case. For two walks P and Q ending at the same vertex, we write P 6≡ Q if P ◦ Q−1 is conflicting and P ≡ Q if P ◦ Q−1 is not conflicting, but both P and Q are implicational7 . The abovementioned lemma implies that (≡) is an equivalence relation. This is a key property in our algorithms. Appendix B presents an alternative axiomatic definition of implicational/conflicting walks and show that any path systems admitting this key property can be expressed as the set of implicational/conflicting walks with 0/1/all constraints. We obtain the following corollaries from Lemma 1. Corollary 1. If a walk W is conflicting, then W −1 is also conflicting. Corollary 2. For three implicational walks P , Q, and R ending at the same vertex, if P ◦ Q−1 is conflicting, then at least one of P ◦ R−1 and R ◦ Q−1 is conflicting. We now introduce a half-integral relaxation8 of 0/1/all Deletion. Let FC,ϕA denote the set of all ϕA -conflicting walks whose internal vertices do not intersect with A9 . When C and ϕA are clear from the context, we simply write F to refer to FC,ϕA . Note that from Corollary 1, we can ignore the direction of walks in F. A function x : V → {0, 21 , 1} is called a half-integral F-cover if x(V (W )) ≥ 1 for every W ∈ F. The size of x is defined as |x| = x(V ). A function y : F → {0, 12 , 1} 7 When at least one of P or Q is not implicational, neither P ≡ Q nor P 6≡ Q holds. Originally we should define this as an LP relaxation, but we discuss it by assuming its half integrality, which is proved via our algorithm and LP duality. 9 This constraint is only for simplicity. From Corollary 2, if P ◦ Q−1 is conflicting for some walks P and Q ending in A, at least one of P and Q is conflicting. Therefore, a maximum packing of conflicting walks that uses none of such walks always exists. 8 9 P is called a half-integral F-packing if for every vertex v ∈ V , it holds that W ∈F 1V (W ) (v)y(W ) ≤ 1. The size of y is defined as |y| = y(F). From the LP-duality, we have |x| ≥ |y| for any pair of halfintegral F-cover x and half-integral F-packing y. Any deletion set X for 0/1/all Deletion must intersect every ϕA -conflicting walk; hence 1X is an integral F-cover10 . Therefore, the size of the minimum half-integral F-cover provides a lower bound on the size of the minimum deletion set. For a half-integral F-cover x, let R(x) denote the set of vertices t such that an implicational walk W with x(V (W )) = 0 ending at t exists. We can prove the following property called persistency by a careful consideration of the results in [19] (see Appendix A for a detailed discussion). Theorem 1. Let C be a set of 0/1/all constraints on a variable set V and ϕA be a partial assignment for a subset A ⊆ V . For any minimum half-integral FC,ϕA -cover x, there exists a minimum deletion set X containing every vertex u with x(u) = 1 but avoiding every vertex in R(x). We say that a minimum half-integral F-cover x0 dominates a minimum half-integral F-cover x if R(x) ( R(x0 ) holds. A minimum half-integral F-cover x is called farthest if there exists no minimum half-integral F-cover dominating x. Suppose that we have an O(T )-time algorithm for computing a farthest minimum half-integral F-cover. From Theorem 1, it is not difficult to obtain an O(d2k T )-time FPT algorithm for 0/1/all Deletion. Moreover, the base d of the exponent can be improved to a constant for several special cases. We give a detailed discussion in Section 5.1. 2.3 Single-Branching Pair and Incremental-Test Oracle As shown in Section 5.2, various important NP-hard problems can be expressed as a special case of 0/1/all Deletion. However, the domain size d is often ω(1) (or even exp(m) for several cases, where m is the number of edges/constraints). Hence, if every constraint is given as the table of size d, the total size of these tables already becomes super-linear. For obtaining linear-time FPT algorithms, we use oracles instead of the explicit expression of the constraints to efficiently check whether a given walk is implicational/conflicting or not. A pair (P, Q) of implicational walks ending at the same vertex is called single-branching if either P ◦ Q−1 forms a simple path or they can be written as P = R ◦ P 0 and Q = R ◦ Q0 for a (possibly zero-length) path R and two walks P 0 and Q0 for which P 0 ◦ Q0−1 forms a simple cycle that is internally disjoint from R. In our algorithm, we need to test whether a given walk is implicational and whether P ◦ Q−1 is conflicting for a given single-branching pair (P, Q). To efficiently answer these queries, we use a tuple (U, I, A, T ), called an incremental-test oracle, of a set U and functions I, A, and T satisfying the following. • Init I : A → U . • Append A : U × Ê → U ∪ {all}. Let A∗ be a function such that ( ( I(s) (s ∈ A) A(A∗ (W ), e) (A∗ (W ) 6= all) ∗ ∗ A ((s)) = and A (W ◦ e) = . all (s 6∈ A) all (A∗ (W ) = all) Then, for any walk W , A∗ (W ) 6= all if and only if W is implicational. 10 Note that the converse may not hold. For example, we have F = ∅ when A = ∅. Therefore, X = ∅ is the minimum integral F-cover. In contrast, X = ∅ may not be a deletion set. Thus, the half-integral relaxation does not always lead to a 2-approximation algorithm. 10 • Test T : U × U → {true, false}. For any single-branching pair (P, Q), T (A∗ (P ), A∗ (Q)) = true if and only if P ◦ Q−1 is conflicting. The running time of the incremental-test oracle is defined as the maximum running time of the three S functions. In general, we can naively implement the incremental-test oracle by setting U := u∈V D(u), I(s) := ϕA (s), A(x, e) = Ce (x), and T (x, y) := true iff x 6= y. We can obtain a constant-time oracle by this naive implementation for several cases including Node Multiway Cut. Meanwhile, the naive implementation takes O(m) time for several other cases including Subset Feedback Vertex Set. We will see in Section 5.2 that we can implement a constanttime oracle for these cases by exploiting the constraint that the inputs to the test function are restrited to single-branching pairs. 3 Half-Integral Packing and Covering We prove the following theorem in this section. Theorem 2. Let C be a set of 0/1/all constraints on variables V and ϕA be a partial assignment for a subset A ⊆ V . Given the primal graph of C, the set A, an incremental-test oracle for (C, ϕA ), and an integer k, we can compute a pair of minimum half-integral FC,ϕA -cover x and maximum half-integral FC,ϕA -packing y with |x| = |y| ≤ k2 or correctly conclude that the size of the minimum half-integral FC,ϕA -cover is at least k+1 2 in O(kmT ) time, where m is the number of constraints and T is the running time of the incremental-test oracle. Our algorithm is based on a simple augmentation strategy summarized as follows. Starting with y(W ) = 0 for every W ∈ F, we repeatedly update a half-integral F-packing y that always consists of only two types of conflicting walks defined in Section 3.1.1. We search an augmenting path/pair (see Section 3.1.2) in each iteration using Algorithm 1 described in Section 3.2. If one is found, we can improve the current F-packing y in linear time by Lemmas 2 and 3 (Augmentation); otherwise, as shown in Section 3.3, we can naturally construct a half-integral F-cover of size |y|, which guarantees the optimality of y with the aid of the LP-duality. Since each augmentation increases |y| by at least 21 , the number of iterations is bounded by k + 1. Algorithm 1 can be implemented in linear time by Lemma 6, which concludes Theorem 2. 3.1 3.1.1 Preliminaries Basic F-Packing In what follows, we focus on F-packings that consist of only two types of conflicting walks. One is a simple path I ∈ F of weight 1, which is called an integral path. The other is a wheel defined as follows and consists of an odd number of conflicting walks of weight 21 . Definition 1. A pair of a simple cycle C = H1 ◦ . . . ◦ Hd of weight 21 and (possibly zero-length) paths {S1 , . . . , Sd } of weight 1 is called a wheel if it satisfies the following conditions (Figure 2). 1. d is an odd positive integer. 2. For any i, Si is a path from A to s(Hi ) = t(Hi−1 ) (where H0 = Hd ) that is internally disjoint from C. 11 𝐻5 𝑆5 𝐻1 𝑆2 𝐴 𝐻4 Spoke 𝑆1 𝐻2 𝑆4 𝑆3 𝐻3 Half-integral cycle Figure 2: Wheel of degree 5. 3. For any distinct i and j, Si and Sj share no vertices. −1 4. For any i, Si ◦ Hi ◦ Si+1 ∈ F, where Sd+1 = S1 . The integer d is called the degree of the wheel, the cycle C is called the half-integral cycle of the wheel, and the paths {S1 , . . . , Sd } are called the spokes of the wheel. Note that a wheel of degree 1 is a closed walk S1 ◦ C ◦ S1−1 ∈ F of weight 21 , and a wheel of degree d ≥ 3 is a sum of d simple paths {S1 ◦ H1 ◦ S2−1 , . . . , Sd ◦ Hd ◦ S1−1 } ⊆ F of weight 21 . A half-integral F-packing y is called a basic F-packing if it is a sum of integral paths and wheels such that each vertex is contained in at most one of the integral paths and the wheels. In our algorithm, a basic F-packing y is dealt with as a weighted graph11 so that we can efficiently update integral paths and wheels in y. We denote by V (y) and E(y) the sets of vertices and (undirected) edges, respectively, that are contained in some integral path or wheel in y (i.e., of positive weights), and particularly by V1 (y) and E1 (y), the sets of those contained in some integral path or spoke in y (i.e., are of weight 1). A walk is called internally disjoint from y if it is internally disjoint from the subgraph (V (y), E(y)). For a basic F-packing y, we define two functions Fy (Forward) and By (Backward) as follows (Figure 3). Let P be a positive-length path contained in an integral path I in y. From Corollary 1, we can assume that I has the same direction as P . We then define paths Fy (P ) and By (P ) such that I = Fy (P ) ◦ P ◦ By (P )−1 holds. For a vertex v contained in a spoke S in y, we denote by Fy (v) the path from t(S) to v along S and by By (v) the path from s(S) to v along S (i.e., Fy (v) ◦ By (v)−1 = S −1 ). For a path P contained in a spoke S in y in the opposite direction to S, we define Fy (P ) := Fy (s(P )) and By (P ) := By (t(P )) (i.e., Fy (P ) ◦ P ◦ By (P )−1 = S −1 ). We omit the subscript y if it is clear from the context. 3.1.2 Augmenting Path/Pair We first define an alternating path to define our augmenting path/pair. Definition 2. For a basic F-packing y, a concatenation of paths P = P1 ◦ · · · ◦ Pp is called a y-alternating path if it satisfies all the following conditions. 11 The weight is naturally defined for each vertex v and edge e by respectively. 12 P W ∈F 1V (W ) (v)y(W ) and P W ∈F 1E(W ) (e)y(W ), 𝐹(𝑣) 𝐹(𝑃) 𝐼 𝑃 𝑆 𝐹(𝑃) 𝑣 𝑆 𝐵(𝑣) 𝐵(𝑃) 𝑃 is contained in an integral path 𝐼 𝑣 is contained in a spoke 𝑆 𝑃 𝐵(𝑃) 𝑃 is contained in a spoke 𝑆 Figure 3: Definition of Fy and By . 1. The edges in E(P ) are distinct (i.e., 1E(P ) (e) ≤ 1 for every e ∈ E). 2. Every vertex in P that is not contained in any integral path or spoke in y appears in P at most once12 (i.e., 1V (P ) (v) ≤ 1 for every v ∈ V \ V1 (y)). 3. s(P ) ∈ A \ V (y). 4. Each Pi is a path of positive length satisfying the following conditions. (a) For any odd i, Pi is internally disjoint from y, and no internal vertex of Pi is in A. (b) For any even i, Pi is contained in an integral path or a spoke in y. In the latter case, Pi has the opposite direction to the spoke. 5. Let us define B(P0 ) := (s(P )). The following conditions are satisfied for any i13 (see Figure 4). (a) If i is odd, B(Pi−1 ) ◦ Pi is implicational. (b) If i is even and Pi is contained in an integral path, then B(Pi−2 )◦Pi−1 6≡ F (Pi ), and none of the Pj ’s are contained in B(Pi ) for j > i. Moreover, if B(Pi−2 ) ◦ Pi−1 6≡ B(Pi ) ◦ Pi−1 , none of the Pj ’s are contained in F (Pi ) for j > i. (c) If i is even and Pi is contained in a spoke, then B(Pi−2 ) ◦ Pi−1 ≡ B(Pi ) ◦ Pi−1 , and none of the Pj ’s are contained in B(Pi ) for j > i. Each Pi is called a segment of P . If P consists of only a single segment, it is called single-segment. A zero-length path P = (s) with s ∈ A \ V (y) is considered as a y-alternating path with zero segments. We define Ty (P ) (Tail) for a y-alternating path P = P1 ◦ · · · ◦ Pp as follows: Ty (P ) := By (Pp−1 ) ◦ Pp if p is odd, and Ty (P ) := By (Pp ) if p is even. Note that Ty (P ) is always implicational. We omit the subscript y if it is clear from the context. Definition 3. A y-alternating path P = P1 ◦ . . . ◦ Pp is called a y-augmenting path if p is odd and one of the following conditions is satisfied. 12 From the other properties of the y-alternating paths, a vertex contained in an integral path or a spoke can appear twice in P . Hence, P may not be a path in the precise sense. 13 Technically, the conditions “none of the Pj ’s are contained in . . . ” mean that P does not admit a shortcut (e.g., if there is some Pj with j > i contained in B(Pi ), we can obtain another y-alternating path P1 ◦· · ·◦Pi−1 ◦W ◦Pj+1 ◦· · · Pp , where W is the path from s(Pi ) to t(Pj ) along the integral path or spoke). As in the case of matroid intersection, we need a shortcut-less alternating path. 13 𝐵(𝑃𝑖−2 ) ≢ 𝑃𝑖−2 𝑃𝑖−1 𝐹(𝑃𝑖−2 ) 𝐹(𝑃𝑖 ) 𝑃𝑖 𝐵(𝑃𝑖−2 ) ≡ 𝑃𝑖−2 𝑃𝑖−1 𝐹(𝑃𝑖−2 ) 𝐵(𝑃𝑖 ) 𝑃𝑖 is contained in an integral path 𝐹(𝑃𝑖 ) 𝑃𝑖 𝐵(𝑃𝑖 ) 𝑃𝑖 is contained in a spoke Figure 4: Conditions for alternating paths. 1. t(P ) ∈ A \ V (y), and T (P ) is conflicting. 2. t(P ) is contained in a half-integral cycle, but in no spokes (i.e., t(P ) ∈ V (y) \ V1 (y)). 3. t(P ) is contained in a spoke S, and the following two conditions are satisfied. (a) T (P ) 6≡ B(t(P )). (b) For any Pj contained in the spoke S, t(P ) is contained in F (Pj ). Definition 4. Let P = P1 ◦ · · · ◦ Pp and Q = Q1 ◦ · · · ◦ Qq be a pair of y-alternating paths ending at the same vertex. (P, Q) is called a y-augmenting pair if it satisfies all the following conditions. 1. P and Q can be written as P = R ◦ P 0 and Q = R ◦ Q0 , respectively, for some walk R such that P 0 and Q0 share no edges. 2. T (P ) 6≡ T (Q). 3. At least one of p and q is odd, and if both of p and q are odd, t(P ) 6∈ V (y). 4. For any Pi contained in an integral path, none of the Qj ’s are contained in B(Pi ) in the opposite direction. Moreover, none of the Qj ’s are contained in F (Pi ) in the same direction if B(Pi−2 ) ◦ Pi−1 6≡ B(Pi ) ◦ Pi−1 . The symmetric condition holds for any Qi contained in an integral path. Note that for any y-alternating path P , the condition 4 is always satisfied against (P, P ). Therefore, in testing the condition 4 against (P, Q), we only need to test pairs (Pi , Qj ) such that at least one of Pi or Qj is not contained in the common prefix R. Using a y-augmenting path/pair, we can improve a basic F-packing y in linear time by the following lemmas, whose proofs are given in Section 3.4. Lemma 2. Given a basic F-packing y and a y-augmenting path, a basic F-packing of size at least |y| + 21 can be constructed in linear time. Lemma 3. Given a basic F-packing y and a y-augmenting pair, a basic F-packing of size |y| + can be constructed in linear time. 1 2 Figure 5 illustrates an example of augmentations. We use the following example which corresponds to the internally-disjoint A0 -paths with A0 = {a, b, c, d}. Every vertex u has the same domain D(u) = {a, b, c, d} and each edge uv has an equality (identity permutation) constraint 14 𝑐 𝑎 𝑑 𝑐 𝑑 𝑐 𝑎 1 𝑎 6 𝑑 5 4 3 2 𝑏 𝐴 𝑏 𝑎 𝑏 𝑐 𝑑 𝑏 𝐴 𝑎 4 7 2 𝑏 𝑏 𝑎 𝑏 𝑎 𝑏 𝑐 𝑑 6 2 𝐴 𝑎 1 𝑏 𝑐 𝑑 𝑏 𝑎 𝑏 𝐴 𝑏 3 𝑎 𝑏 𝐴 5 1 3 𝑏 𝑎 𝑏 𝑏 𝐴 𝑏 𝑎 𝑏 Figure 5: Example of augmentations. Thick solid lines denote edges of weight 1, dashed lines denote edges of weight 12 , and thin solid lines denote edges of weight 0. Blue or green arrows denote segments of augmenting paths/pairs. A number i beside an arrow means that it is the i-th segment. ϕ(u) = ϕ(v). An alphabet beside a vertex u ∈ A shows the value ϕA (u). Any walks starting from A is implicational, and a walk W is conflicting if and only if ϕA (s(W )) 6= ϕA (t(W )). Roughly speaking, for an augmenting path P connecting two distinct vertices in A (i.e., when it satisfies the condition 1 of Definition 3), we take the symmetric difference of E(y) and E(P ) (Figure 5, left). For an augmenting path P ending at a vertex on a wheel (i.e., when it satisfies the condition 2 or 3 of Definition 3), we take the symmetric difference and then decompose the wheel into integral paths (Figure 5, middle). For an augmenting pair (R ◦ P 0 , R ◦ Q0 ) with a common prefix R, we take the symmetric difference of E(y) and E(R), and then introduce a new wheel whose half-integral cycle is P 0 ◦ Q0−1 (Figure 5, right). In addition to this basic augmentation strategy, we need several operations in ensuring that the obtained packing is a basic F-packing. See the proofs in Section 3.4 for more detail. 3.2 Finding Augmenting Path/Pair In this subsection, we propose an algorithm for computing a y-augmenting path or pair. Algorithm 1 describes a rough sketch of the algorithm. We will describe later the details of an efficient implementation using the incremental-test oracle. Note that in this subsection, we prove only the soundness of the algorithm (i.e., the algorithm never returns a path or pair which is not yaugmenting). The completeness of the algorithm (i.e., the algorithm always finds a y-augmenting path or pair if exists) follows from Lemma 7 proved in the next subsection. 15 Algorithm 1 Algorithm for computing a y-augmenting path/pair 1: Initialize a(I) ← 0 and b(I) ← ` for each integral path I of length `. 2: Initialize a(S) ← 0 for each spoke S. 3: for s ∈ A \ V (y) do 4: P (s) ← (s) and X ← {s}. 5: while X 6= ∅ do 6: Pick a vertex u ∈ X and remove u from X. 7: for e = uv ∈ δ(u) \ E(y) such that T (P (u)) ◦ e is implicational do 8: if v is visited then 9: if T (P (u)) ◦ e 6≡ T (P (v)) then return (P (u) ◦ e, P (v)) 10: else if v is contained in an integral path I = (v0 , . . . , v` ) then 11: Let i be the index such that vi = v. 12: if T (P (u)) ◦ e 6≡ (v` , . . . , vi ) then 13: for j ∈ {a(I), . . . , i − 1} do 14: P (vj ) ← (P (u) ◦ e) ◦ (vi , . . . , vj ) and X ← X ∪ {vj }. 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: 29: 30: 31: 32: a(I) ← i. if T (P (u)) ◦ e 6≡ (v0 , . . . , vi ) then for j ∈ {i + 1, . . . , b(I)} do P (vj ) ← (P (u) ◦ e) ◦ (vi , . . . , vj ) and X ← X ∪ {vj }. b(I) ← i. else if v is contained in a spoke S = (v0 , . . . , v` ) then Let i be the index such that vi = v. if T (P (u)) ◦ e 6≡ (v0 , . . . , vi ) then return P (u) ◦ e for j ∈ {a(S), . . . , i − 1} do P (vj ) ← (P (u) ◦ e) ◦ (vi , . . . , vj ) and X ← X ∪ {vj }. a(S) ← i. else if v is contained in a half-integral cycle then return P (u) ◦ e else if v ∈ A then if T (P (u)) ◦ e 6≡ (v) then return P (u) ◦ e else (v 6∈ A ∪ V (y)) P (v) ← P (u) ◦ e and X ← X ∪ {v}. return NO Lemma 4. Any path returned by Algorithm 1 is a y-augmenting path, and any pair returned by the algorithm is a y-augmenting pair. In the algorithm, we hold two indices a(I) and b(I) for each integral path I initialized as a(I) = 0 and b(I) = `, respectively, where ` is the length of I; an index a(S) for each spoke S initialized as a(S) = 0; a set of active vertices X initialized as X = ∅; and a y-alternating path P (v) for each visited vertex v, where we call a vertex v visited if v has been ever pushed into X. We define boundaries as the set of vertices consisting of va(I) and vb(I) for each integral path I = (v0 , . . . , v` ) and va(S) for each spoke S = (v0 , . . . , v` ). We preserve the following invariants during the execution 16 of the algorithm. Lemma 5. The following invariants hold at any iteration of Algorithm 1. 1. For any visited v, the following holds: (a) P (v) is a y-alternating path, (b) V (P (v)) contains only visited vertices or boundaries, and (c) P (v) has an odd number of segments if and only if v 6∈ A ∪ V (y). 2. For any visited u and v with s(P (u)) = s(P (v)), (P (u), P (v)) satisfies the condition 1 of y-augmenting pairs (Definition 4). 3. For any integral path I = (v0 , . . . , v` ), the following holds: (a) a(I) ≤ b(I), (b) vi is visited if and only if i < a(I) or b(I) < i, and (c) for any P (v) with segments P1 ◦· · ·◦Pp and any segment Pi contained in I, Pi is contained in (va(I) , . . . , v0 ) or (vb(I) , . . . , v` ) in these directions; moreover, if B(Pi−2 ) ◦ Pi−1 6≡ B(Pi ) ◦ Pi−1 holds, then s(Pi ) = va(I) = vb(I) holds. 4. For any spoke S = (v0 , . . . , v` ), the following holds: (a) vi is visited if and only if i < a(S), and (b) for any P (v), all the segments of P (v) contained in S are contained in (va(S) , . . . , v0 ). We now prove Lemma 4 and 5. All the invariants are clearly satisfied at the beginning. In each iteration, we pick an arbitrary vertex u from X (line 6). From the invariant 1, P (u) is a y-alternating path. We then iterate over the edges e = uv ∈ δ(u) \ E(y) such that T (P (u)) ◦ e is implicational. When v is already visited and T (P (u)) ◦ e 6≡ T (P (v)) holds, we return a pair (P (u) ◦ e, P (v)) (line 9). Claim 1. (P (v), P (u) ◦ e) returned at line 9 is a y-augmenting pair. Proof. First, we prove e 6∈ E(P (u)), which implies that P (u) ◦ e is a y-alternating path. If e ∈ E(P (u)), e−1 is either the last edge of P (u) or the last edge of an odd segment Pi of P (u) := P1 ◦· · ·◦ Pp . In the former case, T (P (u))◦e = T (P (v))◦e−1 ◦e ≡ T (P (v)) holds, which is a contradiction. In the latter case, from the invariant 2 against (P (u), P (v)), we have P (v) = P1 ◦ · · · ◦ Pi−1 ◦ W , where W is the path satisfying W ◦ e−1 = Pi . Because Pi+1 and Pp are contained in the same integral −1 path or spoke in the same direction, we have B(Pi−1 ) ◦ Pi ≡ B(Pi+1 ) ◦ Pi+1 from the condition 5 −1 of y-alternating paths. Therefore, we have T (P (u)) ◦ e = B(Pp ) ◦ e = B(Pi+1 ) ◦ Pi+1 ◦e ≡ −1 −1 B(Pi−1 ) ◦ Pi ◦ e = B(Pi−1 ) ◦ W ◦ e ◦ e = T (P (v)) ◦ e ◦ e ≡ T (P (v)), which is a contradiction. Next, we prove that (P (v), P (u)◦e) satisfies all the conditions of the y-augmenting pairs (Definition 4). Suppose that s(P (u)) 6= s(P (v)) holds. Then, the vertex v has already been popped from X and, thus, we have P (u) = P (v) ◦ e−1 , which is a contradiction. Therefore, from the invariant 2, the condition 1 is satisfied. The condition 2 is satisfied because T (P (v)) 6≡ T (P (u)) ◦ e = T (P (u) ◦ e). Meanwhile, the condition 3 is satisfied because the number of segments of P (u) ◦ e is odd and because, from the invariant 1, v ∈ V (y) implies that the number of segments of P (v) is even. The condition 4 follows from the invariant 3. 17 When v is not visited, we consider five cases: (Case 1) v is contained in an integral path; (Case 2) v is contained in a spoke; (Case 3) v is contained in a half-integral cycle (but not in spokes); (Case 4) v ∈ A (but not in V (y)); or (Case 5) v 6∈ A ∪ V (y). Note that e is not contained in E(P (u)) when v is not visited. Therefore, P (u) ◦ e is a y-alternating path. In case 3, we return P (u) ◦ e, which is a y-augmenting path satisfying the condition 2 of Definition 3. In case 4, if T (P (u)) ◦ e 6≡ (v) holds, we return P (u) ◦ e, which is a y-augmenting path satisfying the condition 1 of Definition 3, and otherwise, we do nothing. In case 5, we set P (v) ← P (u) ◦ e and insert v into X (line 31). Because e 6∈ E(y) holds, all the invariants are clearly preserved. Finally, we consider the remaining two cases. (Case 1) Let I = (v0 , . . . , v` ) be the integral path containing v and let i be the index such that vi = v. Because v is not visited, a(I) ≤ i ≤ b(I) holds from the invariant 3. If T (P (u)) ◦ e 6≡ (v` , . . . , vi ) holds, we set P (vj ) ← (P (u) ◦ e) ◦ (vi , . . . , vj ) and insert vj into X for each index j ∈ {a(I), . . . , i − 1}, and then update a(I) ← i (lines 12–15). Similarly, if T (P (u)) ◦ e 6≡ (v0 , . . . , vi ) holds, we set P (vj ) ← (P (u)◦e)◦(vi , . . . , vj ) and insert vj into X for each index j ∈ {i+1, . . . , b(I)}, and then update b(I) ← i (lines 16–19). Claim 2. The lines 12–19 preserve all the invariants. Proof. The invariants 1b, 1c, 2, 3a, 3b, and 4 are clearly preserved. Let a(I) and b(I) denote the indices before the updates. For an index j ∈ {a(I), . . . , i − 1}, let P1 ◦ · · · ◦ Pp be the segments of P (vj ). If B(Pp−2 ) ◦ Pp−1 6≡ B(Pp ) ◦ Pp−1 holds, we have T (P (u)) ◦ e = B(Pp−2 ) ◦ Pp−1 6≡ B(Pp ) ◦ Pp−1 = (v0 , . . . , vi ). Therefore, both a(I) and b(I) are updated to i. The same argument applies to the case of j ∈ {i + 1, . . . , b(I)}. Thus, the invariant 3c is preserved. Finally, we prove the invariant 1a by showing that P (vj ) is a y-alternating path for any newly visited vj . The conditions 1–4 of y-alternating paths (Definition 2) are clearly satisfied. Consider the case of j ∈ {a(I), . . . , i − 1}. The proof for the case of j ∈ {i + 1, . . . , b(I)} is symmetric. Let P1 ◦ · · · ◦ Pp be the segments of P (vj ). Because B(Pp−2 ) ◦ Pp−1 = T (P (u)) ◦ e 6≡ (v` , . . . , vi ) = F (Pp ) holds, the condition 5 is satisfied for p. Let Pk be a segment with k < p contained in the same integral path I. From the invariant 3c and because a(I) ≤ j < i ≤ b(I), Pp = (vi , . . . , vj ) is contained in F (Pk ), and B(Pk−2 ) ◦ Pk−1 ≡ B(Pk ) ◦ Pk−1 holds. Therefore, the condition 5 is satisfied for k. (Case 2) Let S = (v0 , . . . , v` ) be the spoke containing v and let i be the index such that vi = v. Because v is not visited, a(S) ≤ i holds from the invariant 4. If T (P (u)) ◦ e 6≡ (v0 , . . . , vi ) holds, we return a path P (u) ◦ e (line 22); otherwise, we set P (vj ) ← (P (u) ◦ e) ◦ (vi , . . . , vj ) and insert vj into X for each index j ∈ {a(S), . . . , i − 1}, and then update a(S) ← i (lines 23–25). Claim 3. P (u) ◦ e returned at line 22 is a y-augmenting path. Proof. Because T (P (u)) ◦ e 6≡ (v0 , . . . , vi ) = B(t(P (u) ◦ e)) holds, the condition 3a of the y-augment paths (Definition 3) is satisfied. From the invariant 4, for any segment Pk of P (u) ◦ e contained in the spoke S, t(P (u) ◦ e) = vi is contained in F (Pk ). Therefore, the condition 3b is satisfied. Claim 4. Lines 23–25 preserve all the invariants. Proof. All the invariants, except for the 1a, are clearly preserved. We prove the invariant 1a by showing that P (vj ) is a y-alternating path for any newly visited vj . The conditions 1–4 of the 18 y-alternating paths (Definition 2) are clearly satisfied. Let P1 ◦ · · · ◦ Pp be the segments of P (vj ). Because B(Pp−2 )◦Pp−1 = T (P (u)◦e) ≡ (v0 , . . . , vi ) = B(Pp )◦Pp−1 holds, the condition 5 is satisfied for p. Let Pk be a segment with k < p contained in the same spoke S. From the invariant 4, Pk must be contained in (va(S) , . . . , v0 ). Therefore, Pp is contained in F (Pk ). Thus, the condition 5 is satisfied for k. Implementation Detail We exploit the incremental-test oracle as follows to achieve the linear-time complexity. For each integral path I = (v0 , . . . , v` ) and for each index i ∈ {0, . . . , `}, we precompute A∗ ((v0 , . . . , vi )) and A∗ ((v` , . . . , vi )). For each spoke S = (v0 , . . . , v` ) and for each index i ∈ {0, . . . , `}, we precompute A∗ ((v0 , . . . , vi )). These precomputations can be done in O(nT ) time, where T is the running time of the oracle. For each visited vertex v, instead of explicitly holding P (v), we hold (1) prev(v) that represents the edge picked at line 7 in the iteration when P (v) is assigned and (2) tail(v) := A∗ (T (P (v))). When the algorithm finds a y-augmenting path or pair, we restore P (v) using the table of prev(v) in O(n) time. For an edge e = uv ∈ δ(u) \ E(y), we have A∗ (T (P (u)) ◦ e) = A(tail(u), e), which can be computed in O(T ) time. For any index j picked at line 13 or 23, we have tail(vj ) = A∗ (T (P (vj ))) = A∗ ((v0 , . . . , vj )), which has been precomputed. For any index j picked at line 17, we have tail(vj ) = A∗ (T (P (vj ))) = A∗ ((v` , . . . , vj )), which has been precomputed. When v 6∈ A ∪ V (y), we have tail(v) = A(tail(u), e). Thus, we can compute tail(w) in O(T ) time for each newly visited vertex w. Next, we show that each of the equivalence tests in the algorithm can be done in O(T ) time. Let e = uv be the edge picked at line 7. First, we consider the equivalence test at line 9. If prev(v) = e−1 holds, we have T (P (u)) ◦ e ≡ T (P (v)); otherwise, (T (P (u)) ◦ e, T (P (v))) forms a single-branching pair from the invariant 2, for which we can test T (P (u)) ◦ e 6≡ T (P (v)) by asking T (A(tail(u), e), tail(v)) in O(T ) time. Next, we consider the equivalence test at line 12. Because (T (P (u) ◦ e), (v` , . . . , vi )) forms a single-branching pair, we can test T (P (u)) ◦ e 6≡ (v` , . . . , vi ) by asking T (A(tail(u), e), A∗ ((v` , . . . , vi ))). This can be done in O(T ) time because A∗ ((v` , . . . , vi )) has been precomputed. The same argument applies to the equivalence tests at lines 16 and 22. The equivalence test at line 29 can be done in O(T ) time by asking T (A(tail(u), e), I(v)). Now, we have shown that tail(v) can be computed in O(T ) time for each visited vertex v, and that all the equivalence tests can be done in O(T ) time for each edge picked at line 7. Because each vertex is pushed into X at most once and because each edge is processed at most twice (in both directions), we obtain the following lemma. Lemma 6. Given an O(T )-time incremental-test oracle, Algorithm 1 runs in O(mT ) time. 3.3 Constructing Half-Integral F-Cover In this subsection, we prove that if Algorithm 1 fails to find a y-augmenting path or pair, we can construct a half-integral F-cover of the same size as follows. Let a and b be the tables used in Algorithm 1. First, we initialize x(v) ← 0 for all v ∈ V . For each integral path I = (v0 , . . . , v` ), we set x(va(I) ) ← 21 and x(vb(I) ) ← 21 if a(I) 6= b(I), and set x(va(I) ) ← 1 if a(I) = b(I). We set x(va(S) ) ← 21 for each spoke S = (v0 , . . . , v` ). From the construction, we have |x| = |y|. We show that the function x is an F-cover. 19 Lemma 7. If Algorithm 1 fails to find a y-augmenting path or pair, the function x constructed as above is a half-integral F-cover of size |y|. We use the following lemma to prove this lemma. Lemma 8. For any implicational walk Q with x(V (Q)) = 0, the vertex t(Q) is visited and T (P (t(Q))) ≡ Q holds. Proof. We prove the lemma by induction on the length of Q. For any vertex s ∈ A with x(s) = 0, s is visited, and T (P (s)) = (s) holds. Therefore, the statement holds when the length of Q is zero. Let Q = Q0 ◦ uv be an implicational walk with x(V (Q)) = 0. From the induction hypothesis, u is visited, and T (P (u)) ≡ Q0 holds. Therefore, we have T (P (u)) ◦ uv ≡ Q0 ◦ uv = Q. If uv 6∈ E(y) and v is contained in a half-integral cycle but in no spokes, the algorithm returns a y-augmenting path at line 27. We consider the following four cases: (Case 1) v 6∈ V (y); (Case 2) uv 6∈ E(y) and v is contained in an integral path; (Case 3) uv 6∈ E(y) and v is contained in a spoke; or (Case 4) uv ∈ E(y). (Case 1) Consider the iteration when u is picked at line 6. If v is already visited, T (P (u)) ◦ uv ≡ T (P (v)) holds because, otherwise, the algorithm returns a y-augmenting pair. Therefore, we have T (P (v)) ≡ T (P (u)) ◦ uv ≡ Q. If v is not visited yet and v ∈ A, we have T (P (u)) ◦ uv ≡ (v). Therefore, we have T (P (v)) = (v) ≡ T (P (u)) ◦ uv ≡ Q. If v is not visited yet and v 6∈ A, then v becomes visited, and we have T (P (v)) = T (P (u)) ◦ uv ≡ Q. (Case 2) Let I = (v0 , . . . , v` ) be the integral path containing v and let i be the index such that vi = v. Consider the iteration when u is picked at line 6. Let a(I) and b(I) denote the values at the beginning of this iteration (hence, x(va(I) ) and x(vb(I) ) might be zero). If i < a(I) holds, then v is already visited, and T (P (v)) = (v0 , . . . , vi ) holds. If T (P (u)) ◦ uv 6≡ (v0 , . . . , vi ) additionally holds, the algorithm returns a y-augmenting pair at line 9. Therefore, we have T (P (v)) = (v0 , . . . , vi ) ≡ T (P (u)) ◦ uv ≡ Q. The same argument applies to the case of i > b(I). We now consider the remaining case that a(I) ≤ i ≤ b(I) holds. Because (v0 , . . . , vi ) 6≡ (v` , . . . , vi ) holds, at least one of T (P (u)) ◦ uv 6≡ (v0 , . . . , vi ) or T (P (u)) ◦ uv 6≡ (v` , . . . , vi ) holds. If both hold, a(I) and b(I) are both set to i after this iteration. Therefore, we have x(v) = 1, which is a contradiction. If T (P (u)) ◦ uv ≡ (v0 , . . . , vi ) holds, a(I) is set to i after this iteration. Because x(v) = 0, a(I) must be greater than i at the end of the algorithm. Therefore, we have T (P (v)) = (v0 , . . . , vi ) ≡ T (P (u)) ◦ uv ≡ Q. The same argument applies to the case of T (P (u)) ◦ uv ≡ (v` , . . . , vi ). (Case 3) Let S = (v0 , . . . , v` ) be the spoke containing v and i be the index such that vi = v. Consider the iteration when u is picked at line 6. Let a(S) denote the value at the beginning of this iteration. If i < a(S) holds, then v is already visited, and T (P (v)) = (v0 , . . . , vi ) holds. If T (P (u)) ◦ uv 6≡ (v0 , . . . , vi ) additionally holds, the algorithm returns a y-augmenting pair at line 9. Therefore, we have T (P (v)) = (v0 , . . . , vi ) ≡ T (P (u)) ◦ uv ≡ Q. We now consider the remaining case that a(I) ≤ i holds. If T (P (u)) ◦ uv 6≡ (v0 , . . . , vi ) holds, the algorithm returns a y-augmenting path at line 22. Therefore, T (P (u)) ◦ uv ≡ (v0 , . . . , vi ) holds and a(I) is set to i after this iteration. Because x(v) = 0, a(I) must be greater than i at the end of the algorithm. Therefore, we have T (P (v)) = (v0 , . . . , vi ) ≡ T (P (u)) ◦ uv ≡ Q. 20 (Case 4) Note that in this case, uv must be contained in an integral path or a spoke because the vertices contained in half-integral cycles are never visited. Let (v0 , . . . , v` ) be the integral path or the spoke containing uv and let i and j be the indices such that vi = v and vj = u. Because u is visited, T (P (u)) is either (v0 , . . . , vj ) or (v` , . . . , vj ), and w.l.o.g., we can assume the former case. If i = j−1, then v is also visited, and we have T (P (v)) = (v0 , . . . , vj−1 ) ≡ (v0 , . . . , vj ) ◦ uv = T (P (u)) ◦ uv ≡ Q. If i = j + 1 and v is not visited, then we have x(v) ≥ 21 , which is a contradiction. Therefore, v is also visited, and we have T (P (v)) = (v0 , . . . , vj+1 ) = (v0 , . . . , vj ) ◦ uv = T (P (u)) ◦ uv ≡ Q. Proof of Lemma 7. Suppose that there exists a conflicting walk W ∈ F with x(V (W )) < 1. Because x(v) ∈ {0, 21 , 1} holds, x(V (W )) is either 0 or 21 . If x(V (W )) = 0, by applying Lemma 8 against W , we have W ≡ T (P (t(W ))) = (t(W )), which is a contradiction. If x(V (W )) = 12 , let v be the vertex with x(v) = 12 on W . By splitting W at v, we obtain two implicational walks Q1 and Q2 1 such that W = Q1 ◦ Q−1 2 and t(Q1 ) = t(Q2 ) = v. Because x(v) = 2 , v is contained in an integral path or a spoke I = (v0 , . . . , v` ). Let i be the index such that vi = v and, w.l.o.g., we can assume that a(I) = i. If I is an integral path, b(I) must be greater than i because, otherwise, we have x(v) = 1. If the length of Q1 is zero, i is either 0 or `. In the latter case, I is an integral path, and ` = i = a(I) < b(I) ≤ `, which is a contradiction. Thus, we have Q1 = (v0 ). If Q1 has a positive length, let Q1 = R ◦ uv. By applying Lemma 8 against R, u is visited, and Q1 = R◦uv ≡ T (P (u))◦uv holds. If uv is contained in I, u must be vi−1 because vi+1 is not visited. Therefore, we have Q1 ≡ T (P (u)) ◦ uv = (v0 , . . . , vi−1 ) ◦ uv = (v0 , . . . , vi ). If uv is not contained in I, consider the iteration when u is picked at line 6. We then have Q1 ≡ T (P (u)) ◦ uv ≡ (v0 , . . . , vi ) because, otherwise, b(I) is set to i, or the algorithm returns a y-augmenting path at line 22. We now have proved that Q1 ≡ (v0 , . . . , vi ) holds. By applying the same argument to Q2 , we have Q2 ≡ (v0 , . . . , vi ). Therefore, we have Q1 ≡ Q2 , which is a contradiction. 3.4 3.4.1 Augmentation Simplification of Alternating Path Before the proofs of Lemmas 2 and 3, we introduce a useful procedure to simplify an augmenting path/pair obtained by Algorithm 1. We first define such an operation in a formal manner and prove the validity just after the definition. Definition 5. For a basic F-packing y and a y-alternating path P = P1 ◦ · · · ◦ Pp with p ≥ 2, the simplification (y 0 , P 0 ) of (y, P ) is defined as follows (Figure 6). • A function y 0 is constructed from y as follows: – if P2 is contained in an integral path I, replace I with an integral path P1 ◦ F (P2 )−1 ; – if P2 is contained in a spoke S, replace S with a spoke P1 ◦ F (P2 )−1 . • A walk P 0 is defined as a concatenation (By (P2 ) ◦ P3 ) ◦ P4 ◦ P5 ◦ · · · ◦ Pp of max{p − 2, 1} paths14 . Lemma 9. For any basic F-packing y and any y-alternating path P = P1 ◦ · · · ◦ Pp with p ≥ 2, the simplification (y 0 , P 0 ) of (y, P ) satisfies the following. 14 When p = 2, P 0 = By (P2 ). 21 ≢ 𝐹(𝑃2 ) ≡ 𝐹(𝑃2 ) 𝑃1 𝑃2 𝑃1 𝑃2 𝑃3 𝑃3 𝐵(𝑃2 ) 𝐵(𝑃2 ) 𝐵 𝑃2 ∘ 𝑃3 𝐵 𝑃2 ∘ 𝑃3 𝑃2 is contained in an integral path 𝑃2 is contained in a spoke Figure 6: Simplification of a basic F-packing and an alternating path. 1. y 0 is a basic F-packing with |y 0 | = |y|. 2. For any even i ≥ 4, the following holds. (a) If Pi is contained in an integral path in y, it is also contained in an integral path in y 0 . Moreover, Fy0 (Pi ) ≡ Fy (Pi ) and By0 (Pi ) ≡ By (Pi ) hold. (b) If Pi is contained in a spoke in y, it is also contained in a spoke in y 0 in the direction toward A. Moreover, Fy0 (Pi ) = Fy (Pi ) and By0 (Pi ) ≡ By (Pi ) hold. 3. P 0 = (By (P2 ) ◦ P3 ) ◦ P4 ◦ P5 ◦ · · · ◦ Pp is a y 0 -alternating path, where By (P2 ) ◦ P3 is the first segment, and Pi (i ≥ 4) is the (i − 2)-th segment 15 . Proof. First, we prove the first claim. When P2 is contained in an integral path I, from the condition 5b of the y-alternating paths (Definition 2), P1 6≡ F (P2 ) holds (i.e., I 0 := P1 ◦ F (P2 )−1 ∈ F). Because P1 is internally disjoint from F (P2 ) and s(P1 ) 6∈ V (y), I 0 is a simple path. Thus, we can replace I with the integral path I 0 . When P2 is contained in a spoke Si of a wheel, from the condition 5c, P1 ≡ B(P2 ) ◦ P2−1 holds. From the definition of the wheel (Definition 1), −1 Si−1 ◦ Hi−1 ◦ Si−1 and Si ◦ Hi ◦ Si+1 are both in F. As B(P2 ) ◦ P2−1 ◦ F (P2 )−1 = Si holds, we have Si0 := P1 ◦ F (P2 )−1 ≡ Si . When the degree of the wheel is at least three, Si−1 ◦ Hi−1 ◦ Si0−1 and −1 are both in F. When the degree of the wheel is one, S10 ◦ H1 ◦ S10−1 is in F. Thus, Si0 ◦ Hi ◦ Si+1 this replacement of the spoke preserves the condition for the wheel. Next, we prove the second claim. We can observe the following for any even i ≥ 4. • If both of P2 and Pi are contained in the same integral path in y, from the condition 5b in Definition 2, Pi is contained in Fy (P2 ), and P1 ≡ By (P2 ) ◦ P2−1 holds. Thus, Pi is contained in the integral path P1 ◦ Fy (P2 )−1 in y 0 . If they have the same direction, P2 is contained in 15 When p = 2, P 0 consists of the single segment By (P2 ). 22 By (Pi ), and we can write By (Pi ) = (By (P2 ) ◦ P2−1 ) ◦ W for some subpath W . Then, it holds that Fy0 (Pi ) = Fy (Pi ) and By0 (Pi ) = P1 ◦ W ≡ (By (P2 ) ◦ P2−1 ) ◦ W = By (Pi ). If they have the opposite direction, P2 is contained in Fy (Pi ), and we can write Fy (Pi ) = (By (P2 ) ◦ P2−1 ) ◦ W for some subpath W . Then, it holds that Fy0 (Pi ) = P1 ◦ W ≡ (By (P2 ) ◦ P2−1 ) ◦ W = Fy (Pi ) and By0 (Pi ) = By (Pi ). • If both P2 and Pi are contained in the same spoke in y, from the condition 5c in Definition 2, Pi is contained in Fy (P2 ), and P1 ≡ By (P2 ) ◦ P2−1 holds. Thus, Pi is contained in the spoke P1 ◦ Fy (P2 )−1 in y 0 . Because both of them are directed toward A, we can write By (Pi ) = (By (P2 ) ◦ P2−1 ) ◦ W for some subpath W . Then, it holds that Fy0 (Pi ) = Fy (Pi ) and By0 (Pi ) = P1 ◦ W ≡ (By (P2 ) ◦ P2−1 ) ◦ W = By (Pi ). • Otherwise, the integral path or the spoke containing Pi does not change. Thus, we have Fy0 (Pi ) = Fy (Pi ) and By0 (Pi ) = By (Pi ). Finally, we prove the third claim. Because By (P2 ) does not contain any Pi , the conditions 1– 4 in Definition 2 are satisfied. From the second claim and the property that A ◦ W ≡ B ◦ W holds for any implicational walks A ◦ W and B satisfying A ≡ B, none of the three equivalence relations appeared in the condition 5 change. As we have seen in the proof of the second claim, E(By0 (Pi )) \ E(By (Pi )) ⊆ E(P1 ) and E(Fy0 (Pi )) \ E(Fy (Pi )) ⊆ E(P1 ) hold. Therefore, for any i ≥ 4, none of the Pj ’s with j > i are newly contained in By0 (Pi ) or Fy0 (Pi ). Thus, the condition 5 is satisfied. We obtain the following corollaries from Lemma 9 by repeatedly applying the simplifying operation. Corollary 3. Given a basic F-packing y, a y-alternating path P = P1 ◦· · ·◦Pp , and an even integer p0 ≤ p, a basic F-packing y 0 of the same size and a y 0 -alternating path P 0 satisfying the following conditions can be constructed in linear time. 1. P 0 can be written as P 0 = (B 0 ◦Pp0 +1 )◦Pp0 +2 ◦· · ·◦Pp for some implicational path B 0 ≡ By (Pp0 ). 2. For any even i ≥ p0 + 2, the following holds. (a) If Pi is contained in an integral path in y, it is also contained in an integral path in y 0 . Moreover, Fy0 (Pi ) ≡ Fy (Pi ) and By0 (Pi ) ≡ By (Pi ) hold. (b) If Pi is contained in a spoke in y, it is also contained in a spoke in y 0 in the direction toward A. Moreover, Fy0 (Pi ) = Fy (Pi ) and By0 (Pi ) ≡ By (Pi ) hold. Corollary 4. Given a basic F-packing y and a y-alternating path P = P1 ◦ · · · ◦ Pp , a basic Fpacking y 0 of the same size and a single-segment y 0 -alternating path P 0 satisfying P 0 ≡ Ty (P ) can be constructed in linear time. 3.4.2 Augmentation by Augmenting Path The subsection aims to prove Lemma 2. 23 𝐴 𝐴 𝐴 𝐴 𝑡 is contained in a half-integral cycle 𝑡 is contained in a spoke Figure 7: Augmentation by a single-segment augmenting path. Proof of Lemma 2. Let P = P1 ◦ · · · ◦ Pp be the given y-augmenting path and t = t(P ). If p > 1, by Corollary 4, we obtain in linear time a basic F-packing y 0 of the same size and a single-segment y 0 alternating path P 0 such that t(P 0 ) = t and P 0 ≡ Ty (P ). We now show that P 0 is a y 0 -augmenting path. If P satisfies the condition 1 or 2 of the y-augmenting paths (Definition 3), P 0 is a y 0 augmenting path because t(P 0 ) = t is still contained in A \ V (y 0 ) or V (y 0 ) \ V1 (y 0 ). If P satisfies the condition 3, the spoke S that contains t in y might not exist in y 0 . However, from the condition 3b, t is still contained in a spoke S 0 in y 0 , which may not be identical to S. By the same argument as in the proof of the second statement of Lemma 9, we have By0 (t) ≡ By (t). Therefore, we have Ty0 (P 0 ) = P 0 ≡ Ty (P ) 6≡ By (t) ≡ By0 (t). Thus, P 0 is a y 0 -augmenting path. We can now concentrate on the case when p = 1 (Figure 7). If the condition 1 is satisfied, P is a conflicting path containing no vertices in V (y). Therefore, we can obtain a basic F-packing of size |y| + 1 by inserting the integral path P . Otherwise, let ({S1 , . . . , Sd }, H1 ◦ · · · ◦ Hd ) be the wheel containing t. W.l.o.g., we can assume that t is contained in Hd or Sd . If t is contained in Hd , let F be the prefix subpath of Hd to t and B be the suffix subpath of Hd from t (i.e., F ◦ B = Hd ). We have Sd ◦ F 6≡ S1 ◦ B −1 because Sd ◦ Hd ◦ S1−1 ∈ F Thus, at least one of Sd ◦ F ◦ P −1 and S1 ◦ B −1 ◦ P −1 is in F, and w.l.o.g., we can assume the former case. Then, we can obtain a basic F-packing of size |y| + 21 by decomposing the wheel into (d − 1)/2 integral paths −1 {S1 ◦ H1 ◦ S2−1 , S3 ◦ H3 ◦ S4−1 , . . . , Sd−2 ◦ Hd−2 ◦ Sd−1 } and by inserting an integral path Sd ◦ F ◦ P −1 . If t is contained in Sd , we have P 6≡ B(t). Then, we can obtain a basic F-packing of size |y| + 12 by decomposing the wheel into (d − 1)/2 integral paths {S1 ◦ H1 ◦ S2−1 , S3 ◦ H3 ◦ S4−1 , . . . , Sd−2 ◦ −1 Hd−2 ◦ Sd−1 } and by inserting an integral path P ◦ B(t)−1 . 24 𝑄1′ 𝑃1′ 𝑅 𝑅 𝑃2 𝐻1 𝑃𝑑 𝐻2 𝑃𝑑−1 𝐴 𝑃𝑑−2 𝑃3 𝑃4 𝑆1 𝑆𝑑 𝑆2 𝐻3 𝑆4 𝐻𝑑−2 𝑆5 𝐻4 𝑃5 𝐻𝑑−1 𝐻5 (a) 𝑄2 𝑃1 𝑄1 𝑃3 𝑆𝑑−1 𝐴 𝑆3 𝑃2 𝐻𝑑 𝑃𝑑 𝐴 𝑃𝑑−1 𝑃4 𝑃5 𝑃6 (b) Figure 8: Applying Lemma 10. 3.4.3 Augmentation by Augmenting Pair This subsection aims to prove Lemma 3. Intuitively, we want to augment y as follows. We first simplify the common prefix R of (P, Q) by applying Corollary 3, and then obtain a new wheel whose half-integral cycle is P 0 ◦ Q0−1 . However, it is not that easy. This approach does not work when P 0 or Q0 intersects with spokes or intersects with the same integral path multiple times. Therefore, we augment y by gradually simplifying the augmenting pair. First, we prove the lemma against a special case. Lemma 10. Given a basic F-packing y and a y-augmenting pair (P = P1 ◦· · ·◦Pp , Q = Q1 ◦· · ·◦Qq ) satisfying all the following conditions, a basic F-packing of size |y| + 21 can be constructed in linear time. 1. None of the Pi ’s are contained in the spokes. 2. None of the integral paths contain multiple Pi ’s. 3. One of the following two conditions is satisfied: (a) p ≥ q = 1 or (b) p ≥ q = 2, P1 = Q1 , and P2 and Q2 are contained in an integral path in the opposite direction. Proof. When (P, Q) satisfies the condition 3a, we further divide the case into the following two cases: p = 1 or p ≥ 2. If p = 1, (P, Q) is a single-branching pair such that P 6≡ Q. Therefore, we can obtain a basic F-packing of size |y| + 12 by introducing a new wheel P ◦ Q−1 of degree one. In p ≥ 2, let R be the common prefix and let us write P1 = R ◦ P10 and Q1 = R ◦ Q01 . We set d := p + 1 and Pd := (t(P )) if p is even; otherwise, we set d := p. We define paths {H1 , . . . , Hd } and {S1 , . . . , Sd } as follows (Figure 8 (a)). • H1 := P10 . • Hi := Pi for i ∈ {2, . . . , d − 1}. • Hd := Pd ◦ Q0−1 1 . 25 • S1 := R. • Si := F (Pi ) for even i ∈ {2, 4, . . . , d − 1}. • Si := B(Pi−1 ) for odd i ∈ {3, 5, . . . , d}. We now show that these paths form a wheel. The first condition of the wheel (Definition 1) is trivially satisfied. As none of the integral paths in y contain multiple Pi ’s, the second and third conditions are satisfied. We can see that the fourth condition is satisfied as follows. • S1 ◦ H1 ◦ S2−1 = R ◦ P10 ◦ F (P2 )−1 = P1 ◦ F (P2 )−1 ∈ F. −1 • Si ◦ Hi ◦ Si+1 = F (Pi ) ◦ Pi ◦ B(Pi )−1 ∈ F for even i ∈ {2, 4, . . . , d − 1}. −1 • Si ◦ Hi ◦ Si+1 = B(Pi−1 ) ◦ Pi ◦ F (Pi+1 )−1 ∈ F for odd i ∈ {3, 5, . . . , d − 2}. −1 = T (P ) ◦ T (Q)−1 ∈ F. • Sd ◦ Hd ◦ S1−1 = B(Pd−1 ) ◦ Pd ◦ Q0−1 1 ◦R Thus, we can obtain a basic F-packing of size |y| + 12 by removing the (d − 1)/2 integral paths intersecting P and by inserting the wheel of degree d. Finally, we consider the case when (P, Q) satisfies the condition 3b. Note that p must be odd from the condition 3 of y-augmenting pairs (Definition 4). Let d := p. We define paths {H1 , . . . , Hd } and {S1 , . . . , Sd } as follows (Figure 8 (b)). • Hi := Pi+1 for i ∈ {1, . . . , d − 1}. • Hd := Q−1 2 . • S1 := P1 . • Si := B(Pi ) for even i ∈ {2, 4, . . . , d − 1}. • Si := F (Pi+1 ) for odd i ∈ {3, 5, . . . , d − 2}. • Sd := B(Q2 ). We now show that these paths form a wheel. The first condition of the wheel is trivially satisfied. The second and third conditions are satisfied because none of the integral paths in y contain multiple Pi ’s. We can see that the fourth condition is satisfied as follows. • S1 ◦ H1 ◦ S2−1 = P1 ◦ P2 ◦ B(P2 )−1 = P1 ◦ (B(P2 ) ◦ P2−1 )−1 = Q1 ◦ F (Q2 )−1 ∈ F. −1 • Si ◦ Hi ◦ Si+1 = B(Pi ) ◦ Pi+1 ◦ F (Pi+2 )−1 ∈ F for even i ∈ {2, 4, . . . , d − 3}. −1 • Si ◦ Hi ◦ Si+1 = F (Pi+1 ) ◦ Pi+1 ◦ B(Pi+1 )−1 ∈ F for odd i ∈ {3, 5, . . . , d − 2}. • Sd−1 ◦ Hd−1 ◦ Sd−1 = B(Pd−1 ) ◦ Pd ◦ B(Q2 )−1 = T (P ) ◦ T (Q)−1 ∈ F. −1 −1 • Sd ◦ Hd ◦ S1−1 = B(Q2 ) ◦ Q−1 2 ◦ P1 = F (P2 ) ◦ P1 ∈ F. Thus, we can obtain a basic F-packing of size |y| + 21 by removing the (d − 1)/2 integral paths intersecting P and by inserting the wheel of degree d. 26 𝑃𝑏+1 𝑃𝑎−1 𝑃𝑏+1 𝑃𝑏 𝑃𝑎 𝑃𝑎−1 𝐹𝑦 (𝑃𝑏 ) 𝑃𝑏−1 𝐵𝑦 (𝑃𝑎 ) 𝐴 𝑃𝑎+1 𝑃𝑏−2 𝑃𝑎+2 𝑊 𝑃𝑏−1 𝐴 𝑃𝑎+1 𝐴 𝑃𝑏−2 𝑃𝑎+2 𝑦 𝑦′ 𝑦ത Figure 9: Applying Lemma 11. Next, we provide two lemmas for weakening the assumptions in the abovementioned lemma. Lemma 11. Given a basic F-packing y and a y-augmenting pair (P = P1 ◦· · ·◦Pp , Q = Q1 ◦· · ·◦Qq ) satisfying all the following conditions, a basic F-packing of size |y| + 12 can be constructed in linear time. 1. None of the Pi ’s are contained in the spokes. 2’. Any two segments of P contained in the same integral path have the same direction. 3. One of the following two conditions is satisfied: (a) p ≥ q = 1 or (b) p ≥ q = 2, P1 = Q1 , and P2 and Q2 are contained in an integral path in the opposite direction. Proof. Note that the conditions 1 and 3 are the same as those in Lemma 10. Besides, if the condition 3b is satisfied, from the condition 5b of y-alternating paths (Definition 2), the integral path containing P2 and Q2 never contain any other segments. We call a segment of P obstructive if it is contained in an integral path containing multiple segments of P . If there exist no obstructive segments, the condition 2 of Lemma 10 is satisfied. Thus, we can obtain a basic F-packing of size |y| + 12 by applying Lemma 10. We repeat the following process while obstructive segments exist. Let Pa be the first obstructive segment and Pb (b > a) be the next segment contained in the integral path containing Pa . From the condition 5b of the y-alternating paths (Definition 2), Pb is contained in F (Pa ), and B(Pa−2 ) ◦ Pa−1 ≡ B(Pa ) ◦ Pa−1 holds. We construct a basic F-packing ȳ of the same size and a ȳ-augmenting pair (P̄ , Q) such that the precondition of this lemma is satisfied and the number of obstructive segments strictly decreases as follows (see Figure 9). First, we construct a basic F-packing y 0 of size |y| − 1 by removing the integral path containing Pa and Pb from y. Observe that P 0 := (By (Pa ) ◦ Pa+1 ) ◦ Pa+2 ◦ · · · ◦ Pb−1 is a y 0 -alternating path. Then, by applying Corollary 4 against y 0 and P 0 , we obtain a basic F-packing y 00 of size |y|−1 and a single-segment y 00 -alternating path P 00 satisfying P 00 ≡ Ty0 (P 0 ) = By0 (Pb−2 )◦Pb−1 = By (Pb−2 )◦Pb−1 . From the condition 5b of the y-alternating path P , By (Pb−2 )◦Pb−1 6≡ Fy (Pb ) holds. From the choice of Pb , Fy (Pb ) is internally disjoint from P 00 . Thus, we can obtain a basic F-packing ȳ of size |y| 27 by introducing a new integral path P 00 ◦ Fy (Pb ). Let W be the path from s(Pa ) to t(Pb ) along the integral path and let P̄ := P1 ◦ · · · ◦ Pa−2 ◦ (Pa−1 ◦ W ◦ Pb+1 ) ◦ Pb+2 ◦ · · · ◦ Pp . Finally, we prove that (P̄ , Q) is a ȳ-augmenting pair satisfying the preconditions of this lemma. From the construction of ȳ and P̄ , the conditions 1–4 of the ȳ-alternating paths (Definition 2) are clearly satisfied. The condition 5b is satisfied because Bȳ (Pa−2 ) ◦ (Pa−1 ◦ W ◦ Pb+1 ) ≡ By (Pa−2 ) ◦ Pa−1 ◦ W ◦ Pb+1 ≡ By (Pa ) ◦ Pa−1 ◦ W ◦ Pb+1 ≡ By (Pb ) ◦ Pb+1 holds. Thus, P̄ is a ȳ-alternating path. Because Tȳ (P̄ ) ≡ Ty (P ) 6= Ty (Q) ≡ Tȳ (Q) holds and (P̄ , Q) satisfies the precondition 3 of the lemma, (P̄ , Q) is a ȳ-augmenting pair. Because no segments of P̄ are newly contained in the spokes in ȳ, the condition 1 of the lemma is satisfied. From the choice of Pa , for any even i ∈ {a+2, . . . , b−2}, By0 (Pi ) contains no segments from {P2 , . . . , Pa−2 }. Therefore, no two segments of P̄ are newly contained in a same integral path in ȳ. Thus, the condition 2’ is satisfied. When (P, Q) satisfies the condition 3a, (P̄ , Q) also satisfies the condition 3a. When (P, Q) satisfies the condition 3b, (P̄ , Q) also satisfies the condition 3b because the integral path containing P2 and Q2 remains in ȳ. Thus, all the conditions in the lemma are satisfied. We can find the pair (Pa , Pb ) by gradually increasing an index i, which is not reset during the repetition, and by searching for Pj contained in F (Pi ) by traversing the integral path. Therefore, each edge is traversed at most once through the whole process. Thus, the total running time is linear in the graph size. Lemma 12. Given a basic F-packing y and a y-augmenting pair (P = P1 ◦· · ·◦Pp , Q = Q1 ◦· · ·◦Qq ), either of a y-augmenting path or a y-augmenting pair (P̄ , Q̄) satisfying all the following conditions can be constructed in linear time. 1. All the segments of Q̄, except for the last one, are contained in P̄ . −1 −1 0 2. P̄ can be written as P̄ = P ◦ Q−1 q ◦ Qq−1 ◦ · · · ◦ Qq 0 for some q . 3. The common prefix of (P̄ , Q̄) contains the common prefix of (P, Q). −1 4. The following two conditions are satisfied for the new segments S := {Q−1 q , . . . , Qq 0 } of P̄ : (a) no segments in S are contained in the spokes, and (b) any two segments in S contained in the same integral path have the same direction. Proof. Initially, the conditions 2–4 are trivially satisfied. We repeat the following process, which preserves these conditions. The condition 1 is satisfied when q becomes one or Qq−1 gets contained in P . 0 (Case 1) If both of p and q are odd, we update P 0 ← P1 ◦ · · · ◦ Pp−1 ◦ (Pp ◦ Q−1 q ) and Q ← 0 0 −1 Q1 ◦ · · · ◦ Qq−1 . Because Pp and Qq share no edges and because T (P ) ◦ T (Q ) = (B(Pp−1 ) ◦ −1 = (B(P −1 = T (P ) ◦ T (Q)−1 holds, (P 0 , Q0 ) is (Pp ◦ Q−1 p−1 ) ◦ Pp ) ◦ (B(Qq−1 ) ◦ Qq ) q )) ◦ B(Qq−1 ) a y-augmenting pair. 0 (Case 2) If p is even and q is odd, we update P 0 ← P1 ◦· · ·◦Pp−1 ◦Pp ◦Q−1 q and Q ← Q1 ◦· · ·◦Qq−1 . −1 = Because Pp and Qq share no edges and because T (P 0 ) ◦ T (Q0 )−1 = (B(Pp ) ◦ Q−1 q ) ◦ B(Qq−1 ) −1 −1 0 0 B(Pp ) ◦ (B(Qq−1 ) ◦ Qq ) = T (P ) ◦ T (Q) holds, (P , Q ) is a y-augmenting pair. 28 𝑃𝑘−1 𝑄𝑞−1 𝑄𝑞 𝑃𝑘−1 𝑊 𝑃𝑘 𝑃𝑘+1 𝑃𝑝 𝑃𝑘−1 𝑄𝑞−1 𝑄𝑞 𝑃𝑘−1 𝑃𝑘 𝑃𝑘+1 𝑊 𝑃𝑘+1 𝑃𝑝 𝑃𝑘−1 𝑃𝑘 𝑄𝑞−1 𝑄𝑞 𝑃𝑝 𝑃𝑘+1 𝑃𝑘 𝑃𝑘−1 𝑃𝑘 𝑃𝑘+1 𝑊 𝑃𝑝 𝑃𝑝 (Case 3) 𝑃𝑘+1 𝑃𝑘 (Case 4 (1)) 𝑃𝑝 (Case 4 (2)) Figure 10: Applying Lemma 12. (Case 3) If p is odd, q is even, and Qq is contained in a spoke, then we search for the nearest segment of P contained in F (Qq ) by traversing the spoke. Note that none of the other segments of Q are contained in F (Qq ) from the condition 5c of the y-alternating paths (Definition 2). Moreover, each segment of P is fully contained in F (Qq ) or internally disjoint from F (Qq ) because Qq shares no edges with P . Thus, each edge is traversed at most once through the entire process, and the total running time of this part is linear in the graph size. If none of the Pi ’s are contained in F (Qq ), then P is a y-augmenting path because t(P ) is contained in the spoke, and T (P ) 6≡ T (Q) = B(t(P )) holds. Otherwise, let Pk be the nearest segment of P contained in F (Qq ); let W be the path from s(Pk ) to t(P ) along the spoke; and let Q̄ := P1 ◦ · · · ◦ Pk−1 ◦ W (Figure 10 (left)). Because T (Q̄) = B(W ) = B(Qq ) = T (Q) holds, (P, Q̄) is a y-augmenting pair. Note that this finishes the process. (Case 4) If p is odd, q is even, and Qq is contained in an integral path, then we search for the nearest segment of P or Q contained in F (Qq ) by traversing the integral path. Because each edge is traversed at most twice (in two directions) through the entire process, the total running time of this part is linear in the graph size. Note that from the condition 5b of the y-alternating path Q (Definition 2), F (Qq ) cannot contain a segment Qk such that Qk has the same direction as Qq or has the opposite direction as Qq and B(Qk−2 ) ◦ Qk−1 6≡ B(Qk ) ◦ Q−1 k holds. If the nearest segment is Pk such that Pk has the same direction as Qq or has the opposite direction as Qq and B(Pk−2 ) ◦ Pk−1 6≡ B(Pk ) ◦ Pk−1 holds, let W be the path from s(Pk ) to t(P ) along the integral path and Q̄ := P1 ◦ · · · ◦ Pk−1 ◦ W (see Figure 10 (right)). Because T (Q̄) = B(W ) = B(Qq ) = T (Q) holds, (P, Q̄) is a y-augmenting pair, and we then finish the process. 29 Otherwise (i.e., (a) no segments are contained in F (Qq ), (b) the nearest segment is Qk that has the opposite direction as Qq and B(Qk−2 ) ◦ Qk−1 ≡ B(Qk ) ◦ Q−1 k holds, or (c) the nearest segment is Pk that has the opposite direction as Qq and B(Pk−2 ) ◦ Pk−1 ≡ B(Pk ) ◦ Pk−1 holds), we update 0 P 0 ← P1 ◦ · · · ◦ Pp ◦ Q−1 q and Q ← Q1 ◦ · · · ◦ Qq−1 . Note that in this case, from the condition 5b of the y-alternating paths (Definition 2) and the condition 4 of y-augmenting pairs (Definition 4), conditions (b) and (c) hold not only against the nearest segment Qk or Pk , but also against any segments contained in F (Qq ). First, we prove that P 0 is a y-alternating path. The first four conditions of the y-alternating paths (Definition 2) are clearly satisfied. The condition 5 is satisfied for i = p + 1 because B(Pp−1 ) ◦ Pp = T (P ) 6≡ T (Q) = B(Qq ) = F (Q−1 q ) holds. For checking the condition 5 against the other Pi ’s, it suffices to show that for any segment Pi contained in the same integral path as Qq , it holds that B(Pi−2 ) ◦ Pi−1 ≡ B(Pi ) ◦ Pi−1 , and Qq is contained in F (Pi ). Let Pi be a segment contained in B(Qq ). From the condition 4 of the y-augmenting pairs (Definition 4), Pi has the same direction as Qq . Thus, Qq is contained in F (Pi ). Therefore, from the condition 4 again, B(Pi−2 ) ◦ Pi−1 ≡ B(Pi ) ◦ Pi−1 holds. Let Pi be a segment contained in F (Qq ). As we have discussed earlier, B(Pi−2 ) ◦ Pi−1 ≡ B(Pi ) ◦ Pi−1 then holds, and Pi has the opposite direction as Qq , implying that Qq is contained in F (Pi ). Next, we prove that (P 0 , Q0 ) is a y-augmenting pair. Conditions 1 and 3 of the y-augmenting pairs (Definition 4) are clearly satisfied. Because T (P 0 ) = B(Q−1 q ) = F (Qq ) 6≡ B(Qq−2 ) ◦ Qq−1 = 0 T (Q ) holds, the condition 2 is satisfied. For checking the condition 4, it suffices to show that (1) none of the Qj ’s with j < q are contained in B(Q−1 q ) = F (Qq ) in the same direction as Qq ; (2) −1 )−1 = F (Q ) ◦ Q , none of the Q ’s with j < q are contained in if B(Pp−1 ) ◦ Pp 6≡ B(Q−1 ) ◦ (Q q q j q q F (Q−1 q ) = B(Qq ) in the opposite direction as Qq ; (3) for any Qj contained in an integral path, Qq is not contained in B(Qj ) in the same direction as Qj ; and (4) if B(Qj−2 ) ◦ Qj−1 6≡ B(Qj ) ◦ Q−1 j holds, Qq is not contained in F (Qj ) in the opposite direction as Qj . All these conditions directly follow from the condition 5b of the y-alternating path Q (Definition 2). Finally, we prove that (P 0 , Q0 ) satisfies the conditions 2–4 of this lemma. Conditions 2 and 3 are clearly satisfied. From the condition 5b of the initial y-alternating path Q, B(Qq ) contains none of the new segments S. Thus, if a segment in S is contained in the same integral path as Qq , it must be contained in F (Qq ). Therefore, it has the same direction as Q−1 q . Finally, we prove Lemma 3 by combining Lemmas 11 and 12. Proof of Lemma 3. First, we apply Lemma 12 against (Q, P ). If we obtain a y-augmenting path, we obtain a basic F-packing of size |y| + 21 by applying Lemma 2. Otherwise, we obtain an updated y-augmenting pair (Q̄, P̄ ) such that all the segments of P̄ , except for the last one, are contained in the common prefix. Next, we apply Lemma 12 against (P̄ , Q̄). We finish if we obtain a y-augmenting path; otherwise, we obtain an updated y-augmenting pair (P̂ , Q̂). Let P̂ := P̂1 ◦· · ·◦ P̂p̂ and Q̂ := Q̂1 ◦· · ·◦ Q̂q̂ . Then, Q̂q̂−1 is contained in P̂q̂−1 . Note that this implies that P̂i = Q̂i holds for any i ∈ {1, . . . , q̂ − 2}. Moreover, P̂q̂−1 = Q̂q̂−1 also holds if q̂ is even. We have q̂ ≥ p̄ − 1 because the common prefix of (P̂ , Q̂) contains the common prefix of (P̄ , Q̄) and because P̂p̄−1 is contained in the common prefix of (P̄ , Q̄). Let us assume that q̂ = p̄ − 1 holds, and q̂ is odd. In this case, P̂q̂ is contained in Q̂q̂ . Therefore, P̂q̂ = Q̂q̂ holds. Then, from the condition 3 of the y-augmenting pair (P̂ , Q̂) (Definition 4), p̂ must be even, and P̂p̂ must be contained in the same integral path as P̂p̄ . However, 30 this violates the condition 5b of the y-alternating path P̂ because B(P̂p̄−2 )◦ P̂p̄−1 = T (Q̂) 6≡ T (P̂ ) = B(P̂p̂ ) = B(P̂p̄ ) ◦ P̂p̄−1 holds. Thus, q̂ ≥ p̄ holds if q̂ is odd. We consider two cases. If q̂ is odd or P̂q̂ and Q̂q̂ have the same direction, let r be the maximum even integer at most q̂. We apply Corollary 3 independently against P̂ and Q̂ and obtain a basic F-packing y 0 of the same size and y 0 -alternating paths P 0 and Q0 . Here, P̂i = Q̂i holds for any i ∈ {1, . . . , r − 1}, and F (P̂r ) = F (Q̂r ) holds. Hence, the basic F-packing obtained by these two applications are the same. We then, thus, use the same symbol y 0 . Because Ty0 (P 0 ) ≡ Ty (P̂ ) 6≡ Ty (Q̂) ≡ Ty0 (Q0 ) holds, (P 0 , Q0 ) is a y 0 -augmenting pair. We now show that (P 0 , Q0 ) satisfies the conditions in Lemma 11. Because Q0 is single-segment, the third condition is satisfied. We have r + 2 = q̂ + 1 > p̄ when q̂ is odd, and we have r + 2 = q̂ + 2 > p̄ when q̂ is even. Therefore, from the condition 4 of Lemma 12, the first and the second conditions hold. Thus, we obtain a basic F-packing of size |y| + 12 by applying Lemma 11 against (P 0 , Q0 ). If q̂ is even and P̂q̂ and Q̂q̂ have the opposite direction, let r := q̂ − 2. We apply Corollary 3 against independently P̂ and Q̂ and obtain a basic F-packing y 0 of the same size and y 0 -alternating paths P 0 and Q0 . Here, P̂i = Q̂i holds for any i ∈ {1, . . . , r}. Hence, the basic F-packings obtained by these two applications are the same. We, thus, use the same symbol y 0 . Because Ty0 (P 0 ) ≡ Ty (P̂ ) 6≡ Ty (Q̂) ≡ Ty0 (Q0 ) holds, (P 0 , Q0 ) is a y 0 -augmenting pair. We now show that (P 0 , Q0 ) satisfies the conditions in Lemma 11. Because we set r := q̂ − 2, the number of segments of Q0 is two. Because P̂q̂−1 = Q̂q̂−1 holds, the first segments of P 0 and Q0 are the same. Because P̂q̂ and Q̂q̂ have the opposite direction, the second segments of P 0 and Q0 have the opposite direction. Thus, the third condition is satisfied. Note that in this case, P̂q̂ and Q̂q̂ are contained in an integral path and, from the condition 5b of the y-alternating paths (Definition 2), the integral path containing P̂q̂ and Q̂q̂ never contain any other segments. Moreover, because r + 4 = q̂ + 2 > p̄, the first and second conditions follow from the condition 4 of Lemma 12. Thus, we obtain a basic F-packing of size |y| + 12 by applying Lemma 11 against (P 0 , Q0 ). 4 Farthest Cover This section provides an algorithm for computing a farthest minimum half-integral F-cover and proves the following theorem. Theorem 3. Let C be a set of 0/1/all constraints on variables V and ϕA be a partial assignment for a subset A ⊆ V . Given the primal graph of C, the set A, an incremental-test oracle for (C, ϕA ), and an integer k, we can compute a pair of a farthest minimum half-integral FC,ϕA -cover x and a maximum half-integral FC,ϕA -packing y with |x| = |y| ≤ k2 or correctly conclude that the size of the minimum half-integral FC,ϕA -cover is at least k+1 2 in O(kmT ) time, where m is the number of constraints, and T is the running time of the incremental-test oracle. We use the following structure of a minimum half-integral F-cover. Lemma 13. The following holds for any maximum basic F-packing y and any minimum halfintegral F-cover x. 1. x(V (I)) = 1 for any integral path I of y. 2. x(V (S)) = 1 2 for any spoke S of y. 31 Proof. For a wheel W , we denote the degree of W by d(W ) and the set of vertices contained in W by ) and V (W ). Note that V (W ) is not a multiset. First, we prove that x(V (W )) is always at least d(W 2 1 if the equality holds, then x(V (S)) = 2 holds for any spoke S of W . Let W be a wheel with a halfintegral cycle H1 ◦ · · · ◦ Hd and spokes {S1 , . . . , Sd }. We define H := V (W ) \ (V (S1 ) ∪ . . . ∪ V (Sd )). We then have x(V (W )) = x(H) + d X x(V (Si )) i=1 d 1X 1 −1 x(V (Si ◦ Hi ◦ Si+1 )) = x(H) + 2 2 i=1 ≥ 1 2 d X 1 i=1 d = . 2 If x(V (W )) = d2 holds, we have x(H) = 0 and x(V (Si )) + x(V (Si+1 )) = 1 for any i ∈ {1, . . . , d}. Because d is odd, this implies that x(V (Si )) = 21 holds for any i ∈ {1, . . . , d}. We now prove the lemma. Because all the integral paths and wheels do not share any vertices, we have X X |x| ≥ x(V (I)) + x(V (W )) I:integral path X ≥ W :wheel 1+ I:integral path X W :wheel d(W ) 2 = |y| = |x|. Therefore, x(V (I)) = 1 and x(V (W )) = W. d(W ) 2 must hold for any integral path I and any wheel Fix an arbitrary maximum basic F-packing. From the above lemma, for every minimum halfintegral F-cover x, we can construct unique indices ax and bx satisfying the following. 1. For every integral path I = (v0 , . . . , v` ), x(vax (I) ) ≥ 21 , x(vbx (I) ) ≥ 2. For every spoke S = (v0 , . . . , v` ), x(vax (S) ) = 1 2 1 2 and ax (I) ≤ bx (I) hold. holds. Therefore, we obtain the following corollary. Corollary 5. A minimum half-integral F-cover x0 dominates a minimum half-integral F-cover x if and only if the following conditions hold. 1. For every integral path I, ax (I) ≤ ax0 (I) ≤ bx0 (I) ≤ bx (I) holds. 2. For every spoke S, ax (S) ≤ ax0 (S) holds. 32 Algorithm 2 Algorithm for computing a farthest minimum half-integral F-cover 1: Compute a maximum basic F-packing y. 2: for each integral path I = (v0 , . . . , v` ) of y do 3: while a(I) < b(I) do 4: Contract (v0 , . . . , va(I)+1 ). 5: if there exists a y-augmenting path or pair then 6: Rewind the search and contraction of this iteration; break 7: while a(I) < b(I) do 8: Contract (v` , . . . , vb(I)−1 ). 9: if there exists a y-augmenting path or pair then 10: Rewind the search and contraction of this iteration; break 11: for each spoke S = (v0 , . . . , v` ) of y do 12: while a(S) < ` do 13: Contract (v0 , . . . , va(I)+1 ). 14: if there exists a y-augmenting path or pair then 15: Rewind the search and contraction of this iteration; break 16: return the minimum half-integral F-cover constructed from the current tables. We use the following operation in our algorithm. Let P be an implicational walk ending at a vertex t. First, we create a new vertex t0 and introduce a constraint ϕ(t) = ϕ(t0 ) (along with a new edge tt0 ). We then insert t0 into A and set ϕA (t0 ) = imp(P ). We call this operation as contracting P. Lemma 14. Let P be an implicational walk ending at a vertex t and let F 0 be the set of conflicting walks after contracting P . Then, any half-integral F 0 -cover x is also a half-integral F-cover, and any half-integral F-cover x satisfying x(V (P ) \ {t}) = 0 is also a half-integral F 0 -cover. Proof. The first claim is trivial because F ⊆ F 0 . We now prove the second claim. Suppose that x is not a half-integral F 0 -cover. Then, there exists a walk W ∈ F 0 with x(V (W )) < 1. Because x is a half-integral F-cover, W must contain the edge t0 t. Therefore, we can write W = t0 t ◦ Q−1 for some implicational walk Q with imp(Q) 6= imp(P ). We then have x(V (P ◦ Q−1 )) = x(V (W )) < 1 and P ◦ Q−1 ∈ F, which contradicts the fact that x is a half-integral F-cover. We now describe the algorithm for computing a farthest minimum half-integral F-cover (see Algorithm 2). We iteratively apply the contraction in the algorithm. We denote the current set of the conflicting walks by F and the original set by Forig . First, we compute a maximum basic F-packing y using the algorithm in Section 3. We keep and reuse the tables (a, b, prev, and tail) used in the last execution of Algorithm 1, which returned NO. We process the integral paths and the spokes of y one by one, whose detail will be described later, by preserving the following invariants. Lemma 15. The following invariants hold at any step of Algorithm 2. 1. y is a maximum basic F-packing. 2. Let x be the minimum half-integral F-cover constructed from the current tables as described in Section 3.3 (i.e., ax = a and bx = b hold). Then, any minimum half-integral Forig -cover dominating x is also a minimum half-integral F-cover dominating x. 33 3. For any processed integral path I, there exists no minimum half-integral F-cover x0 satisfying a(I) < ax0 (I) ≤ bx0 (I) ≤ b(I) or a(I) ≤ ax0 (I) ≤ bx0 (I) < b(I). 4. For any processed spoke S, there exists no minimum half-integral F-cover x0 satisfying a(S) < ax0 (S). When all the integral paths and the spokes are processed, we return the minimum half-integral F-cover x constructed from the current tables. We can easily prove the correctness of the algorithm from these invariants. Lemma 16. When all the integral paths and the spokes are processed, the minimum half-integral F-cover x of G constructed from the current tables is a farthest minimum half-integral Forig -cover. Proof. From the invariants 3 and 4 and Corollary 5, x is a farthest minimum half-integral F-cover. Because Forig ⊆ F and from the invariant 1, x is also a minimum half-integral Forig -cover. Suppose that there exists a minimum half-integral Forig -cover x0 dominating x. Then, from the invariant 2, x0 is a minimum half-integral F-cover dominating x, which contradicts the fact that x is a farthest minimum half-integral F-cover. We now describe how to process the integral paths and spokes. For each integral path I = (v0 , . . . , v` ), we first repeat the following while a(I) < b(I) holds (lines 3–6). Let i := a(I) + 1. We first contract the implicational path (v0 , . . . , vi ). When using the incremental-test oracle, this operation can be done in a constant time by setting I(vi0 ) = A∗ ((v0 , . . . , vi )), which has been precomputed. We then search for a y-augmenting path or pair using Algorithm 1. Instead of searching for a y-augmenting path/pair from scratch by initializing the tables, we restart the search from line 4 of Algorithm 1 by setting s ← vi0 and reusing the current tables. If the restarted search returns NO, we keep the tables updated by the restarted search and continue the repetition. Because vi0 vi 6≡ (v` , . . . , vi ) holds, we have a(I) ≥ i after the search. If the restarted search finds a y-augmenting path or pair, we rewind all the changes in this step (i.e., we restore the tables to the state before the search and remove the edge vi0 vi inserted by the contraction), and then exit the repetition. Note that we do not rewind the changes in the previous steps where the restarted searches returned NO. Claim 5. The restarted search can correctly compute a y-augmenting path or pair if exists. Proof. We can virtually think as follows. Because the order of A \ V (y) at line 3 of Algorithm 1 is arbitrary, we can choose completely the same order as the one used in the last failed search that has constructed the current tables. Note that, from the invariant 3b in Lemma 5, vi was not visited in the last execution. Therefore, the insertion of the edge vi0 vi does not affect the search at all. In the end, the execution reaches to the final iteration of the while loop with s = vi0 , and all the tables are completely the same as the current tables. Thus, instead of running Algorithm 1 from scratch, we can use the restarted search. Claim 6. Lines 3–6 preserve all the invariants. Proof. Because we keep the changes only when the search fails, the invariant 1 is satisfied. The invariant 2 follows from Lemma 14 against P := (v0 , . . . , vi ). The invariants 3 and 4 are trivial. Note that the current integral path I is still under processing. 34 Claim 7. When the repetition of lines 3–6 ends, there exists no minimum half-integral F-cover x0 satisfying a(I) < ax0 (I) ≤ bx0 (I) ≤ b(I). Proof. The repetition ends when a(I) becomes equal to b(I) or when a y-augmenting path or pair is found. The former case is trivial. In the latter case, let F 0 be the set of the conflicting walks after contracting (v0 , . . . , va(I)+1 ). Suppose that there exists a minimum half-integral Fcover x0 satisfying the condition in the lemma. Then, from Lemma 14, x0 is also a half-integral F 0 -cover. In contrast, the size of the maximum half-integral F 0 -packing y 0 is strictly larger than |y| because a y-augmenting path or pair is found. Thus, we have |y 0 | > |y| = |x0 | ≥ |y 0 |, which is a contradiction. After repeating lines 3–6, we repeat the following while a(I) < b(I) holds (lines 7–10). Let i := b(I) − 1. We first contract the implicational path (v` , . . . , vi ), and then we restart the search by setting s ← vi0 . If the restarted search fails, we continue the repetition, and if the restarted search succeeds, we rewind all the changes in this step and exit the repetition. By the same argument as in the case of lines 3–6, all the invariants are preserved and, when the repetition ends, there exists no minimum half-integral F-cover x0 satisfying a(I) ≤ ax0 (I) ≤ bx0 (I) < b(I). Thus, the invariant 3 is satisfied for I when we have finished processing I. Next, we repeat the following for each spoke S = (v0 , . . . , v` ) while a(S) < ` holds (lines 12–15). Let i := a(S) + 1. We first contract the implicational path (v0 , . . . , vi ), and then restart the search by setting s ← vi0 . We continue the repetition if the restarted search fails, and, otherwise, we rewind all the changes in this step and exit the repetition. By the same argument as in the case of integral paths, all the invariants are preserved and, when the repetition ends, there exists no minimum half-integral F-cover x0 satisfying a(S) < ax0 (S). Thus, the invariant 4 is satisfied for S when we have finished processing S. Finally, we analyze the running time of Algorithm 2. The number of inserted edges is O(n), and the number of while loops (lines 3–6, 7–10, or 12–15) is 2k. For each while loop, a series of the restarted searches can be regarded as a single execution of Algorithm 1, which takes O(mT ) time. Therefore, the total running time is O(kmT ). Thus, we obtain Theorem 3. 5 5.1 Linear-Time FPT Algorithms Algorithm for 0/1/all Deletion This section proposes an O(d2k km)-time algorithm for 0/1/all Deletion. Our algorithm is based on the branch-and-bound algorithm in [19]. We exploit the farthest minimum half-integral F-cover and parallel unit-propagation to obtain a linear-time FPT algorithm. Let I = (C, ϕA ) be a pair of 0/1/all constraints C on a variable set V and a partial-assignment ϕA on a subset A ⊆ V . We denote FC,ϕA by FI . For a variable u ∈ V \ A and an element a ∈ D(u), we denote by I[u ← a] a pair (C, ϕA∪{u} ) such that ϕA∪{u} (u) = a. We denote by I − u a pair (C[V \ {u}], ϕA\{u} ) for a variable u ∈ V , where ϕA\{u} is the restriction of ϕA to A \ {u}. We call this operation deleting u. Let Nu := {v ∈ A | uv ∈ E, ϕA does not satisfy Cuv } for a variable u ∈ A. We denote by I/u a pair (C[V \ ({u} ∪ Nu )], ϕA0 ) such that A0 = (A ∪ {v | Cuv (ϕA (u)) 6= all}) \ ({u} ∪ Nu ) and ϕA0 (v) = Cuv (ϕA (u)) for v ∈ A0 \ A. We call this operation fixing u. We can obtain an incremental-test oracle for I/u by setting I(v) = A(I(u), uv) for each v ∈ A0 \ A. We can observe the following. 35 Lemma 17. The following holds for any pair I = (C, ϕA ) and u ∈ V . 1. I admits a deletion set of size k containing u if and only if I − u admits a deletion set of size k − 1. 2. For any half-integral FI−u -cover x0 , the following function x is a half-integral FI -cover: x(u) = 1 and x(v) = x0 (v) for v ∈ V \ {u}. Proof. First, we prove the first claim. For a deletion set X for I containing u, X \ {u} is a deletion set for I − u. For a deletion set X 0 for I − u, X 0 ∪ {u} is a deletion set for I. Next, we prove the second claim. Let W be a walk in FI . We have x(V (W )) ≥ x0 (V (W )) ≥ 1 if W ∈ FI−u . Otherwise, W visits u; therefore, we have x(V (W )) ≥ x(u) = 1. Lemma 18. The following holds for any pair I = (C, ϕA ) and u ∈ A. 1. I admits a deletion set of size k not containing u if and only if I/u admits a deletion set of size k − |Nu |. 2. For any half-integral FI/u -cover x0 , the following function x is a half-integral FI -cover: x(u) = 0, x(v) = 1 for v ∈ Nu , and x(v) = x0 (v) for v ∈ V \ ({u} ∪ Nu ). Proof. First, we prove the first claim. Let X be a deletion set for I not containing u. Because X must contain all of Nu , X \ Nu is a deletion set for I/u of size |X| − |Nu |. For a deletion set X 0 for I/u, X 0 ∪ Nu is a deletion set for I. Next, we prove the second claim. Let W be a walk in FI . If W ∈ FI/u , we have x(V (W )) ≥ x0 (V (W )) ≥ 1. If W visits a vertex in Nu , we have x(V (W )) ≥ 1. Otherwise, we can write W = uv ◦ W 0 . Then, we have W 0 ∈ FI/u , and thus, we have x(V (W )) ≥ x0 (V (W 0 )) ≥ 1. For a minimum half-integral FI -cover x, we define an operation called a persistency reduction as follows. We first delete every vertex in x−1 (1) in an arbitrary order, and then fix every vertex in R(x) in an arbitrary order16 . Because x is a half-integral FI -cover, for any implicational walk W with x(V (W ) \ {t(W )}) = 0 and x(t(W )) ≤ 12 , the value imp(W ) depends only on t(W ). Therefore, the ordering does not affect the result and Nu = ∅ for every fixing. We denote the obtained pair (C[V \ (x−1 (1) ∪ R(x))], ϕA0 ) by I/x. Lemma 19. I admits a deletion set of size k if and only if I/x admits a deletion set of size k − |x−1 (1)|. Proof. From Theorem 1, there exists a minimum deletion set X for I such that x−1 (1) ⊆ X ⊆ V \ R(x). Therefore, from Lemma 17 and 18, the claim holds. Lemma 20. For a farthest minimum half-integral FI -cover x, the restriction of x to V \ (x−1 (1) ∪ R(x)) is the unique minimum half-integral FI/x -cover. Proof. Let x0 be the restriction of x. Then, x0 is a half-integral FI/x -cover because x0 (u) = 12 for every u ∈ A0 . Suppose that there exists a half-integral FI/x -cover z 0 with |z 0 | ≤ |x0 | and z 0 6= x0 . From Lemmas 17 and 18, the following function z is a half-integral F-cover: z(u) = z 0 (u) for u ∈ V \ (x−1 (1) ∪ R(x)), z(u) = 1 for u ∈ x−1 (1), z(u) = 0 for u ∈ R(x). Then, we have |z| = |z 0 | + |x−1 (1)| ≤ |x0 | + |x−1 (1)| = |x|. Therefore, z is a minimum half-integral F-cover dominating x, which is a contradiction. 16 We pick an arbitrary vertex u ∈ R(x) ∩ A and fix u. This changes R(x) and A, and we repeat the process until R(x) becomes the empty set. 36 Algorithm 3 A linear-time FPT algorithm for 0/1/all Deletion 1: procedure Solve(I = (C, ϕA ), k) 2: Compute a farthest minimum half-integral FI -cover x. 3: if |x| > k then return false 4: I ← I/x, k ← k − |x−1 (1)|. 5: if V = ∅ then return true 6: if A 6= ∅ then 7: Pick a vertex u ∈ A. 8: return Solve(I − u, k − 1) ∨ Solve(I/u, k − |Nu |) 9: else 10: Choose a branching set B. 11: if there exists (u, a) ∈ B with FI[u←a] = ∅ then 12: I ← I − R and goto line 5, where R is the set of the implicated variables. 13: else W 14: return (u,a)∈B Solve(I[u ← a], k) Let I = (C, ϕ∅ ) be a pair with A = ∅. A set of pairs B ⊆ {(u, a) | u ∈ V, a ∈ D(u)} is called a branching set for I if it has the following property: any deletion set for I is a deletion set for at least one of I[u ← a] with (u, a) ∈ B. Note that any deletion set for I[u ← a] is a deletion set for I. The running time of our algorithm depends on the choice of branching sets. In general, we can use the following standard choice: pick a vertex u ∈ V and set B := {(u, a) | a ∈ D(u)}. In the next section, we choose different branching sets for problem-specific improvements. Lemma 21. B := {(u, a) | a ∈ D(u)} for some u ∈ V is a branching set. Proof. Let X be a deletion set for I, and let ϕ be a satisfying assignment for C[V \ X]. If u ∈ X, X is a deletion set for every I[u ← a]. Otherwise, X is a deletion set for I[u ← a] with a = ϕ(u). We now provide a linear-time FPT algorithm for 0/1/all Deletion (Algorithm 3). We denote the size of the minimum half-integral FI -cover for a pair I = (C, ϕA ) by c(I). Solve(I, k) is a procedure that returns true if and only if I admits a deletion set of size at most k. We prove the following. Theorem 4. Let I = (C, ϕA ) be a pair of a set C of 0/1/all constraints on a variable set V and a partial assignment ϕA for a subset A ⊆ V . We are given the primal graph of C, the set A, an O(T )-time incremental-test oracle for (C, ϕA ), and an integer k. Under the following assumptions, Algorithm 3 correctly answers whether (C, ϕA ) admits a deletion set of size at most k or not in O(max(2, b)2(k−c(I)) kmT ) time, where b is the integer in the assumption and m is the number of constraints. 1. For any V 0 ⊆ V , we can choose a branching set for (C[V 0 ], ϕ∅ ) of size at most b. 2. For any V 0 ⊆ V and any (u, a) ∈ B for a possible branching set B for (C[V 0 ], ϕ∅ ), we have an O(T )-time incremental-test oracle for (C[V 0 ], ϕ{u} ) with ϕ{u} (u) = a. Note that for 0/1/all Deletion, we assume that each constraint is given as a table of size O(d). Therefore, the naive implementation of the incremental-test oracle runs in a constant time. When 37 using the naive implementation of the incremental-test oracle, the second assumption trivially holds. Thus, by using the standard choice of branching sets, the algorithm runs in O(d2(k−c(I)) km) = O(d2k km) time. First, we compute a farthest minimum half-integral FI -cover x. Because the size of x is a lower bound on the size of the minimum deletion set, if |x| > k holds, there exists no deletion set of size at most k. Otherwise, we have k − c(I) ≥ 0. From Lemma 19, we can apply the persistency reduction and decrease k by |x−1 (1)|. This does not change the difference k − c(I) because c(I) also decreases by |x−1 (1)|. We rename the reduced instance to I = (C, ϕA ) for simplicity of the notation, and let G = (V, E) be the primal graph of C. We rename the restriction of x to V to x. From Lemma 20, x is the unique minimum half-integral FI -cover that maps every vertex in A to 1 2 and every other vertex to 0. This part can be done in O(kmT ) time and we can construct an incremental-test oracle for the new I from the incremental-test oracle for the old I. If A 6= ∅, we pick a vertex u ∈ A and return Solve(I − u, k − 1) ∨ Solve(I/u, k − |Nu |). Claim 8. I admits a deletion set of size at most k if and only if I − u admits a deletion set of size at most k − 1 or I/u admits a deletion set of size at most k − |Nu |. Moreover, both of k − 1 − c(I − u) < k − c(I) and k − |Nu | − c(I/u) < k − c(I) hold. Proof. The first claim follows from Lemma 17 and 18. We now prove the second claim. Suppose that I − u admits a half-integral FI−u -cover z 0 of size at most |x| − 1 = c(I) − 1. Then, from Lemma 17, we can construct a half-integral FI -cover z of size at most |x| with z(u) = 1. Therefore, z is a minimum half-integral FI -cover with x 6= z, which contradicts the uniqueness of x. Suppose that I/u admits a half-integral FI/u -cover z 0 of size at most |x| − |Nu | = c(I) − |Nu |. Then, from Lemma 18, we can construct a half-integral FI -cover z of size at most |x| with z(u) = 0. Therefore, z is a minimum half-integral FI -cover with x 6= z, which contradicts the uniqueness of x. If A = ∅, we choose a branching set B of size at most b. For each (u, a) ∈ B, we check whether FI[u←a] = ∅ or not by the unit-propagation (i.e., by applying Algorithm 1 against the empty packing). If there exists such (u, a), let R be the set of the implicated variables (i.e., R(0) for the empty cover 0 that always returns 0). Let m0 be the number of edges incident to a vertex in R, and suppose that (u, a) is the pair minimizing m0 . The unit-propagation for (u, a) takes only O(m0 T ) time. Hence, by running the unit-propagation for every (u, a) ∈ B in parallel, which is simulated in a single processor in a round-robin fashion, we can find such (u, a) in O(bm0 T ) time. We then set I ← I − R, where I − R is the pair obtained by deleting every vertex v ∈ R, and go back to line 5. Claim 9. I admits a deletion set of size at most k if and only if I − R admits a deletion set of size at most k. Proof. Any deletion set for I is also a deletion set for I − R. Let X be a deletion set for I − R, and let ϕV \(R∪X) be a satisfying assignment for C[V \ (R ∪ X)]. From the construction of R, C[R] admits a satisfying assignment ϕR , and every constraint Cvw with v ∈ R and w 6∈ R is a two-fan (ϕ(v) = a) ∨ (ϕ(w) = b) with a = ϕR (v). Therefore, an assignment ϕV \X such that ϕV \X (v) = ϕR (v) for v ∈ R and ϕV \X (v) = ϕV \(R∪X) (v) for v ∈ V \ (R ∪ X) satisfies C[V \ X]. Thus, X is also a deletion set for I. 38 If FI[u←a] 6= ∅ for all (u, a) ∈ B, we return W (u,a)∈B Solve(I[u ← a]). Claim 10. I admits a deletion set of size at most k if and only if at least one of I[u ← a] admits a deletion set of size k. Moreover, k − c(I[u ← a]) < k − c(I) for every (u, a) ∈ B. Proof. The first claim follows from the definition of the branching set. Because FI[u←a] 6= ∅, we have k − c(I[u ← a]) ≤ k − 21 < k = k − c(I). We now have proved the correctness of the algorithm. Finally, we analyze the running time. Let T (∆) be the running time of Solve(I, k) when ∆ := k − c(I). We can compute a farthest minimum half-integral FI -cover in O(kmT ) time. If A 6= ∅, we branch into two cases and ∆ decreases by at least 21 for each case. Therefore, we have T (∆) ≤ 2T (∆ − 12 ) + O(kmT ). If A = ∅, we search for (u, a) ∈ B with FI[u←a] = ∅ by the parallel unit-propagation. If there exists such (u, a), the parallel unit-propagation takes O(bm0 T ) time, and m decreases by m0 . Therefore, in O(bmT ) time, we reach to the state that either V = ∅ or there exists no such (u, a). In the latter case, we branch into at most b cases, and ∆ decreases by at least 12 for each case. Therefore, we have T (∆) ≤ bT (∆ − 21 ) + O(kmT + bmT ). Thus, we have T (∆) = O(max(2, b)2∆ kmT ). 5.2 Applications to Other Problems Finally, we show that various NP-hard problems can be expressed as a special case of 0/1/all Deletion. Note that we use A = ∅ for every problem other than Node Multiway Cut. We obtain a linear-time FPT algorithm for each problem by giving an incremental-test oracle and a specialized choice of branching sets. Node Unique Label Cover Parameter: k, |Σ| Input: A finite alphabet Σ, a graph G = (V, E), a permutation πe of Σ for every edge e ∈ Ê −1 given as a table of size |Σ|, and an integer k. such that πuv = πvu Question: Is there a pair of set X ⊆ V of at most k vertices and assignment ϕ : V \ X → Σ such that πuv (ϕ(u)) = ϕ(v) for every uv ∈ E[V \ X]? Two-fan Deletion Parameter: k Input: A set of variables V , a set of two-fan constraints on V of the form (ϕ(u) = a)∨(ϕ(v) = b) given as a pair (a, b), and an integer k. Question: Is there a pair of set X ⊆ V of at most k variables and assignment ϕ satisfying every constraint Cuv ∈ C[V \ X]? These two problems are special cases of 0/1/all Deletion such that the set of constraints is limited to permutation or two-fan constraints. Note that the size of the domain is not a parameter for Two-fan Deletion. The naive implementation of the incremental-test oracle for these problems runs in a constant time. Thus, we can solve Node Unique Label Cover in O(|Σ|2k km) time. We can use the following choice of a branching set for Two-fan Deletion: pick a two-fan constraint (ϕ(u) = a) ∨ (ϕ(v) = b) and set B := {(u, a), (v, b)}. Thus, we can solve Two-fan Deletion in O(22k km) = O(4k km) time. Lemma 22. B := {(u, a), (v, b)} for a two-fan constraint (ϕ(u) = a) ∨ (ϕ(v) = b) is a branching set. 39 Proof. Let X be a deletion set for I and let ϕ be an assignment for V \ X satisfying C[V \ X]. If u ∈ X, X is a deletion set for I[u ← a], and if v ∈ X, X is a deletion set for I[v ← b]. Otherwise, at least one of ϕ(u) = a or ϕ(v) = b holds. X is a deletion set for I[u ← a] in the former case, while X is a deletion set for I[v ← b] in the latter case. The next two problems generalize Pseudoforest Deletion [5] in different directions, where a pseudoforest is a graph in which the number of edges is at most the number of vertices for every connected component. A graph is called monochromatically orientable if there exists an edge orientation such that, for every vertex, all the incoming edges are monochromatic. It is known that a graph is a pseudoforest if and only if there exists an edge orientation such that, for every vertex, the number of incoming edges is at most one. Therefore, when every edge has a distinct color, a graph is monochromatically orientable if and only if it is a pseudoforest. Thus, the following problem is a generalization of Pseudoforest Deletion. Monochromatically Orientable Deletion Parameter: k Input: An edge-colored graph G = (V, E) and an integer k. Question: Is there a set X ⊆ V of at most k vertices such that G − X is monochromatically orientable? We can solve Monochromatically Orientable Deletion in O(4k km) time by the following reduction to Two-fan Deletion. Let L be the set of colors. For each vertex v ∈ V , we create a variable v with domain D(v) = L, which represents the color of the incoming edges. We create a two-fan constraint (ϕ(u) = c) ∨ (ϕ(v) = c) for each edge uv ∈ E of color c ∈ L. Lemma 23. A graph G = (V, E) is monochromatically orientable if and only if the corresponding set C of two-fan constraints is satisfiable. Proof. From a monochromatic orientation of G, we can construct a satisfying assignment ϕ as follows: we set ϕ(v) := c for each vertex v ∈ V , where c is the unique color of the incoming edges or an arbitrary color if v has no incoming edges. From a satisfying assignment ϕ, we can construct a monochromatic orientation by orienting each edge uv ∈ E of color c so that the edge is directed toward u if ϕ(u) = c, and toward v otherwise. Another generalization of Pseudoforest Deletion is presented as follows. For a graph G = (V, E) and an edge e = uv ∈ E, contracting e is an operation deleting the edge e and merging u and v into a new vertex e. Note that this operation may create parallel edges (edges uw and vw become parallel edges) and self-loops. If G has parallel edges connecting u and v, the operation only removes one of them, and the rest becomes self-loops. Let S be a subset of edges. A subset X ⊆ V is called a subset feedback vertex set if G − X has no simple cycle passing through an edge of S, or equivalently, the graph obtained from G − X by contracting every edge e ∈ E[V \ X] \ S is a forest. Similarly, we call X a subset pseudoforest deletion set if the graph obtained from G − X by contracting every edge e ∈ E[V \ X] \ S is a pseudoforest. Subset Pseudoforest Deletion Parameter: k Input: A graph G = (V, E), a set S ⊆ E, and an integer k. Question: Is there a set X ⊆ V of at most k vertices such that the graph obtained from G − X by contracting every edge e ∈ E[V \ X] \ S is a pseudoforest? 40 This problem can be expressed as 0/1/all Deletion as follows. Every vertex v has the same domain D(v) = S. We introduce a two-fan constraint (ϕ(u) = e) ∨ (ϕ(v) = e) for every edge e = uv ∈ S. We also introduce an equality (identity permutation) constraint ϕ(u) = ϕ(v) for every edge uv 6∈ S. Let C be the obtained set of constraints. If C[V 0 ] has no two-fan constraints for some V 0 ⊆ V , (C[V 0 ], ϕ∅ ) has a deletion set of size zero. Therefore, we can use the same choice of branching sets as for Two-fan Deletion. Thus, the algorithm runs in O(4k km) time. The correctness of the expression follows from the following lemma. Lemma 24. Let G = (V, E) be a graph with a subset S ⊆ E and let C be the corresponding set of 0/1/all constraints. Then, the graph obtained by contracting every edge e ∈ E \ S is a pseudoforest if and only if C is satisfiable. Proof. We modify C for each contraction of an edge e = uv by removing the constraint Cuv and replacing every occurrence of u and v by e. When we obtain the graph G0 by contracting every edge e ∈ E \ S, we also obtain the set C 0 of constraints corresponding to G0 . Because every contracted edge e = uv has the equality constraint, C is satisfiable if and only if C 0 is satisfiable. From Lemma 23, C 0 is satisfiable if and only if G0 is a pseudoforest. Therefore, C is satisfiable if and only if G0 is a pseudoforest. As mentioned in Section 1.3, Node Multiway Cut is also a special case of our problem. Node Multiway Cut Parameter: k Input: A graph G = (V, E), a set of terminals T ⊆ V , and an integer k. Question: Is there a set X ⊆ V \ T of size at most k such that every terminal in T lies in a different connected component of G − X? This problem can be expressed as 0/1/all Deletion as follows. First, we split each terminal s ∈ T as follows to make s undeletable: for each edge sv ∈ δ(s), we create a new vertex sv and replace the edge sv with sv v. Let G0 = (V 0 , E 0 ) be the obtained graph. We introduce a variable v with D(v) := T for each vertex v ∈ V 0 and an equality (identity permutation) constraint ϕ(u) = ϕ(v) for each edge uv ∈ E 0 . Finally, we set A = {sv | s ∈ T, sv ∈ δ(s)} and ϕA (sv ) := s. We can observe that a minimum deletion set X avoiding every sv always exists because each vertex sv has degree one, and such X is actually a minimum multiway cut. Let I = (C, ϕA ) be the obtained instance. We can construct a multiway cut of size at most 2|x| by rounding up a half-integral FI -cover x. Therefore, we have k − c(I) ≤ 21 k. Because the domain size is |T | = O(n), the naive implementation of the incremental-test oracle runs in a constant time. Because any I = (C[V 0 ], ϕ∅ ) has a deletion set of size zero, we do not need branching sets. Thus, 1 we can solve Node Multiway Cut in O(22· 2 k km) = O(2k km) time. We finally show Group Feedback Vertex Set and its applications. Group Feedback Vertex Set Parameter: k Input: A group Γ = (D, ·) given as an O(TΓ )-time oracle performing the group operation (·), a Γ-labeled graph G = (V, E) with labeling λ : Ê → D with λ(uv) · λ(vu) = 1Γ for every uv ∈ Ê, where 1Γ is the unity of Γ, and an integer k. Question: Is there a set X ⊆ V of at most k vertices such that G − X has a consistent labeling? That is, is there a labeling ϕ : V \ X → D such that ϕ(u) · λ(uv) = ϕ(v) for every uv ∈ E[V \ X]? 41 This problem can be expressed as 0/1/all Deletion because a function πe (a) := a · λ(e) is a permutation. Note that G − X has a consistent labeling if and only if it admits no non-zero cycles (i.e., it admits no cycle (v0 , . . . , v` ) with λ(v0 v1 ) · λ(v1 v2 ) · · · λ(v`−1 v` ) 6= 1Γ ). In contrast to Node Unique Label Cover, the domain size is not a parameter, and each permutation is given not as a table of size |D|, but as an O(TΓ )-time oracle answering a · b for given a, b ∈ D. Therefore, the naive implementation of the incremental-test oracle runs in O(TΓ ) time. We can use the following choice of a branching set: pick a vertex s and set B := {(s, 1Γ )}. Thus, we can solve Group Feedback Vertex Set in O(22k km) = O(4k km) time. Lemma 25. B := {(s, 1Γ )} is a branching set for Group Feedback Vertex Set. Proof. Let X be a deletion set for I, and let ϕ be an assignment for V \ X satisfying C[V \ X]. If s ∈ X, X is a deletion set for I[s ← 1Γ ]; otherwise, ϕ0 such that ϕ0 (v) := ϕ(v) · ϕ(s)−1 is a satisfying assignment for C[V \ X] with ϕ0 (s) = 1Γ . Subset Feedback Vertex Set Parameter: k Input: A graph G = (V, E), a set S ⊆ E, and an integer k. Question: Is there a set X ⊆ V of at most k vertices such that no cycle passes through an edge of S in G − X? This problem can be expressed as Group Feedback Vertex Set as follows. We use a group Γ = (2S , ⊕), where ⊕ is the XOR operator (X ⊕ Y = (X \ Y ) ∪ (Y \ X)). We set λ(e) = {e} for each edge e ∈ S and λ(e) = ∅ for each edge e ∈ E \ S. Then, a cycle is non-zero if and only if it contains an edge in S. The group operation takes O(|S|) = O(m) time. Therefore, the naive implementation of the incremental-test oracle takes O(m) time. We now provide a constant-time incremental-test oracle for (C[V 0 ], ϕ{s} ) with ϕ{s} (s) = ∅. We use the following implementation. U := S ∪ {}. ( e (e ∈ S) I(s) = . A(a, e) = a (e ∈ 6 S). ( true T (a, b) = false (a 6= b) (a = b). For a walk W , the function A∗ (W ) returns the last edge of W contained in S or  if W contains no edges in S. Then, for a single-branching pair (P, Q), we have P 6≡ Q ⇐⇒ the simple cycle contained in P ◦ Q−1 contains an edge in S ⇐⇒ A∗ (P ) 6= A∗ (Q). Therefore, the abovementioned implementation is correct. Thus, we can solve Subset Feedback Vertex Set in O(4k km) time. Non-monochromatic Cycle Transversal Parameter: k Input: An edge-colored graph G = (V, E) and an integer k. Question: Is there a set X ⊆ V of at most k vertices such that G − X contains no nonmonochromatic cycles? This problem can be expressed as Group Feedback Vertex Set as follows. Let L be the set of colors, and let c(e) ∈ L denote the color of an edge e. We use the group Γ = (2V ×L , ⊕). We set λ(e) = {(u, c(e)), (v, c(e))} for each edge e = uv ∈ E. Then, a cycle is non-zero if and only if it is non-monochromatic. The naive implementation of the incremental-test oracle takes O(n|L|) = O(nm) time. We now provide a constant-time incremental-test oracle for (C[V 0 ], ϕ{s} ) with ϕA (s) = ∅. We use the 42 following implementation. ( true U := (V × L) ∪ {, ∗}. I(s) = . T (a, b) = false A(, sv) = (s, c(sv)). (∗ = 6 a 6= b 6= ∗) otherwise.   (w = v ∧ c = c(uv)) ∗ A((w, c), uv) = (w, c) (w = 6 v ∧ c = c(uv))   (u, c(uv)) (c 6= c(uv)). Let W = (v0 , . . . , v` ) be a walk with ` > 0, and let c = c(v`−1 v` ). A suffix (vi , . . . , v` ) is called the longest monochromatic suffix of W if c(vj vj+1 ) = c for every j ≥ i and c(vi−1 vi ) 6= c or i = 0 holds. We can observe that the longest monochromatic suffix of W starts from vi and has the color c if A∗ (W ) = (vi , c), and the longest monochromatic suffix of W forms a monochromatic cycle if A∗ (W ) = ∗. Then, for a single-branching pair (P, Q), we have P 6≡ Q ⇐⇒ the simple cycle contained in P ◦ Q−1 is non-monochromatic ⇐⇒ none of P and Q induces a monochromatic cycle and the longest monochromatic suffixes of P and Q start from different vertices or have different colors ⇐⇒ ∗ = 6 A∗ (P ) 6= A∗ (Q) 6= ∗. Therefore, the abovementioned implementation is correct. Thus, we can solve Non-monochromatic Cycle Transversal in O(4k km) time. References [1] M. A. Babenko. A fast algorithm for the path 2-packing problem. Theory of Computing Systems, 46(1):59–79, 2010. [2] M. L. Balinski. Integer programming: Methods, uses, computations. Management Science, 12(3):253–313, 1965. [3] A. Becker, R. Bar-Yehuda, and D. Geiger. Randomized algorithms for the loop cutset problem. Journal of Artificial Intelligence Research, 12:219–234, 2000. [4] H. L. Bodlaender. A linear-time algorithm for finding tree-decompositions of small treewidth. SIAM Journal on Computing, 25(6):1305–1317, 2006. [5] H. L. Bodlaender, H. Ono, and Y. Otachi. A faster parameterized algorithm for pseudoforest deletion. In Proceedings of the 11th International Symposium on Parameterized and Exact Computation (IPEC 2016), pages 7:1–7:12, 2017. [6] J. Chen, Y. Liu, and S. Lu. An improved parameterized algorithm for the minimum node multiway cut problem. Algorithmica, 55(1):1–13, 2009. [7] M. C. Cooper, D. A. Cohen, and P. G. Jeavons. Characterising tractable constraints. Artificial Intelligence, 65(2):347–361, 1994. [8] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer International Publishing, 2015. [9] M. Cygan, M. Pilipczuk, M. Pilipczuk, and J. O. Wojtaszczyk. On multiway cut parameterized above lower bounds. ACM Transactions on Computation Theory, 5(1):3–11, 2013. 43 [10] R. G. Downey and M. R. Fellows. Parameterized Complexity. Springer Verlag, 2012. [11] J. Edmonds. Paths, trees, and flowers. Canadian Journal of mathematics, pages 449–467, 1965. [12] A. Frank and É. Tardos. An application of simultaneous diophantine approximation in combinatorial optimization. Combinatorica, 7(1):49–65, 1987. [13] S. Fujishige and X. Zhang. New algorithms for the intersection problem of submodular systems. Japan Journal of Industrial and Applied Mathematics, 9(3):369–382, 1992. [14] N. Garg, V. V. Vazirani, and M. Yannakakis. Multiway cuts in node weighted graphs. Journal of Algorithms, 50(1):49–61, 2004. [15] S. Guillemot. FPT algorithms for path-transversal and cycle-transversal problems. Discrete Optimization, 8(1):61–71, 2011. [16] H. Hirai. A dual descent algorithm for node-capacitated multiflow problems and its applications. 2015. arXiv:1508.07065. [17] Y. Iwata. Linear-time kernelization for feedback vertex set. In Proceedings of 44th International Colloquium on Automata, Languages, and Programming (ICALP), pages 68:1–68:14, 2017. [18] Y. Iwata, K. Oka, and Y. Yoshida. Linear-time FPT algorithms via network flow. In Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1749–1761, 2014. [19] Y. Iwata, M. Wahlström, and Y. Yoshida. Half-integrality, LP-branching, and FPT algorithms. SIAM Journal on Computing, 45(4):1377–1411, 2016. [20] D. Lokshtanov, N. S. Narayanaswamy, V. Raman, M. S. Ramanujan, and S. Saurabh. Faster parameterized algorithms using linear programming. ACM Transactions on Algorithms, 11(2):15:1–15:31, 2014. [21] D. Lokshtanov, M. S. Ramanujan, and S. Saurabh. Linear time parameterized algorithms for subset feedback vertex set. In Proceedings of 42nd International Colloquium on the Automata, Languages, and Programming (ICALP), pages 935–946, 2015. [22] D. Lokshtanov, M. S. Ramanujan, and S. Saurabh. A Linear Time Parameterized Algorithm for Directed Feedback Vertex Set. 2016. arXiv:1609.04347. [23] D. Lokshtanov, M. S. Ramanujan, and S. Saurabh. A linear-time parameterized algorithm for node unique label cover. In Proceedings of the 25th Annual European Symposium on Algorithms (ESA), pages 57:1–57:15, 2017. [24] G. Nemhauser and L. Trotter. Vertex packing: structural properties and algorithms. Mathematical Programming, 8:232–248, 1975. [25] G. Pap. A constructive approach to matching and its generalizations. PhD thesis, Eötvös Loránd University, 2006. 44 [26] G. Pap. Packing non-returning A-paths. Combinatorica, 27(2):247–251, 2007. [27] G. Pap. Some new results on node-capacitated packing of A-paths. In Proceedings of the 39th Annual ACM Symposium on Theory of Computing (STOC), pages 599–604, 2007. [28] G. Pap. Packing non-returning A-paths algorithmically. Discrete Mathematics, 308(8):1472– 1488, 2008. [29] G. Pap. Strongly polynomial time solvability of integral and half-integral node-capacitated multiflow problems. Technical report, EGRES Technical Report, TR-2008-12, Eötvös Loránd University, 2008. [30] J.-C. Picard and M. Queyranne. On the structure of all minimum cuts in a network and applications. In Combinatorial Optimization II, volume 13 of Mathematical Programming Studies, pages 8–16. Springer Berlin Heidelberg, 1980. [31] M. S. Ramanujan and S. Saurabh. Linear time parameterized algorithms via skew-symmetric multicuts. In Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1739–1748, 2014. [32] A. Schrijver. Combinatorial optimization: polyhedra and efficiency, volume 24. Springer Science & Business Media, 2002. [33] M. Wahlström. LP-branching algorithms based on biased graphs. In Proceedings of the 28th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1559–1570, 2017. [34] Y. Yamaguchi. Packing A-paths in group-labelled graphs via linear matroid parity. SIAM Journal on Discrete Mathematics, 30(1):474–492, 2016. 45 A Proof of Persistency First, we review the results of [19]. Let D be a set containing a special element ⊥, and let DI := D \ {⊥}. Consider a poset on D such that ⊥ < a for every a ∈ DI and all the other pairs (a, b) with a 6= b are not comparable. Let a u b be a binary operator that returns the minimum of a and b if they are comparable and returns ⊥ otherwise. Similarly, let a t b be a binary operator that returns the maximum of a and b if they are comparable and returns ⊥ otherwise. We define a · b := (a1 · b1 , . . . , an · bn ) for a binary operator (·) and tuples of elements a = (a1 , . . . , an ), b = (b1 , . . . , bn ) ∈ Dn . Each index i herein might have a distinct domain Di containing ⊥, and Dn is an abbreviation for D1 × D2 × . . . × Dn . A function f : Dn → R is called k-submodular if the following inequality holds for any pair of inputs a, b ∈ Dn : f (a) + f (b) ≥ f (a u b) + f (a t b). We denote the restriction of a function f : Dn → R to the domain DIn by fI . Lemma 26 ( [19]). For a k-submodular function f : Dn → R and a minimizer b ∈ Dn of f , there exists a minimizer a ∈ DIn of fI such that ai = bi for every i with bi 6= ⊥. Lemma 27 ( [19]). The following three functions are k-submodular. • For a permutation π on DI , a function pπ : D × D → {0, ∞} defined as ( 0 if (x = y = ⊥) ∨ (x, y ∈ DI ∧ π(x) = y), pπ (x, y) = ∞ otherwise. • For elements a, b ∈ DI , a function ta,b : D × D → {0, ∞} defined as ( 0 if (x = y = ⊥) ∨ (x = a) ∨ (y = b), ta,b (x, y) = ∞ otherwise. • A function e : Dr → {0, 21 , 1} defined as   0 e(x) = 1  1 2 if x1 = x2 = · · · = xr , if ∃i, j such that ⊥ = 6 xi 6= xj 6= ⊥, otherwise. Proof of Theorem 1. Let f : DÊ → R be a function obtained by taking the sum of the following functions. Each variable uv ∈ Ê here has a domain D(u)∪{⊥}. We add the function pπ on (uv, vu) for each permutation constraint π(ϕ(u)) = ϕ(v) of C. We add the function ta,b on (uv, vu) for each two-fan constraint (ϕ(u) = a) ∨ (ϕ(v) = b) of C. Meanwhile, we add the d-ary function e on δ(v) for each variable v ∈ V \ A, where d = |δ(v)| is the degree of v. Similarly for each variable v ∈ A, we add the (d + 1)-ary function e on δ(v) with one argument fixed to ϕ(v) (hence it acts as a d-ary function), where d = |δ(v)|. Then, from Lemma 27, f is k-submodular. We first observe that we can convert an input b ∈ DÊ of f with f (b) < ∞ to a half-integral FC,ϕA -cover x with x(V ) = f (b) and convert a half-integral FC,ϕA -cover x to an input b ∈ DÊ 46 with f (b) ≤ x(V ). For each vertex v, we denote by fv (b) the value of the function e for v, that is, for v ∈ V \ A, fv (b) := e(b|δ(v) ), where P b|δ(v) is the restriction of b to δ(v), and for v ∈ A, fv (b) := e(b|δ(v) , ϕA (v)). We have f (b) = v∈V fv (b) when f (b) < ∞. Let b ∈ DÊ be an input of f with f (b) < ∞. We construct x by setting x(v) = fv (b) for each v ∈ V . To see that x is a (half-integral) FC,ϕA -cover, take a ϕA -conflicting walk W = (v0 , . . . , v` ) with v0 , v` ∈ A. Now, consider a sequence of variables for f along the walk W : v0 v1 , v1 v0 , v1 v2 , v2 v1 , . . . , v`−2 v`−1 , v`−1 v`−2 , v`−1 v` , v` v`−1 . Let bv0 v−1 = ϕA (v0 ) and bv` v`+1 = ϕA (v` ). As the walk P is ϕA -conflicting, we must have an index 0 ≤ i ≤ ` such that ⊥ = 6 bvi vi−1 6= bvi vi+1 6= ⊥ or two indices 0 ≤ i < j ≤ ` such that bvi vi−1 6= ⊥, bvi vi+1 = ⊥, bvj−1 vj = ⊥, and bvj vj+1 6= ⊥. In both cases, x(V (P )) ≥ 1. Next, let x be a half-integral FC,ϕA -cover. Let ϕR(x) be the satisfying assignment for C[R(x)]. We then define an input b for f as follows:   if u ∈ R(x), ϕR(x) (u) buv = Cvu (ϕR(x) (v)) if u 6∈ R(x), v ∈ R(x), and Cvu (ϕR(x) (v)) 6= all,   ⊥ otherwise. We have f (b) < ∞ from the construction. For every vertex u ∈ V with x(u) = 0, we have buv = ϕR(x) (u) for every uv ∈ δ(u) if u ∈ R(x). Otherwise, we have buv = ⊥ for every uv ∈ δ(u). We have fv (b) = 0 in both cases. For any vertex u ∈ V with x(u) = 21 , there exist no edges uv1 , uv2 ∈ δ(u) such that ⊥ = 6 buv1 6= buv2 6= ⊥ because, otherwise, there exists a ϕA -conflicting walk W with x(V (W )) = x(u) = 21 < 1, which is a contradiction. Thus, we have f (b) ≤ x(V ). Now, we prove the claim. Let x be a minimum half-integral FC,ϕA -cover. We then construct a minimizer b ∈ DÊ of f with f (b) = x(V ). From Lemma 26, there is a minimizer a ∈ DIÊ of fI such that auv = buv for every uv ∈ Ê with buv 6= ⊥. From the construction above, this means the existence of the desired deletion set X := {v ∈ V | fv (a) = 1}. B Axiomatic Model We introduce an equivalent formulation of conflicting/implicational walks. The merit of this formulation is that checking whether a set of walks satisfies the conditions below is often easier than finding an explicit expression as 0/1/all constraints. Definition 6. A pair (F, F ∗ ) of (possibly infinite) sets of walks in the same graph is called nice if it satisfies the following conditions. 1. F ⊆ F ∗ . 2. F ∗ is closed under taking a prefix (i.e., for any walk W ∈ F ∗ and any prefix-subwalk P of W , P ∈ F ∗ holds). 3. For two walks P, Q ∈ F ∗ ending at the same vertex, we write P ≡ Q if and only if P ◦Q−1 6∈ F. The relation (≡) then becomes an equivalence relation, that is, (1) P ≡ P , (2) P ≡ Q ⇐⇒ Q ≡ P , and (3) P ≡ Q ∧ Q ≡ R =⇒ P ≡ R hold for every walks P, Q, R ∈ F ∗ ending at the same vertex. 47 4. For any equivalent walks P, Q ∈ F ∗ ending at u and any edge uv ∈ E, P ◦ uv ∈ F ∗ if and only if Q ◦ uv ∈ F ∗ . For two walks P, Q ∈ F ∗ ending at the same vertex, we write P 6≡ Q if and only if P ◦ Q−1 ∈ F. The notations P ≡ Q or P 6≡ Q implicitly imply that P, Q ∈ F ∗ and t(P ) = t(Q). Note that, from the definition, P ∈ F ⇐⇒ P −1 ∈ F always holds. However, P ∈ F ∗ ⇐⇒ P −1 ∈ F ∗ may not hold. When F ∗ is the set of all walks starting from a set of vertices A, (F, F ∗ ) is nice if and only if F is the set of non-returning A-walks (cf. [25, pp. 109–111]). Lemma 28. Let F ∗ be the set of all implicational walks and let F be the set of all conflicting walks. Then (F, F ∗ ) is nice. Proof. Conditions 1, 2, and 4 are trivial from the definition, and we only prove Condition 3. From Lemma 1, for any two walks P, Q ∈ F ∗ ending at the same vertex, P ≡ Q if and only if imp(P ) = imp(Q) holds. Therefore, (≡) is an equivalence relation. Lemma 29. For any nice pair (F, F ∗ ) for a graph G = (V, E), there exist domains for V , 0/1/all constraints for E, and a partial assignment ϕA for some A ⊆ V such that F is exactly the set of all ϕA -conflicting walks and F ∗ is exactly the set of all ϕA -implicational walks. Proof. For each vertex v ∈ V , let D(v) be the classes of equivalent walks in F ∗ ending at v. We denote the class containing W by [W ] ∈ D(v). Let A := {v | (v) ∈ F ∗ } and ϕA (v) := [(v)]. For each edge uv ∈ Ê, we define a subdomain D(uv) ⊆ D(u) and a function fuv : D(uv) → D(v) as follows. 1. For any walk W ∈ F ∗ ending at u, [W ] ∈ D(uv) if and only if W ◦ uv ∈ F ∗ . 2. For any walk W ◦ uv ∈ F ∗ , fuv ([W ]) = [W ◦ uv] holds. Claim 11. D(uv) and fuv are well-defined. Proof. Let W 0 ∈ [W ]. We have W 0 ◦ uv ∈ F ∗ if W ◦ uv ∈ F ∗ (by Condition 4). Suppose that W 0 ◦ uv 6≡ W ◦ uv, which implies W 0 6≡ W ◦ uv ◦ vu. Meanwhile, by Condition 3-(1), W ◦ uv ≡ W ◦ uv holds and, hence, W ◦ uv ◦ vu ≡ W . Thus, we have W 0 6≡ W ◦ uv ◦ vu ≡ W , which is a contradiction. Claim 12. If fuv ([W ]) ∈ D(vu) holds for some class [W ] ∈ D(uv), we have D(u) = D(uv), −1 . D(v) = D(vu), and fuv = fvu Proof. We have fvu (fuv ([W ])) = [W ◦ uv ◦ vu] = [W ] because W ◦ uv ◦ vu ≡ W holds. Suppose that there exists a class [W 0 ] ∈ D(u) \ D(uv) (i.e., W 0 ◦ uv 6∈ F ∗ ). Since [W 0 ] 6= [W ] ∈ D(uv), we have W 0 6≡ W ≡ W ◦ uv ◦ vu, implying W 0 ◦ uv ∈ F ∗ , which is a contradiction. Therefore, we have D(u) = D(uv). By the same argument, we have D(v) = D(vu). Thus, the claim holds. Claim 13. If fuv ([W ]) 6∈ D(vu) holds for some class [W ] ∈ D(uv), there exist classes a ∈ D(u) and b ∈ D(v) such that D(uv) = D(u) \ {a}, D(vu) = D(v) \ {b}, fuv (a0 ) = b for every class a0 ∈ D(uv), and fvu (b0 ) = a for every class b0 ∈ D(vu). 48 Proof. Let b := fuv ([W ]) 6∈ D(vu). Suppose that there exists a class [W 0 ] ∈ D(uv) such that fuv ([W 0 ]) 6= fuv ([W ]). Because W ◦ uv 6≡ W 0 ◦ uv holds, we have W ◦ uv ◦ vu ∈ F ∗ , which is a contradiction. Therefore, we have fuv (a0 ) = b for every class a0 ∈ D(uv). Let [Q] ∈ D(v) \ {b}. Because Q 6≡ W ◦ uv holds, we have Q ◦ vu ∈ F ∗ . Therefore, we have D(vu) = D(v) \ {b}. Pick an arbitrary class b0 ∈ D(vu). If a := fvu (b0 ) ∈ D(uv), from Claim 12, we have D(vu) = D(v), which is a contradiction. Therefore, by the same argument, we have D(uv) = D(u) \ {a} and fvu (b0 ) = a for every class b0 ∈ D(vu). Now, we prove the lemma. We introduce a constraint for each edge uv ∈ E as follows: Pick an arbitrary class c ∈ D(uv). If fuv (c) ∈ D(vu) holds, from Claim 12, the edge uv can be expressed as fuv (ϕ(u)) = ϕ(v) for the permutation fuv . If fuv (c) 6∈ D(vu) holds, from Claim 13, the edge uv can be expressed as (ϕ(u) = a) ∨ (ϕ(v) = b) for some classes a ∈ D(u) and b ∈ D(v). 49
8cs.DS
1 Data-Driven Sparse System Identification arXiv:1803.07753v1 [cs.SY] 21 Mar 2018 Salar Fattahi and Somayeh Sojoudi Abstract—In this paper, we study the system identification porblem for sparse linear time-invariant systems. We propose a sparsity promoting Lasso-type estimator to identify the dynamics of the system with only a limited number of input-state data samples. Using contemporary results on high-dimensional statistics, we prove that Ω(kmax log(m + n)) data samples are enough to reliably estimate the system dynamics, where n and m are the number of states and inputs, respectively, and kmax is the maximum number of nonzero elements in the rows of input and state matrices. The number of samples in the developed estimator is significantly smaller than the dimension of the problem for sparse systems, and yet it offers a small estimation error entrywise. Furthermore, we show that, unlike the recently celebrated least-squares estimators for system identification problems, the method developed in this work is capable of exact recovery of the underlying sparsity structure of the system with the aforementioned number of data samples. Extensive case studies on synthetically generated systems and physical mass-spring networks are offered to demonstrate the effectiveness of the proposed method. I. I NTRODUCTION With their ever-growing size and complexity, real-world dynamical systems are hard to model. Today’s systems are complex and large, often with a massive number of unknown parameters which render them doomed to the so-called curse of dimensionality. Therefore, system operators should rely on simple and tractable estimation methods to identify the dynamics of the system via a limited number of recorded input-output interactions, and then design control policies to ensure the desired behavior of the entire system. The area of system identification is created to address this problem [1]. Despite the long history in control theory, most of the results on system identification deal with asymptotic behavior of the proposed estimation methods [1]–[4]. Although these results shed light on the theoretical consistency of these methodologies, they are not applicable to the finite time/sample settings. In many applications, the dynamics of the system should be estimated under the large dimension-small sample size regime, where the dimension of the states and inputs of the system is overwhelmingly large compared to the number of available input-output data. Under such circumstances, the classical approaches for checking the asymptotic consistency of estimators face major breakdowns. Simple examples of such failures can be easily found in high-dimensional statistics. For instance, given a number of independent and identically Email: [email protected] and [email protected]. Salar Fattahi is with the Department of Industrial Engineering and Operations Research, University of California, Berkeley. Somayeh Sojoudi is with the Departments of Electrical Engineering and Computer Sciences and Mechanical Engineering as well as the Tsinghua-Berkeley Shenzhen Institute, University of California, Berkeley. This work was supported by the ONR grant N00014-17-1-2933, DARPA grant D16AP00002, and AFOSR grant FA955017-1-0163. distributed (i.i.d.) samples with Gaussian distribution, none of the eigenvectors and eigenvalues of the sample covariance matrix are consistent estimators of their true counterparts if the sample size and the dimension of variables grow at the same rate [5]. As another example, it is well-known that the least-squares estimators, which are widely used in system identification problems, seize to exist when the sample size is smaller than the dimension of the system [6]. On the other hand, with the unprecedented interest in datadriven control approaches, such as model-free reinforcement learning, robust, and adaptive control [7]–[9], a question arises as to what the minimum number of input-output data samples should be to guarantee a small error in the estimated model. Answering this question has been the subject of many recent studies on the sample complexity of the system identification problem [10]–[14]. Most of these results are tailored to a specific type of dynamics, depend on the stability of the openloop system, or do not exploit the a priori information on the structure of the system. In this work, the objective is to employ modern results on high-dimensional statistics to reduce the sample complexity of one of the most fundamental problems in control theory, namely the linear time-invariant (LTI) systems with perfect state measurements. This type of dynamical system forms the basis of many classical control problems, such as Linear Quadratic Regulator and Linear Quadratic Gaussian problems. Our results are built upon the fact that, in many practical large-scale systems, the states and inputs exhibit sparse interactions with one another, which in turn translates into a sparse representation of the state space equations of the system. It will be shown that the sparsity of the dynamics enables us to develop an estimator that is guaranteed to achieve infinitesimal estimation error with small number of samples. In particular, we introduce a Lasso-type estimator, i.e. a least-squares estimator augmented by an ℓ1 regularizer, and show that, with an appropriate scaling of the regularization coefficient, its sample complexity is only Ω(kmax log(n+m)), where kmax is maximum number of nonzero elements in the columns of input and state (or system) matrices, and n and m are the dimensions of the states and the inputs, respectively. This is a significant improvement over the recently derived sample complexity Ω(n + m) for the least-squares estimator, in the case where kmax is much smaller than m + n, i.e., when the system is sparse. In this work, we derive upper bounds on the elementwise error of the proposed estimator. In particular,√ we show that the elementwise error decreases at the rate O( log(n + m)/d), where d is the number of available sample trajectories. Although this error√ is not comparable to the operator norm error bound O( n + m/d) for the least-squares estimator introduced in [13], we show tremendous improvements in the 2 accuracy of the developed estimator through different case studies. Another advantage of the proposed Lasso estimator over its least-squares analog is its exact recovery property. More specifically, we show that while the least-squares estimator is unable to identify the sparsity pattern of the input and state matrices for any finite number of samples, the proposed estimator is guaranteed to recover the true sparsity pattern of these matrices with only Ω(kmax log(n + m)) samples. It is worthwhile to mention that this work generalizes the results in [12], where the authors use a similar Lasso estimator to learn the dynamics of a particular type of systems. However, [12] assumes that the system is autonomous (input-free) and inherently stable, both of which will be relaxed in this work. Notations: For a matrix M , the symbols ∥M ∥F , ∥M ∥2 , ∥M ∥1 , and ∥M ∥∞ denote its Frobenius, operator, ℓ1 /ℓ1 , and ℓ∞ /ℓ∞ norms, respectively. Furthermore, κ(M ) refers to its 2-norm condition number, i.e., the ratio of its maximum and minimum singular values. Given integer sets I and J, the notation MIJ refers to the submatrix of M whose rows and columns are indexed by I and J, respectively. Given the sequences f1 (n) and f2 (n), the notations f1 (n) = O(f2 (n)) and f1 (n) = Ω(f2 (n)) imply that there exist c1 < ∞ and c2 > 0 such that f1 (n) ≤ c1 f2 (n) and f1 (n) ≥ c2 f2 (n), respectively. Finally, f1 (n) = o(f2 (n)) is used to show that f1 (n)/f2 (n) → 0 as n → ∞. A zero-mean Gaussian distribution with covariance Σ is shown as N (0, Σ). Given a function f (x), the expression arg min f (x) refers to its minimizer. For a set I, the symbol ∣I∣ denotes its cardinality. II. P ROBLEM F ORMULATION Consider the LTI system x[t + 1] = Ax[t] + Bu[t] + w[t] (1a) where t is the time step, A ∈ Rn×n is the state matrix, and B ∈ Rn×m is the input matrix. Furthermore, x[t] ∈ Rn , u[t] ∈ Rm , and w[t] ∈ Rn are the state, input, and disturbance vectors at time t, respectively. The dimension of the system is defined as m + n. It is assumed that the input disturbance vectors are identically distributed and independent with distribution N (0, Σw ) across different times. In this work, we assume that the matrices A and B are sparse and the goal is to estimate them based on a limited number of sample trajectories, i.e. a sequence {(x(i) [τ ], u(i) [τ ])}Tτ=0 with i = 1, 2, ..., d, where d is the number of available sample trajectories. The ith sample trajectory {(x(i) [τ ], u(i) [τ ])}Tτ=0 is obtained by running the system from t = 0 to t = T and collecting the input and state vectors. The sparsity assumption on A and B is practical in many applications because of two reasons: ● In order to model many large-scale real-world systems accurately, one needs to consider an overwhelmingly large number of internal states. However, it is often the case that the interactions between different states and inputs obey a sparse structure, which translates into a sparsity pattern in state and input matrices. An important example of these types of problems is multi-agent systems, where the agents (subsystems) interact with one another via a sparse communication network. ● More generally, it is well-known that there may not be a unique representation of the state-space model of the system. For example, one can entirely change the state and input matrices in (1) via linear/nonlinear transformations to arrive at a different, but equally accurate state-space equation. Furthermore, the dynamics of the system may have a sufficiently accurate sparse approximation. The question of interest is whether it is possible to design a data-driven method in order to estimate the sparsest state-space representation of the system? Answering this question is crucial, specially in the context of a distributed control problem where the goal is to design a decentralized controller whose structure respects the dynamics of the system [15]–[17]. Given the sample trajectories {(x(i) [τ ], u(i) [τ ])}Tτ=0 for i = 1, 2, ..., d, one can obtain an estimate of (A, B) by solving the following least-squares optimization problem: d T −1 min ∑ ∑ ∥x(i) [t + 1] − (Ax(i) [t] + Bu(i) [t])∥2 2 (2) A,B i=1 t=0 In order to describe the behavior of the least-squares estimator, define ⎡ x(i) [1]T ⎤ ⎡ x(i) [0]T u(i) [0]T ⎤ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ (i) (i) ⋮ ⋮ ⋮ ⎥, X = ⎢ ⎥, Y =⎢ ⎢ (i) ⎥ ⎢ (i) ⎥ T T (i) T ⎢x [T ] ⎥ ⎢x [T −1] ⎥ u [T −1] ⎣ ⎦ ⎣ ⎦ ⎡ w(i) [0]T ⎤ ⎥ ⎢ ⎥ ⎢ ⎥. ⋮ (3) W (i) = ⎢ ⎥ ⎢ (i) T ⎢w [T − 1] ⎥ ⎦ ⎣ for every sample trajectory i = 1, 2, ..., d. Furthermore, let Y , X, and W be defined as vertical concatenations of Y (i) , X (i) , and W (i) for i = 1, 2, ..., d, respectively. Finally, denote Θ = [A B]T as the unknown system parameter and Θ∗ as its true value. Based on these definitions, it follows from (1) that Y =X ⋅Θ+W (4) The system identification problem is then reduced to estimating Θ based on the observation matrix Y and the design matrix X. Consider the following least-squares estimator: Θls = arg min ∥Y − XΘ∥2F (5) Θ One can easily verify the equivalence of (2) and (5). The optimal solution of 5 can be written as Θls = (X T X)−1 X T Y = Θ∗ + (X T X)−1 X T W (6) Notice that Θls is well-defined and unqiue if and only if X T X is invertible, which necessitates d ≥ n+m. The estimation error is then defined as E = Θls − Θ∗ = (X T X)−1 X T W (7) Thus, one needs to study the behavior of (X X) X W in order to control the estimation error of the least-squares estimator. However, since the state of the system at time t is affected by random input disturbances at times 0, 1, ...t − 1, X and W are correlated, which renders (7) hard to analyze. In T −1 T 3 order to circumvent this issue, [18] simplifies the estimator and considers only the state of the system at time T in Y (i) . By ignoring the first T − 1 rows in Y (i) , X (i) , and W (i) , one can ensure that the random matrix (X T X)−1 X T is independent of W . Therefore, it is assumed in the sequel that ⎡x(1) [T ]T ⎤ ⎡x(1) [T −1]T u(1) [T −1]T ⎤ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥, ⋮ ⎥, ⋮ ⋮ Y =⎢ X =⎢ ⎥ ⎢ (d) ⎥ ⎢ (d) T T (d) T ⎢x [T ] ⎥ ⎢x [T −1] u [T −1] ⎥ ⎦ ⎣ ⎦ ⎣ ⎡w(1) [T − 1]T ⎤ ⎥ ⎢ ⎥ ⎢ ⋮ ⎥ W =⎢ (8) ⎥ ⎢ (d) T ⎢w [T − 1] ⎥ ⎦ ⎣ With this simplification, [18] shows that, with input vectors u(i) [t] chosen randomly from N (0, Σu ) for every t = 1, 2, ..., T − 1 and i = 1, 2, ..., d, the least-squares estimator requires at least √ d = Ω(m+n) sample trajectories to guarantee ) with high probability. In what follows, ∥E∥2 = O ( m+n d a regularized estimator will be introduced that exploits the underlying sparsity structure of the system dynamics to significantly reduce the number of sample trajectories for accurate estimation of the parameters. III. M AIN R ESULTS Consider the following variant of (5): 1 (9) Θlasso = arg min ∥X − Y Θ∥2F + λd ∥Θ∥1 Θ 2d where the estimator is now regularized by the sparsitypromoting ℓ1 /ℓ1 penalty term. Under the sparsity assumption on (A, B), we will show that the non-asymptotic statistical properties of Θlasso significantly outperform those of Θls . In particular, the primary objective is to prove that ∥Θlasso −Θ∗ ∥∞ √ decreases at the rate O( log(n + m)/d) with high probability with an appropriate scaling of the regularization coefficient and the sparsity assumption on Θ∗ . The second goal is to show that for a large class of regularization coefficients, only Ω(kmax log(n + m)) sample trajectories are needed to guarantee the uniqueness of the solution and a small estimation error of the Lasso estimator. Here, kmax is the maximum number of nonzero elements in the columns of [A B]T . Comparing this number with the required lower bound Ω(n + m) on the number of sample trajectories for the least-squares estimator, we conclude that the proposed method needs significantly less number of samples when A and B are sparse. The third objective is to prove that this method is able to find the correct sparsity structure of A and B with high probability. In contrast, it will be shown that the solution of the least-squares estimator is fully dense for any finite number of sample trajectories and, hence, it cannot correctly extract the sparsity structures of A and B. We will showcase the superior performance of the Lasso estimator in both sparsity identification and estimation accuracy in simulations. Before presenting the main result of this work, note that x(i) [T −1]=AT −2 Bu(i) [0]+AT −3 Bu(i) [1]+⋯+Bu(i) [T−2] T −2 +A w (i) [0]+A T −3 w (i) [1]+⋯+w (i) [T−2] (10) where, without loss of generality, the initial state is assumed to be zero for every sample trajectory. The results can be readily extended to the case where the initial state is an unknown random vector with Gaussian distribution. Recalling that u(i) [t] and w(i) [t] are i.i.d samples of N (0, Σu I) and N (0, Σw I), respectively, (10) and (8) imply that T Xi,∶ ∼ N (0, Σ̃) (11) where Xi,∶ is the ith row of X and F Σ F T + GΣw GT Σ̃ = [ u 0 0 ] Σu F = [AT −2 B AT −3 B ⋯ B] G = [A T −2 T −3 A ⋯ I] (12a) (12b) (12c) Define Aj as the index set of the rows corresponding to the nonzero elements in the j th column of Θ∗ . The complement of Aj is denoted by Acj . The following assumption plays a key role in deriving the main result of this paper: Assumption 1 (Mutual Incoherence Property). There exists a number γ ∈ (0, 1] such that max {maxc {∥Σ̃iAj (Σ̃Aj Aj )−1 ∥1 }} ≤ 1 − γ 1≤j≤n i∈Aj (13) The mutual incoherence property is a commonly known assumption for the exact recovery of unknown parameters in compressive sensing and classical Lasso regression problems [19]–[22]. This assumption entails that the effect of those submatrices of Σ̃ corresponding to zero (unimportant) elements of Θ on the remaining entries of Σ̃ should not be large. Roughly speaking, this condition guarantees that the unknown parameters are recoverable in the noiseless scenario, i.e. when W = 0. If the recovery cannot be guaranteed in the noise-free setting, then there is little hope for the Lasso estimator to recover the true structure of A and B when the system is subject to noise. Assumption 2. For every 1 ≤ j ≤ n, kj has the order Ω ((n + m)ǫj ) for some ǫj ∈ (0, 1]. Note that this assumption adds a very mild restriction: the growth rate of the number of nonzero elements in each column of Θ∗ should be at least a polynomial function of the dimension of the system. As it will be shown later, this assumption is required to guarantee the consistency of the Lasso estimator in the high-dimensional setting. Next, we define some notations to streamline the presentation of the main theorem. Λmin is used to denote the minimum eigenvalue of Σ̃. Furthermore, define kmax =max ∣Aj ∣, q = max {max ∥(Σ̃−1 Aj Aj)iAj ∥1}, ǫmin= min ǫj 1≤j≤n 1≤j≤n i∈Aj 1≤j≤n (14) We assume that Λmin and q are uniformly bounded away from 0 and +∞. Theorem 1. Given arbitrary constants c1 , c2 > 2 and c3 > 4 1 2ǫmin + 1, suppose that d and λd satisfy the inequalities √ 32c1 σw σG 1 + log(n + m) λd ≥ ⋅ (15) γ2 d 72c2 (16) d ≥ 2 ⋅ κ(Σ̃) ⋅ kmax ⋅ (1 + log(n + m)) γ Then, with probability of at least K1 K2 K3 − − →1 (n + m)c1 −2 (n + m)c2 −2 (n + m)2ǫmin (c3 −1)−1 (17) for some universal constants K1 , K2 , K3 > 0, the following statements hold: ● Statement 1: Θlasso is unique and in addition, 1− where g= ● √ ∥Θlasso −Θ∗ ∥∞ ≤ g 36c3 σw log(kmax ) 8kmax +λd ( √ +q) Λmin d Λmin d (18) (19) Statement 2: The support of Θlasso is a subset of the support of Θ∗ . Furthermore, the supports of Θlasso and Θ∗ are the same if θmin > g, where θmin = min1≤j≤n {mint∈Aj ∣Θtj ∣}. Proof. See Appendix. Theorem 1 states that the support of Θlasso will not contain any false positives with a relatively small number of sample trajectories. Furthermore, it shows that if the decrease rate of the smallest nonzero element in Θ is slow enough, Θlasso will also exclude false negatives. Another important observation can be made from Theorem 1. Notice that the minimum number of required sample trajectories depends on κ(Σ̃). Upon assuming that Σu and Σw are identity matrices, κ(Σ̃) is reduced to κ(F F T + GGT ). F and G are commonly known as finite time controllability matrices for the input and disturbance noise, respectively. Roughly speaking, κ(F F T + GGT ) quantifies the ratio between the eigenvalues of the easiest- and hardest-to-identify modes of the system. Therefore, Theorem 1 implies that only a small number of samples is required to accurately identify the dynamics of the system if all of its modes are easily excitable. The dependency of the estimation error on the modes of the system is also reflected in the non-asymptotic error bound of the least-squares estimator in [18]. This is completely in line with the conventional results on the identifiability of dynamical systems: independent of the proposed method, it is significantly harder to identify the parameters of the system accurately if it possesses nearly-hidden modes. The connection between the identifiability of the system and the number of required sample trajectories to guarantee a small estimation error will be elaborated through different case studies in Section IV. As mentioned before, Theorem 1 implies that, with an appropriate scaling of the regularization coefficient, d = Ω(kmax log(n + m)) suffices to have a unique solution for the Lasso estimation problem and to guarantee its small elementwise error with high probability. The following corollary shows that, with a factor of increase in the number of sample trajectories, the error rate of Lasso estimator becomes significantly smaller in the sparse setting. Corollary 1. Suppose that kmax = o(d1/2 ). Then, d = 2 Ω(kmax ) is enough to obtain √ ⎛ log(n + m) ⎞ ∗ (20) ∥Θlasso − Θ ∥∞ = O d ⎠ ⎝ with probability of at least (17). Proof. According to (15), we have √ ⎛ log(n + m) ⎞ (21) λd = O d ⎠ ⎝ √ Using d = Ω(kmax ), it yields that √ √ ⎛ log(n + m) ⎞ kmax log(n + m) kmax λd ⋅ √ = O( )=O d d ⎝ ⎠ d (22) Combining this with (18) and noting that kmax = O(n + m) concludes the proof. Corollary 1 introduces settings in which our theoretical bounds perform the best. In particular, when Ω((n+m)ǫmin ) = kmin ≤ kmax = O((n + m)ǫmax ) and 0 < ǫmin ≤ ǫmax ≪ 1/2, the Lasso estimator requires a significantly smaller number of samples than Ω(m + n) to guarantee infinitesimal elementwise estimation error. However, notice that Theorem 1 is more general than this corollary and proves the elementwise consistency of the Lasso estimator even in the absence of such assumptions. While it is shown that the proposed estimator can recover the correct sparsity structure of Θ with a relatively small number of sample trajectories, we prove in the next theorem that the least-squares estimator defined as (7) does not extract the correct sparsity structure of Θ for any finite number of sample trajectories. Theorem 2. If A and B are not fully dense matrices, Θls does not recover the support of Θ∗ for any finite number of sample trajectories with probability one. Proof. Define R = ((X T X)−1 X T )T and note that R and W are independent random variables due to the construction of X. Now, suppose that Θ∗ij = 0. We show that with probability zero, it holds that Eij = ∣(Θls )ij − Θ∗ij ∣ = 0 and hence T (Θls )ij = 0. Note that Eij = R∶,i W∶,j . If R∶,i =/ 0, then Eij is a linear combination (with at least one nonzero coefficient) of identically distributed normal random variables with mean zero and variance (Σw )jj . Since R∶,i and W∶,j are independent, we have Eij = 0 with probability zero. Now, assume that R∶,i = 0. This means that the ith row of RT is a zero vector. This, in turn, implies that the ith row of RT X is zero. However, RT X = (X T X)−1 X T X = I, which is a contradiction. This completes the proof. Remark 1. In this work, we considered those LTI systems that are subject to input disturbance with Gaussian distribution. In many areas within control theory, including robust and 5 model predictive control, it is more practical to assume that the input disturbances belong to a bounded set [23], [24]. Moreover, note that this work is based on random generation of inputs from a user-defined Gaussian distribution. Again, this may not be possible since in many applications, the input is constrained to be bounded in order to guarantee the well-being of the entire system. However, the general methodology of this paper is valid in the bounded disturbance-bounded input regime. This is due to the fact that many (although not all) nonasymptotic and high-dimensional properties of estimators involving Gaussian random variables can be carried over to the problems with bounded-support random variables; they benefit from the light-tailed properties of sub-Gaussian probability distributions. Due to the space limitations, these extensions are not included in this work. Similarly, the extension to noisy measurements is left out. IV. N UMERICAL R ESULTS In this section, we illustrate the performance of the Lasso estimator and compare it with its least-square counterpart. We consider two case studies on synthetically generated systems and physical mass-spring networks. The simulations are run on a laptop computer with an Intel Core i7 quad-core 2.50 GHz CPU and 16GB RAM. The reported results are for a serial implementation in MATLAB R2017b and the function lasso is used to solve 9. Define the mismatch error as the total number of false positives and false negatives in the sparsity pattern of the estimator. Moreover, define relative number of sample trajectories (RST) as the number of sample trajectories normalized by the dimension of the system, and relative mismatch error (RME) as the mismatch error normalized by total number of elements in Θ. A. Synthetically Generated Systems Given the numbers n and w, and for each instance of the problem, the state and input matrices are constructed as follows: The diagonal elements of A ∈ Rn×n and B ∈ Rn×n are set to 1 (the dimensions of the inputs and states are chosen to be equal); the elements of the first w upper and lower diagonals of A and B are set to 0.3 or −0.3 with equal probability. Furthermore, at each row of A, another w elements are randomly chosen from the elements not belonging to the first w upper and lower diagonals and set to 0.3 or −0.3 with equal probability. The structure of [A B] is visualized in Figure 1 for w = 2, 3, 4. Based on this procedure, the number of nonzero elements at each row of [A B] varies between 3w + 2 and 5w + 2. Assuming that Σu = I and Σw = 0.5I, the mutual incoherence property is satisfied for all constructed instances. To verify the developed theoretical results, λd is set to √ 2(1 + log(n + m)) (23) d Note that this choice of λd does not require any additional fine-tuning and is at most a constant factor away from the lower bound introduced in (15). In the first set of experiments, we consider the mismatch error of Θlasso with respect to the number of sample trajectories and for different system dimensions. The length of the time horizon T is set to 3. The results are illustrated in Figure 2a for n + m equal to 200, 600, 1200, and 2000. In all of these test cases, w is chosen in such a way that the number of nonzero elements in each column of Θ is between (n + m)0.3 and (n + m)0.4 . It can be observed that, as the dimension of the system increases, a higher number of sample trajectories is required to have a small mismatch error in the Lasso estimator. Conversely, the required value of RST to achieve a small RME reduces as the dimension of the system grows. More precisely, RST should be at least 1.80, 1.13, 0.37, and 0.20 to guarantee RME ≤ 0.1%, when m + n is equal to 200, 600, 1200, and 2000, respectively. In the next set of experiments, we consider the mismatch error for different time horizons T = 3, 4, ..., 7, when fixing m+ n and w to 600 and 2, respectively. As mentioned before, large values of T tend to inflate the easily identifiable modes of the system and suppress the nearly hidden ones, thereby making it hard to obtain an accurate estimation of the parameters. It is pointed out that κ(F F T + GGT ) is a good indicator of the gap between these modes. This relationship is clearly reflected in Figures 2b and 2c. As can be observed in Figure 2b, 330 sample trajectories are enough to guarantee RME ≤ 0.1% for T = 3. However, for T = 7, RME cannot be reduced below 0.42% with 1000 sample trajectories. To further elaborate on this dependency, Figure 2c is used to illustrate the value of κ(F F T + GGT ) with respect to T in a log-log scale. One can easily verify that κ(F F T + GGT ) associated with T = 7 is 485 times greater than this parameter for T = 3. Finally, we study the Lasso estimator for different percolumn number of nonzero elements in Θ and compare its accuracy to the least-squares estimator. Fixing T = 3 and m+n = 600, Figure 3a depicts the mismatch error of the Lasso estimator when the maximum number of nonzero elements at each column of Θ ranges from 7 (corresponding to w = 1) to 27 (corresponding to w = 5). Not surprisingly, the required number of samples to achieve a small mismatch error increases as the number of nonzero elements in each column of Θ grows. On the other hand, the least-squares estimator is fully dense in all of these experiments, regardless of the number of sample trajectories. To have a better comparison between the two estimators, we consider the 2-norm of the estimation errors normalized by the 2-norm of Θ∗ , for different number of nonzero elements in each column of Θ∗ . As it is evident in Figure 3b, the Lasso estimator significantly outperforms the least-squares estimator for any number of sample trajectories. Furthermore, the least-squares estimator is not defined for d < 600. B. Mass-Spring Systems In this case study, we conduct simulations on mass-spring networks with different sizes to elucidate the performance of the Lasso estimator on physical systems. Consider N identical masses connected via a path of springs. Figure 4 exemplifies this system for N = 2. The state-space equation of this system in continuous domain can be written as ẋc (t) = Ac x(t) + Bc u(t) (24) 6 0 50 100 150 200 250 300 0 100 200 300 400 500 600 nz = 6526 (a) w = 2 (b) w = 3 Fig. 1: The sparsity structure of the matrix [A B] for w = 2, 3, 4. 10 4 5000 1.5 1 0.5 0 500 1000 1500 10 T=3 T=4 T=5 T=6 T=7 4000 cond(FFT+GGT) n+m = 200 n+m = 600 n+m = 1200 n+m = 2000 Number of mismatches Number of mismatches 2 (c) w = 4 3000 2000 10 10 5 10 0 1000 0 2000 200 400 Number of samples 600 800 1000 4 6 Number of samples (a) Mismatch error for different d 8 10 12 14 T (b) Mismatch error for different T (c) Condition number for different T Fig. 2: (a) The mismatch error with respect to the number of sample trajectories for different system dimensions, (b) the mismatch error with respect to the number of sample trajectories for different time horizons, (c) the condition number of F F T + GGT with respect to the time horizon. k max = 12, (n+m)0.39 10000 0.44 k max = 17, (n+m) 8000 k max = 22, (n+m)0.48 0.52 6000 k max = 27, (n+m) 4000 8 6 k max = 12, Lasso k max = 12, LS k max = 17, Lasso k max = 17, LS k max = 22, Lasso k max = 22, LS 4 2 2000 0 10 k max = 7, (n+m)0.30 Estimation error Number of mismatches 12000 200 400 600 800 1000 Number of samples 0 200 400 600 800 1000 Number of samples (a) Mismatch error for different kmax (b) Estimation error for different kmax Fig. 3: (a) The mismatch error with respect to the number of sample trajectories for different per-column number of nonzero elements in Θ∗ , (b) the normalized estimation error for Lasso and least-square (abbreviated as LS) estimators with respect to the number of sample trajectories. where Ac ∈ R2N ×2N , Bc ∈ R2N ×N , and xc (t) consists of two parts: the first N elements correspond to the location of the masses while the second N elements capture their velocity. With the unit masses and spring constants, we have Ac = [ 0 S I ], 0 0 Bc = [ ] I (25) where S ∈ Rn×n is a tridiagonal matrix whose diagonal elements are set to −2 and its first upper and lower diagonal elements are set to 1 [25]. This continuous system is discretized using the forward Euler method with sampling time of 0.2s. Similar to the previous case, Σu and Σw are set to I and 0.5I, respectively. Furthermore, T is set to 3 and λd is chosen as 23. The mutual incoherence property holds in all of these simulations. Notice that RST is equal to 3d/N in this case study, since n = 2N and m = N . Figure 5a depicts the required RST to achieve RME ≤ 0.1% for different number of masses. If N = 30, the minimum number of sample trajectories to guarantee a small mismatch error is 5 times the dimension of the system, whereas this number is dropped to 0.02 for N = 500. Furthermore, Figure 5b shows the normalized estimation errors of the Lasso and leastsquares estimators for fixed N = 300 and T = 3. Similar to the previous case study, the proposed estimator results in significantly smaller estimation errors in medium- and largesampling regime, while it remains as the only viable estimator when the number of available sample trajectories is smaller than 900. V. C ONCLUSION We consider the problem of identifying the parameters of linear time-invariant (LTI) systems. In many real-world 7 Fig. 4: Mass-spring system with two masses problems, the state-space equation describing the evolution of the system admits a sparse representation, due to localized or internally limited interactions of states and inputs. In this work, we leverage this property and introduce a Lasso-type estimator to identify the parameters of the system. Using modern high-dimensional statistics, we derive sharp nonasymptotic bounds on the minimum number of input-state data samples to guarantee a small elementwise estimation error. In particular, upon defining n and m as the respective numbers of states and inputs, and kmax as the maximum number of nonzero elements in the rows of input and state matrices, we prove that Ω(kmax log(n + m)) data samples suffice to ensure infinitesimal elementwise estimation error and exact recovery of the sparsity structure of the system. Through different case studies on synthetically generated systems and mass-spring networks, we demonstrate substantial improvements in the accuracy of the proposed estimator, compared to its wellknown least-squares counterpart. [18] S. Dean, H. Mania, N. Matni, B. Recht, and S. Tu, “On the sample complexity of the linear quadratic regulator,” arXiv preprint arXiv:1710.01688, 2017. [19] P. Zhao and B. Yu, “On model selection consistency of lasso,” Journal of Machine learning research, vol. 7, no. Nov, pp. 2541–2563, 2006. [20] N. Meinshausen and P. Bühlmann, “High-dimensional graphs and variable selection with the lasso,” The annals of statistics, pp. 1436–1462, 2006. [21] D. L. Donoho, “For most large underdetermined systems of linear equations the minimal 1-norm solution is also the sparsest solution,” Communications on pure and applied mathematics, vol. 59, no. 6, pp. 797–829, 2006. [22] E. Candes and J. Romberg, “Sparsity and incoherence in compressive sampling,” Inverse problems, vol. 23, no. 3, p. 969, 2007. [23] D. Q. Mayne, M. M. Seron, and S. Raković, “Robust model predictive control of constrained linear systems with bounded disturbances,” Automatica, vol. 41, no. 2, pp. 219–224, 2005. [24] G. E. Dullerud and F. Paganini, A course in robust control theory: a convex approach. Springer Science & Business Media, 2013, vol. 36. [25] F. Lin, M. Fardad, and M. R. Jovanović, “Design of optimal sparse feedback gains via the alternating direction method of multipliers,” IEEE Transactions on Automatic Control, vol. 58, no. 9, pp. 2426–2431, 2013. [26] S. N. Negahban and M. J. Wainwright, “Simultaneous support recovery in high dimensions: Benefits and perils of block ℓ1 /ℓ∞ -regularization,” IEEE Transactions on Information Theory, vol. 57, no. 6, pp. 3841– 3863, 2011. [27] M. J. Wainwright, “Sharp thresholds for high-dimensional and noisy sparsity recovery using ℓ1 -constrained quadratic programming (lasso),” IEEE transactions on information theory, vol. 55, no. 5, pp. 2183–2202, 2009. R EFERENCES [1] L. Ljung, “System identification,” in Signal analysis and prediction. Springer, 1998, pp. 163–173. [2] L. Ljung, “Convergence analysis of parametric identification methods,” IEEE transactions on automatic control, vol. 23, no. 5, pp. 770–783, 1978. [3] R. Pintelon and J. Schoukens, System identification: a frequency domain approach. John Wiley & Sons, 2012. [4] E.-W. Bai, “Non-parametric nonlinear system identification: An asymptotic minimum mean squared error estimator,” IEEE Transactions on automatic control, vol. 55, no. 7, pp. 1615–1626, 2010. [5] I. M. Johnstone, “On the distribution of the largest eigenvalue in principal components analysis,” Annals of statistics, pp. 295–327, 2001. [6] D. R. Cox and D. V. Hinkley, Theoretical statistics. CRC Press, 1979. [7] S. Ross and J. A. Bagnell, “Agnostic system identification for modelbased reinforcement learning,” arXiv preprint arXiv:1203.1007, 2012. [8] S. Sadraddini and C. Belta, “Formal guarantees in data-driven model identification and control synthesis,” in 21st ACM International Conference on Hybrid Systems: Computation and Control. ACM, 2018. [9] Z. Hou and S. Jin, “Data-driven model-free adaptive control for a class of mimo nonlinear discrete-time systems,” IEEE Transactions on Neural Networks, vol. 22, no. 12, pp. 2173–2188, 2011. [10] E. Weyer, R. C. Williamson, and I. M. Mareels, “Finite sample properties of linear model identification,” IEEE Transactions on Automatic Control, vol. 44, no. 7, pp. 1370–1383, 1999. [11] E. Weyer, “Finite sample properties of system identification of arx models under mixing conditions,” Automatica, vol. 36, no. 9, pp. 1291– 1299, 2000. [12] J. Pereira, M. Ibrahimi, and A. Montanari, “Learning networks of stochastic differential equations,” in Advances in Neural Information Processing Systems, 2010, pp. 172–180. [13] S. Dean, H. Mania, N. Matni, B. Recht, and S. Tu, “On the sample complexity of the linear quadratic regulator,” arXiv preprint arXiv:1710.01688, 2017. [14] S. Tu, R. Boczar, A. Packard, and B. Recht, “Non-asymptotic analysis of robust control from coarse-grained identification,” arXiv preprint arXiv:1707.04791, 2017. [15] G. Darivianakis, S. Fattahi, J. Lygeros, and J. Lavaei, “High-performance cooperative distributed model predictive control for linear systems,” to appear in American Control Conference, 2018. [16] Y.-S. Wang, N. Matni, and J. C. Doyle, “Localized lqr optimal control,” in IEEE 53rd Conference on Decision and Control, 2014, pp. 1661– 1668. [17] G. Fazelnia, R. Madani, A. Kalbat, and J. Lavaei, “Convex relaxation for optimal distributed control problems,” IEEE Transactions on Automatic Control, vol. 62, no. 1, pp. 206–221, 2017. A PPENDIX In this section, we present the proof of Theorem 1. First, note that (4) can be reformulated as the set of linear equations Y∶,j = XΘ∶,j + W∶,j ∀j = 1, ..., n (26) where Y∶,j , Θ∶,j , and W∶,j are the j th column of Y , Θ, and W , respectively. Based on this definition, consider the following set of Lasso regression subproblems: Θjlasso = arg min Θ∶,j ∈Rd Furthermore, define 1 ∥Y∶,j − XΘ∶,j ∥22 + λd ∥Θ∶,j ∥1 2d γj = 1 − max {∥Σ̃iAj (Σ̃Aj Aj )−1 ∥1 } c i∈Aj (27) (28) for every j = 1, 2, ..., n. Note that we have 0 < γj ≤ 1 due to Assumption 1. The following important lemma holds for the newly defined subproblems (27). Lemma 1. Given arbitrary constants c1 , c2 , c3 > 1 and for every 1 ≤ j ≤ n, suppose that d and λd satisfy ¿ Á 32c1 ⋅ (Σw )jj ⋅ σG 1 + log(n + m) À λd ≥ Á ⋅ (29) γj2 d 72c2 d ≥ 2 ⋅ κ(Σ̃) ⋅ kj ⋅ (1 + log(n + m)) (30) γj Then, with probability at least 1 − 3 exp ( − (c1 − 1)(1 + log(n + m))) − 2 exp ( − (c2 − 1)(1 + log(n + m))) kj − 2 exp ( − 2(c3 − 1)(log(kj ))) − 6 exp (− ) → 1 2 the following statements hold: (31) 8 20 5 Estimation error d/(n+m) 4 3 2 1 0 100 200 300 400 500 Number of masses 15 10 5 0 500 1000 1500 Number of samples (a) Relative mismatch error (b) Normalized estimation error Fig. 5: (a) The minimum RST to guarantee RME ≤ 0.1% with respect to the number of masses, (b) the normalized estimation error for Lasso and least-square estimators with respect to the number of sample trajectories. 1. Θjlasso is unique and in addition, √ 8kj 36c3 (Σw )jj log(kj ) j ∥Θlasso −Θ∶,j ∥∞ ≤ +λd ( √ +q) Λmin d Λmin d ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ gj (32) 2. The support of Θjlasso is a subset of the support of Θ∶,j . Furthermore, the supports of Θjlasso and Θ∶,j are identical j j if θmin > gj , where θmin = mint∈Aj ∣Θtj ∣. Proof. The proof is based a the primal-dual witness approach introduced in [26], [27]. The details are omitted due to the space restrictions. Proof of Theorem 1: First, we present the sketch of the proof in a few steps: 1. We decompose the multivariate lasso regression problem (9) into n disjoint lasso regression subproblems defined in (27). 2. For each of these Lasso regression subproblems, we consider the event that Statements 1 and 2 of Theorem 1 hold. 3. We consider the intersection of these n events and show that they hold simultaneously with a probability that is lower bounded by (17). Step 1: (9) can be rewritten as follows: n Θlasso = arg min ∑ ( Θ j=1 1 ∥X∶,j − Y Θ∶,j ∥22 + λ∥Θ∶,j ∥1 ) 2d (33) The above optimization problem can be naturally decomposed into n disjoint lasso regression subproblems in the form of (27). Step 2: Assume that (30) and (29) hold for every 1 ≤ j ≤ n. Upon defining Tj as the event that the first and second statements of Lemma 1 hold, one can write P(Tj ) ≥ 1 − 3 exp ( − (c1 − 1)(1 + log(n + m))) − 2 exp ( − (c2 − 1)(1 + log(n + m))) − 2 exp ( − 2(c3 − 1)(log(kj ))) − 6 exp (− For every 1 ≤ j ≤ n. kj ) 2 (34) Step 3: Consider the event T = T1 ∩ T2 ∩ ⋯ ∩ Tn . Based on the Frechet inequality, we have P(T1 ) + P(T2 ) + ⋯ + P(Tn ) − (n − 1) ≤ P(T ) (35) Combining (34) and (35) yields that P(T ) ≥ 1 − 2 ∑ exp ( − 2(c3 − 1)(log(kj ))) n ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ j=1 (a) − 3n exp ( − (c1 − 1)(1 + log(n + m))) ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ (b) − 2n exp ( − (c2 − 1)(1 + log(n + m))) ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ (c) kj ) 2 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ − 6n exp (− (36) (d) Denote N as the event that Statements 1 and 2 of the theorem hold. Note that the lower bounds introduced in (16) and (15) for d and λd guarantee the validity of (30) and (29) for every 1 ≤ j ≤ n. Furthermore, based on (18) and (32), we have g ≥ gj for every 1 ≤ j ≤ n. This implies that P(T ) ≤ P(N ). It remains to show that the right hand side in (36) is lower bounded by (17). We have n (a) ≤ 2 ⋅ (37a) 2ǫ (n + m) min (c3 −1) n (37b) (b) ≤ 3 exp(−(c1 − 1)) (n + m)c1 −1 n (c) ≤ 2 exp(−(c2 − 1)) (37c) (n + m)c2 −1 where (a) is due to the assumption kmin = Ω((n + m)ǫmin ) and (b) and (c) follow from a simple calculation. In light of Assumption 2, one can write kmax = Ω((n+m)ǫmax ) for some ǫmax > 0. This yields that (d) is dominated by the right hand sides of (37a), (37b), and (37c). The proof is completed by 1 + 1 are enough to ensure noting that c1 , c2 > 2 and c3 > 2ǫmin that (a), (b), (c), and (d) converge to zero as the dimension of the system grows.
3cs.SY
arXiv:1802.10107v1 [math.CO] 27 Feb 2018 DOMINANCE PHENOMENA: MUTATION, SCATTERING AND CLUSTER ALGEBRAS NATHAN READING Abstract. An exchange matrix B dominates an exchange matrix B 0 if the signs of corresponding entries weakly agree, with the entry of B always having weakly greater absolute value. When B dominates B 0 , interesting things happen in many cases (but not always): the identity map between the associated mutation-linear structures is often mutation-linear; the mutation fan for B often refines the mutation fan for B 0 ; the scattering (diagram) fan for B often refines the scattering fan for B 0 ; and there is often an injective homomorphism from the principal-coefficients cluster algebra for B 0 to the principal-coefficients cluster algebra for B, preserving g-vectors and sending the set of cluster variables for B 0 (or an analogous larger set) into the set of cluster variables for B (or an analogous larger set). The scope of the description “often” is not the same in all four contexts and is not settled in any of them. In this paper, we prove theorems that provide examples of these dominance phenomena. Contents 1. Introduction 1.1. Mutation-linear maps 1.2. Mutation fans 1.3. Scattering diagrams 1.4. Ring homomorphisms between cluster algebras 1.5. Related work 2. Mutation-linear algebra 2.1. Partial linear structures 2.2. Mutation-linear structures 2.3. Mutation-linear maps 2.4. Mutation-linear algebra and universal coefficients 2.5. Examples of mutation-linear maps 3. Mutation-linearity and refinement of the mutation fan 3.1. Connecting Phenomenon I and Phenomenon II 3.2. Erasing edges 3.3. The 2 × 2 case 3.4. Acyclic finite type 3.5. Resection of marked surfaces 4. Refinement of scattering fans 4.1. Scattering fans 4.2. Scattering fans refine mutation fans 2 2 4 6 7 10 10 10 12 14 17 17 19 19 20 20 23 25 38 39 40 2010 Mathematics Subject Classification. Primary: 13F60, 52C99; Secondary 05E15, 20F55, 57Q15. This paper is based upon work partially supported by the National Science Foundation under Grant Numbers DMS-1101568 and DMS-1500949. 1 2 NATHAN READING 4.3. The 2 × 2 case 5. Morphisms of cluster algebras 5.1. General considerations 5.2. The 2 × 2 case 5.3. Acyclic finite type 5.4. Resection of surfaces 41 42 42 44 49 50 Acknowledgments References 67 67 1. Introduction An exchange matrix is a skew-symmetrizable integer matrix. Given n × n exchange matrices B = [bij ] and B 0 = [b0ij ], we say B dominates B 0 if for each i and j, we have bij b0ij ≥ 0 and |bij | ≥ |b0ij |. This paper explores the consequences of the dominance relationship between B and B 0 for the mutation-linear algebra, mutation fans, scattering diagrams, and principal-coefficients cluster algebras associated to B and B 0 . We present our results by describing several phenomena that often occur when B dominates B 0 . We call these “phenomena” because the hypotheses are not yet nailed down and because there are negative examples (where a phenomenon does not occur). The goal of this paper is to prove theorems that give compelling and surprising examples of the phenomena. By providing examples of the phenomena, we wish to establish that something real and nontrivial is happening, with an eye towards two potential benefits: In one direction, we anticipate that researchers from the various areas will apply their tools to find additional examples of the phenomena, necessary and/or sufficient conditions for the phenomena to occur, and/or additional dominance phenomena. In the other direction, since the phenomena discussed here concern fundamental aspects of matrix mutation, the geometry of scattering diagrams, and the commutative algebra of cluster algebras, we anticipate that the phenomena will lead to insights in the various areas where matrix mutation, scattering diagrams, and cluster algebras are fundamental. The rest of the introduction is devoted to describing the phenomena and stating the main results. 1.1. Mutation-linear maps. The study of mutation-linear algebra was initiated in [22] and continued in [23, 24, 2]. The “mutation” in “mutation-linear” is matrix mutation in the sense of cluster algebras. (See [10, Definition 4.2] or Section 2.2.) To put mutation-linear algebra into context, consider the formulation of (ordinary) linear algebra as the study of linear relations on a module. For example, the notion of basis is defined in terms of the existence and non-existence of certain linear relations.P As another example, aPmap λ : V → V 0 is linear if for every linear relation i∈S ci vi in V , the sum i∈S ci λ(vi ) is a linear relation in V 0 . In the same sense, mutation-linear algebra is the study of B-coherent linear 1 n relations. P A B-coherent linear relation among vectors in R is a linear relation i∈S ci vi with the following property: If the vi are placed as coefficient 1Here in the introduction, we sweep a technicality under the rug. See Definition 2.5. DOMINANCE PHENOMENA 3 rows under B and the resulting extended exchange matrix is subjected to an arbitrary sequence of mutations to produce new coefficient rows vi0 , then the relation P 0 B is shorthand for the mutation-linear i∈S ci vi = 0 holds. The notation R n n structure on R , meaning the set R together with the collection of B-coherent linear relations. Mutation-linear algebra is closely tied to cluster algebras. For example, finding a basis for RB is the same thing as finding a cluster algebra for B with universal (geometric) coefficients [22, Theorem 4.4]. The key mutation-linear B notion in this paper is the notion of a mutation-linear map. A map P λ from R B0 to R P is mutation-linear if for every B-coherent linear relation i∈S ci vi , the sum i∈S ci λ(vi ) is a B 0 -coherent linear relation. A mutation-linear map can be seen as a defining a functor from the category of geometric cluster algebras with exchange matrix B, under coefficient specialization, to the same category for B 0 . (See Section 2.4.) From the definition of a mutation-linear map, it is not clear whether interesting mutation-linear maps exist. The first phenomenon points out that they often do. Phenomenon I (Identity is mutation-linear). Suppose B and B 0 are exchange matrices such that B dominates B 0 . In many cases, the identity map from RB to 0 RB is mutation-linear. It is easy to find non-examples of Phenomenon I; the surprise is that the phenomenon occurs as much as it does. A simple non-example is when B comes from a quiver with three vertices and three arrows, forming a directed cycle, and B 0 is defined by deleting one arrow from the cycle. On the positive side, we verify the phenomenon for B acyclic of finite type. We prove the following theorem using results of [25]. Theorem 1.1. Suppose B is an acyclic exchange matrix of finite type, and suppose B 0 is another exchange matrix dominated by B. Then the identity map from RB to 0 RB is mutation-linear. Replacing R with Q in Phenomenon I, we obtain a weaker phenomenon, which we will call the “rational version” of Phenomenon I. (This rational version is equivalent to the “integer version,” where R is replaced with Z. See Section 2.3.) We define a simple operation, called resection, on triangulated surfaces that produces a dominance relation among signed adjacency matrices. We verify the rational version of Phenomenon I in the case of resection of some surfaces. Resection and the Null Tangle Property are defined in Section 3.5. A null surface is a once-punctured monogon, once-punctured digon, or unpunctured quadrilateral. Theorem 1.2. Given a marked surface (S, M) with a triangulation T , perform a resection of (S, M) compatible with T and let T 0 be the triangulation induced by T on the resected surface (S0 , M0 ). Suppose also that every component of (S, M) and (S0 , M0 ) either has the Null Tangle Property or is a null surface. Then the identity 0 map from QB(T ) to QB(T ) is mutation-linear. At the time of writing, the only surfaces known to have the Null Tangle Property (or defined to be null surfaces) are those whose connected components are of the following types: a disk with 0, 1, or 2 punctures, an annulus with 0 or 1 punctures, a sphere with three boundary components and no punctures, a torus with one puncture and no boundary components, and a sphere with four punctures and no boundary components. (See [24, Theorem 3.2] for the once-punctured torus, [2, 4 NATHAN READING Theorem 1.1] for the four-punctured sphere, and [23, Theorem 7.4] for the other surfaces.) Interestingly, this list is closed under resection. This list includes the surfaces of finite type—those whose associated cluster algebra is of finite type. There are no surfaces known (or conjectured) not to have the Null Tangle Property except the null surfaces. Work is in progress by Shira Viel [31] to generalize and extend Theorem 1.2 to the case of orbifolds in the sense of [7]. As another example of Phenomenon I, we say that B 0 = [b0ij ] is obtained from  B = [bij ] by erasing edges if b0ij ∈ 0, bij for every pair i, j. In this case, B dominates B 0 . The reference to edges refers to a graph on the indices of B with an edge connecting two indices i and j if and only if bij 6= 0: The edges “erased” are i—j such that b0ij = 0 6= bij . We can describe B 0 by specifying which edges in the graph of B are erased to obtain B 0 . Theorem 1.3. Suppose the indices of B are written as a disjoint union I ∪ J and suppose B 0 is obtained by erasing all edges of B that connect indices in I to indices 0 in J. Then the identity map from RB to RB is mutation-linear.   Our final example of Phenomenon I concerns 2 × 2 exchange matrices B = 0b a0 with ab ≤ 0, and illustrates the subtlety of the phenomenon. The matrix B is of finite type if and only if ab ≥ −3 and of affine type if and only if ab = −4. Otherwise, we say B is of wild type. The following theorem completely describes Phenomenon I for 2 × 2 exchange matrices. Theorem 1.4. Suppose B and B 0 are 2 × 2 exchange matrices. 0 (1) If B does not dominate B 0 , then id : RB → RB is not mutation-linear. 0 (2) If neither B nor B 0 is wild then id : RB → RB is mutation-linear if and 0 only if B dominates B . 0 (3) If exactly one of B and B 0 is wild, then id : RB → RB is mutation-linear if and only if either  (a) B 0 = h00 00 , or i (b) B 0 = ∓10 ±10 and B dominates B 0 but agrees with B 0 except in one position. 0 (4) If both B and B 0 are wild, then id : RB → RB is not mutation-linear. 1.2. Mutation fans. The mutation fan FB of an exchange matrix B is a fan that encodes the piecewise-linear geometry of mutations of B. (See Definition 2.11.) Given fans F and F 0 with ∪F = ∪F 0 , we say that F refines F 0 if every cone of F is contained in a cone of F 0 , or equivalently if every cone of F 0 is a union of cones of F. If the mutation-linear structure RB is sufficiently well-behaved, then the 0 assertion that the identity is mutation-linear from RB to RB is equivalent to the assertion that the mutation fan FB refines the mutation fan FB 0 (Proposition 3.1). Specifically, “sufficiently well-behaved” means that RB admits a cone basis in the sense of Definition 2.17. We are not aware of any B such that RB does not admit a cone basis, and we expect that the following phenomenon is equivalent to Phenomenon I. Phenomenon II (Refinement of mutation fans). Suppose B and B 0 are exchange matrices such that B dominates B 0 . In many cases, the mutation fan FB refines the mutation fan FB 0 . DOMINANCE PHENOMENA 5 In all of the examples and non-examples described above for Phenomenon I, a cone basis is known to exist for RB . In particular, our proof of Theorem 1.1 proceeds by proving the following equivalent theorem. Theorem 1.5. Suppose B is an acyclic exchange matrix of finite type, and suppose B 0 is another exchange matrix dominated by B. Then the mutation fan FB refines the mutation fan FB 0 . We prove Theorem 1.5 by first relating the mutation fan to the Cambrian fan and then appealing to the analogous refinement theorem for Cambrian fans, proved in [25]. Remark 1.6. This refinement relation among Cambrian fans was the origin of the present study of dominance. It was discovered though lattice-theoretic investigations of the weak order [25], following a clue in work of Simion [30]. This is another example where the lattice theory of the weak order on finite Coxeter groups provides a guide to discovering much more general phenomena. (Compare [20, 21, 28, 29].) Given a triangulation T of a marked surface (S, M) having the Null Tangle Property (or a weaker property called the Curve Separation Property), the rational part of the mutation fan FB(T ) (Definition 3.2) is a fan FQ (T ) called the rational quasi-lamination fan. The cones of FQ (T ) are the real spans of shear coordinates of pairwise compatible collections of certain curves called allowable curves. Since each of these shear coordinates is an integer vector, each cone of FQ (T ) is a rational B cone. Furthermore, we will see that every point  in Q n is contained in some cone of n FQ (T ). Thus if |T | = n, then FQ (T ) ∩ Q = C ∩ Q : C ∈ FQ (T ) is a complete fan in Qn . We prove Theorem 1.2 by proving the following theorem, which is equivalent by Proposition 3.3. Since FQ (T ) is the rational part of FB(T ) , the theorem provides an example of a “rational version” of Phenomenon II. Theorem 1.7. Given a marked surface (S, M) and a triangulation T with |T | = n, perform a resection of (S, M) compatible with T to obtain (S0 , M0 ) and let T 0 be the triangulation of (S0 , M0 ) induced by T . Suppose also that every component of (S, M) and (S0 , M0 ) either has the Null Tangle Property or is a null surface. Then FQ (T ) ∩ Qn refines FQ (T 0 ) ∩ Qn . As in the case of mutation-linear maps, the work of Viel [31] is expected to generalize and extend this example of Phenomenon II coming from orbifolds. Proposition 3.1, mentioned above, also says that, independent of any hypothesis 0 on B (e.g. that RB admits a cone basis), if the identity map RB → RB is mutationlinear, then FB refines FB 0 . Thus we obtain the following theorem as a consequence of Theorem 1.3. Theorem 1.8. Suppose the indices of B are written as a disjoint union I ∪ J and suppose B 0 is obtained by erasing all edges of B that connect indices in I to indices in J. Then FB refines FB 0 . Finally, the results of [22, Section 9] show that RB admits a cone basis for every 2 × 2 exchange matrix B, so Theorem 1.4 implies the following theorem (in light of Proposition 3.1 as above). Theorem 1.9. Suppose B and B 0 are 2 × 2 exchange matrices. (1) If B does not dominate B 0 , then FB does not refine FB 0 . 6 NATHAN READING (2) If neither B nor B 0 is wild then FB refines FB 0 if and only if B dominates B 0 . 0 (3) If exactly one  0 0 of  B and B is wild, then FB refines FB 0 if and only if either 0 (a) B = h 0 0 , or i (b) B 0 = ∓10 ±10 and B dominates B 0 but agrees with B 0 except in one position. (4) If both B and B 0 are wild, then FB does not refine FB 0 unless B = B 0 . 1.3. Scattering diagrams. The mutation fan is related to the cluster scattering diagram of [14]. We will not give a complete definition of cluster scattering diagrams here, but will give enough of the definition and quote results that let us connect them to mutation fans. The (principal coefficients) cluster scattering diagram Scat(B) associated to an exchange matrix B is a certain collection W of walls—codimension-1 cones—together with some additional algebraic data that we will mostly ignore here. (A technical point: The cluster scattering diagram is defined only up to a notion of equivalence, and we take an equivalence-class representative with minimal support.) In a way that is made more precise in Section 4, the walls cut the ambient space into convex cones, and these cones, together with their faces, comprise a complete fan [26, Theorem 3.1] called the (cluster) scattering fan and denoted by ScatFan(B). (As pointed out in [4, Section 3], the cluster scattering diagram is expected to essentially coincide with the fan defined by the spherical semi-invariants picture of [17].) Phenomenon III (Refinement of scattering fans). Suppose B and B 0 are exchange matrices such that B dominates B 0 . In many cases, ScatFan(B) refines ScatFan(B 0 ). Most, but not all, of our examples and nonexamples of Phenomenon III come from the relationship between ScatFan(B) and FB . The most general statement about that relationship is the following theorem. Theorem 1.10. The scattering fan ScatFan(B) refines the mutation fan FB for any exchange matrix B. A comparison of results of [4, Section 3] and [22, Section 9] shows that when B is 2 × 2, the scattering fan ScatFan(B) coincides with the mutation fan FB if and only if B is of finite or affine type. (See [4, Section 3] and [22, Section 9] and compare [26, Section 6]. See also Theorem 3.7.) When B is of finite type, both ScatFan(B) and FB are known to coincide with the g-vector fan for B T , so they coincide. (The fact for ScatFan(B) follows from [14, Lemma 2.10] and [14, Corollary 5.9]. The transpose is not mentioned in [14], but arises from a difference in conventions as explained in [26, Remark 4.2]. The fact for mutation fans is proved in [22, Section 10].) Conjecture 1.11. Given an exchange matrix B, the scattering fan ScatFan(B) coincides with the mutation fan FB if and only if either (i) B is 2 × 2 and of finite or affine type or (ii) B is n × n for n > 2 and of finite mutation type. Conjecture 1.11 is suggested by the 2 × 2 and finite-type examples mentioned above, other small examples, and a characterization of mutation-finiteness given in [6, Theorem 2.8]: For n > 2, an n × n exchange matrix is mutation finite if and only if every matrix [bij ] in its mutation class has bij bji ≥ −4 for all i 6= j. (Recall DOMINANCE PHENOMENA 7 also that a 2 × 2 exchange matrix is of finite or affine type if and only if the product of its two off-diagonal entries is greater than or equal to −4.) Since ScatFan(B) and FB coincide in finite type as explained above, Theorem 1.5 implies the following example of Phenomenon III. Theorem 1.12. Suppose B is an acyclic exchange matrix of finite type, and suppose B 0 is another exchange matrix dominated by B. Then the scattering fan ScatFan(B) refines the scattering fan ScatFan(B 0 ). Also because ScatFan(B) and FB coincide in finite type, the non-example for Phenomenon II (where B comes from a cyclically-directed triangle and B 0 is obtained by deleting one arrow) is a non-example of Phenomenon III as well. More generally, as an immediate consequence of Theorem 1.10, we have the following result which lets us obtain examples of Phenomenon III from examples of Phenomenon II. Theorem 1.13. Suppose that B dominates B 0 and that ScatFan(B 0 ) coincides with FB 0 . If Phenomenon II occurs for B and B 0 , then Phenomenon III also occurs. Theorem 1.13 has the following consequence. Corollary 1.14. Suppose B is an exchange matrix of finite type, and suppose B 0 is another exchange matrix dominated by B such that Phenomenon II occurs. Then the scattering fan ScatFan(B) refines the scattering fan ScatFan(B 0 ). Since finite mutation fans coincide with their rational parts, and since non-null surfaces of finite type have the Null Tangle Property [23, Theorem 7.4], we also have the following corollary. Corollary 1.15. Given a marked surface (S, M) of finite type with a triangulation T , perform a resection of (S, M) compatible with T . Let T 0 be the triangulation induced by T on the resected surface (S0 , M0 ). Then ScatFan(B(T )) refines ScatFan(B(T 0 )). Our final example of Phenomenon III is a departure from the correspondence with Phenomenon II. For any 2 × 2 matrix, the full-dimensional rational cones of the scattering fan coincide with the full-dimensional rational cones of the mutation fan (because both are known to coincide with the g-vector cones of clusters). In wild type, the closure of the complement of the full-dimensional rational cones is a single irrational cone of the mutation fan. It is expected (see [14, Example 1.30]) that every rational ray in this irrational cone is a wall of the scattering diagram. If this expectation is correct, then in particular every ray in the irrational cone is a distinct cone of the scattering fan. We call the expectation that every ray in the irrational cone is a distinct cone of the scattering fan the Discreteness Hypothesis for the purpose of stating the following theorem, which is proved in Section 4.3. Theorem 1.16. Suppose B and B 0 are 2 × 2 exchange matrices. Assuming the Discreteness Hypothesis for 2 × 2 scattering fans, the scattering fan ScatFan(B) refines the scattering fan ScatFan(B 0 ) if and only if B dominates B 0 . 1.4. Ring homomorphisms between cluster algebras. Another phenomenon surrounding dominance of exchange matrices is the existence of ring homomorphisms between cluster algebras with different exchange matrices of the same size, preserving g-vectors. This thought-provoking surprise suggests, in particular, that 8 NATHAN READING the category of principal-coefficients cluster algebras and g-vector-preserving ring homomorphisms may reward study. (Or perhaps one should consider a category with more morphisms, namely ring homomorphism that act linearly on g-vectors.) We start with the acyclic, finite-type version of the phenomenon. Write A• (B) for the principal-coefficients cluster algebra associated to B and recall from above that when B is of finite type, the mutation fan FB coincides with the g-vector fan for B T . When B is acyclic and of finite type and B dominates B 0 , Theorem 1.5— applied to B T and (B 0 )T —implies that the set of g-vectors of cluster variables in A• (B 0 ) is a subset of the set of g-vectors of cluster variables in A• (B). Thus there is a natural inclusion of the sets of cluster variables, sending each cluster variable for B 0 to the cluster variable for B having the same g-vector. We will see below, with additional details, that this inclusion extends to an injective ring homomorphism. The principal-coefficients cluster algebra for an n×n exchange matrix is a subring of the ring of Laurent polynomials in indeterminates x1 , . . . , xn , with coefficients (ordinary) integer polynomials in indeterminates y1 , . . . , yn . The g-vector is a Zn grading of the cluster algebra determined by setting the g-vector of xk equal to the standard unit basis vector ek and the g-vector of yk equal to the negative of the k th column of the exchange matrix. A cluster monomial is a monomial in the cluster variables in some cluster. For each integer vector λ ∈ Zn (still restricting our attention to finite type), there exists a unique cluster monomial whose g-vector is λ. Writing A• (B 0 ) in terms of primed indeterminates x0i and yj0 , we observe that a ring homomorphism from A• (B 0 ) to A• (B) is entirely determined by its values on x01 , . . . , x0n and y10 , . . . , yn0 . Define zk , for each k, to be the cluster monomial whose th g-vector of B minus the k th column of B 0 . We define a set map  0is the k0 column 0 0 νz on x1 , . . . , xn , y1 , . . . , yn by (1.1) νz (x0k ) = xk νz (yk0 ) = yk zk for all k = 1, . . . , n, and extend νz to a ring homomorphism from A• (B 0 ) to the ring of Laurent polynomials in x1 , . . . , xn , with coefficients integer polynomials in y1 , . . . , yn . The homomorphism νz has a nice reformulation in terms of F polynomials. Define ŷi to be yi xb11i · · · xbnni . As a consequence of [12, Corollary 6.3], every cluster monomial z is xg11 · · · xgnn times a polynomial Fz in the ŷi , where (g1 , . . . , gn ) is the g-vector of z. The map νz sends each ŷk0 to ŷk · Fzk . A priori, it is not clear that the image of A• (B 0 ) under νz is even contained in A• (B), but surprising things often happen: Expected Theorem 1.17. Suppose B is an acyclic exchange matrix of finite type, and suppose B 0 is another exchange matrix dominated by B. Then the map νz is an injective, g-vector-preserving ring homomorphism from A• (B 0 ) to A• (B) and sends each cluster variable in A• (B 0 ) to a cluster variable in A• (B). We have verified Theorem 1.17 computationally for n × n exchange matrices with n ≤ 8, thus in particular handling the exceptional types. (See Section 5.3.) Furthermore, Theorem 1.20, below, in particular verifies the theorem when B is acyclic of type A or D. We call this an “expected theorem” because work in progress by Viel [31], already mentioned, is expected to complete the proof of the theorem by handling types B and C. DOMINANCE PHENOMENA 9 To generalize Theorem 1.17 beyond finite type, we need a generalization of the set of cluster monomials, and a natural choice is the theta functions arising from scattering diagrams [14]. Due to differences in conventions already mentioned (and discussed in [26, Remark 4.2]), there is a global transpose, so we work with the scattering diagram ScatT (B) = Scat(B T ). (The reason we write ScatT (B) is psychological: We think of B as our primary object, so we want to think of the global transpose not as a modification of B, but rather as a choice of conventions for creating a scattering diagram.) For each integer vector, there is a Laurent polynomial called a theta function obtained as a sum of monomials arising from broken lines in ScatT (B). We write zi for the theta function of the vector obtained as the k th column of B minus the k th column of B 0 and define νz as in (1.1). The cluster variables are naturally generalized by what we call ray theta functions. There is one ray theta function for each rational ray of the scattering fan ScatFanT (B) (the fan defined by ScatT (B)): It is the theta function of the shortest integer vector in the ray. Phenomenon IV (Injective homomorphisms). In many cases, when B dominates B 0 , the map νz is an injective, g-vector-preserving ring homomorphism from A• (B 0 ). In a smaller set of cases, νz sends each ray theta function for B 0 to a ray theta function for B. In some infinite types, some of the theta functions are not in the cluster algebra. The relationship between theta functions and the cluster algebra is complicated and is discussed at length in [14]. Because of this issue, we have been deliberately vague about the target of the ring homomorphism in Phenomenon IV. We now discuss additional instances of Phenomenon IV, beginning with the 2×2 case. Theorem 1.18. If B and B 0 are 2 × 2 exchange matrices such that B dominates B 0 , with B of finite or affine type, then both parts of Phenomenon IV occur. Theorem 1.19. If B and B 0 are 2 × 2 exchange matrices such that B dominates B 0 , with B 0 of finite type, then νz is an injective, g-vector-preserving ring homomorphism from A• (B 0 ) to A• (B). Assuming the Discreteness   Hypothesis,  νz sends 0 b and B 0 = 0 d with ray theta functions to ray theta functions unless B = c 0 a 0  cd = −3 and 1 6∈ |a|, |b| . Theorem 1.19 explains why Phenomenon IV references a “smaller set of cases.” Further understanding of Phenomenon IV in the 2 × 2 case is limited, for now, by a lack of detailed constructions of scattering diagrams in the “wild” 2 × 2 cases. We next describe an example of Phenomenon IV from the surfaces model. Theorem 1.20. Suppose (S, M) is a once-punctured or unpunctured disk and suppose T is a triangulation of (S, M) such that B(T ) is acyclic. If B(T ) dominates B 0 , then both parts of Phenomenon IV occur. Theorem 1.20 is precisely the cases of Theorem 1.17 where the Cartan matrix associated to B is of type A or type D. (For the Cartan matrix associated to B, see Section 3.4.) In Section 5.4, we prove Theorem 1.20 by observing that every matrix B 0 dominated by B(T ) can be obtained by a resection and then establishing Phenomenon IV for each relevant resection. Although in proving some of the cases, we generalize slightly to handle some of the non-acyclic cases, our treatment of the 10 NATHAN READING surfaces case suggests that acyclicity, at least in some local sense near the changed entries in the exchange matrix, is essential to Phenomenon IV. The arguments given in this paper can be extended to deal with most resections of twice-punctured disks (surfaces of affine type D), providing a step towards a version of Theorem 1.17 where B is of affine type. The cost is an increase in complexity, and we have not pursued the extension here. At the end of Section 5.4, we discuss the difficulties involved in generalizing Theorem 1.20 by allowing more general (S, M). 1.5. Related work. Several other lines of research touch tangentially on the ideas presented in this paper. The notion of dominance appears in [15, 16] as an integral part of the definition of seed homomorphisms. In [15, 16], the condition is broader in the sense that it allows passing to submatrices and/or making a global sign change. Also, the condition in [15, 16] is applied to extended exchange matrices, not just to exchange matrices. Some surface-cutting constructions similar to resection have appeared in [1, 16], but resection is different because it aims to do something different, namely to construct dominance relations without restricting to submatrices (and without “freezing” variables). All of the references cited just above are concerned primarily with notions of morphisms of cluster algebras, but, to the author’s knowledge, the homomorphisms proposed here do not fit into any category of cluster algebras already proposed. One category, proposed in [1], features rooted cluster morphisms. These require that initial cluster variables (including “frozen” variables) map to cluster variables or to integers and disallow exchangeable (“non-frozen”) initial cluster variables mapping to frozen variables. In our examples, the map restricts to a bijection on (non-frozen) initial cluster variables, but the frozen variables (which we call coefficients and denote by yi ) may map to other elements of the cluster algebra. In [15], it is shown that each rooted cluster morphism defines a seed homomorphism, or, loosely speaking, a dominance relationship in the broader sense allowing submatrices and/or a sign change as above. Other notions of morphisms between cluster algebras include the coefficient specializations of [12, Section 12] and [22, Section 3] and the quasi-homomorphisms of [13], both of which fix the exchange matrix B. Plan for the rest of the paper. The rest of this paper is devoted to proving the theorems stated above that provide examples of dominance phenomena. We begin by defining mutation-linear algebra and proving or quoting key results in Sections 2.1–2.2. We then consider the phenomena in the order in which they were introduced above, except that we consider Phenomena I and II together. 2. Mutation-linear algebra In this section, we quote definitions and results surrounding mutation-linear algebra, prove some new preliminary results, and discuss some examples of mutationlinear maps. 2.1. Partial linear structures. To put the notion of mutation-linear algebra into context, we briefly explore a more general notion that we will call a “partial linear structure.” We do not propose at this time a systematic study of partial linear structures, primarily because we are aware of only one interesting class of examples (the mutation-linear structures). However, it is hoped that the idea behind mutation-linear structures will be clarified by this brief foray into greater generality. DOMINANCE PHENOMENA 11 To make the notion of partial linear structures completely clear, we first formulate the usual notion of linear algebra entirely in terms of linear relations, saying exactly nothing surprising in the process. Let M be a module over a ring R (with P R having a multiplicative identity 1). Consider formal expressions of the form i∈S ci xi , where S is some finite indexing set, each ciPis an element of R, and each xi is an element of M . As usual, the expression i∈S ci xi is a linear relation on M if it evaluates (using the action of R on M and the addition operation in M ) to the zero element of M . Say a linear relation is trivial if it is empty or can be reduced to the empty relation by repeated applications of combining like terms (i.e. replacing ax + bx by cx where a + b = c) and deleting terms of the form 0x. Since addition in M is commutative, we consider linear relations up to commutativity, so that for example c1 x1 + c2 x2 and c2 x2 + c1 x1 are considered to be the same linear relation. Let Rel(M, R) be the set of linear relations on M with coefficients in R. A set A ⊆ M is independent (in the usual linear-algebraic sense) if every linear relation among elements of A is trivial. A set A ⊆ M spans M (in the usual linearalgebraic sense) if, for every element x ∈ M , there exists a linear relation writing x as a linear combination of elements of A. Now suppose M is a module over R and M 0 is a module over R0 . A map λ : M → M 0 induces a map on formal sums. Specifically, reusing the name λ for P P the induced map, λ i∈S ci xi is defined to be x∈S rx λ(x). A map is linear, in the usual sense, if and only if every linear relation in λ(Rel(M, R)) ⊆ Rel(M 0 , R0 ). In a partial linear structure, we modify all linear-algebraic constructions by ignoring all linear relations not in some fixed subset of Rel(M, R). Definition 2.1 (Partial linear structure). A partial linear structure is a triple (M, R, V) such that M be a module over a ring R (with identity) and V is a subset of Rel(M, R), satisfying the following conditions: (i) (Empty relation.) The set V contains the empty relation (the relation with empty indexing set). P (ii) (Irrelevance P of zeros.) For any x ∈ M , a relation 0x + i∈S ci xi is in V if and only if i∈S ci xi is in V, P (iii) (Combining like terms.) P If c = a + b in R, then ax + bx + i∈S ci xi is in V if and only if cx + i∈S P ci xi is in V. P (iv) (Scaling.) If c ∈ R and if i∈S ci xi is in V, then i∈S di xi is in V, where each di is cci . P P (v) (Formal addition.) If i∈S ci xi and j∈T dj yj are in V, then the formal P P sum i∈S ci xi + j∈T dj yj is in V. Example 2.2 (The trivial partial linear structure). Taking V0 to be the set of all trivial relations, we obtain a partial linear structure on (M, R, V0 ). In light of conditions (i) and (ii) of Definition 2.1, every partial linear structure (M, R, V) has V0 ⊆ V. The definition of a partial linear structure easily implies the following properties. Proposition 2.3. Suppose (M, R, V) is a partial linear structure. (vi) (Tautology.) V contains all relations for x ∈ M . P of the form 1x + (−1)x P (vii) (Formal substitution.) If a + i∈S (−ci )xi and da + j∈T dj yj are in P P V, then i∈S (dci )xi + j∈T dj yj is in V. 12 NATHAN READING We will call the relations in V the valid linear relations. Given a partial linear structure, new versions of the usual linear-algebraic notions can be obtained by substituting the set V for the set Rel(M, R) in the definitions. For example, we define a set A ⊆ M to be independent in (M, R, V) if every valid linear relation among elements of A is trivial. A set A ⊆ M spans P (M, R, V) if, for every element x ∈ M , there exists a valid linear relation x + i∈S (−ci )xi with {xi : i ∈ S} ⊆ A. A basis for (M, R, V) is an independent spanning set. Theorem 2.4. If (M, R, V) is a partial linear structure and R is a field, then (M, R, V) admits a basis. The proof of Theorem 2.4 follows the usual non-constructive proof of the existence of a basis for an arbitrary vector space, using Zorn’s lemma. This proof, for the special case of mutation-linear algebra defined in Section 2.2, is given in [22, Proposition 4.6]. We omit it here. Given partial linear structures (M, R, V) and (M 0 , R0 , V 0 ), a map λ : M → M 0 is linear , with respect to the partial linear structures, if the induced map on linear relations restricts to a map from V to V 0 . That is, the map is linear with respect to the partial linear structures if every valid linear relation is mapped to a valid linear relation. 2.2. Mutation-linear structures. We now review the definition of a partial linear structure that we will call a mutation-linear structure. An exchange matrix is an n × n skew-symmetrizable integer matrix B = [bij ]. Skew-symmetrizability means that there exist positive integers d1 , . . . , dn with di bij = −dj bji for all i, j ∈ [n] and implies in particular that bij and bji are either both zero or have strictly opposite signs. The fundamental notion in mutation-linear algebra is the notion of a B-coherent linear relation: a linear relation preserved by a family of piecewise-linear maps, called mutation maps, depending on B. To define B-coherent linear relations, we first need to define matrix mutation and the mutation maps. Let R be Z or any subfield of R. In particular, if R is a field, then it contains Q as a subfield. We call R the underlying ring . Let B̃ be an (n + 1) × n matrix whose top n rows agree with B and whose (n + 1)st row is a vector in Rn . For each k = 1, . . . , n, the mutation of B̃ at index k is B̃ 0 = µk (B̃) with entries given by  −bij if i = k or j = k; (2.1) b0ij = bij + sgn(bkj ) [bik bkj ]+ otherwise. Here [x]+ means max(0, x). The operation µk is an involution (i.e. µk (µk (B̃)) = B̃). We also use the symbol µk to denote the map B 7→ µk (B) given by the same formula. Given a sequence k = kq , . . . , k1 , the notation µk means µkq ◦ µkq−1 ◦ · · · ◦ µk1 . Note that the sequence k is indexed with the first entry in the sequence on the right. An  exchange matrix B is called mutation finite if the set µk (B) , where k ranges over all sequences of indices of B, is finite. Given B and a sequence k of indices, the mutation map ηkB : Rn → Rn is defined as follows: Given a ∈ Rn , let B̃ be the (n + 1) × n matrix with B in the top n rows and a in the bottom row. Then ηkB (a) is defined to be the bottom row of µk (B̃). This is a piecewise linear homeomorphism from Rn to itself, with inverse B ηk1q+1 ,...,kq . To see why, it may be helpful to consider the case where k consists of a single index k. In this case, if a = (a1 , . . . , an ), then ηkB (a) = (a01 , . . . , a0n ) with DOMINANCE PHENOMENA 13 each a0j given by (2.2)  −ak    aj + ak bkj 0 aj =  aj − ak bkj   aj if j = k; if j 6= k, ak ≥ 0 and bkj ≥ 0; if j 6= k, ak ≤ 0 and bkj ≤ 0; otherwise. When k is the empty sequence, ηkB is the identity map. For a nonempty sequence k = kq , kq−1 , . . . , k1 , if we define B1 = B and Bi+1 = µki (Bi ) for i = 1, . . . , q, then (2.3) B B q−1 ηkB = ηkBq ,kq−1 ...,k1 = ηkqq ◦ ηkq−1 ◦ · · · ◦ ηkB11 Since the entries of B are integers, the mutation maps on Rn restrict to a map Rn → Rn for any underlying ring R. These maps also commute with scaling by a nonnegative element c ∈ R, and have the property that (2.4) ηkB (a) = −ηk−B (−a). Definition 2.5 (B-coherent linear relation). Given a finite set S, vectors (vi : i ∈ P S) in Rn , and elements (ci : i ∈ S) of R, the formal expression i∈S ci vi is a B-coherent linear relation with coefficients in R if the equalities X (2.5) ci ηkB (vi ) = 0, and i∈S (2.6) X ci min(ηkB (vi ), 0) = 0 i∈S hold for every sequence k = kq , . . . , k1 of indices. Here min takes the minimum in each component separately. Since, in particular, (2.5) holds when k is the empty sequence, a B-coherent linear relationPis in particular a linear relation in the usual sense. Recall that a linear relation i∈S ci vi is trivial if it is empty or can be reduced to the empty relation by repeated applications of combining like terms (i.e. replacing av + bv by cv where a + b = c) and deleting terms of the form 0v. Example 2.6. It  to produce linear relations that are not B-coherent. For  is easy example, if B = −10 10 , the relation 1 · [ 1 0 ] + 1 · [ −1 0 ] is not B-coherent because η1B ([ 1 0 ]) = [ −1 1 ] but η1B ([ −1 0 ]) = [ 1 0 ]. On the other hand, one can check that the relation 1 · [ 1 0 ] + 1 · [ 0 1 ] + (−1) · [ 1 1 ] is B-coherent. Definition 2.7 (Mutation-linear structure). Given any underlying ring R and any exchange matrix B, the mutation-linear structure is the partial linear structure RB = (Rn , R, V) where V is the set of B-coherent relations over R. It is easy to check that RB is indeed a partial linear structure in the sense of Definition 2.1. Since RB is a partial linear structure, we can do “mutation-linear algebra” in R . That is, we can study the notions of basis and of linear maps with respect to the partial linear structure (Rn , R, V). A set A ⊆ RB is independent if every B-coherent linear map among elements of A is trivial. The set A isPan spanning if for every v ∈ Rn , there exists a B-coherent linear relation v − i∈S ci vi with {vi : i ∈ S} ⊆ A. A basis for RB is a set that is both independent and spanning. Theorem 2.4 implies that, when R is a field, a basis exists for RB . See [22, Proposition 4.6] for a proof in the mutation-linear case. We have no proof that a basis exists for ZB for every exchange matrix B, but also we have no example of an exchange matrix B such that ZB has no basis. B 14 NATHAN READING Remark 2.8. A basis for RB is called an “R-basis for B” in [22], and similarly,“Rindependent sets for B” and “R-spanning sets for B,” because the notion of a mutation-linear structure RB was not explicitly named in [22]. 2.3. Mutation-linear maps. Suppose B and B 0 are exchange matrices and con0 sider maps from RB to RB . These are simply maps from Rn to Rm if B is n × n 0 and B 0 is m × m, but we retain the notation RB and RB because the mutation0 B linear structure is essential in what follows.PA map λ : RP → RB induces a map on formal linear combinations that sends i∈S ci vi to i∈S ci λ(vi ). The map 0 λ : RB → RB is mutation-linear if every B-coherent linear relation with coefficients in R is sent by λ to a B 0 -coherent linear relation with coefficients in R. We first notice that mutation-linear maps commute with nonnegative scaling. 0 Proposition 2.9. Suppose λ : RB → RB is a mutation-linear map. If v ∈ RB and w = cv for c a nonnegative element of R, then λ(w) = cλ(v). Proof. Since each mutation map ηkB commutes with nonnegative scaling, the rela0 tion cv+(−1)w is B-coherent. Thus if λ : RB → RB is a mutation-linear map, then 0 cλ(v) + (−1)λ(w) is a B -coherent linear relation. In particular cλ(v) + (−1)λ(w) is a linear relation, so λ(w) = cλ(v).  We omit the proof of the following easy proposition. Proposition 2.10. Suppose R0 and R1 are underlying rings with R0 ⊆ R1 . If 0 0 λ : R1B → R1B is mutation-linear and restricts to a map from R0B to R0B , then the restriction is mutation-linear. 0 On the other hand, given a mutation-linear map λ : ZB → ZB , one can extend 0 λ to a map λ : QB → QB by clearing denominators in the usual way, and λ can be shown to be mutation-linear as well. In other cases, it is not clear whether mutation-linear maps can be extended to larger underlying rings. Definition 2.11 (The mutation fan FB ). To define the mutation fan, we first define an equivalence relation ≡B on RB by setting a1 ≡B a2 if and only if sgn(ηkB (a1 )) = sgn(ηkB (a2 )) for every sequence k of indices. Here sgn(a) denotes the vector of signs (−1, 0, or 1) of the entries of a. The equivalence classes of ≡B are called B-classes. The closures of B-classes are called B-cones. These are closed convex cones [22, Proposition 5.4], meaning that they are closed under nonnegative scaling and addition. The set of B-cones and their faces constitute a complete fan [22, Theorem 5.13] called the mutation fan. (In this context, a fan is a collection of convex cones, closed under taking faces, such that the intersection of any two cones is a face of each. Although we have no examples of B-cones that are not polyhedral cones, the possibility has not been ruled out. Thus to refer to a “face” in this definition, we must use the general definition of a face of a convex body, rather than the more special notion of a face of a polyhedron.) We point out a helpful fact, which is [22, Proposition 7.1]. Proposition 2.12. F−B = −FB . The following is part of [22, Proposition 5.5]. Proposition 2.13. For any sequence k of indices, a set C is a B-cone if and only if ηkB (C) is a µk (B)-cone. DOMINANCE PHENOMENA 15 P The formal expression i∈S ci vi is a B-local linear relation if it is a linear relation in the usual sense and if {vi : i ∈ S} is contained in some B-cone. The following is [22, Proposition 5.9]. Proposition 2.14. Every B-local linear relation is a B-coherent linear relation. A collection of vectors in Rn is sign-coherent if for any k ∈ [n], the k th coordinates of the vectors in the collection are either all nonnegative or all nonpositive. The following is [22, Proposition 5.30]. Proposition 2.15. A set C ⊆ Rn is contained in some B-cone if and only if the set ηkB (C) is sign-coherent for every sequence k of indices in [n]. 0 Proposition 2.16. If λ : RB → RB is a mutation-linear map and C is any cone of FB , then the restriction of λ to C ∩ RB is a linear map (in the usual sense) into some cone of FB 0 . Proof. We argue the case where C is a B-cone. Since every cone of FB is a face of some B-cone, the result then follows for arbitrary cones of FB . Let v and w be elements of C ∩RB and let x = v+w. Then x is in C ∩RB as well. The linear relation v + w + (−1)x is a B-local linear relation, and thus B-coherent by Proposition 2.14. Mutation-linearity of λ implies that λ(v) + λ(w) + (−1)λ(x) is a B 0 -coherent linear relation, and thus in particular a linear relation in the usual sense. Thus λ(x) = λ(v) + λ(w), and this, together with Proposition 2.9, shows that the restriction of λ to C ∩ RB is linear. Suppose for the sake of contradiction that λ does not map C ∩ RB into a B 0 cone. Then in particular, there are two elements v and w of C ∩ RB such that λ(v) and λ(w) are not contained in a common B-cone. Proposition 2.15 says that 0 0 there exists a sequence k of indices and an index j such that ηkB λ(v) and ηkB λ(w) 0 strictly disagree in the sign of their j th entry. Without loss of generality, ηkB λ(v) 0 has strictly positive j th entry and ηkB λ(w) has strictly negative j th entry. Again, 0 0 let x = v + w. Write hv , hw and hx for the j th entries of ηkB λ(v), ηkB λ(w), and 0 ηkB λ(x), respectively. As above, λ(v) + λ(w) + (−1)λ(x) is a B 0 -coherent linear 0 relation. Thus if the j th entry of ηkB λ(x) is nonnegative, then (2.5) and (2.6) imply that hv + hw − hx = 0 and hv − hx = 0, so that hw = 0. Similarly, if the j th 0 entry of ηkB λ(x) is negative, then (2.6) implies that hv = 0. In either case, we have reached a contradiction, because hv and hw strictly disagree in sign.  We wish to prove the converse of Proposition 2.16, but we will need an additional hypothesis. (The additional hypothesis are necessary for our proof, but we know of no counterexample to the converse without the additional hypothesis.) Definition 2.17. A cone basis for RB is a independent set U in RB such that, for every B-cone C, the R-linear span of U ∩ C contains Rn ∩ C. By [22, Proposition 6.4], a cone basis for RB also spans RB , and thus is a basis for RB . In fact, a stronger property than spanning is easily proved. Proposition 2.18. Suppose a cone basisP U exists for RB . Then for every a ∈ RB , there exists a B-local linear relation a + i∈S (−ci )vi such that each vi is in U . Proof. If C is any B-cone containing a, then the P definition of a cone basis says that a is a linear combination (in the usual sense) i∈S ci vi of elements of U ∩ C.  16 NATHAN READING Remark 2.19. We have no proof that every mutation-linear structure RB has a cone basis. (Indeed, as mentioned, we have no proof that ZB admits even a basis for every B.) However, we know of no example of a mutation-linear structure RB that does not admit a cone basis. A stronger requirement on a basis is that it be positive. An basis U for RB n is P positive if, for every a ∈ R , there exists a B-coherent linear relation a + i∈S (−ci )bi such that all of the ci are nonnegative elements of R. Our main use for positive bases in this paper is the following fact, which allows us to quote results from other papers where positive bases are constructed. This fact is a part of [22, Proposition 6.7]. Proposition 2.20. If U is a positive basis for RB , then it is a cone basis for RB . Remark 2.21. In contrast to the situation for cone bases (as explained in Remark 2.19), there are known examples of mutation-linear structures that do not admit positive bases. For example, [22, Proposition 9.9] says that when B = 0b a0 with ab < 4, no positive basis exists for ZB or for QB . We know of no example where RB does not admit a positive basis. The existence of a cone basis for RB greatly simplifies mutation-linear algebra. Proposition 2.22. If a cone basis exists for RB , then a linear relation is Bcoherent if and only if it can be reduced to a (possibly empty) formal sum of B-local relations by a sequence of changes, each of which adds a term with coefficient zero or un-combines like terms. Proof. Proposition 2.14 implies that any formal sum of B-local relations is Bcoherent. Deleting a term with coefficient zero or combining like terms preserves B-coherence, so any linear relation that can be reduced to a formal sum of B-local relations by adding termsP and/or un-combining like terms is B-coherent. On the other hand, let i∈S ci vP i be a B-coherent linear relation. For each i ∈ S, write a B-local linear relation vi + j∈Ti (−dij )wij such that each wij is an element of the cone basis. (We P can doPthis by Proposition 2.18.) By Proposition 2.3(vii) (applied several times), i∈S j∈Ti (ci dij )wj is a B-coherent linear relation. Since the wj are a basis, this relation is trivial, meaning that either it is empty or it can be reduced to the empty relation by repeated applications of combining like terms (replacing bx by cx where a + b = c) and deleting terms of the form 0x. Pax +P Therefore also i∈S j∈Ti (−ci dij )wj is trivial. P Now, starting again withPthe B-coherent linear combination i∈S ci vi , we can P add in the trivial relation i∈S j∈Ti (−ci dij )wj , by a sequence of additions of terms withPcoefficient zero of like terms. The result can be  P and/or un-combinings rewritten c v + (−c d )w using commutativity. (Recall that we i i i ij j i∈S j∈Ti P consider linear relations up to commutativity.) Each ci vi + j∈Ti (−ci dij )wj is P B-local because vi + j∈Ti (−dij )wij is.  We now prove a converse to Proposition 2.16, when a cone basis exists for RB . 0 Proposition 2.23. Suppose RB admits a cone basis. Then λ : RB → RB is mutation-linear if and only if, for any B-cone C, the restriction of λ to C ∩ RB is a linear map (in the usual sense) into some B 0 -cone. DOMINANCE PHENOMENA 17 Proof. One direction is Proposition 2.16. For the other direction, consider a BP P coherent linear relation R = i∈S ci vi and let λ(R) be the formal sum i∈S ci λ(vi ). Proposition 2.22 says that R can be reduced to a formal sum of B-local relations by adding terms with coefficient zero or un-combining like terms. If the corresponding changes are made to λ(R), by the hypothesis on λ, the result is a formal sum of B 0 -local linear relations, which is therefore B 0 -coherent by Proposition 2.14. Then λ(R) is B-coherent because undoing the changes (i.e. recombining like terms and deleting zero terms) preserves B-coherence.  2.4. Mutation-linear algebra and universal coefficients. It was mentioned in Section 1 that by [22, Theorem 4.4], finding a basis for RB is the same thing as finding a cluster algebra for B with universal (geometric) coefficients. We now elaborate on that remark, and describe how mutation-linear maps figure into the picture. For more details and background, see [22, Sections 2–4]. For a fixed exchange matrix B, we consider a category GeomR (B) whose objects are cluster algebras of geometric type over R in the sense of [22, Section 2], all having the same initial exchange matrix B and all having the same initial cluster. This notion of geometric type is broader than the usual notion in [12], because it allows extended exchange matrices to have infinitely many coefficient rows, and these coefficient rows are allowed to have entries in R, which may be larger than Z. The arrows in GeomR (B) are coefficient specializations, similar to those defined in [12, Section 12], but with an extra topological requirement necessitated by the possible infinity of rows. (When there are countably many rows, the requirement is that the coefficient specialization be continuous in a “formal power series” topology on the coefficient semifield.) Let B be an exchange matrix and let I and K be arbitrary index sets. Let B̃1 be an extended exchange matrix with exchange matrix B and coefficient rows (ai : i ∈ I). Let B̃2 be an extended exchange matrix with exchange matrix B and coefficient rows (bk : k ∈ K). A coefficient specialization between cluster algebras P A(B̃1 ) and A(B̃2 ) amounts to a collection of B-coherent linear relations bk + i∈Sk (−cik )ai , one for each k ∈ K, expressing the coefficient row bk of B̃ 0 as a B-coherent linear combination of coefficient rows of B̃. 0 Given a mutation-linear map λ : RB → RB , we define a functor from GeomR (B) 0 to GeomR (B ) by sending each extended exchange matrix B̃ with coefficient rows (ai : i ∈ I) to the exchange matrix B̃ 0 with coefficient rows (λ(ai ) : i ∈ I). A coefficient specialization between A(B̃1 ) and A(B̃2 ) defined by B-coherent linear P relations bk + i∈Sk (−cik )ai is sent to the coefficient specialization A(B̃10 ) and P A(B̃20 ) defined by B 0 -coherent linear relations λ(bk ) + i∈Sk (−cik )λ(ai ). The latter linear relations are B 0 -coherent because λ is mutation-linear. Indeed, given an 0 arbitrary map λ : RB → RB , the construction described here yields a functor if and only if λ is mutation-linear. 2.5. Examples of mutation-linear maps. Phenomenon I considers bijective mutation-linear maps that are not isomorphisms, because their inverses are not mutation-linear (unless B 0 = B). By way of contrast, here we mention some examples of mutation-linear maps, specifically mutation-linear isomorphisms, surjections and injections. We omit the straightforward proofs. In Section 3, we return to the topic of dominance and discuss Phenomenon I. 18 NATHAN READING 2.5.1. Mutation-linear isomorphisms. The easiest examples of mutation-linear isomorphisms are the mutation maps themselves. Proposition 2.24. For any exchange matrix B, any underlying ring R and any sequence k of indices, the mutation map ηkB : RB → Rµk (B) is a mutation-linear isomorphism. We mention a few more mutation-linear isomorphisms. Proposition 2.25. For any exchange matrix B and any underlying ring R, the antipodal map x 7→ −x is a mutation-linear isomorphism from RB to R−B . P Proof. Given a B-coherent linear map i∈S ci vi , we show that the expression P c (−v ) is (−B)-coherent. Given any sequence i i Pi∈S P k of integers, the expression c η (−v ) can be rewritten, using (2.4), as i k i i∈S i∈S (−ci )ηk (vi ), which is 0 by P the B-coherence of i∈S ci vi . Thus the antipodal map is mutation-linear from RB to R−B . It is self-inverse, and it is also mutation-linear from from R−B to RB .  Given a permutation π of the indexing set of B = [bij ], let π(B) be [bπ(i)π(j) ] and let π also denotes the linear map sending eπ(i) to ei , where ei is the vector with 1 in the ith entry and zeros in the other positions. Proposition 2.26. For any exchange matrix B and any underlying ring R, the linear reindexing map π : RB → Rπ(B) is a mutation-linear isomorphism. Let B = [bij ] and B 0 = [b0ij ] be exchange matrices. Then B 0 is a rescaling of B if there exists a diagonal matrix Σ = diag(σ1 , . . . , σn ) with positive entries such σ that B 0 = Σ−1 BΣ. Equivalently, b0ij = σji bij for all i and j. As verified in [22, Proposition 7.6], the matrix Σ can always be taken to have integer entries. An exchange matrix may or may not admit any nontrivial rescaling. More information on rescaling exchange matrices is found in [22, Section 7]. In particular, [22, Proposition 7.8.4] immediately implies the following fact. Proposition 2.27. Let B be an exchange matrix and take the underlying ring R to be a field. Suppose B 0 is a rescaling of B. Write B 0 = Σ−1 BΣ for Σ having entries 0 in R. Then the map v 7→ (vΣ) : RB → RB is a mutation-linear isomorphism. 0 When R = Z, the map v 7→ (vΣ) is still mutation-linear from ZB to ZB . It is injective, but need not be surjective. 2.5.2. Surjective mutation-linear maps. For any subset I of the indices of B, define BI to be the matrix obtained from B by deleting row j and column j for all j 6∈ I. Retain the original indexing of BI , so that it is indexed by I. Let ProjI : RB → RBI be the usual projection (ignoring the j th coordinate for each j 6∈ I). Proposition 2.28. Let B be an exchange matrix. For any subset I of the indices of B, the map ProjI : RB → RBI is mutation-linear. We also mention an ad hoc construction of a surjective mutation-linear map by “wrapping”   matrix   a “larger” mutation fan around a “smaller” one. The exchange B = −30 10 has eight maximal cones, while the mutation fan for B 0 = 00 00 has four. (See Figure 1.) We will see later, combining Theorem 3.5 and Proposition 2.20, that RB admits a cone basis. One can construct a piecewise-linear branched double-cover 0 λ : RB → RB that maps each cone of FB linearly to a cone of FB 0 and thus is mutation-linear by Proposition 2.23. This construction admits many variations. DOMINANCE PHENOMENA 19 2.5.3. Injective mutation-linear maps. As mentioned above, nontrivial rescalings, when R = Z are mutation-linear and injective but possibly not bijective. Much 0 stranger injective mutation-linear maps can be found from RB to RB when B is n × n and B 0 is m × m with n < m. In this case, by Proposition 2.23, one might, for example, map FB injectively into the n-skeleton of FB 0 , sending each cone linearly to a cone. A less strange special case is when B = (B 0 )I in the sense of Section 2.5.2, taking the natural injection into the subspace of RB indexed by I. Results of [22, Section 8] can be used to show that such an injection is mutation-linear, at least in the case where (B 0 )I is of finite type. 3. Mutation-linearity and refinement of the mutation fan In this section, we discuss Phenomena I and II. We begin by observing (in Proposition 3.1), the close connection between Phenomenon I and Phenomenon II. Similarly, in Proposition 3.3, we connect rational versions of the two phenomena. In the rest of this section, we gather examples of the two phenomena. 3.1. Connecting Phenomenon I and Phenomenon II. A fan F refines a fan F 0 with ∪F 0 = ∪F if every cone of F contained in a cone of F 0 , or equivalently if every cone of F 0 is a union of cones of F. The following proposition is an immediate consequence of Propositions 2.16 and 2.23. 0 Proposition 3.1. If the identity map RB → RB is mutation-linear, then FB re0 fines FB 0 . If RB admits a cone basis, then the identity map RB → RB is mutationlinear if and only if FB refines FB 0 . In the case of resection of surfaces, where we deal with rational versions of the phenomena, we need a rational version of Proposition 3.1. In light of Proposition 2.10, for fixed B and B 0 , the statement that the identity map from RB to 0 RB is mutation-linear is equivalent to the statement that the identity map from 0 RB to RB is mutation-linear for any underlying ring R. Thus in particular the conclusion of Theorem 1.2 is formally weaker than Phenomenon I. We now develop a geometric formulation of the weaker assertion. Definition 3.2 (Rational part of a fan). Suppose F and F̃ are fans in Rn satisfying the following conditions: (i) Each cone in F̃ is the nonnegative R-linear span of finitely many rational vectors. (ii) Each cone in F̃ is contained in a cone of F. (iii) For each cone C of F, there is a unique largest cone (under containment) among cones of F̃ contained in C. This largest cone contains Qn ∩ C. Then F̃ is called the rational part of F. The rational part of a given F might not exist, but if it exists, it is unique. (See [22, Definition 6.9] for examples of non-existence and a proof of uniqueness, in the more general context where Q is replaced by any underlying ring R.) We have defined the rational part of a real fan to be another real fan. Instead, n one n might want to o consider the “rational part” of a fan as a fan in Q , by taking n : n n C ∩ Q C ∈ F̃ . We write F̃ ∩ Q for this fan in Q . 20 NATHAN READING Proposition 3.3. Suppose that B and B 0 are n × n exchange matrices, that F̃B is the rational part of FB , and that F̃B 0 is the rational part of FB 0 . If the identity 0 map QB → QB is mutation-linear, then F̃B ∩ Qn refines F̃B 0 ∩ Qn . Assuming also that QB admits a cone basis, the identity map is mutation-linear if and only if F̃B ∩ Qn refines F̃B 0 ∩ Qn . 0 Proof. Suppose the identity map QB → QB is mutation-linear. Suppose C̃ is any cone of F̃B . By definition of the rational part of a fan, there exists a cone C of FB containing C̃. Proposition 2.16 says that C ∩ Qn is contained in some cone C 0 of FB 0 . There exists a cone C̃ 0 of F̃B 0 contained in C 0 and containing C 0 ∩ QB . Thus C̃ ∩ QB ⊆ C̃ 0 . We have shown that F̃B ∩ Qn refines F̃B 0 ∩ Qn . Assuming now that QB admits a cone basis, suppose conversely that F̃B ∩Qn refines F̃B 0 ∩Qn . Given any B-cone C, there is a cone C̃ of F̃B with C ∩QB ⊆ C̃ ⊆ C. Since F̃B refines F̃B 0 as fans in QB , there exists a cone of F̃B 0 containing C̃. Since some cone C 0 of FB 0 contains C̃ 0 , we have satisfied the hypothesis of Proposition 2.23, for λ equal to the identity map, so the identity map is mutation-linear.  3.2. Erasing edges. We now prove Theorems 1.3 and 1.8. Both theorems suppose that the indices of B are written as a disjoint union I ∪ J and that B 0 = [b0ij ] is obtained by erasing all edges of B = [bij ] that connect indices in I to indices in J. That means that every b0ij = b0ji = 0 for all i ∈ I and j ∈ J and every other entry of B 0 agrees with the corresponding entry of B. The conclusion of Theorem 1.3 is that 0 the identity map RB → RB is mutation-linear. The conclusion of Theorem 1.8 is that FB refines FB 0 . We prove Theorem 1.3 directly. Theorem 1.8 then follows immediately by Proposition 3.1. The proof of Theorem 1.3 uses the following observation, which follows from (2.1) and (2.2) and a simple induction. We continue the notation of Section 2.5.2. Proposition 3.4. Suppose I is a subset of the indices of B and suppose k is a sequence of indices in I. Then (µk (B))I = µk (BI ). Furthermore, if v ∈ RB , then ηkBI (ProjI (v)) = ProjI (ηkB (v)). P Proof of Theorem 1.3. Suppose i∈S ci vi is a B-coherent linear P relation. Considering Proposition 3.4 for all sequences k of indices in I, we see that i∈S ci ProjI (vi ) P is a BI -coherent linear relation. Similarly, i∈S ci ProjJ (vi ) is BJ -coherent. P We will show that i∈S ci vi is a B 0 -coherent relation. Each vector x ∈ RB can be written x = ProjI (x) + ProjJ (x). Since b0ij = 0 whenever i ∈ I and 0 j ∈ J and because BI = BI0 and BJ = BJ0 , we see from (2.2) that ηkB (x) = 0 BI B ηk ProjI (x) + ProjJ (x) if k ∈ I and ηkB (x) = ProjI (x) + ηk J ProjJ (x) if k ∈ J. Let k be any sequence of indices of B, let kI be the subsequence of k consisting of indices k consisting of indices in J. J be the subsequence ofP P in I,B 0and let kP Then ci ηk (vi ) = ci ηkBII (ProjI (vi )) + i∈S ci ηkBJ (ProjJ (vi )). Since i∈S i∈S P P and i∈S ci ProjJ (vi ) is a BJ i∈S ci ProjI (vi ) is a BI -coherent linear relation P 0 coherent linear relation, we conclude that i∈S ci ηkB (vi ) = 0.  3.3. The 2 × 2 case. In this section, we prove Theorem 1.9, which characterizes when Phenomenon II occurs for 2 × 2 exchange matrices. As explained in the introduction, results of [22, Section 9] combined with Proposition 3.1 then immediately imply Theorem 1.4. DOMINANCE PHENOMENA B= 0 0 B= 00  01 −1 0  B=  01 −2 0 21  B=  01 −3 0  Figure 1. Mutation fans FB for 2 × 2 exchange matrices of finite type   Suppose B = 0b a0 is a 2 × 2 exchange matrix. We now briefly review the description of the mutation fan of B given in [22, Section 9]. Since FB is complete and contained in R2 , we can describe itcompletely   by listing its rays. If a = b = 0, then FB has rays ± 10 and ± 01 . Otherwise, there are two possible sign patterns for B, either a > 0 and b < 0 or a < 0 and b > 0. Figure 1 shows the mutation fans in the case of finite type (i.e. when ab > −4) with the additional condition that 0 ≤ a ≤ −b. pictures are  The   rays   occurring  in these  spanned by the vectors ± 10 , ± 01 , −11 , −12 , −23 , and −13 . The remaining finite type cases can be recovered via two observations: First, Proposition   2.12  says  that F−B is the antipodal opposite of FB , and second, passing from 0b a0 to a0 0b reflects the mutation fan through the line x1 = x2 . In the cases where ab ≤ −4, we define P0 = P1 = 0 and, for m ≥ 2 ( −abPm−1 − Pm−2 if m is even, or (3.1) Pm = Pm−1 − Pm−2 if m is odd. This is a recursion, given in [26, (6.1)], for the polynomials defined in [22, (9.5)]. The first several values of Pm (a, b) are shown here. m Pm 0 1 1 1 2 −ab − 1 3 −ab − 2 4 a2 b2 + 3ab + 1 5 a2 b2 + 4ab + 3 Define, for k ≥ 0, vectors h i  k (a,b)  sgn(a)P −aPk+1 (a,b) i vk = vk (a, b) = h −bPk (a,b)   sgn(b)P (a,b) k+1 v∞ = v∞ (a, b) = h if k is odd. i √ 2 sgn(a) −ab √ √ −a( −ab+ −ab−4) h i −bPk+1 (a,b)   sgn(b)P k (a,b) i wk = wk (a, b) = h sgn(a)Pk+1 (a,b)   −aP (a,b) k w∞ = w∞ (a, b) = if k is even, or h if k is even, or if k is odd. i √ √ −b( −ab+ −ab−4) √ . 2 sgn(b) −ab Still assuming ab ≤ −4, the rays of FB are spanned by the vectors: n    o ± 10 , ± 01 ∪ {vk : k = 0, . . .} ∪ {v∞ } ∪ {wk : k = 0, . . .} ∪ {w∞ } . It is apparent that a mutation fan FB in R2 refines a mutation fan FB 0 in R2 if and only if the set of rays of FB 0 is contained in the set of rays of FB . Since the 22 NATHAN READING Slopes B  0 1 −1 0  −1  0 1 −2 0  −1, − 12  0 2 −1 0  −2, −1  0 1 −3 0  −1, − 12 , − 23 , − 13  0 3 −1 0  −3, −2, − 32 , −1 Table 1. Relevant slopes of FB for B of finite type, a > 0 and b < 0 s0 , s1 , s2 , s3 . . . t0 , t 1 , t 2 , t 3 . . . s∞ = t∞  −1, − 43 , − 32 , − 58 , . . . − 41 , − 13 , − 38 , − 52 , . . . − 21  0 2 −2, − 23 , − 34 , − 54 , . . . − 21 , − 23 , − 34 , − 54 , . . . −1  0 4 −1, − 34 , − 23 , − 85 , . . . −4, −3, − 83 , − 25 , . . . −2 B    0 1 −4 0 −2 0 −1 0 Table 2. Some relevant slopes of FB for B of affine type, a > 0 and b < 0     rays spanned by ± 10 and ± 01 are in every mutation fan, we can ignore these  rays. If B 0 = 00 00 , then FB refines FB 0 for every B. Otherwise, it is apparent that FB cannot refine FB 0 unless B and B 0 have weakly the same  sign pattern. By Proposition 2.12, we need only consider the case where B = 0b a0 with a > 0 and b < 0 and B 0 has weakly the same sign pattern. Having made this reduction, we also may as well consider, rather than containment of rays, containment of the set of slopes of rays. We ignore slopes 0 and ∞ and call the slopes of the remaining rays the relevant slopes. In finite type, the relevant slopes are shown in Table 1. For ab ≤ −4 (and continuing with a > 0 and b < 0), the relevant slopes are  − aPk+1 (a,b) if k is even, or Pk (a,b) sk (a, b) = P (a,b)  k+1 if k is odd. bPk (a,b) √ √ a( −ab + −ab − 4) √ s∞ (a, b) = − 2 −ab   Pk (a,b) if k is even, or bPk+1 (a,b) tk (a, b) = aP (a,b) − k if k is odd. Pk+1 (a,b) √ 2 −ab √ t∞ (a, b) = √ . b( −ab + −ab − 4) All of these slopes are negative. The slopes sk increase and limit to s∞ . The slopes tk decrease and limit to t∞ , and we have s∞ ≤ t∞ . When ab = −4, t∞ and s∞ are equal and rational, but for ab < −4, t∞ and s∞ are distinct and irrational. Some DOMINANCE PHENOMENA 23 of these slopes are shown in Table 2 for the affine types. We now prove the characterization of Phenomenon II for 2×2 exchange matrices. Proof of Theorem 1.9. A brief inspection of Tables 1 and 2 is enough to verify Theorem 1.9(2) and the non-wild case of Theorem 1.9(1). We next check Theorem 1.9(3) and the case of Theorem 1.9(1) where exactly one of B and B 0 is wild. If B 0 is wild and B is not, then FB 0 has two irrational relevant slopes, while FB has only rational  relevant slopes, so FB does not refine   FB 0 . Now consider the case where B = 0b a0 is wild and B 0 is not. If B 0 = 00 00 , then FB refines FB 0 . Now consider the case where B 0 = −10 10 . If a > 2 and b < −2, then √ √ √ 2 −ab a( −ab + −ab − 4) √ √ < −1 < √ = t∞ , s∞ = − 2 −ab b( −ab + −ab − 4) and thus −1 is not a relevant slope of FB , so FB does not refine FB 0 . If a = 1, then s0 (a, b) = −a = −1, so FB refines FB 0 , and similarly if b = −1, then  0 1t0 (a, b) = 1 0 0 . Next, consider the case where B = −1, so F refines F = B B −2 0 . Since b   B 0 dominates −10 10 , by Theorem 1.9(2) we conclude that FB does not refine FB 0 except possibly when a = 1 or b = −1. But if a = 1, then b ≤ −5, so √ √ √ a( −ab + −ab − 4) 1 2 −ab √ √ s∞ = − <− < √ = t∞ , 2 2 −ab b( −ab + −ab − 4) and thus − 12 is not a relevant slope of FB , so FB does not refine FB 0 . If b = −1, then a ≥ 5, and we check similarly   that s∞ < −2 < t∞ , so that FB does not refine FB 0 . The case where B 0 = −10 20 is dealt with by the symmetric  0 2  Since  0 1  argument. 0 every other non-wild exchange matrix B dominates either −2 0 or −1 0 , we see by Theorem 1.9(2) that no non-wild B 0 can have FB refine FB 0 . We have verified Theorem 1.9(3).     Finally, suppose B = 0b a0 and B 0 = d0 0c are distinct wild 2 × 2 exchange matrices. Since FB and FB 0 each have exactly two irrational slopes, for a refinement relation to exist in either direction, in particular, these two slopes must be the same in FB as in FB 0 . The conditions that the two slopes are the same in each fan can be rewritten, after some manipulation, as p p abd − a2 b2 d2 + 4abd2 = bcd − b2 c2 d2 + 4b2 cd and p p abc + a2 b2 c2 + 4abc2 = acd + a2 c2 d2 + 4a2 cd We know that these slopes are irrational (in the case ab < −4), so the square roots in the above equations are all irrational. No two distinct irrational square roots of integers can differ by a rational number, so the square roots on each side must be equal, and thus the integers on each side must be equal. That is, abd = bcd and abc = acd, and therefore a = c and b = d. We have verified Theorem 1.9(4) and the final case of Theorem 1.9(1).  3.4. Acyclic finite type. An exchange matrix B is acyclic if, after an appropriate reindexing if necessary, it has the property that bij ≤ 0 whenever i > j. The exchange matrix B is of finite type if every cluster algebra with B as its initial exchange matrix has finitely many cluster variables. For details and a classification of cluster algebras of finite type, see [11]. We now prove Theorem 1.1, which establishes Phenomenon I when B is an acyclic exchange matrix of finite type. 24 NATHAN READING The following theorem is [22, Theorem 10.12]. The theorem as stated in [22] refers to the Standard Hypotheses, as do several other results quoted below. These are that B has the property that the nonnegative orthant O = (R≥0 )n is a B-cone, and that each member of a family of related matrices has the same property. It was conjectured [22, Conjecture 8.1] that every exchange matrix has this property. That conjecture was shown in [22, Proposition 8.9] to be equivalent to a standard conjecture often called “sign-coherence of c-vectors,” now a theorem of [14]. Thus we remove all references to the Standard Hypotheses when we quote from [22]. Theorem 3.5. If RB is a mutation-linear structure such that B is of finite type, then the g-vectors associated to B T constitute a positive basis for RB . It will not be necessary here to define the g-vectors. For our purposes, it is only important that a positive basis exists. The following corollary is immediate by Propositions 2.20 and 3.1. Corollary 3.6. If B is an exchange matrix of finite type and B 0 is an exchange 0 matrix of the same size, then the identity map RB → RB is mutation-linear if and only if FB refines FB 0 . The following theorem is established as part of the proof of [22, Theorem 10.12], but is unfortunately not stated separately as a numbered result in [22]. Theorem 3.7. If B is an exchange matrix of finite type, then FB consists of the g-vector cones associated to B T and their faces. We can again proceed without defining g-vectors, this time because we merely wish to concatenate Theorem 3.7 with a theorem below about Cambrian fans. Starting with an exchange matrix B, if one replaces each 0 on the diagonal with 2 and turns all positive off-diagonal entries negative, the resulting matrix Cart(B) is a symmetrizable (generalized) Cartan matrix . A Cartan matrix is of finite type if and only if it is positive definite. By results of [11], Cart(B) is of finite type if and only if B is acyclic and of finite type. The Coxeter element c associated to B and the Cambrian fan associated to (Cart(B), c) are defined in many places (e.g. in [27, Section 4], [28, Section 9], [29, Section 5]), but we leave them undefined here, because we only use them to continue concatenating results. It is enough to say that the Cambrian fan is a certain refinement of the Coxeter fan for Cart(B), the collection of hyperplanes normal to a root system for Cart(B). Specifically, we construct a root system Φ(Cart(B)) for Cart(B) in a real n-dimensional vector space V and identify V with Rn by identifying e1 , . . . , en with the simple co-roots. The following theorem was conjectured in [27, Section 10], where it was proved in a special case. It was proved in [32, Theorem 1.10] and later as [29, Corollary 5.16]. Theorem 3.8. Suppose B is an acyclic exchange matrix of finite type and let c be the associated Coxeter element. Then the Cambrian fan associated to (Cart(B), c) consists of the g-vector cones associated to B and their faces. Theorems 3.7 and 3.8 combine to say the following: Theorem 3.9. Suppose B is an acyclic exchange matrix of finite type and let c be the associated Coxeter element. Then FB T is the Cambrian fan associated to (Cart(B), c). DOMINANCE PHENOMENA 25 A Cartan matrix A = [aij ] dominates a Cartan matrix A0 = [a0ij ] if aij ≥ a0ij for all i and j. If is easy to see that if B dominates B 0 , then Cart(B) dominates Cart(B 0 ). The following is [25, Proposition 1.10]. Proposition 3.10. Suppose A and A0 are symmetrizable Cartan matrices such that A dominates A0 . If Φ(A) and Φ(A0 ) are both defined with respect to the same simple roots αi , then Φ(A) ⊇ Φ(A0 ) and Φ+ (A) ⊇ Φ+ (A0 ). The following theorem, which is [25, Theorem 1.11], relies of Proposition 3.10 as it applies to the dual root system Φ∨ (A) to Φ(A). Theorem 3.11. Suppose A and A0 are Cartan matrices such that A dominates A0 and suppose W and W 0 are the associated groups, both generated by the same set S. Suppose c and c0 are Coxeter elements of W and W 0 respectively that can be written as a product of the elements of S in the same order. Choose a root system Φ(A) and a root system Φ(A0 ) so that the simple co-roots are the same for the two root systems. Construct the Cambrian fan for (A, c) by coarsening the fan determined by the Coxeter arrangement for Φ(A) and construct the Cambrian fan for (A0 , c0 ) by coarsening the fan determined by the Coxeter arrangement for Φ(A0 ). Then the Cambrian fan for (A0 , c0 ) is a coarsening of the Cambrian fan for (A, c). Whereas the codimension-1 faces of the Cambrian fan for (A, c) are orthogonal to co-roots (i.e. elements of Φ∨ (A)), the Cambrian fan for (A0 , c0 ) is obtained by removing all codimension-1 faces orthogonal to elements of Φ∨ (A) \ Φ∨ (A0 ). The orthogonality appearing in Theorem 3.11 and below in Theorem 3.12 is the standard inner product on Rn defined in terms of the basis of the ei and has nothing to do with the Euclidean inner products associated to A and A0 . We wish to translate Theorem 3.11 into a more detailed version of Theorem 1.1. Theorem 3.9 identifies the Cambrian fan as FB T but by Proposition 2.12, we can equally well consider the relationship between F−B T and F−(B 0 )T . Passing from B to −B T corresponds to passing from Cart(B) to Cart(B)T while preserving c, and transposing the Cartan matrix corresponds to switching the roles of Φ and Φ∨ . Thus in rewriting Theorem 3.11 as the following theorem, we identify simple roots with the ei rather than simple co-roots. Theorem 3.12. Suppose B and B 0 are acyclic exchange matrices of finite type such that B dominates B 0 and let A = Cart(B) and A0 = Cart(B 0 ). Realize the root systems Φ(A) and Φ(A0 ) in Rn such that for each i, the simple root αi for A, the simple co-root αi0 for A0 and the unit basis vector ei all coincide. Each codimension-1 face of FB is orthogonal to an element of Φ(A), and FB 0 is obtained from FB by removing all codimension-1 faces orthogonal to roots in Φ(A) \ Φ(A0 ). Theorem 3.12 completes the proof of Theorem 1.5. Theorem 1.1 follows immediately by Proposition 3.1. 3.5. Resection of marked surfaces. We next show that rational versions of Phenomena I and II occur in some cases where B is the signed adjacency matrix of a surface. For the sake of brevity, we refrain from repeating here all of the background material found in [23]. (See also the seminal works [8, 9] on cluster algebras and surfaces.) We do, however, give the most basic background. Unjustified assertions in this basic background are established in [8]. 26 NATHAN READING Let S be a surface obtained from a compact, oriented surface without boundary by deleting a finite collection of open disks whose closures are pairwise disjoint. Usually, S is required to be connected, but this is merely an “irreducibility” criterion and is not essential for what we do here. Indeed, since we will consider an operation on S that may disconnect it, we must allow disconnected surfaces. However, compactness implies that S has finitely many components. The boundaries of the removed disks are called boundary components. Choose a finite set M of points in S called marked points. Each boundary component must contain at least one marked point. The marked points in the interior of S are called punctures. The marked points on boundary components cut the boundary components into curves called boundary segments. None of the connected components of (S, M) may be unpunctured monogons, unpunctured digons, unpunctured triangles, or spheres with fewer than 4 punctures. Typically, one excludes once-punctured monogons as well. Here, we allow once-punctured monogons, but in many of the definitions that follow, we have to single out once-punctured monogons as a special case. An arc in (S, M) is a curve in S with endpoints in M, with the following restrictions: An arc may not intersect itself, except that its endpoints may coincide. An arc must be disjoint from M and from the boundary of S, except at its endpoints. An arc may not bound an unpunctured monogon. Finally, we disallow arcs that define, together with a single boundary segment connecting the endpoints, an unpunctured digon. We consider arcs up to isotopy relative to M. Arcs α and γ are incompatible if they intersect in S \ M, and if the intersection cannot be removed by (independently) isotopically deforming the arcs. If two arcs are not incompatible, they are compatible. A triangulation of (S, M) is a maximal collection of distinct pairwise compatible arcs. Each (S, M) admits at least two triangulations, except a once-punctured monogon, which admits exactly one triangulation, consisting of the unique arc. Every triangulation of a given marked surface has the same number of arcs. These arcs divide S into triangles, each of which has 1, 2, or 3 distinct vertices and 2 or 3 distinct sides. A self-folded triangle is a triangle with 2 distinct sides. As a surface in its own right, a self-folded triangle is a once-punctured monogon with an arc α connecting the vertex of the monogon to the puncture. From the interior, this appears as a triangle, but two of the sides of the triangle are α. We call α the fold edge of the self-folded triangle. The signed adjacency matrix of a triangulation T of (S, M) is a matrix B(T ) = [bαβ ]α,β∈T indexed by the arcs of T . The definition of the entries bα,β is complicated by the possible presence of self-folded triangles. If T has no self-folded triangles, one can simplify what follows by taking πT to be the identity map. In general however, we define a map πT on the arcs of T , fixing each arc except for the arcs that are fold edges of self-folded triangles. If α is the fold edge of a self-folded triangle P in T , then πT (α) is the other edge of the triangle. The entry bαβ is the sum ∆ b∆ αβ over all triangles ∆ of T which are not self-folded of the quantities b∆ αβ    1 if ∆ has sides πT (α) and πT (β) such that πT (α)     is immediately followed by πT (β) in clockwise order.   = −1 if ∆ has sides πT (α) and πT (β) such that πT (α)    is immediately followed by πT (β) in counterclockwise order.      0 otherwise. DOMINANCE PHENOMENA 27 The matrix B(T ) is a skew-symmetric integer matrix, and thus an exchange matrix. If S has multiple connected components, then B(T ) has a block-diagonal form with a diagonal block for each component. Given a surface (S, M), the exchange matrix B(T ) is the zero matrix for some T if and only if B(T ) is the zero matrix for every T , if and only if every connected component of (S, M) is a null surface (a once-punctured monogon, once-punctured digon, or unpunctured quadrilateral). In general, one considers tagged arcs and tagged triangulations. In this section, we are primarily interested in the exchange matrices, so we appeal to [8, Proposition 12.3], which says that every exchange matrix arising from a marked surface can be obtained from an (ordinary) triangulation. Therefore we put off defining tagged arcs until Section 5.4. An allowable curve is a curve in S of one of the following forms: • a closed curve; • a curve whose two endpoints are unmarked boundary points, • a curve having one endpoint an unmarked boundary point, with the other end spiraling (clockwise or counterclockwise) into a puncture, or • a curve spiraling into (not necessarily distinct) punctures at both ends. However, an allowable curve may not • have any self-intersections, • be contractible in S \ M, • be contractible to a puncture, • have two endpoints on the boundary and be contractible to a portion of the boundary containing zero or one marked points. • have both endpoints are the same boundary segment and cut out, together with the portion of the boundary between its endpoints, a once-punctured disk, or • have two coinciding spiral points and cut out a once-punctured disk. We consider allowable curves up to isotopy relative to M. Essentially, two allowable curves are compatible if they are non-intersecting, but one kind of intersection is allowed. Suppose two allowable curves are identical except that, at one (and only one) end of the curve, they spiral opposite directions into the same point. Then these two curves are compatible unless they are contained in a component of (S, M) that is a once-punctured monogon. (There are exactly two allowable curves in a once-punctured monogon, and by convention the two are incompatible.) A (rational) quasi-lamination is a collection of pairwise compatible allowable curves and an assignment of a positive rational weight to each curve. We require the curves in the quasi-lamination to be distinct up to isotopy. The set of curves that appear in a quasi-lamination L is called the support of L and written supp(L). Given a quasi-lamination L and a triangulation T , the shear coordinate of L with respect to T is a vector b(T, L) =P (bγ (T, L) : γ ∈ T ) indexed by the arcs γ in T . We define bγ (T, L) to be the sum λ∈L wλ bγ (T, λ) over the curves λ in L, where wλ is the weight of λ in L and bγ (T, λ) is defined as follows. First, suppose γ is contained in two distinct triangles of T ◦ (rather than being the fold edge of a self-folded triangle). Since λ is defined up to isotopy, we can assume that each time λ crosses an arc α of T , it does not cross α again in the opposite direction until it has crossed some other arc β of T . (In the case where λ spirals into 28 NATHAN READING +1 −1 Figure 2. Computing shear coordinates a puncture that is incident to only one arc α, is crosses α repeatedly, but always in the same direction.) The quantity bγ (T, λ) is the sum, over each intersection of λ with γ, of a number in {−1, 0, 1} that depends on which other arcs (or boundary segments) λ visits immediately before and after the intersection with γ. (The other arcs or boundary segments are the other edges of the two triangles containing α.) Figure 2 shows the situations in which this number is 1 or −1. It is 0 otherwise. In these pictures, γ is the diagonal of the square and λ0 is the vertical or horizonal line intersecting the square. The quadrilaterals represented in Figure 2 might have fewer than 4 distinct vertices and 4 distinct sides. As a concrete example, opposite sides of the square may be identified to make a torus. Or, one of the triangles may be self-folded (with folding edge not equal to γ). Next, consider the case where γ is the fold edge of a self-folded triangle and let p be the marked point that is incident only to γ and not to any other arc in the triangulation. If λ spirals into p, then let λ0 be the allowable curve obtained from λ by reversing the direction of the spiral into p. Otherwise let λ0 = λ. Let γ 0 be the other edge (besides γ) of the self-folded triangle. Define bγ (T, λ) = bγ 0 (T, λ0 ). We can do this unless γ is contained in a component of (S, M) that is a once-punctured monogon, in which case γ 0 is a boundary segment. In this case, we define bγ (T, λ) to be +1 if λ spirals into the puncture counterclockwise or −1 if it spirals in clockwise. The following is [23, Theorem 4.4], which is essentially a rephrasing of [9, Theorem 13.6]. (In [23, Theorem 4.4], once-punctured monogons are not allowed, but the proof for once-punctured monogons is easy.) Recall that, as a set, QB(T ) is Qn when B(T ) is n × n. Theorem 3.13. Fix a tagged triangulation T . Then the map L 7→ b(T, L) is a bijection between rational quasi-laminations and QB(T ) . A tangle is any collection of allowable curves (with no requirement of compatibility) and an assignment of an integer weights (with no requirement of nonnegativity). The curves must be distinct up to isotopy. A tangle Ξ can be given shear coordinates b(T, Ξ) just as a quasi-lamination can: as a weighted sum of the shear coordinates of each curve in Ξ. A null tangle in (S, M) is a tangle Ξ with b(T, Ξ) = 0 for all tagged triangulations T . (Recall that we have not defined tagged triangulations. Since we won’t work closely with the Null Tangle Property in this paper, we can safely continue to put off the definition of tagged triangulations until Section 5.4.) For technical reasons, the definition is different if (S, M) has no boundary components and exactly one puncture. In that case, a null tangle is a tangle Ξ such that b(T, Ξ) is the zero vector for all ordinary triangulations T with all tags plain. A tangle is trivial if all of its curves have weight zero. We say (S, M) has the Null Tangle Property if every null tangle in (S, M) is trivial. DOMINANCE PHENOMENA 29 One point of the Null Tangle Property is that it allows us to construct a positive basis for QB(T ) using allowable curves. The following is a [23, Theorem 7.3], restricted to ordinary triangulations (as opposed to a tagged triangulations), with a minor correction (as explained in Remark 3.15, below). Recall that a null surface is a once-punctured monogon, once-punctured digon, or unpunctured quadrilateral. Theorem 3.14. Suppose R is Z or Q and let T be a triangulation of (S, M). If (S, M) is not a null surface, then the following are equivalent: (i) (S, M) has the Null Tangle Property. (ii) The shear coordinates of allowable curves form a basis for QB(T ) . (iii) The shear coordinates of allowable curves form a positive basis for QB(T ) . (By Proposition 2.20, this is also a cone basis for QB(T ) .) If (S, M) is a null surface, then it fails (i) and satisfies (ii) and (iii). Remark 3.15. Theorem 3.14 should be taken as a correction to [23, Theorem 7.3]. Once-punctured monogons are not allowed in [23], and [23, Theorem 7.3] takes a tagged triangulation T , but [23, Theorem 7.3] should be corrected by dealing separately with the quadrilateral and digon as in Theorem 3.14. The assertions of Theorem 3.14 for null surfaces are easy exercises. The correction to [23, Theorem 7.3] contradicts [23, Theorem 7.4] for the quadrilateral and digon, but does so in a way that preserves the truth of [23, Corollary 7.5], and therefore preserves all of the results of the paper on universal coefficients. The error arose because the proof of [23, Proposition 7.9] used [23, Proposition 2.3], whose hypotheses rule out the quadrilateral and digon (despite the assertion in the paragraph before [23, Proposition 3.5]). The main constructions and results of [23] are all valid, even for the quadrilateral and digon, but some some auxiliary results are wrong as stated for the quadrilateral and digon. We say (S, M) has the Curve Separation Property if, given incompatible allowable curves λ and λ0 , there exists a tagged triangulation T and an arc γ ∈ T such that bγ (T, λ) and bγ (T, λ0 ) have strictly opposite signs. (Again, for technical reasons, the definition is different if (S, M) has no boundary components and exactly one puncture. In this case we require that T be an ordinary triangulation.) The Null Tangle Property implies the Curve Separation Property [23, Corollary 7.14]. Furthermore, it is an easy exercise to check that every null surface satisfies the Curve Separation Property. (We know of no surfaces that fail the Curve Separation Property, but it remains unproven for surfaces with no boundary components and exactly one puncture.) The point of the Curve Separation Property is that is allows us to understand the rational part of the mutation fan FB . (See Definition 3.2.) Given a triangulation T of (S, M), the rational  quasi-lamination fan FQ (T ) is the collection of cones of the form SpanR≥0 b(T, λ) : λ ∈ Λ such that Λ is a set of pairwise compatible allowable curves. The following is a special case of [23, Theorem 4.10]. Theorem 3.16. Suppose T is a triangulation of (S, M). Then FQ (T ) is a rational simplicial fan. It is the rational part of FB(T ) if and only if (S, M) has the Curve Separation Property. We now introduce an operation, called resection, on marked surfaces that induces a dominance relation on signed adjacency matrices. 30 NATHAN READING −→ α pα α Figure 3. Resection of a surface Definition 3.17 (Resecting a surface at an arc). The resection operation on surfaces is illustrated in Figure 3. Suppose α is an arc in (S, M) connecting marked points p1 and p2 . Place a new marked point pα in the interior of S close to α. Draw a curve in S connecting p1 to p2 and forming a digon with pα in its interior, but containing no other marked points. Draw two more curves inside the digon, one connecting p1 to pα and the other connecting p2 to pα . This cuts the digon into two triangles. Remove the interior of the one that does not have α as an edge. If p1 is on a boundary segment, then resulting surface must be cut at the point p1 in order to satisfy the requirement that the boundary components be circles. In this case p1 becomes two marked points. Similarly, if p2 is on a boundary segment, then the surface must be cut at the point p2 . The resulting marked surface is a resection of (S, M) at α, and accordingly we use the verb “resect” and the noun “resection” to describe passing from (S, M) to the resected surface. The resected surface S0 may be disconnected even if S is connected. Typically, there are two possible resections at α, one for each side of α. However, we disallow any resections that leave a connected component that is an unpunctured triangles, because unpunctured triangles are disallowed in the definition of a marked surface. Definition 3.18 (Resection at a collection of arcs). More generally, a resection of (S, M) is a marked surface obtained by performing any collection of resections at arcs, including possibly resecting the same arc on both sides. This is well-defined up to isotopy as long as the arcs in question are compatible and the resections are not disallowed as described above. Definition 3.19 (Resection compatible with a triangulation). Fix a triangulation T of (S, M). A resection of (S, M) is compatible with T if • Each arc that is resected is an arc of T . • For each arc α that is resected, the point pα is placed in the interior of a triangle of T bounded by α and the new curves that define the resection never leave the interior of that triangle except possibly at their endpoints. • If an arc α that is resected has an endpoint at a puncture q, then either both endpoints of α are at q or at least one more arc β with an endpoint at q is resected. In the latter case, we require that pα and pβ are in different triangles of T . These requirements imply in particular that a resected arc α may not be the fold edge of a self-folded triangle of T . Equivalently, a resected arc α is contained in two distinct triangles of T . The resection must also satisfy the requirements of Definition 3.17 by not cutting off an unpunctured triangle. Definition 3.20 (Triangulation induced on the resected surface). If (S0 , M0 ) is a resection of (S, M) that is compatible with T , then each arc in T is also an arc DOMINANCE PHENOMENA 31 α α −→ β γ α α β γ  α β p = γ p γ β β γ 0 2 −2 −2 0 2 2 −2 0 α  α β γ  β γ 0 1 −1 −1 0 2 1 −2 0  Figure 4. An example of resection α α γ −→ γ α = β γ p β β p α α β γ  β γ 0 1 −1 −1 0 2 1 −2 0 α  α β γ  β γ 0 1 −1 −1 0 1 1 −1 0  Figure 5. Another example of resection in (S0 , M0 ). These arcs cut (S0 , M0 ) into triangles, defining a triangulation T 0 of (S0 , M0 ) called the triangulation induced on (S0 , M0 ) by T . Example 3.21. Figures 4 and 5 show two examples of resection (at a single arc) compatible with a triangulation, showing the induced triangulation in each case. The signed adjacency matrix of each triangulation is also shown. In Figure 4, the surface is a torus with one puncture, obtained by identifying opposite edges of a square. The puncture is the point at the corners of the square. In Figure 4, the resection is at α, and in Figure 5, the resection is at γ. Remark 3.22. The marked points and the arcs in a triangulation T define a graph (which is not necessarily simple, in that it can have multiple edges and loops). Choosing a resection compatible with T is equivalent to choosing an edge-induced subgraph and then orienting the subgraph. (We take the convention that, moving along the edge with the orientation, the new point pα is on the right.) Definition 3.19 in particular requires that any leaf (degree-1 vertex) of the subgraph must be on the boundary of S, and also, together with Definition 3.17 makes some restrictions on how the subgraph may be oriented. 32 NATHAN READING β δ β δ −→ α γ ǫ α γ ǫ Figure 6. An illustration for the proof of Proposition 3.23 γ γ α β δ −→ α β δ Figure 7. Another illustration for the proof of Proposition 3.23 Proposition 3.23. Given a marked surface (S, M) with a triangulation T , perform a resection of (S, M) compatible with T and let T 0 be the triangulation induced by T on the resected surface. Then B(T ) dominates B(T 0 ). Proof. We will prove two assertions that together amount to a version of the proposition with weaker hypotheses. First, if α is an arc in T not incident to the puncture in a once-punctured digon and T 0 is the triangulation obtained by resecting at α according to the rule in the second bullet point in Definition 3.19, then B(T ) dominates B(T 0 ). The part of T right around α is illustrated in Figure 6. Since α is not incident to the puncture in a once-punctured digon, β 6= δ and γ 6= , and α is distinct from the other four arcs. The signed adjacency matrix B(T 0 ) is obtained from B(T ) as follows. If β is an arc (rather than a boundary segment), then decrease the αβ-entry by 1 and increase the βα-entry by 1. If γ is an arc, then increase the αγ-entry by 1 and decrease the γα-entry by 1. Second, if α and β are distinct arcs in T , both incident to the puncture in a once-punctured digon and T 0 is the triangulation obtained by resecting at α and at β according to the rule in the second bullet point in Definition 3.19 and with pα and pβ in different triangles of T , then B(T ) dominates B(T 0 ). The situation is illustrated in Figure 7. The αβ entry in the signed adjacency matrix is 0 before and after resection. The arcs δ and γ do not coincide, because if they do, the unresected surface is a sphere with only 3 punctures. Thus if γ is an arc, resection decreases the absolute value of the βγ and γβ entries by 1, and if δ is an arc, resection decreases the absolute value of the αδ and δα entries by 1. (If neither γ nor δ is an arc, then the signed adjacency matrix is unchanged.)  Remark 3.24. Not every exchange matrix dominated by B(T ) can be obtained by resecting the associated surface (even for the broader class of resections allowed in the proof of Proposition 3.23). Not even every skew-symmetric exchange matrix dominated by B(T ) can be obtained. This is because, in some cases, resection must change four entries of B(T ). Specifically, in the labelling shown in Figure 6, when neither β nor γ is a boundary segment, the entries bαβ , bαγ , bβα , and bγα are all DOMINANCE PHENOMENA 33 changed. In this case, it is impossible to change only the entries bαβ and bβα , even though the result would be an exchange matrix dominated by B(T ). Our next task is to prove Theorems 1.2 and 1.7, starting with the latter. Recall from the introduction that FQ (T ) is the fan whose cones are the real spans of shear coordinates of collections of pairwise compatible allowable curves. Theorem 1.7 supposes that T is a triangulation of (S, M), that (S0 , M0 ) and T 0 are the marked surface and triangulation obtained from T by a resection compatible with T , and that every component of (S, M) and (S0 , M0 ) either has the Null Tangle Property or is a null surface. The conclusion of the theorem is that FQ (T ) ∩ Qn refines FQ (T 0 ) ∩ Qn . Proof of Theorem 1.7. Theorem 3.13 says that L 7→ b(T, L) is a bijection between rational quasi-laminations and Qn and that L0 7→ b(T 0 , L0 ) is a bijection between rational quasi-laminations and Qn . Thus there is a bijection sending each rational quasi-lamination L on (S, M) to a rational quasi-lamination L0 on (S0 , M0 ) such that b(T, L) = b(T 0 , L0 ). For each C ∈ FQ (T ), the set C ∩ Qn is the set of shear coordinates of all nonnegative rational weightings on some collection Λ of pairwise-compatible allowable curves in (S, M). We prove the theorem by showing that, for each such Λ, there is a collection Λ0 of pairwise-compatible allowable curves in (S0 , M0 ) such that for any nonnegative rational weighting on Λ, there is a nonnegative rational weighting on Λ0 giving the same shear coordinates. Thus C ∩ Qn is contained in the cone of FQ (T 0 ) consisting of shear coordinates of nonnegative rational weightings on Λ0 . We can almost, but not quite, construct Λ0 by resecting at one arc at a time; we will need to appeal to the third condition in Definition 3.19. We consider a resection at α. Let αβγ be the triangle where the point pα is placed in the process of resecting at α. The situation is illustrated in Figure 8. Notice that α cannot coincide with β or with γ, because if so, α has an endpoint incident to only one arc and resecting at α violates the third condition in Definition 3.19. However, possibly β and γ coincide, so that this is a self-folded triangle. We will leave this possibility to the end of the proof. Until then, we are free to compute shear coordinates at α, β, and γ without invoking the special rule for shear coordinates on the fold edge of a self-folded triangle. Assume as in the definition of shear coordinates that pairwise compatible isotopy representatives of the curves in Λ have been chosen so that no curve crosses an arc and then immediately doubles back to cross in the opposite direction. Consider all intersections of the curves in Λ with the triangle αβγ. A single curve in the quasi-lamination may intersect this triangle many times, and even infinitely many times if it spirals into a vertex of the triangle. In the first step of constructing Λ0 from Λ, we only alter the curves inside the triangle. Up to isotopy of curves in (S, M), we can assume that no curve connecting α and β (inside the triangle) separates the point p from the arc γ. The same statement holds with α, β, and γ cyclically permuted. Thus we can assume that p is situated with respect to the curves as illustrated in the left picture of Figure 8. Furthermore, we can assume that the curves intersect the new boundary component as illustrated in the middle picture of Figure 8. In particular, the boundary component does not intersect any of the curves connecting β to γ. Now remove from each curve its intersection with the new boundary component, as illustrated in the right picture of Figure 8. This cuts many curves in Λ into smaller pieces. The result is a 34 NATHAN READING β α β α p γ β α p γ p γ Figure 8. The identity map in terms of quasi-laminations Figure 9. Replacing a bad curve with two compatible allowable curves collection Λ0 of curves that may not be a quasi-lamination because it may contain some curves that are not allowable and because we have not yet checked that the curves are pairwise compatible. However, if we let each piece inherit a weight from the original weighted collection and allow ourselves to compute shear coordinates of the new weighted collection, including the bad curves, we obtain again the shear coordinates of the original collection. Since the curves in Λ are pairwise compatible, and because we chose pairwise compatible isotopy representatives such that no curve crosses an arc and then immediately doubles back to cross in the opposite direction, each bad curve that occurs fits one of the following two descriptions: • It has two endpoints on a boundary segment and is contractible to a portion of the boundary containing one marked point. • It has both endpoints on the same boundary segment and, together with the portion of the boundary between its endpoints, it cuts out a once-punctured disk. The first type of bad curve can occur infinitely many times when one or more curves in Λ spirals around an endpoint of α. Since each such curve has all shear coordinates zero, we can we delete all such curves from the collection without changing the shear coordinates. The second type of bad curve makes a nonzero contribution to the shear coordinates of the collection. We replace each such bad curve by two curves that start on the same boundary component as the bad curve and spiral around the puncture that the bad curve encloses. The two are identical except for spiraling opposite ways into the puncture. The total of the shear coordinates of these two curves equals the shear coordinates of the bad curve. Furthermore, they are are allowable, are compatible, and, since the original set λ is pairwise compatible, are compatible with the remaining curves. Figure 9 represents the bad curve (solid) and the two allowable curves (dashed) that replace it. It is possible that distinct curves in the new collection may coincide up to isotopy. In this case, we delete repetitions of curves and adjust weights accordingly. All of DOMINANCE PHENOMENA β α 35 β α p γ β α p p γ γ Figure 10. A problem ruled out by Definition 3.19 α p β α p β Figure 11. The identity map on quasi-laminations for β = γ these modifications have preserved the shear coordinates of the weighed collection. Call the new collection Λ0 . It remains only to show that the curves in Λ0 are pairwise compatible. If all curves in Λ are pairwise non-intersecting, then all curves in Λ0 are pairwise nonintersecting and therefore pairwise compatible. Recall, however, that two allowable curves that intersect are compatible if and only if they are identical expect for spiraling in the opposite direction at exactly one of their endpoints. Consider two such curves in Λ. If they spiral into a point other than an endpoint of α, then they are cut into one or more curves in Λ0 that remain compatible. If they spiral into an endpoint of α, but not just after passing through the triangle αβγ, then they are cut into infinitely many pieces, most of which are discarded, and the remaining pieces are non-intersecting. However, if they pass through the triangle αβγ just before spiraling into an endpoint of α (and if we forget the third condition in Definition 3.19), then the resulting non-discarded pieces may intersect, as illustrated in the left two pictures of Figure 10. However, the third condition in Definition 3.19 requires that some other arc incident to the endpoint of α is also resected, so this problem never occurs. See the right picture of Figure 10. Recall that we left the case where β and γ coincide to the end of the proof. In this case, we can’t compute shear coordinates at β = γ without invoking the special rule for fold edges of self-folded triangles. The resection cuts off a once-punctured digon (the self-folded triangle), triangulated by the single arc β = γ. The other triangle (besides the self-folded αβγ) having α as an edge is not self-folded, because if so, (S, M) is a three-times-punctured sphere. Thus the situation is as pictured in Figure 11. In the left picture of the figure, the curves that come from below and spiral counterclockwise into the puncture contribute positively to the shear coordinate of β with respect to T . The curves that come from below, go around 36 NATHAN READING the puncture, and return downwards also contribute positively. Curves that come from above and spiral clockwise (not pictured) or curves that come from above, go around the puncture and return upwards (also not pictured) would contribute negatively. (Curves from below that spiral clockwise and curves from above that spiral counterclockwise contribute zero.) We place the point p as shown in the figure (or similarly if instead there are curves from above that go around the puncture and return upwards). By the same construction outlined above, we construct a pairwise compatible collection of allowable curves in the components of (S0 , M0 ) aside from the once-punctured monogon cut off by the resection. The construction leaves some pieces of curves in the once-punctured monogon, as indicated in the right picture of Figure 11. There are 7 types of curves, and we need to distinguish them according to their behavior in the digon of T from which the monogon was cut. First, there are 4 types of curves with spirals: two spiral directions, with curves originating from the top or bottom of the digon. We delete curves from below that spiral clockwise and curves from above that spiral counterclockwise, and retain the other two types. (Since the original collection of curves was pairwise compatible, at most one of the other two types is present.) Second, there are 3 types of curves that enter the monogon, cross β, and leave the digon. The first type come from curves that cross the digon from top to bottom; we delete these. The second type come from curves that come from the bottom of the digon, go around the puncture, and exit the bottom of the digon. We replace each such curve with a new curve (with the same weight) that enters the monogon and spirals counterclockwise into the puncture. The third type come from curves that come from the top of the digon, go around the puncture, and exit the top of the digon. Each of these is replaced with a curve that enters the monogon and spirals clockwise into the puncture. Again, we delete repetitions of curves and adjust weights accordingly. These weighted curves, together with the curves we constructed outside of the monogon, are pairwise compatible and have the same shear coordinates as the original weighed collection. This is the desired new collection Λ0 .  Recall that, under the same hypotheses as Theorem 1.7, Theorem 1.2 asserts 0 that the identity map QB → QB is mutation-linear. Proof of Theorem 1.2. By Theorem 3.14 and Proposition 2.20, QB(T ) admits a cone basis. Theorem 3.16 says that FQ (T ) is the rational part of FB(T ) and FQ (T 0 ) is 0 the rational part of FB(T 0 ) . By Proposition 3.3, the identity map QB → QB is mutation-linear.  Example 3.25. We illustrate the shear-coordinate-preserving bijection from rational quasi-laminations in a surface to rational quasi-laminations in a resected surface for the simplest nontrivial resection: Resecting an arc in a triangulated pentagon. The resection is shown in Figure 12. The quasi-laminations of the pentagon are represented in the left column of Figure 13. Each is given by two curves, each with a weight (labeled a or b in the figure). The right column shows the image of each quasi-lamination under the bijection. The five rows represent the five maximal cones in the mutation fan for the pentagon. The mutation fan for the union of two squares has four maximal cones, one of which is the union of two cones of the fan for the pentagon. DOMINANCE PHENOMENA −→ 37 = Figure 12. A resection of the pentagon a 7→ a 7→ a b 7→ b b b a b b a b a 7→ b a 7→ b a b b a b a Figure 13. The bijection between quasi-laminations on the pentagon and quasi-laminations on the union of two squares 38 NATHAN READING Figure 14. The mutation fan for the once-punctured torus The proof of Theorem 1.2 works because the bijection on quasi-laminations is simple in the forward direction (the direction we described). Specifically, the curves appearing in the quasi-lamination L0 depend only on the curves in L. By contrast, in the inverse direction, weights on curves in L0 must be taken into account in order to determine what curves appear in L. For example, it seems nearly hopeless to describe the inverse maps in the following example, which continues Example 3.21. Example 3.26. Figures 14, 15, and 16 show the mutation fans associated to the three triangulated surfaces of Example 3.21. Each figure is a projection of a fan in R3 obtained as follows: The intersection of the fan with a sphere about the origin is a decomposition of the sphere, which is projected stereographically to the plane to produce the picture shown. Each figure is projected in the same way, making visible the refinement relationship guaranteed by Theorem 1.2. (The figures are placed so that paging through the electronic version of this paper highlights the relationship.) 4. Refinement of scattering fans In this section, we review the definition of the scattering fan and discuss examples of Phenomenon III. Recall that Theorem 1.12 establishes Phenomenon III for acyclic finite type and that Corollary 1.15 establishes Phenomenon III for surfaces of finite (not necessarily acyclic) type. It was explained in the introduction how both of these follow from cases of Phenomenon II in light of Theorem 1.10, the assertion that the scattering fan refines the mutation fan. In this section, we validate these results by proving Theorem 1.10. We also establish an additional example of Phenomenon III, namely the 2 × 2 case, by proving Theorem 1.16. DOMINANCE PHENOMENA 39 Figure 15. The mutation fan for an annulus 4.1. Scattering fans. We will not define the scattering fan completely, but instead, we will indicate what kind of an object it is and quote results that describe its important properties. Scattering fans are fans defined by scattering diagrams, as we now describe. For more on scattering diagrams, see [14], or for an exposition more suited to the goals of this paper, see [26]. We work with principal coefficients, so that the initial data for a scattering diagram amounts to an exchange matrix. A scattering diagram D is a collection of walls. Each wall is a pair (d, fd ), where d is a codimension-1 rational cone in Rn , normal to a nonzero vector with nonnegative entries and fd is a formal power series in n variables with constant term 1. The scattering diagram may have infinitely many walls, but for each k ≥ 1 only finitely many walls have nonzero coefficients on terms of degree d with 0 < d ≤ k. To each sufficiently generic path γ in Rn , we associate the sequence of walls crossed by γ, and using this sequence and the exchange matrix B, we define a certain automorphism of a larger formal power series ring (in 2n variables). (Some issues arise when D is infinite, but these are resolved using the finiteness requirement for each k and taking a limit.) A scattering diagram is consistent if the automorphism arising from a path depends only on the endpoints of the path. Two scattering diagrams D and D0 are equivalent if any (sufficiently generic) path defines the same automorphism with respect to both scattering diagrams. We only care about scattering diagrams up to equivalence. Given a scattering diagram D and a normal vector n, the rampart associated to n is the union of all walls of D that are normal to n. Given p ∈ Rn , let RamD (p) be the set of ramparts of D containing p. We say points p and q in Rn are Dequivalent if there is a path from p to q on which RamD ( · ) is constant. The 40 NATHAN READING Figure 16. The mutation fan for a hexagon closures of D-equivalence classes are called D-cones. Let Fan(D) be the set of all D-cones and their faces. This is a complete fan [26, Theorem 3.1]. The (principal coefficients) cluster scattering diagram Scat(B) associated to an exchange matrix B is defined to be the unique (up to equivalence) consistent scattering diagram satisfying certain conditions. We won’t give full details, but one of the conditions is that each coordinate hyperplane in Rn , paired with a specific formal power series (in fact a very simple polynomial) determined by B, is a wall of Scat(B). The other condition uses B to place limits on the other walls that can appear. We write ScatFan(B) for Fan(Scat(B)), the (principal coefficients) scattering fan for B. 4.2. Scattering fans refine mutation fans. Our proof of Theorem 1.10 consists of a simple observation about mutation fans and a rephrasing of a result of [14] on scattering diagrams. First, the observation. Proposition 4.1. Consider a family of fans in RB , with one fan EB 0 for each B 0 obtained as µk (B) for some sequence k of indices. Suppose that each cone of each EB 0 is a sign-coherent set. Suppose also that, for each B 0 , each sequence k, and each 0 cone C of EB 0 , the set ηkB (C) is contained in a cone of Eµk (B 0 ) . Then ScatFan(B) refines the mutation fan FB . Proof. Let C be a cone of EB . By hypothesis, for any sequence k, the set ηkB (C) is contained in some cone of Eµk (B) and thus ηkB (C) is sign-coherent. Now Proposition 2.15 says that C is contained in some B-cone.  Since each coordinate hyperplane is a wall of Scat(B) by definition, each cone of ScatFan(B) is contained in some closed coordinate orthant of RB . The following result is [26, Corollary 4.5]. DOMINANCE PHENOMENA 41 Proposition 4.2. For any sequence k of indices, the mutation map ηkB is a piecewise linear isomorphism from ScatFan(B) to ScatFan(µk (B)). We see that ScatFan(B) satisfies the hypotheses of Proposition 4.1. Thus the proof of Theorem 1.10 is complete. 4.3. The 2 × 2 case. We now prove Theorem 1.16, which says that in the 2 × 2 case, there is a refinement relation among scattering diagrams if and only if there is a corresponding dominance relation among exchange matrices. We make use of the background from [22, Section 9] (already quoted in Section 3.3) on the mutation  fan FB for a 2 × 2 exchange matrix B = 0b a0 . The Discreteness Hypothesis in the statement of Theorem 1.16 is the assertion that, for ab < −4, every slope weakly between s∞ (a, b) and t∞ (a, b) is a relevant slope (in the sense of Section 3.3) for the scattering fan ScatFan(B).     Proof of Theorem 1.16. Suppose B = 0b a0 and B 0 = d0 0c are 2 × 2 exchange matrices. In neither is wild, then since the scattering fan coincides with the mutation fan in the non-wild 2 × 2 case, Theorem 1.9(2) implies that ScatFan(B) refines ScatFan(B 0 ) if and only if B dominates B 0 . If B 0 is wild and B is not, then ScatFan(B 0 ) has many irrational relevant slopes, while ScatFan(B) has only rational slopes, so ScatFan(B) does not refine ScatFan(B 0 ). If neither is of finite type, we first show that dominance is necessary for refinement. Indeed, suppose that B does not dominate B 0 , so that either a < c or b > d. If a < c, then the smallest relevant slope of ScatFan(B) is s0 (a, b) = −a, while the smallest relevant slope of ScatFan(B 0 ) is s0 (c, d) = −c. Thus in particular the smallest relevant slope of ScatFan(B 0 ) is not a relevant slope of ScatFan(B), so ScatFan(B) does not refine ScatFan(B 0 ). Similarly, if b > d, then the two largest relevant slopes are t0 (a, b) = 1b and t0 (c, d) = d1 , and thus ScatFan(B) does not refine ScatFan(B 0 ). Continuing in the case where neither B nor B 0 is of finite type, we show that dominance is sufficient for refinement. It is enough to consider two cases: the case where c = a and d = b + 1 and the case where c = a − 1 and d = b. First, consider the case where c = a and d = b + 1. In this case, s0 (a, b) = −a = −c = s0 (c, d). If a > 1, then one can show that s∞ (a, b) < s1 (c, d) and t0 (c, d) < t∞ (a, b). (Using the fact that a > 0 and b < b + 1 < 0, both of these inequalities can be shown to be equivalent to (1 − a)b > a, which holds because a > 1 and b < −1.) Now the Discreteness Hypothesis implies that every relevant slope of ScatFan(B 0 ) is a relevant slope of ScatFan(B) as desired. −ab−2 If on the other hand, a = 1, then s2 (a, b) = −a −ab−1 = −b−2 b+1 and s1 (c, d) = −cd−1 1 1 = −a(b+1)−1 = −b−2 d b+1 b+1 . Furthermore, t1 (a, b) = −a −ab−1 = b+1 and t0 (c, d) = 1 b+1 . To complete the proof in the case where a > 1, we show that s∞ (a, b) < s2 (c, d) and t1 (c, d) < t∞ (a, b). (Both equations are shown to be equivalent to b < −4, which holds because a = 1 and a(b + 1) ≤ −4.) Again the Discreteness Hypothesis now implies that every relevant ray of ScatFan(B 0 ) is a relevant ray of ScatFan(B). In the case where c = a − 1 and d = b, we argue similarly or appeal to symmetry. We have proven the theorem in the case where both B and B 0 are non-wild and in the case where both B and B 0 are of infinite type. It thus remains to prove the theorem in the case where one is of finite type and the other is of wild type. If B is of finite type and B 0 is of wild type, then B does not dominate B 0 and 42 NATHAN READING ScatFan(B) does not refine ScatFan(B 0 ) because the latter has irrational relevant slopes while the former does not. Finally, suppose B is of wild type and B 0 is of finite type. If B dominates B 0 , then there is an exchange matrix B 00 of affine type such that B 00 dominates B 0 and B dominates B 00 . Concatenating the results we have proved, we see that ScatFan(B) refines ScatFan(B 0 ). If B = 0b a0 does not dominate B 0 = d0 0c , then there are two possibilities: Either a < c or b > d. Notice that in every case (as shown in Table 1), the smallest relevant slope of ScatFan(B 0 ) is −c. The smallest relevant slope of ScatFan(B) is s0 (a, b) = −a. Thus if a < c, ScatFan(B) does not refine ScatFan(B 0 ). Similarly, the largest relevant slope of ScatFan(B 0 ) is d1 and the largest relevant slope of ScatFan(B) is 1b . Thus also if b > d, ScatFan(B) does not refine ScatFan(B 0 ). We have checked that in every 2 × 2 case, ScatFan(B) refines ScatFan(B 0 ) if and only if B dominates B 0 .  5. Morphisms of cluster algebras In this section, we prove results and describe computer verifications that provide examples of Phenomenon IV. 5.1. General considerations. Let B be an n × n exchange matrix and let x = x1 , . . . , xn and y = y1 , . . . , yn be indeterminates. Let L stand for Z[x±1 , y] = ±1 Z[x±1 1 , . . . , xn , y1 , . . . , yn ], the ring of Laurent polynomials in x with coefficients integer polynomials in y. Let K stand for the field of rational functions in x and y. A seed is a pair (B̃, (v1 , . . . , vn )), where B̃ is a 2n × n integer matrix (an extended exchange matrix ) whose top n rows are an exchange matrix and (v1 , . . . , vn ) is an ordered n-tuple of elements of K. The tuple (v1 , . . . , vn ) is called a cluster . (This is a special case of the definition of a seed of geometric type [12].) Given a seed (B̃, (v1 , . . . , vn )) and k ∈ {1, . . . , n}, we can mutate to obtain a new seed µk (B̃, (v1 , . . . , vn )) = (B̃ 0 , (v10 , . . . , vn0 )), where B̃ 0 = µk (B̃) is described in (2.1), vi0 = vi for i 6= k, and vk is described by the exchange relation (5.1) vk vk0 = n Y i=1 [bik ]+ vi n Y i=1 [b(n+i)k ]+ yi + n Y i=1 [−bik ]+ vi n Y [−b(n+i)k ]+ yi , i=1 where the bik are entries of Bt and [b]+ means  max(0, b). We take as an initial seed the pair BI , (x1 , . . . , xn ) , where B is an exchange matrix and I is the identity matrix. A cluster variable with to this   respect initial seed is an entry in a cluster that can be obtained from BI , (x1 , . . . , xn ) by an arbitrary sequence of mutations. Typically there are infinitely many cluster variables. If not, then B is of finite type. The principal coefficients cluster algebra A• (B) is the subring of K generated by all cluster variables. Different authors have chosen different conventions about whether the variables y are inverted in A• (B). Very importantly to what follows, we chose above not to invert the y and thus, by the Laurent Phenomenon [10, Theorem 3.1], the cluster algebra A• (B) is a subring of L. Write Var• (B) for the set of cluster variables in A• (B). Consider a Zn -grading of L = Z[x± , y] given by setting the degree of each xk to be the standard basis vector ek and setting the degree of each yk to be the negative of the k th column of B. If an element x of L is homogeneous with respect to this grading, then its Zn -degree is called the g-vector of x. Each cluster variable in A• (B) is homogeneous in this grading and thus has a g-vector [12, Proposition 6.1]. DOMINANCE PHENOMENA 43 To consider homomorphisms, we need another cluster algebra, which we will construct just as above, but with primes on every symbol. Thus B 0 is an n0 × n0 matrix, L0 is the ring of Laurent polynomials in x0 = x01 , . . . , x0n0 with coefficients polynomials in y0 = y10 , . . . , yn0 0 . Also, A• (B 0 ) is the cluster algebra with initial exchange matrix B 0 , initial cluster x0 and principal coefficients (in y0 ) at B 0 , with y0 not inverted, Var• (B 0 ) is the set of cluster variables of A• (B 0 ), etc. Given a set map ν from x0 , y0 to L, there is a unique ring homomorphism from 0 L to K agreeing with ν on x0 , y0 . We reuse the symbol ν for the homomorphism. Given a Laurent polynomial p ∈ L0 , we obtain ν(p) by replacing each x0i and yj0 by ν(x0i ) and ν(yj0 ) respectively in any expression for p as a rational function of x0 and y0 . The following observation is immediate from the definition of A• (B 0 ).  Proposition 5.1. Suppose ν : x0 , y0 → L is a set map and reuse the symbol ν for the extension to a ring homomorphism from L0 to K. If ν(Var• (B 0 )) ⊂ A• (B), then ν restricts to a ring homomorphism from A• (B 0 ) to A• (B). Given a set map ν : Var• (B 0 ) ∪ {y} → A• (B), each exchange relation in A• (B 0 ) is mapped to a equation (valid or not) relating elements of A• (B).  Proposition 5.2. Suppose ν is a set map from Var• (B 0 ) ∪ y0 to A• (B). If ν maps every exchange relation of A• (B 0 ) to a valid relation in A• (B), then ν extends to a ring homomorphism from A• (B 0 ) to A• (B).  Proof. Restrict the set map ν to x0 , y0 and consider the extension of ν to a ring homomorphism ν : L0 → F . Suppose z 0 is a cluster variable in A• (B 0 ). Then z 0 is in L0 . We claim that ν(z 0 ) = ν(z 0 ). We argue by induction on the smallest number k of mutation steps required to obtain z 0 from the initial cluster x0 . Given a sequence of k mutations producing z 0 from x0 , the last mutation defines an exchange relation writing z 0 in terms of the yi and in terms of cluster variables wi0 that can be reached from x0 in fewer mutations. Since ν maps each exchange relation to a valid relation, we obtain an expression for ν(z 0 ) in terms of the quantities ν(wi0 ). By induction, each wi has ν(wi0 ) = ν(wi0 ). Thus ν(wi0 ) can be obtained from any expression for wi0 by replacing each x0i and yj0 by ν(xi ) and ν(yj ). We have found an expression for z 0 such that replacing each x0i and yj0 by ν(xi ) and ν(yj ) yields ν(z 0 ). Therefore ν(z 0 ) = ν(z 0 ). Since we know that ν sends every cluster variable of A• (B 0 ) to an element of A• (B), we have verified the hypotheses of Proposition 5.1 and obtained a homomorphism from A• (B 0 ) to A• (B). Furthermore, this homomorphism agrees with  the given set map from Var• (B 0 ) ∪ y0 to A• (B). We will often use the following criterion to checking that νz is injective. Proposition 5.3. Suppose B dominates B 0 . If there is at most one index k such that there exists i with bik < b0ik ≤ 0, then the homomorphism νz is injective. Proof. Up to reindexing, we may as well assume that n is the unique index such that there exists i with bin < b0in ≤ 0. (If there is no pair ik such that bik < b0ik ≤ 0, then an easier version of the following argument works.) For all k < n, we can take zk to be a cluster monomial, specifically, a monomial in the initial cluster h variables i ∂ ν(x0i ) 0 0 x1 , . . . , xn . Writing xn+i for yi and xn+i for yi , the Jacobian matrix is ∂xj upper-triangular. The first n diagonal entries are 1. The next n−1 diagonal entries are (possibly trivial) monomials in x1 , . . . , xn . The last diagonal entry is ∂y∂ (yn zn ), n 44 NATHAN READING which is not zero because yn is not a factor of zn . (This is immediate from the definition in either the theta-function case or the surfaces case.) The Jacobian determinant is the product of these diagonal entries, and is therefore not zero. Non-vanishing of the Jacobian determinant is a well-known criterion for injectivity (see, for example, [5, Theorem 2.2]).  5.2. The 2 × 2 case. We now prove Theorems 1.18 and 1.19, which provide examples of Phenomenon IV in the 2 × 2 case. Although there is some overlap in the two theorems, we prove them mostly separately, The advantage is that we prove Theorem 1.18 (almost) entirely using the basic definition of cluster variables, with (almost) no theta functions (and without the need for the Discreteness Hypothesis). Theorem 1.18 starts with a 2 × 2 exchange matrix B of finite or affine type and another 2 × 2 exchange matrix B 0 dominated by B and asserts that νz is an injective ring homomorphism from A• (B 0 ) to A• (B), preserving g-vectors and taking each ray theta function in ScatFanT (B 0 ) to the theta function for the same ray in ScatFanT (B). We will see that Theorem 1.18 reduces to the following proposition. Proposition 5.4. Under the hypotheses of Theorem 1.18, the map νz sends each cluster variable of A• (B 0 ) to a cluster variable of A• (B) or, in the case where B is of affine type, possibly to the theta function of the limiting ray. When B is of affine type, the only ray theta function that is not a cluster variable is the ray theta function for the limiting ray. Furthermore, since B 0 is strictly dominated by B, it is of finite type, so all of its ray theta functions are cluster variables. Thus Proposition 5.4 verifies that νz takes each ray theta function in ScatFanT (B 0 ) to the theta function for the same ray in ScatFanT (B). Necessarily, each zk is a power of a cluster variable with g-vector [ ±1 0 ]or [ 0 ±1 ]. Assuming Proposition 5.4, Proposition 5.1 implies that the map νz is a ring homomorphism from A• (B 0 ) to A• (B) preserving g-vectors and Proposition 5.3 implies that νz is injective. We have verified that Proposition 5.4 implies Theorem 1.18. 0 To prove Proposition   5.4, we can, by symmetry (transposing   both B and B ), assume that B = a0 0b with a < 0 and b > 0 and B 0 = 0c d0 with c ≤ 0 and d ≥ 0. Furthermore, if B dominates B 0 , which dominates B 00 and we check the proposition for B and B 0 , and also check it for B 0 and B 00 , then in particular the map from A• (B 0 ) to A• (B) will take the cluster variables of A• (B 0 ) with g-vectors [ ±1 0 ] and [ 0 ±1 ] to the cluster variables of A• (B) with the same g-vectors. Thus the composition of the maps from A• (B 00 ) to A• (B 0 ) and from A• (B 0 ) to A• (B) will coincide with the map directly from A• (B 00 ) to A• (B), and will affect cluster variables of A• (B 00 ) as desired. In light of these facts and Theorem 1.9, we need to check the following cases.     Case 1: B 0 =  00 00 and B = a0 0b .  Case 2: B 0 =  −10 10  and B =  −10 0b . Case 3: B 0 =  −10 10  and B =  a0 01 . Case 4: B 0 =  −10 20  and B =  −10 30 . Case 5: B 0 =  −10 20  and B =  −20 20 . Case 6: B 0 =  −20 10  and B =  −30 10 . Case 7: B 0 =  −20 10  and B =  −20 20 . Case 8: B 0 =  −10 30  and B =  −10 40 . Case 9: B 0 = −30 10 and B = −40 10 . DOMINANCE PHENOMENA 45 We define cluster variables indexed by integers, starting with x1 and x2 and defining the remaining xi so that each pair xi , xi+1 forms a cluster. In particular, each xi is related to xi+2 by the exchange relation 5.1. Since we are assuming that a < 0 and b > 0, we have the following formulas for cluster variables: (5.2) xb1 y2 + 1 x2 −a x y1 + 1 = 0 x1 b x + y2 = −1 x0 −a x + y1 y2−a = −2 x−1 x0 = (5.3) x−1 (5.4) x−2 (5.5) x−3 (5.6) x−4 = (5.7) (5.8) (5.9) xb−3 + y1b y2−ab−1 x−2 y1 + x−a 2 x3 = x1 y b y2 + xb3 x4 = 1 x2 y1−ab−1 y2−a + x−a 4 x5 = x3 In fact, all of these formulas except the formulas for x−4 and x5 are valid under the weaker assumption that a ≤ 0 and b ≥ 0. Thus, since we also have c ≤ 0 and d ≥ 0, those formulas hold for cluster variables x0i with −3 ≤ i ≤ 4, replacing a by c, replacing b by d, and replacing each xj by x0j in the formulas. We will use the notation x∞ (a, b) for the ray theta function of the limiting ray when ab = −4. A simple scattering diagram computation (quoted later as Proposition 5.18) in ScatT (B) yields the following formulas. (5.10) (5.11) (5.12) y1 + y1 y2 x21 + x22 x1 x2 x41 y12 y2 + x22 + 2x2 y1 + y12 x∞ (−1, 4) = x21 x2 2 2 y1 y2 x1 + y1 + 2y1 y2 x1 + x42 x∞ (−4, 1) = x1 x22 x∞ (−2, 2) = (The formula for x∞ (−2, 2) is the subject of [4, Example 3.8], but note that due to the global transpose, we switch the indices 1 and 2 to relate (5.10) to [4, Example 3.8].) The proof of Proposition 5.4 (and thus Theorem 1.18) follows from a sequence of lemmas that describe where νz sends various cluster variables. In each case, the lemma follows from a computation that can be checked by hand or with a computer-algebra system. In every case, z1 = xc−a and z2 = xb−d 0 1 .     Lemma 5.5. If B = a0 0b and B 0 = 0c d0 with a ≤ c ≤ 0 and b ≥ d ≥ 0, then νz (x00 ) = x0 and νz (x0−1 ) = x−1 . 46 NATHAN READING   Lemma 5.5 is enough for Case 1 because for B 0 = 00 00 , the cluster variables in  A• (B 0 ) are x0−1 , x00 , x01 , x02 . Lemma 5.5 also reduces the checking for other cases.     Lemma 5.6. If B = a0 0b and B 0 = a0 d0 with b ≥ d ≥ 0, then νz (x03 ) = x3 .     Lemma 5.7. If B = a0 0b and B 0 = 0c 0b with a ≤ c ≤ 0, then νz (x0−2 ) = x−2 .   When B 0 = −10 10 , the only cluster variable not accounted for in Lemma 5.5 is x0−2 = x03 . Thus Cases 2–3 are handled  0by  Lemmas 5.5, 5.6, and 5.7. 2 The cluster variables for B 0 = −1 accounted for by Lemma 5.5 are  0 not  x0−2 and x03 . In Case 4, where B = −10 30 , the variable x03 is accounted for by Lemma 5.6. The variable x−2 is accounted for by the following lemma, and we are finished with Case 4. h i   Lemma 5.8. If B = −10 d+1 and B 0 = −10 d0 with d ≥ 1, then νz (x0−2 ) = x−3 . 0 In Case 5, the variable x0−2 is accounted for by Lemma 5.7, and we complete the verification of Case 5 by proving the following lemma.     Lemma 5.9. If B = −20 20 and B 0 = −10 20 , then νz (x03 ) = x∞ (2, −2).   Similarly, the cluster variables for B 0 = −20 10 not accounted for by Lemma 5.5  are x0−2 and x03 . In Case 6, B = −30 10 , so x0−2 is accounted for by Lemma 5.7. The following lemma finishes Case 6 by accounting for x03 . 0 1  0 1 0 0 Lemma 5.10. If B = c−1 0 and B = c 0 with c ≤ −1, then νz (x3 ) = x4 . In Case 7, the variable x03 is accounted for by Lemma 5.6, and we complete the case with the following lemma.     Lemma 5.11. If B = −20 20 and B 0 = −20 10 , then νz (x0−2 ) = x∞ (2, −2).     In Case 8, when B 0 = −10 30 and B = −10 40 , the only cluster variables in A• (B 0 ) not accounted for by Lemmas 5.5, 5.6, and 5.8 are x0−3 and x04 . The following two lemmas account for these.     Lemma 5.12. If B = −10 40 and B 0 = −10 30 , then νz (x04 ) = x5 .     Lemma 5.13. If B = −10 40 and B 0 = −10 30 , then νz (x0−3 ) = x∞ (4, −1).     Finally, in Case 9, when B 0 = −30 10 and B = −40 10 , Lemmas 5.5, 5.7 and 5.10 account for all of the cluster variables except x0−3 and x04 . We complete the case with the following two lemmas.     Lemma 5.14. If B = −40 10 and B 0 = −30 10 , then νz (x0−3 ) = x−4 .     Lemma 5.15. If B = −40 10 and B 0 = −30 10 , then νz (x04 ) = x∞ (1, −4). We have completed the proof of Proposition 5.4 and thus of Theorem 1.18. We now turn to Theorem 1.19. This theorem again starts with a 2 × 2 exchange matrix B dominating B 0 , this time with B 0 of finite type. It asserts that νz is an injective, g-vector-preserving ring homomorphism from A• (B 0 ) to A• (B) and that, under certain conditions reproduced below in Proposition 5.16, and assuming the Discreteness Hypothesis, νz sends ray theta functions to ray theta functions. We will see that Theorem 1.19 reduces to the following proposition. DOMINANCE PHENOMENA 47 Proposition 5.16. If B and B 0 are 2×2 exchange matrices such that B dominates B 0 , with B 0 of finite type, then νz takes every cluster variable of A• (B 0 ) to a theta function for B plus an element of A• (B). In fact, variable  νz takes every  cluster  of A• (B 0) to a theta function for B unless B = a0 0b and B 0 = 0c d0 with cd = −3 and 1 6∈ |a|, |b| Suppose we have proven Proposition 5.16. It is not true in general that every ray theta function for B is an element of A• (B), but combining [3, Theorem 1.18], [14, Theorem 0.12], and [14, Proposition 0.14], we see that it is true in this case. Thus in particular νz takes every cluster variable of A• (B 0 ) to an element of A• (B). We apply Proposition 5.1 to see that νz is a ring homomorphism from A• (B 0 ) to A• (B) preserving g-vectors. Again, Proposition 5.3 implies that νz is injective. To prove the assertion about ray theta functions mapping to ray theta functions, first note that, since B 0 is of finite type, all of its ray theta functions are cluster variables. Thus Proposition 5.16 verifies (except in the excluded cases) that νz takes each ray theta function in ScatFanT (B 0 ) to the theta function for the same ray in ScatFanT (B). Since we assume the Discreteness Hypothesis, Theorem 1.16 implies that these theta functions are in fact ray theta functions. We see that Proposition 5.16 implies Theorem 1.19. To prove Proposition 5.16, again  by  the symmetry of transposing both  B  and B 0 , we can continue to take B = a0 0b with a < 0 and b > 0 and B 0 = 0c d0 with c ≤ 0 and d ≥ 0 and cd ≥ −3. We need to check six cases:   Case 1: B 0 =  00 00 . Case 2: B 0 =  −10 10 . Case 3: B 0 =  −10 20 . Case 4: B 0 =  −20 10 . Case 5: B 0 =  −10 30 . Case 6: B 0 = −30 10 . We write ϑ[m1 ,m2 ] for the theta function with respect to B associated to the integer vector [m1 , m2 ]. We write ϑ0[m1 ,m2 ] for the theta function with respect to B 0 . In each of the six cases, the cluster variables x0−1 , x00 , x01 , x02 , are sent (according to Lemma 5.5 and by definition) by νz to x−1 , x0 , x1 , x2 , which in turn are ϑ[±1,0] and ϑ[0,±1] . This observation completes Case 1 and leaves from 1 to 4 remaining cluster variables to check in the other cases. For convenience, we call these remaining cluster variables diagonal cluster variables here. We handle the remaining cases in several lemmas, some of which are slightly more general: Unless explicitly stated, the lemmas do not need the assumption that cd ≥ −3. The lemmas are proved using the following theta function computations, which are [26, Proposition 6.14], [26, Proposition 6.16], and [26, Example 6.18]. As before, the notation [b]+ means max(0, b). Proposition 5.17. If −b ≤ m1 ≤ 0 and m2 ≥ 0, then −m X1 −m1  [−m −ai]+ [m2 +ai]+ 1 ϑ[m1 ,m2 ] = xm y1i x0 2 x2 . 1 i i=0 Proposition 5.18. If m1 < −b and 0 ≤ m2 < −a, then   −m m2  X1 X −m1 − bj m2 i j −m2 −ai m1 +bj 1 m2 ϑ[m1 ,m2 ] = xm x + y1 y2 x0 x1 . 1 2 i j i=1 j=0 48 NATHAN READING Proposition 5.19. If a = −3 and b = 1, then −2 −1 3 2 3 −2 ϑ[−2,3] = x−2 1 x2 + 2y1 x1 + 3y1 y2 x1 + y1 x0 x1 . All theta function computations in the lemmas below use Proposition 5.17 unless otherwise stated. We use νz (x00 ) = x0 (Lemma 5.5) throughout the proofs. Most of the lemmas below have essentially the same proof. The only subtlety in most of the lemmas is to notice that they are not subject to the complications occurring in Lemma 5.24. Thus after the first of these lemmas, we omit the proofs that are b−d essentially the same. Again, in every case, z1 = xc−a 0  and z2 = x1 . 0 1 The only diagonal cluster variable for B 0 = −1 0 has g-vector [−1, 1] and this equals ϑ0[−1,1] . The following lemma thus completes Case 2.     Lemma 5.20. If B = a0 0b and B 0 = 0c d0 with a ≤ c ≤ −1 and b ≥ d ≥ 1, then νz (ϑ0[−1,1] ) = ϑ[−1,1] .  Proof. We compute ϑ0[−1,1] = (x01 )−1 x02 + y10 (x00 )−1−c , so that νz (ϑ0[−1,1] ) equals  x−1 x2 + y1 x−1−a = ϑ[−1,1] .  1 0   When B 0 = −10 20 , there is exactly one diagonal cluster variable not covered by Lemma 5.20, with g-vector [−2, 1]. The following lemma completes Case 3.     Lemma 5.21. If B = a0 0b and B 0 = 0c d0 with a ≤ c ≤ −1 and b ≥ d ≥ 2, then νz (ϑ0[−2,1] ) = ϑ[−2,1] .   When B 0 = −20 10 , the one diagonal cluster variable not covered by Lemma 5.20 has g-vector [−1, 2]. The following lemma completes Case 4.     Lemma 5.22. If B = a0 0b and B 0 = 0c d0 with a ≤ c ≤ −2 and b ≥ d ≥ 1, then νz (ϑ0[−1,2] ) = ϑ[−1,2] .   When B 0 = −10 30 , the diagonal cluster variables have g-vectors [−1, 1], [−2, 1], [−3, 1], and [−3, 2]. The first two of these are covered by Lemmas 5.20 and 5.21, and the other two are covered by the following two lemmas, where we also encounter the first excluded case of Theorem 1.19.     Lemma 5.23. If B = a0 0b and B 0 = 0c d0 with a ≤ c ≤ −1 and b ≥ d ≥ 3, then νz (ϑ0[−3,1] ) = ϑ[−3,1] .     Lemma 5.24. If B = a0 0b and B 0 = −10 d0 with a ≤ −1 and b ≥ d ≥ 3, then ( ϑ[−3,2] if a = −1, or 0 νz (ϑ[−3,2] ) = b−3 ϑ[−3,2] + 3y1 y2 x−2−a x if a ≤ −2. 0 1  Proof. We compute ϑ0[−3,2] = (x01 )−3 (x02 )2 + 3y10 x02 + 3(y10 )2 + (y10 )3 x00 , so that −1−a 2 νz (ϑ0[−3,2] ) = x−3 x2 + 3y12 x0−2−2a + y13 x0−2−3a ). 1 (x2 + 3y1 x0 We also compute ϑ[−3,2] = x−3 1 3   X 3 i [−2−ai]+ [2+ai]+ y x x2 i 1 0 i=0 ( 2 2 3 x−3 1 (x2 + 3y1 x2 + 3y1 + y1 x0 ) = −2−a 2 x−3 + 3y12 x0−2−2a + y13 x0−2−3a ) 1 (x2 + 3y1 x0 if a = −1, or if a ≤ −2. DOMINANCE PHENOMENA 49 We see that if a = −1 then νz (ϑ0[−3,2] ) = ϑ[−3,2] and if a ≤ −2 then νz (ϑ0[−3,2] ) − ϑ[−3,2] is −1−a −2−a b−3 3y1 x−3 x2 − x−2−a ) = 3y1 x−2−a x−3 x1 .  1 (x0 0 0 1 (x0 x2 − 1) = 3y1 y2 x0   01 0 This completes Case 5, leaving only Case 6: B = −3 0 . In this case, the g-vectors of diagonal cluster variables are [−1, 1], [−1, 2], [−1, 3], and [−2, 3]. Lemmas 5.20 and 5.22, together with the following two lemmas, take care of Case 6 and thus complete the proof of Proposition 5.16 and thus Theorem 1.19.     Lemma 5.25. If B = a0 0b and B 0 = 0c d0 with a ≤ c ≤ −3 and b ≥ d ≥ 1, then νz (ϑ0[−1,3] ) = ϑ[−1,3] .     Lemma 5.26. If B = a0 0b and B 0 = −30 10 with a ≤ −3 and b ≥ 1, then ( ϑ[−2,3] if b = 1, or 0 νz (ϑ[−2,3] ) = ϑ[−2,3] + 3y1 y2 x0−3−a xb−2 if b ≥ 2. 1 Proof. Proposition 5.19 says that ϑ0[−2,3] = (x01 )−2 (x02 )3 + 2y10 (x01 )−2 + 3y10 y20 (x01 )−1 + (y10 )2 (x00 )3 (x01 )−2 , so −3−a −2 3 νz (ϑ0[−2,3] ) = x−2 x1 + 3y1 y2 x0−3−a xb−2 + y12 x0−3−2a x−2 1 x2 + 2y1 x0 1 1 . If b = 1, then (since the case where a = −3 is a tautology) we can apply Proposition 5.18 to compute   2 X 3  X 2−j 3 i j −3−ai −2+j −2 3 ϑ[−2,3] = x1 x2 + y y x x1 i j 1 2 0 i=1 j=0 −3−a −2 3 2 −3−2a −2 = x−2 x1 + 3y1 y2 x−3−a x−1 x1 1 x2 + 2y1 x0 0 1 + y1 x0 = νz (ϑ0[−2,3] ). If b ≥ 2, then we can apply Proposition 5.17 to compute −3−a 3 ϑ[−2,3] = x−2 + y12 x0−3−2a ) = νz (ϑ0[−2,3] ) − 3y1 y2 x−3−a x1b−2 .  1 (x2 + 2y1 x0 0 5.3. Acyclic finite type. In this section, we point out a simplification in the description of the map νz in the cases where B is acyclic. (Recall the definition of acyclic from Section 3.4.) We then describe how the simpler description of νz allows for shortcuts in the computations that verify Theorem 1.17 (Phenomenon IV for acyclic finite-type exchange matrices) up to 8 × 8 matrices. The following easy observation is a well-known feature of the acyclic case. Proposition 5.27. Suppose B is an acyclic exchange matrix. Then for each i ∈ {1, . . . , n}, there exists a cluster variable ai with g-vector −ei . For each k ∈ {1, . . . , n}, there exists a cluster containing ai if bik > 0 and containing xi if bik < 0. Here ei is the ith standard unit basis vector in Rn . Recall that the g-vector is the Zn -grading of A• (B) such that the g-vector of xk is ek and the g-vector of yk is the negative of the k th column of B. A cluster monomial is a monomial in the cluster variables in some cluster. Using Proposition 5.27, we prove the following fact, which lets us factor the map νz in the acyclic case. Our notation νz does not explicitly show the dependence on 0 B and B 0 . We temporarily make the notation more explicit by writing νzB ,B . 50 NATHAN READING Proposition 5.28. Suppose B is acyclic and dominates B 0 , which dominates B 00 . 00 0 0 If νzB ,B sends cluster variables of A• (B 00 ) to cluster variables of A• (B 0 ) and νzB ,B sends cluster variables of A• (B 00 ) to cluster variables of A• (B 0 ), then the compoνz νz A• (B) sends cluster variables of A• (B 00 ) to cluster A• (B 0 ) −→ sition A• (B 00 ) −→ 00 variables of A• (B). This composition equals νzB ,B : A• (B 00 ) → A• (B). Proof. The first statement is immediate. Proposition 5.27 implies that the elements 00 0 00 0 ziB ,B in the definition of νzB ,B are cluster monomials, and similarly for the 0 0 0 elements ziB ,B in the definition of νzB ,B . Furthermore, νzB ,B acts on the cluster 00 0 monomials ziB ,B simply by mapping each cluster variable to the corresponding cluster variable.  Suppose B is an acyclic exchange matrix of finite type, and that B dominates B 0 . Theorem 1.17 asserts that νz is an injective, g-vector-preserving ring homomorphism from A• (B 0 ) to A• (B), sending each cluster variable in A• (B 0 ) to a cluster variable in A• (B). To verify the theorem, we first reduce to the case of an irreducible exchange matrix B. (An exchange matrix is reducible if it can be written in block-diagonal form with more than one block.) Furthermore, we will see that it is enough to check the special case where B dominates B 0 but they differ either in exactly one position or they differ in that B has ±1 in a pair of symmetric positions where B 0 has zero. To check the special case, we need only check that νz sends each cluster variable to a cluster variable and then apply Propositions 5.1 and 5.3. If we have checked the special case, then in general, we can find a sequence of exchange matrices, starting at B 0 and ending at B such that each two adjacent matrices in the sequence belong to the special case. Proposition 5.28 ensures that, 0 composing the maps νz at each step in the sequence, we obtain the map νzB ,B , which therefore has the desired properties. We have verified the special case computationally (and thus we deduce the general case) for exchange matrices B that are 8 × 8 or smaller. Recall that acyclic exchange matrices of finite type are exactly those such that Cart(B) is a Cartan matrix of finite type. Since we have proved the theorem up to 8 × 8 exchange matrices, the theorem is proved whenever Cart(B) is of exceptional finite type (E, F, or G). Below, we prove the theorem in types A and D using the surfaces model. Work in progress by Viel [31], already mentioned, is expected to complete the proof of Theorem 1.17 by handling types B and C. 5.4. Resection of surfaces. In this section, we give more background on the surfaces model (building on Section 3.5) and prove Theorem 1.20. A tagged arc is an arc that does not cut out a once-punctured monogon and that has each endpoint incident to a puncture marked (or “tagged”) either notched or plain, with the condition that if both ends of the arc are at the same puncture, they must have the same tagging. Two tagged arcs are compatible if either (1) the corresponding untagged arcs are distinct and compatible and the two arcs have the same tagging at any endpoint they have in common, or (2) both correspond to the same untagged arc, which has two distinct endpoints and is not contained in a component of (S, M) that is a once-punctured monogon, and the taggings of the two tagged arcs disagree at exactly one endpoint. (The two tagged arcs in a once-punctured monogon are not compatible.) A tagged triangulation is a maximal collection of pairwise compatible tagged arcs. Each tagged triangulation DOMINANCE PHENOMENA 51 has the same number of tagged arcs. The operation of reversing all taggings at a given puncture sends a tagged triangulation to a tagged triangulation. By iterating this operation, we can remove all notched taggings except that some punctures are incident to two tagged arcs with opposite taggings at that puncture (compatible as in (2) above). When notches have been thus maximally removed, the resulting tagged triangulation corresponds to an ordinary triangulation as follows: Each arc without notches becomes an ordinary arc. Each arc with a notch (necessarily only on one end) becomes a loop with endpoints at the unnotched end, tracing around the tagged arc. This loop becomes the non-fold edge of a self-folded triangle, as shown in Figure 17. −→ Figure 17. The last step in making an ordinary triangulation from a tagged triangulation Passing from a tagged triangulation to an ordinary triangulation as described above does not change the associated exchange matrix [8, Definition 9.6]. Thus in proving statements about exchange matrices arising from surfaces, we may as well assume that T is an ordinary triangulation, but we pass freely between T and the corresponding tagged triangulation as in Figure 17. The definition of shear coordinates can be extended to define shear coordinates of a quasi-lamination L with respect to a tagged triangulation T . To compute b(T, L), we perform the operation of reversing all taggings at a given puncture until as many notches are removed as possible. For each reversal of taggings, the quasi-lamination L is also altered by reversing the directions of all spirals into that puncture, obtaining a new quasi-lamination L0 . The new tagged triangulation corresponds to an ordinary triangulation T 0 as in Figure 17. The shear coordinate bγ (T, L) for a tagged arc γ in T is defined to be the shear coordinate bγ 0 (T 0 , L0 ), where γ 0 is the arc in T 0 corresponding to γ. In a marked surface (S, M) with tagged triangulation T , the cluster variables in A• (B(T )) are in bijection with tagged arcs. (In fact, when S has no boundary and |M| = 1, the cluster variables are in bijection with the tagged arcs having only plain tags, but we will not need to consider that case in this paper.) We write xγ for the cluster variable associated to a tagged arc γ. The coefficients yi can be identified with the certain allowable curves obtained from the arcs in the triangulation T . Specifically, if γ is a tagged arc in T , then the elementary lamination Lγ associated to γ is a curve that agrees with γ except very close to the endpoints. If γ has an endpoint that is on a boundary component, then as one follows Lγ toward the endpoint, it misses the endpoint by veering slightly to the right. If γ is tagged plain at a puncture, then Lγ again misses to the right and then spirals counterclockwise into the puncture. If γ is tagged notched at a puncture, then Lγ misses to the left and spirals clockwise. We  write LT for the set Lγ : γ ∈ T of elementary laminations associated to T . It is easily verified that, for each tagged arc γ in T , the shear coordinates of Lγ are 1 in the γ position and 0 elsewhere. 52 NATHAN READING Since cluster variables are in bijection with tagged arcs, elements of A• (B(T )) can be represented (non-uniquely) as sums of terms, each term being a monomial in the xα for tagged arcs α (tagged plain when (S, M) is a once-punctured surface with no boundary components) times a monomial in the LT . The tagged arcs are also in bijection with non-closed allowable curves, and we write κ for this bijection. Explicitly, the bijection κ is described as follows: Suppose γ is a tagged arc. As a curve, κ(γ) agrees with γ except very close to the endpoints. If γ has an endpoint that is on a boundary component, then as one follows κ(γ) toward the endpoint, it misses the endpoint by veering slightly to the left. If γ is tagged plain at a puncture, then κ(γ) again misses to the left and then spirals clockwise into the puncture. (Thus the map κ is like the map taking γ to the elementary lamination Lγ , except with right and left reversed.) If γ is tagged notched at a puncture, then κ(γ) misses to the right and spirals counterclockwise. The map κ induces a bijection from tagged triangulations to maximal sets of pairwisecompatible non-closed allowable curves (which are in turn in bijection with maximal cones in the rational quasi-lamination fan FQ (T )). In [23, Proposition 5.2], it is shown that for a tagged arc γ, the g-vector of the cluster variable xγ is −b(T, κ(γ)), the negative of the shear coordinates of κ(γ) with respect to T . Suppose (S, M) is a marked surface with a triangulation T , suppose (S0 , M0 ) is obtained by a resection compatible with T , and suppose T 0 is the triangulation induced by T . Supposing also that every component of (S, M) and (S0 , M0 ) either has the Null Tangle Property or is a null surface, Theorem 1.7 implies that every rational ray of rational quasi-lamination fan FQ (T 0 ) is also a ray of FQ (T ). Since rational rays of the rational quasi-lamination fan are spanned by the shear coordinates of allowable curves, we see that for every allowable curve λ0 in (S0 , M0 ), there is a (unique) allowable curve λ in (S, M) such that b(T, λ) = b(T 0 , λ0 ). Our goal is to prove Theorem 1.20, and accordingly we now restrict our attention to the case where (S, M) is a once-punctured or unpunctured disk and T is a triangulation such that B(T ) is acyclic. (Thus also (S0 , M0 ) is a union of oncepunctured or unpunctured disks and B(T 0 ) is acyclic.) This simplifies the situation considerably, not least because in this case there are no closed allowable curves. In particular, the allowable curves are in bijection with the tagged arcs by the map κ described above. We define a map χ on tagged arcs by letting χ send a tagged arc γ 0 in (S0 , M0 ) to the tagged arc γ in (S, M) such that κ(γ 0 ) and κ(γ) have the same shear coordinates. Equivalently, χ sends γ 0 to the tagged arc γ such that the g-vector of the cluster variable xγ in A• (B(T )) equals the g-vector of xγ 0 in A• (B(T 0 )). Accordingly, we re-use the symbol χ for the map on cluster variables in A• (B(T )) sending x0γ to xχ(γ 0 ) for a tagged arc γ 0 in (S0 , M0 ). Furthermore, we   extend χ to a map on Var• (B(T 0 )) ∪ y0 by letting it agree with νz on y0 . One can check that the requirement that B(T ) is acyclic implies that (1) every triangle of T has at least one edge on the boundary or has two edges inside a oncepunctured digon that has an edge on the boundary and (2) the puncture is contained in a once-punctured digon of T , with one side of the digon being a boundary edge. To prove Theorem 1.20, we do not need to check every possible resection. Instead, since every resection gives a dominance relation and since Theorem 1.20 is an assertion about matrices dominated by B(T ), it is enough to check every matrix dominated by B(T ). Furthermore, by Proposition 5.28, we need only check for DOMINANCE PHENOMENA 53 each pair bαβ = −bβα = ±1 of entries in B(T ), that the theorem holds when B 0 is obtained by setting bαβ and bβα to zero. To accomplish this, it is enough to consider three cases: Case 1. (S0 , M0 ) is obtained by resecting a single arc α ∈ T with two distinct endpoints, both on the boundary, the point pα used to construct the resection is in a triangle of T having exactly one edge on the boundary, the component of (S0 , M0 ) containing the arc α is an unpunctured disk, and the other component is a once-punctured or unpunctured disk. Case 2. (S0 , M0 ) is obtained by resecting an arc α ∈ T that is the non-fold edge of a self-folded triangle in T , the self-folded triangle is contained in a once-punctured digon with exactly one edge on the boundary, the point pα is in the self-folded triangle, the component of (S0 , M0 ) containing α is an unpunctured disk, and the other component is a once-punctured monogon. Case 3. (S0 , M0 ) is obtained by resecting two arcs incident to the puncture in a once-punctured digon of T , with one side of the digon being a boundary edge, one component of (S0 , M0 ) is an unpunctured disk with at least 5 boundary points, and the other is an unpunctured quadrilateral. (The resection must be as pictured in Figure 7, or the mirror image of that figure.) We prove each case separately. In fact, in each case, we prove a proposition that establishes more cases of Phenomenon IV than are covered in Theorem 1.20. Specifically, in each case, we prove a proposition that makes no requirement of acyclicity. The hypotheses on will imply, however, that the resection does not erase any edges (in the sense of Theorem 1.3) that are contained in oriented cycles. In each case, the outline is as follows: (1) Explicitly describe the map χ on tagged arcs. (In every case, χ maps each tagged arc in T 0 to the corresponding tagged arc in T , and we won’t mention these cases separately in each proof.) n o  (2) Explicitly describe the map χ (i.e. νz ) on y0 = L0ζ : ζ ∈ T 0 . Here L0ζ means the elementary lamination in (S0 , M0 ) as opposed to in (S, M). Recall that νz (L0ζ ) = Lζ zζ , where zζ is the cluster monomial whose g-vector is the ζ-column of B(T ) minus the ζ-column of B(T 0 ). (3) Check that χ takes every exchange relation in A• (B(T 0 )) to an exchange relation in A• (B(T )). (4) Conclude by Proposition 5.2 that χ extends to a ring homomorphism.  (Since χ agrees with νz on Var• (B(T 0 )) ∪ y0 , it coincides with νz .) (5) Apply Proposition 5.3 to conclude that νz is injective. We begin with the following proposition, which in particular proves Theorem 1.20 in Case 1. Proposition 5.29. Suppose (S, M) is a once-punctured or unpunctured disk and suppose (S0 , M0 ) and T 0 are obtained from (S, M) and T by a resection compatible with T , along a single arc α with distinct endpoints, both on the boundary. Suppose also that the point pα used to construct the resection is in a triangle of T having exactly one edge on the boundary and that the puncture in (S0 , M0 ) is in the component not containing the arc α. Then both parts of Phenomenon IV occur. 54 NATHAN READING p1 α̃ γ γ̃ β s resect −−−→ α p1 pα β p2 r γ s p1 q1 α q2 p2 q1 p1 ζ′ χ ← − pα ζ p1 χ ← − γ p2 p2 q1 s ζ r q2 p2 r ζ′ r p1 γ pα p2 Figure 18. Illustrations of the map χ on tagged arcs, Case 1 Proof. We use the labels shown in the first row of Figure 18. Some of the arcs in T , and the corresponding arcs in T 0 , are shown and labeled. For generality, neither β nor γ is pictured as a boundary segment, although by hypothesis exactly one of them them is a boundary segment. Given a tagged arc ζ 0 in (S0 , M0 ), we explicitly describe ζ = χ(ζ 0 ). If neither endpoint of ζ 0 is pα or q2 , then ζ is obtained from ζ 0 by the natural inclusion (preserving taggings) taking (S0 , M0 ) minus the triangle p1 p2 pα into (S, M). If an endpoint of ζ 0 is pα , then we delete the part of ζ 0 contained in the triangle p1 p2 pα and include what remains of ζ 0 into (S, M) The included arc is attached to the point r that is closest to p2 (moving from p2 while keeping the surface on the right), as illustrated at the second row of Figure 18. The curves κ(ζ 0 ) and κ(ζ) are shown dotted in the figure. If an endpoint of ζ 0 is at q2 , then there are two cases, depending on whether β or γ is a boundary segment. If γ is a boundary segment, then ζ is the inclusion of ζ 0 . If β is a boundary segment, then the inclusion of ζ 0 is cut where it crosses β, the piece incident to p2 is discarded, and the remaining piece is connected to s, the point that is closest to p1 , moving from p1 keeping the surface on the right. (This case is illustrated in the third row of Figure 18. In the pictures, the curves ζ, χ(ζ), ζ 0 , and χ(ζ 0 ) are shown “dangling” because they might end at the boundary or at the marked point.) n o  We next describe explicitly the map χ (or νz ) on y0 = L0ζ : ζ ∈ T 0 by describing the elements zζ . As before, r ∈ M is the marked point closest to p2 along the boundary and s ∈ M is the marked point closest to p1 along the boundary. If β is a boundary segment, then γ̃ = γ and zα = xγ and zγ = xα̃ . If γ is a boundary segment, then zα = xγ̃ and zβ = xα . We now show that χ takes every exchange relation of A• (B(T 0 )) to an exchange relation in A• (B(T )). We can deal with such an exchange relation without having to DOMINANCE PHENOMENA 55 ν α γ̃ ν χ δ ← − δ̃ pα µ α µ r Figure 19. Arcs in exchange relations, Case 1 pin down whether β or γ is a boundary segment, because in any case, neither L0β nor L0γ is involved in the exchange relation. If neither of the tagged arcs being exchanged is α, then L0α is also not involved. Mapping the exchange relation to (S, M) by χ, we obtain precisely the exchange relation that exchanges the corresponding arcs in (S, M). If one of the arcs being exchanged is α, then the situation is as illustrated in the right picture of Figure 19. The exchange relation is Y Y s0 0 s0 0 x0α x0δ = x0µ (L0ζ )[bα (T ,Lζ )]+ + x0ν (L0ζ )[−bα (T ,Lζ )]+ , ζ∈T 0 ζ∈T 0 where Ts0 is any triangulation of (S0 , M0 ) containing the arcs α and whichever of µ and ν are not boundary segments. (If µ and/or ν is a boundary segment, then we set xµ and/or xν equal to 1.) We have bα (Ts0 , L0α ) = 1. Since the allowable curves in LT 0 are pairwise compatible, no curve ζ ∈ L0T 0 has bα (Ts0 , L0ζ ) < 0. Thus we can rewrite the exchange relation as Y s0 0 x0α x0δ = x0µ L0α (L0ζ )[bα (T ,Lζ )]+ + x0ν . ζ∈T 0 \{α} Applying χ and remembering that L0β and L0γ do not appear in the product, we obtain the formal relation Y [bα (Ts 0 ,L0ζ )]+ xα xδ̃ = xµ xγ̃ Lα Lζ + xν . ζ∈T 0 \{α} Take Ts to be a triangulation of (S, M) agreeing with Ts0 on the inclusion of the right component of (S0 , M0 ) minus the triangle p1 p2 p into (S, M) and containing γ 0 . We see that bα (Ts0 , L0ζ ) = bα (Ts, Lζ ) for every ζ ∈ T 0 , so that the relation becomes Y [b (Ts ,L )] ζ + xα xδ̃ = xµ xγ̃ Lζ α + xν , ζ∈T which is an exchange relation in A• (B(T )), as illustrated in the left picture of Figure 19. To assist in considering further cases, we recast the above treatment pictorially. The top row of Figure 20 shows the exchange relation in (S0 , M0 ) in the case where one of the arcs being exchanged is α, with each element x0ζ represented by the arc ζ. Aside from L0α (shown dotted), the elementary laminations involved in the relation are not shown. The next row of the figure shows the analogous representation of the image of the relation under χ. Now consider an exchange relation in the component of (S0 , M0 ) not containing p. First suppose that β is a boundary segment. If neither tagged arc being exchanged is incident to p2 , then L0γ does not appear in the exchange relation and χ acts by inclusion on the tagged arcs involved in the exchange relation and acts 56 NATHAN READING ν δ α = + L′α µ χ↓ χ↓ χ↓ ν α δ̃ = γ̃ + Lα µ r r r = Lγ′ + r χ↓ χ↓ = Lα α̃ Lβ′ = χ↓ χ↓ + + χ↓ = Lβ′ χ↓ + Figure 20. The action of χ on some exchange relations, Case 1 trivially on each elementary lamination in the exchange relation. The result is an exchange relation in A• (B(T )). If one of the tagged arcs being exchanged is incident to p2 , then L0γ appears in the exchange relation if and only if the other arc being exchanged is incident to r. When the other arc being exchanged is not incident to r, as χ moves the endpoints of arcs at p2 to s, the exchange relation is taken to an DOMINANCE PHENOMENA 57 exchange relation in A• (B(T )). (The exchange relation in A• (B(T )) still does not involve Lγ , and also does not involve Lζ for any arc ζ from the other component of S0 .) When the other arc being exchanged is incident to r, as χ moves the endpoints of arcs at p2 to s, the exchange relation picks up a factor xα̃ in one of its right-side terms, as illustrated in the third and fourth rows of Figure 20. (In (S0 , M0 ) the boundary segment β is positioned relative to the exchange relation exactly where α̃ is positioned relative to the exchange relation in (S, M).) In all of these cases where β is a boundary segment, there are various forms the exchange relations can take, depending on how the exchanged arcs are positioned relative to the puncture and the boundary. (For example, one or more of the arcs appearing on the right side may be a boundary segment, thus disappearing from the relation. Or, one of the arcs on the right side may be replaced by a pair of arcs to the puncture coinciding except for opposite taggings at the puncture. The latter happens when the two arcs being exchanged share an endpoint.) These details are preserved when χ is applied. We see (as illustrated in the fourth row of Figure 20 that χ maps the exchange relation in A• (B(T 0 )) to an exchange relation in A• (B(T )). Next, suppose γ is a boundary segment. If neither tagged arc being exchanged is incident to p2 , then Lβ does not appear in the exchange relation and χ acts by inclusion on the tagged arcs involved in the exchange relation and acts trivially on each elementary lamination in the exchange relation. The result is again an exchange relation in A• (B(T )). If one of the tagged arcs being exchanged is incident to p2 , then Lγ appears in the exchange relation if and only if the other arc being exchanged is incident to p1 . This time, χ does not move the endpoints of arcs at p2 , but again χ takes the exchange relation to an exchange relation in A• (B(T )), as illustrated in the fifth and sixth rows of Figure 20. The rest of the proof follows the outline given above. (We can apply Proposition 5.3 because of the hypothesis that either β or γ is a boundary segment.)  The following proposition proves Theorem 1.20 in Case 2, again relaxing the global requirement of acyclicity. Proposition 5.30. Suppose (S, M) is a once-punctured disk and (S0 , M0 ) is obtained by resecting an arc α ∈ T that is the non-fold edge of a self-folded triangle in T that is in turn contained in a once-punctured digon having exactly one edge on the boundary. Suppose also that the point pα is in the self-folded triangle. Then both parts of Phenomenon IV occur. Proof. We describe χ on tagged arcs. The situation is shown in the first row of Figure 21, where again for generality neither γ nor β is pictured as a boundary arc, although by hypothesis, one of them is. The point labeled r is the closest marked point to p (in that direction) on the boundary. A tagged arc ζ 0 in the component of (S0 , M0 ) containing α is mapped into (S, M) by the natural inclusion unless it has an endpoint at p1 or pα . If ζ 0 has an endpoint at p1 or pα , then the inclusion is extended to p or r or the puncture, tagged plain or notched there. The details depend also on whether β or γ is a boundary segment, as illustrated in the second through fifth rows of Figure 21. (In the fifth row, if β and ζ 0 share an endpoint, then ζ is an arc from r to the puncture, tagged plain.) There are two tagged arcs in the once-punctured monogon. The one that is tagged plain maps into (S, M) by the natural inclusion and remains tagged plain 58 NATHAN READING γ̃ r α δ p1 γ p γ resect α̃ −−−→ δ pα q α β β p2 ζ p γ ζ′ pα q γ ζ p χ ← − γ χ ← − ζ′ q γ ζ′ r χ ← − p pα q ζ β β r β β χ ← − p pα q ζ′ ζ r ζ p χ ← − ζ′ q Figure 21. Illustrations of the map χ, Case 2 at the puncture. The one that is tagged notched maps to an arc from r to the puncture, still tagged notched, as shown in the last row of Figure 21. If β is a boundary segment, then zγ = xδ and zδ = xγ̃ . If γ is a boundary segment, then γ̃ = β, so that zβ = xα̃ and zδ = xγ̃ = xβ . We now show that χ takes every exchange relation of A• (B(T 0 )) to an exchange relation in A• (B(T )). The only exchange relation in the once-punctured monogon is shown in Figure 22. In either case (whether β or γ is a boundary segment), χ sends this exchange relation to an exchange relation as illustrated in Figure 22. DOMINANCE PHENOMENA = 59 + ′ Lδ χ↓ χ↓ r χ↓ r = γ̃ + Lδ Figure 22. The exchange relation in the monogon, Case 2 Lα′ = χ↓ + χ↓ = = χ↓ χ↓ + Lα Lγ′ + χ↓ = χ↓ Lγ + Figure 23. Exchange relations, Case 2 (β a boundary segment) 60 NATHAN READING Lγ′ = χ↓ + χ↓ = χ↓ Lγ = Lα′ Lγ′ χ↓ + + χ↓ = χ↓ Lγ Lα + Figure 24. More exchange relations, Case 2 (β a boundary segment) Now suppose β is a boundary segment and consider an exchange relation in the component containing α. The elementary lamination L0δ does not participate in the exchange relation. There are four cases, based on whether L0α and (independently) L0γ participate in the exchange relation. We observe that L0α participates if and only if an arc incident to pα is exchanged with an arc incident to p1 . Similarly, L0γ participates if and only if exactly two of p1 , pα , and p2 occur as endpoints of the two arcs being exchanged. The cases where neither L0α nor L0γ appears is easy. The remaining cases are illustrated in Figures 23 and 24. Finally, suppose γ is a boundary segment and consider an exchange relation in the component containing α. Again, L0δ does not participate in the exchange relation, and we break into cases based on whether L0α and L0β participate in the exchange relation. Again, L0α participates if and only if an arc incident to pα is exchanged with an arc incident to p1 . We observe that L0γ participates if and only if exactly two of s, p1 , and pα , occur as endpoints of the two arcs being exchanged, where as before s is the marked point closest to p1 . Again, the case where neither L0α nor L0β appears is easy, and the remaining cases are illustrated in Figures 25 and 26. The rest of the proof follows the outline given above.  DOMINANCE PHENOMENA s 61 s Lα′ = χ↓ s + χ↓ s χ↓ s s = + Lα s s Lβ′ = χ↓ s + χ↓ s χ↓ s s = Lβ s + s Lα′ = χ↓ + Lβ′ χ↓ s χ↓ s = s Lβ Lα + Figure 25. Exchange relations, Case 2 (γ a boundary segment) Finally, we prove a proposition that covers Case 3. 62 NATHAN READING s s = s + Lβ′ χ↓ χ↓ s χ↓ s s = + Lβ Figure 26. More exchange relations, Case 2 (γ a boundary segment) s p1 α p2 resect −−−→ γ p1 q1 δ q3 α q2 β p3 pα p2 pβ β γ p3 s ζ χ ← − ζ′ ζ χ ← − ζ′ Figure 27. Illustrations of the map χ on tagged arcs, Case 3a Proposition 5.31. Suppose (S, M) is a once-punctured disk with at least 3 marked points on the boundary with triangulation T . Suppose (S0 , M0 ) is obtained by a resection compatible with T , resecting two arcs incident to the puncture in a oncepunctured digon of T , with one of the edges of the digon being a boundary segment. Then both parts of Phenomenon IV occur. Proof. There are two cases, depending on which of the two arcs in the digon is cut off from the rest of the disk. We will refer to the case where α is cut off as Case 3a and refer to the other case as Case 3b. The situation is illustrated in Figure 27, which also, along with Figures 28 and 29, also illustrates the map χ on tagged arcs. (To see why this resection cuts off a quadrilateral, refer back to Figure 7. Cases 3a DOMINANCE PHENOMENA 63 ζ χ ← − ζ′ χ ← − ζ′ ζ Figure 28. More illustrations of the map χ on tagged arcs, Case 3a p1 ǫ p1 α p2 resect −−−→ γ β p3 q2 q1 β q3 pα α pβ p2 γ p3 ζ χ ← − ζ′ ζ χ ← − ζ′ χ ← − ζ ζ′ Figure 29. Illustrations of the map χ on tagged arcs, Case 3b and 3b correspond to which of the arcs δ or γ in Figure 7 is a boundary segment.) In Case 3a, the element zα is xδ and zγ is xα . In Case 3b, zβ is xγ and zγ is x . In Case 3a, χ takes the unique exchange relation involving arcs in the quadrilateral component to an exchange relation in A• (B(T )), as illustrated in Figure 30. An exchange relation in A• (B(T 0 )) involves L0β if and only if it exchanges an arc 64 NATHAN READING = Lα′ χ↓ + χ↓ s χ↓ s Lα = + = Lβ′ χ↓ + χ↓ = Lβ χ↓ + Figure 30. Exchange relations, Case 3a with endpoint p2 and an arc with endpoint pβ . An exchange relation involves L0γ if and only if exactly two of p1 , p2 , and pβ occur as endpoints of the two arcs being exchanged. In each case, χ maps the exchange relation to an exchange relation, as illustrated in Figure 31. Similarly, in Case 3b, χ maps the unique exchange relation in the quadrilateral to an exchange relation, as illustrated in Figure 32. An exchange relation in A• (B(T 0 )) involves L0α if and only if it exchanges an arc with endpoint p1 and an arc with endpoint pα . An exchange relation involves L0γ if and only if exactly two of p1 , pα , and p2 occur as endpoints of the two arcs being exchanged. Again, χ takes each exchange relation to an exchange relation, as illustrated in Figure 33. Again, the proof concludes as outlined above.  This completes the proof of Theorem 1.20. It appears that there exists a generalization of Theorem 1.20 beyond the surfaces of finite type. In particular, the surfaces of affine type (the unpunctured annulus and the twice-punctured disk) should yield to arguments like those given here. One difficulty with such generalizations is the further proliferation of cases; some way of arguing more generally would help. But there are also some theoretical difficulties that arise because, outside of finite type, we are forced to consider cluster-algebra DOMINANCE PHENOMENA = + Lγ′ χ↓ 65 χ↓ = χ↓ + Lγ = L′γ χ↓ + χ↓ = + Lγ = χ↓ χ↓ Lβ′ L′γ + χ↓ = Lβ Lγ χ↓ + Figure 31. More exchange relations, Case 3a elements associated to closed allowable curves in addition to tagged arcs. First, and most easily dealt with, there is currently no generalization of [23, Proposition 5.2], giving the g-vector of such closed-curve elements in terms of shear coordinates. (Such a generalization should follow easily from [19, Proposition 6.2] in the case where (S, M) has no punctures, and we hope soon to provide a full generalization.) 66 NATHAN READING = + Lβ′ χ↓ χ↓ = χ↓ + Lβ Lα′ = + χ↓ χ↓ χ↓ Lα = + Lα′ = Lγ′ χ↓ + χ↓ = Lα Lγ χ↓ + Figure 32. Exchange relations, Case 3b Second, and more seriously, outside of the finite-type case, the map χ may map a tagged arc to a closed allowable curve. Thus χ cannot map exchange relations to exchange relations in every case, but one would have to show that, instead, χ maps some exchange relations to skein relations [19, 18]. (See [19, Appendix A] for an in-depth discussion of the difficulties associated with skein relations.) DOMINANCE PHENOMENA = χ↓ Lγ′ 67 + χ↓ = χ↓ + Lγ = χ↓ Lγ′ + χ↓ = Lγ χ↓ + Figure 33. More exchange relations, Case 3b Acknowledgments. The author thanks Man Wai “Mandy” Cheung, Salvatore Stella, and Shira Viel for helpful conversations. References [1] I. Assem, G. Dupont, and R. Schiffler, On a category of cluster algebras. J. Pure Appl. Algebra 218 (2014), no. 3, 553–582. [2] E. Barnard, E. Meehan, N. Reading, and S. Viel, Universal geometric coefficients for the four-punctured sphere. To appear in Ann. Comb. [3] A. Berenstein, S. Fomin, A. Zelevinsky, Cluster algebras. III. Upper bounds and double Bruhat cells. Duke Math. J. 126 (2005) no. 1, 1–52. [4] M. W. Cheung, M. Gross, G. Muller, G. Musiker, D. Rupel, S. Stella, and H. Williams, The greedy basis equals the theta basis: a rank two haiku. J. Combin. Theory Ser. A 145 (2017), 150–171. [5] R. Ehrenborg and G.-C. Rota, Apolarity and canonical forms for homogeneous polynomials. European J. Combin. 14 (1993), no. 3, 157–181. [6] A. Felikson, M. Shapiro, and P. Tumarkin, Cluster algebras of finite mutation type via unfoldings. Int. Math. Res. Not. IMRN 2012 no. 8, 1768–1804. [7] A. Felikson, M. Shapiro, and P. Tumarkin, Cluster algebras and triangulated orbifolds. Adv. Math. 231 (2012), no. 5, 2953–3002. [8] S. Fomin, M. Shapiro, and D. Thurston, Cluster algebras and triangulated surfaces. I. Cluster complexes. Acta Math. 201 (2008), no. 1, 83–146. 68 NATHAN READING [9] S. Fomin and D. Thurston, Cluster algebras and triangulated surfaces. Part II: Lambda lengths. Preprint, 2008. (arXiv:1210.5569) [10] S. Fomin and A. Zelevinsky, Cluster algebras. I. Foundations. J. Amer. Math. Soc. 15 (2002), no. 2, 497–529. [11] S. Fomin and A. Zelevinsky, Cluster algebras II: Finite type classification. Invent. Math. 154 (2003), 63–121. [12] S. Fomin and A. Zelevinsky, Cluster Algebras IV: Coefficients. Compos. Math. 143 (2007), 112–164. [13] C. Fraser, Quasi-homomorphisms of cluster algebras. Adv. in Appl. Math. 81 (2016), 4077. [14] M. Gross, P. Hacking, S. Keel, and M. Kontsevich, Canonical bases for cluster algebras. Preprint, 2014. (arXiv:1411.1394) [15] Min Huang, Fang Li, and Yichao Yang On Structure of cluster algebras of geometric type I: In view of sub-seeds and seed homomorphisms. Preprint, 2015. arXiv:1509.01050 [16] Min Huang and Fang Li, On structure of cluster algebras of geometric type, II: Green’s equivalences and paunched surfaces. Pure Appl. Math. Q. 11 (2015), no. 3, 451–490. [17] K. Igusa, K. Orr, G. Todorov, and J. Weyman, Picture groups of finite type and cohomology in type An . Preprint, 2016. (arXiv:1609.02636) [18] G. Musiker and L. Williams, Matrix formulae and skein relations for cluster algebras from surfaces. Int. Math. Res. Not. IMRN 2013, no. 13, 2891–2944. [19] G. Musiker, R. Schiffler, and L. Williams, Bases for cluster algebras from surfaces. Compos. Math. 149, (2013) 217–263. [20] N. Reading, Cambrian lattices. Adv. Math. 205 (2006) no. 2, 313–353. [21] N. Reading, Clusters, Coxeter-sortable elements and noncrossing partitions. Trans. Amer. Math. Soc. 359 (2007), no. 12, 5931–5958. [22] N. Reading, Universal geometric cluster algebras. Math. Z. 277 no. 1–2 (2014), 499–547. [23] N. Reading, Universal geometric cluster algebras from surfaces. Trans. Amer. Math. Soc. 366 (2014), 6647–6685. [24] N. Reading, Universal geometric coefficients for the once-punctured torus. Sém. Lothar. Combin. B71e (2015), 30 pp. [25] N. Reading, Lattice homomorphisms between weak orders. Preprint, 2017. [26] N. Reading, Scattering diagrams and scattering fans. Preprint, 2017. [27] N. Reading and D. E. Speyer, Cambrian fans. J. Eur. Math. Soc. (JEMS), 11 no. 2, 407-447. [28] N. Reading and D. E. Speyer, Sortable elements in infinite Coxeter groups Trans. Amer. Math. Soc. 363 (2011) no. 2, 699-761. [29] N. Reading and D. Speyer, Combinatorial frameworks for cluster algebras. Int. Math. Res. Notices. 2016 no. 1 (2016), 109–173. [30] R. Simion, Noncrossing partitions, Formal power series and algebraic combinatorics (Vienna, 1997). Discrete Math. 217 (2000), no. 1-3, 367–409. [31] S. Viel, Work in progress, 2017. [32] S. Yang and A. Zelevinsky Cluster algebras of finite type via Coxeter elements and principal minors, Transformation Groups 13 (2008), no. 3–4, 855–895.
0math.AC
RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 arXiv:1601.02911v2 [math.AG] 26 Apr 2016 GIANFRANCO CASNATI This paper is dedicated to the memory of A.T. Lascu. Abstract. Let F ⊆ P3 be a smooth determinantal quartic surface which is general in the Nöther–Lefschetz sense. In the present paper we give a complete classification of locally free sheaves E of rank 2 on F such that h1 F, E(th) = 0 for t ∈ Z. 1. Introduction and Notation Throughout the whole paper we will work over an algebraically closed field k of characteristic 0 and we will denote by P3 the projective space of dimension 3 over k. If F ⊆ P3 is a smooth surface we set OF (h) := OP3 (1) ⊗ OF . A vector bundle E on F is called aCM (i.e. arithmetically Cohen–Macaulay) if h1 F, E(th) = 0 for t ∈ Z. When deg(F ) ≤ 2, the well–known theorems of Horrocks and Knörrer (see [21], [1] and the references therein) provide a complete classification of aCM bundles on F . As deg(F ) increases a complete picture is missing. It is thus interesting to better understand aCM bundles on surfaces with deg(F ) ≥ 3. In order to simplify our study we restrict our attention to bundles satisfying some further technical non–restrictive conditions. First, notice that the property of being aCM is invariant up to tensorizing by OF (th), thus we  can consider in our  study only initialized bundles, i.e. bundles E such that h0 F, E(−h) = 0 and h0 F, E 6= 0. Moreover, we can obviously focus onto indecomposable bundles, i.e. bundles which do not split in a direct sum of bundles of lower rank. When deg(F ) = 3 some scattered results are known. E.g., M. Casanellas and R. Hartshorne proved that each smooth cubic surface supports families of arbitrary dimension of non–isomorphic, indecomposable initialized aCM bundles (see [4] and [5]). Moreover, D. Faenzi classified completely aCM bundles of rank 1 and 2 on such an F in [12]. A key point of all the quoted papers is the complete knowledge of Pic(F ). The case deg(F ) ≥ 4 is wide open. In this case the structure of Pic(F ) is not uniquely determined, thus it could be very difficult to classify initialized indecomposable aCM bundles E also in low ranks, unless assuming true some extra hypothesis on F . E.g. if Pic(F ) is generated by OF (h) and E has rank 2 on F with c1 (E) ∼ = OF (ch), then 3 − deg(F ) ≤ c ≤ deg(F ) − 1, as essentially proved by C. Madonna (see [19]: see also [6] where some slight improvement is explained). We are mostly interested in the case deg(F ) = 4. In this case F is a K3 surface, thus a lage amount of informations is available on Pic(F ). E.g. K. Watanabe in [24] classified initialized aCM bundles of rank 1 on such an F . 2010 Mathematics Subject Classification. Primary 14J60; Secondary 14J45. Key words and phrases. Vector bundle, Cohomology. The author is member of the GNSAGA group of INdAM. 1 2 GIANFRANCO CASNATI In the beautiful paper [9] E. Coskun, R.S. Kulkarni, Y. Mustopa dealt with the case of rank 2 initialized indecomposable bundles E on F with c1 = 3h. In [8] the authors gave a complete description of bundles with c1 = 2h: it is perhaps interesting to notice that in this case there is a large family of quartic surfaces, the determinantal ones (i.e. surfaces whose equation can be expressed as the determinant of a 4 × 4 matrix with linear entries), supporting bundles whose non–zero sections do not vanish in the expected dimension. The present paper is motivated by our interest in the classification of rank 2 initialized indecomposable bundles E on a determinantal quartic surface F ⊆ P3 . In this case F contains a projectively normal smooth sextic curve A of genus 3. It is thus obvious that Pic(F ) must contain the rank 2 lattice Λ generated by h and A. Moreover, we actually have Pic(F ) = Λ if F is general in the sense of Nöther–Lefschetz, i.e. it represents a point of |OP3 (4)| outside a countable union of suitable proper subvarieties, due to some classic results proved in [18]. Each F satisfying the above property will be called a general determinantal quartic surface. Theorems 6.2 and 6.5 give the complete classification of the bundles we are interested in on such an F . For reader’s benefit we summarize such a classification in the following simplified statement. For the complete statement see Theorems 6.2 and 6.5. Main Theorem. Let E be an initialized indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F ⊆ P3 . Then one of the following asseritions holds. (i) c1 ∈ { −h, A−h, 2h−A }: then c2 = 1. The zero locus E of a general s ∈ H 0 F, E) is a point. (ii) c1 = 0: then c2 = 2. The zero locus E ⊆ P3 of a general s ∈ H 0 F, E) is a 0–dimensional subscheme contained in exactly one line. (iii) c1 = h: then c2 is 3, 4, or 5. The zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional subscheme of degree c2 of a linear subspace of P3 of dimension c2 − 2. (iv) c1 ∈ { A, 3h − A }: then c2 is 3, or 4. In this case the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional subscheme. (v) c1 ∈ { A + h, 4A − h }: then c2 = 8. In this case the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional subscheme which is contained in exactly two linearly independent quadrics. (vi) c1 = 2h: then c2 = 8. In this case one of the two following cases occur: (a) the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional subscheme which is the base locus of a net of quadrics; (b) the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional subscheme lying on exactly a twisted cubic; (c) the zero locus E of each non–zero s ∈ H 0 F, E) is a divisor in either |A|, or |3h − A|. (vii) c1 = 3h: then c2 = 14. In this case the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional subscheme. Moreover, all the above cases actually occur on F . We now explain briefly the structure of the paper. In Section 2 we recall all the facts on a smooth quartic surface that we will need in the paper. In Section 3 we specialize to the case of determinantal quartic surfaces. In Section 4 we list a series of examples. In Section 5 we deal with the zero–locus of the sections of initialized indecomposable aCM bundles of rank 2 on F , proving that it is 0–dimensional, but the two cases which we mentioned RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 3 before. Finally, in Section 6 we prove the above Main Theorem, showing that the list of examples in Section 4. The author would like to thank F. Galluzzi for several helpful discussions and the referee, whose comments helped to considerably improve the exposition of the paper. For all the notations and unproven results we refer to [13]. 2. Preliminary results In this section we will recall some preliminary facts about vector bundles on a smooth quartic surface F ⊆ P3 , most of them collected from several papers (see [10], [22], [24] and [14]). By adjunction in P3 we have ωF ∼ = OF , i.e. each such an F is a K3 surface, thus we can make use of all the results proved in [22]. The first important fact is that Serre duality for each locally free sheaf F on F becomes   hi F, F = h2−i F, F ∨ , i = 0, 1, 2. Moreover Riemann–Roch theorem on F is    c1 (F)2 − c2 (F). (1) h0 F, F + h2 F, F = h1 F, F + 2rk(F) + 2 In particular, if D is a divisor with D2 ≥ −2, then either effective.  D 0or −D is necessarily  2 If D is an effective divisor on F , then h F, OF (D) = h F, OF (−D) = 0. Moreover    h1 F, OF (D) = h1 F, OF (−D) = h0 D, OD − 1 (see [22], Lemma 2.2). It follows that for each irreducible divisor D on F , the dimension of |D| and the arithmetic genus of D satisfy (2)  D2 , h0 F, OF (D) = 2 + 2 deg(D) = Dh, pa (D) = 1 + D2 , 2 (see [22], Paragraph 2.4). In particular, the unique irreducible and reduced fixed divisors D on F satisfy D2 = −2, thus they are smooth rational curves. We summarize the other helpful results we will need in the following statement. Proposition 2.1. Let F ⊆ P3 be a smooth quartic surface. For each effective divisor D on F such that |D| has no fixed components the following assertions hold. (i) D2 ≥ 0 and OF (D) is globally generated. (ii) If D2 > 0, then  the general element of |D| is irreducible and smooth: in this case h1 F, OF (D) = 0. (iii) If D2 = 0, then there is an irreducible curve D with pa (D) = 1 such that OF (D) ∼ = OF (eD) where e − 1 := h1 F, OF (D) : in this case the general element of |D| is smooth. Proof. See [22], Proposition 2.6 and Corollary 3.2.  Now we turn our attention to aCM line bundles on F recalling a very helpful result from [24]. A first obvious fact is that for each divisor D on F , then OF (D) is aCM if and only if the same is true for OF (−D). The main result from [24] is the following statement. 4 GIANFRANCO CASNATI Proposition 2.2. Let F ⊆ P3 be a smooth quartic surface. For each effective divisor D on F with D 6= 0, then OF (D) is aCM if and only if one of the following cases occurs. (i) D2 = −2 and 1 ≤ hD ≤ 3. (ii) D2 = 0 and 3 ≤ hD ≤ 4. (iii) D2 = 2 and hD = 5.   (iv) D2 = 4, hD = 6 and h0 F, OF (D − h) = h0 F, OF (2h − D) = 0. Proof. See [24], Theorem 1.1.  Notice that the general smooth quartic surface F ⊆ P3 does not support initialized aCM line bundles besides OF . Indeed, in this case, Pic(F ) is generated by OF (h), due to Nöther–Lefschetz theorem. In particular, for each line bundle OF (D) on F one has that both Dh and D2 are positive multiples of 4.  Recall that a curve D ⊆ P3 is called aCM if h1 P3 , ID|P3 (th) = 0 for each t ∈ Z (see [20], Lemma 1.2.3). A curve D ⊆ P3 is projectively normal if and only if it is aCM and smooth. We recall the following helpful fact. Lemma 2.3. Let F ⊆ P3 be a smooth quartic surface. For each effective divisor D on F , then OF (D) is aCM if and only if the curve D is aCM in P3 . Proof. We have the exact sequence 0 −→ IF |P3 −→ ID|P3 −→ ID|F −→ 0.  Since IF |P3 ∼ ID|F ∼ OF (−D), it follows that h1 P3 , ID|P3 (t) = 0 if and = OP3 (−4) and =   only if h1 F, OF (D − th) = h1 F, ID|F (th) = 0.  We conclude this section by recalling some further facts facts on vector bundles on a smooth quartic surface F ⊆ P3 .LLet F be a vector  bundle on F : we can consider it as a sheaf over P3 , thus H∗0 (F) := t∈Z H 0 F, F(th) is also a module over the polynomial ring S := k[x0 , x1 , x2 , x3 ]. For the following fact see [4], Theorem 3.1. Lemma 2.4. Let F be an initialized aCM bundle of rank r on a smooth quartic surface F ⊆ P3 . The minimal number of generators of H∗0 (F) as a module over S is at most 4r. Definition 2.5. Let F be an initialized aCM bundle of rank r on a smooth quartic surface F ⊆ P3 .  We say that F is Ulrich if it is initialized, aCM and h0 F, F = 4r. Remark 2.6. If F is Ulrich, then it is globally generated by Lemma 2.4. More generally, Ulrich bundles on F of rank r are exactly the bundles F on F such that H∗0 (F) has linear minimal free resolution over S (see [4], Proposition 3.7). Sheafifying such a resolution we thus obtain an exact sequence 0 −→ OP3 (−1)⊕4r −→ OP⊕4r −→ F −→ 0. 3 3. The general smooth determinantal quartic surfaces In this section we will summarize some facts about the smooth determinantal quartic surfaces in P3 . We will denote by f the polynomial defining F . RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 5 Proposition 3.1. Let F ⊆ P3 be a smooth quartic surface. The following assertions are equivalent. (i) The polynomial f defining F is the determinant of a 4 × 4 matrix Φ with linear entries. (ii) There exists a line bundle OF (A) such that the sheafified minimal free resolution of H∗0 (OF (A)) has the form ϕ 0 −→ OP3 (−1)⊕4 −→ OP⊕4 3 −→ OF (A) −→ 0. (3) (iii) There exists an initialized aCM line bundle OF (A) with A2 = 4. (iv) There exists a smooth integral sextic curve A ⊆ F of genus 3 which is not hyperelliptic. (v) There exists a smooth integral sextic curve A ⊆ F of genus 3 which is projectively normal in P3 . (vi) There exists a smooth integral sextic curve A ⊆ F of genus 3 such that   h0 F, OF (2h − A) = h0 F, OF (A − h) = 0. Proof. Assertion (i) is equivalent to assertions (ii), (iv), (v) due to Corollaries 1.8, 6.6 and Proposition 6.2 of [2]. Assertions (iii) and (vi) are equivalent thanks to Proposition 2.2. Assertion (ii) implies (iii). Indeed, we have A2 = χ(OF (A)) = 4(χ(OP3 ) − χ(OP3 (−1))) = 4, 2  thus A2 = 4. Conversely, h0 F, OF (A) = 4 for each initialized aCM line bundle with A2 = 4. Hence, OF (A) is Ulrich, thus it fits into Sequence (3) (see e.g. Proposition 3.7 of [4]).  2+ Remark 3.2. Let F be determinantal, so that we have Sequence (3). Its dual is tϕ 0 −→ OP3 (−1)⊕4 −→ OP⊕4 3 −→ OF (3h − A) −→ 0,  because Ex t1P3 OF (A), OP3 ∼ = OF (4h − A). Thus also the general element in |3h − A| is a curve enjoying the same properties as A. Notice that OF (3h − A) 6∼ = OF (A), because (3h − A)A = 14 6= 4 = A2 . (4) From the above results we deduce that if the polynomial f defining the smooth quartic surface F is the determinant of a 4 × 4 matrix with linear entries, then Pic(F ) contains the even sublattice Λ generated by h and A such that h2 = A2 = 4 and hA = 6. By definition the general element in |h| is an irreducible smooth plane curve of degree 4. Thanks to Proposition 3.1 the general element in |A| is an integral smooth projectively normal curve of degree 6 and genus 3(see Equalities (2)). Conversely, let A ⊆ P3 be an integral projectively normal sextic of genus 3. We know that each smooth quartic surface F ⊆ P3 containing A is necessarily determinantal.  The minimal degree of a surface containing such an A is 3, because h0 F, OF (2h − A) = 0. Moreover OF (4h − A) is globally generated (see e.g. Sequence (4)), thus the general element in |OF (4h − A)| is integral. It follows that Pic(F ) = Λ due to [18], Theorem II.3.1, if F is general in the Nöther–Lefschetz sense. From now on, we will assume that the surface F has the above property and we will call it a general determinantal quartic surface. When we will consider such a surface F , we will always denote by OF (A) a line bundle corresponding to a projectively normal sextic of genus 3 on F . Notice that there is an 6 GIANFRANCO CASNATI ambiguity because, as we pointed out above, the line bundle OF (3h − A) has the same properties as OF (A). In order to complete the picture, we recall that not all determinantal quartic surfaces F ⊆ P3 satisfy rk(Pic(F )) = 2. E.g. a smooth quartic surface F ⊆ P3 containing a twisted cubic curve T and a line E without common points is determinantal (see [11], Exercise 4.14). Indeed the union of T with any plane cubic linked to E define a linear system whose general element A is a projectively normal sextic of genus 3. It is easy to check that h, T , E are linearly independent in Pic(F ), thus rk(Pic(F )) ≥ 3 in this case. In [7] the authors construct many examples of determinantal quartic surfaces F ⊆ P3 with rk(Pic(F )) = 3 and show that even the classification of indecomposable Ulrich bundles of rank 2 on them seems to be highly non–trivial. Remark 3.3. A first interesting fact is that if D ∈ |xh + yA|, then (5) D2 = 4x2 + 12xy + 4y 2 , Dh = 4x + 6y, DA = 6x + 4y. More precisely it is very easy to check the following assertions: (i) D2 ≡ 0 (mod 16) if and only if both x and y are even; (ii) D2 ≡ 4 (mod 8) if and only if x and y are not simultaneously even. The two above assertions imply that D2 is always a multiple of 4. Thus Equality (1) imply that χ(OF (D)) is even. Moreover, Equalities (5) imply that the equation D2 = 0 has no non–trivial integral solutions, thus: (iii) D2 = 0 if and only if D = 0. Irreducible and reduced fixed divisors D on F satisfy D2 = −2 (see the last Equality (2)), which is not a multiple of 4. It follows that there are no fixed divisors on F . In particular DU ≥ 0 for each divisor U ⊆ F , hence (see Equalities (5)): (iv) DU ≥ 0 is even for each divisor U ⊆ F . Combining the above assertions with Proposition 2.1, we deduce the following consequence: (v) each element in |h|, |A|, |3h − A| is certainly integral (but maybe singular). Looking at Remark 3.3 we know that the following result holds. Lemma 3.4. Let F ⊆ P3 be a general determinantal quartic surface. If the divisor D on F is effective, then |D| has no fixed components.  In particular for each effective divisor D on F we have D2 ≥ 0 and h0 F, OF (D) ≥ 2. Moreover, if D 6= 0, then from Remark 3.3 we know that D2 6= 0, whence D2 ≥ 4 and h0 F, OF (D) ≥ 4. Finally, the ampleness of h implies Dh ≥ 2. The above discussion partially proves the following result. Lemma 3.5. Let F ⊆ P3 be a general determinantal quartic surface. The divisor D on F is effective if and only if either D = 0, or Dh ≥ 2 and D2 ≥ 4. Proof. The ‘only if’ part was proved above. Assume that Dh ≥ 2 and D2 ≥ 4. Either D or −D is effective, because D2 ≥ −2. We have (−D)h = −2, thus −D is not effective. we conclude that D is effective.  Remark 3.6. Thus D ∈ |xh + yA| is effective if and only if either (x, y) = (0, 0), or x2 + 3xy + y 2 ≥ 1, 2x + 3y ≥ 1. RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 7 The above inequalities give the closed connected subset of the affine plane with coordinates x and y not containing the point (x, y) = (0, 0) and whose border is the graph of the function √ −3x + 5x2 + 4 y= 2 2 (the ‘upper branch’ of the hyperbola x + 3xy + y 2 = 1). Proposition 2.1 and the above Lemmas 3.4 and 3.5 prove the following corollary. Corollary 3.7. Let F ⊆ P3 be a general determinantal quartic surface. The divisor D on F is effective if and only if OF (D) is globally generated. 4. Examples of rank 2 aCM bundles In this section we briefly recall Hartshorne–Serre correspondence, listing some examples of rank 2 aCM bundles on general determinantal quartic surfaces. Let E be a rank 2 vector bundle on a smooth quartic surface F ⊆ P3 with Chern classes ci := ci (E), i = 1, 2. If s ∈ H 0 F, E is non–zero, then its zero–locus E := (s)0 ⊆ F has codimension at most 2. We can write (s)0 = E ∪ D where E has pure codimension 2 (orit is empty) and D is a possibly zero divisor. We thus obtain a section σ ∈ H 0 F, E(−D) vanishing exactly on E. The Koszul complex of σ gives the following exact sequence (6) 0 −→ OF (D) −→ E −→ IE|F (c1 − D) −→ 0. The degree of E is c2 (E(−D)). Bertini’s theorem for the sections of a vector bundle guarantees that  if E is globally generated (e.g. if E is Ulrich), then D = 0 for the general s ∈ H 0 F, E .  Assume that the divisorial part D of (s)0 is not empty for general s ∈ H 0 F, E . We can define a rational map H 0 F, E) 99K Pic(F ) defined by s 7→ OF (D). Since Pic(F ) is discrete, it follows that such a map is constant. Thus the divisorial part of the zero–loci of the sections in a suitable non–empty subset of H 0 F, E) are linearly equivalent. The above construction can be reversed. To this purpose we give the following well– known definition. Definition 4.1. Let F ⊆ P3 be a smooth quartic surface and let G be a coherent sheaf on F . We say that a locally complete intersection subscheme E ⊆ F of dimension zero is 0 Cayley–Bacharach (CB for short) with respect to G if, for each  E ⊆ E of degree deg(E)−1, 0 0 the natural morphism H F, IE|F ⊗ G → H F, IE 0 |F ⊗ G is an isomorphism. For the following result see Theorem 5.1.1 in [15]. Theorem 4.2. Let F ⊆ P3 be a smooth quartic surface, E ⊆ F a locally complete intersection subscheme of dimension 0. Then there exists a vector bundle E of rank 2 on F with det(E) = L and having a section s such that E = (s)0 if and only if E is CB with respect to L.  If such an E exists, then it is unique up to isomorphism if h1 F, L∨ = 0. Let F ⊆ P3 be a smooth quartic surface. We recall that in [9] the author proved the existence on F of aCM bundles E with c1 (E) = 3h and c2 (E) = 14. To this purpose the authors proved the existence of suitable sets of 14 points which are CB with respect to OF (3h). 8 GIANFRANCO CASNATI Similarly, in [8] the author proved the existence on F of aCM bundles E with c1 (E) = 2h and c2 (E) = 8. Essentially two cases are possible. In the first case, the zero–locus of a general section of E has pure dimension 0: this case occurs without further hypothesis on F . In the second case, the zero–loci of non–zero sections of E define a linear system |D| 0 whose general element  is a smooth integral sextic of genus 3 such that h F, OF (D −h) = 0 h F, OF (2h − D) = 0. In particular F must be a determinantal quartic surface in this second case. Due to the results listed in the previous section we know that this case can occur only if F is determinantal. In particular it occurs when F is a general determinantal quartic surface F . We now construct examples of rank 2 aCM bundles starting from suitable sets of points E ⊆ F . Recall that for each closed subscheme X ⊆ F we have an exact sequence (7) 0 −→ IX|F −→ OF −→ OX −→ 0. In what follows we will present a series of examples of initialized indecomposable aCM rank 2 bundles on a quartic surface F ⊆ P3 with plane section h. We stress that the bundles defined in Examples 4.3 and 4.4 exist on each smooth quartic surface. Example 4.3. Let E ⊆ F be a 0–dimensional subscheme arising as follows: • m = 3 E is the complete intersection of a line and a cubic surface; • m = 4 E is the complete intersection of two conics inside a plane; • m = 5 E is the degeneracy locus of a 5 × 5 skew symmetric matrix ∆ with linear entries. It is clear that E is CB with respect to OF (h) in the first two cases. In the third one, a minimal free resolution of H∗0 (OE ) over P3 is given by the Buchsbaum–Eisenbud complex (see [3]). Thus again E is CB with respect to OF (h) thanks to Lemma (4.2) of [23]. Theorem 4.2 thus yields the existence of a rank 2 bundle E fitting into Sequence (6) with c1 = h, c2 = m and D = 0, i.e. (8) 0 −→ OF −→ E −→ IE|F (h) −→ 0, The cohomology of the above sequence twisted by OF (−h) gives    h2 F, E = h0 F, E ∨ = h0 F, E(−h) = 0, i.e. E is initialized and the cohomology of Sequence (8) yields the exact sequence.    0 −→ H 1 F, E −→ H 1 F, IE|F (h) −→ H 2 F, OF −→ 0.   By definition h0 F, IE|F (h) = 5 − m, whence h1 F, IE|F (h) = 1, from the cohomology  of Sequence (7). We conclude that h1 F, E = 0. The cohomology of Sequence (8) twisted by OF (th) gives   h1 F, E(th) = h1 F, IE|F ((t + 1)h) , t ≥ 1. In the cases m = 3, 4 the minimal free resolution of H∗0 (OE ) over P3 is given by a Koszul complex. In the case m = 5, it is given by the Buchsbaum–Eisenbud complex as already pointed out above. In any case it is easy to checkthat the Hilbert function of E is  (1, m − 1, m, . . . ), hence the natural map H 0 F, OF (th) → H 0 E, OE (th) is surjective RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 9  for t ≥ 1, thus h1 F, IE|F (th) = 0, in the same range (see the cohomology of Sequence  (7)). We conclude that h1 F, E(th) = 0 for t ≥ 0. By duality    h1 F, E(th) = h1 F, E ∨ (−th) = h1 F, E((−t − 1)h) = 0, −t − 1 ≥ 0, i.e. t ≤ −1. We conclude that E is aCM. If E ∼ = OF (U ) ⊕ OF (V ), then U + V = h and c2 = U V which is even (see Remark 3.3). Moreover both OF (U ) and OF (V ) are aCM and at least one of them must be initialized, the other being either initialized or without sections. Let OF (U ) be initialized. If U = A, then V = h − A, thus U V = 2 < 3 ≤ c2 . The case U = 3h − A can be excluded in the same way. If U = 0, then V = h, a contradiction. We conclude that E is indecomposable. Example 4.4. Let E ⊆ F be a set of m = 1, 2 points. Then E is trivially CB with respect to OF ((m − 2)h), thus there is a rank 2 bundle E fitting into Sequence (6) with c1 = (m − 2)h, c2 = m and D = 0. As in the Example 4.3 one easily checks that such an E is aCM, initialized and indecomposable. For all the following examples we obviously need to restrict to a determinantal quartic surface F ⊆ P3 . As usual h denotes the class of the plane section and A the class of a projectively normal integral sextic curve of genus 3 on F . Example 4.5. Let E ⊆ F be a single point. The line bundle OF (A − h) is not effective, thus E is again trivially CB with respect to OF (A − h). In particular there is a rank 2 bundle E fitting into an exact sequence of the form (9) 0 −→ OF −→ E −→ IE|F (A − h) −→ 0,  The cohomology of Sequence (9) twisted by OF (−h) gives h0 F, E(−h) = 0, because OF (A − 2h) is not effective. Thus E is initialized. The same  sequence twisted by OF (th) also yields the existence of an injective map H 1 F, E(th) → H 1 F, IE|F (A + (t − 1)h) for each t. The cohomology of Sequence (7) twisted by OF (A + (t − 1)h) yields h1 F, IE|F (A + (t −   1)h) = 0 for t ≥ 1. Hence h1 F, E(th) = 0, for t ≥ 1.  The cohomology of Sequence (7) twisted by OF (−th) yields h1 F, IE|F (−th) = 0 for t ≤ 0. Sequence (9) yields the existene of surjective maps   H 1 F, OF ((1 − t)h − A) → H 1 F, E ∨ (−th) , t ≤ 0.   Since OF (A) is aCM, the same is true for OF (−A), thus h1 F, E(th) = h1 F, E ∨ (−th) = 0 for t ≤ 0. We conclude that E is aCM. If E ∼ = OF (U ) ⊕ OF (V ), then c2 = U V which is even (see Remark 3.3). Since c2 = 1, it follows that E is indecomposable. Finally c1 = A − h by construction. Substituting A with 3h − A in the above construction we are also able to construct an example of initialized indecomposable aCM bundle of rank 2 with c1 = 2h − A. We can construct two more bundles. Indeed, E ∨ (h) is indecomposable and aCM too. Moreover the cohomology of Sequence (9) suitably twisted also implies that E ∨ (h) is initialized. Computing the Chern classes of E ∨ (h) we obtain c1 = 3h − A and c2 = 3. Again the substitution of A with 3h − A gives an example of initialized indecomposable aCM bundle of rank 2 with c1 = A and c2 = 3. 10 GIANFRANCO CASNATI Example 4.6. The curve A ⊆ F has genus 3 and it is not hyperelliptic (see Proposition 3.1), thus it supports infinitely many complete g41 . Fix a smooth divisor E in one of these ⊕2 g41 . The natural composite map OF⊕2 → OA → OA (E) is surjective on global sections and its kernel is a vector bundle F on F . The bundle E := F ∨ is called the Lazarsfeld–Mukai bundle of OA (E). By definition we have the exact sequence (10) 0 −→ E ∨ −→ OF⊕2 −→ OA (E) −→ 0. Clearly c1 = A and c2 = 4. In what follows we will show that E is aCM.  quickly  1 1 ∨ The cohomology Sequence (10) yields h F, E = h F, E = 0 by definition. Twisting Sequence (10) by OF (−th), taking its cohomology we also obtain h1 F, E(−th) = 0 for  t ≥ 1, because h0 A, OA (E−th) = 0, due to the inequality deg(OA (E−th)) = 4−6t ≤ −2 in the same range.  We have Ex t1F OA (E), OF ∼ = = ωA (−E). Adjunction formula on F thus implies ωA ∼ OA (A), hence the dual of Sequence (10) is (11) 0 −→ OF⊕2 −→ E −→ OA (A − E) −→ 0. Thus, also for t ≥ 1, we obtain    h1 F, E(th) ≤ h1 A, OA (A − E + th) = h0 A, OA (E − th) = 0. We conclude that E is aCM. The  cohomology of Sequence (11) twisted by OF (−h) and the vanishing h0 A, OA (E − h) = 0 (see above) also imply that E is initialized. Assume that E ∼ = OF (U ) ⊕ OF (V ). We thus have U + V = A and U V = 4. As in Example 4.3 we can assume that OF (U ) is aCM and initialized and that OF (V ) is either initialized or without sections. If U = A, then V = 0, whence U V = 0 a contradiction. Similarly U 6= 0. If U = 3h − A, then V = 2A − 3h and again U V = 0. The same argument with 3h − A instead of A yields the existence of initialized indecomposable aCM bundles of rank 2 with c1 = 3h − A and c2 = 4. Also in this case we can construct two more bundles. Indeed E ∨ (h)  is indecomposable  and aCM too. Moreover, Equality (1) and the vanishing h2 F, E ∨ (th) = h0 F, E(−th) = 0, t = 1, 2, return   h0 F, E ∨ (h) = 0, h0 F, E ∨ (2h) = 6. It follows that E ∨ (2h) is initialized. For such a bundle c1 = 4h − A and c2 = 8. The substitution of A with 3h − A gives an example of initialized indecomposable aCM bundle of rank 2 with c1 = A + h and c2 = 8. It is interesting to notice that the bundle E (hence also E ∨ (2h)) described above has an interesting property. Indeed, one can prove using Lemma 2.4 of [16] that both OF (A), and OF (3h − A) are very ample on F , thus they define two other different embeddings F ⊆ P3 with respective hyperplane line bundles OF (A) and OF (3h − A). Proposition 2.3 of [25] guarantees that E is aCM also for these embeddings: indeed, it is the bundle with m = 4 of Example 4.3 with respect to these embeddings. 5. On the zero–locus of general sections In this section we will assume that F ⊆ P3 is a general determinantal quartic surface. Lemma 5.1. Let E be an indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F⊆ P3 . If s ∈ H 0 F, E , then (s)0 6= ∅. RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 11 Proof. Assume that (s)0 = ∅. Then Sequence (6) becomes 0 −→ OF −→ E −→ OF (c1 ) −→ 0,   where h1 F, OF (c1 ) = h1 F, OF (−c1 ) 6= 0, because E is indecomposable. The coho mology of the above sequence gives a monomorphism H 1 F, OF (c1 ) → H 2 F, OF , thus h1 F, OF (c1 ) = 1. Remark 3.3 and Equality (1) imply that χ(OF (c1 )) is even, thus either OF (c1 ) or OF (−c1 ) is effective. Hence c21 ≥ 4 (see Lemma 3.5), thus h1 F, OF (c1 ) = 0 (see Lemma 3.4 and Proposition 2.1), a contradiction.     Trivially h1 F, E ∨ (2h) = 0: moreover h2 F, E ∨ (h) = h0 F, E(−h) = 0. The above vanishings prove the following lemma. Lemma 5.2. Let E be an initialized, indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F ⊆ P3 . Then E ∨ (3h) is regular in the sense of Castelnuovo–Mumford. In particular, both E ∨ (3h) and OF (6h − c1 ) are globally generated.  Let E be as above. If s ∈ H 0 F, E is a non–zero section, then (s)0 6= ∅. From now on we will assume that (s)0 = E ∪ D where E has pure dimension 0 (or it is empty) and D is an effective divisor.  Notice that the cohomology of Sequence (6) gives an injective map H 0 F, OF (D−h) → H 0 F, E(−h) , thus OF (D) is initialized too, or, in other words, OF (D−h) is not effective. Now assume that c1 = xh + yA is effective. It follows from Lemma 5.2 and Remark 3.6 that c1 is either 0, or 6h, or the inequalities y 2 + 3xy + x2 ≥ 1, 2x + 3y ≥ 1, y 2 − 3(6 − x)y + (6 − x)2 ≥ 1, 2(6 − x) − 3y ≥ 1, must hold. Thus, in the latter case, the point (x, y) must lie in the region delimited by the graphs of the two functions p √ 3(6 − x) − 5(6 − x)2 + 4 −3x + 5x2 + 4 , y= . y= 2 2 By drawing the graphs of the above functions (we made such computations using the software Grapher), we obtain Picture 1 below. 3 2 1 -2 -1 0 1 2 3 4 -1 -2 -3 Figure 1 5 6 7 8 12 GIANFRANCO CASNATI The values of c1 we are interested in are in the region bounded by the graphs of the two functions above, plus the two points (0, 0) and (6, 0) (see above). Thus c1 is in the following list: (12) 0, A, 2A, h, h + A, 2h, 2h + A, 3h − A, 3h, 3h + A, 4h − A, 4h, 5h − A, 5h, 6h − 2A, 6h − A, 6h. We are ready to prove the main result of the present section. Theorem 5.3. Let E be an initialized indecomposable aCM bundle of rank 2 on a general 3 determinantal quartic  surface F ⊆ P . 0 Let s ∈ H F, E and (s)0 = E ∪ D be as above. Then exactly one of the following cases occurs: (i) c1 − D is effective; (ii) E is a point, D = 0 and c1 is not effective; (iii) E = ∅, D ∈ |A| ∪ |3h − A| and c1 = 2h. Proof. Let c1 − D be effective. It is obviously true that c1 is effective, thus case (ii) above cannot occur. If c1 = 2h and D ∈ |A| ∪ |3h − A|, then c1 − D is not effective, because (2h − A)2 = (A − h)2 = −4, (see Lemma 3.5). It turns out that case (iii) above cannot occur as well. In order to complete the proof we can thus assume that c1 − D is not effective proving that case either (ii), or (iii) must hold. The cohomology of Sequence (7) for X := E twisted by OF (c1 − D) implies  h1 F, IE|F (c1 − D) ≥ deg(E). The cohomology of Sequence (6) implies    h1 F, IE|F (c1 − D) ≤ h2 F, OF (D) = h0 F, OF (−D) . If D = 0, then c1 is not effective. We have E 6= ∅ due to Lemma 5.1. Moreover   1 ≤ deg(E) ≤ h1 F, IE|F (c1 ) ≤ h0 F, OF = 1, thus equality must hold. It follows that E is a point. If D 6= 0, then   deg(E) ≤ h1 F, IE|F (c1 − D) ≤ h0 F, OF (−D) = 0, because D is effective and non–zero. Thus E = ∅ and Sequence (6) becomes 0 −→ OF (D) −→ E −→ OF (c1 − D) −→ 0.  The cohomology ofthe above sequence gives surjective maps H 0 F, OF (c1 − D + th) → H 1 F, OF (D + th) , for t ∈ Z. It follows that h1 F, OF (D + th) = 0 for each t ≤ 0. Moreover D is effective, thus h1 F, OF (D + th) = 0 also for t ≥ 1. We conclude that D is aCM. We already know that OF (D) is initialized, because the same is true for E. Due to Proposition 2.2 and Lemma 3.5 the only initialized aCM line bundles OF (xh + yA) 6∼ = OF 2 on F satisfy the equalities (xh + yA) = 4 and (xh + yA)h = 6, which are equivalent to the system x2 + 3xy + y 2 − 1 = 2x + 3y − 3 = 0. The unique solutions of this system correspond exactly to A and 3h − A. RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 13 If c1 is effective, then it is in the List (12). Taking into account that c1 − D is not effective, it follows that D is non–zero, whence E = ∅ from the above discussion. The pair (c1 , D) must be either (2h, A), or (2h, 3h − A), or it is in the following list; (0, 3h − A), (0, A), (h + A, 3h − A), (3h − A, A), (A, 3h − A), (2h + A, 3h − A), (4h − A, A), (2A, 3h − A), (h, 3h − A), (h, A), (5h − A, A), (6h − 2A, A). In order to complete the proof of the theorem we have to show that all the cases in the above list cannot occur. We will examine the cases one by one. Let (c1 , D) = (0, A). Sequence (6) becomes 0 −→ OF (A) −→ E −→ OF (−A) −→ 0,  which splits, because h1 F, OF (2A) = 0, contradicting the indecomposability of E. The same argument can be used to exclude also the case (0, 3h − A). Let (c1 , D) = (h, A). Sequence (6) becomes 0 −→ OF (A) −→ E −→ OF (h − A) −→ 0.   Since (2A − = −4, we deduce that h0 F, OF (2A − h) = h1 F, OF (2A − h) = 0. Equality (1) thus yields h1 F, OF (2A − h) = 0, hence the above sequence must split, again a contradiction. The same argument also excludes the case (h, 3h − A). 2 Let (c1 , D) = (A,3h − A). Since (2A − 3h)2 = (3h that both  − 02A) = −20, we deduce  0 2 h F, OF (2A − 3h) = 0, and h F, OF (2A − 3h) = h F, O (3h − 2A) = 0. It follows F    1 1 2 from Equality (1)that h F, OF (2A−3h) = 8. Since h F, OF (A) = h F, OF (A) = 0, it follows h1 F, E = h1 F, OF (2A − 3h) 6= 0 from the cohomology of Sequence (6). The same argument can be also used to exclude the cases (2A, 3h − A), (A + h, 3h − A), (3h − A, A), (4h − A, A), (6h − 2A, A). In the cases (A + 2h, 3h − A), (5h − A, A) we must modify slightly the above argument by taking the cohomology of Sequence (6) twisted by OF (−h).    Remark 5.4. We have h1 F, OF (2A − 2h) = h1 F, OF (4h − 2A) = 6, thus there exist indecomposable extensions h)2 0 −→ OF (A) −→ E −→ OF (2h − A) −→ 0, 0 −→ OF (3h − A) −→ E −→ OF (A − h) −→ 0. Both OF (A) and OF (−A) are aCM, thus the same is true for the above bundles E and it is easy to check that such bundles E are initialized as well. As pointed out in [8], Theorem 4.1 and Remark 4.2, we know that such bundles are indecomposable and that each non–zero section of the above bundles vanishes on a divisor in |A| and |3h − A| respectively. 6. Proof of the main Theorem In this section we will prove the main theorem stated in the introduction. We will first give the proof when c1 is not effective, then we will examine the case of effective c1 .  6.1. The case of non–effective c1 . If s ∈ H 0 F, E and (s)0 = E ∪ D, thanks to the previous Theorem 5.3, we know that D = 0 and E is a point. Lemma 6.1. Let E be an initialized indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F ⊆ P3 . If c1 is not effective, then c1 + h is effective. 14 GIANFRANCO CASNATI Proof. If c1 is not effective, then D = 0. Twisting Sequence (6) by OF (−c1 + th), we obtain (13) 0 −→ OF (D − c1 + th) −→ E ∨ (th) −→ IE|F (th − D) −→ 0. The cohomology of Sequence (13) with t = −1 yields    h0 F, OF (c1 + h) = h2 F, OF (−c1 − h) ≥ h1 F, IE|F (−h) . We also know that deg(E) = 1, hence the cohomology of Sequence (7) twisted by OF (−h)  1 yields h F, IE|F (−h) = 1.  Assume that c1 = xh + yA. It follows from Lemmas 5.2 and 6.1 that either c1 = −h, or the inequalities y 2 + 3(x + 1)y + (x + 1)2 ≥ 1, 2 2 y − 3(6 − x)y + (6 − x) ≥ 1, 2(x + 1) + 3y ≥ 1, 2(6 − x) − 3y ≥ 1, must hold. We obtain a picture similar to Figure 1, with the bottom branch of hyperbola, shifted by −1 in the vertical direction. The admissible values of c1 correspond either to (−1, 0), or to the integral points (x, y) in the bounded region such that either y 2 +3xy+x2 ≤ 0 or 2x + 3y ≤ 0, because c1 is not effective. Thus c1 is in the following list: (14) − h, 3A − 2h, A − h, 2A − h, 2h − A, 5h − 2A, 7h − 3A. Theorem 6.2. Let E be an initialized indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F ⊆ P3 . If c1 is not effective, then c1 ∈ { −h, A − h, 2h − A }. The zero locus E of a general s ∈ H 0 F, E) is a point. Moreover all these cases occur on F , and E is never Ulrich. Proof. We know that the above three cases occur thanks to Examples 4.4 and 4.5. Looking at List (14) it remains to exclude the cases 3A − 2h, 2A − h, 5h − 2A, 7h − 3A. Recall that D = 0 and deg(E) = 1.   2 = −20, thus h0 F, O (3h − 2A) = h1 F, O (3h − 2A) = 0, hence We have (3h − 2A) F F   h1 F, OF (3h−2A) = 8 (see Equality (1)). Thus we obtain h1 F, IE|F (3h−2A) = 9 from the cohomology of Sequence (7) twisted by OF (3h − 2A). The cohomology of Sequence (6) gives the exact sequence    0 −→ H 1 F, E −→ H 1 F, IE|F (3h − 2A) −→ H 2 F, OF , hence h1 F, E) ≥ 8, a contradiction. One can argue similarly when c1 = 7h −3A.  0 Let c1 = 2A − h. We trivially have h F, IE|F (h) = 3 and h1 F, IE|F (h) = 0, hence the cohomology of Sequence (6) twisted by OF (2h − 2A) gives the exact sequence    H 0 F, IE|F (h) −→ H 1 F, IE|F (2h − 2A) −→ H 1 F, E(h) −→ 0.  Since (2h − 2A)2 = −16, as usual we deduce that h1 F, IE|F (2h − 2A) ≥ 6, hence  h1 F, E(h) ≥ 3 necessarily, a contradiction. The same argument excludes the case c1 = 5h − 2A. Finally, notice that if c1 is not effective, then E is not globally generated, thus it cannot be Ulrich (see Remark 2.6).  Remark 6.3. Let E be as in the statement above. We know that it fits into Sequence (6) with D = 0. Since    h1 F, OF (h) = h1 F, OF (h − A) = h1 F, OF (A − 2h) = 0 RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 15 it follows that such a bundle is necessarily unique once the point is fixed. Thus E actually coincides with the bundles constructed in Examples 4.4 and 4.5.  6.2. The case of effective c1 . If s ∈ H 0 F, E and (s)0 = E ∪ D, thanks to the previous Theorem 5.3, we know that c1 − D is effective, unless c1 =  2h and D 6= 0. Lemma 2.4 and Remark 2.6 imply that 1 ≤ h0 F, E ≤ 8, because E is initialized. Moreover equality holds if and only if E is Ulrich. If this is the case D = 0 for a general choice of s (see again Remark 2.6).  Assume D 6= 0. We have 1 ≤ h0 F, E ≤ 7, then the cohomology of Sequence (6) and  0 2 the effectiveness of D give h F, OF (D) ≤ 7 and h F, OF (D) = 0. Proposition 2.1 and Lemma 3.5 also give h1 F, OF (D) = 0. It follows from Equality (1) that D2 ≤ 10. We deduce that D2 = 4 by Lemma 3.5 and Remark 3.3. If D ∈ |xh + yA|, then x2 + 3xy + y 2 = 1. Moreover 2x + 3y ≥ 1 (see Remark 3.6), hence √ −3x + 5x2 + 4 . y= 2 Taking into account List (12), an analysis case by case of the inequality (c1 − D)h ≥ 2 shows that the possible cases with effective c1 − D are exactly the ones listed in the following table. Table A c1 6h 6h − A 6h − 2A 5h 5h − A 4h 4h − A 3h + A 3h 3h − A 2h + A 2h h+A h 2A A 0 D D − c1 + h 0, A, 3h − A −5h, A − 5h, −2h − A 0, A, 3h − A A − 5h, 2A − 5h, −2h 0, 3h − A 3h − 6A, A − 2h 0, A, 3h − A −4h, A − 4h, −h − A 0, 3h − A A − 4h, −h 0, A, 3h − A −3h, A − 3h, −A 0, 3h − A A − 3h, 0 0, A, 3h − A −2h − A, −2h, h − 2A 0, A, 3h − A −2h, A − 2h, h − A 0, 3h − A A − 2h, h 0, A −h − A, −h 0 −h 0, A −A, 0 0 0 0, A h − 2A, h − A 0, A h − A, h 0 0 Is D − c1 + h effective? y/n. n, n, n n, n, n n, n n, n, n n, n n, n, n n, y n, n, n n, n, n n, y n, n n n, y y n, n n, y y The importance of the two last columns will be clear in a while. We know from Equality (1) that  c2 h2 F, E(−h) = 8 + 1 − c1 h − c2 . 2  Looking at Table A one deduces from Sequence (13) with t = 0 that h2 F, E = 0 if (c1 , D) is neither (3h − A, 3h − A), nor (A, A), nor (0, 0). (15)   c2 h0 F, E + h2 F, E = 4 + 1 − c2 , 2 16 GIANFRANCO CASNATI Looking at Table A one also  deduces from Sequence (13) with t = 1 that if D 6= 0, then 0 ∨ F, E(−h) = h F, E (h) = 0 in all the listed cases but (c1 , D) in the following short list: h2 (16) (4h − A, 3h − A), (3h − A, 3h − A), (h + A, A), (A, A), (0, 0). Proposition 6.4. Let E be an initialized indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F ⊆ P3 . Then exactly one of the following cases occurs:  (i) the zero–locus of a general section s ∈ H 0 F, E has pure dimension 0;  (ii) c1 = 2h and the zero locus of each non–zero s ∈ H 0 F, E is in either |A|, or |3h − A|.  Proof. Assume D 6= 0, so that 1 ≤ h0 F, E ≤ 7. Let (c1 , D) be  not in List (16). By 0 combining the above informations we obtain the equality h F, E = c1 h − 4, hence 5 ≤ c1 h ≤ 11, which turns out to be impossible by computing c1 h from Table A. Thus we have to restrict to (c1 , D) in List (16). In the case (0, is noth 0) there 0 2 0 ∨ ing to prove. In the case (A, = 1,  A) we deduce h F, E = 4, h F, E = h F, E h2 F, E(−h) = h0 F, E ∨ (h) = 4 Sequences (6) and (13). The first Equality (15) yields c2 = 1, the second c2 = 0, a contradiction. The case (3h − A, 3h − A) can be excluded similarly.   In the case (A+h, A) we deduce h2 F, E(−h) = h0 F, E ∨ (h) = 1 (from the cohomology of Sequence (13) with t = 1). The second Equality (15)  gives c2 = 7. The cohomology of Sequence (7) twisted by OF (h) gives h1 F, IE|F (h) ≥ 3. Thus the cohomology of  Sequence (6) implies h1 F, E ≥ 3 because OF (A) is aCM. The case (4h − A, 3h − A) can be excluded similarly.  We are now ready to give the complete classification of the bundles we are interested in when c1 is effective. For reader’s benefit we recall that a closed subscheme X ⊆ P3 is aG (i.e. arithmetically Gorenstein) if it is aCM and ωX ∼ = OX ⊗ OP3 (α) for a suitable α ∈ Z. Closed subschemes in P3 of degree at most 2 and dimension 0 are always aG. This is no longer true when the degree is d ≥ 3. Anyhow, if the Hilbert function of the homogeneous coordinate ring SE of a 0-dimensional, aG, closed subscheme E ⊆ P3 is (1, n1 , . . . , nσ−1 , d, . . . ) with nσ−1 < d, then E is CB with respect to OP3 (σ) (see [17], Theorem 1.1). Theorem 6.5. Let E be an initialized indecomposable aCM bundle of rank 2 on a general determinantal quartic surface F ⊆ P3 . If c1 is effective, then one of the following holds. (i) c1 = 0: then c2 = 2. The zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional aG subscheme contained in exactly one line. (ii) c1 = h: then c2 is 3, 4, or 5. The zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional aG subscheme of degree c2 of a linear subspace of P3 of dimension c2 − 2. (iii) c1 ∈ { A, 3h − A }: then c2 is 3, or 4. In this case the zero locus E of a general s ∈ H 0 F, E) is 0–dimensional subscheme which is not aG. (iv) c1 ∈ { A + h, 4A − h }: then c2 = 8. In this case the zero locus E of a general s ∈ H 0 F, E) is 0–dimensional subscheme which is not aG and which is contained in exactly one pencil of quadrics. RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 17 (v) c1 = 2h: then c2 = 8. In this case one of the two following cases occur: (a) the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional aG subscheme which is the base locus of a net of quadrics; (b) the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional aG subscheme lying on exactly one twisted cubic; (c) the zero locus E of each non–zero s ∈ H 0 F, E) is a divisor in either |A|, or |3h − A|. (vi) c1 = 3h: then c2 = 14. In this case the zero locus E of a general s ∈ H 0 F, E) is a 0–dimensional aG subscheme which is CB with respect to OF (3h). Moreover, all the above cases actually occur on F , and E is Ulrich if and only if c1 = 3h. Proof. Recall that the zero–locus E of a general section of E has pure codimension 2, but possibly in the case c1 = 2h (see Proposition 6.4). With this restriction in mind, we will prove the statement by an analysis case by case with the help of Table A. We first assume that h − c1 is effective. Looking at the last column of Table A this is equivalent to c1 ∈ { 0, h }, because D = 0. ∼ ∨ Let c1 = 0. In this  case E = E . Computing the cohomology of Sequence (6) we obtain 0 2 h F, E = h F, E = 1. The first equality (15) thus gives c2 = 2. As already pointed out, E is necessarily aG with Hilbert function (1, 1, . . . , 1, 2, . . . ). The homogeneous coordinate ring SE is generated in degree 1, thus there is only a single 1 in the sequence: in particular E is contained in exactly two linearly independent planes, hence in a single line. This case actually occurs as proved in Example 4.4. Let c1 = h. In this case the cohomology of Sequence (6) implies    h0 F, E = h0 F, IE|F (h) + 1 ≤ h0 F, OF (h) + 1 = 5.    Moreover h2 F, E = h0 F, E ∨ = h0 F, E(−h) = 0. The first equality (15) thus gives   c2 = 6 − h0 F, E = 5 − h0 F, IE|F (h) .  Thus c2 ∈ { 1, 2, 3, 4, 5 }. If c2 = 1 (resp. 2), we would have h0 F, IE|F (h) = 4 (resp. 3). Nevertheless, in this case E is a point which is necessarily contained in at most three planes (resp. a closed subscheme of degree 2, which is necessarily contained in at most two planes). Thus we obtain a contradiction. When c2 ≥ 3, then E is contained in exactly 5 − c2 linearly independent planes, i.e. in a linear subspace of dimension c2 − 2 of P3 . Since E must be CB with respect to OF (h), it follows that its closed subschemes of degree c2 − 1 are not contained in smaller linear subspaces, thus E is aG (see e.g. [23], Lemma (4.2)). These cases occur as proved in Example 4.3. From now on we will assume that h − c1 , hence −c1 , is not effective. Again we know from the last column of Table A which are  the possible  values of c1 . The cohomology of Sequence (13) with t = 0implies h2 F, E = h0 F, E ∨ = 0. We know that h0 F, E ≤ 8. By combining this restriction with the Equalities (15) we obtain the folowing chain of inequalities  c2 c21 c2 − 4 ≤ c2 = 8 + 1 − c1 h − h2 F, E(−h) ≤ 8 + 1 − c1 h, (17) 2 2 2 whence c1 h ≤ 12. Looking at List (12) one can immediately exclude all the cases but A, h + A, 2A, 2h, 3h − A, 3h, 4h − A, 6h − 2A. We will now examine these cases. Sequence (13) with t = 1 gives    H := h2 F, E(−h) = h0 F, E ∨ (h) ≤ h0 F, IE|F (h) ≤ 4. 18 GIANFRANCO CASNATI Let c1 = 2A. In this case Inequalities (17) force c2 = 4, hence H = 0. We will  check that 0 E is the intersection of two divisors in |A|, by checking that h F, IE|F (A) = 2: indeed A2 = 4 and the elements of |A| are integral (see Remark 3.3). Notice that the  coho0 0 mology of Sequence (6) twisted by OF (−A) gives h F, IE|F (A) = h F, E(−A) . Since    h2 F, E(−A) = h0 F, E(−A) , it follows that 2h0 F, E(−A) = 4, thanks to Equality (1). The above discussion proves the existence of an exact sequence of the form 0 −→ OF −→ OF (A)⊕2 −→ IE|F (2A) −→ 0, ⊕2 ∼ besides Sequence  (6). Thanks to Theorem 4.2 we deduce that E = OF (A) , because 1 h F, OF (−2A) = 0. The same argument allows us to exclude the case c1 = 6h − 2A as well.  Let c1 = A (the case c1 = 3h−A can be handled similarly), hence h0 F, E = 6−c2 ≤ 6. Inequalities (17) imply H ∈ { 0, 1, 2, 3, 4 } and c2 = 4 − H. The case H = 4 cannot occur, because E 6= ∅ (see Lemma 5.1). The case H = 3 cannot occur too: indeed, in this case, E should be a point, thus it cannot be CB with respect to OF (A), because this line bundle is globally generated. Assume H = 2. In this case E is a 0–dimensional closed subscheme of degree 2. As  already pointed out above we have h0 F, IE|F (h) = 2. Thus, computing the cohomology of Sequence (13) for t = 0 and 1, one checks that E ∨ (h) is initialized. Moreover, it is also aCM and indecomposable (by hypothesis) because the same is true for E. We have c1 (E ∨ (h)) = 2h − A and c2 (E ∨ (h)) = 0. We thus deduce from Theorem 6.2 that this case cannot occur. The same argument when H = 1 implies that E ∨ (h) is initialized indecomposable aCM with c1 (E ∨ (h)) = 2h − A and c2 (E ∨ (h)) = 1. As pointed out at the end of Example 4.5 we can construct a bundle with all the above properties, thus such an E exists. Moreover,  one easily checks h0 F, IE|F (h) = h0 F, E ∨ (h) = 1, thus E is not contained in a line, hence E is not aG in this case. Let H = 0. In this case c1 = A and c2 = 4. We can construct a bundle   with these properties, as shown in Example 4.6. We have h0 F, IE|F (h) = h0 F, E ∨ (h) = 0 as well. Thus E is not contained in any plane, hence its Hilbert function is (1, 4, . . . ). Since every subscheme of E of degree 3 trivially lies in a plane, it follows that E is not CB with respect to OP3 (1), hence it cannot be aG. Let c1 = h + A (the case c1 = 4h − A can be handled similarly). Inequalities (17) imply H ∈ { 0, 1, 2 } and c2 = 8 − H. Consider the exact sequence 0 −→ IF |P3 (t) −→ IE|P3 (t) −→ IE|F (th) −→ 0 We have that OF (−A) is aCM and IF |P3 (t) ∼ = OF (t − 4). Thus the cohomologies of Sequence (13) and of the above one for t = 0, 1, 2, imply that the Hilbert function of E is (1, c2 − 4, c2 , . . . ). In particular we obtain (1, 4, 8, . . . ), (1, 3, 7, . . . ), (1, 2, 6, . . . ) if H = 0, 1, 2 respectively. Since SE is generated in degree 1, looking at the growth from degree 1 to 2, we deduce that the only admissible Hilbert function is the first one, hence H = 0. It follows that c2 = 8 and that E is contained in exactly one pencil of quadrics. In particular, E is not CB with respect to OP3 (2), thus E cannot be aG. Moreover, h0 F, E = 6. A bundle with these properties can be constructed as shown end of Example 4.6.  at the  0 F, E(−h) = 0, thus Equality (1) for Let c1 = 2h. In this case h2 F, E(−h) = h  χ(E(−h)) yields c2 = 8, hence h0 F, E = 4. The complete description and the existence of initialized, indecomposable aCM bundles of rank 2 on F with c1 = 2h and c2 = 8 RANK TWO ACM BUNDLES ON GENERAL DETERMINANTAL QUARTIC SURFACES IN P3 19 is in [8], Proposition 2.1 and Theorems 6.2, 7.1, 4.1, where the three family of bundles corresponding to the cases (a), (b), (c) are described in details.  Notice that in all the above cases we showed that h0 F, E ≤ 7, hence the considered E are not Ulrich.  Let c1 = 3h. In this case H = 0 and c2 = 14, thus h0 F, E = 8, thanks to the first Equality (15): hence E is Ulrich. We know from Theorem 1.1 of [9] the existence of such kind of bundles on each smooth quartic surface.  Remark 6.6. Let E be as in the statement above. As in Remark 6.3 one easily checks that the bundles with c1 ∈ { 0, h, 2h, 3h } are uniquely determined by E, thus they actually coincide with the bundles constructed in Examples 4.4 and 4.3, or with the bundles described in [8] and [9] respectively. Similarly, when either c1 ∈ { A, 3h − A } and c2 = 3, or c1 ∈ { A + h, 4h − A }, then E is as at the end of Example either 4.5, or 4.6. Let us consider the case c1 = A more in details (analogous arguments hold in the cases   c1 = 3h − A). The first Equality (15) implies h0 F, E = 6 − c2 , hence h0 F, IE|F (A) = 5 − c2 ≥ 1, thus there is a curve in C ∈ |A| containing E. If C is smooth, the cohomology of a suitable twist of the exact sequence 0 −→ IC|F −→ IE|F −→ IE|C −→ 0,  and the isomorphisms IC|F ∼ (−E) yield h0 C, OC (C − E) = 1. = OF (−C), IE|C ∼ = OC  Riemann–Roch theorem on C implies h0 C, OC (E) = 2. Thus OC (E) is a complete gc12 on C. In particular the bundle E is constructed exactly as in either Example 4.6, if c2 = 4, or at the end of Example 4.5, if c2 = 3. An analogous construction holds when c1 = A + h (or 4h − A). Indeed, in this case c2 = 8. As in the previous case one h0 F, IE|F (A) ≥ 1. If there is a smooth curve C ∈ |A + h|, then we can argue as above that OC (E) is a complete g81 on C. References [1] V. Ancona, G. Ottaviani: Some applications of Beilinson’s theorem to projective spaces and quadrics. Forum Math. 3 (1991), 157–176. [2] A. Beauville: Determinantal hypersurfaces. Michigan Math. J. 48 (2000), 39–64. [3] D.A. Buchsbaum, D. Eisenbud: Algebra structures for finite free resolutions, and some structure theorems for ideals of codimension 3. Amer. J. Math. 99 (1977), 447–485. [4] M. Casanellas, R. Hartshorne: ACM bundles on cubic surfaces. J. Eur. Math. Soc. 13 (2011), 709–731. [5] M. Casanellas, R. Hartshorne, F. Geiss, F.O. Schreyer: Stable Ulrich bundles. Int. J. of Math. 23 1250083 (2012). [6] G. Casnati: On rank two aCM bundles. Preprint. [7] G. Casnati, F. Galluzzi: Stability of rank 2 Ulrich bundles on projective K3 surfaces. Work in progress. [8] G. Casnati, R. Notari: Examples of rank two aCM bundles on smooth quartic surfaces in P3 . Preprint arXiv:1601.02907 [math.AG]. To appear in Rend. Circ. Mat. Palermo. [9] E. Coskun, R.S. Kulkarni, Y. Mustopa: Pfaffian quartic surfaces and representations of Clifford algebras. Doc. Math. 17 (2012), 1003–1028. [10] E. Coskun, R.S. Kulkarni, Y. Mustopa: The geometry of Ulrich bundles on del Pezzo surfaces. J. Algebra 375 (2013), 280–301. [11] I.V. Dolgachev: Classical algebraic geometry. A modern view. Cambridge University Press (2012). [12] D. Faenzi: Rank 2 arithmetically Cohen–Macaulay bundles on a nonsingular cubic surface. J. Algebra 319 (2008), 143–186. [13] R. Hartshorne: Algebraic geometry. G.T.M. 52, Springer (1977). [14] D. Huybrechts: Lectures on K3 surfaces. To appear: draft notes available at http://www.math.uni-bonn.de/people/huybrech/K3Global.pdf . 20 GIANFRANCO CASNATI [15] D. Huybrechts, M. Lehn: The geometry of moduli spaces of sheaves. Second edition. Cambridge Mathematical Library, Cambridge U.P. (2010). [16] A. Knutsen: Smooth curves on projective K3 surfaces. Math. Scand. 90 (2002), 215–231. [17] M. Kreuzer: On 0–dimensional complete intersections. Math. Ann. 292 (1992), 43–58. [18] A. Lopez: Noether–Lefschetz theory and the Picard group of projective surfaces. Memoirs A.M.S. 89 (1991). [19] C. Madonna: A splitting criterion for rank 2 vector bundles on hypersurfaces in P4 . Rend. Sem. Mat. Univ. Pol. Torino 56 (1998), 43–54. [20] J.C. Migliore: Introduction to liaison theory and deficiency modules. Progress in Mathematics 165, Birkhäuser (1998). [21] C. Okonek, M. Schneider, H. Spindler: Vector bundles on complex projective spaces. Progress in Mathematics 3, (1980). [22] B. Saint–Donat: Projective models of K–3 surfaces. Amer. J. Math. 96 (1974), 602–639. [23] F.O. Schreyer: Syzygies of canonical curves and special linear series. Math. Ann. 275 (1986), 105–137. [24] K. Watanabe: The classification of ACM line bundles on a quartic hypersurfaces on P3 . Geom. Dedicata 175 (2015) 347–353. [25] K. Watanabe: On the splitting of Lazarsfeld–Mukai bundles on K3 surfaces. J. Algebra 447 (2016), 445–454. Gianfranco Casnati, Dipartimento di Scienze Matematiche, Politecnico di Torino, c.so Duca degli Abruzzi 24, 10129 Torino, Italy e-mail: [email protected]
0math.AC
Linear-Time In-Place DFS and BFS in the Restore Model Frank Kammer and Andrej Sajenko arXiv:1803.04282v1 [cs.DS] 12 Mar 2018 THM, University of Applied Sciences Mittelhessen, Germany {frank.kammer,andrej.sajenko}@mni.thm.de Abstract We present an in-place depth first search (DFS) and an inplace breadth first search (BFS) that runs in the restore model in linear time. To obtain our results we show several linear-time in-place graph transformations from one representation to another representation. Keywords space efficient, depth first search, breadth first search, cross pointer, restore model 1 Introduction Because of the rapid growth of the data sizes in nowadays applications, algorithms that are designed to efficiently utilize space are becoming more and more important. Another reason for the need for such algorithms is the limitation in the memory sizes of tiniest devices. Several models of computation have been considered for the case when writing in the input area is restricted. In the multi-pass streaming model [17] the input is assumed to be held in a read-only sequentially-accessible media, and the main optimization target is the number of passes an algorithm makes over the input. In the read-only word RAM [13] the input is assumed to be stored on a read-only randomly-accessible media, and arithmetic operations on operands that fit in a computer word are assumed to take constant time each. Another model allows the input to be permuted but not destroyed [5]. A variant of the latter model is called the restore model [8] where the input array is allowed to be modified during the process of answering a query, but it has to be restored to its original state afterwards. There are several algorithms for the read-only word RAM, e.g., for sorting [4,18], geometric problems [1,3,12], or graph algorithms [2,7,9,11,14,15,16]. Unfortunately, there is not much hope to find an in-place graph algorithm—i.e., an algorithm on graphs that uses only O(1) extra words—in the read-only RAM model since there is a lower bound for the reachability problem [10], i.e., the problem to find out if two given vertices of a given graph are in the same connected component. The lower bound essentially says that we can solve reachability in polynomial time only if we have roughly Θ(n) bits of working memory. Therefore, we focus on a different model: the restore model on a word RAM where we assume that the word size is at least ⌈log2 (N )⌉ with N being is the 1 size of the input. We are not aware of an in-place DFS of BFS in this model. However, Chakraborty et al. [6] introduced another model where the adjacency arrays of a graph can be only rotated, but a restoration is not required. In their model, they recently showed that one can run an in-place depth first search (DFS) and a breath first search (BFS) in O(nℓ2 + m) time if the graph has n vertices, m edges and if the algorithm constructs a DFS/BFS tree of height ℓ. They use a graph representation consisting of n + 4m + 2 words. 1.1 Our contribution We use the restore model to show linear-time in-place algorithms for both DFS and BFS that runs on a graph with a representation consisting of only n + m + 2 words on directed graphs and n + 2m + 2 words on undirected graphs (each undirected edge occurs with both endpoints) where n is the number of vertices and m the number of edges of the graph. To show our results we use different representations of the graph that all need the same number of words. We start to show that we can change the representations with a linear-time in-place algorithm. 2 Representation Graph algorithms usually do not specify the input format of a given graph since linear time and a linear number of words of working memory are usually sufficient to convert between any two reasonable adjacency-list representations—e.g., reorder the adjacency arrays with radix sort. However, having only o(n log n) bits of working memory, we have to be more specific about the input format and also the transformation from one input format to another is not trivial. In our standard representation (Fig. 1), we first store the number n of vertices and a table of pointers with one entry for each vertex. Subsequently, we store the total size of the adjacency arrays. In addition, we assume that the adjacency array of vertex i is stored before the adjacency array of vertex i + 1 for all i = 1, . . . , n − 1 and that all vertices inside an adjacency array are also stored in ascending order. Another representation is the cross-pointer representation (Fig. 2), which is identical to the standard representation except that the adjacency arrays store pointers to the cross-edges instead of vertex names. In other words, a word storing a vertex v in the adjacency array of a vertex u is replaced by the pointer to the word storing u in the adjacency array of v. The cross-pointer representation is useful for many space-efficient graph algorithms [11,15,16]. Moreover, we introduce in this paper the adjacency-array begin-pointer representation or short the begin-pointer representation (Fig. 3). We obtain the representation by replacing each vertex name v in the standard representation by a pointer to the beginning of the adjacency-array of vertex v. 2 A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 5 7 9 12 14 17 12 2 5 1 3 4 2 4 2 3 5 1 4 n 2m T Figure 1: Standard representation of a graph with m undirected or 2m directed edges. A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 5 7 9 12 14 17 12 9 17 7 12 14 10 15 11 13 18 8 16 n 2m T Figure 2. Cross-pointer representation of the graph from Figure 1. A 0 1 2 3 4 5 5 7 9 12 14 17 n T 6 7 8 9 10 11 12 13 14 15 16 17 18 12 9 17 7 12 14 17 9 14 9 12 7 14 2m Figure 3. Begin-pointer representation of the graph from Figure 1. 3 3.1 Transformations of our Representations Cross-Pointer and Begin-Pointer Representation To compute the cross-pointer representation, the idea is to store some information in the table T of pointers that helps our algorithm, but allows us finally to restore the table. For our algorithm, we use the invariant that the pointer T [v] of each vertex v in the table T points to the first neighbor of v in the adjacency array that is not yet replaced. The invariant is true in the beginning of the algorithm. We consider the vertices v in increasing order and iterate through the neighbors in their adjacency arrays also in increasing order. We replace the vertex names with the positions of their corresponding cross-edges as follows. The replacement takes a vertex u at a position pu and lookup the position pv = T [u] of vertex v in the adjacency array of u. By the invariant, T [u] points to the smallest neighbor of u that has not been replaced yet in u’s adjacency array, and this neighbor is exactly v. Now we replace the vertex names with the position of its corresponding edge, i.e., set A[pu ] = pv , A[pv ] = pu . To reestablish the invariant again, we increment T [v] and T [u]. Afterwards, we check if T [v] points at a word > n or if T [u] = T [u + 1]. If both check passes, continue with the next vertex in the adjacency array of v. Otherwise, we reached the end of v’s adjacency array and we proceed with vertex v + 1 unless v = n. To start with vertex v + 1, we start with the vertex in A[T [v + 1]]. 3 After this process, each pointer T [v] points to the word behind the end of v’s adjacency array, which is the beginning of the next adjacency array. Therefore, we can restore the pointers in T by using the previous vertex pointer and set T [v] := T [v − 1] for v = n, . . . , 2. Finally set T [1] := n + 2. Lemma 1. There is an in-place transformation from the standard representation to the cross-pointer representation that runs in linear time. The begin-pointer representation can be computed very easily. Iterate over all adjacency arrays and replace each entry A[i] := T [A[i]]. Lemma 2. There is an in-place transformation from the standard representation to the begin-pointer representation that runs in linear time. 3.2 Restoring the Standard Representation The goal is now to replace the pointers with the vertex name of the adjacency array they point into. To find out the vertex name of an adjacency array we write the vertex name into the first position of its adjacency array. Since we do not want to lose the values stored there, we move them temporally inside T . Of course, we lose the pointers stored there, but we can reconstruct them by knowing the positions where the vertex names are stored. Because the vertex names are lower than a cross-pointer, we are still able to distinguish between them and to notice whenever an adjacency array ends and a new begins. We call a cross-pointer or a begin-pointer representation swapped whenever each adjacency array has a vertex name and T [u] stores an information that is usually part of u’s adjacency array. For examples, see Fig. 4 and 5. Clearly, we can change a representation into a swapped one by iterating once through T . For the reverse direction, we have to iterate over all adjacency arrays to find the vertex names and write their positions into T . Lemma 3. There is an in-place transformation that swaps and unswaps a representation in linear time. For a simpler description of the algorithm we assume that an access to the adjacency array always returns a pointer and Name[p] returns the vertex name if it is stored in A[p]. Moreover, if we say that we store a vertex name v in A or store v at some position of some adjacency array, then we mean that the value from there is moved to T [v] first and subsequently v is stored. Formally, we distinguish between the raw access to A storing the real data and an access via A[i] with reading and writing in A[i] means reading and writing in A[i] if and only if A[i] = 0 or A[i] > n, otherwise the access is to T [A[i]]. In addition, we have Name[i] = A[i] if and only if A[i] 6= 0 and A[i] ≤ n, otherwise it is null. In our example shown in Fig. 4, we have A[7] = 9, A[9] = 7 and Name[7] = 1. To compute the standard representation from a swapped begin-pointer representation is easy. If the given representation is not swapped, then make it swapped. Iterate then over all adjacency arrays and store in each entry A[i] the vertex name Name[A[i]]. Finally, unswap the representation. 4 A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 5 9 7 10 11 8 12 1 17 2 12 14 3 15 4 13 18 5 16 n 2m T Figure 4. Swapped cross-pointer representation of the graph in Fig. 1. A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 5 9 7 9 9 7 12 1 17 2 12 14 3 14 4 12 17 5 14 n T 2m Figure 5. Swapped begin-pointer representation of the graph in Fig. 1. Lemma 4. There is an in-place transformation from the begin-pointer representation to the standard representation that runs in linear time. To compute the standard representation from a swapped cross-pointer representation, we use an invariant similar to the one used in the previous subsection. The vertex name is stored in the first word of an adjacency array that has not yet been replaced by a vertex name. The invariant is true in the beginning of the algorithm. We iterate over all pointers in all adjacency arrays in increasing order with respect to the stored position in A. While considering position q in the adjacency array of a vertex v we replace the pointer p = A[q] as follows. Determine the vertex name u of the adjacency array to which p points. By the invariant, u = A[p]. Set A[q] = u and A[p] = v. Store the vertex names of v and u in A[q + 1] and A[p + 1], respectively. Thus, the invariant holds again. After moving the vertex names, check if A[q + 1] > n. If so, continue with the next pointer in the adjacency array of v. Otherwise, we reached the end of v’s adjacency array. Then skip over all words smaller than n (these words are already replaced by vertex names) until reaching the end of A or finding in A an entry v + 1, which is the position in A of the first not-converted pointer in (v + 1)’s adjacency array. Lemma 5. There is an in-place transformation from the cross-pointer representation to the standard representation that runs in linear time. 4 Depth-First Search As usual, we extend our DFS by the calls of user implemented functions as follows. Initially all vertices of a graph are unvisited, also called white. Whenever a DFS visits a vertex v for the first time, we call visit(v). Within this call, we color v gray and execute preprocess(v). Then we iterate over all outgoing edges 5 of v. Whenever the other endpoint u of such an edge is white we call preexplore(v, u). If u turns black and we backtrack to v, we run postexplore(v, u). When finally everything is explored from v, we call postprocess(v) and color v black. Using suitable implementations for the four function, we can easily output, e.g., the vertices in pre-, post-, or inorder with respect to our constructed DFS tree. 4.1 DFS on Directed Graphs To run our DFS on directed graphs, we use the swapped begin-pointer representation. To run an in-place DFS, we need to know the colors of the vertices and, if a vertex turns black, the vertex where we have to backtrack. To store the information, we assume for the time being that every vertex has at least two edges. Let s be the start vertex of our DFS. Note that, in the swapped begin-pointer representation, every pointer in an adjacency array points to a word with a value at most n. Our invariant of our algorithm is that a vertex v is white if and only if v 6= s and the second entry in the adjacency array points to a word with a value at most n. Thus, the invariant is initially true. Note also that the invariant allows us to test in constant time if a vertex is white. Assume that our DFS currently visits vertex u and we iterate through u’s adjacency array. We check for each neighbor v of u if it is white. If so, we visit v. Let p be the position of the pointer to v’s adjacency array in u’s adjacency array. Then, v’s adjacency array begins at position q = A[p]. To visit u, the idea is to set A[p] = A[q + 1] and A[q + 1] = p. Note that the second entry A[q + 1] of v’s adjacency array now points into u’s adjacency array and is thus a value larger n. By the setting we additionally have the information to backtrack from v to u since A[q + 1] is the pointer to the position in u’s adjacency array such that we can continue our iteration for u. However, there is a technical issue one has to take care of. By the definition of A, if a vertex w is discovered from the pointer x at the first position q in v’s adjacency array, then storing something in A[q] means that we write it to the table T [v]. Moreover, the reverse pointer in v’s adjacency array can not point to q (or our invariant would be false). Instead we point to q + 1. Furthermore, if a vertex w′ is discovered from the second pointer p in v’s adjacency array, then we can not use the second position in the array to store a pointer from x′ —otherwise we would loose the pointer used to backtrack from v. Therefore, the idea is, after returning from w, to swap the pointer w and w′ in v’s adjacency array and to explore again the first pointer in the array. Whenever we return from a vertex to the second position of an adjacency array, we have to check additionally if the first two pointers in the adjacency array are swapped. If so, we undo the swap and continue with the third pointer in the array. Otherwise, swap as described above. For an illustration of the algorithm, see Fig. 6. Before we present the details of our algorithm, let us simplify its description. We define A∗ [i] as a redirection to A[i] exactly if i is not the second entry of an adjacency array or if the second entry still points to the beginning of an 6 p q q v p q ′ v p q ′ v p q ′ v p q x v p q x v . . . backtrack and swap x and x′p v q v x visit v u v x visit w u v z u . . . backtrack v x swap x and x v u u v z′ x u u visit w′′ v x u . . . backtrack from w′′ v x u . . . backtrack from v v x x x x x ′′ w x ′′ w x ′′ w q+1 x ′′ w z+1 x p z x p x p ′ x′ visit w′ x x ′ u x′ p x′′ z′ x′ p x′′ w′ q+1 x′′ p x′ v p q x′ v p q ′ v p q q v x′ +1 x′′ x w′ x′′ w′′ z ′′ x′′ p z ′′ w′′ q+2 p x′′ w′′ z ′′ +1 Figure 6: Animation of our DFS on directed graph visiting vertices u, v, x, return to v, visit x′ , return to v, visit x′′ , return to v and return to u. adjacency array, i.e, (Name[i − 1] = null) ∨ (0 < A[i] ≤ n). Otherwise, p = A[i] is a back pointer to a parent into a position where the actual edge is stored. Because we never point to the first position inside an adjacency array, we have to run an extra check if p points to the second position (Name[p − 1] 6= null). In this case we find our value at the first position and accesses to A∗ [i] must be redirected to A[p − 1] = A[A[i] − 1]. Otherwise, A∗ [i] redirects to A[p] = A[A[i]]. 7 In detail, our DFS runs as follows. Starting with a vertex v init we search for the vertex in O(m) time, i.e., search for the position p with A[p] = v init . Then we call visit(p). – visit(p): (Visit the vertex whose adjacency array starts at position p.) In the swapped begin pointer representation, v = Name[p] is always the vertex name. First, preprocess(v). Finally, start iterating through the neighbors starting from position p by executing nextNeighbor(p), but skipping over the first check. – nextNeighbor(p): (Follows the edge at position p in A∗ if the opposite endpoint of the edge is white. Otherwise iterate to the next position p + 1.) First, we check if the edge belongs to the same adjacency array, i.e., check if Name[p] = null. If not, we reached the next adjacency array and we return to the parent by lookup the largest q ≤ p such that Name[q] 6= null and by calling gotoParent(q). Second, we check if the edge at p points to a white vertex and run the non-recursive procedure isWhite(A∗ [p])=true. If so, call gotoChild(p). Otherwise, call nextNeighbor(p + 1) again. – isWhite(p): (Return true exactly if the vertex v whose adjacency array begins at position p is white.) Recall that we have to check if the second word of v’s adjacency array points to a value > n. Consequently, return A[A[p + 1]] > n. – gotoChild(p): (We are going to discover a new child via an edge e stored at position p and color it implicitly gray.) First we determine the position q = A∗ [p] where e points to. Second we get the name of the neighbor v = Name[q]. Third, we are going to create a pointer to backtrack later and replace the edge to v with the second edge of v’s adjacency array. Therefore, remember the second edge x = A[q + 1] of v. Now, store the edge inside A∗ [p] := x and create a back pointer from v’s second word into its parent’s adjacency array as follows: Set A[q + 1] := p + 1 if and only if p is the first position inside the parent’s adjacency array, i.e, if Name[p] 6= null. Otherwise set A[q +1] := p. Finally, visit the neighbor by executing visit(q). – gotoParent(q): (From a child v go to its parent where q is the beginning of v’s adjacency array and p = A[q + 1] is a pointer to the adjacency array of the parent.) Before going to the parent we have to restore the edges that we modified when we went to the child. Thus, we first restore the child’s edge that was temporary stored in the parent’s adjacency array, but let it point one edge further to guarantee the invariant. In detail, we set A[q + 1] := A[p] + 1. Then we restore the pointer to the child by setting A[p] := q. It remain to check the technical issue described above. For this, check if p points at the second position of the array, i.e, check if Name[p − 1] 6= null. If not, call nextNeighbor(p + 1). If the check is true, we have to check if we left the parent over the first or the second edge. Therefore, check if the first two words inside the parent’s adjacency array are swapped, i.e, check if A∗ [p − 1] < A∗ [p]. If so, we left over the first edge of the parent. Thus, swap A∗ [p − 1] and A∗ [p] and call nextNeighbor(p). Otherwise we left over the second edge and we swap A∗ [p − 1] and A∗ [p] back and call nextNeighbor(p + 1). 8 – findNextStartVertex: To check for another unvisited vertex we iterate through all adjacency arrays by using a global variable. If the invariant fails for a vertex v whose adjacency array starts at position p, run visit(p). It remains to describe what we have to do for vertices of degree smaller than 2. If such a vertex is a start vertex, store the information in a global variable so that we do not need a second entry in the adjacency array. In addition, whenever we reach a vertex v of degree 1, then we will neither proceed from v to any other vertex nor reach v from another vertex again. We therefore need no manipulation of v’s adjacency array. As a consequence, we can redefine our invariant such that vertices of degree smaller than 2 are excluded from the invariant. Finally note that ignoring the calls for the user-defined functions as well as isWhite, which is not recursive, and findNextStartVertex we only make tail-calls and consequently require no recursion stack. Theorem 6. There is a in-place DFS for directed graphs in the restore model that runs in O(n+m) time on n-vertex m-edge graphs with a representation consisting of n + m + 2 words and that calls the user defined functions preprocess and postprocess. Of course, the algorithm above can also run on an undirected graph. However, on such graphs, we can also call the user defined functions preexplore and postexplore as we see in the next section. 4.2 DFS on Undirected Graphs To run our DFS on undirected graphs, we use the swapped cross-pointer representation. The DFS works as follows. Whenever we use a cross-pointer to move to a new vertex by calling visit(q), we jump into the middle at position q of an adjacency array of a white vertex. From there, we search for the vertex name v. Recall that we find the name at the first position of the adjacency array. We then call preprocess(v). In addition, replace all cross-edges of v’s edges by zero (avoid so visits of v from descendants of v) and all zero entries at v’s adjacency array by n + 1 (avoid so visits of v from proper ancestors of v’s parent). Then we iterate over all non-zero pointers p starting from position q+1 to the last pointer in v’s adjacency array, then jumping to the first pointer until we reach position q − 1. Note that we can easily test if we reached the end of v’s adjacency array by reading the value v + 1 in A, which is the name of the next vertex. For each such pointer p, restore the cross edge at position p (so that we can backtrack later), determine the new vertex name as described above and call preexplore(v, u). Moreover, set A[p] = v so that we know v’s name when returning to v. Then, call visit(p). The call returns the position p such that our current visit of vertex v does not need to store p explicitly. Reading A[p], we relearn the name of v and call postexplore(u, v). Then we step to the next non-zero pointer in v’s adjacency array unless we reach position q with 0 < A[q] ≤ n. If we reach q, we can backtrack and inform the previous vertex that we used the edge at position q. Fig. 7 and 8 sketch two snapshots of a DFS. 9 A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 5 0 7 0 0 8 12 1 17 2 12 14 3 15 4 13 18 5 16 n 2m T Figure 7. A DFS starts at vertex 2. A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 5 9 2 6 6 8 12 1 1 2 12 14 3 0 4 13 18 5 5 n T 2m Figure 8: The DFS continues from Fig. 7 and visits vertices 2,1,5 and 4 in this order. To start our algorithm below, we can either call findNextComponent to output all DFS trees or one can output only the vertices of one DFS tree with start vertex by storing the start vertex in a global variable s and calling visit(q) where q is the first position in the adjacency array of s. – visit(q): (Visit a vertex v via an edge that is in v’s adjacency array at position q.) First, compute the boundaries of the adjacency array containing A[q], i.e, determine the largest a ≤ q and the smallest b ≥ q with Name[a] 6= null and Name[b] 6= null. Then, v = Name[a] is our vertex name and we call preprocess(v) and exit. Second, eliminate the possibility to visit v again by iterating over p ∈ [a, b]/{q}, replace all A[p] = 0 by A[p] = n + 1, check if there is a crosspointer stored in A[p] > n and destroy it with A[A[p]] := 0. Finally, start iterating through the neighbors starting from position q + 1 by calling nextNeighbor(v, q + 1). – nextNeighbor(v, p): (Finds the next unvisited neighbor and visit it using v as the name of the current vertex and p is the current position in A.) First, check if A[p] belongs to v’s adjacency array, i.e., check Name[p] 6= v + 1. If not, iterate to the largest position q < p such that Name[q] = v and set p := q. Second, check if A[p] contains the back edge to the parent of v, i.e, check if A[A[p]] ≤ n. Also check if Name[p] = s = v, i.e, check if we currently finished the start vertex. If one of the two checks succeeds, all neighbors of v are visited and we return by exiting and calling gotoParent(v, p). If A[p] = 0 or A[p] = n + 1, A[p] initially stored a back or forward edge, respectively, and we can ignore it, i.e, call nextNeighbor(v, p + 1). Otherwise, A[p] is a pointer to an unexplored adjacency array. Peek inside this adjacency array at position A[p] and iterate back using a pointer q until finding the vertex name u = Name[q] 6= null. Then, call preexplore(v, u) and explore u by calling gotoChild(v, p). 10 – gotoChild(v, p): (We are going to discover a new child of vertex v via an edge e stored at position p.) Restore the cross edge e′ of e so that the child gets a pointer to the edge e by setting A[A[p]] = p. Remember the position of edge e′ in q = A[p] and write the vertex name v at A[p] such that we can quickly find out the name of the parent of the child when returning to the parent. Visit the child by calling visit(q). – gotoParent(v, p): (Go to the parent of vertex v using the pointer stored at position p) Go to the parent at the position q = A[p], remember the vertex name of the parent u = A[q], set A[q] = p and A[p] = 0 to restore the edge to the child and destroy the back-edge of the child. Call postexplore(v, u) and continue handling the remaining neighbors of the parent by calling nextNeighbor(u, q + 1). – findNextComponent: Check for another connected components by iterating through all adjacency arrays to find one that does not contain a single pointer p with (A[p] = 0) ∨ (A[p] = n + 1). If there is one, go to the largest position q ≤ p of the adjacency array in A to store the name of the corresponding vertex in some global variable s = Name[q] and to call visit(q). Otherwise, we are done. Since we destroyed only one of each pair of cross pointers, we can easily restore them by iterating over all adjacency arrays. If we find a pointer (A[p] 6= 0)∧ (A[p] 6= n + 1) at a position p, then follow it and set A[A[p]] = p. So we have restored the cross-pointer representation. As in the previous subsection, ignoring the calls for the user-defined functions and the function findNextComponent we make only tail-calls and require consequently no recursion stack. Theorem 7. There is an in-place DFS for undirected graphs in the restore model that runs in O(n + m) time on n-vertex m-edge graphs with a representation consisting of n + 2m + 2 words and that calls all user defined functions. 5 Breadth-First Search As usual for a BFS, our algorithm runs in rounds and, in round z − 1, all vertices of distance z from a start vertex are added into a new list. Then our algorithm can always iterate through a list of vertices and for each such vertex u, we iterate through u’s adjacency array. For a simpler description, assume that all vertices are initially white and whenever a vertex is added into the BFS tree, then it turns light gray. If we are in the round where the vertex is processed, the vertex is dark gray. After adding its white neighbors into the BFS tree, the vertex turns black. The following description works on directed and undirected graphs. To implement our BFS we make use of the following observation. In the standard representation all words in T are stored in ascending order. Our idea is to partition them in regions such that the most significant bits of the words are equal per region. We use this to create a shifted representation of T by ignoring the most significant bits and shifting the words in T together (Lemma 8) such 11 that we have a linear number of bits free to store a c-color choice dictionary [14, Theorem 7.6] D as demonstrated in Fig. 9. We encapsulate the read access to the words stored packed in T through a new data structure T . The details of the access are described in the proof of Lemma 8. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 p1 p2 2 choice dictionary posit. & c′ packed words Figure 9. Shifted representation with c-color choice dictionary. Lemma 8. Let c > 0 be a constant. Having an array of n ≥ 2c+1 w ordered words we can pack it in linear time with an in-place algorithm such that we have cn unused bits free and that we still can access all elements of the array in constant time. With a similar linear time in-place algorithm, we can unpack the words. Proof. The idea is to partition the words into parts such that each pair of words in a part has the same c′ = c + 1 significant bits. Since the sequence is ordered, ′ we iterate over all words and look for the positions where one of the most 2c − 1 significant bits change. During the construction we remember all these positions in the working memory. Now, the most significant c′ bits of each word are equal per region. We treat them as unused space and shift every of the packed words of w − c′ bit each together. Stored packed, they occupy n(w − c′ ) bits in total such that it leaves ′ c′ n bits free to use. We use the last 2c words to store c′ and all the positions. ′ c′ ′ Thus, c n − 2 w ≥ c n − n = cn bits remain free. For implementing a function read(i) that reads the ith original word, we have to identify its current position that can be distributed between two words, to cut its bits out of the two words and to use the remembered position reconstruct its most significant bits. For the following description assume that the bits of a word are numbered from 0 (least significant) to w − 1 (most significant). In detail, the ith word in T originally stored at position w(v − 1) was shifted exactly c(v − 1) bits and now starts after x = (w − c)(v − 1) bits, i.e., it starts with bit y = (x mod w) in the word ((x div w) − 1) and consists of the next w − c bits. Using suitable shift operations we can get the ith word in constant time. To reconstruct its most significant bits, scan the last c′ words to determine the part to which i belongs. To restore to the standard representation of the array, we store c′ and the positions in the working memory. Afterwards, we iterate over the words backwards and set T [i] = read(i) for all i ∈ [1, n].  12 Before we now obtain our linear time BFS, we want to remark that the shifted representation can not be used to run a standard DFS since a stack for the DFS can require Θ(n log n) bits on n-vertex graphs. We first prepare the shifted representation of our graph. Then we can use the free bits to implement a c-color choice dictionary D in which we store the colors of the vertices and to iterate over dark gray vertices in constant time per vertex. As described in [14], the c-color choice dictionary provides the following functions. – iterate.init(c): Prepare an iteration over all entries with the color c. – iterate.next(): Returns the next entry that is colored with the color c. – iterate.more(): Returns true exactly when iterate.next can return an entry colored as color c. – setColor(c, v): Colors an entry v with the color c. – color(v): Returns the color of the entry v. To start our BFS at vertex v, we first initialize the c-color choice dictionary D for four colors with all vertices being initially white. Remember in a global variable a round counter z = 0 to output the round number for each vertex. Then, color the root vertex v light gray by calling D.color(lightGray, v). Finally, we start to process the whole DFS-tree as follows. Whenever the current round counter z is even, we look for light gray vertices and color their white neighbors dark gray and if z is odd we do vice versa. We explain how to process the vertices if z is even. For an odd z, simple switch the meaning of light gray and dark gray. First, we call D.iterate.init(lightGray) to prepare the iteration over the light gray vertices. Now we iterate over all light gray vertices using the D.iterate.more function to check if there is another light gray vertex left to process and D.iterate.next to get the next light gray vertex v. As long as we have a light gray vertex v we output z as its round number, color all its white neighbors dark gray, and color v black. For this, we iterate over v’s adjacency array starting at position p = T [v] and ending at q = T [v + 1] − 1 where T [n + 1] is a pointer to the word strictly behind our graph representation. For every neighbor u = A[j] with p ≤ j ≤ q we check if D.color(u) = white and if so we color u dark gray by calling D.setColor(darkGray, u), otherwise we ignore it. After the iteration over v’s adjacency array we call D.setColor(black, v). If the iteration over the light-gray vertices ends, we reiterate over all lightgray vertices again and color them black to finish the current round. If we could color a vertex dark gray in the current round, then there are vertices left to process. We increase z by one and start a new round by iterating now over the dark gray colored vertices as described. If we could not color any vertices dark gray, we have no vertices left to process and finish the BFS. By Lemma 8, we can restore to the standard representation. Theorem 9. There is an in-place BFS for directed graphs in the restore model that runs in O(n + m) time on n-vertex m-edge graphs with a representation consisting of n + m + 2 words. 13 5.1 BFS on Undirected Graphs In the following, we present another linear-time in-place BFS that only works on undirected graphs. For the following BFS, do not use a shifted representation. Instead, we use the swapped begin-pointer representation where each word is a pointer or a vertex name. In addition, we can use the redundancy of the edges (each undirected edge is stored as a directed edge from its endpoints) to store the depth of each vertex in the BFS-tree constructed. Consider the situation that we are in round z and currently consider a vertex u and we want to add a white neighbor v into the list for round z + 1. Then the idea is to replace the entry u for the edge {v, u} in the adjacency array of v by some extra information that helps us to build a list consisting of all vertices of round z + 1. Moreover, we have to guarantee that we can use the extra information to test quickly if a vertex is white or not. In detail, we use the invariant that a vertex v is white if and only if v is not the start vertex and the second entry in the adjacency array points to value that is not zero and at most n. Note that we do not distinguish in our invariant if a vertex is gray or black. A vertex is dark or light gray exactly if it is part of the list over which we currently iterate or which we currently build, respectively. Therefore, we move the extra information at the beginning of v’s adjacency array by swapping the information of two edges in the adjacency array of v. In detail, assume that v’s adjacency array starts in A at position p and the edge {u, v} in the adjacency array of u is stored at position q. We then store the value of A[p + 1] in A[q] and set A[p + 1] to the first position of the adjacency array of the last vertex that was added to round z + 1 or to 0 if there is no such vertex. The swap is also sketched in Fig. 10. Finally, output v as a vertex in depth z + 1, color it gray and proceed with the next vertex in u’s adjacency array. If the iteration over all vertices in round z is done, proceed with the vertices in the list for round z + 1. Note that, after choosing one start vertex and coloring it gray, the invariant is initially true. In the first round, the list consists of only that start vertex. i k u p i k u p p p+1 ... v j p p+1 ... v x j q ... i j q ... j o p x x+1 ... o p w x x+1 ... w Figure 10: A white vertex v (above) is discovered from a vertex u and turns gray (below). The previous discovered vertex is w. A[x + 1] points to the vertex discovered before w. 14 Moreover, the pointer stored for the lists are values larger than n so that the invariant holds again after coloring v gray. The only value we have to store is an iterator over the current list and two pointers to the last vertex added into the list for the current and the next round. In detail our BFS runs as follows. Choose a start vertex v = Name(p) whose adjacency array begins at index p. Then run bfs(p). – bfs(p): (Initially starts the BFS on a vertex at the position p.) In a global variable remember the start vertex v start = Name[p] and the current round number z = 0. Output the vertex v start as a vertex of depth 0. Increment the round number z by one and call addNeighborsOfVertex(p, 0) to create a list of adjacency arrays of vertices of depth 1. The call returns a pointer p′ to the first such adjacency array of the list or 0 there is none. If p′ is 0, we can discover no more vertices from v start and we finish. Otherwise, we run the following two steps. First, we call runRound(p′ ) that returns another pointer p′′ . Second, we increment z by one and set p′ = p′′ . We repeat these two steps until p′′ = 0. Then the whole BFS tree is output. – addNeighborsOfVertex(i, r): (Extends a list of adjacency arrays starting at position r by the adjacency array of white neighbors of the vertex u = Name(i) and returns a pointer the first adjacency array in the list or 0 if there is none.) Iterate over the whole adjacency array of the vertex u using a variable k until Name[k + 1] = u + 1 ∨ k > n + m + 1, but ignore the second entry i + 1 for every u 6= v start since the adjacency array of the start vertex is not linked and the second entry is still a pointer to a neighbor. In each iteration step, whenever a pointer p = A[k] points at a white vertex v = Name(p), we add it into the list for the next round. In detail, check if isWhite(p) is true. If so, lookup the position of u (more exactly the position of the begin pointer to u) in v’s adjacency array, i.e, the smallest position q ′ ≥ p such that A[q ′ ] = i. Swap the words at position p + 1 and q ′ . Now the edge pointing at u is stored at the second position and can be overwritten to store a link to the currently first adjacency array in the list as follows. Write the pointer r into the second word at index p + 1 and set the start pointer r := p such that r is pointing at the start the list again. After the iteration, return r. – runRound(p): (Outputs every vertex of the list of adjacency array using p as a pointer to the first adjacency array and returns a pointer to a list for the next round or 0 if there is none.) Use a global variable r to store a pointer at a linked list of adjacency arrays for the vertices of the next round and set it initially to 0. In fact, we reuse the next-to-last such pointer. We now iterate over all adjacency arrays in the list starting with p. In each iteration check if we reached the end of the list by checking if p = 0. If so return r. Otherwise, first output the vertex v = Name[p] as a vertex of depth z. Then, link the white neighbors of v by calling addNeighborsOfVertex(p, r) that returns a pointer r′ pointing at the start of a new linked list. Finally, set r = r′ and p = A[p + 1] to proceed with the next vertex of the current round. Continue with the next round. 15 – isWhite(p): (Return true if the adjacency array starting at position p was never visited.) An adjacency array was never visited if it does not belong to the start vertex and is not linked. Therefore, return true exactly if both Name(p) 6= v start and the second pointer points to a word x = A[A[p + 1]] with 0 < x ≤ n. Notice that excluding the functions isWhite and bfs, which both are not recursive, all other function calls are tail-recursive. Consequently, we do not need a recursion stack. To restore the begin-pointer representation, we iterate over all adjacency arrays and for each array, set the second entry to 0 and undo the swap by interchanging the second smallest pointer in the array with the second entry of the adjacency array. Then we run the transformation to the standard representation, but leave the zero entries unchanged. We so get a standard representation— except one entry in each adjacency array is a zero. And finally, we run the transformation to the cross-pointer representation to turn the zeros into the correct cross pointers. Two changes in the transformation: 1.) Step over a zero to the next vertex name. 2.) When a vertex v points to a vertex w = A[T [v]], but A[T [w]] < v. Let u be the vertex in A[T [w]]. Then the edge {u, w} was skipped due to a zero. Search once in the adjacency array of u for the position p with the 0. Set A[T [w]] = p as well as A[p] = T [w] and increment T [w] by one. Theorem 10. There is an in-place BFS for undirected graphs in the restore model that runs in O(n + m) time on n-vertex m-edge graphs with a representation consisting of n + 2m + 2 words. We want to remark that we can also store the depth of each vertex inside the representation. After iterating through the current list and building the next list, we only have to iterate again through the current list and replace all pointers of the list by the current round number. Nevertheless, we can restore the representation as described above. 6 Acknowledgments Andrej Sajenko acknowledge support of the Deutsche Forschungsgemeinschaft (DFG), project SpaceeffGA (KA 4663/1-1). References 1. Tetsuo Asano, Kevin Buchin, Maike Buchin, Matias Korman, Wolfgang Mulzer, Günter Rote, and André Schulz. Reprint of: Memory-constrained algorithms for simple polygons. Comput. Geom. Theory Appl., 47(3, Part B):469–479, 2014. 2. Tetsuo Asano, Taisuke Izumi, Masashi Kiyomi, Matsuo Konagaya, Hirotaka Ono, Yota Otachi, Pascal Schweitzer, Jun Tarui, and Ryuhei Uehara. Depth-first search using O(n) bits. In Proc. 25th International Symposium on Algorithms and Computation (ISAAC 2014), volume 8889 of LNCS, pages 553–564. Springer, 2014. 16 3. Luis Barba, Matias Korman, Stefan Langerman, Rodrigo I. Silveira, and Kunihiko Sadakane. Space-time trade-offs for stack-based algorithms. In Proc. 30th International Symposium on Theoretical Aspects of Computer Science (STACS 2013), volume 20 of LIPIcs, pages 281–292. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2013. 4. Paul Beame. A general sequential time-space tradeoff for finding unique elements. SIAM J. Comput., 20(2):270–277, 1991. 5. Hervé Brönnimann, John Iacono, Jyrki Katajainen, Pat Morin, Jason Morrison, and Godfried Toussaint. Space-efficient planar convex hull algorithms. Theor. Comput. Sci., 321(1):25–40, 2004. 6. Sankardeep Chakraborty, Anish Mukherjee, Venkatesh Raman, and Srinivasa Rao Satti. Frameworks for designing in-place graph algorithms. CoRR, abs/1711.09859, 2017. 7. Sankardeep Chakraborty, Venkatesh Raman, and Srinivasa Rao Satti. Biconnectivity, chain decomposition and st-numbering using O(n) bits. In Proc. 27th International Symposium on Algorithms and Computation (ISAAC 2016), volume 64 of LIPIcs, pages 22:1–22:13. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2016. 8. Timothy M. Chan, J. Ian Munro, and Venkatesh Raman. Selection and sorting in the “restore” model. In Proc. 25th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2014), pages 995–1004. SIAM, 2014. 9. Samir Datta, Raghav Kulkarni, and Anish Mukherjee. Space-Efficient Approximation Scheme for Maximum Matching in Sparse Graphs. In Proc. 41st International Symposium on Mathematical Foundations of Computer Science (MFCS 2016), volume 58 of LIPIcs, pages 28:1–28:12. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2016. 10. Jeff Edmonds, Chung Keung Poon, and Dimitris Achlioptas. Tight lower bounds for st-connectivity on the NNJAG model. SIAM J. Comput., 28(6):2257–2284, 1999. 11. Amr Elmasry, Torben Hagerup, and Frank Kammer. Space-efficient basic graph algorithms. In Proc. 32nd International Symposium on Theoretical Aspects of Computer Science (STACS 2015), volume 30 of LIPIcs, pages 288–301. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2015. 12. Amr Elmasry and Frank Kammer. Space-efficient plane-sweep algorithms. In Proc. 27th International Symposium on Algorithms and Computation (ISAAC 2016), volume 64 of LIPIcs, pages 30:1–30:13. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2016. 13. Greg N. Frederickson. Upper bounds for time-space trade-offs in sorting and selection. J. Comput. Syst. Sci., 34(1):19–26, 1987. 14. Torben Hagerup and Frank Kammer. Succinct choice dictionaries. Computing Research Repository (CoRR), arXiv:1604.06058 [cs.DS], 2016. 15. Torben Hagerup, Frank Kammer, and Moritz Laudahn. Space-efficient Euler partition and bipartite edge coloring. Theor. Comput. Sci., to appear, 2018. 16. Frank Kammer, Dieter Kratsch, and Moritz Laudahn. Space-Efficient Biconnected Components and Recognition of Outerplanar Graphs. In Proc. 41st International Symposium on Mathematical Foundations of Computer Science (MFCS 2016), volume 58 of LIPIcs, pages 56:1–56:14. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2016. 17. J. I. Munro and M. S. Paterson. Selection and sorting with limited storage. Theor. Comput. Sci., 12(3):315–323, 1980. 17 18. Jakob Pagter and Theis Rauhe. Optimal time-space trade-offs for sorting. In Proc. 39th Annual IEEE Symposium on Foundations of Computer Science (FOCS 1998), pages 264–268. IEEE Computer Society, 1998. 18
8cs.DS
arXiv:1512.01120v3 [math.AC] 12 May 2016 GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE (WITH AN APPENDIX BY DRISS BENNIS) LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG Abstract. For a commutative ring R and a faithfully flat R-algebra S we prove, under mild extra assumptions, that an R-module M is Gorenstein flat if and only if the left S-module S ⊗R M is Gorenstein flat, and that an R-module N is Gorenstein injective if and only if it is cotorsion and the left S-module HomR (S, N ) is Gorenstein injective. We apply these results to the study of Gorenstein homological dimensions of unbounded complexes. In particular, we prove two theorems on stability of these dimensions under faithfully flat (co-)base change. 1. Introduction Auslander and Bridger’s [2] notion of G-dimension for finitely generated modules over noetherian rings was generalized and dualized by Enochs and collaborators, who introduced the Gorenstein projective, Gorenstein injective, and Gorenstein flat dimension of modules over associative rings. In the treatment by Christensen, Frankild, and Holm [6], these invariants were considered for complexes with bounded homology. It is, however, possible to define Gorenstein projective dimension for unbounded complexes: This was done by Veliche [21], and the dual case of Gorenstein injective dimension was treated by Asadollahi and Salarian [1]. Inspired by [21], we propose a definition of Gorenstein flat dimension for unbounded complexes; it coincides with the one introduced by Iacob [15] whenever the latter is defined. Our main results are two theorems on stability of Gorenstein homological dimensions of complexes under faithfully flat change of base. 1.1 Theorem. Let R be a commutative coherent ring and let S be a faithfully flat R-algebra that is left GF-closed. For every R-complex M there is an equality GfdR M = GfdS (S ⊗L R M) . In particular, an R-module M is Gorenstein flat if and only if the S-module S ⊗R M is Gorenstein flat. This result is part of Theorem 6.7; the condition that S is left GF-closed is discussed in 5.10; it is satisfied if S is right coherent. Date: 12 May 2016. 2010 Mathematics Subject Classification. 13D05; 13D02. Key words and phrases. Gorenstein injective dimension, faithfully flat co-base change, Gorenstein flat dimension, faithfully flat base change. This research was partly supported by NSA grant H98230-14-0140 (L.W.C.), by NSFC grant 11301240 and SRF for ROCS, SEM (L.L.). Part of the work was done in the summer of 2015 when L.W.C. and L.L. visited Nanjing University supported by NSFC grant 11371187 (PI N.Q. Ding). 2 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG 1.2 Theorem. Let R be a commutative noetherian ring with splf R < ∞, i.e. every flat R-module has finite projective dimension, and let S be a faithfully flat R-algebra. For every R-complex N with Hi (N ) = 0 for all i ≫ 0 there is an equality GidR N = GidS RHomR (S, N ) . In particular, an R-module N is Gorenstein injective if and only if the S-module HomR (S, N ) is Gorenstein injective and ExtiR (S, N ) = 0 holds for all i > 0. This result is part of Theorem 4.12. The condition splf R < ∞ is discussed in 4.9; it is satisfied if R has finite Krull dimension or cardinality at most ℵn for some integer n > 0. We notice that the condition implies vanishing of ExtiR (S, N ) for i ≫ 0. ’ A recent result of Štovı́ček [20] implies that every Gorenstein injective module is cotorsion. It allows us to prove the next result, which appears as Theorem 4.5. 1.3 Theorem. Let R be commutative noetherian and let S be a faithfully flat R-algebra. An R-module N is Gorenstein injective if and only if it is cotorsion and the S-module HomR (S, N ) is Gorenstein injective. This result compares to the statement about modules in Theorem 1.2. By significantly relaxing the conditions on the rings, the results of this paper improve results obtained by Christensen and Holm [7], by Christensen and SatherWagstaff [11], and by Liu and Ren [22]. Details pertaining to Theorems 1.1 and 1.2 are given in Remarks 4.13 and 6.8; the trend is that the rings in [11, 22] are assumed to be commutative noetherian and, more often than not, of finite Krull dimension. The paper is organized as follows: In section 2 we set the notation and recall some background material. Sections 3–4 focus on the Gorenstein injective dimension, and Sections 5–6 deal with the Gorenstein flat dimension. Section 7 has some closing remarks and, finally, an appendix by Bennis answers a question raised in an earlier version of this paper. 2. Complexes Let R be a ring with identity. We consider only unitary R-modules, and we employ the convention that R acts on the left. That is, an R-module is a left R-module, and right R-modules are treated as modules over the opposite ring, denoted R◦ . 2.1 Complexes. Complexes of R-modules, R-complexes for short, is our object of study. Let M be an R-complex. With homological grading, M has the form M ∂i+1 ∂M · · · −→ Mi+1 −−−−→ Mi −−i−→ Mi−1 −→ · · · ; i M to get one switches to cohomological grading by setting M i = M−i and ∂M = ∂−i ∂ i−1 ∂i M −→ M i+1 −→ · · · . · · · −→ M i−1 −−− −→ M i −−M For n ∈ Z the symbol M>n denotes the quotient complex of M with (M>n )i = Mi for i > n and (M>n )i = 0 for i < n. The subcomplexes B(M ) and Z(M ) of boundaries and cycles, the quotient complex C(M ) of cokernels, and the subquotient complex H(M ) of homology all have GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 3 zero differentials; their modules are given by M Bi (M ) = Im ∂i+1 = −i−1 Im ∂M = B−i (M ) Zi (M ) = Ker ∂iM = −i Ker ∂M = Z−i (M ) Ci (M ) = Mi / Bi (M ) = M −i / B−i (M ) = C−i (M ) Hi (M ) = Zi (M )/ Bi (M ) = Z−i (M )/ B−i (M ) = H−i (M ) The complex M is called acyclic if H(M ) = 0, i.e. Hi (M ) = 0 holds for all i ∈ Z. 2.2 Morphisms. A morphism α : M → N of R-complexes is a family of R-linear maps {αi : Mi → Ni }i∈Z that commute with the differentials. That is, for all i ∈ Z one has αi−1 ∂iM = ∂iN αi . A morphism α : M → N maps boundaries to boundaries and cycles to cycles, so it induces a morphism H(α) : H(M ) → H(N ). If the induced morphism H(α) is bijective, then α is called a quasi-isomorphism. Such morphisms are marked by the symbol ‘≃’. 2.3 Resolutions. An R-complex I is called semi-injective if each module I i is injective, and the functor HomR (−, I) preserves quasi-isomorphisms. Dually, an Rcomplex P is called semi-projective if each module Pi is projective, and the functor HomR (P, −) preserves quasi-isomorphisms. An R-complex F is called semi-flat if each module Fi is flat, and the functor − ⊗R F preserves quasi-isomorphisms. Every semi-projective complex is semi-flat. For the following facts see Avramov and Foxby [3]. • Every R-complex M has a semi-projective resolution in the following sense: There is a quasi-isomorphism π : P → M , where P is a semi-projective complex; moreover, π can be chosen surjective. • Every R-complex M has a semi-injective resolution in the following sense: There is a quasi-isomorphism ι : M → I, where I is semi-injective; moreover, ι can be chosen injective. 2.4 The derived category. In the derived category over R, the objects are R≃ complexes, and the morphisms are equivalence classes of diagrams • ←−− • −→ • of morphisms of R-complexes. The isomorphisms in the derived category are classes represented by diagrams with two quasi-isomorphisms; and they are also marked by the symbol ‘≃’. The derived tensor product, − ⊗L R −, and the derived Hom, RHomR (−, −), are functors on the derived category; their values on given R-complexes are computed by means of the resolutions described above. As for modules one sets ExtiR (M, N ) = Hi (RHomR (M, N )) for R-complexes M and N and i ∈ Z. 3. Gorenstein injective modules and cotorsion For definitions and standard results on Gorenstein homological dimensions our references are Holm’s [14] and the monograph [12] by Enochs and Jenda. We start by recalling the definition [12, def. 10.1.1] of a Gorenstein injective module. 3.1. A complex U of injective R-modules is called totally acyclic if it is acyclic and the complex HomR (I, U ) is acyclic for every injective R-module I. 4 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG An R-module G is called Gorenstein injective (for short, G-injective) if there exists a totally acyclic complex U of injective R-modules with Z0 (U ) ∼ = G. 3.2. Every R-module has an injective resolution, so to prove that a module N is Gorenstein injective it suffices to verify the following: (1) ExtiR (I, N ) = 0 holds for all i > 0 and every injective R-module I. (2) N has a proper left injective resolution. That is, there exists an acyclic complex of R-modules U + = · · · → U −2 → U −1 → N → 0 with each U i injective, such that HomR (I, U + ) is acyclic for every injective R-module I. 3.3. An R-module X is called cotorsion if one has Ext1R (F, X) = 0 (equivalently, Ext>0 R (F, X) = 0) for every flat R-module F . ’ Štovı́ček [20, cor. 5,9] shows that for every module N that has a left injective resolution, one has ExtiR (F, N ) = 0 for all i > 0 and every R-module F of finite flat dimension. In particular, every Gorenstein injective module is cotorsion. This provides for the following improvement of [7, Ascent table II.(h)] 3.4 Lemma. Let R → S be a ring homomorphism such that fdR S and fdR◦ S are finite. For a Gorenstein injective R-module G, the S-module HomR (S, G) is Gorenstein injective. Proof. Let U be a totally acyclic complex of injective R-modules with Z0 (U ) ∼ = G. The complex HomR (S, U ) consists of injective S-modules, and it is acyclic as fdR S is finite; see 3.3. One has Z0 (HomR (S, U )) ∼ = HomR (S, G), so it remains to show that the complex ∼ HomR (J, U ) HomS (J, HomR (S, U )) = is acyclic for every injective S-module J. As fdR◦ S is finite, an injective S-module has finite injective dimension over R; see [3, cor. 4.2]. Acyclicity of HomR (J, U ) now follows from [6, lem. 2.2].  3.5. Let N be an R-complex. A complete injective resolution of N is a diagram ι υ N −→ I −−→ U where ι is a semi-injective resolution, υ i is bijective for all i ≫ 0, and U is a totally acyclic complex of injective R-modules. The Gorenstein injective dimension of an R-complex N is defined [1, def. 2.2] as   ι υ N −→ I −−→ U   GidR N = inf g ∈ Z is a complete injective resolution with .   υ i : I i → U i bijective for all i > g The next result is dual to (parts of) [21, thm. 3.4], and the proof is omitted.1 1 The statement [1, thm. 2.3] is also meant to be the dual of [21, thm. 3.4], but it has an unfortunate typographical error, reproduced in [22, lem. 1]: In parts (ii) and (iii) the kernel has been replaced by the cokernel in the same degree. GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 5 3.6 Proposition. Let N be an R-complex and g be an integer. The following conditions are equivalent. (i) GidR N 6 g. (ii) Hi (N ) = 0 holds for all i > g, and there exists a semi-injective resolution ≃ N −−→ I such that the module Zg (I) is Gorenstein injective. ≃ (iii) Hi (N ) = 0 holds for all i > g, and for every semi-injective resolution N −−→ I the module Zg (I) is Gorenstein injective.  The next result improves [22, cor. 9] by removing assumptions that R and S should be commutative noetherian with dim R < ∞. 3.7 Proposition. Let R → S be a ring homomorphism such that fdR S and fdR◦ S are finite. For every R-complex N one has GidS RHomR (S, N ) 6 GidR N . Proof. We may assume that H(N ) is non-zero, otherwise there is nothing to prove. ≃ Let N −−→ I be a semi-injective resolution; the S-complex HomR (S, I) is semiinjective and isomorphic to RHomR (S, N ) in the derived category. If GidR N is finite, say g, then by Proposition 3.6 one has Hi (I) = 0 for i > g and Z = Zg (I) is a G-injective R-module. Now 3.3 yields Hg+n (HomR (S, I)) = ExtnR (S, Z) = 0 for all n > 0. Moreover, the S-module Zg (HomR (S, I)) ∼ = HomR (S, Zg (I)) is G-injective by Lemma 3.4, so GidS RHomR (S, N ) 6 g holds by Proposition 3.6.  4. Faithfully flat co-base change Throughout this section, R is a commutative ring and S is an R-algebra. We are primarily concerned with the following setup. 4.1. Let S be a faithfully flat R-algebra; there is then a pure exact sequence (4.1.1) 0 −→ R −→ S −→ S/R −→ 0 ; that is, S/R is a flat R-module. Let I be an injective R-module. The induced sequences 0 −→ I −→ S ⊗R I −→ S/R ⊗R I −→ 0 and 0 −→ HomR (S/R, I) −→ HomR (S, I) −→ I −→ 0 are split exact so, as an R-module, I is a direct summand of S ⊗R I and of HomR (S, I). 4.2 Lemma. Let S be a faithfully flat R-algebra and let N be an R-module. If ExtiR (S, N ) = 0 holds for all i > 0, then the following conditions are equivalent. (i) ExtiR (I, N ) = 0 holds for all i > 0 and every injective R-module I. (ii) ExtiS (J, HomR (S, N )) = 0 holds for all i > 0 and every injective S-module J. Proof. Let X be an S-module; for every i > 0 there are isomorphisms ExtiS (X, HomR (S, N )) = Hi (RHomS (X, HomR (S, N ))) ∼ = Hi (RHomS (X, RHomR (S, N ))) ∼ = Hi (RHomR (S ⊗L X, N )) S ∼ = Hi (RHomR (X, N )) = ExtiR (X, N ) ; 6 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG >0 the first isomorphism follows by the vanishing of ExtR (S, N ), and the second is Hom-tensor adjointness in the derived category. Under the assumptions, every injective S-module is injective as an R-module, so it is evident from the computation that (i) implies (ii). For the converse, let I be an injective R-module and recall from 4.1 that it is a direct summand of the injective S-module HomR (S, I).  For later application, we recall a fact about cotorsion modules. 4.3. Let X be a cotorsion R-module. For every flat R-module F it follows by Hom-tensor adjointness that HomR (F, X) is cotorsion. We recall the notion of an injective precover, also known as an injective right approximation. 4.4. Let M be an R-module. A homomorphism ϕ : E → M is an injective precover of M , if E is an injective R-module and every homomorphism from an injective R-module to M factors through ϕ. Every R-module has an injective precover if and only if R is noetherian; see [12, thm. 2.5.1]. In the proof of the next theorem, the noetherian hypothesis on R is used to ensure the existence of injective precovers. 4.5 Theorem. Let R be commutative noetherian and let S be a faithfully flat R-algebra. An R-module N is Gorenstein injective if and only if it is cotorsion and the S-module HomR (S, N ) is Gorenstein injective. Proof. The “only if” part of the statement follows from 3.3 and Lemma 3.4. For the converse, note that Lemma 4.2 yields (1) Ext>0 R (I, N ) = 0 holds for every injective R-module I. To prove that N is G-injective, it is now sufficient to show that N has a proper left injective resolution; see 3.2. As N is cotorsion, application of HomR (−, N ) to (4.1.1) yields an exact sequence 0 → HomR (S/R, N ) → HomR (S, N ) → HomR (R, N ) → 0. There is also an exact sequence of S-modules 0 → G → U → HomR (S, N ) → 0 where U is injective and G is G-injective. The composite U → HomR (S, N ) → N is surjective. Since R is noetherian, there exists an injective precover ϕ : E → N . The module U is injective over R, so a surjective homomorphism factors through ϕ, whence ϕ is surjective. Now consider the commutative diagram 0  HomR (S/R, N ) (2) 0 // G // U 0  // K  // E  // HomR (S, N ) // 0  // N // 0 ϕ  0 GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 7 To construct a proper left injective resolution of N , it suffices to show that K has the same properties as N , including (1); that is, K is cotorsion, HomR (S, K) is G-injective over S, and Ext>0 R (I, K) = 0 holds for every injective R-module I. Injective S-modules are injective over R, so by 3.3 the G-injective S-module G is cotorsion over R. Further, HomR (S/R, N ) is cotorsion as S/R is flat; see 4.3. Let F be a flat R-module; application of HomR (F, −) to (2) yields a commutative diagram HomR (F, HomR (S, N )) // 0  HomR (F, N )  // Ext1R (F, K) // 0  0, from which one concludes Ext1R (F, K) = 0. That is, K is cotorsion. >0 Let I be an injective R-module. One has Ext>0 R (I, E) = 0 and ExtR (I, N ) = 0 by (1), so the exact sequence in cohomology associated to 0 → K → E → N → 0, yields Ext>1 R (I, K) = 0. Finally, HomR (I, −) leaves the sequence exact, because E → N is an injective precover, and Ext1R (I, K) = 0 follows. The exact sequence 0 → HomR (S, K) → HomR (S, E) → HomR (S, N ) → 0 shows that HomR (S, K) has finite Gorenstein injective dimension over S, at most 1. Lemma 4.2 yields Ext>0 R (J, HomR (S, K)) = 0 for every injective S-module J, so HomR (S, K) is G-injective by [14, thm. 2.22].  4.6 Lemma. Let R be commutative noetherian. Let 0 → L → Q → C → 0 be a pure exact sequence of R-modules with L 6= 0 free and Q faithfully flat. For an R-module N with ExtiR (Q, N ) = 0 for all i > 0 the next conditions are equivalent. (i) ExtiR (I, N ) = 0 holds for all i > 0 and every injective R-module I. (ii) ExtiR (I, HomR (Q, N )) = 0 holds for all i > 0 and every injective R-module I. Moreover, if N is cotorsion, then these conditions are equivalent to (iii) ExtiR (I, HomR (F, N )) = 0 holds for all i > 0, every injective R-module I, and every flat R-module F . Proof. Let F be a flat R-module with Ext>0 R (F, N ) = 0. For each integer i > 0 and R-module X one has ExtiR (X, HomR (F, N )) = Hi (RHomR (X, HomR (F, N ))) ∼ = Hi (RHomR (X, RHomR (F, N ))) ∼ = Hi (RHomR (F ⊗L X, N )) R ∼ = Hi (RHomR (F ⊗R X, N )) = ExtiR (F ⊗R X, N ) ; the first isomorphism holds by the assumption on F , the second by Hom-tensor adjointness in the derived category, and the third by flatness of F . Let I be an injective R-module. Since R is noetherian, the module F ⊗R I is injective, so it is immediate from the computation above that (i) implies (ii) and, if N is cotorsion, also the stronger statement (iii). 8 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG To show that (ii) implies (i), let I be an injective R-module and consider the exact sequence 0 → L ⊗R I → Q ⊗R I → C ⊗R I → 0. As R is noetherian, the module L ⊗R I is injective; thus the sequence splits, whence I is isomorphic to a direct summand of Q ⊗R I. The computation above yields >0 Ext>0 R (Q ⊗R I, N ) = 0, and ExtR (I, N ) = 0 follows as Ext functors are additive.  The next result improves [7, Ascent table I.(d)]. Part (iii) applies, in particular, to the setting where Q is a faithfully flat R-algebra; cf. (4.1.1). 4.7 Proposition. Let R be commutative noetherian. For an R-module N the following conditions are equivalent. (i) N is Gorenstein injective. (ii) HomR (F, N ) is Gorenstein injective for every flat R-module F . (iii) N is cotorsion and HomR (Q, N ) is Gorenstein injective for some faithfully flat R-module Q that contains a non-zero free R-module as a pure submodule. Proof. Assume first that N is G-injective and let U be a totally acyclic complex of injective modules with Z0 (U ) ∼ = N . Let F be a flat module. G-injective modules are cotorsion, see 3.3, so the complex HomR (F, U ) is acyclic. Moreover, it is a complex of injective modules, and for every injective module I the complex HomR (I, HomR (F, U )) ∼ = HomR (F ⊗R I, U ) is acyclic, as F ⊗R I is injective by the assumption that R is noetherian. This proves the implication (i) =⇒ (ii). That (ii) implies (iii) is trivial, just take F = R = Q and recall 3.3. To prove that (iii) implies (i), let Q be a faithfully flat R-module with a free pure submodule L 6= 0, and assume that HomR (Q, N ) is G-injective. In particular, Ext>0 R (I, HomR (Q, N )) = 0 holds for every injective R-module I, so by Lemma 4.6 (1) Ext>0 R (I, N ) = 0 holds for every injective R-module I. To prove that N is G-injective, it is now sufficient to show that it has a proper left injective resolution; see 3.2. To this end, consider the pure exact sequence (2) 0 −→ L −→ Q −→ C −→ 0 , and notice that the module C is flat as Q is flat. Since HomR (Q, N ) is G-injective, there is an exact sequence of R-modules 0 → G → E → HomR (Q, N ) → 0, where E is injective and G is G-injective. As N is cotorsion, applying HomR (−, N ) to (2) GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 9 yields another exact sequence, and the two meet in the commutative diagram, 0  HomR (C, N ) (3) 0 // G // E  // HomR (Q, N ) // 0 0  // H // E  // HomR (L, N ) // 0  0. The first step towards construction of a proper left injective resolution of N is to show that H has the same properties as N , including (1). Claim 1: The module H is cotorsion, HomR (Q, H) is G-injective, and one has >0 ExtR (I, H) = 0 for every injective R-module I. Proof: Since the module G is G-injective, it is also cotorsion, and so is the module HomR (C, N ) as C is flat; see 3.3 and 4.3. Let F be a flat R-module and apply HomR (F, −) to (3) to get the commutative diagram HomR (F, HomR (Q, N )) // 0  HomR (F, HomR (L, N ))  // Ext1R (F, H) // 0  0, which shows that Ext1R (F, H) vanishes, whence H is cotorsion. Notice that one has HomR (L, N ) ∼ = N Λ for some index set Λ. Let I be an injec>0 tive module; by (1) one has ExtR (I, N Λ ) = 0, so the exact sequence in cohomology induced by the last non-zero row of (3) yields Ext>1 R (I, H) = 0. Now we argue that 1 also ExtR (I, H) vanishes. As G is G-injective, one has Ext>0 R (I, G) = 0; by the assumptions on N and (1), Lemma 4.6 applies to yield Ext>0 R (I, HomR (C, N )) = 0. Applying HomR (I, −) to (3) one thus gets the commutative diagram (4) HomR (I, HomR (Q, N )) // 0  HomR (I, HomR (L, N ))  // Ext1R (I, H) // 0  0, which yields Ext1R (I, H) = 0. It now follows from Lemma 4.6 that Ext>0 R (I, HomR (Q, H)) = 0 holds for every injective R-module I. To prove that the module HomR (Q, H) is G-injective it is thus, by [14, thm. 2.22], enough to show that it has finite Gorenstein injective 10 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG dimension. However, that is immediate as application of HomR (Q, −) to last nonzero row of (3) yields the exact sequence 0 −→ HomR (Q, H) −→ HomR (Q, E) −→ HomR (Q, N Λ ) −→ 0 , where HomR (Q, E) is injective, and HomR (Q, N Λ ) ∼ = HomR (Q, N )Λ is G-injective; see [14, thm. 2.6]. This finishes the proof of Claim 1. In the commutative diagram below, the second non-zero row is the last non-zero row from (3), and π is a canonical projection. 0  ′ NΛ (5) 0 // H // E  // N Λ 0  // K // E  // N // 0 π // 0  0 To construct a proper left injective resolution of N , it is sufficient to prove that K has the same properties as N , including (1). Claim 2: The module K is cotorsion, HomR (Q, K) is G-injective, and one has >0 ExtR (I, K) = 0 for every injective R-module I. Proof: Apply the Snake Lemma to the diagram (5) to get the exact sequence ′ 0 → H → K → N Λ → 0. By assumption N is cotorsion, and H is cotorsion by Claim 1; it follows that K is cotorsion. Moreover, by (1) and Claim 1 one has >0 >0 (I, H) = 0 = Ext>0 ExtR R (I, N ) and hence ExtR (I, K) = 0 for every injective module I. The induced sequence ′ 0 −→ HomR (Q, H) −→ HomR (Q, K) −→ HomR (Q, N )Λ −→ 0 ′ is exact, as H is cotorsion. The modules HomR (Q, N )Λ and HomR (Q, H) are G-injective by assumption and Claim 1, so by [14, thm. 2.6] also HomR (Q, K) is G-injective. That finishes the proof of Claim 2.  4.8 Corollary. Let R be commutative noetherian. For every flat R-module F and for every R-complex N there is an inequality GidR RHomR (F, N ) 6 GidR N . Proof. We may assume that H(N ) is non-zero, otherwise there is nothing to prove. ≃ Let N −−→ I be a semi-injective resolution and assume that GidR N = g is finite. By Proposition 3.6 the module Z = Zg (I) is G-injective, and arguing as in the proof of 3.7 one gets Hi (RHomR (F, N )) = 0 for all i > g. The complex, HomR (F, I) is semi-injective and isomorphic to RHomR (F, N ) in the derived category. From Proposition 4.7 it follows that the module Zg (HomR (F, I)) ∼ = HomR (F, Zg (I)) is G-injective, and then Proposition 3.6 yields GidR RHomR (F, N ) 6 g.  4.9. For the statement of the next corollary we recall the invariant splf R = sup{proj.dimR F | F is a flat R-module} . GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 11 If R is noetherian of finite Krull dimension d, then one has splf R 6 d by works of Jensen [16, prop. 6], and Gruson and Raynaud [19, thm. II.(3.2.6)]. If R has cardinality at most ℵn for some integer n, then a theorem of Gruson and Jensen [13, thm. 7.10] yields splf R 6 n + 1. Osofsky [18, 3.1] gives examples of rings for which the splf invariant is infinite. Since an arbitrary direct sum of flat R-modules is flat, the invariant splf R is finite if and only if every flat R-module has finite projective dimension. 4.10 Proposition. Let R be commutative noetherian with splf R < ∞. Let S be a faithfully flat R-algebra and N be an R-module with ExtiR (S, N ) = 0 for all i > 0. If HomR (S, N ) is Gorenstein injective as an R-module or as an S-module, then N is Gorenstein injective. ≃ Proof. Set n = splf R. Let N −−→ E be an injective resolution. As every flat R-module has projective dimension at most n, it follows by dimension shifting that the cosyzygy Z = Zn (E) is cotorsion. Vanishing of Ext>0 R (S, N ) implies that HomR (S, −) leaves the sequence 0 → N → E 0 → · · · → E n−1 → Z → 0 exact. The class of G-injective modules is injectively resolving; see [14, thm. 2.6]. If HomR (S, N ) is G-injective over R, it thus follows that HomR (S, Z) is G-injective over R. From Proposition 4.7 it follows that Z is G-injective over R, so N has finite Gorenstein injective dimension. Finally, Lemma 4.6 yields Ext>0 R (I, N ) = 0 for every injective R-module I, so N is G-injective by [14, thm. 2.22]. If HomR (S, N ) is G-injective over S, then it follows as above that HomR (S, Z) is G-injective over S. From Theorem 4.5 it follows that Z is G-injective over R, so N has finite Gorenstein injective dimension. Finally, it follows from Lemma 4.2 that  Ext>0 R (I, N ) = 0 holds for every injective R-module I, so N is G-injective. 4.11 Lemma. Let R be commutative noetherian and let S be a faithfully flat R-algebra. For every R-complex N of finite Gorenstein injective dimension one has GidS RHomR (S, N ) = GidR N = GidR RHomR (S, N ) . Proof. We may assume that H(N ) is non-zero, otherwise there is nothing to prove. Set g = GidR N ; Proposition 3.7 and Corollary 4.8 yield GidS RHomR (S, N ) 6 g ≃ and GidR RHomR (S, N ) 6 g. Let N −−→ I be a semi-injective resolution; the complex HomR (S, I) is isomorphic to RHomR (S, N ) in the derived category, and it is semi-injective as an R-complex and as an S-complex. By [1, thm. 2.4] there exists an injective R-module E with ExtgR (E, N ) 6= 0. As an R-module, E is a direct summand of the injective S-module HomR (S, E), see 4.1, so one gets Extg (HomR (S, E), RHomR (S, N )) ∼ = Hg (HomS (HomR (S, E), HomR (S, I))) S ∼ = Hg (HomR (HomR (S, E), I)) ∼ Extg (HomR (S, E), N ) 6= 0 , = R and hence GidS RHomR (S, N ) > g by [1, thm. 2.4]; i.e. equality holds. As an R-module, E is a direct summand of S ⊗R E, see 4.1, so one gets Extg (E, RHomR (S, N )) ∼ = Hg (HomR (E, HomR (S, I))) R ∼ = Hg (HomR (S ⊗R E, I)) ∼ Extg (S ⊗R E, N ) 6= 0 , = R and hence GidR RHomR (S, N ) > g by [1, thm. 2.4]; i.e. equality holds.  12 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG 4.12 Theorem. Let R be commutative noetherian with splf R < ∞ and let S be a faithfully flat R-algebra. For every R-complex N with Hi (N ) = 0 for all i ≫ 0 there are equalities GidS RHomR (S, N ) = GidR N = GidR RHomR (S, N ) . Proof. In view of Lemma 4.11 it is sufficient to prove that GidR N is finite if RHomR (S, N ) has finite Gorenstein injective dimension as an R-complex or as an S-complex. Assume, without loss of generality, that Hi (N ) = 0 holds for i > 0, and ≃ let N −−→ I be a semi-injective resolution. The complex HomR (S, I) is isomorphic to RHomR (S, N ) in the derived category, and it is semi-injective as an R-complex and as an S-complex. If GidS RHomR (S, N ) is finite, then one has GidS RHomR (S, N ) 6 g for some g > 0. By Proposition 3.6 the S-module Zg (HomR (S, I)) ∼ = HomR (S, Zg (I)) is Gn g g+n injective and one has ExtR (S, Z (I)) = H (HomR (S, I)) = 0 for all n > 0. Now follows from Proposition 4.10 that Zg (I) is G-injective, so GidR N 6 g holds. The same argument applies if GidR RHomR (S, N ) is finite.  The equivalent of this theorem in absolute homological algebra is [10, thm. 2.2] about injective dimension of a complex N . In that statement, the a priori assumption of vanishing of Hi (N ) for i ≫ 0 is missing, though it is applied in the proof in the same manner as above. However, [10, thm. 2.2] is correct as stated, and an even stronger result is proved in [9, cor. 3.1]. 4.13 Remark. Theorem 4.5 improves [11, lem. 1.3.(a)] by removing assumptions that S should be commutative and noetherian with dim S < ∞. The first equality in Theorem 4.12 compares to the equality in [11, thm. 1.7] as follows: There is no assumption of finite Krull dimension of R, and S is not assumed to be commutative nor noetherian. The inequality in [11, thm. 1.7] is subsumed by [22, cor. 9] and compares to Proposition 3.7 as discussed there. In [22, thm. 11] the first equality in Theorem 4.12 is proved without conditions on the homology of N but under the assumption that S is finitely generated as an R-module. Note that this assumption implies that R is an algebra retract of S. 5. Gorenstein flat dimension Recall the definition [12, def. 10.3.1] of a Gorenstein flat module. 5.1. A complex T of flat R-modules is called totally acyclic if it is acyclic and the complex I ⊗R T is acyclic for every injective R◦ -module I. An R-module G is called Gorenstein flat (for short, G-flat) if there exists a totally acyclic complex T of flat R-modules with C0 (T ) ∼ = G. 5.2. Every module has a projective resolution, so to prove that an R-module M is Gorenstein flat it suffices to verify the following: ◦ (1) TorR i (I, M ) = 0 holds for all i > 0 and every injective R -module I. (2) There exists an acyclic complex T + = 0 → M → T−1 → T−2 → · · · of Rmodules with each Ti flat, such that I ⊗R T + is acyclic for every injective R◦ -module I. The next result is a non-commutative version of [7, Ascent table I.(a)]. GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 13 5.3 Lemma. Let R → S be a ring homomorphism such that fdR S and fdR◦ S are finite. For a Gorenstein flat R-module G, the S-module S ⊗R G is Gorenstein flat. Proof. Let T be a totally acyclic complex of flat R-modules with C0 (T ) ∼ = G. The complex S ⊗R T consists of flat S-modules, and it is acyclic as fdR◦ S is finite; see [6, lem. 2.3]. One has C0 (S ⊗R T ) ∼ = S ⊗R G, so it suffices to show that the complex J ⊗S (S ⊗R T ) ∼ = J ⊗R T is acyclic for every injective S ◦ -module J. As fdR S is finite, every injective S ◦ -module has finite injective dimension over R◦ ; see [3, cor. 4.2]. Acyclicity of J ⊗R T hence follows from [6, lem. 2.3].  In standard homological algebra, the study of flat modules and flat dimension is propelled by flat–injective duality: An R-module F is flat if and only if the R◦ module HomZ (F, Q/Z) is injective. A similar Gorenstein flat–injective duality is only known hold if R is right coherent, so that is a convenient setting for studies of Gorenstein flat dimension; see for example [14, thm. 3.14]. As Bennis demonstrates [4, thm. 2.8], one can do a little better and get a wellbehaved theory of Gorenstein flat dimension, as long as the class of Gorenstein flat modules is projectively resolving. Rings with that property have become known as left GF-closed; we recall the definition in 5.10. Iacob [15] defined a notion of Gorenstein flat dimension for unbounded complexes, but only for complexes over left GF-closed rings. Here we give a definition, inspired by Veliche’s [21, def. 3.1]; it applies to complexes over any ring and coincides with Iacob’s for left GF-closed rings. To get started we recall some terminology from Liang [17]. 5.4. Let M be an R-complex; a semi-flat replacement of M is a semi-flat R-complex F such that there is an isomorphism F ≃ M in the derived category. Every complex has a semi-projective resolution and hence a semi-flat replacement; see 2.3. A Tate flat resolution of M is a pair (T, F ) where T a totally acyclic complex of flat Rmodules and F ≃ M a semi-flat replacement with T>g ∼ = F>g for some g ∈ Z. 5.5 Definition. Let M be an R-complex. The Gorenstein flat dimension of M is given by GfdR M = inf{g ∈ Z | (T, F ) is a Tate flat resolution of M with T>g ∼ = F>g } . 5.6. Let M be an R-complex; the following facts are evident from the definition. (a) One has GfdR M < ∞ if and only if M admits a Tate flat resolution. (b) M is acyclic if and only if one has GfdR M = −∞. (c) GfdR (Σn M ) = GfdR M + n holds for every n ∈ Z.2 A Gorenstein flat R-module M 6= 0 clearly has has GfdR M = 0 per Definition 5.5. In Remark 5.13 we compare 5.5 to competing definitions in the literature. 5.7 Proposition. Let M be an R-complexes and g ∈ Z. The following conditions are equivalent. (i) GfdR M 6 g. (ii) Hi (M ) = 0 holds for all i > g, and there exists a semi-flat replacement F ≃ M such that Ci (F ) is Gorenstein flat for every i > g. 2 Here Σn M is the complex with (Σn M ) = M Σn M = (−1)n ∂ M for all i ∈ Z. i i−n and ∂i i−n 14 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG (iii) Hi (M ) = 0 holds for all i > g, and there exists a semi-flat replacement F ≃ M such that Cg (F ) is Gorenstein flat. Proof. If GfdR M 6 g holds, then there exists a Tate flat resolution (T, F ) of M with T>g ∼ = F>g . As T is acyclic and F ≃ M one has Hi (M ) ∼ = Hi (F ) ∼ = Hi (T ) = 0 ∼ for all i > g, and Ci (F ) = Ci (T ) is G-flat for every i > g. Thus (i) implies (ii). Clearly, (ii) implies (iii); to finish the proof assume that Hi (M ) = 0 holds for all i > g and that there exists a semi-flat replacement F ≃ M such that Cg (F ) is G-flat. As one has H(F ) ∼ = H(M ) it follows that the complex F ′ = · · · −→ Fg+1 −→ Fg −→ Cg (F ) −→ 0 is acyclic; i.e. it is an augmented flat resolution of Cg (F ). For every injective ′ R◦ -module I one has TorR >0 (I, Cg (F )) = 0, so the complex I ⊗R F is acyclic. Moreover, there exists an acyclic complex of R-modules T ′ = 0 −→ Cg (F ) −→ Tg−1 −→ Tg−2 −→ · · · with each Ti flat, such that I ⊗R T ′ is acyclic for every injective R◦ -module I. Now let T be the complex obtained by splicing together F ′ and T ′ at Cg (F ); that is T = · · · −→ Fg+1 −→ Fg −→ Tg−1 −→ Tg−2 −→ · · · . It is a totally acyclic complex of flat R-modules, so (T, F ) is a Tate flat resolution  of M with T>g ∼ = F>g , and so one has GfdR M 6 g. In the next statement ‘fd’ means flat dimension as defined in [3, sec. 2.F]. 5.8 Proposition. Let M be an R-complex; the following assertions hold. (a) One has GfdR M 6 fdR M ; equality holds if fdR M < ∞. (b) One has GidR◦ HomZ (M, Q/Z) 6 GfdR M ; equality holds if R is right coherent. Proof. (a): For an acyclic R-complex M one has GfdR M = −∞ = fdR M , and the inequality is trivial if fdR M = ∞. Now assume that H(M ) 6= 0 and fdR M = g holds for some g ∈ Z. It follows from [3, thm. 2.4.F] that Hi (M ) = 0 holds for all i > g, and that there exists a semi-flat replacement F ≃ M such that Cg (F ) is flat, in particular G-flat. Thus one has GfdR M 6 g by Proposition 5.7. To prove that equality holds, assume towards a contradiction that GfdR M 6 g − 1 holds. One then has Hi (M ) = 0 for all i > g −1 and there exists a semi-flat replacement F ≃ M such that Cg−1 (F ) is G-flat; see Proposition 5.7. Consider the exact sequence 0 −→ Cg (F ) −→ Fg−1 −→ Cg−1 (F ) −→ 0 . The module Cg (F ) is flat, so one has fdR (Cg−1 (F )) ≤ 1, and then Cg−1 (F ) is flat; see Bennis [5, thm. 2.2]. Thus one has fdR M ≤ g − 1, which is a contradiction. (b): For an acyclic R-complex M one has GidR◦ HomZ (M, Q/Z) = −∞ = GfdR M , and the inequality is trivial if GfdR M = ∞. Now assume that H(M ) 6= 0 and GfdR M = g holds for some g ∈ Z. By Proposition 5.7 one has Hi (M ) = 0 for all i > g, and there exists a semi-flat replacement F ≃ M such that Cg (F ) is G-flat. The R◦ -complex HomZ (F, Q/Z) is semi-injective and yields a semi-injective ≃ resolution HomZ (M, Q/Z) −−→ HomZ (F, Q/Z); see [3, 1.4.I]. For all i > g one has Hi (HomZ (M, Q/Z)) ∼ = HomZ (Hi (M ), Q/Z) = 0 , and by [14, prop. 3.11] the R◦ -module Zg (HomZ (F, Q/Z)) ∼ = HomZ (Cg (F ), Q/Z) is G-injective, so one gets GidR◦ HomZ (M, Q/Z) 6 g from Proposition 3.6. GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 15 Assume now that g = GidR◦ HomZ (M, Q/Z) is finite. Let F ≃ M be a semi-flat replacement; as above it yields a semi-injective resolution ≃ HomZ (M, Q/Z) −−→ HomZ (F, Q/Z) . By Proposition 3.6 one has HomZ (Hi (M ), Q/Z) ∼ = Hi (HomZ (M, Q/Z)) = 0 for all g ∼ i > g and the module HomZ (Cg (F ), Q/Z) = Z (HomZ (F, Q/Z)) is G-injective. It follows that Hi (M ) = 0 holds for all i > g, and under the assumption that R is right coherent it follows from [14, prop. 3.11] that Cg (F ) is G-flat. Thus one has  GfdR M 6 g by Proposition 5.7, and GidR◦ HomZ (M, Q/Z) = GfdR M holds. The next result is proved by standard applications of the Snake Lemma. 5.9 Lemma. Let 0 → X → Y → Z → 0 be an exact sequence of R-complexes and let n ∈ Z. The following assertions hold. (a) If is (b) If is Hn (Z) = 0, then the sequence 0 → Cn−1 (X) → Cn−1 (Y ) → Cn−1 (Z) → 0 exact. Moreover if Hn (Y ) = 0, then the converse holds. Hn (X) = 0, then the sequence 0 → Zn+1 (X) → Zn+1 (Y ) → Zn+1 (Z) → 0 exact. Moreover if Hn (Y ) = 0, then the converse holds.  5.10. Recall from [4] that R is called left GF-closed if the class of Gorenstein flat R-modules is closed under extensions (equivalently, it is projectively resolving). The class of left GF-closed rings strictly contains the class of right coherent rings. The next result appears as [15, lem. 3.2] with the additional hypothesis that R is left GF-closed; that hypothesis is, however, not needed. 5.11 Lemma. Let 0 → F → G → H → 0 be an exact sequence of R-modules. If F is flat, G is Gorenstein flat, and HomZ (H, Q/Z) is Gorenstein injective, then H is Gorenstein flat. Proof. It is argued in the proof of [15, lem. 3.2] that there is an exact sequence 0 → H → X → Y → 0 with X flat and Y G-flat, so H is G-flat by [4, lem. 2.4].  5.12 Proposition. Let M be an R-complex and g ∈ Z. Conditions (i) and (ii) below are equivalent and imply (iii). Moreover, if R is left GF-closed, then all three conditions are equivalent. (i) Hi (M ) = 0 holds for all i > g, and for every semi-flat replacement F ≃ M the cokernel Cg (F ) is Gorenstein flat. (ii) Hi (M ) = 0 holds for all i > g, and there exists a semi-projective resolution ≃ P −−→ M such that the cokernel Cg (P ) is Gorenstein flat. (iii) GfdR M 6 g. Proof. Clearly (i) implies (ii), and (ii) implies (iii) by Proposition 5.7. ≃ (ii) =⇒ (i): Let P −−→ M be a semi-projective resolution such that Cg (P ) is G-flat. Let F ≃ M be a semi-flat replacement. There is by [3, 1.4.P] a quasi≃ isomorphism ϕ : P −−→ F , and after adding to P a projective precover of F , we may assume that ϕ is surjective. One then has an exact sequence 0 → K → P → F → 0. The kernel K is semi-flat and acyclic, so the module Ci (K) is flat for every i ∈ Z; see Christensen and Holm [8, thm. 7.3]. The sequence (1) 0 −→ Cg (K) −→ Cg (P ) −→ Cg (F ) −→ 0 16 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG is exact by Lemma 5.9(a). The module Cg (P ) is G-flat, so to prove that Cg (F ) is G-flat, it is by Lemma 5.11 enough to prove that HomZ (Cg (F ), Q/Z) is Ginjective. By Propositions 5.7 and 5.8(b) one has GidR◦ HomZ (M, Q/Z) 6 g. The R◦ -complex HomZ (F, Q/Z) is semi-injective and yields a semi-injective resolution ≃ HomZ (M, Q/Z) −−→ HomZ (F, Q/Z); see [3, 1.4.I]. It now follows from Proposition 3.6 that the module HomZ (Cg (F ), Q/Z) ∼ = Zg (HomZ (F, Q/Z)) is G-injective. Thus, the conditions (i) and (ii) are equivalent. We now assume that R is left GF-closed and verify the implication (iii) =⇒ (ii). By Proposition 5.7 one has Hi (M ) = 0 for all i > g, and there exists a semi-flat replacement F ≃ M with ≃ Cg (F ) G-flat. Choose a surjective semi-projective resolution P −−→ F ; it yields by ≃ [3, 1.4.P] a semi-projective resolution P −−→ M . As above one gets an sequence (1), where now Cg (K) is flat and Cg (F ) is G-flat; it follows that Cg (P ) is G-flat as R is left GF-closed.  For commentary on the proof, see Remark 5.16 at the end of the section. 5.13 Remark. From Proposition 5.12 and [15, thm. 1] it is clear that for complexes over a left GF-closed ring, Definition 5.5 agrees with Iacob’s definition [15, def. 3.2]. In this setting, it thus extends the definitions in [14, def. 3.9] and [6, 1.9] of Gorenstein flat dimension for modules and complexes with bounded below homology; see [15, rmk. 3]. Though they potentially differ over rings that are not GF-closed, we use the same symbol, ‘Gfd’, for our notion and the one from [6, 1.9]. It causes no ambiguity, as the results we prove in the remainder of this paper are valid for either notion. When needed, we make remarks to that effect. The next result improves [22, thm. 12] by removing assumptions that R and S should be commutative noetherian. 5.14 Proposition. Let R → S be a ring homomorphism such that fdR S and fdR◦ S are finite. For every R-complex M one has GfdS (S ⊗L R M ) 6 GfdR M . Proof. We may assume that H(M ) is non-zero, otherwise there is nothing to prove. Assume that GfdR M = g holds for some g ∈ Z. By Proposition 5.7 one has Hi (M ) = 0 for all i > g, and there exists a semi-flat replacement F ≃ M with Cg (F ) G-flat. The S-complex S ⊗R F is semi-flat and isomorphic to S ⊗L R M in the derived category. For n > 0 one has Hg+n (S ⊗L M ) ∼ = TorR (S, Cg (F )) = 0 , = Hg+n (S ⊗R F ) ∼ R n where vanishing follows from [6, lem. 2.3] as fdR◦ S is finite. Finally, the module Cg (S ⊗R F ) ∼ = S ⊗R Cg (F ) is G-flat by Lemma 5.3, so GfdS (S ⊗L R M ) 6 g holds by Proposition 5.7.  5.15 Remark. For an R-complex M with Hi (M ) = 0 for i ≪ 0 the inequality in Proposition 5.14 is also valid with the definition of Gorenstein flat dimension from [6, 1.9]. Indeed, the argument for [7, (4.4)] also applies when R is not commutative. 5.16 Remark. From the proof of (iii) =⇒ (ii) in Proposition 5.12, consider the exact sequence (5.16.1) 0 −→ Cg (K) −→ Cg (P ) −→ Cg (F ) −→ 0 GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 17 where the module Cg (K) is flat and Cg (F ) is Gorenstein flat. Under the assumption that R is left GF-closed, the module in the middle, Cg (P ), is Gorenstein flat. Without that assumption, the best we know is to consider the pull-back diagram 0 0  G  G 0 // Cg (K)  // X  // T // 0 0 // Cg (K)  // Cg (P )  // Cg (F ) // 0  0  0 where T is flat and G is Gorenstein flat. It follows that X is flat, so Cg (P ) has Gorenstein flat dimension at most 1, per Definition 5.5 as well as per [14, def. 3.9]. Notice that the module Cg (K) in (5.16.1) is not just Gorenstein flat but flat. In an earlier version of this paper we asked if a ring is GF-closed provided that in every exact sequence of R-modules 0 → F → M → G → 0 with F flat and G Gorenstein flat also M is Gorenstein flat. A positive answer to this question is provided in the Appendix by Driss Bennis. 6. Faithfully flat base change In this section, R is commutative and S is an R-algebra. 6.1 Lemma. Let S be a faithfully flat R-algebra. For an R-module M the following conditions are equivalent. (i) TorR i (I, M ) = 0 holds for all i > 0 and every injective R-module I. (ii) TorSi (J, S ⊗R M ) = 0 holds for all i > 0 and every injective S ◦ -module J. Proof. For every S ◦ -module X and every i > 0 one has R L L ∼ ∼ TorSi (X, S ⊗R M ) ∼ = Hi (X ⊗L S (S ⊗R M )) = Hi (X ⊗R M ) = Tori (X, M ) . As S is flat over R, every injective S ◦ -module is an injective R-module, so (i) implies (ii). Let I be an R-module and recall from 4.1 that it is a direct summand of the injective S ◦ -module HomR (S, I); now (i) follows from (ii).  For use in the next proof, we recall the notion of a flat preenvelope, also known as a flat left approximation. 6.2. Let M be an R-module. A homomorphism ϕ : M → F is a flat preenvelope of M , if F is a flat R-module and every homomorphism from M to a flat R-module factors through ϕ. Every R-module has a flat preenvelope if and only if R is coherent; see [12, thm. 5.4.1]. 6.3 Lemma. Let R be coherent and let S be a faithfully flat R-algebra that is left GF-closed. An R-module M is Gorenstein flat if and only if the S-module S ⊗R M is Gorenstein flat. 18 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG Proof. The “only if” part is a special case of Lemma 5.3. To prove “if”, assume that S ⊗R M is a G-flat S-module. Now TorS>0 (J, S ⊗R M ) = 0 holds for every injective S ◦ -module J, so by Lemma 6.1 one has (1) TorR >0 (I, M ) = 0 for every injective R-module I. Thus, 5.2.(1) holds and we proceed to verify 5.2.(2). Let ϕ : M → F be a flat preenvelope; it exists because R is coherent. The Gflat module S ⊗R M embeds into a flat S-module, and M embeds into S ⊗R M as S is faithfully flat; cf. (4.1.1). A flat S-module is also flat as an R-module, so M embeds into a flat R-module, and it follows that ϕ is injective. Consider the associated exact sequence ϕ η = 0 −→ M −−→ F −→ C −→ 0 . Let I be an injective R-module. The induced sequence I ⊗R η is exact if and only if HomZ (I ⊗R η, Q/Z) ∼ = HomR (η, HomZ (I, Q/Z)) is exact. The R-module HomZ (I, Q/Z) is flat, as R is coherent, so exactness of HomR (η, HomZ (I, Q/Z)) holds because ϕ is a flat preenvelope. Notice that exactness of I ⊗R η and (1) yield (2) TorR >0 (I, C) = 0 for every injective R-module I. To construct the complex in 5.2.(2), it is now sufficient to show that S ⊗R C is Gflat over S. Consider the following push-out diagram in the category of S-modules where H is flat and G is G-flat. 0 0 0  // S ⊗R M  // S ⊗R F // S ⊗R C // 0 0  // H  // X // S ⊗R C // 0  G  G  0  0 Since S is left GF-closed, the module X is G-flat. To conclude from Lemma 5.11, applied to the diagram’s second non-zero row, that S ⊗R C is G-flat, we need to verify that the S ◦ -module HomZ (S ⊗R C, Q/Z) is G-injective. Application of HomZ (−, Q/Z) to the second row shows that the module has finite Gorenstein injective dimension (at most 1). Let J be an injective S ◦ -module and i > 0 be an integer. The isomorphism below is Hom-tensor adjointness, and vanishing follows from Lemma 6.1 in view of (2) ExtiS ◦(J, HomZ (S ⊗R C, Q/Z)) ∼ = HomZ (TorSi (J, S ⊗R C), Q/Z) = 0 . It follows that HomZ (S ⊗R C, Q/Z) is G-injective; see [14, thm. 2.22].  It is straightforward to verify that for a flat R-module F and a Gorenstein flat R-module G, the module F ⊗R G is Gorenstein flat; see [7, Ascent table I.(a)]. The next lemma provides a partial converse; it applies, in particular, to the setting where Q is a faithfully flat R-algebra; cf. (4.1.1). GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 19 6.4 Proposition. Let R be coherent and M be an R-module. If Q ⊗R M is Gorenstein flat for some faithfully flat R-module Q that contains a non-zero free R-module as a pure submodule, then M is Gorenstein flat. Proof. As Q is faithfully flat, vanishing of TorR i (I, Q ⊗R M ) implies vanishing of R TorR i (I, M ), so one has Tor>0 (I, M ) = 0 for every injective R-module I. Thus 5.2.(1) is satisfied and we proceed to construct the complex in 5.2.(2). Let L 6= 0 be a free pure submodule of Q; application of − ⊗R M to the associated pure embedding yields an embedding L ⊗R M → Q ⊗R M . By assumption, the module Q ⊗R M embeds into a flat R-module, whence L ⊗R M , and therefore M , embeds into a flat R-module. Since R is coherent, there exists a flat preenvelope, ϕ : M → F , and it is necessarily injective. Consider the associated exact sequence 0 → M → F → C → 0. As in the proof of Lemma 6.3 one gets (1) TorR >0 (I, C) = 0 for every injective R-module I. To construct the complex in 5.2.(2) it is now sufficient to show that Q ⊗R C is G-flat. It is immediate from the induced exact sequence 0 −→ Q ⊗R M −→ Q ⊗R F −→ Q ⊗R C −→ 0 that Q ⊗R C has finite Gorenstein flat dimension. For every injective R-module I one gets TorR >0 (I, Q ⊗R C) = 0 from (1), so Q ⊗R C is G-flat by [14, thm. 3.14].  6.5 Proposition. For every flat R-module F and for every R-complex M there is an inequality GfdR (F ⊗R M ) 6 GfdR M. Proof. We may assume that H(M ) is non-zero, otherwise there is nothing to prove. Assume that GfdR M = g holds for some g ∈ Z. By Proposition 5.7 one has Hi (M ) = 0 for all i > g, and there exists a semi-flat replacement P ≃ M with Cg (P ) G-flat. The complex F ⊗R P is semi-flat and isomorphic to F ⊗R M in the derived category. For i > g one has Hi (F ⊗R M ) ∼ = F ⊗R Hi (M ) = 0, and the module Cg (F ⊗R P ) ∼ F ⊗ C (P ) is G-flat by [7, Ascent table I.(a)], so = R g GfdR (F ⊗R M ) 6 g holds by Proposition 5.7.  6.6 Remark. For an R-complex M with Hi (M ) = 0 for i ≪ 0 the inequality in Proposition 6.5 is by [7, (4.4)] valid for Gorenstein flat dimension defined in [6, 1.9]. 6.7 Theorem. Let R be commutative coherent, let S be a faithfully flat R-algebra, and let M be an R-complex. There is an equality GfdR M = GfdR (S ⊗R M ) , and if S is left GF-closed also an equality GfdS (S ⊗R M ) = GfdR M . Proof. We may assume that H(M ) is non-zero, otherwise there is nothing to prove. Consider the first equality. The inequality “>” is a special case of Proposition 6.5. For the opposite inequality, assume GfdR (S ⊗R M ) = g holds for some g ∈ Z. ≃ Choose a semi-projective resolution P −−→ M ; the induced quasi-isomorphism S ⊗R P → S ⊗R M is a semi-flat replacement. From Proposition 5.12 it follows 20 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG that S ⊗R Hi (M ) ∼ = Hi (S ⊗R M ) = 0 holds for all i > g and that the module S ⊗R Cg (P ) ∼ = Cg (S ⊗R P ) is G-flat. Since S is faithfully flat, it follows that Hi (M ) = 0 holds for all i > g, and by Proposition 6.4 the module Cg (P ) is G-flat. By Proposition 5.12 one now has GfdR M 6 g, and this proves the first equality. A parallel argument applies to establish the second equality; only one has to invoke Proposition 5.14 instead of 6.5 and Lemma 6.3 instead of 6.4.  6.8 Remark. The first equality in Theorem 6.7 improves the second equality in [11, prop. 1.9] by removing the requirements that R be semi-local and S be commutative noetherian. The equality in [11, thm. 1.8] are subsumed by [22, thm. 15], and the second equality in Theorem 6.7 improves [22, thm. 15] by removing the assumption that S is commutative and by relaxing the conditions on R and S from noetherian to coherent/left GF-closed. 7. Closing remarks Research on Gorenstein homological dimensions continues to be guided by the idea that every result about absolute homological dimensions should have a counterpart in Gorenstein homological algebra. In a curious departure from this principle, the behavior of Gorenstein injective dimension under faithfully flat base change was understood [11, thm. 1.7] years before the absolute case [10, thm. 2.2]. The main results of the present paper have brought the (co-)basechange results for Gorenstein dimensions in closer alignment with the results for absolute homological dimensions by relaxing the conditions on rings and complexes. It would still be interesting to remove the a priori assumption of homological boundedness from Theorem 4.12 to align it even closer with [9, cor. 3.1]. Acknowledgments This paper was written during visits by L.L. to Texas Tech University—while F.K. was a student there—and by L.W.C. and L.L. to Nanjing University; we gratefully acknowledge the hospitality of both institutions. In particular, we thank Nanqing Ding, our host at Nanjing, for inviting and supporting us. Appendix. GF closed rings by Driss Bennis Recall that a ring R is said to be left GF-closed if in every exact sequence of Rmodules 0 → G → M → G′ → 0 with G and G′ Gorenstein flat, also the middle module is Gorenstein flat. No example is known of a ring that is not GF-closed. From [4, lem. 2.5] it is known that the middle module is Gorenstein flat in every exact sequence 0 → G → M → F → 0 with F flat and G Gorenstein flat. A.1 Theorem. For a ring R the following conditions are equivalent. (i) R is left GF-closed. (ii) In every exact sequence of R-modules 0 → F → M → G → 0 with F flat and G Gorenstein flat, the module M is Gorenstein flat. Proof. The implication (i) =⇒ (ii) is trivial. To prove that (ii) implies (i), let 0 −→ A −→ B −→ C −→ 0 GORENSTEIN DIMENSIONS OF UNBOUNDED COMPLEXES AND CHANGE OF BASE 21 be an exact sequence of R-modules where A and C are G-flat. The goal is to prove that B is G-flat. For every injective R◦ -module I one evidently has TorR >0 (I, B) = 0. Thus, 5.2.(1) holds and we proceed to verify 5.2.(2). Since A is G-flat, there is an exact sequence 0 −→ A −→ F ′ −→ G′ −→ 0 with F ′ flat and G′ G-flat. From the push-out diagram 0 0 0  // A  // B // C // 0 0  // F ′  // X // C // 0  G′  G′  0  0 one gets by (ii) that X is G-flat, and so there exists an exact sequence 0 −→ X −→ F −→ G −→ 0 with F flat and G G-flat. Consider the second non-zero row in the push-out diagram 0 0 0 // B  // X  // G′ // 0 0 // B  // F  // B ′ // 0  G  G  0  0. The module B ′ has the same properties as B; indeed G and G′ are G-flat, and hence ′ ◦ one has TorR >0 (I, B ) = 0 for every injective R -module I. Thus, by repeating the above process one can construct the complex in 5.2.(2).  References [1] Javad Asadollahi and Shokrollah Salarian, Gorenstein injective dimension for complexes and Iwanaga-Gorenstein rings, Comm. Algebra 34 (2006), no. 8, 3009–3022. MR2250584 [2] Maurice Auslander and Mark Bridger, Stable module theory, Memoirs of the American Mathematical Society, No. 94, American Mathematical Society, Providence, R.I., 1969. MR0269685 [3] Luchezar L. Avramov and Hans-Bjørn Foxby, Homological dimensions of unbounded complexes, J. Pure Appl. Algebra 71 (1991), no. 2-3, 129–155. MR1117631 [4] Driss Bennis, Rings over which the class of Gorenstein flat modules is closed under extensions, Comm. Algebra 37 (2009), no. 3, 855–868. MR2503181 [5] Driss Bennis, A note on Gorenstein flat dimension, Algebra Colloq. 18 (2011), no. 1, 155–161. MR2753665 22 LARS WINTHER CHRISTENSEN, FATIH KÖKSAL, AND LI LIANG [6] Lars Winther Christensen, Anders Frankild, and Henrik Holm, On Gorenstein projective, injective and flat dimensions—A functorial description with applications, J. Algebra 302 (2006), no. 1, 231–279. MR2236602 [7] Lars Winther Christensen and Henrik Holm, Ascent properties of Auslander categories, Canad. J. Math. 61 (2009), no. 1, 76–108. MR2488450 [8] Lars Winther Christensen and Henrik Holm, The direct limit closure of perfect complexes, J. Pure Appl. Algebra 219 (2015), no. 3, 449–463. MR3279365 [9] Lars Winther Christensen and Srikanth B. Iyengar, Tests for injectivity of modules over commutative rings, preprint arXiv:1508.04639 [math.AC]; 8 pp. [10] Lars Winther Christensen and Fatih Köksal, Injective modules under faithfully flat ring extensions, Proc. Amer. Math. Soc. 144 (2016), no. 3, 1015–1020. MR3447655 [11] Lars Winther Christensen and Sean Sather-Wagstaff, Transfer of Gorenstein dimensions along ring homomorphisms, J. Pure Appl. Algebra 214 (2010), no. 6, 982–989. MR2580673 [12] Edgar E. Enochs and Overtoun M. G. Jenda, Relative homological algebra, de Gruyter Expositions in Mathematics, vol. 30, Walter de Gruyter & Co., Berlin, 2000. MR1753146 [13] Laurent Gruson and Christian U. Jensen, Dimensions cohomologiques reliées aux foncteurs lim(i) , Paul Dubreil and Marie-Paule Malliavin Algebra Seminar, 33rd Year (Paris, 1980), ←− Lecture Notes in Math., vol. 867, Springer, Berlin, 1981, pp. 234–294. MR0633523 [14] Henrik Holm, Gorenstein homological dimensions, J. Pure Appl. Algebra 189 (2004), no. 1-3, 167–193. MR2038564 [15] Alina Iacob, Gorenstein flat dimension of complexes, J. Math. Kyoto Univ. 49 (2009), no. 4, 817–842. MR2591118 [16] Christian U. Jensen, On the vanishing of lim(i) , J. Algebra 15 (1970), 151–166. MR0260839 ←− [17] Li Liang, Tate homology of modules of finite Gorenstein flat dimension, Algebr. Represent. Theory 16 (2013), no. 6, 1541–1560. MR3127346 [18] Barbara L. Osofsky, Homological dimension and cardinality, Trans. Amer. Math. Soc. 151 (1970), 641–649. MR0265411 [19] Michel Raynaud and Laurent Gruson, Critères de platitude et de projectivité. Techniques de “platification” d’un module, Invent. Math. 13 (1971), 1–89. MR0308104 ’ [20] Jan Štovı́ček, On purity and applications to coderived and singularity categories, preprint, arXiv:1412.1601v1 [math.CT]. [21] Oana Veliche, Gorenstein projective dimension for complexes, Trans. Amer. Math. Soc. 358 (2006), no. 3, 1257–1283. MR2187653 [22] Liu Zhongkui and Ren Wei, Transfer of Gorenstein dimensions of unbounded complexes along ring homomorphisms, Comm. Algebra 42 (2014), no. 8, 3325–3338. MR3196047 L.W.C. Texas Tech University, Lubbock, TX 79409, U.S.A. E-mail address: [email protected] URL: http://www.math.ttu.edu/~ lchriste F.K. Lewis University, Romeoville, IL 60446, U.S.A. E-mail address: [email protected] L.L. Lanzhou Jiaotong University, Lanzhou 730070, China E-mail address: [email protected]
0math.AC
PathNet: Evolution Channels Gradient Descent in Super Neural Networks Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha† , Andrei A. Rusu, Alexander Pritzel, Daan Wierstra Google DeepMind, London, UK. † Google Brain arXiv:1701.08734v1 [cs.NE] 30 Jan 2017 [email protected] ABSTRACT For artificial general intelligence (AGI) it would be efficient if multiple users trained the same giant neural network, permitting parameter reuse, without catastrophic forgetting. PathNet is a first step in this direction. It is a neural network algorithm that uses agents embedded in the neural network whose task is to discover which parts of the network to re-use for new tasks. Agents are pathways (views) through the network which determine the subset of parameters that are used and updated by the forwards and backwards passes of the backpropogation algorithm. During learning, a tournament selection genetic algorithm is used to select pathways through the neural network for replication and mutation. Pathway fitness is the performance of that pathway measured according to a cost function. We demonstrate successful transfer learning; fixing the parameters along a path learned on task A and re-evolving a new population of paths for task B, allows task B to be learned faster than it could be learned from scratch or after fine-tuning. Paths evolved on task B re-use parts of the optimal path evolved on task A. Positive transfer was demonstrated for binary MNIST, CIFAR, and SVHN supervised learning classification tasks, and a set of Atari and Labyrinth reinforcement learning tasks, suggesting PathNets have general applicability for neural network training. Finally, PathNet also significantly improves the robustness to hyperparameter choices of a parallel asynchronous reinforcement learning algorithm (A3C). Keywords Giant networks, Path evolution algorithm, Evolution and Learning, Continual Learning, Transfer Learning, MultiTask Learning, Basal Ganglia 1. INTRODUCTION A plausible requirement for artificial general intelligence is that many users will be required to train the same giant ACM ISBN X-XXXXX-XX-X/XX/XX. DOI: XXX neural network on a multitude of tasks. This is the most efficient way for the network to gain experience, because such a network can reuse existing knowledge instead of learning from scratch for each task. To achieve this, we propose that each user of the giant net be given a population of agents whose job it is to learn the user-defined task as efficiently as possible. Agents will learn how best to re-use existing parameters in the environment of the neural network by executing actions within the neural network. They must work in parallel with other agents who are learning other tasks for other users, sharing parameters if transfer is possible, learning to update disjoint parameters if interference is significant. Each agent may itself be controlled by an arbitrarily complex reinforcement learning algorithm, but here we chose the very simplest possible ‘agent’, a unit of evolution [4]. The framework for AGI described above includes aspects of transfer learning [22], continual learning [16] and multitask learning [2]. Our work shares a motivation with a recent paper “Outrageously large neural networks” in which the authors write that “the capacity of a neural network to absorb information is limited by its number of parameters” [19]. If a standard neural network is naively trained training cost scales quadratically with model width, whereas PathNet theoretically has constant computation speed with respect to the network width because only a fixed-size subset of the larger network is used for the forwards and backwards pass at any time (although there is no guarantee that more training may not be required in some cases). Our work is related also to “Convolutional Neural Fabrics” in which connection strengths between modules in the fabric are learned, but where (unlike PathNet) the whole fabric is used all the time [18]. This paper introduces PathNet, a novel learning algorithm with explicit support for transfer, continual, and multitask learning. Figure 1 illustrates the algorithm in action. The first task A is Pong and the second task B is Alien; both are trained consecutively for 80M timesteps each. The purple lines in Box 1 of the figure shows all 64 randomly initialized paths through the neural network model at the beginning of Pong training. A tournament selection genetic algorithm is then used to evolve paths, where during a fitness evaluation the path is trained for a few game episodes by gradient descent using a reinforcement learning algorithm. Thus, evolution and learning are taking place simultaneously, with evolution only guiding where gradient descent should be applied to change the weight and bias parameters. Box 2 shows the population converging (with many paths overlapping with each other) as performance improves. As perfect performance is achieved the population converges to a single path as shown in Box 3. Box 4 shows that the converged single pathway persists until the end of the training session. At this point the task switches to task B (Alien) and the optimal path for Pong gets ’fixed’, i.e. the modules on that path have their weights and biases frozen. Box 5 shows the fixed path as heavy dark red lines alongside a new randomly initialized population of paths in light blue. The new population of paths evolve and converge on Alien by Box 8. After 160M steps, the optimal path for Alien was fixed, shown as dark blue lines in Box 9. linear, followed by a transfer function; rectified linear units are used here. For each layer the outputs of the modules in that layer are summed before being passed into the active modules of the next layer. A module is active if it is present in the path genotype currently being evaluated (see below). A maximum of N distinct modules per layer are permitted in a pathway (typically N = 3 or 4). The final layer is unique and unshared for each task being learned. In the supervised case this is a single linear layer for each task, and in the A3C case each task (e.g. Atari game) has a value function readout and a policy readout (see [12] for a complete description of the A3C algorithm used). 2.2 Pathway Evolution: Serial and Parallel PathNets evolve a population of pathways through a neural network that scaffolds and channels any desired gradientdescent-based learning algorithm towards a limited subset of the neural network’s parameters and then fixes these parameters after learning so that functionality can never be lost; it resembles progressive neural networks, in that catastrophic forgetting is prevented by design [17]. In progressive neural networks the topology determining transfer is hardwired rather than learned, whereby the first neural network is trained on the source task and then a second neural network is trained on the target task which receives inputs from the first neural network, which has its weights fixed. PathNets allow the relationships between the original ‘columns’ and later ‘columns’ to be evolved, where a column is one deep neural network. P genotypes (pathways) are initialized randomly, each genotype is at most a N by L matrix of integers, which describe the active modules in each layer in that pathway. In the serial supervised implementation, a binary tournament selection algorithm is implemented in series as follows. A random genotype is chosen, and its pathway is trained for T epochs, its fitness being the negative classification error during that period of training. Then another random genotype is chosen and its pathway trained for T epochs. A copy of the winning pathway’s genotype overwrites the losing pathways genotype. The copy of the winning pathway’s genotype is then mutated by choosing independently each element with a probability of 1/[N × L] and adding an integer in the range [−2, 2] to it. A local neighbourhood was used to promote spatial localization of network functionality. Two examples of PathNets were investigated, a serial implementation on two supervised learning tasks where the PathNet is trained by stochastic gradient descent, and a parallel implementation on reinforcement learning tasks (Atari and Labyrinth games) where the PathNet is trained by the Async Advantage Actor-Critic (A3C) . A3C is an efficient distributed reinforcement learning algorithm which runs on multiple CPUs with e.g. 64 asynchronously updated workers that simultaneously share and update the parameters of a single network[12]. Positive transfer from a source task to a target task is demonstrated in all four domains, compared to single fixed path controls trained from scratch and after fine-tuning on the first task. In the A3C (reinforcement learning) case, all 64 genotypes are evaluated in parallel, one by each of the 64 workers. Therefore pathways restrict the simultaneous updates of parameters by the workers to only their specific subsets, as opposed to the standard A3C algorithm in which all workers update all parameters. The fitness of a genotype is the return accumulated over the T episodes that a worker played using that genotypes pathway. While a worker is evaluating, it writes a large negative fitness to the shared fitness array, so that no genotypes wins a tournament until it has been evaluated. Once the worker has finished T episodes, it chooses B other random genotypes and checks if any of those genotypes have returned a fitness of at least its own fitness. If at least one has, then the highest fit genotype overwrites the current worker’s genotype, and is mutated as above. If no other worker had a genotype with fitness greater than this workers own genotype, then the worker reevaluates its own genotype. The concept of the PathNet was first conceived of within the framework of Darwinian Neurodynamics as an attempt to envisage how evolutionary algorithms could be implemented in the brain [5]. However, in that original work both the topology and the weights of the path were evolved and there was no gradient descent learning [4]. Performance was comparable with a standard genetic algorithm on combinatorial optimization problems. Here we show performance superior to A3C and stochastic gradient descent for transfer learning. 2. 2.1 METHODS PathNet Architecture A PathNet is a modular deep neural network having L layers with each layer consisting of M modules. Each module is itself a neural network, here either convolutional or 2.3 Transfer Learning Paradigm Once task A has been trained for a fixed period of time or until some error threshold has been reached, the best fit pathway is fixed, which means its parameters are no longer allowed to change. All other parameters not in an optimal path are reinitialized. We found that without reinitialization transfer performance did not exceed that of fine-tuning. In the A3C case (but not the supervised learning cases) the original best fit pathway is always active during the forwards pass of the network, in addition to the newly evolving pathway, but its parameters are not modified by the backwards pass. A new set of random pathways is then initialized and evolved/trained on task B. In both the supervised and reinforcement settings, pathNet is compared with two alternative setups: an independent learning control where the target task is learned de novo, and a fine-tuning control where the second task is learned with the same path that learned the first task (but with a new value function and policy readout). 2.4 Binary MNIST classification tasks A binary MNIST classification involves distinguishing two classes of MNIST digits from one another, for example 5 verses 6 [11]. To make the task more difficult, salt and pepper noise of 50% is added to the MNIST digits. A transfer experiment involves training and evolving paths on the first task until perfect classification on the training set is achieved. At this point, a new population of path genotypes is initialized and evolved on the second task until perfect performance on the training set is achieved. Between tasks the following modifications are made to the underlying network. The parameters contained in the optimal path evolved on the first task are fixed, and all other parameters are reset to their random initial values. The reported outcome measures are the training times required to reach this classification accuracy. The overall PathNet consists of L = 3 layers. Each layer contains M = 10 linear units with 20 neurons each followed by rectified linear units. Between layers the activations of the units are summed. Reading out from the final layer is a separate linear layer for each task. A maximum of 3 of these units per layer can be included in a single pathway. The genotype describing this pathway is a 3×3 matrix of integers in the range [1, 10]. A population of 64 of these genotypes was generated uniformly randomly at the start of each task. The overall algorithm is as follows. Two paths are chosen randomly for evaluation. The evaluation of one path involves training with stochastic gradient descent with learning rate 0.0001 on 50 mini-batches of size 16. The fitness of that pathway is the proportion of correct examples classified on the training set during this training period. Once the fitness of both pathways has been calculated, the pathway with the lower fitness is overwritten by a copy of the pathway with the higher fitness and mutated. Mutation takes place with equal probability 1/(3 × 3) per element of the genotype, adding a new random integer from range[−2, 2] to the current value of that element. This is a binary tournament selection algorithm (B = 2) [7]. 2.5 CIFAR and SVHN classification tasks The larger version of the above network is used to train on CIFAR and cropped SVHN [13] of standard size 28 × 28 withL = 3 and M = 20 modules per layer of 20 neurons each, and with pathways that may contain up to 5 modules per layer. In this case the networks were not trained to some fixed accuracy, but for a fixed period of 500 generations. Each generation consisted of the evaluation of two pathways, each for 50 mini-batches of size 16 images. The performance measure here was the accuracy achieved after this fixed training time. Evidence for positive transfer in this case is given by the second task showing a higher final accuracy compared to when it was learned from scratch. 2.6 Atari games We tested whether there was a speedup in learning a second (target) game after having played either Pong, RiverRaid or Seaquest as a source game. The target games were as follows: Alien, Asterix, Boxing, Centipede, Gopher, Hero, JamesBond, Krull, RoadRunner, StarGunner, WizardofWor. These are the same games presented by the authors of Progressive Neural Networks. In this case the A3C algorithm was used with 64 workers running in parallel. The first game is learned for 80M (or 40M) timesteps of training in total across all workers. During this time, workers are evolving pathways through the network. After the first game is learned the population of paths is reinitialized and the workers play the second game for 40M timesteps in total. Between tasks the following modifications are made to the underlying network. The parameters contained in the optimal path evolved on the first task are fixed, and all other parameters are reset to their random initial values. Evolution takes place according to a distributed asynchronous tournament selection algorithm which works as follows. The population of pathways and their fitnesses is stored in a central parameter server. Each worker accesses its own particular index into this population of genotypes corresponding to that worker’s id. While it is first evaluating that pathway it writes -1000 to the fitness array at a position corresponding to it’s id. After T = 10 episodes that worker writes the total sum of rewards obtained into the fitness array. It then observes the fitnesses of B = 20 other random workers and if any worker has a higher fitness than itself, it chooses the highest fitness worker in that set, and copies the pathway genotype to its own location with mutation, and resets its own fitness to -1000. However, if no other worker of the B chosen has a path better than its own, then its own pathway is re-evaluated without resetting the fitness obtained from the previous evaluation. Mutation is as before, with the minor modification that if a location is chosen for mutation, then there is a 20% chance (during the second task) that a module from the optimal pathway evolved in the first task will be incorporated into the genotype. In the variant of PathNet used for this task, the optimal path evolved in the first task is always active during the forwards pass, in this sense, this version resembles progressive nets more closely than the supervised versions presented above. The architecture of the network used consists of a core of L = 4 layers, each with a width of M = 10 modules: namely, 3 convolutional layers where each module consists of 8 kernels, and a final layer of fully connected modules of 50 hidden nodes each. Between layers the feature maps and activations are summed. Reading out from the final layer are two linear layers for each game, one encoding a value function and the other encoding an action policy. Figure 2 shows the network architecture: The first three layers’ modules are shown as green boxes; final layer of modules shown as purple boxes; the between-layer summing modules are shown as blue boxes; and active modules specified by the pathway as red boxes. Readout units not involved in the PathNet itself are shown as circles on the right. 2.7 Labyrinth Games We investigated the performance of a very similar PathNet architecture on 3 Labyrinth games. Labyrinth is a 3D first person game environment [10]. We used the same settings used for Atari games. We chose three games: ’Laser Tag Chasm’ (lt chasm), ’Seek Avoid Arena 01’ (seekavoid arena) and ’Stairway to Melon 01’ (stairway to melon). lt chasm takes place in a square room with four opponents that must be tagged for points. A chasm in the center of the room must be avoided and shield pick-ups exist that the player must jump to reach. seekavoid arena is a 3D room containing apples and lemons. The player must pick up the apples whilst avoiding the lemons. stairway to melon offers the player two options; the player can either collect a set of small rewards (apples) or they can opt to take a punishment (lemon) in order to reach a melon which results in a higher net reward. The PathNet architecture used for Labyrinth was identical to the one used for Atari with the exception of a moduleduplication mechanism. In the previous models there has not been any actual copying of parameters, only copying of views over a fixed set of parameters. In the Labyrinth model we enabled PathNet to copy the weights of modules to other modules within the same layer, emulating Net2Net [3] or a distillation [9] like operation. To aid this, we measure the extent to which a module contributes to the fitness of all the paths it is in by taking a sliding mean over the fitness of any path which contains that module. Using this measure it is possible to bias a weight-copying operator such that currently more globally useful modules are more likely to be copied across into other modules and hence into other paths. The hyperparameter module duplication rate determines the rate at which modules would be allowed to duplicate themselves. For the Labyrinth games both the first and second tasks were trained for 40M timesteps. 3. 3.1 RESULTS Binary MNIST Classification Figure 3 shows that with PathNet, learning a source MNIST classification task helps speed up learning in the target MNIST classification task (mean time to solution = 167 generations); greater than the speedup achievable by fine-tuning a fixed path that had initially been trained on the source task (mean time to solution = 229), and also compared to de novo learning of the target task from scratch (mean time to solution = 195). PathNet (bottom graph) learns in fewer generations with less data than both the fine-tuning (middle) and independent learning controls (top). The control networks consisted of exactly the same learning algorithm, but with no evolution of paths, and only one fixed maximum size pathway through the network. The total speedup ratio compared to the independent controls was 1.18. Videos showing performance of PathNet can be obtained online at https://goo.gl/oVHMJo. They reveal that the modules in early layers of the big network converge in the population of pathways the quickest, followed by later layers, there being much more exploration and training of all the modules in the final layer. Many modules in the final layer contribute to high fitness, whereas only a few modules in the first layer do so. Thus, a population provides an elegant solution to the exploration /exploitation trade-off in a layer specific manner. Analysis did not reveal that the speedup ratio was correlated with path overlap as measured by the number of modules in the original optimal path that were present in the population of paths at the end of the second task. This suggests that speedup can be obtained by PathNet both determining when there should be overlap and when there should not be overlap. The fact that on average fine-tuning is slower than independent task learning in MNIST binary classification tasks implies that generally overlap would not be expected to provide speedup. Thus, PathNet can be seen to have done its job by controlling the amount of overlap properly, see Figure 4 3.2 CIFAR and SVHN In this experiment we compare only the accuracy PathNet obtains after a short fixed number of 500 generations. A fully connected network of this size is generally insufficient to perform well on these datasets. After 500 generations, when learning from scratch with PathNet, cSVHN and CIFAR are learned to 25.5% and 35.3% accuracy on average, see Figure 5. But after the first task has been learned to this accuracy, learning the second task is faster, so when cSVHN and CIFAR are learned as the second task with PathNet, then accuracies of 35.7% and 39.8% are achieved respectively. Thus, both cSVHN and CIFAR are learned faster with PathNet when learned second rather than first. 3.3 Atari Games Transfer performance on PathNet is compared with fixed maximum-size-path de novo training and fine-tuning controls for which a hyperparameter sweep was conducted involving learning rates [0.001,0.0005,0.0001] and entropy costs [0.01, 0.001, 0.0001]. The following hyperparameters were investigated for PathNet: evaluation time T [1,10,50], mutation rate [0.1, 0.01, 0.001] and tournament size B [2, 10, 20]. In Figure 6 the performance of PathNet (blue) is compared with independent and fine-tuning controls (green) on 4 target games, having learned RiverRaid first. In all cases the top 5 runs of a hyperparameter search are shown. We generally found that strong selection with tournament sizes of B = 10, T = 10 game episodes per evaluation, and low mutation rates of 0.01-0.001 were optimal, allowing rapid convergence of paths to a single path, followed by exploration of small variants to a path, thus focusing learning on a few parameters, with occasional exploration of novel bypasses. PathNet is superior to controls in these 4 cases. Figure 7 shows that PathNet is on average superior to independent learning and fine-tuning over the Atari games we investigated. Compared to a control of 1.0, fine tuning achieves a 1.16 speedup on average whereas PathNet achieves a 1.33 times speedup. Results for transfer on more Atari games can be seen in Figure 12. 3.4 Labyrinth Games PathNet transfer between the three labyrinth games lt chasm, seekavoid arena and stairway to melon is compared with fixed maximum-size-path de novo training and fine-tuning controls for which a hyperparameter sweep was conducted involving mutation rates [ 0.1, 0.01, 0.001], module duplication rate [ 0.0, 0.05, 0.1, 0.5] (per episode completed by worker 0) and tournament size B [2, 10]. The learning rate was fixed at 0.001, entropy cost at 0.0001 and evaluation time T at 13. Figure 9 shows the mean of the training performance for the top 5 experiments for both PathNet and the fixed-path controls. Each row represents a different source game and each column a different target game. Where the source game and target game are the same the graph shows the results of learning the game de novo from scratch. Figure 11 shows that in some cases the module duplication operation produces improved performance compared to standard PathNet. In several cases (transfer to lt chasm and transfer from lt chasm to seekavoid arena) PathNet learns the second task faster than fine-tuning. Interestingly, PathNet also performs better than fine-tuning when learning stairway to melon and seekavoid arena from scratch. Results of the best runs from a hyperparameter search are summarized in Figure 10. Here performance is evaluated by measuring the area under the learning curve (average score per episode during training), rather than final score. The numbers in the table show the relative performance of an architecture learning a target task (each column) compared with an independent baseline with a fixed maximum size path trained from scratch only on the target task. The controls are labelled as ’Fixed path de novo from scratch’ (top row) and is 1 for each column (target task). A ratio in a column >1 represents the speedup when learning that column’s target task, and <1 is a slowdown. The three rows below the de novo control show the fine-tuning results between the three games. The first row in the PathNet results show performance learning the individual games from scratch and the three rows below that show the PathNet transfer results between the three games. On transferring to lt chasm both fine-tuning and PathNet perform worse than the control de novo learning. On the the other two games both exhibit positive transfer performance. The average performance ratio for fine-tuning for transfer across all the game combinations is 1.00 (2 d.p.), essentially no faster than learning from scratch. The average performance ratio for PathNet is 1.26 (2 d.p.); largely due to the good performance transferring from seekavoid arena to stairway to melon. We also compared PathNet to independent and fine-tuning controls over the same sweep of 243 hyperparameters as used in the Atari experiments described above. On the Labyrinth level seekavoid arena in which the agent must collect apples but avoid lemons we found that the PathNet had significantly higher mean performance than control runs, both when learning seekavoid arena from scratch compared to the de novo controls, and for relearning from the same task, compared to fine-tuning from a network that had previously learned seekavoid arena, see Figure 8. 4. CONCLUSION PathNet extends our original work on the Path Evolution Algorithm [4] to Deep Learning whereby the weights and biases of the network are learned by gradient descent, but evolution determines which subset of parameters is to be trained. We have shown that PathNet is capable of sustaining transfer learning on at least four tasks in both the supervised and reinforcement learning settings. PathNet may be thought of as implementing a form of ‘evolutionary dropout’ in which instead of randomly dropping out units and their connections, dropout samples or ‘thinned networks’ are evolved [21]. PathNet has the added advantage that dropout frequency is emergent, because the population converges faster at the early layers of the network than in the later layers. PathNet also resembles ‘evolutionary swapout’ [20], in fact we have experimented with having standard linear modules, skip modules and residual modules in the same layer and found that path evolution was capable of discovering effective structures within this diverse network. PathNet is related also to recent work on convolutional neural fabrics, but there the whole network is always used and so the principle cannot scale to giant networks [18]. Other approaches to combining evolution and learning have involved parameter copying, whereas there is no such copying in the current implementation of PathNet [1][3]. Whilst we have only demonstrated PathNet in a fairly small network, the principle can scale to much larger neural networks with more efficient implementations of pathway gating. This will allow extension to multiple tasks. We also wish to try PathNet on other RL tasks which may be more suitable for transfer learning than Atari, for example continuous robotic control problems. Further investigation is required to understand the extent to which PathNet may be superior to using fixed paths. Firstly, a possibility is that mutable paths provide a more useful form of diverse exploration in RL tasks [15]. Secondly, it is possible that a larger number of workers can be used in A3C because if each worker can determine which parameters to update, there may be selection for pathways that do not interfere with each other. We are still investigating the potential benefits of module duplication. see Supplementary Video https://goo.gl/oVHMJo. Using this measure it is possible to bias the mutation operator such that currently more globally useful modules are more likely to be slotted into other paths. Further work is also to be carried out in multi-task learning which has not yet been addressed in this paper. Finally, it is always possible and sometimes desirable to replace evolutionary variation operators with variation operators learned by reinforcement learning. A tournament selection algorithm with mutation is only the simplest way to achieve adaptive paths. It is clear that more sophisticated methods such as policy gradient methods may be used to learn the distribution of pathways as a function of the long term returns obtained by paths, and as a function of a task description input. This may be done through a softer form of gating than used by PathNet here. Furthermore, a population (ensemble) of soft gate matrices may be maintained and an RL algorithm may be permitted to ’mutate’ these values. The operations of PathNet resemble those of the Basal Ganglia, which we propose determines which subsets of the cortex are to be active and trainable as a function of goal/subgoal signals from the prefrontal cortex, a hypothesis related to others in the literature [8] [14] [6]. 5. ACKNOWLEDGMENTS Thanks to Hubert Soyer, Arka Pal, Gabriel Dulac-Arnold, Gabriel Barth-Maron, Matteo Hessel, Alban Rustrani, Stephen Gaffney, Joel Leibo, Eors Szathmary 6. REFERENCES [1] J. E. Auerbach, C. Fernando, and D. Floreano. Online extreme evolutionary learning machines. In Artificial Life 14: Proceedings of the Fourteenth International Conference on the Synthesis and Simulation of Living Systems, number EPFL-CONF-200273, pages 465–472. The MIT Press, 2014. [2] R. Caruana. Multitask learning. In Learning to learn, pages 95–133. Springer, 1998. [3] T. Chen, I. Goodfellow, and J. Shlens. Net2net: Accelerating learning via knowledge transfer. arXiv preprint arXiv:1511.05641, 2015. [4] C. Fernando, V. Vasas, E. Szathmáry, and P. Husbands. Evolvable neuronal paths: a novel basis for information and search in the brain. PloS one, 6(8):e23534, 2011. [5] C. T. Fernando, E. Szathmary, and P. Husbands. Selectionist and evolutionary approaches to brain function: a critical appraisal. Frontiers in computational neuroscience, 6:24, 2012. [6] M. J. Frank, B. Loughry, and R. C. OâĂŹReilly. Interactions between frontal cortex and basal ganglia in working memory: a computational model. Cognitive, Affective, & Behavioral Neuroscience, 1(2):137–160, 2001. [7] I. Harvey. The microbial genetic algorithm. In Advances in artificial life. Darwin Meets von Neumann, pages 126–133. Springer, 2011. [8] T. E. Hazy, M. J. Frank, and R. C. O’Reilly. Towards an executive without a homunculus: computational models of the prefrontal cortex/basal ganglia system. Philosophical Transactions of the Royal Society B: Biological Sciences, 362(1485):1601–1613, 2007. [9] G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. [10] M. Jaderberg, V. Mnih, W. M. Czarnecki, T. Schaul, J. Z. Leibo, D. Silver, and K. Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. [11] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. [12] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu. Asynchronous methods for deep reinforcement learning. arXiv preprint arXiv:1602.01783, 2016. [13] Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Ng. Reading digits in natural images with unsupervised feature learning. [14] R. C. O’Reilly and M. J. Frank. Making working memory work: a computational model of learning in the prefrontal cortex and basal ganglia. Neural computation, 18(2):283–328, 2006. [15] I. Osband, C. Blundell, A. Pritzel, and B. Van Roy. Deep exploration via bootstrapped dqn. arXiv preprint arXiv:1602.04621, 2016. [16] M. B. Ring. Continual Learning in Reinforcement Environments. PhD thesis, University of Texas at Austin, 1994. [17] A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer, J. Kirkpatrick, K. Kavukcuoglu, R. Pascanu, and R. Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. [18] S. Saxena and J. Verbeek. Convolutional neural fabrics. arXiv preprint arXiv:1606.02492, 2016. [19] N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean. Outrageously large neural networks: A sparsely-gated mixture-of-experts layer. 2017. [ICAR submission]. [20] S. Singh, D. Hoiem, and D. Forsyth. Swapout: Learning an ensemble of deep architectures. arXiv preprint arXiv:1605.06465, 2016. [21] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014. [22] M. E. Taylor and P. Stone. An introduction to intertask transfer for reinforcement learning. AI Magazine, 32(1):15, 2011. 4 7 2 5 8 3 6 Score 1 9 Pong Alien 8e7 steps 16e7 steps Figure 1: A population of randomly initialized pathways (purple lines in Box 1) are evolved whilst learning task A, Pong. At the end of training, the best pathway is fixed (dark red lines in Box 5) and a new population of paths are generated (light blue lines in Box 5) for task B. This population is then trained on Alien and the optimal pathway that is evolved on Alien is subsequently fixed at the end of training, shown as dark blue lines in Box 9. Figure 2: The PathNet used for learning Atari and Labyrinth games consists of a 4 layer network with 10 (or sometimes 15) modules in each layer. The first three layers’ modules are convolutional 2D kernels with 8 kernels per module (green boxes in figure), kernel sizes (8, 4, 3), and strides (4, 2, 1) from 1st to 3rd layers respectively, and with the final layer consisting of fully connected linear layers of 50 neurons each (purple boxes). After each module there is a rectified linear unit. Between layers the feature maps are summed before being passed to the modules of the subsequent layer (blue boxes). Typically a maximum of 4 modules per layer are permitted to be included in a pathway (shown as red boxes), otherwise evolution would simply grow the pathway to include the whole network as this would improve fitness by growing the number of parameters that can be tuned by learning. C1 + C2 Fine-tuning control Frequency C1 +C2|C1 Independent control P1 + P2|P1 PathNet Generations to achieve 0.998 accuracy Overlap measure Figure 3: PathNet is capable of supporting transfer learning in binary MNIST classification. The top figure shows C1 + C2 which is the sum of learning times over the source and target tasks required to learn both tasks to 0.998 accuracy, with a separate maximum sized fixed path for each task, and thus constitutes the independent learning control; mean = 195 generations. The middle figure shows C1 + C2|C1, which is the time required to learn both tasks where the second task is not learned from scratch, but by fine tuning the same maximum sized fixed path as was used to learn the first task; mean = 229 generations. The bottom graph shows the sum of times required to learn the source and target tasks using PathNet; mean = 167 generations. Speedup Ratio Figure 4: Overlap (y-axis) is not correlated with speedup ratio (C1 + C2)/(P1 + P2|P1) suggesting that PathNet discovers how much overlap to use based on whether the tasks could benefit from overlap or not. cSVHN → CIFAR = cSVHN Frequency = CIFAR CIFAR → cSVHN Correct classifications out of test set of size 1000 Figure 5: Top graph shows a histogram of accuracies when learning cSVHN first and CIFAR second. The bottom graph shows a histogram of accuracies when learning CIFAR first and cSVHN second. Learning either task second with PathNet results in improved performance, suggesting that PathNet is able to use weights in the optimal path found in task A, when learning task B. Figure 6: PathNet shows positive transfer from River Raid to Centipede, Boxing, Hero and James Bond. The graphs show the reward over the first 40 million steps of training. Blue shows the results from the best five hyperparameter settings of PathNet out of 243. Compare these to the best five hyperparameter setting runs for independent learning (red) and fine-tuning (green) controls out of 45. wizard_of_wor Star_gunner Robotank Road_Runner Krull JamesBond Hero Gopher Centipede Boxing Asterix Alien de novo Fine-Tuning Pathnet Figure 7: We report results by showing the best run from a hyperparameter search. Performance is evaluated by measuring the area under the learning curve (average score per episode during training), rather than final score. The transfer score is then defined as the relative performance of an architecture compared with a independent baseline (control) with a fixed maximum size path, trained only on the target task (top row). This ratio is >1 if there is speedup, and <1 if there is slowdown. We present transfer score curves for selected source-target games, and summarize all such pairs in this transfer matrix. The next 3 rows show fine-tuning controls, and the final three rows show PathNet. Green means positive transfer, and blue means negative transfer. PathNet PathNet Control Control Figure 8: PathNet outperforms the de novo control and the fine-tuning control on learning seekavoid arena for the first time, on average over all hyperparameters explored, out of 243 experiments. It also outperforms fine tuning when seekavoid arena [10] is learned a second time, with a new value and policy readout, with the same set of 243 hyperparameter settings. Figure 9: Means of the best 5 (out of a size 24 parameter sweep) training runs for PathNet compared with the fine-tuning (off diagonal) and independent path (diagonal) controls for the three Labyrinth levels considered. The diagonal results show the performance of PathNet when learning from scratch compared to the independent fixed path control. Off-diagonal results show the performance of pathNet during transfer. Figure 10: Results of the best runs from a hyperparameter search. The numbers in the table show the relative performance of an architecture learning a target task compared with an independent baseline with a fixed maximum size path trained from scratch only on the target task. A number >1 (green) represents improved performance and <1 (blue) is worse performance. The top row (’Fixed path de novo from scratch’) are the controls, i.e. the performance of the fixed-path network learning the target task (each column) from scratch. The rows below show fixed-path performance on fine-tuning and PathNet, learning de novo from scratch, and PathNet performance transferring from task A to task B. Figure 11: The contribution of Module Duplication on performance. Graphs show the mean performance across all hyperparameters for PathNet with different module duplication rates [0, 0.05. 0.1, 0.5] per episode completed by worker 0. For some tasks (e.g. learning seekavoid arena module duplication has been beneficial. Figure 12: Transfer results on various Atari games. The graphs show the reward over the first 40 million steps of training. Blue shows the results from the best five hyperparameter settings of PathNet out of 243. Compare these to the best five hyperparameter setting runs for independent learning (red) and fine-tuning (green) controls out of 45.
9cs.NE
arXiv:1508.01041v2 [cs.CE] 20 Apr 2016 Implementation of the Log-Conformation Formulation for Two-Dimensional Viscoelastic Flow Kristian Ejlebjerg Jensen [email protected] Department of Micro- and Nanotechnology Technical University of Denmark Fridolin Okkels Department of Micro- and Nanotechnology Technical University of Denmark Peter Szabo [email protected] Department of Chemical and Biochemical Engineering Technical University of Denmark April 21, 2016 Abstract We have implemented the log-conformation method for two-dimensional viscoelastic flow in COMSOL, a commercial high-level finite element package. The code is verified for an Oldroyd-B fluid flowing past a confined cylinder. We are also able to describe the well-known bistability of the viscoelastic flow in a crossslot geometry for a FENE-CR fluid, and we describe the changes required for performing simulations with the Phan-Thien-Tanner (PTT), Giesekus and FENE-P models. Finally, we calculate the flow of a FENECR fluid in a geometry with three in- and outlets. The implementation is included in the supplementary material, and we hope that it can inspire new as well as experienced researchers in the field of differential constitutive equations for viscoelastic flow. 1 Introduction The flow of viscoelastic fluids in complex geometries has many practical applications due to the fact that any dissolved elastic component, be it polymers or biological molecules, will give rise to viscoelastic effects. Such effects require the use of models that take the fluid’s memory of past deformations into account. Differential constitutive equations is able to do just that, and they provide a good quantitative agreement with experiments [1]. Complex geometries call for the use of numerical methods, and therefore significant effort has been devoted towards various discretization techniques as well as model reformulations. The appearance of singularities, that caused codes to break down in smooth geometries at moderate elasticity, led to the definition of a ”High Weissenberg Number Problem” (HWNP). This was effectively solved, when loss of positive definiteness for the conformation tensor was recognized as the cause of the singularity and the log-conformation method was introduced as a remedy [2]. 1 Commercial numerical tools provide a simple workflow due to the integration of geometry description, unstructured mesh generation, discretization, solvers, and post-processing, which makes the treatment of complex geometries effortless compared to research grade code, but to our knowledge there does not exist any commercial tool that implements the log-conformation method. This is perhaps due to the fact that the method involves the calculation of eigenvectors and eigenvalues for the conformation tensor, which complicates the formulation of the constitutive equation. In two (and three) dimensions explicit expressions exists for these quantities, and it is thus possible to formulate the governing equations solely in terms of functions recognized by commercial simulation packages. We have included a COMSOL implementation of the Oldroyd-B model with log-conformation reformulation in the supplementary material, and we hope this can help other researchers working with differential constitutive equations. The article is structured in four parts: First we discuss implementation details of the log-conformation method for the most simple constitutive equations. We then calculate the flow of an Oldroyd-B fluid past a confined cylinder using COMSOL for comparison with other works [2]. As a third point we demonstrate the ability to predict bistable behavior of a FENE-CR fluid in a cross-slot geometry [3]. Finally we consider a geometry with three in- and outlets, which has not previously been described in the context of viscoelastic flow, at least to our knowledge. 2 Log-Conformation Implementation In the so-called dumbbell models, the elastic part of a viscoelastic fluid is approximated by two springconnected point masses – an elastic dumbbell. The orientation and elongation of this dumbbell is described using the end-to-end vector a with the equilibrium length aeq , see figure 1. The conformation tensor, A, is Figure 1: An elastic dumbbell is illustrated with its two spring-connected point masses, and the end-to-end vector describing orientation and extension. the statistical average, h...i of the dyadic product between the end-to-end vector and itself normalized with the squared equilibrium length, A = ha ⊗ ai . a2eq (1) The equation system for the Oldroyd-B model is written below. We neglect the effect of temperature and compressibility, but keep inertia for the sake of generality.   τ s z }| {   Dv   ρ = ∇ · −pI + ηs ∇ v + (∇ v)T +τ e  (2) Dt ∇·v τe 0 = 0  ηp = A−I λ 1 = − (A − I) λ  DA  + − A · ∇ v + (∇ v)T · A Dt (3) (4) (5) where D is the material derivative, ρ is the density, v is the velocity vector, t is time, p is the pressure ηs is the solvent viscosity, ηp is the elastic viscosity, λ is the relaxation time, τ s is the solvent stress tensor and 2 τ e is the elastic stress tensor. The Navier-Stokes equation (2) guarantees that Newton’s 2nd law is obeyed, while the continuity equation (3) insures mass conservation. The expression for the elastic stress tensor in equation (4) is best understood in light of the conformation tensor definition (1) with ηp /λ acting as a spring constant. The evolution of the conformation tensor is described in equation (5) with relaxation on the first line balanced by the upper convected time derivative, which consists of convection in terms of the total derivative together with rotation and extension of the conformation tensor as imposed by the velocity gradient in square brackets. When the velocity vector is approximated by C 0 continuous polynomials, the velocity gradient, ∇ v becomes discontinous. This can cause numerical difficulties [4], and therefore aC 0 continuous approximation  G is usually constructed for use in equation (5). Furthermore the zero, ηp ∇ v + (∇ v)T − G − GT is added on the right-hand side of equation (2) to preserve the elliptic nature of the equation for small solvent viscosities in what is called the discrete elastic viscous stress splitting (DEVSS). Similarly, we use streamline upwind diffusion (SUPG) to introduce an elliptic component in the constitutive equation (5), see section 2.2. The conformation tensor is symmetric, and the point of the log-conformation method is to exploit this property to rewrite equation (5) to the form Ds − Π(s, G) = 0, Dt where s is the matrix logarithm of the conformation tensor. They are related by the following expression = R · e s · RT   v1nx −v1ny = v1ny v1nx  λ  e 1 0 = 0 eλ2 A R es A11 2 2 = v1nx eλ1 + v1ny eλ2 A12 = v1nx v1ny eλ1 − eλ2 A22 2 2 = v1nx eλ2 + v1ny eλ1  where λ1 and λ2 are the eigenvalues of s, while v1nx and v1ny are the normalized eigenvector components belonging to λ1 . The orthonormal rotation matrix, R holds the components of the eigenvectors, but we exploit that the second eigenvector is just the transpose of the first. Denoting the components of s by s11 , s12 and s22 , the eigenvalues become1   q 1 λ1 = 2 s22 + s11 − (s22 − s11 )2 + 4s212 and   q 1 λ2 = 2 s22 + s11 + (s22 − s11 )2 + 4s212 , and the components of the first eigenvector are  λ1 − s22 s12 v1 =  v1x .q v1y .q  . The eigenvector is normalized v1n =  2 + v2 v1x 1y 2 + v2 v1x 1y  . The special case s12 = 0 is handled by using R = I, when |s12 | ≤  with  = 10−12 . 1 There also exist explicit expressions for the three-dimensional case. 3 In the following we will express the reaction term Π as R · Ω · RT as described in [5]. The continuous velocity gradient is rewritten in the principal frame G̃ = RT · G · R G̃11 = 2 2 v1nx G11 + v1nx v1yx (G12 + G21 ) + v1ny G22 G̃12 = 2 2 v1nx G21 + v1nx v1yx (G22 − G11 ) − v1ny G12 G̃21 = 2 2 v1nx G12 + v1nx v1yx (G22 − G11 ) − v1ny G21 G̃22 = 2 2 v1nx G22 − v1nx v1yx (G12 − G21 ) + v1ny G11 . Note that this matrix is not symmetric, and that we have chosen G12 to refer to the derivative of the velocity vector x-component with respect to the y-coordinate. The diagonal components Ω11 and Ω22 are Ω11 = Ω22 = 1 − e−λ1 and λ 1 − e−λ2 . 2G̃22 − λ 2G̃11 − These expressions are specific to the Oldroyd-B model, but it is straight forward to write the expressions for the Giesekus, Phan-Thien-Tanner, FENE-P or FENE-CR models instead [5]. Giesekus : where PTT(exp.) : where (6) 1 + αGie. (ex − 1) x (e − 1) λ 1 ,λ2 ) 2G̃11 − kPTT (λ λ kPTT (λ1 ,λ2 ) , 2G̃22 − λ Ω11 = Ω22 = kPTT (λ1 , λ2 ) = ePTT (λ1 +λ2 −3) : where FENE − CR 2G̃11 − fGie. (λ1 ) 2G̃22 − fGie. (λ2 ) fGie. (x) = where FENE − P Ω11 = Ω22 = Ω11 = 2G̃11 − Ω22 = 2G̃22 − k(λ1 , λ2 ) = : (7) )eλ1 −1 k(λ1 ,λ2 λ k(λ1 ,λ2 )eλ2 −1 λ , 1 1 − (λ1 + λ2 )/max2 λ1 2G̃11 − k(λ1 , λ2 ) e λ−1 , λ2 2G̃22 − k(λ1 , λ2 ) e λ−1 1 k(λ1 , λ2 ) = 1 − (λ1 + λ2 )/a2max Ω11 = Ω22 = In the case of the FENE-P and FENE-CR models, the amax is a maximum extension, i.e. Tr(A) < a2max . The off-diagonal component, Ω12 , equals   (  λ1 −λ2 λ1 λ2 e G̃ + e G̃ , |λ1 − λ2 | >  12 21 λ λ e 1 −e 2 Ω12 = G̃12 + G̃21 , |λ1 − λ2 | ≤  where we have reused the numerical parameter for the eigenvectors to handle the special case of identical eigenvalues. We are now ready to compute the reaction term Π12 2 2 = v1nx Ω11 − 2v1nx v1ny Ω12 + v1ny Ω22  2 2 = v1nx − v1ny Ω12 + v1nx v1ny (Ω11 − Ω22 ) Π22 2 2 = v1nx Ω22 + 2v1nx v1ny Ω12 + v1ny Ω11 Π11 To ease modification of the implementation, we have opted to construct functions for each of the above expressions instead of writing out Π in terms of expressions involving s and G only. 4 With respect to visualization it is worth noting that s and A are positive definite tensors, and one can thus think of them as ”fields of ellipses”. In other words, the conformation tensor has an extension in every point given by its eigenvectors and eigenvalues. The trace of the conformation tensor is the sum of the extensions in the principal directions given by the eigenvectors, while the determinant is the product of these extensions. The trace and determinant equal the sum and product of the eigenvalues, respectively, and they are invariant with respect to rotation of the coordinate system, which make them interesting in the context of post-processing/visualization2 . The last invariant is the angle(s) related to the eigenvectors, but the eigenvector associated with the largest eigenvalue is normally oriented with the streamlines, so this angle provides little insight. Furthermore the trace and determinant often look very similar, at least on a logarithmic scale, and therefore we just plot the trace. 2.1 Non-Dimensional Equations We define the dimensionless spatial coordinates, velocity, time, pressure and stress (x̃, ṽ, t̃, p̃ and τ̃ e ), x = Lchar x̃, v = p = and τ e = Lchar t̃, vchar vchar (ηs + ηp ) p̃ Lchar vchar (ηs + ηp ) τ̃ , Lchar e vchar ṽ, t= which give rise to the following dimensionless constants Re = β = ρvchar Lchar , η ηs , ηp + ηs We = λ vchar Lchar and The Weissenberg number, We, describes the magnitude of elastic to viscous effects. When either the Weissenberg number goes to zero or the solvent to total viscosity ratio, β goes to unity, the model approaches that of a Newtonian fluid. Contrarily, the value β = 0 corresponds to the upper convected Maxwell model, which is characterized by particularly strong viscoelastic effects. Note that we prefer to define a characteristic pressure in terms of a characteristic velocity rather than the other way around due the fact that we intend to impose fixed flow rates rather than fixed driving pressures (as in [6]). We have implemented the above approach to the log-conformation formulation for two dimensions in COMSOL Multiphysics version 4.3a, and the online version of the manuscript includes a MATLAB script for computation of the flow of an Oldroyd-B fluid past a confined cylinder using the COMSOL LiveLink interface. We use default settings for everything, except we force update of the Jacobian every time a time step is saved. 2.2 The Oldroyd-B model in Weak Form Converting the strong form of the Oldroyd-B model (2-5) to weak form is trivial for the equations that relate to the elastic stress (4) and continuity equation (3) as one just multiplies with test functions for the elastic stress and pressure, respectively. The evolution equation for the conformation tensor (5) requires SUPG stabilization, which means that the equation is multiplied3 with Atest + (2hmesh /vchar )v · ∇Atest rather than Atest , where hmesh is a characteristic size for the mesh. In other words a small amount of stream-wise diffusion is added to allow computation, but since the amount of diffusion is scaled with the mesh, the implementation still converges towards the solution without any artificial diffusion. Finally there is the momentum part of the Navier-Stokes equation (2), which is multiplied with a test function for the P Q that det(A) = i eλi = e iλi = eTr(s) , where λi are the eigenvalues of s. that the test functions for elastic stress and conformation tensor are multiplied element wise (Hadamard product), denoted with ◦. 2 Note 3 Note 5 velocity and integrated by parts: Z  0 = Ω    Dṽ ˜ · ṽtest dΩ ∇ · τ̃ s + τ̃ e − Re Dt̃ Z 0 = + + ˜ ṽtest dΩ − τ̃ s ◦ ∇ Ω  Z  Dṽ −1 ˜ − Re − Da α̃ṽ · ṽtest dΩ ∇τ̃ e Dt̃ Ω Z τ̃ s · n̂ · ṽtest dl (8) (9) ∂Ω The point is that the solvent stress is integrated by parts (Divergence theorem), but the elastic stress is not. This approach has been followed by other researchers so we have adopted it as well. The advantage is that the specification of τ̃ s on open boundaries is slightly more simple than τ̃ s + τ̃ e , but in fact we have tested both without finding any difference. At outlets we thus usually impose normal velocity, zero pressure and τ̃ s = −Ip̃. This can also be used for pressure driven inlets, but in that case one should supplement the constraint on the velocity with one on the conformation tensor. We impose the no-slip(ṽ = 0) at walls, which causes COMSOL to automatically enforce ṽtest = 0, such that the boundary integral in equation (9) vanishes. 3 Confined Cylinder The confined cylinder geometry is shown in figure 2, and it is a popular benchmark geometry due to the absence of geometric singularities. It can be used to verify an implementation by comparing the drag over the cylinder with a reference. We exploit the symmetry of the problem by modeling only half of the domain. Figure 2: The confined cylinder geometry with a (dashed) symmetry axis. The fluid comes in from the left, flows past the obstacle and exits through the boundary on the right (t̂ is the unit tangent vector). The cylinder radius R, and average inlet velocity are used as characteristic length and velocity, respectively. The blocking ratio is set at 2 (H = 2R), and the outlet length, L1 , is fixed at 10R. In agreement with references [2] the solvent to total viscosity ratio is taken as 0.59, and inertia is neglected (Re = 0). The dimensionless drag is computed as Z Drag = −2 τ̃ · n̂ · x̂T dl, with cylinder h i ˜ + (∇ṽ) ˜ T + τ̃ τ̃ = −p̃I + ηs ∇ṽ e where n̂ is the outward pointing normal vector. We impose the no-slip boundary condition ṽ = 0 at the walls, and fully developed boundary conditions 6 at the inlet ṽinlet A11,inlet A12,inlet A22,inlet 3 2  1 − (ỹ/2)2 x̂  2 ∂ (ṽ · x̂) 9 = 1+2 We = 1 + 8 ỹ 2 We2 ∂ ỹ ∂ (ṽ · x̂) 3 = We = − 4 ỹWe and ∂ ỹ = 1 = (10) Note that the intermediate expressions can be used in the case of a pressure driven setup. At the outlet we impose zero pressure together with zero normal viscous stress. poutlet = τ̃ s,outlet · n̂ = 0 and  −Ip · n̂ = 0 (11) At the symmetry line we impose zero normal velocity. ṽsym · n̂ = 0 3.1 (12) Solution Details In order to find a steady solution one can go through the following steps. #1 Solve for the case of a Newtonian fluid. #2 Use #1 as initial condition for a transient simulation running for 10 relaxation times. #3 Use #2 as initial guess for a non-linear solver to find a steady solution. We however choose to vary the Weissenberg number slowly using a regularized step function, st, such that a quasi steady state is achieved. We st(t̃) Westart + (Weend − Westart ) st(t̃), where  , t̃ < t̃start  0 3 = 0.5 + 1.5t̄ − 2 (t̄ ) , t̃start ≤ t̃ < t̃end  1 , t̃end ≤ t̃ = t̄ = t̃ − (t̃start + t̃end )/2 t̃end − t̃start (13) This way we compute a quasi-steady solution to a range of Weissenberg numbers using a single transient simulation. COMSOL uses a fully implicit transient scheme with adaptive stepsize, so slow variation of the Weissenberg number does not increase the total computation time [7]. 3.2 Results Since the time over which the Weissenberg number is varied is a numerical parameter, we choose to scale it with the mesh size, so convergence towards the steady case is achieved, i.e. Tstep = t̃end − t̃start = 8000(0.07/hmesh ) Figure 3 shows the drag as a function of the Weissenberg number for three different discretizations indicating convergence up to We = 0.6. We observe good agreement with the reference drag although the simplicity of the chosen mesh gives rise to a high computational cost. It is well known, that convergence at We = 0.7 is difficult, even with anisotropic mesh adaption [8]. In fact experimental [9] as well as theoretical evidence [10] exist to suggest, that a transition to a three dimensional flow pattern occurs, in other words the two 7 Figure 3: The drag on a confined cylinder is plotted as a function of the Weissenberg number for three discretizations together with the data of another study [2]. Figure 4: The trace of the conformation tensor is plotted on a logarithmic scale for the confined cylinder in the case of We = 0.51 and We = 0.7 (183k DOF). The acceleration of the flow in the wake of the cylinder generates a strand of highly elongated dumbbells. dimensional symmetric flow becomes unstable. The three-dimensional solution has the flow alternating between going above and below the cylinder, and one can think of this solution as having a weaker extensional character in the wake of the cylinder. When a viscoelastic fluid enters a region, where the extension rate times the relaxation time reaches a critical value, the dumbbell extension starts to grow exponentially in time. This kind of non-linearity explains the significant difference between the wake of extension at We = 0.51 and We = 0.7 illustrated in figure 4. 4 Bistable Cross-Slot   The FENE-CR model differs from the Oldroyd-B model in the sense, that λ is replaced by λ 1 − Trace(A)/a2max . In other words the relaxation time decreases as the dumbbell extension increases. This puts an upper bound a2max , on the trace of the conformation tensor corresponding to a maximum extension. The modification prevents an unbounded extension in free stagnation points such as the center of the cross-slot geometry shown in figure 5. This geometry is known to exhibit bistability experimentally [11] as well as numerically [12]. In this case we keep the average inlet velocity as characteristic velocity and the channel width H as 8 Figure 5: The cross-slot geometry is shown with the fluid coming in from the left and right, diverging at the center, before exiting through to the upper and lower outlets leaving a free stagnation point (red) in the center of the geometry. characteristic length scale. We impose the no-slip boundary condition at walls, but we have been unable to initialize the transient simulation with the correct expressions for the conformation tensor at the inlets. Therefore we have used s = I at the inlets rather than values based on the expressions listed in Appendix I. Furthermore we had to change the damping strategy of the nonlinear method from the default (constant) to automatic. At the outlet we impose the same boundary conditions (11) as for the cylinder. We compute the dissipation as an integral over the entire computation domain, Z h i ˜ + (∇ṽ) ˜ T dΩ. φ̃ = τ̃ : ∇ṽ (14) Ω where : is the Frobenius product. We choose Re = 0, a2max = 100 and β = 0.2 and the same transient solution procedure as described in section 3.1. We define the center of the domain, Ωc = x̃ ∈ |x̃|∞ < 0.5, as we find that the vorticity integrated over this is a good asymmetry parameter. We thus plot the square of this versus the Weissenberg number in figure 6(a). The system becomes bistable at We = 0.54 , which gives rise to a kink in the dissipation as shown in figure 6(b). The solution before the kink [figure 7(a) and (d)] has vertical, horizontal as well as 180-degree rotational symmetry, while the solutions after the kink [7(b), (c), (e) and (f)] only have 180-degree rotational symmetry, and we will refer to these solutions with reduced symmetry as ”asymmetric solutions”. The phenomenon can be understood by considering the velocity magnitude and conformation tensor trace as plotted in figures 7(a-c) and (d-e) respectively. The dumbbells are extended along the vertical axis, which causes a damping that delays the merging of the flows, and this gives rise to extra shear and thus also dissipation. When the solution becomes asymmetric, the damping is moved to the side of the channel, which lowers the dissipation. One can also think of the asymmetry as a flow pattern with less extension in the stagnation point, similar to the three-dimensional flow for the confined cylinder. 4 Based on personal correspondence we have concluded that the results of [3] are based on the FENE-MCR model, so no reference exists for the FENE-CR model. 9 Figure 6: The square of the vorticity integrated over the center of the domain Ωc is plotted as function of the Weissenberg number in (a). In (b) it is the dissipation that is plotted, which shows that the point of bistability occurs (We = 0.5) at the point of maximum dissipation and thus also maximum hydraulic resistance. 5 Three Inlets In this section we introduce the geometry shown in figure 8. It consists of a hexagon with in- and outlets attached in an alternating fashion. The geometry always has a vertical symmetry axis, but depending on whether a certain angle is equal to π/3 or not, it can also have 120-degree rotational symmetry. We define the distance to the center as in- and outlet lengths, and set equal inlet flow rates as well as outlet pressures. We set Lin = 6H and Lout = 8L, β = 0.1 and s = I at the inlets, but keep all other parameters identical to those used for the cross-slot. That is, the channel width and average inlet velocity are still used as characteristic length and velocity, respectively. If the angle is equal to π/3 radians, a single stagnation point exists in the center of the geometry regardless of whether we consider Stokes flow or just the regime of low elasticity. Two stagnation points however appear, if the angle is different from π/3 radians. Figure 9 shows the solution for angles of π/3, π/3.5 and π/4 radians at We = 0.37 (low) and We = 0.66 (moderate). For an angle of π/3 radians, the regime of low elasticity (a) gives a solution with 120-degree rotational- and vertical symmetry, while for moderate elasticity a solution with only 120-degree rotation symmetry appears (b). This latter solution involves the clockwise flow, but by symmetry we can argue that there must be an identical solution with counter-clockwise flow. The symmetry reduction also occurs for an angle of π/3.5 (c-d), in the sense that we go from only vertical symmetry to no symmetry. Regardless of the Weissenberg number, there are two stagnations points even though the flow rate between them is small at moderate elasticity (d). In other words turning up the elasticity causes the flow rate between the stagnation points to decrease significantly. As shown in (e-f) this is different for an angle of π/4 radians, as the elasticity tends to move the stagnation points apart and increase the flow 10 Figure 7: The magnitude of the velocity vector is plotted for three Weissenberg numbers in the cross geometry in (a-c), while red stream lines are plotted on top of the trace of the conformation tensor on a logarithmic scale in (d-f). Note that we plot streamlines as contours of the stream function instead of integrating the velocity field. Figure 8: The geometry with three in- and outlets is illustrated with the different length scales and angles involved. For an angle of π/3 the geometry has a symmetry axis through every channel as well as 120-degree rotational symmetry, while an angle different from π/3 breaks all symmetries except the one around the vertical axis. rate between them in what we call the ”straight flow solution”, that is the solution has identical symmetry properties at low and moderate elasticity. Using artificial initial conditions as listed in Appendix II, one can show that the clockwise, counterclockwise and the straight flow all exist as stable solutions at moderate elasticity for angles between π/4 and π/3.5. This means, that it is the connectivity with the solution at low elasticity that changes. Furthermore there is an unstable solution with a vertical symmetry axis for the angles π/3 and π/3.5. We calculate the integral of the vorticity around the center Ωo = x ∈ ||x||2 < 0.5, 11 Figure 9: The trace of the conformation tensor is plotted on a logarithmic scale together with streamlines (red) for two Weissenberg numbers in a geometry with three inlets. The case of an angle equal to π/3 (a-b), π/3.5 (c-d) and π/4 (d-e) radians is illustrated (see figure 8). Although not apparent from the streamlines, the upper outlet receives fluid from all three inlets in (d), and the flow between the stagnation points is reduced with a factor of 4.5 in (d) compared to (c). Any asymmetry around the vertical axis in (a), (c) and (e) is due to the mesh. as a measure of asymmetry and plot this as a function of the Weissenberg number for five angles in figure 10 (a), while the dissipation is plotted in figure 10 (b). It is clear that the point of maximum dissipation coincides with the point of bistability for the three largest angles (π/3, π/3.5, π/3.75), as it was also the case for the cross-slot in section 4. The dissipation for the two smaller angles also have maxima, but we do not believe they play the same role as the maxima for the smaller angles. This is due to the fact that the Weissenberg number at which the maximum occurs seems to decreases for both the rotation and straight flow solutions. This means that there does not exists an angle for which the two kind of maxima in figure 10 (b) meet, although there must exist an angle where all three solutions are connected to the solution at low elasticity. 6 Conclusion We have described the viscoelastic Oldroyd-B model with log-conformation reformulation in detail, and included a COMSOL implementation in the supplementary material. We have showed, that the code agrees with a reference in a benchmark geometry. Furthermore we have demonstrated that the code can be modified to describe the flow of a FENE-CR fluid, and that it can reproduce bistability in the cross-slot geometry. Finally we computed the flow in a geometry with three in- and outlets, which appears to feature three stable solutions. The connectivity of the solutions with the solution at low elasticity seems to depend on an angle in the geometry. 12 Figure 10: The square of the vorticity integrated over the center of the domain Ωo is plotted as functions of the Weissenberg number for five different angles in (a), while the dissipation is shown (b). The point of bistability coincides with the point of maximum dissipation for the three larger angles, as it was also the case for the cross-slot in figure 6. The maximum dissipation of the two smaller angles is indicated with crosses. References [1] F. Baaijens, S.H.A. Selen, H.P.W. Baaijens, G.W.M. Peters, and H.E.H. Meijer. Viscoelastic flow past a confined cylinder of a low density polyethylene melt. Journal of non-newtonian fluid mechanics, 68 (2-3):173–203, 1997. [2] M.A. Hulsen, R. Fattal, and R. Kupferman. Flow of viscoelastic fluids past a cylinder at high weissenberg number: stabilized simulations using matrix logarithms. Journal of Non-Newtonian Fluid Mechanics, 127(1):27–39, 2005. [3] G.N. Rocha, R.J. Poole, M.A. Alves, and P.J. Oliveira. On extensibility effects in the cross-slot flow bifurcation. Journal of Non-Newtonian Fluid Mechanics, 156(1):58–69, 2009. [4] A. Fortin, R. Guenétte, and R. Pierre. On the discrete evss method. Computer Methods in Applied Mechanics and Engineering, 189(1):121–139, 2000. [5] A. Kane, R. Guénette, and A. Fortin. A comparison of four implementations of the log-conformation formulation for viscoelastic fluid flows. Journal of Non-Newtonian Fluid Mechanics, 164(1-3):45–50, 2009. ISSN 0377-0257. [6] K. Ejlebjerg Jensen, P. Szabo, and F. Okkels. Topology optimization of viscoelastic rectifiers. Applied Physics Letters, 100(23):234102–234102, 2012. 13 [7] COMSOL Multiphysics. 4.3a user’s guide, 2013. [8] R. Guénette, A. Fortin, A. Kane, and J.F. Hétu. An adaptive remeshing strategy for viscoelastic fluid flow simulations. Journal of Non-Newtonian Fluid Mechanics, 153(1):34–45, 2008. [9] G.H. McKinley, R.C. Armstrong, and R.A. Brown. The wake instability in viscoelastic flow past confined circular cylinders. Philosophical Transactions: Physical Sciences and Engineering, 344(1671):265–304, 1993. [10] M. Sahin. Parallel large-scale simulation of viscoelastic fluid flow instabilities. In 17th International Workshop on Numerical Methods for Non-Newtonian Flows, 2012. [11] P.E. Arratia, CC Thomas, J. Diorio, and JP Gollub. Elastic instabilities of polymer solutions in crosschannel flow. Physical review letters, 96(14):144502, 2006. [12] RJ Poole, MA Alves, and PJ Oliveira. Purely elastic flow asymmetries. Physical review letters, 99(16): 164503, 2007. Appendix I: FENE-CR Inlet Conditions The developed flow conditions for a FENE-CR fluid 3 5 is listed in the following equations  1 − (ỹ/2)2 n̂ ṽinlet = −2 A11,inlet = a2max − 1   p a2max a2max − a4max + 8χ2 (a2max − 2) + 4χ2   A11,inlet + 1 A12,inlet = χ 1 − a2max A22,inlet = 1, where ∂ (ṽ · x̂) = 3We(ỹ/4)(n̂ · x̂) χ = We ∂ ỹ Appendix II: Artificial Initial Conditions for the Geometry with Three In- and Outlets One can find stable solutions not connected to the solution at low elasticity using artificial initial conditions. A rotating solution for an angle of π/4 (see figure 8) can in example be found by imposing a rotating force in the center of the geometry,  F =  (x̂ỹ − ŷx̃)/2 0 , ||r̃|| , ||r̃|| ≤ 0.5/ sin(π/6) , > 0.5/ sin(π/6) where r̃ is the position vector and || · · · || is the euclidean norm. This is then multiplied with a step function as given by equation (13), such that the force vanishes beyond some critical time. Alternatively one might find a solution going straight through a geometry with an angle of π/3.5 using an upward force,  ŷ(0.25 − x̃2 ) , ||r̃|| ≤ 0.5/ sin(π/6) ↑ F = . 0 , ||r̃|| > 0.5/ sin(π/6) 5 Note that one has to set A11,inlet = 1 when χ2 < . 14
5cs.CE
1 The MAC Resolvability Region, Semantic Security and Its Operational Implications arXiv:1710.02342v2 [cs.IT] 22 Jan 2018 Matthias Frey, Igor Bjelaković and Sławomir Stańczak Technische Universität Berlin Abstract—Building upon previous work on the relation between secrecy and channel resolvability, we revisit a secrecy proof for the multiple-access channel (MAC) from the perspective of resolvability. In particular, we refine and extend the proof to obtain novel results on the second-order achievable rates. Then we characterize the resolvability region of the memoryless MAC by establishing a conceptually simple converse proof for resolvability that relies on the uniform continuity of Shannon’s entropy with respect to the variational distance. We discuss the operational meaning of the information theoretic concept of semantic security from the viewpoint of Ahlswede’s General Theory of Information Transfer and show interesting implications on the resilience against a large class of attacks even if no assumption can be made about the distributions of the transmitted messages. Finally we provide details on how the resolvability results can be used to construct a wiretap code that achieves semantic security. I. I NTRODUCTION With an increasing number of users and things being connected to each other, not only the overall amount of communication increases, but also the amount of private and personal information being transferred. This information needs to be protected against various attacks. In fact, for some emerging applications such as e-health services where sensitive medical data might be transmitted using a Body Area Network, the problem of providing secrecy guarantees is a key issue. As discovered by Csiszár [12] and later more explicitly by Bloch and Laneman [10] and investigated by Yassaee and Aref [45] for the multiple-access case, the concept of channel resolvability can be applied to provide such guarantees; in essence, an effective channel is created by an encoder in such a way that the output observed by a wiretapper is almost independent of the transmitted messages, which renders the output useless for the wiretapper. Moreover, the concept can also be used in point-to-point settings as a means of exploiting channel noise to convey randomness in such a way that the distribution observed at the receiver can be accurately controlled by the transmitter. In this paper, we explore channel resolvability in a multipleaccess setting in which there is no communication between the transmitters, yet they can control the distribution observed at the receiver in a non-cooperative manner. More precisely, given a multiple-access channel with an ideal or desired The work was supported by the German Research Foundation (DFG) under grant STA864/7-1 and by the German Federal Ministry of Education and Research under grant 16KIS0605. This paper was presented in part at the 4th Workshop on Physical-Layer Methods for Wireless Security held during the 2017 IEEE Conference on Communications and Network Security [21]. X ∼ qX W Y ∼ qY qZ|X,Y Z ∼ qZ Fig. 1. Idealized input-output model of the multiple-access channel. M1 M2 C1 C2 Xn Yn W n qZ n |X n ,Y n Z n ∼ pZ n |C1 ,C2 Fig. 2. Resolvability codebooks are used as an input to the MAC channel to create a channel output at the receiver that well approximtes the given ideal output distribution (i.e. the output looks as if n independent instances of the output in Figure 1 were observed). output distribution under some input distribution (see Figure 1), the resolvability problem for this multiple-access channel consists in finding for each transmitter a finite sequence of codewords (called a codebook) such that if each transmitter chooses a codeword at random and independently of the other transmitter, then the MAC output well approximates the given ideal output distribution. The coding setting is illustrated in Figure 2 where each codeword is a tuple of n elements of the input alphabet and n is referred to as the block length. The first part of the paper deals with the resolvability problem for the memoryless MAC, while in the second part we show how the concept of channel resolvability can be exploited to achieve strong operational secrecy guarantees over a wiretap MAC. A. Literature To the best of our knowledge, the concept of approximating a desired output distribution over a communication channel using as little randomness as possible at the transmitter was first introduced by Wyner [43], who used normalized KullbackLeibler divergence as a similarity measure between the actual and the desired output distribution. Han and Verdú [25] introduced the term channel resolvability for a similar concept; however, instead of the Kullback-Leibler divergence, they used the variational distance as a metric and showed, among others, the existence of a codebook that achieves an arbitrarily small variational distance by studying the expected variational distance of a random codebook. A stronger result stating that the probability of drawing an unsuitable random codebook is doubly exponentially small is due to Cuff [14]. Related results were proposed before by 2 Csiszár [12] and by Devetak [16] for the quantum setting, who based his work on the non-commutative Chernoff bound [3]. Further secrecy results that are either based on or related to the concept of channel resolvability can be found in Hayashi [27], Bloch and Laneman [10], Hou and Kramer [29], and Wiese and Boche [42], who applied Devetak’s approach to a multipleaccess setting. Cuff [14] presented a result on the secondorder rate; a related result was proposed by Watanabe and Hayashi [41]. Resolvability for MACs has been explored by Steinberg [38] and later by Oohama [36]. Explicit lowcomplexity codebooks for the special case of symmetric MACs have been proposed by Chou, Bloch and Kliewer [11]. Converse results for channel resolvability have been proposed by Han and Verdú [25] and, for the multiple-access channel, by Steinberg [38]. However, since these studies consider channels with memory and allow for arbitrary input and output distributions (in particular, they need not be identical or independent across channel uses), these results are not directly transferrable to the memoryless case considered in this paper. Converse proofs for the memoryless single-user case were given by Wyner [43] and later by Hou [30]; however, instead of the variational distance considered in this study, they assumed the Kullback-Leibler divergence as a similarity measure so that the results of [30] cannot be applied directly either. Security concepts from cryptography have been redefined as secrecy concepts for the information theoretic setting by Bellare, Tessano and Vardy in [7] and [6], where the authors also discussed the interrelations between these concepts; however, the proposed schemes involve some cryptographic elements, such as shared secrets between transmitter and legitimate receiver or shared (public) randomness. To the best of our knowledge, there are currently no results explicitly linking the concept of semantic security to the operational meaning in communication theory. Wiretap codes that achieve semantic security over the Gaussian channel have been proposed by Ling, Luzzi, Belfiore and Stehlé [33]. Thangaraj [39] showed how to construct semantically secure wiretap codes for the single-user channel and already noted the link to channel resolvability. Goldfeld, Cuff and Permuter [23][22] showed how channel resolvability techniques can be used to obtain wiretap codebooks that achieve semantic security over the single-user channel; the technique had been previously sketched by Cuff [15]. Liang and Poor [32], Ekrem and Ulukus [18], as well as Liu and Kang [34] provided converse results for related multipleaccess wiretap channels, but these results do not carry over in a straightforward fashion to the converse results for the channel model we consider in this paper to study achievable secrecy rates. II. OVERVIEW AND M AIN C ONTRIBUTIONS The main contributions of this study can be summarized as follows: 1) We characterize the resolvability region of the discrete memoryless multiple-access channel. Section IV proves the direct theorem and a result on the second-order rate, while Section V deals with the converse theorem. R2 I(Y ; Z|V, X) resolvability region capacity region I(Y ; Z|V ) R1 I(X; Z|V ) I(X; Z|V, Y ) Fig. 3. The MAC capacity region and the resolvability region as characterized in Theorem 1. 2) In Section VI we show how resolvability can be exploited to achieve strong secrecy and semantic security, and discuss operational implications of semantic security. In the remainder of this section, we discuss our results in more detail and explain their implications in the context of secrecy and security. We start with a basic definition and refer the reader to Section III for further definitions. Given a multiple-access channel W = (X , Y, Z, qZ|X,Y ) and an output distribution qZ , we call a rate pair (R1 , R2 ) achievable if there are sequences of codebooks (C1,n )n≥1 , (C2,n )n≥1 of rates R1 and R2 , respectively, and each of block length n such that lim kpZ n |C1,n ,C2,n − qZ n kTV = 0 . n→∞ The resolvability region SW,qZ is defined as the closure of the set of all achievable rate pairs. The main resolvability result in this work is the following characterization of SW,qZ . Theorem 1. Let W = (X , Y, Z, qZ|X,Y ) be a channel and qZ an output distribution. Let [ ′ := SW,q {(R1 , R2 ) :I(X, Y ; Z|V )≤R1 + R2 Z pV ,qX|V ,qY |V ∧ I(X; Z|V ) ∧ I(Y ; Z|V ) ≤R1 ≤R2 }, where pV ranges over all probability mass functions on the alphabet V := {1, . . . , |Z| + 3} and qX|V , qY |V range over all conditional probability distributions such that, for every v ∈ V and z ∈ Z, XX qZ (z) = qX|V (x|v)qY |V (y|v)qZ|X,Y (z|x, y), x∈X y∈Y and the mutual information is computed with respect to the induced joint probability mass function. Then ′ = SW,qZ . SW,q Z 3 X qZlegit |X,Y by distinguishing security, has some very strong operational implications, some of which we explore from the point of view of Ahlswede’s General Theory of Information Transfer [1]. Zlegit W Y qZtap |X,Y III. N OTATION , D EFINITIONS Ztap Fig. 4. Channel resolvability problem as part of the wiretap coding problem. Figure 3 illustrates the resolvability region as characterized in this theorem and the well-known capacity region of the MAC. ′ Theorem 8 in Section V yields the converse part SW,q ⊇ Z SW,qZ of the above theorem. The proof is remarkably simple and mainly relies on the fact that entropy is continuous with respect to the variational distance. ′ The direct part SW,q ⊆ SW,qZ follows from a stronger Z statement proven in Section IV. In particular, the statement shows that not only does a sequence of codebook pairs achieving the rates exist, but if they are constructed randomly, the probability of drawing a bad pair of codebooks vanishes doubly exponentially with increasing block length. Precisely, we show the following. Theorem 2. Let W = (X , Y, Z, qZ|X,Y ) be a given channel; let qX and qY be input distributions, R1 > I(X; Z) (1) R2 > I(Y ; Z) (2) R1 + R2 > I(X, Y ; Z). (3) Then there exist γ1 , γ2 > 0 such that for sufficiently large block length n, the codebook distributions of block length n and rates R1 and R2 satisfy  PC1 ,C2 kpZ n |C1 ,C2 − qZ n kTV > exp(−γ1 n) ≤ exp (− exp (γ2 n)) . To prove this theorem, we revisit the proof in [42], thereby focusing on channel resolvability. We use a slightly different technique than that in [14], which we extend to the multipleaccess case to provide an explicit statement and a more intuitive proof for a result only implicitly contained in [42]. In an excursus, we show how a slight refinement of the same technique similarly as in [14] yields also a partial second-order result. In Section VI, we discuss different notions of secrecy and the interconnection between them. In particular, in addition to the classical concept of strong secrecy, we review the notions of distinguishing security and semantic security, both of which originate from the field of cryptography. Distinguishing security allows us to decompose the problem for wiretap coding into a problem of coding for the MAC and a channel resolvability problem as shown in Figure 4. The concept of semantic security, on the other hand, which is implied AND P REREQUISITES The operations log and exp use Euler’s number as a basis, + and all information quantities are given in nats. [·] denotes the maximum of its argument and 0. A channel W = (X , Y, Z, qZ|X,Y ) is given by finite input alphabets X and Y, a finite output alphabet Z and a collection of probability mass functions qZ|X,Y on Z for each pair (x, y) ∈ X × Y. The random variables X, Y and Z represent the two channel inputs and the channel output, respectively. Input distributions for the channel are probability mass functions on X and Y denoted by qX and qY , respectively. We define an induced joint distribution qX,Y,Z on X × Y × Z by qX,Y,Z (x, y, z) := qX (x)q Y (y)qP Z|X,Y (z|x, y) and the output P distribution qZ (z) := x∈X y∈Y qX,Y,Z (x, y, z) is the marginal distribution of Z. By a pair of codebooks of block length n ≥ 1 and rates R1 ≥ 0 and R2 ≥ 0, we mean finite sequences exp(nR ) exp(nR ) C1 = (C1 (m))m=1 1 and C2 = (C2 (m))m=1 2 , where n the codewords C1 (m) ∈ X and C2 (m) ∈ Y n are finite sequences of elements of the input alphabets. We define a probability distribution PC1 ,C2 on these codebooks as i.i.d. drawings in each component of each codeword according to qX and qY , respectively. Accordingly, we define the output distribution induced by C1 and C2 on Z n by pZ n |C1 ,C2 (z n ) := exp(−n(R1 + R2 )) exp(nR1 ) exp(nR2 ) · X m1 =1 X m2 =1 qZ n |X n ,Y n (z n |C1 (m1 ), C2 (m2 )). Given probability distributions P and Q on a finite set A with mass functions p and q, respectively, and positive α 6= 1, the Rényi divergence of order α of P from Q is defined as X 1 Dα (P ||Q) := log p(a)α q(a)1−α . α−1 a∈A Furthermore, we define the variational distance between P and Q (or between their mass functions) as X 1X + kp − qkTV := |p(a) − q(a)| = [p(a) − q(a)] . 2 a∈A a∈A Given random variables A, B and C distributed according to rA,B,C , we define the (conditional) information density as i(a; b) := log rB|A,C (b|a, c) rB|A (b|a) , i(a; b|c) := log . rB (b) rB|C (b|c) The (conditional) mutual information is the expected value of the (conditional) information density. We use H(A) and H(A|B) to denote Shannon’s (conditional) entropy. The following inequality was introduced in [8] and [20]; we use a refinement here which follows e.g. from [5]. Theorem 3 (Berry-Esseen Inequality). Given a sequence (Ak )nk=1 of i.i.d. copies of a random variable A on the reals 4 3 with EA = 0, EA2 = σ 2 < ∞ and E |A| = ρ < ∞, define Ā := (A1 + · · · + An )/n. dis√ Then the cumulative √ tribution Rfunctions F (a) := P(Ā n/σ ≤ a) of Ā n/σ and a Φ(a) := −∞ 1/(2π) exp(−x2 /2)dx of the standard normal distribution satisfy ρ ∀a ∈ R : |F (a) − Φ(a)| ≤ 3 √ . σ n r and a family (rb′ ′ )b′ ∈B′ of elements of P such that for each k ∈ {1, . . . , n}, Z X fk (rb )P (db) = fk (rb′ ′ )r(b′ ). (4) We further use variations of the concentration bounds introduced in [28]. In this section, we prove Theorem 2. Together with the observation that the resolvability region SW,qZ is convex, by a standard time sharing argument, this yields the direct part of Theorem 1 as an immediate consequence. First, we show a slightly weaker version of Theorem 2. Theorem 4 (Chernoff-Hoeffding Bound). Suppose A = P n k=1 Ak , where the random variables in the sequence (Ak )nk=1 are independently distributed with values in [0, 1] and EA ≤ µ. Then for 0 < δ < 1,  2  δ P(A > µ(1 + δ)) ≤ exp − µ . 3 This version can e.g. be found in [17, Ex. 1.1]. We also use an extension of the Chernoff-Hoeffding bound for dependent variables due to Janson [31, Theorem 2.1], of which we state only a specialized instance that is used in this paper. P Theorem 5 (Janson [31]). Suppose A = nk=1 Ak , where the random variables in the sequence (Ak )nk=1 take values in [0, 1] and can be partitioned into χ ≥ 1 sets such that the random variables in each set are independently distributed. Then, for δ > 0,   δ2 . P(A ≥ EA + δ) ≤ exp −2 χ·n We also use [13, Lemma 2.7] which provides bounds for the entropy difference of random variables that are close to each other in terms of their variational distance. Note that the definition of the variational distance in [13] differs from the one used in this paper by a factor of 1/2, and we state the lemma conforming with our definitions. Lemma 1. Let A and B be random variables on an alphabet A, distributed according to probability mass functions rA and rB , respectively. If krA − rB kTV = δ ≤ 1/4, then 1 δ |H(A) − H(B)| ≤ − δ log . 2 2|A| Remark 1. Since the function −δ/2 · log(δ/2|A|) is nondecreasing (in particular) for 0 ≤ δ ≤ 1/4, the assumption can be weakened to krA − rB kTV ≤ δ ≤ 1/4. Furthermore, for the proof of the converse theorem, we need a lemma that is based on the Fenchel-Eggleston-Carathéodory theorem and appeared first in [4] and of which we state the version from [19, Appendix C]. Lemma 2. Let A be a finite set and B an arbitrary set. Suppose that P is a connected compact subset of probability mass functions on A and consider a family (rb )b∈B of elements of P. Suppose further that (fk )nk=1 are continuous functions mapping from P to the reals and B is a random variable on B distributed according to some probability measure P . Then there exist a random variable B ′ on some alphabet B ′ of cardinality at most n with probability mass function b′ ∈B′ B IV. P ROOF OF THE D IRECT T HEOREM Theorem 6. Suppose W = (X , Y, Z, qZ|X,Y ) is a channel, qX and qY are input distributions, R1 > I(X; Z|Y ) R2 > I(Y ; Z). Then there exist γ1 , γ2 > 0 such that for large enough block length n, the codebook distributions of block length n and rates R1 and R2 satisfy  PC1 ,C2 kpZ n |C1 ,C2 − qZ n kTV > exp(−γ1 n) ≤ exp (− exp (γ2 n)) . (5) The full Theorem 2 can then be proven with a time sharing argument which we detail at the end of this section. In the proof of Theorem 6, we consider two types of typical sets: n := {(xn , y n , z n ) : i(xn ; z n |y n ) ≤ n(I(X; Z|Y ) + ε)} T1,ε n T2,ε := {(y n , z n ) : i(y n ; z n ) ≤ n(I(Y ; Z) + ε)}. We split the variational distance in atypical and typical parts as follows, where Patyp,1 , Patyp,2 and Ptyp (z n ) are defined by (6), (7) and (8) shown on the next page. kpZ n |C1 ,C2 − qZ n kTV +  X pZ n |C1 ,C2 (z n ) n −1 = qZ n (z ) qZ n (z n ) z n ∈Z n X + ≤Patyp,1 + Patyp,2 + qZ n (z n ) [Ptyp (z n ) − 1] . (9) z n ∈Z n Remark 2. The denominator of the fraction is almost surely not equal to 0 as long as the numerator is not equal to 0. We implicitly let the summation range only over the support of the denominator, as we do in all further summations. So the theorem can be proven by considering typical and atypical terms separately. But first, we prove two lemmas to help us to bound the typical and the atypical terms. Lemma 3 (Bound for typical terms). Given a block length n, ε > 0, 0 < δ < 1, random variables A, B and C on finite alphabets A, B and C respectively with joint probability mass exp(nR) function rA,B,C , a rate R and a codebook C = (C(m))m=1 with each component of each codeword drawn i.i.d. according to rA , for any bn ∈ B n and cn ∈ C n , we have 5 exp(nR1 ) exp(nR2 ) Patyp,1 := X exp(−n(R1 + R2 )) X exp(−n(R1 + R2 )) m1 =1 z n ∈Z n Patyp,2 := X X X exp(−n(R1 + R2 )) m2 =1 m1 =1 P̌ := PC  exp(nR) X X m1 =1 exp(nR1 ) exp(nR2 )  n qZ n |X n ,Y n (z n |C1 (m1 ), C2 (m2 ))1(C1 (m1 ),C2 (m2 ),zn )∈T / 1,ε (6) X n qZ n |X n ,Y n (z n |C1 (m1 ), C2 (m2 ))1(C2 (m2 ),zn )∈T / 2,ε (7) m2 =1 exp(nR1 ) exp(nR2 ) z n ∈Z n Ptyp (z n ) := X exp(−nR) m=1 m2 =1 qZ n |X n ,Y n (z n |C1 (m1 ), C2 (m2 )) n 1(C (m ),C (m ),z n )∈T n 1(C2 (m2 ),zn )∈T2,ε 1 1 2 2 1,ε qZ n (z n ) (8) A ⊆ X n × Y n × Z n , δ > 0, µ ≥ P((X n , Y n , Z n ) ∈ A) as well as rates R1 and R2 and codebooks distributed according to PC1 ,C2 defined in Section III, we have rC n |An ,B n (cn |C(m), bn ) rC n |B n (cn |bn )  · 1(C(m),bn,cn )∈Tεn > 1 + δ  X P̂ := PC1 ,C2  δ2 ≤ exp − exp(−n(I(A; C|B) + ε − R)) , 3  where the typical set is defined as exp(−n(R1 + R2 )) z n ∈Z n exp(nR1 ) exp(nR2 ) X X m2 =1 m1 =1 Tεn := {(an , bn , cn ) : i(an ; cn |bn ) ≤ n(I(A; C|B) + ε)}. qZ n |X n ,Y n (z n |C1 (m1 ), C2 (m2 )) ! 1(C1 (m1 ),C2 (m2 ),zn )∈A > µ(1 + δ) (10) ≤ exp(−2δ 2 µ2 exp(n min(R1 , R2 ))). Proof. We have  exp(nR) X  exp(−n(I(A; C|B) + ε)) P̌ = PC Proof. We have  exp(nR1 ) exp(nR2 ) m=1 n · P̂ =PC1 ,C2  n rC n |An ,B n (c |C(m), b ) · 1(C(m),bn ,cn )∈Tεn rC n |B n (cn |bn )  > exp(−n(I(A; C|B) + ε − R))(1 + δ) . By the definition of Tεn in (10), the summands are at most 1, and furthermore, the expectation of the sum can be bounded as  exp(nR) X exp(−n(I(A; C|B) + ε)) EC  m=1 exp(nR) ≤ X  rC n |An ,B n (cn |C(m), bn ) 1(C(m),bn,cn )∈Tεn  · rC n |B n (cn |bn ) exp(−n(I(A; C|B) + ε)) m=1 rC n |An ,B n (cn |C(m), bn ) rC n |B n (cn |bn ) = exp(−n(I(A; C|B) + ε − R)). · EC   Now applying Theorem 4 to the above shows the desired probability statement and completes the proof. Lemma 4 (Bound for atypical terms). Given a channel W = (X , Y, Z, qZ|X,Y ), input distributions qX and qY , some set X m1 =1 X m2 =1 X z n ∈Z n qZ n |X n ,Y n (z n |C1 (m1 ), C2 (m2 ))1(C1 (m1 ),C2 (m2 ),zn )∈A ! > exp(n(R1 + R2 ))(µ + µδ)  exp(nR1 ) exp(nR2 ) ≤PC1 ,C2  X m1 =1 n X m2 =1 X z n ∈Z n qZ n |X n ,Y n (z |C1 (m1 ), C2 (m2 ))1(C1 (m1 ),C2 (m2 ),zn )∈A !    n n n > exp n(R1 + R2 ) P((X , Y , Z ) ∈ A) + µδ exp(2n(R1 + R2 ))µ2 δ 2 exp(n max(R1 , R2 )) exp(n(R1 + R2 )) 2 2 = exp(−2δ µ exp(n min(R1 , R2 ))),  ≤ exp −2  where the inequality follows from Theorem 5 by observing that the innermost sum is confined to [0, 1], the two outer summations together have exp(n(R1 + R2 ) summands which can be partitioned into exp(n(max(R1 , R2 )) sets with exp(n min(R1 , R2 )) independently distributed elements each, and the overall expectation of the term is exp(n(R1 + R2 )P((X n , Y n , Z n ) ∈ A). We now have all the technical ingerdients needed to finish the proof of Theorem 6. 6 Proof of Theorem 6. In order to bound Patyp,1 , we observe that for any α > 1, we can bound n / T1,ε ) PX n ,Y n ,Z n ((X n , Y n , Z n ) ∈  n n qZ n |X n ,Y n (Z |X , Y n ) = PX n ,Y n ,Z n qZ n |Y n (Z n |Y n ) (11) = PX n ,Y n ,Z n ≤EX n ,Y n ,Z n n n n qZ n |X n ,Y n (Z |X , Y ) qZ n |Y n (Z n |Y n ) (12) α−1 > exp(n(α − 1)(I(X; Z|Y ) + ε))  α−1 ! qZ n |X n ,Y n (Z n |X n , Y n ) qZ n |Y n (Z n |Y n ) · exp(−n(α − 1)(I(X; Z|Y ) + ε)) = exp n(α − 1) · (Dα PX,Y,Z ||PX|Y PZ|Y PY  − I(X; Z|Y ) − ε) ≤ exp(−nβ),  (13)  (14) (15) (16) where (16) holds as long as β < (α − 1)(I(X; Z|Y ) + ε − Dα PX,Y,Z ||PX|Y PZ|Y PY ). We can achieve this for sufficiently small β > 0 as long as α > 1 and I(X; Z|Y ) + ε − Dα PX,Y,Z ||PX|Y PZ|Y PY > 0. In order to choose an α > 1 such that the latter requirement holds, note that since our alphabets are finite, the Rényi divergence is also finite and thus it is continuous and approaches the Kullback-Leibler divergence for α tending to 1 [40], which is in this case equal to the mutual information term. n and We apply Lemma 4 with A = (X n × Y n × Z n ) \ T1,ε δ = 1 to obtain PC1 ,C2 (Patyp,1 > 2 exp(−nβ)) ≤ exp(−2 exp(n(min(R1 , R2 ) − 2β))). (17) Proceeding along similar lines of reasoning including another n application of Lemma 4 with A = X n × ((Y n × Z n ) \ T2,ε ) and δ = 1, we show that if β > 0 is small enough, PC1 ,C2 (Patyp,2 > 2 exp(−nβ)) ≤ exp(−2 exp(n(min(R1 , R2 ) − 2β))). (18) As for the typical term, we first observe that for any fixed y n and z n , we can apply Lemma 3 with A = X, B = Y , C = Z and δ = exp(−nβ) to obtain PC1 (Ptyp,1 (y n , z n ) > 1 + exp(−nβ))   1 ≤ exp − exp(−n(I(X; Z|Y ) + ε + 2β − R1 )) , (19) 3 where we used exp(nR1 ) n n Ptyp,1 (y , z ) := X exp(−n(R1 )) m1 =1 · q n n (z |C1 (m1 ), y ) n . 1(C1 (m1 ),yn ,zn )∈T1,ε (20) qZ n |Y n (z n |y n ) Z n |X n ,Y n y n ∈Y n and bound for arbitrary but fixed z n  > exp(n(I(X; Z|Y ) + ε))  We define a set of codebooks \ Czn := {C1 : Ptyp,1 (y n , z n ) ≤ 1 + exp(−nβ)} (21) P̃ := PC1 ,C2 (Ptyp (z n ) > 1 + 3 exp(−nβ) | C1 ∈ Czn ) in (78) to (81) in the appendix. We can now put everything together as shown in (82) to (84) in the appendix. What remains is to choose γ1 and γ2 such that (5) holds. First, we have to choose ε and β small enough such that δ̂ := min min(R1 , R2 ) − 2β, R1 − 2β − ε − I(X; Z|Y ),  R2 − 2β − ε − I(Y ; Z) is positive. Since there have so far been no constraints on β and ε except that they are positive and sufficiently small, such a choice is possible provided R1 > I(X; Z|Y ) and R2 > I(Y ; Z). The theorem then follows for large enough n by choosing positive γ2 < δ̂ and γ1 < β. Obviously, we can reverse the roles of X and Y in Theorem 6. In order to prove the full Theorem 2, however, we need time sharing between the two corner points, as we detail in the following proof. Proof of Theorem 2. We are given R1 and R2 satisfying (1), (2) and (3). In this proof, we show how to find a time sharing parameter λ̂ with the following properties: R1 > λ̂I(X; Z) + (1 − λ̂)I(X; Z|Y ) R2 > λ̂I(Y ; Z|X) + (1 − λ̂)I(Y ; Z) We can then conclude the proof by applying Theorem 6 twice: First we apply it with a block length of (1 − λ̂)n and then we apply it with reversed roles of X and Y and a block length of λ̂n. To this end, we define functions mapping from [0, 1] to the reals R1 (λ) := λI(X; Z) + (1 − λ)I(X; Z|Y ) R2 (λ) := λI(Y ; Z|X) + (1 − λ)I(Y ; Z). Note that R1 (λ) and R2 (λ) are continuous in λ, R1 (0) = I(X; Z|Y ), R1 (1) = I(X; Z), R2 (0) = I(Y ; Z) and R2 (1) = I(Y ; Z|X). If R1 > I(X; Z|Y ) or R2 > I(Y ; Z|X), there is nothing to prove since Theorem 6 can be applied directly. So we assume R1 ≤ I(X; Z|Y ) and R2 ≤ I(Y ; Z|X), which means that R1 (λ) is strictly decreasing and R2 (λ) is strictly increasing in λ and we can find λ1 , λ2 ∈ [0, 1] such that R1 (λ1 ) = R1 and R2 (λ2 ) = R2 . Next, we note that λ1 < λ2 because under the assumption λ1 ≥ λ2 , we can obtain the contradiction R1 + R2 ≤ R1 (λ2 ) + R2 (λ2 ) = I(X, Y ; Z). So we can find λ̂ with λ1 < λ̂ < λ2 . We then have R1 > R1 (λ̂) and R2 > R2 (λ̂), as required. 7 A. Second Order Rates Although the probability statement in Theorem 2 is quite strong in the sense that it guarantees a doubly exponentially low probability of drawing a bad pair of codebooks, a weak spot of the theorem as stated is that the constants γ1 and γ2 are guaranteed to be positive, but can be arbitrarily small depending on the channel, the input distribution and the codebook rates and that the theorem is only valid for sufficiently large n. So in particular, if we are given a fixed n (even a very large one), we cannot draw any conclusions for the error probabilities of codebooks of block length n. However, a closer examination of the proof shows that stronger statements are possible in this case. In this subsection, we give details how the proof of Theorem 2 can be modified so that it yields the following second-order result. Theorem 7. Given a channel W = (X , Y, Z, qZ|X,Y ), input distributions qX and qY , ε ∈ (0, 1), let the central second and absolute third moment of i(X; Z|Y ) be V1 and ρ1 , respectively; analogously, we use V2 and ρ2 to denote the central second and absolute third moment of i(Y ; Z). Suppose the rates R1 , R2 depend on n in the following way: r log n V1 −1 Q (ε) + c (22) R1 = I(X; Z|Y ) + n rn V2 −1 log n R2 = I(Y ; Z) + Q (ε) + c , (23) n n where Q := 1 − Φ with Φ as defined in the statement of Theorem 3, and c > 1. Then, for any d ∈ (0, c − 1), we have  PC1 ,C2 kpZ n |C1 ,C2 − qZ n kTV >    3 1 +√ (µ1 + µ2 ) 1 + √ n n   2 2 2 min(µ1 , µ2 ) exp(n min(R1 , R2 )) ≤ 2 exp − n   1 + 2 exp n(log|Z| + log|Y|) − nc−d−1 , 3 where for both k = 1 and k = 2,   ρk d log n −1 + 3√ µk := Q Q (ε) + √ nVk Vk2 n tends to ε for n → ∞. An immediate consequence of this theorem is the following corollary that follows by applying the theorem with the roles of X and Y reversed. Corollary 1. Theorem 7 holds with (22) and (23) replaced by r log n V1 −1 Q (ε) + c R1 = I(X; Z) + n n r V2 −1 log n Q (ε) + c R2 = I(Y ; Z|X) + n n and V1 , ρ1 , V2 and ρ2 redefined to be the second and third moments of i(X; Z) and i(Y ; Z|X), respectively. Remark 3. The question of how the achievable second-order rates behave near the line connecting these two corner points should be a subject of further research. n and Proof of Theorem 7. We consider the typical sets T1,ε 1 n T2,ε2 , where for k = 1, 2, we choose εk > 0 to be r log n Vk −1 Q (ε) + d . (24) εk := n n The definitions (6), (7) and (8) change accordingly. In order to bound Patyp,1 , we use Theorem 3 to obtain n ) / T1,ε PX n ,Y n ,Z n ((X n , Y n , Z n ) ∈ 1 ! n X 1 = PX n ,Y n ,Z n (i(Xk ; Zk |Yk ) − I(X; Z|Y )) > ε1 n k=1  r  n ρ1 + 3 √ = µ1 . ≤ Q ε1 V1 2 V1 n √ An application of Lemma 4 with δ = 1/ n yields    1 PC1 ,C2 Patyp,1 > µ1 1 + √ n   2 2µ1 ≤ exp − exp(n min(R1 , R2 )) . (25) n Reasoning along similar lines shows n ) ≤ µ2 / T2,ε PY n ,Z n ((Y n , Z n ) ∈ 2 so that a further application of Lemma 4 yields    1 PC1 ,C2 Patyp,2 > µ2 1 + √ n   2 2µ ≤ exp − 2 exp(n min(R1 , R2 )) . (26) n For the typical term, we use the definitions (20) and (21) with n , and observe that for any fixed y n and z n , the typical set T1,ε 1 we can√apply Lemma 3 with A = X, B = Y , C = Z and δ = 1/ n to obtain   1 PC1 Ptyp,1 (y n , z n ) > 1 + √ ) n   1 ≤ exp − exp(−n(I(X; Z|Y ) + ε1 − R1 )) . (27) 3n Now proceeding in a similar manner as in (78) to (81) shows   3 PC1 ,C2 Ptyp (z n ) > 1 + √ | C1 ∈ Czn n   1 ≤ exp − exp(−n(I(Y ; Z) + ε2 − R2 )) , 3n √ where there is no assumption on n because 1/ n ≤ 1 for all n ≥ 1. The theorem then follows from (85) to (88) in the appendix. 8 V. C ONVERSE T HEOREM In this section, we establish the following theorem which implies the converse part of Theorem 1. Theorem 8. Let W = (X , Y, Z, qZ|X,Y ) be a channel, qX , qY input distributions (with induced output distribution qZ ) and (C1,n )n≥1 , (C2,n )n≥1 sequences of codebooks of block lengths n and rates R1 and R2 , respectively, such that lim kpZ n |C1,n ,C2,n − qZ n kTV = 0. n→∞ Then there are random variables X on X , Y on Y, Z on Z and V on alphabet V = {1, . . . , |Z| + 3} with a joint probability mass function pV pX|V pY |V qZ|X,Y such that the marginal distribution of Z is qZ and I(X, Y ; Z|V ) ≤ R1 + R2 I(X; Z|V ) ≤ R1 I(Y ; Z|V ) ≤ R2 . even if the same probability mass functions have already been introduced using a different notation before; e.g. we will write pZ̃ n instead of pZ n |C1 ,C2 . Additionally, we define a random variable V̂ that is uniformly distributed on {1, . . . , n} and independent of the other random variables. We furthermore define X̂ := X̃V , Ŷ := ỸV and Ẑ := Z̃V , noting that the conditional distribution of Ẑ given X̂ and Ŷ is qZ|X,Y . We observe that for any k ∈ {1, . . . , n}, an application of the triangle inequality yields 1X pZ̃k (z) − qZ (z) kpZ̃k − qZ kTV = 2 z∈Z = (28) ≤ pẐ − qZ TV ≤δ δ ≤ R1 + R2 2|Z| δ ≤ R1 I(X̂; Ẑ|V̂ ) + δ log 2|Z| δ I(Ŷ ; Ẑ|V̂ ) + δ log ≤ R2 . 2|Z| I(X̂, Ŷ ; Ẑ|V̂ ) + δ log (32) (33) Proof of Lemma 5. The codebooks induce random variables X̃ n , Ỹ n and Z̃ n with joint probability mass function X exp(−nR1 )1C1 (m)=xn X exp(−nR2 )1C2 (m)=yn . m=1 exp(nR2 ) n pY n |C2 (y ) = m=1 Note that for notational convenience, we will throughout this proof use the definition of pX̃ n ,Ỹ n ,Z̃ n and the definitions implied through the marginals of this probability mass function 1X pZ̃k − qZ n ≤ δ, (36) k=1 n X k=1 pZ̃k − qZ (37) TV TV (38) (39) where (37) is by definition of V̂ and Ẑ, (38) follows by the triangle inequality and (39) by applying (36). Next, we bound nI(X̂, Ŷ ; Ẑ|V̂ ) − I(X̃ n , Ỹ n ; Z̃ n ) ≤ −nδ log δ 2|Z| (40) as shown in (89) through (93) in the appendix. Taking into consideration that n(R1 + R2 ) ≥ H(X̃ n , Ỹ n ) ≥ I(X̃ n , Ỹ n ; Z̃ n ) and that (n + 1)/n ≤ 2, this proves (33). Furthermore, we note that nI(X̂; Ẑ|V̂ ) − I(X̃ n ; Z̃ n )  = nI(X̂, Ŷ ; Ẑ|V̂ ) − I(X̃ n , Ỹ n ; Z̃ n )  − nI(Ŷ ; Ẑ|X̂, V̂ ) − I(Ỹ n ; Z̃ n |X̃ n ) δ ≤ −nδ log , 2|Z| where pX n |C1 (x ) = TV = 1 ≤ n = pX n |C1 (xn )pY n |C2 (y n )qZ n |X n ,Y n (z n |xn , y n ), exp(nR1 ) z ∈Z n pẐ − qZ pX̃ n ,Ỹ n ,Z̃ n (xn , y n , z n ) = pX n ,Y n ,Z n |C1 ,C2 (xn , y n , z n ) n 1 X |p n (z n ) − qZ n (z n )| 2 n n Z̃ Now, we can bound (34) (35) (pZ̃ n (z n ) − qZ n (z n )) = kpZ̃ n − qZ n kTV ≤ δ. Lemma 5. Let W = (X , Y, Z, qZ|X,Y ) be a channel, qX , qY input distributions (with induced output distribution qZ ) and C1 , C2 a pair of codebooks of rates R1 , R2 and block length n such that 1 kpZ n |C1 ,C2 − qZ n kTV ≤ δ ≤ . (31) 4 Then there are random variables X̂ on X , Ŷ on Y, Ẑ on Z and V̂ on V := {1, . . . , n} with a joint probability mass function pV̂ pX̂|V̂ pŶ |V̂ qẐ|X̂,Ŷ such that qẐ|X̂,Ŷ = qZ|X,Y and X z∈Z z n ∈Z n zk =z (29) (30) We first prove the following lemma, which states conclusions we can draw from the existence of a codebook pair for a fixed block length n. Theorem 8 then follows by considering the limit n → ∞, as we detail at the end of this section. 1X 2 (41) where the inequality step follows by (40) and the observation that nI(Ŷ ; Ẑ|X̂, V̂ ) − I(Ỹ n ; Z̃ n |X̃ n ) is nonnegative, because I(Ỹ n ; Z̃ n |X̃ n ) n   X = H(Z̃k |X̃ n , Z̃ k−1 ) − H(Z̃k |X̃ n , Ỹ n , Z̃ k−1 ) (42) = k=1 n  X k=1  H(Z̃k |X̃ n , Z̃ k−1 ) − H(Z̃k |X̃k , Ỹk ) (43) 9 ≤ n  X k=1  H(Z̃k |X̃k ) − H(Z̃k |X̃k , Ỹk ) (44) = nI(Ŷ ; Ẑ|X̂, V̂ ), where (42) follows by the entropy chain rule, (43) follows from the fact that Z̃k is conditionally independent of the other random variables given X̃k and Ỹk and (44) follows because conditioning does not increase entropy. Analogous reasoning exchanging the roles of X and Y yields δ . (45) nI(Ŷ ; Ẑ|V̂ ) − I(X̃ n ; Z̃ n ) ≤ −nδ log 2|Z| In a similar way as we obtained (33) from (40), we can derive the bounds (34) and (35) from (41) and (45), respectively. In order to ensure convergence in the limit n → ∞, we need a slight refinement of Lemma 5. Corollary 2. Lemma 5 holds with V replaced by some (fixed) V ′ with |V ′ | ≤ |Z| + 3. Proof. We apply Lemma 2 with A := X × Y, B := V, letting P be the set of all probability distributions on X × Y corresponding to independent random variables X on X and Y on Y. Define rv (x, y) := pX̂|V̂ (x|v)pŶ |V̂ (y|v) = n X 1 XX rv (x, y)qZ|X,Y (z|x, y) n k=1 x∈X y∈Y = pẐ (z), implying that (32) is preserved and completing the proof of the corollary. We are now ready to put everything together and prove our main converse result. Proof of Theorem 8. We write δn := kpZ n |C1,n ,C2,n − qZ n kTV , and apply Corollary 2 with C1 := C1,n , C2 := C2,n and δ := δn to obtain random variables Xn , Yn , Zn and Vn satisfying conditions (32), (33), (34) and (35). We assume without loss of generality that Vn are all on the same alphabet V with cardinality |Z| + 3 and observe that the space of probability distributions on a finite alphabet is compact. So we can pick an increasing sequence (nk )k≥1 such that pVnk converges to a probability mass function pV , pXnk |Vnk =v converges for all v ∈ V to a probability mass function pX|V =v and pYnk |Vnk =v converges for all v ∈ V to a probability mass function pY |V =v . The random variables X, Y , Z and V are then defined by the joint probability mass function pV pX|V pY |V qZ|X,Y . (32) implies that the marginal distribution of Z is qZ , (33) implies (28), (34) implies (29) and (35) implies (30). and f1 (rv ) := I(X̂; Ẑ|V̂ = v) f2 (rv ) := I(Ŷ ; Ẑ|V̂ = v) f3 (rv ) := I(X̂, Ŷ ; Ẑ|V̂ = v) XX f4,z (rv ) := rv (x, y)qZ|X,Y (z|x, y), x∈X y∈Y where f4,z defines a set of |Z| functions. Note that the mutual information terms depend only on rv and are continuous. By Lemma 2, we have a random variable V ′ distributed according to r on some alphabet of cardinality at most |Z| + 3 and probability mass functions (rv′ ′ )v′ ∈V ′ corresponding to random variables independently distributed on X and Y. These probability mass functions induce random variables X ′ and Y ′ that depend on V ′ , but are conditionally independent given V ′ . We further define a random variable Z ′ on Z depending on X ′ and Y ′ through qZ|X,Y . We have, by (4), X I(X ′ , Y ′ ; Z ′ |V ′ ) = r(v ′ )I(X ′ , Y ′ ; Z ′ |V ′ = v ′ ) = v ′ ∈V ′ n X k=1 1 I(X̂, Ŷ ; Ẑ|V̂ = k) n = I(X̂, Ŷ ; Ẑ|V̂ ), and, by similar arguments, I(X ′ ; Z ′ |V ′ ) = I(X̂; Ẑ|V̂ ) and I(Y ′ ; Z ′ |V ′ ) = I(Ŷ ; Ẑ|V̂ ), so properties (33), (34) and (35) are preserved for the new random variables. Furthermore, also by (4), we have for each z ∈ Z, X XX pZ ′ (z) = r(v ′ ) rv′ ′ (x, y)qZ|X,Y (z|x, y) v ′ ∈V ′ x∈X y∈Y VI. C HANNEL R ESOLVABILITY AND S ECRECY In this section, we summarize some existing notions of secrecy and a few facts about their relations to each other, and also discuss their operational meaning. We then show explicitly how our channel resolvability results can be applied to derive a secrecy result for the multiple-access wiretap channel that not only implies strong secrecy, but has even stronger operational properties than the classical notion of strong secrecy. We start by providing additional definitions in Subsection VI-A, then we discuss different notions of secrecy and interconnections between them in Subsection VI-B. Then we show how our resolvability results can be used to achieve secrecy over the multiple-access channel. To this end, we first repeat a well-known result on achievable rates for the multipleaccess channel in Subsection VI-C which is the second crucial ingredient besides resolvability for the secrecy theorem we prove in Subsection VI-D. A. Additional Definitions In contrast to the other sections, in this section we will also look at probability distributions on the alphabet of possible messages. We therefore use M1 := {1, . . . , exp(nR1 )} and M2 := {1, . . . , exp(nR2 )} to denote the message alphabets; for convenience we denote by M := M1 × M2 the message alphabet of both transmitters. M1 , M2 and M = (M1 , M2 ) are random variables on M1 , M2 and M, respectively, to denote the messages chosen by the transmitters to be sent and pM denotes a probability mass function on M describing the distribution of M . 10 By a wiretap channel W = (Wlegit , Wtap ), we mean a pair of channels, where Wlegit = (X , Y, Zlegit , qZlegit |X,Y ), which we call the legitimate receiver’s channel and Wtap = (X , Y, Ztap , qZtap |X,Y ), which we call the wiretapper’s channel, share the same input alphabets. A wiretap code of block length n for the multiple-access channel consists of two collections of probability mass functions pX n |M1 for each m1 ∈ M1 and pY n |M2 for each m2 ∈ M2 (called the encoders), as well as a decoder n dn : Zlegit → M. The encoders and the wiretap channel together induce for each m ∈ M a wiretapper’s output n distribution on Ztap . In this paper, we will consider a special class of wiretap encoders: By a pair of wiretap codebooks of block length n, information rates R1 , R2 and randomness rates L1 , L2 , we mean functions C1 : {1, . . . , exp(nR1 )}×{1, . . . , exp(nL1 )} → X n and C2 : {1, . . . , exp(nR2 )} × {1, . . . , exp(nL2 )} → Y n . Given input distributions qX and qY , we define a probability distribution PC1 ,C2 on the set of all possible codebooks such that every component of every function value is independently distributed according to qX or qY , respectively. Such a pair of codebooks induces encoders in the sense defined above: For k ∈ {1, 2}, transmitter k picks (independently of the other transmitter and without cooperation) a message mk ∈ {1, . . . , exp(nRk )} and draws uniformly at random ℓk ∈ {1, . . . , exp(nLk )}. The transmitted codeword is then Ck (mk , ℓk ). The wiretapper’s output distribution induced by these encoders is n n |C (m ,·),C (m ,·) (z ) := exp(−n(L1 + L2 )) pZtap 1 1 2 2 exp(nL1 ) exp(nL2 ) X ℓ1 =1 X ℓ2 =1 n n |X n ,Y n (z |C1 (m1 , ℓ1 ), C2 (m2 , ℓ2 )). qZtap B. Secrecy Concepts and their Operational Meanings In this subsection, we give a short overview over some notions of information-theoretic secrecy. We adapt the notation in the sources we are citing in such a way that it conforms with the notation we have been using in this paper, which is specific to the multiple-access wiretap channel, but the concepts apply also to (and have originally been introduced for) wiretap channels with only one transmitter. Before we give formal definitions of secrecy, we first take a look at what operational implications we want a good notion of secrecy to have. In [1], Ahlswede generalized the communication task in the following way: The transmitter picks a message M ∈ M, encodes it and transmits the codeword over a channel. The receiver then aims not necessarily at decoding the exact message M , but rather has a partition {M1 , . . . , Mℓ } of M, i.e. disjoint subsets of M with ∪ℓk=1 Mk = M. Its task is then to reconstruct k such that M ∈ Mk . The transmitter does not know the exact partition Π that defines the receiver’s task, but has only a subset {Π1 , . . . , Πt } of the set of all possible partitions of M and has to encode the message in such a way that the receiver can decode with high probability no matter which partition in this set defines its task, or put differently, we assume there is one receiver for each partition and we want each receiver to be able to decode with high probability. Ahlswede gave a list of specializations of this very general concept that he deemed potentially useful for real-world communication systems. Two of these are: 1) t = 1 and the only partition Π1 is the partition of singletons {{m} : m ∈ M}. Here the receiver’s task amounts to Shannon’s classical decoding problem. 2) t = |M| and the set of partitions is n o {m}, M \ {m} : m ∈ M . This corresponds to the problem of identification codes introduced by Ahlswede and Dueck [2]. These decoding tasks can not only be understood as having to be performed by a legitimate receiver in cooperation with the transmitter, but can also be seen as possible attacks carried out by a wiretapper, where each possible attack is defined by a partition on M. For instance, we could deem the wiretapper’s attack to have been successful only if it is able to decode the exact message which corresponds to item 1), or, by an alternative definition, we might choose a particular message m that the wiretapper has to identify corresponding to one of the partitions of item 2) and thus obtain a much weaker definition of what constitutes a successful attack. The best thing we might wish for in this context would be that the wiretapper be unable to perform attacks corresponding to any possible partition of the message set. In fact, demanding this is equivalent to insisting that the wiretapper shall not be able to decode any function of the transmitted message (because a function induces a partition of its domain via the preimages of singletons in its range). Shannon introduced the concept of information theoretic secrecy and required for his condition of perfect secrecy [37] n that I(M ; Ztap ) = 0. In Shannon’s cipher system, both legitimate receiver and wiretapper get the same channel output, but transmitter and legitimate receiver share a key, which is unknown to the wiretapper. Later, Wyner introduced the wiretap channel, where the wiretapper’s channel output is degraded with respect to the legitimate receiver’s channel output, but wiretapper and legitimate receiver do not share a key. In order to be more amenable to analysis, he weakened Shannon’s perfect secrecy criterion and introduced the concept of weak secrecy [44]. A sequence of codes for increasing block length n is said n to achieve weak secrecy if I(M ; Ztap )/n approaches 0 as n tends to infinity, where M is assumed to be uniformly distributed over the message set. It turns out that a sequence of codes can achieve weak secrecy while the total amount of information leaked still tends to infinity with the block length approaching infinity. Therefore, the notion of strong secrecy was introduced by Maurer [35] which demands that n I(M ; Ztap ) shall approach 0 as n tends to infinity, where M is still assumed to be distributed uniformly. Strong secrecy has been shown to imply that full decoding is only possible with an error probability that approaches the probability of guessing correctly without any channel output at all and it has also been shown to provide some resilience against identification attacks [9], however, some messages 11 might remain identifiable and no method was proposed of ensuring that any particular given message m not be identifiable. This leads us to the concept of semantic security, which originates from the field of cryptography [24], but has also been redefined to yield an information-theoretic secrecy criterion [7][6]. Following the discussion in [6] and adapting to our notation, we say that a sequence of codebooks achieves semantic security if  n n (M ∈ f (Z max PM,Ztap max tap )) n →Π Π,PM f :Ztap  − max PM (M ∈ g) (46) g∈Π tends to 0 with n → ∞, where Π ranges over all partitions of M and PM over all possible probability distributions of M ∈ M. The positive term can be seen as the probability that the wiretapper solves its decoding task successfully given its channel output and an optimal decoder, while the second term can be seen as subtracting the probability that the decoding task is solved without knowledge of the wiretapper’s channel output (i.e. by optimal guessing). This notion has a wide range of operational implications in the sense that it is impossible for the wiretapper to carry out an attack corresponding to any of Ahlswede’s general communication tasks, as we show in Examples 1 and 2. Remark 4. (46) remains unchanged even if we allow stochastic wiretapper’s decoders, i.e. if we let f range over functions n to probability mass functions on Π. that map from Ztap On the other hand, it is not easy to prove directly that a particular wiretap code achieves semantic security. We therefore define a notion of secrecy that is more amenable to analysis, namely, we say that a sequence of wiretap codebooks achieves distinguishing security if max m1 ,m2 ,m̃1 ,m̃2 n |C (m ,·),C (m ,·) kpZtap 1 1 2 2 n |C (m̃ ,·),C (m̃ ,·) kTV − pZtap 1 1 2 2 (47) approaches 0 with n tending to infinity. Remark 5. Due to the triangle inequality, this notion of distinguishing security is equivalent to n |C (m ,·),C (m ,·) − pkTV max kpZtap 1 1 2 2 m1 ,m2 approaching 0, where p is an arbitrary probability mass n function on Ztap that does not change with m1 or m2 . In [6], these and several other related concepts of information-theoretical secrecy and their interrelations are discussed. It turns out that distinguishing and semantic security are equivalent and strictly stronger than strong secrecy. On the other hand, if the definition of strong secrecy is strengthened n ) approaches 0, where pM to demand that maxpM I(M ; Ztap ranges over all distributions of M (not just uniform), then it becomes equivalent to the notions of distinguishing and semantic security. We will present here versions of the proofs adapted to our notation only for the implications that seem most important to us in this context, namely that distinguishing security, which we show in Theorem 10 can be achieved over the multiple-access channel, implies both strong secrecy (Lemma 6) and semantic security (Lemma 7). We now give two examples to show how the concept of semantic security implies resilience to the Ahlswede attacks discussed before. Example 1 is a special case of Example 2, but we lay it out explicitly anyway because we believe that it illustrates the ideas involved in a particularly accessible way and that the type of attack it describes could be particularly relevant in practice. Example 1 (Identification attacks). Suppose the wiretapper’s n objective is to determine from an observation of Ztap whether the transmitted message was m or not, i.e. the attack we consider can be described by the partition Π = {{m}, M\{m}}. To this end, we assume that it has found some decoding n function f : Ztap → Π. Without assuming any probability distribution on M, we can define two types of error  n n E1 := PZtap f (Ztap ) = M \ {m} M = m  n n E2m̂ := PZtap f (Ztap ) = {m} M = m̂ , where E2m̂ is defined for each m̂ 6= m. Note that if the wiretapper is allowed to guess stochastically, it can trivially achieve any values for E1 and E2m̂ that satisfy E1 + E2m̂ = 1 (where E2m̂ can be chosen the same for all m̂), even without n observing Ztap . On the other hand, we now show that if (46) is bounded above by δ > 0 and we fix an arbitrary m̂ 6= m, the wiretapper cannot achieve much lower error probabilities n E1 and E2m̂ , even given the observation of Ztap . We first note that if PM is defined by the probability mass function ( 1 , m̃ = m ∨ m̃ = m̂ pM (m̃) := 2 0, otherwise, then PM (M ∈ g) = 1/2 regardless of the choice of g. We can thus bound n n (M ∈ f (Z δ > PM,Ztap tap )) − max PM (M ∈ g) g∈Π (48) 1 n n (m ∈ f (Z = PZtap tap )|M = m) 2 (49) 1 1 n n (m̂ ∈ f (Z + PZtap tap )|M = m̂) − 2 2 1 m̂ = (1 − E1 + 1 − E2 − 1), (50) 2 where (48) is by assumption valid for all PM , (49) follows by plugging in PM as defined above and (50) follows by plugging in the definitions of E1 and E2m̂ . We can express this equivalently as E1 + E2m̂ > 1 − 2δ, which shows that as (46) tends to 0, the wiretapper’s minimum error probabilities approach values that can be achieved by stochastic guessing without observing any channel output at all. Example 2 (General Ahlswede Attacks). As a second, slightly more involved example, we show the operational meaning of semantic security for attacks defined by more general 12 partitions. Let Π = {M1 , . . . , Mℓ } be a partition of M. Instead of assuming that the wiretapper can determine the correct partition element based on any transmitted message, we make a weaker assumption, namely that there is (at least) one message in each partition element that can be correctly identified. Formally, we say that the attacker has n a decoding function f : Ztap → Π and we fix messages m1 ∈ M1 , . . . , mℓ ∈ Mℓ . For all k ∈ {1, . . . , ℓ}, the error probabilities associated with partition element Mk and message mk ∈ Mk is then defined as  n n Ekmk := PZtap f (Ztap ) 6= Mk M = mk . We observe that if stochastic guessing is allowed, the wiretapper can achieve any combination of error probabilities that satisfies E1m1 + · · · + Eℓmℓ = ℓ − 1. We consider PM defined by the probability mass function ( 1 , m = mk for some k pM (m) := ℓ 0, otherwise, noting that PM (M ∈ g) = 1/ℓ regardless of the choice of g. Assuming that (46) is upper bounded by δ > 0, we have n n (M ∈ f (Z δ > PM,Ztap tap )) − max PM (M ∈ g) (51) ℓ X 1 1 n n (f (Z PZtap = tap ) = Mk |M = mk ) − ℓ ℓ (52) g∈Π = k=1 ℓ X k=1 1 1 (1 − Ekmk ) − ℓ ℓ k=1 (53) which shows that also in this more general case, as (46) tends to 0, the wiretapper’s error probabilities approach the values that can be achieved by stochastic guessing without observing any channel output at all, even if these error probabilities are only assumed to hold for one message in every partition element. Lemma 6. If distinguishing security holds, i.e. (47) apn ). In particular, proaches 0, then so does maxpM I(M ; Ztap strong secrecy holds. Proof. Fix an arbitrary distribution pM on M and observe that (54) (55) where m̃ is chosen as the element in M that minimizes n H(Ztap |M = m̃) and the last inequality follows from Lemma 1, as long as δ < 1/4, where δ= X (m1 ,m2 )∈M n |C (m ,·),C (m ,·) pM (m1 , m2 )pZtap 1 1 2 2 TV pM (m1 , m2 ) (m1 ,m2 )∈M n |C (m ,·),C (m ,·) − pZ n |C (m̃ ,·),C (m̃ ,·) pZtap 1 1 2 2 1 2 2 tap 1 TV . Note that the latter variational distance term is clearly bounded by (47), and so (55) approaches 0 if (47) does. Lemma 7. (46) is upper bounded by (47). In particular, distinguishing security implies semantic security. Proof. Fix a partition Π of M and a probability mass function pM on M that realize the outer maximum in (46) as well n as a decoding function f : Ztap → Π and a guess g ∈ Π that realize the inner maxima. We can then rewrite (46) as P1 + P2 − P3 by steps as shown in (94) through (97) in the appendix. Observe that P2 = PM (M ∈ / g) and that, taking into account that g minimizes PM (M ∈ / g) over the elements of Π, X X pM (m̃1 , m̃2 ) P3 = ≥ Ekmk > ℓ − 1 − ℓδ, n n n I(M ; Ztap ) ≤ H(Ztap ) − H(Ztap |M = m̃) 1 δ ≤ − δ log n, 2 2 |Ztap | ≤ X (m̃1 ,m̃2 )∈M where (51) is by assumption valid for all PM , (52) follows by plugging in PM and (53) follows by plugging in the definitions of the Ekmk . We can express this equivalently as ℓ X n |C (m̃ ,·),C (m̃ ,·) − pZtap 1 1 2 2 n z n ∈Ztap n n n |C (m̃ ,·),C (m̃ ,·) (z )PM (M ∈ )) pZtap / f (Ztap 1 1 2 2 X X pM (m̃1 , m̃2 ) (m̃1 ,m̃2 )∈M n z n ∈Ztap n n |C (m̃ ,·),C (m̃ ,·) (z )PM (M ∈ pZtap / g)) 1 1 2 2 = PM (M ∈ / g). Observe further that, optimizing the outcome of the indicator function in P1 for each term in the innermost sum individually, we get X pM (m1 , m2 )pM (m̃1 , m̃2 ) P1 ≤ (m1 ,m2 ),(m̃1 ,m̃2 )∈M n |C (m ,·),C (m ,·) − pZ n |C (m̃ ,·),C (m̃ ,·) · pZtap 1 1 2 2 1 2 2 tap 1 TV , which is clearly upper bounded by (47). Summarizing, we have seen that P2 − P3 ≤ 0 and P1 is upper bounded by (47) and consequently (46) is upper bounded by (47). C. Coding for the Multiple-Access Channel The capacity region of the multiple-access channel is well known and proofs can be found in almost any textbook on information theory, however, usually only the existence of a suitable codebook is proven. We need here the slightly stronger statement that the probability of randomly choosing a bad codebook is exponentially small, which requires only a slight modification of these proofs, which is, however, normally not given in the textbooks explicitly. For the sake of completeness, we state here the result that we need and present a proof that is similar to the one given in [26]. Theorem 9. Suppose W = (X , Y, Z, qZ|X,Y ) is a channel, qX and qY are input distributions, R1 < I(X; Z|Y ), 13 R2 < I(Y ; Z|X) and R1 + R2 < I(X, Y ; Z). Then there exist decoding functions dn : Z n → {1, . . . , exp(nR1 )} × {1, . . . , exp(nR2 )} and γ1 , γ2 > 0 such that for large enough block length n, the codebook distributions of block length n and rates R1 and R2 satisfy PC1 ,C2 (E > exp(−γ1 n)) ≤ exp (−γ2 n) , (56) where we define the average decoding error probability as E := nR2 exp nR1 exp X X m1 =1 exp(−n(R1 + R2 )) m2 =1  PZ n dn (Z n ) 6= (m1 , m2 )  X n = C1 (m1 ), Y n = C2 (m2 ) . Proof. We define typical sets n T1,ε := {(xn , y n , z n ) : i(xn ; z n |y n ) ≥ n(I(X; Z|Y ) − ε)} n T2,ε := {(xn , y n , z n ) : i(y n ; z n |xn ) ≥ n(I(Y ; Z|X) − ε)} n T3,ε := {(xn , y n , z n ) : i(xn , y n ; z n ) ≥ n(I(X, Y ; Z) − ε)} n n n Tεn := T1,ε ∩ T2,ε ∩ T3,ε and a corresponding joint typicality decoder  n n  (m1 , m2 ), (C1 (m1 ), C2 (m2 ), z ) ∈ Tε dn (z n ) := for unique (m1 , m2 )   (1, 1), otherwise. So in order to correctly decode a received message, we need the received sequence to be jointly typical with the encoded messages, while it must not be jointly typical with any other pair of codewords. We formalize these notions by defining the probabilities of three error events corresponding to the former being false as in (57) for k ∈ {1, 2, 3} and the probabilities of error events corresponding to the latter being false as in (58), (59) and (60). We observe by the union bound and the definition of dn that E≤ nR2 exp nR1 exp X X m1 =1 (62) ! (63) > exp(n(α − 1)(I(X; Z|Y ) − ε))  ≤EX n ,Y n ,Z n qZ n |X n ,Y n (Z n |X n , Y n ) qZ n |Y n (Z n |Y n ) α−1 ! where (63) follows for all α < 1 by substituting the definition n of T1,ε and reformulating, (64) follows by Markov’s inequality, (65) follows by reformulating and substituting the definition of Rényi divergence, and (66) follows for sufficiently small but positive β by taking α < 1 close enough to 1 such that the exponent in (65) is negative. This is possible by observing that the Rényi divergence approaches the Kullback-Leibler divergence for α → 1 and is finite and thus continuous since our alphabets are finite [40]. Moreover, we have 1,1 EC1 ,C2 Eatyp,1 (67) n n n n T1,ε  ≤ exp(nR1 )PY n ,Z n PX n (X , Y , Z ) ∈ (68) X X X n n n = exp(nR1 ) qX n (x )qY n ,Z n (y , z ) xn ∈X n y n ∈Y n z n ∈Z n n · 1(xn ,yn ,zn )∈T1,ε = exp(nR1 ) X X X (69) n n  m1 ,m2 m1 ,m2 m1 ,m2 + + Etyp,3 + Etyp,2 Etyp,1  m1 ,m2 m1 ,m2 m1 ,m2 Eatyp,1 + Eatyp,2 + Eatyp,3 1,1 1,1 1,1 +EC1 ,C2 Etyp,2 +EC1 ,C2 Etyp,2 EC1 ,C2 E ≤ EC1 ,C2 Etyp,1 1,1 1,1 1,1 +EC1 ,C2 Eatyp,3 +EC1 ,C2 Eatyp,2 +EC1 ,C2 Eatyp,1 We next bound each of these summands above by exp(−nβ) for sufficiently large n and sufficiently small but positive β. To this end, we observe (61) n qX n ,Y n ,Z n (x , y , z ) xn ∈X n y n ∈Y n z n ∈Z n n n n n · exp(−i(x ; z |y ))1(xn ,yn ,zn )∈T1,ε ≤ exp(−nβ), m2 =1 (64) · exp(−n(α − 1)(I(X; Z|Y ) − ε))  = exp n(α − 1)   · Dα PX,Y,Z ||PX|Y PZ|Y PY − I(X; Z|Y ) + ε (65) ≤ exp(−nβ) (66) ≤ exp(n(R1 − I(X; Z|Y ) + ε)), exp(−n(R1 + R2 )) and thus, taking into cosideration that since the codewords are m1 ,m2 m1 ,m2 have expectations and Eatyp,k identically distributed, Etyp,k that do not depend on m1 or m2 , we get 1,1 EC1 ,C2 Etyp,1   n / T1,ε = PX n ,Y n ,Z n (X n , Y n , Z n ) ∈   qZ n |X n ,Y n (Z n |X n , Y n ) α−1 = PX n ,Y n ,Z n qZ n |Y n (Z n |Y n ) (70) (71) (72) where (68) is an application of the union bound, (70) follows by reformulating and substituting the definition of information n density, (71) follows by the definition of T1,ε and (72) follows for all sufficiently small β > 0 as long as ε is so small that R1 − I(X; Z|Y ) + ε < 0. We skip the calculations showing that 1,1 1,1 ≤ exp(−nβ) , EC1 ,C2 Eatyp,k EC1 ,C2 Etyp,k for k ∈ {2, 3}, sufficiently large n and sufficiently small β > 0, since these statements follow in a straightforward manner by calculations very similar to the ones shown above. Putting everything together and applying Markov’s inequality, we get that for sufficiently large n, sufficiently small β > 0, γ1 < β, and γ2 < β − γ1 , PC1 ,C2 (E > exp(−γ1 n)) ≤ EC1 ,C2 (E) exp(γ1 n) 14   m1 ,m2 n / Tk,ε X n = C1 (m1 ), Y n = C2 (m2 ) := PZ n (X n , Y n , Z n ) ∈ Etyp,k   m1 ,m2 n X n = C1 (m1 ), Y n = C2 (m2 ) Eatyp,1 := PZ n ∃m̃1 6= m1 : (C1 (m̃1 ), Y n , Z n ) ∈ T1,ε   m1 ,m2 n Eatyp,2 := PZ n ∃m̃2 6= m2 : (X n , C1 (m̃2 ), Z n ) ∈ T2,ε X n = C1 (m1 ), Y n = C2 (m2 )   m1 ,m2 n X n = C1 (m1 ), Y n = C2 (m2 ) Eatyp,3 := PZ n ∃m̃1 6= m1 , m̃2 6= m2 : (C1 (m̃1 ), C1 (m̃2 ), Z n ) ∈ T3,ε ≤ 6 exp(−n(β − γ1 )) ≤ exp(−γ2 n), I(X; Ztap |V ) < I(X; Zlegit|Y, V ) − R1 I(Y ; Ztap |V ) < I(Y ; Zlegit |X, V ) − R2 I(X, Y ; Ztap |V ) < I(X, Y ; Zlegit |V ) − (R1 + R2 ). (73) (74) (75) Then there exist randomness rates L1 , L2 , decoding functions n dn : Zlegit →{1, . . . , exp(nR1 )} × {1, . . . , exp(nL1 )}× {1, . . . , exp(nR2 )} × {1, . . . , exp(nL2 )} and γ1 , γ2 > 0, such that for sufficiently large n, the probability Pbad under the wiretap codebook distribution of drawing a bad pair of codebooks is less than exp(−nγ2 ). A pair of codebooks is considered to be bad if it does not satisfy all of the following: 1) (Distinguishing Security). For any m1 , m2 , we have n |C (m ,·),C (m ,·) − qZ n kTV ≤ exp(−γ1 n), kpZtap 1 1 2 2 i.e. the wiretap codebooks achieve distinguishing security and the variational distance vanishes exponentially. 2) (Average decoding error). We have E ≤ exp(−nγ1 ), where nL1 exp nR2 exp nL2 exp nR1 exp X X X X ℓ1 =1 (60) L1,2 = −I(Y ; Zlegit |X, V ) + R2 + I(X, Y ; Ztap |V ) Theorem 10. Suppose W = (Wlegit , Wtap ) is a wiretap channel, V is a random variable, qX|V and qY |V are input distributions and R1 and R2 are rates satisfying m2 =1 (59) L1,1 = I(X; Ztap |V ) D. Achievable Secrecy Rates under Distinguishing Security m1 =1 (58) where concluding the proof of the theorem. E := (57) ℓ2 =1 exp(−n(R1 + R2 + L1 + L2 ))  PZ n dn (Z n ) 6= (m1 , ℓ1 , m2 , ℓ2 )  X n = C1 (m1 , ℓ1 ), Y n = C2 (m2 , ℓ2 ) . is the average decoding error; i.e. the legitimate receiver is able to (on average) reconstruct both the messages and the randomness used at the transmitters with exponentially vanishing error probability. Proof. Choose L1 such that max(L1,1 , L1,2 ) < L1 < min(L1,1 , L1,2 ), (76) L1,1 = I(X; Zlegit |Y, V ) − R1 L1,2 = −I(Y ; Ztap |V ) + I(X, Y ; Zlegit |V ) − (R1 + R2 ). To see that this is possible, we note that L1,1 < L1,1 is equivalent to (73), L1,1 < L1,2 and L1,2 < L1,1 both follow from (75), and L1,2 < L1,2 follows from (74) and (75). Then, choose L2 such that max(L2,1 , L2,2 ) < L2 < min(L2,1 , L2,2 ), (77) where L2,1 = I(X, Y ; Ztap |V ) − L1 L2,2 = I(Y ; Ztap |V ) L2,1 = I(Y ; Zlegit |X, V ) − R2 L2,2 = I(X, Y ; Zlegit |V ) − L1 − (R1 + R2 ). To see that this is possible, we note that L2,1 < L2,1 is equivalent to L1 > L1,2 , which we have by (76), L2,1 < L2,2 is equivalent to (75), L2,2 < L2,1 is equivalent to (74) and L2,2 < L2,2 is equivalent to L1 < L1,2 , which we also have by (76). Summarizing, with these choices, L1 , L2 have the properties I(X; Ztap |V ) < L1 < I(X; Zlegit|Y, V ) − R1 I(Y ; Ztap |V ) < L2 < I(Y ; Zlegit |X, V ) − R2 . I(X, Y ; Ztap |V ) < L1 + L2 < I(X, Y ; Zlegit |V ) − R2 . We can look at the thus defined pair of random wiretap codebooks in two ways. On the one hand, they can be understood as a pair of codebooks of rates R1 + L1 and R2 +L2 . They fulfill all the requirements of Theorem 9 (noting that by time sharing, the theorem is also valid for the convex closure specified in the theorem statement), and so we get γ1′ , γ2′ > 0 such that PC1 ,C2 (E > exp(nγ1′ )) ≤ exp(nγ2′ ). On the other hand, we can look at them as a colexp(nR ) lection (C1 (m1 , ·))m1 =1 1 of codebooks of rate L1 and exp(nR ) (C2 (m2 , ·))m2 =1 2 of codebooks of rate L2 , respectively. Therefore, each pair (m1 , m2 ) corresponds to a pair of codebooks fulfilling the requirements of Corollary 2 (again noting the possibility of time sharing), and so we get γ1′′ , γ2′′ > 0 such that for all m1 , m2 15  PC1 ,C2 kpZ n |C1 (m1 ,·),C2 (m2 ,·) − qZ n kTV > exp(−γ1′′ n) ≤ exp (− exp (γ2′′ n)) . Choosing γ1 := min(γ1′ , γ1′′ ), we can apply the union bound to get Pbad ≤ PC1 ,C2 (E > exp(nγ1′ ))+ exp(nR1 ) exp(nR2 ) X m1 =1 X m2 =1  PC1 ,C2 kpZ n |C1 (m1 ,·),C2 (m2 ,·) − qZ n kTV  > exp(−γ1′′ n) ≤ exp(γ2′ n) + exp(n(R1 + R2 ) − exp(γ2′′ n)) ≤ exp(γ2 n) for sufficiently large n, as long as we choose γ2 < γ2′ . R EFERENCES [1] R. Ahlswede. General theory of information transfer: Updated. Discrete Applied Mathematics, 156(9):1348–1388, 2008. [2] R. Ahlswede and G. Dueck. Identification via channels. Trans. Inf. Theory, 35(1):15–29, 1989. [3] R. Ahlswede and A. Winter. Strong converse for identification via quantum channels. Trans. Inf. Theory, 48(3):569–579, 2002. [4] Rudolf Ahlswede and Janos Korner. Source coding with side information and a converse for degraded broadcast channels. IEEE Transactions on Information Theory, 21(6):629–637, 1975. [5] P. Beek. An application of Fourier methods to the problem of sharpening the Berry-Esseen inequality. Prob. Theo. Rel. Fields, 23(3):187–196, 1972. [6] M. Bellare, S. Tessaro, and A. Vardy. A cryptographic treatment of the wiretap channel. arXiv:1201.2205, 2012. [7] M. Bellare, S. Tessaro, and A. Vardy. Semantic security for the wiretap channel. In Advances in Cryptology–CRYPTO 2012, pages 294–311. Springer, 2012. [8] A.C. Berry. The accuracy of the Gaussian approximation to the sum of independent variates. Trans. Am. Math. Soc., 49(1):122–136, 1941. [9] I. Bjelaković, H. Boche, and J. Sommerfeld. Secrecy results for compound wiretap channels. Problems Inform. Transmission, 49(1):73– 98, 2013. [10] M. R. Bloch and J. N. Laneman. Strong secrecy from channel resolvability. Trans. Inf. Theory, 59(12):8077–8098, 2013. [11] R. A. Chou, M. R. Bloch, and J. Kliewer. Low-complexity channel resolvability codes for the symmetric multiple-access channel. In Information Theory Workshop, pages 466–470. IEEE, Nov 2014. [12] I. Csiszár. Almost independence and secrecy capacity. Problems Inform. Transmission, 32(1):40–47, 1996. [13] I. Csiszár and J. Körner. Information theory: coding theorems for discrete memoryless systems. Cambridge University Press, 2011. [14] P. Cuff. Soft covering with high probability. In ISIT, pages 2963–2967. IEEE, 2016. [15] Paul Cuff. A stronger soft-covering lemma and applications. In Communications and Network Security (CNS), 2015 IEEE Conference on, pages 40–43. IEEE, 2015. [16] I. Devetak. The private classical capacity and quantum capacity of a quantum channel. Trans. Inf. Theory, 51(1):44–55, 2005. [17] D. P. Dubhashi and A. Panconesi. Concentration of measure for the analysis of randomized algorithms. Cambridge University Press, 2009. [18] E. Ekrem and S. Ulukus. On the secrecy of multiple access wiretap channel. In 2008 46th Annual Allerton Conference on Communication, Control, and Computing, pages 1014–1021, Sept 2008. [19] Abbas El Gamal and Young-Han Kim. Network information theory. Cambridge university press, 2011. [20] C.-G. Esseen. On the Liapounoff limit of error in the theory of probability. Almqvist & Wiksell, 1942. [21] M. Frey, I. Bjelaković, and S. Stańczak. Mac resolvability: First and second order results. In 2017 IEEE Conference on Communications and Network Security (CNS): The Workshop on Physical-Layer Methods for Wireless Security, pages 569–573, 2017. [22] Z. Goldfeld, P. Cuff, and H. H. Permuter. Semantic-security capacity for the physical layer via information theory. In 2016 IEEE International Conference on Software Science, Technology and Engineering (SWSTE), pages 17–27, June 2016. [23] Z. Goldfeld, P. Cuff, and H. H. Permuter. Semantic-security capacity for wiretap channels of type ii. IEEE Transactions on Information Theory, 62(7):3863–3879, July 2016. [24] S. Goldwasser and S. Micali. Probabilistic encryption. Journal of computer and system sciences, 28(2):270–299, 1984. [25] T. S. Han and S. Verdú. Approximation theory of output statistics. Trans. Inf. Theory, 39(3):752–772, 1993. [26] Te Sun Han. Information-spectrum methods in information theory, volume 50. Springer Science & Business Media, 2013. [27] M. Hayashi. General nonasymptotic and asymptotic formulas in channel resolvability and identification capacity and their application to the wiretap channel. Trans. Inf. Theory, 52(4):1562–1575, 2006. [28] W. Hoeffding. Probability inequalities for sums of bounded random variables. J. Am. Stat. Ass., 58(301):13–30, 1963. [29] J. Hou and G. Kramer. Effective secrecy: Reliability, confusion and stealth. In ISIT, pages 601–605. IEEE, 2014. [30] Jie Hou. Coding for relay networks and effective secrecy for wire-tap channels. PhD thesis, Universitätsbibliothek der TU München, 2014. [31] S. Janson. Large deviations for sums of partly dependent random variables. Random Structures & Algorithms, 24(3):234–248, 2004. [32] Y. Liang and H. V. Poor. Multiple-access channels with confidential messages. IEEE Transactions on Information Theory, 54(3):976–1002, March 2008. [33] C. Ling, L. Luzzi, J.-C. Belfiore, and D. Stehlé. Semantically secure lattice codes for the gaussian wiretap channel. Trans. Inf. Theory, 60(10):6399–6416, 2014. [34] N. Liu and W. Kang. The secrecy capacity region of a special class of multiple access channels. In 2011 IEEE International Symposium on Information Theory Proceedings, pages 623–627, July 2011. [35] Ueli M. Maurer. The Strong Secret Key Rate of Discrete Random Triples, pages 271–285. Springer US, Boston, MA, 1994. [36] Y. Oohama. Converse coding theorems for identification via multiple access channels. In ISIT, pages 1650–1654, July 2013. [37] C. E. Shannon. Communication theory of secrecy systems. Bell Labs Technical Journal, 28(4):656–715, 1949. [38] Y. Steinberg. Resolvability theory for the multiple-access channel. Trans. Inf. Theory, 44(2):472–487, Mar 1998. [39] A. Thangaraj. Coding for wiretap channels: Channel resolvability and semantic security. In 2014 IEEE Information Theory Workshop (ITW 2014), pages 232–236, Nov 2014. [40] T. van Erven and P. Harremoës. Rényi divergence and Kullback-Leibler divergence. Trans. Inf. Theory, 60(7):3797–3820, July 2014. [41] S. Watanabe and M. Hayashi. Strong converse and second-order asymptotics of channel resolvability. In ISIT, pages 1882–1886. IEEE, 2014. [42] M. Wiese and H. Boche. An achievable region for the wiretap multipleaccess channel with common message. In ISIT, pages 249–253. IEEE, 2012. [43] A. Wyner. The common information of two dependent random variables. Trans. Inf. Theory, 21(2):163–179, 1975. [44] A. D. Wyner. The wire-tap channel. Bell Labs Technical Journal, 54(8):1355–1387, 1975. [45] M. H. Yassaee and M. R. Aref. Multiple access wiretap channels with strong secrecy. In ITW, pages 1–5. IEEE, Aug 2010. 16 A PPENDIX P̃ = X Cˆ2  PC2 (C2 = Cˆ2 )PC1 ,C2   exp(nR2 ) X exp(−nR2 ) m2 =1 qZ n |Y n (z n |C2 (m2 )) n 1(C2 (m2 ),zn )∈T2,ε qZ n (z n )  (78) Ptyp,1 (C2 (m2 ), z n ) > 1 + 3 exp(−nβ) | C1 ∈ Czn , C2 = Cˆ2  exp(nR2 ) qZ n |Y n (z n |C2 (m2 )) n 1(C2 (m2 ),zn )∈T2,ε qZ n (z n ) m2 =1 Cˆ2  1 + 3 exp(−nβ) > | C1 ∈ Czn , C2 = Cˆ2  1 + exp(−nβ)   exp(nR2 ) X qZ n |Y n (z n |C2 (m2 )) n > 1 + exp(−nβ) 1(C2 (m2 ),zn )∈T2,ε ≤ PC2  exp(−nR2 ) qZ n (z n ) m2 =1   1 ≤ exp − exp(−n(I(Y ; Z) + ε + 2β − R2 )) 3 ≤ X PC2 (C2 = Cˆ2 )PC1 ,C2  X exp(−nR2 ) (79) (80) (81) (78) follows from the law of total probability, (79) is a consequence of the condition C1 ∈ Czn , (80) results from an application of the law of total probability and the assumption that n is sufficiently large such that exp(−nβ) ≤ 1. Finally, (81) follows from Lemma 3 with A = Y , C = Z, B a deterministic random variable with only one possible realization and δ = exp(−nβ).  PC1 ,C2 kpZ n |C1 ,C2 − qZ n kTV > 7 exp(−nβ) ≤PC1 ,C2 (Patyp,1 > 2 exp(−nβ)) + PC1 ,C2 (Patyp,2 > 2 exp(−nβ)) X + / Czn ) + PC1 ,C2 (Ptyp (z n ) > 1 + 3 exp(−nβ) | C1 ∈ Czn )) (PC1 (C1 ∈ (82) (83) z n ∈Z n   1 ≤2 exp(−2 exp(n(min(R1 , R2 ) − 2β))) + |Z|n |Y|n exp − exp(−n(I(X; Z|Y ) + ε + 2β − R1 )) 3   1 + |Z|n exp − exp(−n(I(Y ; Z) + ε + 2β − R2 )) 3 (84) (83) follows from (9) and the union bound. (84) is a substitution of (17), (18), (19) and (81).     3 1 +√ PC1 ,C2 kpZ n |C1 ,C2 − qZ n kTV > (µ2 + µ1 ) 1 + √ n n       1 1 + PC1 ,C2 Patyp,2 > µ2 1 + √ ≤PC1 ,C2 Patyp,1 > µ1 1 + √ n n    X 3 / Czn ) + PC1 ,C2 Ptyp,1 (C2 (m2 ), z n ) > 1 + √ | C1 ∈ Czn + PC1 (C1 ∈ n z n ∈Z n     2µ2 2µ2 ≤ exp − 1 exp(n min(R1 , R2 )) + exp − 2 exp(n min(R1 , R2 )) n n     1 1 n n n + |Y| |Z| exp − exp(−n(I(X; Z|Y ) + ε1 − R1 )) + |Z| exp − exp(−n(I(Y ; Z) + ε2 − R2 )) 3n 3n     1 c−d−1 2 min(µ21 , µ22 ) exp(n min(R1 , R2 )) + 2 exp n(log|Z| + log|Y|) − n ≤ 2 exp − n 3 (85) (86) (87) (88) (86) results from (9) and the union bound, (87) follows by substituting (25), (26) and (27). (88) follows by substituting (22), (23) and (24), as well as elementary operations. 17 nI(X̂, Ŷ ; Ẑ|V̂ ) − I(X̃ n , Ỹ n ; Z̃ n ) = n X X X X k=1 x∈X y∈Y z∈Z − = pX̃k ,Ỹk ,Z̃k (x, y, z) log X X X n X H(Z̃k ) X pX̃ n ,Ỹ n ,Z̃ n (xn , y n , z n ) log qZ|X,Y (zk |xk , yk ) − H(Z̃ n ) (90) H(Z̃k ) − H(Z̃ n ) k=1 n X (89) k=1 k=1 xn ∈X n y n ∈Y n z n ∈Z n n X qZ n |X n ,Y n (z n |xn , y n ) pZ̃ n (z n ) pX̃k ,Ỹk ,Z̃k (x, y, z) log qZ|X,Y (z|x, y) + k=1 x∈X y∈Y z∈Z n X X X = pX̃ n ,Ỹ n ,Z̃ n (xn , y n , z n ) log xn ∈X n y n ∈Y n z n ∈Z n n X X X X − qZ|X,Y (z|x, y) pZ̃k (z) (91)     1 1 δ δ ≤ H(Zk ) − H(Z ) + n − δ log + − δ log 2 2|Z| 2 2|Z n | k=1   δ ≤ n −δ log 2|Z| n (92) (93) (92) follows by (31), (36) and Lemma 1. max X Π,PM =  max P X n →Π f :Ztap (m1 ,m2 )∈M − n M,Ztap n z n ∈Ztap X X X  − max PM (M ∈ g) X n z n ∈Ztap n n |C (m ,·),C (m ,·) (z )1(m ,m )∈f (z n ) pZtap 1 1 2 2 1 2 n z n ∈Ztap (m1 ,m2 ),(m̃1 ,m̃2 )∈M − (95) n n |C (m̃ ,·),C (m̃ ,·) (z ) · 1(m̃ ,m̃ )∈g pM (m̃1 , m̃2 )pZtap 1 1 2 2 1 2 pM (m1 , m2 )pM (m̃1 , m̃2 ) X (94) g∈Π n n |C (m ,·),C (m ,·) (z ) · 1(m ,m )∈f (z n ) pM (m1 , m2 )pZtap 1 1 2 2 1 2 n (m̃1 ,m̃2 )∈M z n ∈Ztap = (M ∈ n f (Ztap )) n ! (96) n |C (m̃ ,·),C (m̃ ,·) (z ) · (1(m ,m )∈f (z n ) + 1(m ,m )∈f n − 1(m̃ ,m̃ )∈g pZtap ) 1 1 2 2 1 2 1 2 / (z ) 1 2 / = P1 + P2 − P3 (97) with the definitions X P1 := pM (m1 , m2 )pM (m̃1 , m̃2 ) n z n ∈Ztap (m1 ,m2 ),(m̃1 ,m̃2 )∈M P2 := X (m̃1 ,m̃2 )∈M P3 := X pM (m̃1 , m̃2 ) · (m1 ,m2 ),(m̃1 ,m̃2 )∈M X  X  n n n |C (m ,·),C (m ,·) (z ) − pZ n |C (m̃ ,·),C (m̃ ,·) (z ) pZtap 1 1 2 2 1 2 2 tap 1 n n |C (m̃ ,·),C (m̃ ,·) (z )1(m̃ ,m̃ )∈g pZtap 1 1 2 2 1 2 / (98) ·1(m1 ,m2 )∈f (zn ) (99) n z n ∈Ztap pM (m1 , m2 )pM (m̃1 , m̃2 ) X n z n ∈Ztap n n |C (m̃ ,·),C (m̃ ,·) (z )1(m ,m )∈f n . pZtap 1 1 2 2 1 2 / (z ) (100)
7cs.IT
Admissible Time Series Motif Discovery with Missing Data Yan Zhu, ‡Abdullah Mueen, Eamonn Keogh University of California, Riverside, ‡University of New Mexico [email protected], [email protected], ‡[email protected] Abstract The discovery of time series motifs has emerged as one of the most useful primitives in time series data mining. Researchers have shown its utility for exploratory data mining, summarization, visualization, segmentation, classification, clustering, and rule discovery. Although there has been more than a decade of extensive research, there is still no technique to allow the discovery of time series motifs in the presence of missing data, despite the welldocumented ubiquity of missing data in scientific, industrial, and medical datasets. In this work, we introduce a technique for motif discovery in the presence of missing data. We formally prove that our method is admissible, producing no false negatives. We also show that our method can “piggy-back” off the fastest known motif discovery method with a small constant factor time/space overhead. We will demonstrate our approach on diverse datasets with varying amounts of missing data. 1 INTRODUCTION Time series motifs are short approximately repeated patterns within a longer time series dataset. The fact that such patterns are conserved often suggests underlying structure and regularities that can be exploited in many ways. Some examples include: rule discovery [16][13], forecasting [10], or building better classifiers [2]. However, despite over a decade of active research, there is no known method to allow the discovery of the motif in the presence of missing data. Paradoxically, in spite of improvements in sensor technology, missing data is becoming more prevalent. This is because sensors are now so cheap that we are willing to deploy them in hostile environments with intermittent and unreliable bandwidth [17]. In many cases, sensors have become “throwable” and disposable [7]. Figure 1 shows an example of a motif in a music processing domain [12]. Note that both occurrences of the motif contain sections of missing data. At 34 seconds At 75 seconds 0 100 200 Figure 1: A four-second long motif that appears in the pitch contour time series of a Cypriot folk song, Kotsini Trantafillia (Red Rose-tree). Note that both occurrences have multiple instances of missing data [12]. Here the data could be missing for one of two reasons. It might be meaningful, i.e., the vocalist may not have produced a sound at these times. It is also possible that the missing data reflects poor audio quality, loose wires, etc. We are agnostic to such issues in this work. We simply note that in either case, there is no motif discovery algorithm defined in the literature for such data. More generally, in time series data mining, missing data is often handled by filling in the missing values with some interpolation or imputation method [8] and running the analysis unaltered. However, as we shall demonstrate, no matter what interpolation or imputation method is used, and no matter what current motif discovery algorithm is used, this can result in false negatives. To solve this problem, we introduce a novel algorithm that does not allow false negatives. The method may allow occasional false positives, but since the discovered motifs are typically examined by the human eye [11][13][25], or some subsequent analysis, the false positives (if any) can be filtered out at a later stage. We call our algorithm Motif Discovery with Missing Data, MDMS. Our MDMS algorithm is built on top of the recently introduced Matrix Profile data structure [22]. Matrix Profile-based motif discovery has been shown to have following advantageous features, all of which are inherited by our proposed MDMS: • It is simple and parameter-free. In contrast, the more general motif discovery algorithms require building and tuning spatial access methods and/or hash functions. • It requires an inconsequential space overhead, just O(n) with a small constant factor. • While the exact algorithm is scalable, for extremely large datasets we can compute the results in a timely fashion, allowing ultra-fast approximate solutions. • Having computed the motifs for a dataset, we can incrementally update them efficiently. Thus for reasonable arrival rates (< 20Hz), we can effectively maintain exact motifs on streaming data forever. • The algorithm is easy to parallelize, both on multicore processors and in distributed systems. The rest of the paper is organized as follows. Section 2 reviews related work and introduces the necessary background materials and definitions. In Section 3, we introduce our algorithms. Section 4 gives a detailed empirical evaluation. Finally, in Section 5 we offer conclusions and directions for future work. 2 RELATED WORK AND BACKGROUND Time series motif discovery has been an active area of research for over a decade [3][11][25]. However, to the best of our knowledge, there is no existing algorithm that can find motifs in the presence of missing data. In principle, any motif discovery algorithm could be used with missing data, if we use some imputation algorithm to fill in the missing values. There are hundreds of imputation algorithms in the literature (see [21] and the references therein) to choose from, but no matter which one we use, we may obtain false negatives with respect to the oracle data (the true underlying data, without missing values). To see this let us consider a simple example. Suppose we have a dataset that is composed of just three (sub)sequences: A={0,2,0,2}, B={0,2,0,2}, C={0,-1,0,2} Clearly the pair A|B is a perfect motif. Now suppose that the second value in A is missing. The most obvious imputation technique is interpolation from the two neighbors of the missing data point, giving us Amiss = {0,0,0,2}. As we can see in Figure 2, this one change means that we no longer discover the pair A|B as a perfect motif, but instead we are lead to believe that the pair A|C is the best motif in the dataset. C B B C A Amissing Figure 2: left) A contrived dataset in which the pair A|B is a perfect motif. right) If A had its second value replaced by the most common imputation algorithm, we would fail to discover A|B as the motif. While this is a trivial example, it is easy to see that no matter what imputation algorithm is used, using a simple adversarial augment we can always construct datasets for which the classic motif algorithms would produce false negatives. This problem is even more severe with complex datasets that contain a lot of highfrequency patterns and noise, for example seismology datasets. In [26] the authors showed two repeating earthquake patterns that appeared approximately 14 years apart. We have found that if there were just a handful of missing data points in one of the earthquake samples, we would be unable to detect a match between them with any common imputation method, as such high-frequency and noisy data defies the assumptions that most imputation techniques assume. Figure 2 showed that imputation methods can produce possible false negatives even if we have a random single value missing in the data. Moreover, another disadvantage of current imputation methods is that they cannot predict block-missing data well [24]. In some circumstances, due to malfunction of the sensor or other anomaly factors, we may lose reading from a sensor at consecutive timestamps (instead of sparse, single missing timestamps at random locations of the data). This is often called block-missing data. In [24] the authors proposed a state-of-the-art spatial-temporal imputation method to predict the block-missing data by learning from not only the real-valued reading of the same sensor, but also from the reading of several geographically nearby sensors. The information from other sensors greatly improved their imputation accuracy. However, this method does not apply when we only have access to one single sensor, or when all the sensors contain block-missing data at the same timestamps due to regional power outage or communication errors. Finally, note that our problem is not artificial or contrived in anyway. The literature is replete with examples of data analysts frustrated by the inability to perform motif discovery in the presence of missing data. For example, a recent paper studies recurrent water consumption behavior by Australian consumers [19]. The authors observe “A small proportion of all hourly readings are missing…, probably due to server failure or maintenance”. The authors observe “A small proportion of all hourly readings are missing…, probably due to server failure or maintenance.” The authors realize that any imputation method used here has a risk of producing false negatives by noting that “…hourly water consumption is highly unpredictable, we ignored the points of missing hours for the routine discovery, rather than approximating missing readings.” However, their solution of ignoring some data runs the risk of missing interesting patterns. 2.1 Dismissing Apparent Solutions In this section, we continue the discussion of related work, while explicitly dismissing some proposed solutions to the task at hand. The last decade has seen several distance measures for handling uncertain time series, including PROUD [23], DUST [15], PBRQ, MUNICH etc. One might believe that these measures could be used to replace the Euclidean Distance subroutine in an existing motif discovery algorithm. However, we believe this is not possible for the following reasons: • These methods assume not missing data, but uncertain data. For example, they can address the situation where an observation is not known precisely but comes from some known distribution. In [15] for example, they explicitly model the normal, exponential, and uniform error distributions. However, for generality, we do not wish to make such strong assumptions. • Even if we assume that we could somehow avail of an existing uncertain distance measure, none of them are metrics (only measures). However, all speedup techniques for exact motif discovery that we are aware of require and exploit metric properties [3][11]. This suggests we must resort to a O(n2m) brute force search (n is the length of the time series, m the motif length). However, our proposed method is O(n2). As m may be in the thousands (see Section 4), this suggests a three order of magnitude time difference. • Finally, we want to be able to guarantee that our search produces no false negatives in the face of missing data. To the best of our knowledge, no existing uncertain time series similarity measures can support this requirement. The reader may not appreciate why our task-at-hand is hard, because the analogue problem with strings is trivial. Suppose we are asked to compare the following text strings “Norwegian blue” and “Norwegian wood” under the Hamming Distance, and we encounter missing values, represented here by “*”. Norwe*ian_blu* N*rwe*ian_wood We can easily compute both the lower and upper bound of the Hamming Distance. In the former case, we would assume that all the missing values in one word are the same as their counterparts in the other word. Given that the only letters we can be sure are differing are “woo” vs “blu,” and we have a lower bound of three. In the latter case, we would assume that all the missing values in one word are different than their counterparts in the other word. These three pessimistic differences combined with the three observed differences give us an upper bound of six. However, consider the time series version of this problem. Suppose we have the following two time series: [ 0.5, 0.1, ***, ***, -0.6, -0.7, 0.0 ] [ 0.3, 0.1, 1.1, ***, -0.6, ***, 0.1 ] One might consider applying similar logic here. For example, accumulating 0.2 error (i.e. |0.5 - 0.3|) from the first pair of numbers, then 0.0 error (i.e. |0.1 - 0.1|) from the second pair, etc. However, the critical difference is that the time series must be normalized before comparison. This is because, aside from the rare and well understood exceptions [5], it is meaningless to compare time series without normalizing them first. This presents a problem as z-normalization (the most common normalization technique [4]) requires us to know the exact mean and standard deviation of the data, which are undefined when we have a single missing data point. Thus, for any pair of corresponding known points (for example the 0.5 and the 0.3 in the above), it is possible that the true (had we known the mean and standard deviation of the data to allow the correct normalization) difference between them could stay the same, increase, or decrease by arbitrary amounts. This suggests that producing tight upper or lower bounds will be nontrivial. 2.2 Definitions and Notations In this section, we introduce all of the necessary notations and definitions needed to explain our solution. Some of the text is redundant with [26], but is included for completeness. We begin by defining the data type of interest, time series: Definition 1: A time series T is a sequence of real-valued numbers ti: T = t1, t2, ..., tn, where n is the length of T. For motif discovery, we are not interested in the global properties of time series, but in the local regions, known as subsequences: Definition 2: A subsequence Ti,m of a time series T is a continuous subset of the values from T of length m starting from position i. Formally, Ti,m = ti, ti+1,…, ti+m-1, where 1 ≤ i ≤ n-m+1. We can generalize both definitions to allow for the possibly that at least one value is missing. For clarity, we differentiate such time series (including subsequences) with a “bar.” To keep both in a common notation and for implementation purposes (i.e. Matlab, etc.), we use NaNs as a placeholder for missing values. Definition 3: A missing value time series 𝑇 is a sequence of values that are either real-valued numbers or NaNs, 𝑡̅i: 𝑇 = 𝑡̅1, 𝑡̅2, ..., 𝑡̅n, where n is the length of 𝑇. In the rest of the paper, we assume that T is the actual time series of 𝑇 before the missing values were created by some process. That is to say, we would have obtained T instead of 𝑇 if the sensors were functioning properly. We do not know T precisely, but we have 𝑡i̅ = ti (1 ≤ i ≤ n) if 𝑡̅i ≠NaN. The following restates definitions introduced with the STOMP framework [26], but are reviewed for clarity. We can take any subsequence from a time series and compute its distance to all sequences. We call an ordered vector of such distances a distance profile: Definition 4: A distance profile Di of time series T is a vector of the Euclidean distances between a given query subsequence Ti,m and each subsequence in the time series T. Formally, Di = [di,1, di,2,…, di,n-m+1], where di,j (1 ≤ i, j ≤ n-m+1) is the distance between Ti,m and Tj,m. We assume that the distance is measured by Euclidean distance between z-normalized subsequences [4]. Unless otherwise stated, in the rest of the paper, we always use di,j to represent the znormalized Euclidean distance between Ti,m and Tj,m. We are interested in finding the nearest neighbors of all subsequences in T, as the closest pairs are the classic definition of time series motifs [3][11]. Note that by definition, the ith location of distance profile Di is zero and very close to zero just before and after this location. In the literature, such matches are called trivial matches [11]. We avoid such matches by ignoring an “exclusion zone” of length m/4 before and after the location of the query. In practice, we set di,j to infinity (i-m/4 ≤ j ≤ i+m/4) while evaluating Di. We use a vector called a matrix profile to represent the distances between all subsequences and their nearest neighbors: Definition 5: A matrix profile P of time series T is a vector of the Euclidean distances between each subsequence Ti,m and its nearest neighbor (closest match) in time series T. Formally, P = [min(D1), min(D2),…, min(Dn-m+1)], where Di (1 ≤ i ≤ n-m+1) is the distance profile Di of time series T. This vector is called the matrix profile, because one way to compute it would be to compute the full distance matrix of all pairs of subsequences in time series T, and then evaluate the minimum value of each column. Figure 3 illustrates both a distance profile and a matrix profile created on the same raw time series T. Q, query of length m T, synthetic data D, a distance profile Note that |D| = |T|-|Q|+1 m/4 m/4 0 2,500 T, synthetic data P, a matrix profile Note that |P| = |T|-|Q|+1 0 2,500 Figure 3: top) One distance profile (Definition 4) created from a random subsequence Q of T. If we created distance profiles for all possible subsequences of T, the element-wise minimum of this set would be the matrix profile (Definition 5) shown at (bottom). Note that the two lowest values in P are at the location of the 1st motif [3][11]. The ith element in the matrix profile P tells us the Euclidean distance from the subsequence Ti,m to its nearest neighbor in the time series T. However, it does not tell us where that neighbor is located. This information is recorded in a companion data structure called the matrix profile index. Definition 6: A matrix profile index I of time series T is a vector of integers: I=[I1, I2, … In-m+1], where Ii=j if di,j = min(Di). By storing the neighboring information this way, we can efficiently retrieve the nearest neighbor of query Ti,m by accessing the ith element in the matrix profile index. For clarity of presentation, we have confined this work to the single dimensional case; however, nothing about our work intrinsically precludes generalizations to multidimensional data. To briefly summarize this section: we can create two meta time series, the matrix profile and the matrix profile index, to annotate a time series T with the distance and location of all its subsequences’ nearest neighbors within itself. For a time series T of length n, the recently introduced STOMP algorithm [26] is able to compute the two meta files with a mere O(n2) time complexity and O(n) space complexity, which enables a fast motif discovery in a massive time series. However, the STOMP algorithm is not applicable to any time series 𝑇 with missing values (NaNs). Here we claim our MDMS algorithm can solve the missing data problem with the same time and space complexity. We leave the detailed discussion of the algorithm to Section 3. 2.3 Pseudo Missing Data Before introducing our solution to the missing data problem, we take the time to point out that the problem is more general than one may assume when we consider the generalization of the Pseudo Missing Data (PMD). Informally, we define PMD as data that technically is not missing, but effectively is. Figure 4 illustrates three kinds of PMD frequently encountered. (A) Spike 100 50 0 -50 -100 (B) Noise (C) Single Byte Maximum Value EOGright 0 400 800 1200 Figure 4: A snippet of an Electrooculogram (EOG) exhibits three kinds of pseudo missing data In Figure 4.A, we see a “spike.” Given what we know about this domain, it is inconceivable that the human eye could move fast enough to produce this data, so it is clearly an artifact. Likewise, in Figure 4.B, the dramatic increase in variance suggests that this section of data is not likely to faithfully represent the underlying physical process. Finally, in Figure 4.C the perfectly flat plateau is not reflective of reality, but is simply a region where the physical process exceeds the 8-bit precision available to record it. In all three cases, the best thing to do would be to treat the data as missing. Note that this decision is domain-dependent; there are clearly domains where a short spike represents some physical event, or where a perfectly flat plateau represents a physical limitation, which is not a quirk of the hardware/software use. 3 ALGORITHMS 3.1 An Intuitive Preview We begin by previewing our solution. As shown in [22][26], if T is used to compute a matrix profile then finding the motifs is trivial. The location of the smallest pair of values in the matrix profile is also the location in T of the optimal motif pair. Moreover, other definitions of motifs, such as the top-K motifs or range motifs [11][22], can also be easily extracted from the matrix profile. Given this, our solution to the missing data problem is to build a special matrix profile using 𝑇. This special matrix profile will be very similar to the true matrix profile, and be a (in general, very tight) lower bound for it. If we use the existing motif extraction algorithms [22][26] to pull out motifs from this matrix profile, we may have false positives, but we will have no false negatives [5]. Thus most of our contribution outlined below is to show how we can build this special matrix profile. 3.2 Lower Bound Matrix Profile To create the special matrix profile data structure, our MDMS algorithm evaluates the z-normalized Euclidean distance between every pair of subsequences within a missing value time series 𝑇. Depending on whether or not the subsequences have missing values, we may encounter three different situations. Assume that the pair of subsequences under consideration is 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 . • • • Case 1: Neither 𝑇𝑖,𝑚 nor 𝑇𝑗,𝑚 contain any missing value (NaN). Normally, this applies to most subsequence pairs within time series 𝑇 if 𝑇 contains more real-valued numbers than NaNs. The traditional exact z-normalized Euclidean distance between 𝑇𝑖,𝑚 or 𝑇𝑗,𝑚 can be evaluated in this case. Case 2: 𝑇𝑖,𝑚 contains missing values (NaNs) while 𝑇𝑗,𝑚 does not, or vice versa. Case 3: Both 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 have missing values (NaNs). In cases 2 and 3, the exact z-normalized Euclidean distance 𝑑𝑖,𝑗 between 𝑇𝑖,𝑚 or 𝑇𝑗,𝑚 cannot be evaluated. However, we can 𝐿𝐵 evaluate a lower bound of the distance between 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 , 𝑑𝑖,𝑗 , 𝐿𝐵 such that 𝑑𝑖,𝑗 ≤ min⁡(𝑑𝑖,𝑗 ). That is to say, no matter what the missing values in 𝑇𝑖,𝑚 or 𝑇𝑗,𝑚 are, we guarantee that 𝑑𝑖,𝑗 , the actual 𝐿𝐵 distance between 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 , is no less than 𝑑𝑖,𝑗 . Now we are ready to redefine Definition 4 in the context of missing values. We keep those 𝑑𝑖,𝑗 values corresponding to Case 1 𝐿𝐵 unchanged, and use 𝑑𝑖,𝑗 to replace the 𝑑𝑖,𝑗 values corresponding to Cases 2 and 3. We will then obtain a lower bound distance profile: Definition 7: A lower bound distance profile 𝐷𝑖𝐿𝐵 of a missing value time series 𝑇 is a vector 𝐷𝑖𝐿𝐵 =[ ⁡𝑑̅𝑖,1 , 𝑑̅𝑖,2 ,…, 𝑑̅𝑖,𝑛−𝑚+1 ,], where 𝑑̅𝑖,𝑗 =𝑑𝑖,𝑗 (1 ≤ i, j ≤ n-m+1) if neither 𝑇𝑖,𝑚 nor 𝑇𝑗,𝑚 contains 𝐿𝐵 NaNs (Case 1), and 𝑑̅𝑖,𝑗 =𝑑𝑖,𝑗 (1 ≤ i, j ≤ n-m+1) otherwise. Similarly, we can redefine Definition 5 and Definition 6 in the context of missing values. Definition 8: A lower bound matrix profile 𝑃 𝐿𝐵 of a missing value time series 𝑇̅ is a vector of the lower bound Euclidean distances between each subsequence 𝑇̅𝑖,𝑚 and its nearest possible neighbor (closest possible match) in 𝑇̅ . Formally, 𝑃 𝐿𝐵 = [min( 𝐷1𝐿𝐵 ), 𝐿𝐵 min(𝐷2𝐿𝐵 ),…, min(𝐷𝑛−𝑚+1 )], where 𝐷𝑖𝐿𝐵 (1 ≤ i ≤ n-m+1) is the lower bound distance profile 𝐷𝑖𝐿𝐵 of the time series T. Definition 9: A lower bound matrix profile index 𝐼𝐿𝐵 of 𝑇̅ is a 𝐿𝐵 vector of integers: 𝐼𝐿𝐵 =[𝐼1𝐿𝐵 , 𝐼2𝐿𝐵 , … 𝐼𝑛−𝑚+1 ], where 𝐼𝑖𝐿𝐵 =j if 𝑑̅𝑖,𝑗 = 𝐿𝐵 min(𝐷𝑖 ). The lower bound matrix profile gives us an optimistic approximation of the distance between every subsequence and its nearest possible neighbor in 𝑇̅. This approximation may produce false positives, but it will not allow false negatives. We believe this is the best approach for the task at hand: as [26] shows, the cost to filter out false positive motifs is very low once we have the matrix profile, but we cannot afford the occurrence of any false negatives, since they may include the most important patterns in the time series. In the following sections, we will introduce the lower bounds corresponding to Cases 2 and 3 respectively; then we will introduce our MDMS algorithm, which evaluates a lower bound of the matrix profile. 3.3 The Lower Bound Euclidean Distance 3.3.1 Case 2 Let us first consider Case 2, where 𝑇𝑖,𝑚 contains missing values (NaNs) while 𝑇𝑗,𝑚 does not, or vice versa. Without loss of generality, for now we assume 𝑇𝑖,𝑚 is the subsequence that contains missing values, and 𝑇𝑗,𝑚 is the subsequence without missing values. Figure 5 shows a visual example of this case. 𝑇𝑖,𝑚 𝑇𝑗,𝑚 0 40 𝑀 80 𝑅 Figure 5: top) A subsequence with missing values. bottom) A subsequence without missing values. Here 𝑀 = {𝑘⁡|⁡1 ≤ 𝑘 ≤ 𝑚⁡and⁡𝑡𝑖+𝑘−1 = 𝑁𝑎𝑁⁡} is a set of the locations of missing values within 𝑇𝑖,𝑚 , and 𝑅 = {𝑘⁡|⁡1 ≤ 𝑘 ≤ 𝑚⁡and⁡𝑡𝑖+𝑘−1 ≠ 𝑁𝑎𝑁⁡and⁡𝑡𝑗+𝑘−1 ≠ 𝑁𝑎𝑁⁡} is the intersection of the real-valued locations within 𝑇𝑖,𝑚 and those within 𝑇𝑗,𝑚 . To evaluate the lower bound distance of the two subsequences, first we need to z-normalize them [4][11][22][26]. We assume that 𝜇𝑖 and⁡𝜎𝑖 are the mean and standard deviation of 𝑇𝑖,𝑚 , 𝜇𝑗 and⁡𝜎𝑗 are the mean and standard deviation of 𝑇𝑗,𝑚 . Note that because 𝑇𝑖,𝑚 has missing values, we cannot evaluate 𝜇𝑖 ⁡and 𝜎𝑖 ; however, we can treat them as variables. Assume 𝑑𝑖,𝑗 is the distance between 𝑇𝑗,𝑚 and the oracle subsequence of 𝑇𝑖,𝑚 , we can easily obtain a lower bound distance of 𝑑𝑖,𝑗 by ignoring the missing part 𝑀 in Figure 5: 2 𝑡𝑖+𝑘−1 − 𝜇𝑖 𝑡𝑗+𝑘−1 − 𝜇𝑗 𝐿𝐵 𝑑𝑖,𝑗 = √min ∑ ( − ) 𝜇𝑖 ,𝜎𝑖 𝜎𝑖 𝜎𝑗 𝑘∈𝑅 2 𝐿𝐵 Assume 𝑓1 = 𝑑𝑖,𝑗 . We can linearly transform 𝑓1 as: (1) 2 𝜎𝑅 2 𝑡𝑖+𝑘−1 −𝜇 𝐿𝐵 𝑓1 = ⁡𝑑𝑖,𝑗 = ( 𝑗 ) 𝑚𝑖𝑛 ∑𝑘∈𝑅( 𝜎𝑗 𝜎 𝜇,𝜎 − 𝑡𝑗+𝑘−1 −𝜇𝑗𝑅 2 ) 𝜎𝑗𝑅 (2) 𝑅 𝑅 is shown in Figure 5. We assume 𝑇𝑖,𝑚 is the real-valued part of 𝑅 𝑇𝑖,𝑚 , 𝑇𝑗,𝑚 is the subset of 𝑇𝑗,𝑚 corresponding to 𝑅, 𝜇𝑖𝑅 and 𝜎𝑖𝑅 are 𝑅 the mean and standard deviation of 𝑇𝑖,𝑚 , 𝜇𝑗𝑅 and 𝜎𝑗𝑅 are the mean 𝑅 and standard deviation of 𝑇𝑗,𝑚 . 𝑅 Figure 6 visualizes (2). Note that 𝑇𝑗,𝑚 is z-normalized in (2), so its 𝐿𝐵 offset and scale are fixed in Figure 6; to obtain 𝑑𝑖,𝑗 , we would like 𝑅 to adjust 𝜇 (corresponding to the offset of 𝑇𝑖,𝑚 ) and 𝜎 𝑅 (corresponding to the scale of 𝑇𝑖,𝑚 ), such that the Euclidean 𝑅 𝑅 distance between 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 is minimized. scale 𝑅 𝑇𝑖,𝑚 𝐿𝐵 We can see from (6) that ⁡𝑑𝑖,𝑗 is dependent on 𝜎𝑗 (controlled by the 𝐿𝐵 missing part of 𝑇𝑗,𝑚 in Figure 7): the larger 𝜎𝑗 , the smaller ⁡𝑑𝑖,𝑗 . 𝐿𝐵 Note that 𝜎𝑗 can be as large as +∞; in that case ⁡𝑑𝑖,𝑗 becomes zero. This is a very undesirable lower bound, as any pair of missing value subsequences can be reported as a motif, even if they look very different from each other. Figure 8 shows an example of this. offset 𝑅 𝑇𝑗,𝑚 0 Figure 7). A diligent reader may have noticed that the lower bound 𝐿𝐵 expression 𝑑𝑖,𝑗 in (5) subsumes (1), the lower bound expression in Case 2. We can visualize this in Figure 7: if we remove 𝑀𝑗 , the problem is transformed to Case 2. Therefore, we can directly substitute (3), the result of Case 2, into (5): 1 𝜎𝑗𝑅 √|𝑅| min 𝑖𝑓⁡𝑞𝑖,𝑗 ≤ 0 𝜎𝑗 𝜎𝑗 𝐿𝐵 (6) ⁡𝑑𝑖,𝑗 = 1 2 𝜎𝑗𝑅 √|𝑅|(1 − 𝑞𝑖,𝑗 ) min 𝑖𝑓⁡𝑞𝑖,𝑗 > 0 𝜎𝑗 𝜎𝑗 { 40 80 𝑅 Figure 6: Different setting of 𝝁⁡and 𝝈 changes the offset and the scale of 𝑻𝑹𝒊,𝒎 . Note that the offset and scale of 𝑻𝑹𝒋,𝒎 are fixed. By solving 𝜕𝑓1 𝜕𝜇 = 0 and 𝜕𝑓1 𝜕𝜎 = 0, and substituting 𝜎 and 𝜇 back into (2), we have: 𝜎𝑗𝑅 𝜎𝑗 𝐿𝐵 ⁡𝑑𝑖,𝑗 = 𝜎𝑗𝑅 { 𝜎𝑗 √|𝑅| 𝑖𝑓⁡𝑞𝑖,𝑗 ≤ 0 (3) 2 ) 𝑖𝑓⁡𝑞𝑖,𝑗 > 0 √|𝑅|(1 − 𝑞𝑖,𝑗 Here⁡𝑅 is the intersection of the real-valued locations within 𝑇𝑖,𝑚 and those within 𝑇𝑗,𝑚 , 𝑞𝑖,𝑗 is the Pearson Correlation Coefficient 𝑅 𝑅 between 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 : 𝑞𝑖,𝑗 = ∑𝑘∈𝑅 𝑡𝑗+𝑘−1 𝑡𝑖+𝑘−1 − |𝑅|𝜇𝑖𝑅 𝜇𝑗𝑅 (4) |𝑅|𝜎𝑖𝑅 𝜎𝑗𝑅 The analysis of Case 2 is now complete. Let us turn to Case 3, where both 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 contain missing values. 3.3.2 Case 3 Figure 7 shows an example of Case 3. 𝑇𝑖,𝑚 0 40 Figure 8: Two subsequences with missing values. The real-valued parts of the subsequences look very different from each other, but if we fill the missing parts with infinitely large numbers, the z-normalized Euclidean distance of the two subsequences will become zero. Fortunately, sensor readings normally have physical limits. The accelerometer values on an iPhone 7 are limited to ± 8g (± 78.48 m/s2); virtually all medical sensors come with carefully specified limits to meet regulations (i.e., EU directive 93/42/EEC mandates that a pediatric lung ventilator monitor produces values in the range of 0 to 125cmH20), etc. Therefore, we can assume that the missing values in 𝑇𝑗,𝑚 are bounded by [𝑉𝑚𝑖𝑛 , 𝑉𝑚𝑎𝑥 ]. With this bound, we can derive the following inequality for 𝜎𝑗 2 (we refer the interested reader to [18] for the complete derivation): 𝑅𝑗 𝜎𝑗 2 ≤ 2 𝐶 2 ∑𝑘∈𝑅𝑗 𝑡𝑗+𝑘−1 + |𝑅𝑗 | (𝐵 − 𝜇𝑗 𝐴) + 4 𝑚 40 𝑅 Figure 7), 𝜇𝑗 𝑗 is the mean of the real-valued part of 𝑇𝑗,𝑚 ,𝐶 = 𝑉𝑚𝑎𝑥 − 𝑉𝑚𝑖𝑛 ,⁡𝐵 = 𝑉𝑚𝑎𝑥 𝑉𝑚𝑖𝑛 , 𝐴 = 𝑉𝑚𝑎𝑥 + 𝑉𝑚𝑖𝑛 . In practice, we set 𝑉𝑚𝑖𝑛 and 𝑉𝑚𝑎𝑥 as the minimum and maximum value of the realvalued part of 𝑇𝑗,𝑚 . 𝐿𝐵 We can now evaluate⁡𝑑𝑖,𝑗 by substituting (7) back into (6): 𝑀𝑗 𝐿𝐵 𝑑𝑖,𝑗 ={ 80 𝑅 𝑅𝑗 2 Here 𝑓𝐿𝐵 (𝑗) = 𝜎𝑅𝑗 ⁄[ Figure 7: Two subsequences with missing values. As both 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 have missing values, their means (𝜇𝑖 , 𝜇𝑗 ) and standard deviations (𝜎𝑖 ,⁡𝜎𝑗 ) can be arbitrary values. We have: 𝑡𝑖+𝑘−1 −𝜇𝑖 𝑑𝑖,𝑗 ≥ 𝑚𝑖𝑛 √𝑚𝑖𝑛 ∑𝑘∈𝑅 ( 𝜇𝑗 ,𝜎𝑗 𝜇𝑖 ,𝜎𝑖 𝜎𝑖 − 𝑡𝑗+𝑘−1 −𝜇𝑗 𝜎𝑗 2 𝐿𝐵 ) = 𝑑𝑖,𝑗 (5) Here 𝑅 = {𝑘⁡|⁡1 ≤ 𝑘 ≤ 𝑚, 𝑡𝑖+𝑘−1 ≠ 𝑁𝑎𝑁, 𝑡𝑗+𝑘−1 ≠ 𝑁𝑎𝑁} is the intersection of the real-valued locations within 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 (see (7) Here 𝑅𝑗 is a set of the locations of all the real values within 𝑇𝑗,𝑚 (see 𝑇𝑗,𝑚 0 80 √|𝑅|𝑓𝐿𝐵 (𝑗) 𝑖𝑓⁡𝑞𝑖,𝑗 ≤ 0 2 )𝑓𝐿𝐵 (𝑗) 𝑖𝑓⁡𝑞𝑖,𝑗 > 0 √|𝑅|(1 − 𝑞𝑖,𝑗 𝐶2 4 (8) 𝑅𝑗 + ∑𝑘∈𝑅𝑗 𝑡𝑗+𝑘−1 2 +|𝑅𝑗 |(𝐵−𝜇𝑗 𝐴) ]. Note that 𝑓𝐿𝐵 (𝑗) 𝑚 is based on 𝑅𝑗 , the real-valued part of 𝑇𝑗,𝑚 (recall Figure 7). However, as in Case 3 both 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 have missing values, we can analogously derive a similar lower bound expression as (8) 𝐿𝐵 based on 𝑇𝑖,𝑚 , and set the larger expression as 𝑑𝑖,𝑗 . 2 Formally, let 𝑓𝐿𝐵 (𝑝) = 𝜎𝑅𝑝 ⁄[ 𝐶2 4 𝑅 + ∑𝑘∈𝑅𝑝 𝑡𝑝+𝑘−1 2 +|𝑅𝑝|(𝐵−𝜇𝑝 𝑝 𝐴) the lower bound distance for Case 3 is: 𝑚 ] , then √|𝑅|⁡𝑚𝑎𝑥⁡(𝑓𝐿𝐵 (𝑖), 𝑓𝐿𝐵 (𝑗)) 𝐿𝐵 ⁡𝑑𝑖,𝑗 ⁡= { 2 )𝑚𝑎𝑥⁡(𝑓𝐿𝐵 (𝑖), 𝑓𝐿𝐵 (𝑗)) √|𝑅|⁡(1 − 𝑞𝑖,𝑗 • 𝑖𝑓⁡𝑞𝑖,𝑗 ≤ 0 𝑖𝑓⁡𝑞𝑖,𝑗 > 0 (9) The analysis of Case 3 is now complete. Finally, for completeness, let us briefly discuss Case 1. 3.3.3 Case 1 As neither 𝑇𝑖,𝑚 nor 𝑇𝑗,𝑚 contain any missing value in Case 1, we 𝐿𝐵 set 𝑑𝑖,𝑗 as the exact z-normalized Euclidean distance between 𝑇𝑖,𝑚 and 𝑇𝑗,𝑚 , using the following equation ([26]): 𝐿𝐵 𝑑𝑖,𝑗 = √2𝑚(1 − 𝑞𝑖,𝑗 ) (10) Note that all three cases use the same expression of 𝑞𝑖,𝑗 in (4). Now that we have the lower bound distance for any subsequence pair in 𝑇, we can also evaluate the lower bound matrix profile. 3.4 The MDMS Algorithm The STOMP algorithm [26] can obtain the matrix profile of a time series that is free of missing values, in O(n2) time with only O(n) space; as we will now show, in the face of missing data, our MDMS algorithm can obtain the lower-bound matrix profile with the same time and space complexity. TABLE I summarizes the algorithm. TABLE I. MDMS ALGORITHM Algorithm MDMS(T, m) Input: A missing value time series T, subsequence length m Output: Lower bound matrix profile P and the associated lower bound matrix profile index I of T 1 n ← Length(T), len ← n-m+1 2 vmax ← SlidingMax(T), vmin ← SlidingMin(T) Z ← PadZero(T), B ← OneZero(T), X← ElementWiseSquare(Z) 3 μz, σz ← ComputeMeanStd(Z, m) // see [14] 4 μb, σb ← ComputeMeanStd(B, m) // see [14] 5 QZ ← SlidingDotProduct(Z[1:m], Z), QZ_first ← QZ //see [26] 6 QB ← SlidingDotProduct(B[1:m], B), QB_first ← QB //see [26] 7 BZ ← SlidingDotProduct(B[1:m], Z), BZ_first ← BZ //see [26] 8 ZB ← SlidingDotProduct(Z[1:m], B), ZB_first ← ZB//see [26] 9 10 BX ← SlidingDotProduct(B[1:m], X), BX_first ← BX //see [26] 11 XB ← SlidingDotProduct(X[1:m], B), XB_first ← XB //see [26] 12 P ← CalculateLBDistance (n, m, vmax, vmin, QZ, QB, BZ, ZB, BX, XB, μz, σz, μb, σb, i), I← ones // initialization // in-order evaluation 13 for i = 2 to len for j= len downto 2 // update dot product, see [26] 14 QZ[j] ← QZ[j-1]-Z[i-1]×Z[j-1]+Z[i+m-1]×Z[j+m-1] 15 QB[j] ← QB[j-1]-B[i-1]×B[j-1]+B[i+m-1]×B[j+m-1] 16 BZ[j] ← BZ[j-1]-B[i-1]×Z[j-1]+B[i+m-1]×Z[j+m-1] 17 ZB[j] ← ZB[j-1]-Z[i-1]×B[j-1]+Z[i+m-1]×B[j+m-1] 18 BX[j] ← BX[j-1]-B[i-1]×X[j-1]+B[i+m-1]×X[j+m-1] 19 XB[j] ← XB[j-1]-X[i-1]×B[j-1]+X[i+m-1]×B[j+m-1] 20 end for 21 QZ[1]←QZ_first[i], QB[1]←QB_first[i], BZ[1]←ZB_first[i] 22 ZB[1]←BZ_first[i], BX[1]←XB_first[i], XB[1]←BX_first[i] 23 D ← CalculateLBDistance (n, m, vmax, vmin, QZ, QB, BZ, ZB, 24 BX, XB, μz, σz, μb, σb, i) P, I ← ElementWiseMin(P, I, D, i) 25 26 end for 27 return P, I Before discussing the algorithm in detail, we first need to introduce three important auxiliary time series (shown in line 3), 𝑍, 𝑋 and 𝐵. • We define 𝑍 = 𝑧1 , 𝑧2 , … 𝑧𝑛 , such that 𝑧𝑖 = 𝑡𝑖 if 𝑡𝑖 ≠ 𝑁𝑎𝑁 , and 𝑧𝑖 = 0 if 𝑡𝑖 = 𝑁𝑎𝑁. We can simply obtain 𝑍 by filling zeros in the locations of 𝑇 where the data is missing. • We define 𝑋 = 𝑥1 , 𝑥2 , … 𝑥𝑛 = 𝑧1 2 , 𝑧2 2 , … 𝑧𝑛 2 . We define 𝐵 = 𝑏1 , 𝑏2 , … 𝑏𝑛 , such that 𝑏𝑖 = 1 if 𝑡𝑖 ≠ 𝑁𝑎𝑁 , and 𝑏𝑖 = 0 if 𝑡𝑖 = 𝑁𝑎𝑁 . We can see that 𝐵 indicates the locations of the real-valued numbers and missing values in 𝑇. With these three auxiliary time series and the techniques introduced in [14][26], we can evaluate any lower bound distances introduced in the last section in O(1) time with O(n) space. The MDMS algorithm is very similar to the STOMP framework [26]. In line 2 we evaluate the maximum and minimum values of the real-value part of every subsequence in T. Lines 4-5 evaluate the mean and standard deviation of every subsequence in 𝑍 and 𝐵, lines 6-11 initialize all the dot product vectors. We initialize the lower bound matrix profile P and matrix profile index I in line 12. Lines 13-26 iteratively evaluate the lower bound distance profile D, and update P and I if necessary. The CalculateLBDistance algorithm in lines 12 and 24 is shown in TABLE II. TABLE II. ALGORITHM TO CALCULATE LOWERBOUND DISTANCE PROFILE Algorithm CalculateLBDistance (n, m, vmax, vmin, QZ, QB, BZ, ZB, BX, XB, μz, σz, μb, σb, i) Input: the length n of time series T, the subsequence length m, the maximum/minimum possible value vector vmax/vmin, dot product vectors QZ, QB, BZ, ZB, BX, XB, means and standard deviations μz, σz, μb, σb of time series Z and B, current subsequence index i Output: Lower bound distance profile D 1 for j= 1 to n-m+1 2 ui ← ZB[j] / QB[j], uj ← BZ[j] / QB[j] //𝜇𝑖𝑅 and 𝜇𝑗𝑅 3 vi← XB[j] / QB[j] - ui ^2, vj← BX[j] / QB[j] - uj ^2 //𝜎𝑖𝑅 and 𝜎𝑗𝑅 q ← (QZ[j] / QB[j]- ui × uj) / sqrt(vi × vj) // (4) 4 if QB[j]==m then // Case 1, |𝑅| = 𝑚 5 6 D[j] ←2×m×(1-q) // (10) else 7 if max(μb[i],μb[j])==1 then //Case 2 8 if μb[i]>μb[j] then vo←vi, v←σz[i]^2 9 else vo←vj, v←σz[j]^2 10 end if 11 if q<=0 then D[j] ← QB[j] × vo / v // (3) 12 else D[j] ← QB[j] × vo / v × (1 - q^2) 13 end if 14 else //Case 3 15 v1←vmax[i], v2←vmin[i], C←v1–v2, B←v1×v2, A←v1+v2 16 ur←μz[i]/μb[i], vr←(σz[i]^2+μz[i]^2)/μb[i]-ur^2 17 f1=vi/(ub[i]×(vr+B+ur×(ur–A))+C^2/4) 18 v1←vmax[j], v2←vmin[j], C←v1–v2, B←v1×v2, A←v1+v2 19 ur←μz[j]/μb[j], vr←(σz[j]^2+μz[j]^2)/μb[j]-ur^2 20 f2=vj/(ub[j]×(vr+B+ur×(ur–A))+C^2/4) 21 if q<=0 then D[j] ← QB[j] × max(f1, f2) // (9) 22 else D[j] ← QB[j] × max(f1, f2) × (1 - q^2) 23 end if 24 end if 25 end if 26 27 end for 28 return D The CalculateLBDistance algorithm evaluates all n-m+1 lower bound distance values in D with equations (3), (9) and (10). Line 4 evaluates (4). Case 1 is handled in lines 5-6. Lines 8-14 handle Case 2, and Case 3 is evaluated in lines 15-24. We can see that each loop of the CalculateLBDistance algorithm in TABLE II (lines 2-26) can be evaluated in O(1) time, so the time complexity of CalculateLBDistance is O(n). The space needed to store all the vectors in TABLE I is O(n), and each loop in lines 1425 of TABLE I takes O(n) time. Therefore, the time complexity of MDMS algorithm is O(n2) and the space complexity is O(n), the same as STOMP [26]. Furthermore, we can see that unlike most imputation algorithms [21], our MDMS algorithm is extremely model-free and parameterfree. The only inputs to the algorithm are the time series and the subsequence length. In the next section, we will use two case studies to demonstrate the robustness and efficacy of our ultra-fast, parameter-free motif discovery algorithm in the face of missing data. 4 EXPERIMENTAL EVALUATION We begin by noting that all the code and data used in this work are archived in perpetuity at [18]. In the following two case studies, we consider both random-missing data (data with sparsely located, random missing timestamps) and block-missing data (data with consecutive missing timestamps). For each case study, we compare our method with the commonly used strawman in the literature to handle missing data: linear imputation, as shown in Figure 9. besides the two deep valleys). This shows the robustness of our algorithm in the face of random missing data. Next, we consider the “larger gaps” (or the block-missing data) case. Here, instead of removing individual data points, we removed two consecutive blocks of length 25. As shown in Figure 10 (orange curve), the shape of the lower bound matrix profile still looks very similar to that of the oracle matrix profile (red curve). We see only two deep valleys at the vicinity of the motifs, so no false positive patterns are discovered. The result demonstrates that our lower bound matrix profile is robust against producing false positives. To test the robustness of our algorithm against false negatives, we removed a block of missing data at the center of the second earthquake pattern. The length of the missing block is 400, which is 20% of the subsequence length. In Figure 11, we compared our lower-bound matrix profile result with the matrix profile generated by linear imputation. We removed a block of 400 data points here imputation Raw data Figure 9: To evaluate our method, we compare our result with that of linear imputation. Oracle Matrix Profile Matrix Profile (MDMS) 4.1 Case Study: Seismological Data Repeated pattern (i.e. motif) discovery is a fundamental tool in seismology, which allows for the discovery of foreshocks, triggered earthquakes, swarms, volcanic activity, and induced seismicity [25]. However, this domain is replete with missing data. For example, a classic paper notes “A frequent dilemma in spectral analysis (in seismology) is the incompleteness of the data record, either in the form of occasional missing data or as larger gaps” (our emphasis) [1]. In this experiment, we demonstrate that we can handle both cases. We consider a dataset for which we know the answer from external sources. On April 30th, 1996, there was an earthquake of magnitude 2.12 in Sonoma County, California. Then, on December 29 th, 2009, about 13.6 years later, there was another earthquake with a similar magnitude. To allow the results to be visualized in a single plot, we edited this data such that the two earthquakes happen just 15 seconds apart. We set the subsequence length as 2,000, which corresponds to one second of data. As shown in Figure 10 (red curve), when there is no missing data the matrix profile correctly discovers the locations of the two earthquakes. Raw data Oracle Matrix Profile Matrix Profile (Missing points) 50 data points missing Matrix Profile (Missing blocks) 50 data points missing 0 25s Figure 10: A raw seismograph contrived such that two earthquakes from the same region happen 15 seconds apart. The matrix profile computed with no missing data (red curve) finds the true event, as does MDMS even in the presence of missing points (green curve) or missing blocks (orange curve). To test our algorithm for the “occasional missing data” case, we randomly deleted 50 data points. As Figure 10 (green curve) shows, the matrix profile is still minimized at the correct location, and there are no false positives (no other small values in the matrix profile Matrix Profile (imputation) Oracle Matrix Profile 0 25s Figure 11: We removed 400 consecutive data points at the center of the second earthquake pattern. The oracle matrix profile computed with no missing data (black curve) finds the true event, as does MDMS (red curve) even in the presence of a large missing block. The Matrix Profile generated after linear imputation (green curve) fails to capture the minimum points within the oracle matrix profile. We can see that the lower-bound matrix profile generated by our MDMS algorithm (red curve) agrees closely with the oracle matrix profile (black curve) at the vicinity of the two earthquake patterns, while the matrix profile generated after linear imputation (green curve) shows a high value at these locations. 0 2000 0 2000 Figure 12: The first motif found by the MDMS algorithm (right) in the presences of a large missing block is identical with the first motif found in the oracle data (left). As a result, the MDMS algorithm successfully captures the 1 st motif (as shown in Figure 12) even in the presence of a large missing block within one earthquake pattern, while the imputation method misses the 1st motif within the oracle data. This illustrates two major strength of our algorithm over imputation methods. Firstly, our algorithm does not allow false negatives. Secondly, our algorithm is more robust to large missing blocks as it does not change the data, while imputation method can change the data a lot. In the next case study, we will further demonstrate the robustness of MDMS in the presence of missing blocks. 4.2 Case Study: Activity Data from Video Time series extracted from video often has missing data reflecting “frame drops” due to bandwidth congestion [6]. To test our algorithm in this context, we examine the activity dataset of [20]. This dataset consists of a 13.3 minute 10-fps video sequence of an actor performing various activities. From this data, the original authors extracted 721 channels of the optical flow time series, and the length of each time series is 8,000. We consider the time series corresponding to the 533th channel, which is suggestive of the structure in places but is noisy. The data is shown in Figure 13.top, the subsequence length is 120. From the oracle matrix profile in Figure 13 (black curve), we can extract the 1st motif in the oracle data. To test the performance of the MDMS algorithm, here we remove 12 consecutive data points in the center of one of the 1st motif patterns. In Figure 13, we compare our lower-bound matrix profile result (red curve) with the matrix profile generated by linear imputation (green curve). We removed two blocks of length p Data Oracle Matrix Profile p=100 1st motif p=400 1st motif p=700 1st motif p=800 1st motif We removed a block of 12 data points here Data Oracle Matrix Profile 1st motif Matrix Profile (MDMS) 1st motif Matrix Profile (Imputation) The deep valley disappears 0 8000 Figure 13: A raw activity time series. We removed 12 consecutive data points in one of the 1st motif patterns in the time series. The oracle matrix profile computed with no missing data (black curve) finds the true motif starting at the 540th and the 622nd data points. With the presence of 12 missing data points, the MDMS algorithm finds the same motif as the oracle result (red curve), starting at 520th and 602nd data points. The Matrix Profile generated after linear imputation (green curve) fails to capture the two deep valleys within the oracle matrix profile and thus misses the 1st motif. We can see that the oracle matrix profile (black curve) shows two apparent valleys at the locations of the 1st motif, as does the lower bound matrix profile generated by MDMS (red curve). The 1st motif discovered by the MDMS algorithm (shown in Figure 14.right) is identical to the oracle motif (shown in Figure 14.left). In contrast, the matrix profile generated by imputation does not have these valleys, and thus misses the 1st motif of the oracle data. Figure 15: Lower bound matrix profiles corresponding to different missing block lengths. We removed 2 blocks of length p from the seismograph. The oracle matrix profile (black curve) finds the true motif. For p=100, p=400 and p=700, MDMS is able to find the true event as the 1st motif. When p=800, MDMS finds a false positive as the 1st motif. Note that the removed blocks are not within the two repeated earthquake patterns. As a result, the lower bound matrix profiles are the same as the oracle matrix profile at the vicinity of the two repeated patterns, while lower at other locations. In other words, it is easier to detect false positives with such setting. We can see that when p=100 (5% of the subsequence length m), the side valleys in the oracle Matrix Profile become deeper, and two more side valleys show up at the vicinity of the removed blocks. As p increases, all the side valleys become deeper and deeper. For p=100, p=400 and p=700, we are able to find the true event as the 1st motif with MDMS. However, when p=800 (40% of m), the 1st motif (corresponding to the minimum point of the lower bound matrix profile) is no longer the true event. We show this false positive motif pair in Figure 16. 0 2000 Figure 16: The 1st motif found by the MDMS algorithm when p=800. 0 120 0 120 Figure 14: The first motif found by the MDMS algorithm (right) is identical to the first motif within the oracle data (left), despite a small phase shift. We can see from Figure 14 that though a large portion of the blue pattern is missing, our MDMS algorithm still finds it very similar with the red pattern. This example further demonstrates that our algorithm is robust against missing true motif patterns. 4.3 Quantifying the Robustness of MDMS As MDMS evaluates the lower bound matrix profile, it naturally does not allow false negatives, but it can produce false positives. Here we perform two “stress tests” to evaluate the robustness and limitations of our MDMS algorithm against false positives. We use the seismograph data in Figure 10 again for the stress test. The subsequence length in this dataset is m=2,000. We first test the sensitivity of MDMS over the length of missing blocks. Here we remove two missing blocks of length p, located at 7.5s and 15s respectively, from the data. In Figure 15, we show how the lower bound matrix profile varies as we increase p. The two subsequences are both at the vicinity of the second missing block in Figure 15. With a close inspection we can see why this pair is reported as the 1st motif by MDMS. If we fill in the missing part of each subsequence with their counterpart in the other subsequence (shown by the arrows in Figure 16), the two can be very similar to each other. Since MDMS does not allow false negatives, it will capture and report this possible matching pair. Furthermore, Figure 16 implies that p cannot be larger than 50% of the subsequence length, otherwise MDMS will be able to find a perfect match (with one subsequence missing the first 50% and the other missing the second 50%). When p > 700 (35% of the subsequence length), we are already very close to this limit. Therefore, in Figure 15 we can see very deep side valleys at the vicinity of the missing blocks, and we are prone to detect false positives. When p ≤ 400 (20% of the subsequence length), the two main valleys corresponding to the true events dominate, so we will not detect false positives. We have demonstrated that MDMS is robust against discovering false positives when there are two missing blocks, and when the length of the two blocks are within a reasonable range. Next, we “stress test” the sensitivity of MDMS over the total number of missing values in the data. We again use the seismograph dataset in Figure 10, which consist of 50,000 data points. The lower bound matrix profile results are shown in Figure 17. 10% 10% 20% 20% 30% 30% 40% 40% Figure 17: Lower bound matrix profiles corresponding to various percentage of data missing. left) random-missing data right) blockmissing data. In the first run, we randomly selected 5,000, 10,000, 15,000 and 20,000 points to remove from the data. From Figure 17.left, we can see that the scale of the matrix profile decreased as more points are missing. However, even when 40% of the data is missing, the two valleys corresponding to the true events (recall Figure 10) still dominate. In the second run, we removed blocks of length 200 from the data. The missing blocks were uniformly distributed, and the number of missing blocks increased from 25 to 100. Figure 17.right shows that even when 30% of the data is missing, the two main valleys still dominate. When 40% of the data is missing, the two main valleys are no longer apparent, but we can still find the true events as the 1st motif. The experiment demonstrates that MDMS is robust against detecting false positives even if a large percentage of data is missing. 5 CONCLUSIONS We introduced what we believe to be the first time series motif discovery algorithm that can find motifs in the presence of missing data. The algorithm has the same time and space complexity as the fastest known algorithm for motif discovery [26]. We formally proved the admissibility of our algorithm, it does not produce any false negatives. Experimental results show that our algorithm is also robust against false positives even when a large portion of the data is missing. Because our algorithm is based on creating a special version of the matrix profile [22][26], our work may have implications for other algorithms that can exploit the matrix profile, including discord discovery and time series joins. The lower bounds introduced can also be used to accelerate various length motif discovery. We leave such considerations to future work. 6 REFERENCES [1] Baisch, S. and Bokelmann, G.H. Spectral analysis with incomplete time series: an example from seismology. Computers and Geosciences, 25.7(1999): 739-750. [2] Buza, K., and Schmidt-Thieme, L. Motif-based classification of time series with bayesian networks and svms. In Advances in Data Analysis, Data Handling and Business Intelligence 2009: 105-114. [3] Chiu, B., Keogh, E., and Lonardi, S. Probabilistic discovery of time series motifs. ACM SIGKDD 2003: 493-498. [4] Ding, H., Trajcevski, G., Scheuermann, P., Wang, X., and Keogh, E. J. Querying and Mining of Time Series Data: Experimental Comparison of Representations and Distance Measures. PVLDB 1.2 (2008): 1542-1552. [5] Faloutsos, C., Ranganathan, M., Manolopoulos, Y. Fast subsequence matching in time-series databases. SIGMOD 1994: 419-429. [6] Gangadharan, D. et al. Video Quality Driven Buffer Sizing via Frame Drops. RTCSA. 1 (2011): 319-328. [7] Gao, H., Bi, S., Zhang, R. and Tang, S., The design of a throwable two-wheeled reconnaissance robot. In 2012 IEEE Robotics and Biomimetics (ROBIO), pp 2150-2155. [8] Honaker, J. and King, G. What to do about missing values in time‐series cross‐section data. American Journal of Political Science, 54.2 (2010): 561-581. [9] Kelly, G. Body temperature variability: a review of the history of body temperature and its variability due to site selection, biological rhythms, fitness, and aging. Med Rev. 11.4 (2006): 278. [10] McGovern, A., Rosendahl, D.H., Brown, R.A. and Droegemeier, K.K. Identifying predictive multi-dimensional time series motifs: an application to severe weather prediction. Data Mining and Knowledge Discovery, 22.1 (2011): 232-258. [11] Mueen, A., Keogh, E., Zhu, Q., Cash, S. and Westover, B. Exact Discovery of Time Series Motifs. SDM 2009: 473-484. [12] Neocleous, A., Petkov, N., Schizas, C.: Finding repeating Stanzas in monophonic folk songs of Cyprus. 6th Cyprus Workshop on Signal Processing and Informatics 2013. pp 26. [13] Patnaik, D., Manish, M., Sharma, R. K., and Ramakrishnan, N. Sustainable Operation and Management of Data Center Chillers Using Temporal Data Mining. KDD 2009: 1305-1314. [14] Rakthanmanon, T., et al. Searching and Mining Trillions of Time Series Subsequences under Dynamic Time Warping. KDD 2012: 262-270. [15] Sarangi, S. R. and Murthy, K. DUST: a generalized notion of similarity between uncertain time series. KDD 2010: 383-392. [16] Shokoohi-Yekta, M., Chen, Y., Campana, B., Hu, B., Zakaria, J., and Keogh, E. Discovery of meaningful rules in time series. KDD 2015: 1085-1094. [17] Song, W.Z., Huang, R., Xu, M., Ma, A., Shirazi, B. and LaHusen, R. Air-dropped sensor network for real-time highfidelity volcano monitoring. In Proceedings of the 7th international conference on Mobile systems, applications, and services, ACM, 2009, 305-318. [18] Supporting Webpage: https://sites.google.com/site/motifmissingdata/ [19] Wang, J., Cardell-Oliver, R. and Liu, W. An Incremental Algorithm for Discovering Routine Behaviours from Smart Meter Data. Knowledge-Based Systems, 113 (2016): 61-74. [20] Veeraraghavan, A., Chellappa R., and Roy-Chowdhury, A. The function space of an activity. Computer Vision and Pattern Recognition. 1 (2006): 959-968. [21] Weber, M. and Denk, M. Imputation of Cross-Country Time Series: Techniques and Evaluation. In European Conference on Quality in Official Statistics, 2010. [22] Yeh, C.C.M. et al. Matrix Profile I: All Pairs Similarity Joins for Time Series: A Unifying View that Includes Motifs, Discords and Shapelets. ICDM 2016: 1317-1322. [23] Yeh, M.Y., Wu, K. L., Yu, P. S. and Chen, M. S. PROUD: a probabilistic approach to processing similarity queries over uncertain data streams. EDBT 2009: 684-695 [24] Yi, X., Zheng, Y., Zhang, J. and Li, T. ST-MVL: Filling Missing Values in Geo-sensory Time Series Data. IJCAI 2016: 2704-2710 [25] Yoon, C., O’Reilly, O., Bergen, K., and Beroza, G. Earthquake Detection through Computationally Efficient Similarity Search. Sci. Adv. 11.1(2015): e1501057. [26] Zhu, Y. et al. Matrix Profile II: Exploiting a Novel Algorithm and GPUs to Break the One Hundred Million Barrier for Time Series Motifs and Joins, ICDM 2016: 739-748.
2cs.AI
Shape-based defect classification for Non Destructive Testing Gianni D’Angelo, Salvatore Rampone University of Sannio Dept. of Science and Technology Benevento, Italy {dangelo, rampone}@unisannio.it Abstract— The aim of this work is to classify the aerospace structure defects detected by eddy current non-destructive testing. The proposed method is based on the assumption that the defect is bound to the reaction of the probe coil impedance during the test. Impedance plane analysis is used to extract a feature vector from the shape of the coil impedance in the complex plane, through the use of some geometric parameters. Shape recognition is tested with three different machine-learning based classifiers: decision trees, neural networks and Naive Bayes. The performance of the proposed detection system are measured in terms of accuracy, sensitivity, specificity, precision and Matthews correlation coefficient. Several experiments are performed on dataset of eddy current signal samples for aircraft structures. The obtained results demonstrate the usefulness of our approach and the competiveness against existing descriptors. Keywords— Non-destructive testing (NDT); learning algorithm; signature-based classifier; content-based image retrieval (CBIR); shape geometric descriptor (SGD); eddy current testing (ECT). I. INTRODUCTION In aircraft manufacturing and maintenance, NonDestructive Testing (NDT) is widely used for evaluating the property of materials, components, systems, without causing damage during the analysis. In aerospace industry, NDT is not only necessary but plays a crucial role in the risk management. There are many NDT technologies that are available for aircraft inspections. The most common methods are based on visual and optical testing: optical holography, X-ray, ultrasonic wave, infrared detection, X-ray and ultrasonic C-scan. Eddy Current Testing (ECT) [1] is one of the NDT methods that are widely used in detecting surface or subsurface crack of conductive materials. Recent technologies have made ECT more powerful and useful in quality assurance. Modern ECT techniques are based on low-cost methods and are easily applicable for the inspection where premature failures could give rise to economic or related to human life problems. However, most of the ECTs in maintenance, are currently conducted by human inspectors, whose individual experiences may yield to differences in the interpretation of the results. Generally, the human operators take decision on the acceptability of the piece under test by evaluating typical bidimensional graphs that depict the reactance versus the resistance of the coil impedance of the ECT probe. Reliable human performance is vital to inspections and tests. Inadequate human performance could lead to missed defects and inaccurate reports, with potentially serious safety and cost consequences. In addition, manpower assigned to such tasks imply significant recurrent costs and it is time consuming in itself. The presence of defects and damages in aerospace structure pose a significant threat to the integrity of the structures and also in the reliability and safety of the inspections. For this reason, the accuracy of diagnosis of aerospace materials cannot be left only to the human capacity but also entrusted to mathematical models and advanced methods of data processing [2]. Different solutions have been proposed to address this issue [3-5]. The main methods are based on two principal approaches: the model-based and the model-free techniques. The first refers to the use of computer simulators to determine the material electromagnetic parameters by measuring the probe impedance and relying on mathematical models of probe geometry [6]. The second are based on the use of signal analysis acquired during the test. In this approach, the methods of spectrum analysis and pattern recognition are often used in multi-parametric control for data processing [7]. The signals acquired from the sensors are processed in the frequency domain to extract the magnetic field information. In fact, the presence of a defect causes an asymmetricity in the induced signals that can be highlighted in the frequency domain by analyzing the signal harmonics [8]. In [9] the authors proposed a "model-free" method based on the spectrum analysis and on a proper algorithm that uses a machine learning technique [10]. However, the application of these methods requires sophisticated techniques for processing signals that, for a large amount of date, lead to have long process time [11]. These do not allow the automation of the test and deprive them of the same dynamism typical of a system able to adapt itself to changes in the parameters of the testing system at run-time. NDT should be performed with methods able to collect the most comprehensive information about new defects, expand existed base of defects and increase diagnostics system precision in runtime. The key to a successful testing system is to choose the right features extraction method that represent the defect as accurately and uniquely as possible in a short time. Recently, the image analysis and pattern recognition techniques in NDT are being increasingly used to increase the objectivity, consistency and efficiency of the testing [12, 13]. This new “model-free” based approach involves extracting of the hidden useful “knowledge” [14-16] embedded in the images representing the ECT results. The fast growth of multimedia data [17], mainly due to the wide spread of digital devices [18], has lead to the developing of high-performance techniques in the field of image retrieval. These techniques are generally known as Content-Based Image Retrieval (CBIR) systems [19]. CBIR systems perform the image retrieval through a similarity process which is defined in terms of visual features with more objectiveness. The Shape Geometric Description (SGD) [20] of the objects, represented in images, is one of the most significant properties used in CBIR tasks. One of the main advantages of this approach is the performance in terms of velocity. In fact, the feature data for each of the visual attributes of each image is very much smaller in size compared to the image data. In this paper we propose a CBIR system to recognize the image that assumes the impedance of the EC probe coil in the complex plane for a specific defect. We make use of some geometric parameters to descript the shape of the probe coil impedance. Shape recognition is tested on the automated classification of eddy current signatures with three different machine-learning based classifiers: decision trees, neural networks and Naive Bayes. The performance of the proposed detection system are measured in terms of accuracy, sensitivity, specificity, precision and Matthews correlation coefficient. The paper is organized as follows. In Section 2 we briefly give a description of the image retrieval issue. Discussions related to the features extraction from the EC impedance shape in the complex plane are presented in section 3. In Section 4 we explain the SGD-based proposed experiments and we show the related results. Last section is devoted to the conclusions. II. There are many different implementations of CBIR, so that the CBIR has attracted the attention of researchers across several disciplines. Nevertheless, the key to a good retrieval system is to choose the right features that better represent the images while minimizing the computation complexity. B. Shape geometric descriptor (SGD) The shape descriptor aims to measure geometric attributes of an image. There are many different kinds of shape matching methods, and the progress in improving the matching rate has been substantial in recent years. However, these descriptors are categorized into two main groups: region-based shape descriptors and contour-based shape descriptors [29]. The first method uses all the pixel information within a shape region of a image. Common region-based methods make use of moment descriptors [30] that include: geometric moments, Legendre moments, Zernike moments and others [31]. Contour-based approaches use only the information related to the boundary of a shape region and do not consider the shape interior content. These include Fourier descriptor, Wavelet descriptors, curvature scale space and shape signatures [32]. In Fig. 1 is depicted typical geometric parameters for the shape signatures. They include: Area (A), perimeter (P), centroid (G), orientation angle (α), principal inertia axes, width (W), length (L) and surfaces of symmetry (Si) for an equivalent ellipse image region. IMAGE RETRIEVAL A. CBIR Content Based Image Retrieval (CBIR) is an actively researched area in computer vision whose goal is to find images similar in visual content to a given query from an image dataset [21]. Image analysis can be based on several distinct features such as color [22], texture [23], shape [24] or any other information that can better describe the image. A typical CBIR system extracts the features from each image in the dataset and stores them in the database. Then, when similar images are searched using a “query” image (the image is used as a query), a feature vector is first extracted from this query image, then a distance between the calculated vector and the database image features is computed. Typical distance metrics between the feature vectors include: Canberra distance, Euclidean distance, Manhattan metric, Minkowski metric and others [25]. If the calculated distance is small, then the images compared are considered similar. Compared with the traditional methods, which represent image contents by keywords, the CBIR systems are faster and more efficient. The main advantage of CBIR system is that it uses image features rather than image itself. For these reason the application areas are numerous and different: remote sensing, geographic information systems, weather forecasting, medical imaging [26] and in the last few years also in image search on the Internet [27, 28]. Fig. 1. Typical geometric parameters. From these base parameters some advanced parameters (not changing when the original object is submitted to translation, scale changes and rotations) can be derived. They include [33]: Compactness: C=4πA/P2. It represents the ratio of the shape area to the area of a circle having the same perimeter. Elongation: E=L/W. It is defined by the ratio of the length to the width of the minimal rectangle surrounding the object called also the minimal bounding box. Rectangularity: R=A/(L x V). It represents how rectangular a shape is. It is equal to the ratio of the shape area to the area of its minimal bounding box. Eccentricity: It represents the measure of aspect ratio. It is obtained from the ratio of the minor axis to the major axis in the object equivalent ellipse. Convexity: It is defined as the ratio of perimeters of the convex hull over that of the original contour. III. PROPOSED FEATURE EXTRACTION METHOD We investigate the application of the CBIR techniques to characterize the image that assumes the impedance of the EC probe coil in the complex plane for a specific defect. Eddy current testing is based on the physics phenomenon of electromagnetic induction [34]. In an eddy current probe, an alternating current flows through a wire coil and generates an oscillating magnetic field. If the probe and its magnetic field are taken close to a conductive material like a metal test piece, a circular flow of electrons known as an eddy current will begin to move through the metal like swirling water in a stream. That eddy current flowing through the metal will in turn generate its own magnetic field, which will interact with the coil and its field through mutual inductance. Changes in metal thickness or defects like near-surface cracking will interrupt or alter the amplitude and pattern of the eddy current and the resulting magnetic field. This in turn affects the movement of electrons in the coil by varying the electrical impedance of the coil. In this approach, the presence of damage is characterized by the changes in the signature of the resultant signal that propagates through the structure. So, changing material parameters corresponds to a particular output impedance that is characterized by a specific shape in the complex plane. We use the shape of the impedance in the complex plane as signatures to identify changes in the test piece. In Fig. 2 a typical shape of coil impedance in the complex plane for an aluminum sample with notch perpendicular of width 0.3 mm and depth 1 mm is shown. L W α Fig. 2. Typical shape of coil impedance in the complex plane. The dotted line is principal inertia axis. (L,W,α) is the feature vector. As depicted, we intercept the principal inertia axis and we use the set (L,W,α), composed by the length (L), width (W) and orientation angle (α) of the shape, as feature vector that characterizes a defect. These features are used as input to machine learning based classifiers. IV. EXPERIMENTS AND RESULTS In order to evaluate the suitability and usefulness of the proposed set of features to describe shape of the coil impedance in the complex plane, we selected twelve types of defects, with 20 shape for each category (class). The classification performance was evaluated using a ten-fold cross-validation method. A. Sample data The data used in the study refers to a subset of a database with EC signal samples for aircraft structures [35]. The overall database is divided in 4 parts. The first contains 240 records acquired on an aluminum sample with notches of width 0.3 mm, depth 0.4, 0.7, 1, and 1.5 mm perpendicular, depth 0.4, 0.7, 1, and 1.5 mm with an angle of 30 degrees, 0.7, 1 and 1.5 mm with an angle of 60 degrees and 1.5 mm with an angle of 45 degrees. The second refers to 150 records, notches of width 0.2 mm, depth 1, 3 and 5 mm, both perpendicular and 45 degrees orientation of a stainless steel structure. Third database refers to two-layer aluminum aircraft structure with rivets, two notches below the rivets in the first layer (width 0.2 mm, length 2.5 mm, angle 90 degrees and 30 degrees) and two in the second layer (width 0.2 mm, length 2.5 mm and 5 mm, angle 90 degrees), two defect-free rivets. The fourth sample refers to four-layer aluminum structure (layer thickness 2,5 mm) with rivets containing 4 notches (width 0.2 mm, length 2.5 mm, angle 90 deg ) below the rivets in the first, second, third or fourth layer, four defect-free rivets. In this work we used the first part of the database that contains 240 records acquired on an aluminum sample with notches which differ in the depth and angle. Each record contains 4096 samples, an sampling frequency of 10KHz and two canals (real and imaginary part) for each acquired measure. B. Pre-Processing To obtain an image suitable for feature extraction, we performed a noise reduction by removing the irrelevant information from the shapes. This was accomplished by detecting and extracting the image regions of interest, cropping them through their bounding box. As depicted in Fig. 2 the most noise is concentrated on the top-right side which represents a high value for both real and imaginary part of the coil impedance. This is confirmed by also the spectrum analysis of each single channel of the samples [9]. Using Matlab program, the upper-right side of the image was removed through a sorting and cutting procedures acting on the raw data. At this stage we had to guarantee that all the features computed were independent from noise. Then, by using the Image Processing toolbox, the centroid, the principal inertia axis and then the feature set (L,W,α) were calculated for each record. To take a look at the separation of the 12 classes provided by the proposed feature vectors, in Fig. 3 a tridimensional scatter graph is reported. There is a good separation among classes. This indicates that the feature vectors obtained are able to yield good classification from the classifiers. assumes the impedance of the EC probe coil in the plane complex as signatures to identify changes in the test piece. We have made use of information related to the boundary of the image through a feature vector composed by only three geometric parameters: length, width and orientation angle of the shape. To evaluate the suitability and usefulness of the proposed set of features we tested them with three different machine-learning based classifiers: J48 decision trees, Multilayer Perceptron neural network and Naive Bayes. The results, evidencing an accuracy rate of almost 100%, demonstrates the effectiveness of our approach. ACKNOWLEDGMENTS Fig. 3. Tridimensional scatter graph of the proposed feature vectors for 12 classes. C. Classifiers To evaluate the proposed shape descriptor system as recognition of structure defect we tested it with three different machine-learning based classifiers: J48 decision trees, Multilayer Perceptron neural networks and Naive Bayes [36]. The classification performance of each classifier is evaluated using the ten-fold cross-validation method [37], a model validation technique for assessing how the classification results will generalize for an independent data set. The 12 different classes are derived from the overall dataset by comparing each defect to the others. Then, all the data have been randomly divided into 10 disjoint subsets (folders), each containing approximately the same amount of instances. In each experiment, nine folders have been used as training data, while the remaining folder is used as validation. This process has been repeated 10 times, for each different choice of the validation folder. The performance results, reported in Table I as mean values, are presented in terms of accuracy, sensitivity, specificity, precision and Matthews correlation coefficient [38]. TABLE I. Classifier J48 Naive Bayes Multilayer Perceptron EXPERIMENTAL RESULTS Mean values of the “ten-fold cross validation” results Accuracy Sensitivity Specificity Precision Matthews 0.96 0.74 0.98 0.81 0.75 0.95 0.68 0.97 0.67 0.63 0.98 0.85 0.99 0.89 0.84 The experiment results confirm the good class separation shown in Fig. 3. In particular, the performance parameters related to the neural network were very high, near to 1. V. CONCLUSIONS In this paper we have investigated the application of a "CBIR" method to characterize aerospace structure defects based on eddy current testing. We have used the shape that This work has been supported in part by Distretto Aerospaziale della Campania (DAC) in the framework of the CERVIA project - PON03PE_00124_1. REFERENCES [1] X.-M. Pei, H.-S. Liang, Y.-M. Qia, "A frequency spectrum analysis method for eddy current nondestructive testing," Proceedings of the IEEE International Conference on Machine Learning and Cybernetics, Vol. 3, 2002. [2] S. Yacout,M. Meshreki, H. Attia,"Monitoring and Control of Machining Process by Data Mining and Pattern Recognition," Proceedings of the IEEE International Conference on Complex, Intelligent and Software Intensive Systems, (CISIS), pp. 106-113, July 2012. [3] Y. Nagaya, H. Endo, T. Uchimoto, T. Takagi, H. J. Jung, and L. Bartolomeo, “Crack shape reconstruction using ECT camera system,” in Proc. SICE Annu. Conf., Sapporo, Japan, , Vol. 3, pp. 2760–2763, 2004. [4] Z. Chen, K. Aoto, and K. Miya, “Reconstruction of cracks with physical closure from signals of eddy current testing,” IEEE Trans. Magn., vol. 36, pt. 1, no. 4, pp. 1018–1022, Jul. 2000. [5] G. Preda, B. Cranganu-Cretu, F. I. Hantila, O. Mihalache, Z. Chen, and K. Miya, “Nonlinear FEM-BEM formulation and model-free inversion procedure for reconstruction of cracks using pulse eddy currents,” IEEE Trans. Magn., Vol. 38, n. 2, pp. 1241–1244, Mar. 2002. [6] E. Todorov, “Measurement of electromagnetic properties of heat exchanger tubes with model-based eddy current technique,” Elsevier NDT&E International, Vol.48, pp.70–76, 2012. [7] C.H. Chen (Ed.), Signal Processing and Pattern Recognition in Nondestructive Evaluation of Materials. Springer, Berlin. Proceedings of the NATO Advanced Research Workshop on Signal Processing and Pattern Recognition in Nondestructive Evaluation of Materials, held at the Manoir St-Castin, Lac Beauport, Quebec, Canada, August 19-22, 1987. [8] I.V. Kuts, A.G. Protasov, I. Iu. Lysenko , O.L. Dugin. "Pulsed Eddy Current Non-Destructive Testing of the Coating Thickness," 11th European Conference on Non-Destructive Testing (ECNDT 2014), Prague, Czech Republic, October 6-10, 2014. [9] G.D’Angelo, S. Rampone, “Diagnosis of aerospace structure defects by a HPC implemented soft computing algorithm,” IEEE International Workshop on Metrology for Aerospace, Benevento, Italy, May 29-30, 2014. [10]G.D’Angelo, S. Rampone, “Towards a HPC-oriented parallel implementation of a learning algorithm for bioinformatics applications,” BMC Bioinformatics 2014, 15 (Suppl 5):S2. [11] V.S. Eremenko, О. Gileva, "Application of linear recognition methods in problems of nondestructive testing of composite materials," International scientific conference on Electromagnetic and acoustic methods of nondestructive testing of materials and products, LЕОТЕSТ2009. [12] R.P. Pein, J. Lu, J. Birger Stav, Q. Xu, M. Uran, L. Mráz, "Automatic digital x-ray control with contentbased image retrieval techniques," 10th International Conference of the Slovenian Society for Non Destructive Testing - Application of Contemporary Non-Destructive Testing in Engineering -, Ljubljana, Slovenia, pp.167-176, September 1-3, 2009. [13] N. Nacereddine, D. Ziou, L. Hamami, "Fusion-based shape descriptor for weld defect radiographic image retrieval," International Journal of Advanced Manufacturing Technology, Vol. 68, n. 9-12, pp 2815-2832, 2013. [14] G.D’Angelo, S. Rampone, “Discovering Hidden Information in Big Data: a Soft computing Approach,” VII International Meeting in Dynamics of Socio-Economic System, Dyses 2014, Seville, Spain, September 16-18, 2014. [15] G.D’Angelo, S. Rampone, “A proposal for advanced services and data processing aiming the territorial intelligence development,” World Scientific Proceedings First International Workshop “Global Sustainability Inside and Outside the Territory", Benevento, 14 February 2014. [16] U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth, “From Data Mining to Knowledge Discovery in Databases.” Advances in knowledge discovery and data mining, American Association for Artificial Intelligence, Menlo Park, CA, 1996. [17]R.J. Kauffman, A.A.Techatassanasoontorn, "The Global Diffusion Patterns of Successive Technology Generations: Modeling Analog and Digital Wireless Phone Growth," IEEE International Conference on Information and Communication Technology and Development, ICTD2006 4085540, pp. 266-276, 2006. [18] A. Xia Liu, Y. Wang, X. Chen, X. Jiang, "Understanding the diffusion of mobile digital content: a growth curve modelling approach," ACM Journal Information Systems and e-Business Management archive, Vol.12, n. 2, pp. 239-258, 2014. [19] R. Datta, D. Joshi, J. Li, J.Z. Wang, "Image retrieval: Ideas, influences, and trends of the new age," Journal ACM Computing Surveys (CSUR), Vol. 40, n. 2, 2008. [20] Sarfraz and M. Ridha “Content-Based Image Retrieval Using Multiple Shape Descriptors”, IEEE/ACS International Conference On Computer Systems and Applications, pp. 730-737, 2007. [21]V.N. Gudivada, V.V. Raghavan, “Content-based image retrieval systems”, IEEE Computer, Vol. 28, N. 9, 18-22, Sep. 1995. [22] M. Swain and D. Ballard, “Color indexing,” Intl. Journal Computer Vision, Vol. 7, n.1, pp. 11-32, 1991. [23] R. M. Haralick, “Texture features for image classification,” IEEE Trans. on Sys. Man and Cyb., 1990. [24] B.M. Mehtre, “Shape measures for content based image retrieval: a comparison”, Information Proc. Management, Vol. 33, n.3, pp.319-337, 1997. [25] M. Kokare, B.N. Chatterji, P.K. Biswas, "Comparison of similarity metrics for texture image retrieval," IEEE Region 10 Annual International Conference, Proceedings/TENCON, Vol. 3, pp. 571-575, 2003. [26] L.R. Long, S. Antani, T.M. Deserno, and G.R. Thoma, "Content-Based Image Retrieval in Medicine: Retrospective assessment, state of the art, and future directions," Int J Healthcare Inf Syst Inform., Vol.4, n.1, pp 1-16, 2009. [27] Dingyuan Xia, Pian Fu, Chaobing Huang, Yu Wang, "Trend of ContentBased Image Retrieval on the Internet," IEEE Fifth International Conference on Image and Graphics, pp. 733 - 738, 2009. [28] Chun-Rong Su, Jiann-Jone Chen, "Content-Based Image Retrieval On reconfigurable Peer-to-Peer networks," IEEE 14th International Workshop on Multimedia Signal Processing (MMSP), pp- 343 - 348, 2012. [29] B.M. Mehtre, M.S. Kankanhalli, W. F. Lee, "Shape measures for content based image retrieval: A comparison," Information Processing & Management, Vol. 33, n. 3, pp. 319-337, 1997. [30] G. Doretto and Y. Yao, "Region moments: fast invariant descriptors for detecting small image structures," IEEE Conference on Computer Vision and Pattern Recognition (CVPR) San Francisco, CA, pp. 30193026, June 2010. [31] H. Kim, J. Kim, "Region-based shape descriptor invariant to rotation, scale and translation," Signal Processing: Image Communication, Vol.16, pp 87–93, 2000. [32] A. Amanatiadis, V.G. Kaburlasos, A. Gasteratos, S.E. Papadakis, "Evaluation of shape descriptors for shape-based image retrieval," Image Processing, IET, Vol.5, n.5, pp.493-499, August 2011. [33] M. Yang, K. Kpalma, J. Ronsin, "A Survey of Shape Feature Extraction Techniques," [https://hal.archives-ouvertes.fr/hal-00446037/document], HAL archives-ouvertes.fr [34] J. García-Martín, J. Gómez-Gil and E. Vázquez-Sánchez, "NonDestructive Techniques Based on Eddy Current Testing," Sensors, vol.11, pp. 2525-2565. [35] EC data - manual inspection. [http://measure.feld.cvut.cz/usr/staff/smid/datasets/]. Department of Measurement Czech Technical University in Prague, Faculty of Electrical Engineering. Online, last access January 8, 2015. [36] H.J. Jiawei, M. Kamber, J. Pei, Data Mining: Concepts and Techniques: Morgan Kaufmann Publishers, 2006. [37] I. Witten, E. Frank: Data Mining: Practical Machine Learning Tools and Techniques, Morgan Kaufmann, 2005. [38] G. Holmes, A. Donkin, I.H. Witten, "WEKA: a machine learning workbench," IEEE Conference on Intelligent Information Systems, Proceedings of the 1994 Second Australian and New Zealand, 1994.
5cs.CE
Maximum Matching in Two, Three, and a Few More Passes Over Graph Streams arXiv:1702.02559v3 [cs.DS] 21 Apr 2017 Sagar Kale1 and Sumedh Tirodkar2 1 Dartmouth 2 TIFR, College. Email: sag (at) cs.dartmouth.edu Mumbai. Email: sumedh.tirodkar (at) tifr.res.in Abstract We consider the maximum matching problem in the semi-streaming model formalized by Feigenbaum, Kannan, McGregor, Suri, and Zhang [13] that is inspired by giant graphs of today. As our main result, we give a two-pass (1/2 + 1/16)-approximation algorithm for triangle-free graphs and a two-pass (1/2 + 1/32)-approximation algorithm for general graphs; these improve the approximation ratios of 1/2 + 1/52 for bipartite graphs and 1/2 + 1/140 for general graphs by Konrad, Magniez, and Mathieu [20]. In three passes, we are able to achieve approximation ratios of 1/2 + 1/10 for triangle-free graphs and 1/2 + 1/19.753 for general graphs. We also give a multi-pass algorithm where we bound the number of passes precisely—we give a (2/3 − ε)-approximation algorithm that uses 2/(3ε) passes for triangle-free graphs and 4/(3ε) passes for general graphs. Our algorithms are simple and combinatorial, use O(n log n) space, and (can be implemented to) have O(1) update time per edge. For general graphs, our multi-pass algorithm improves the best known deterministic algorithms in terms of the number of passes: • Ahn and Guha [1] give a (2/3 − ε)-approximation algorithm that uses O(log(1/ε)/ε2 ) passes, whereas our (2/3 − ε)-approximation algorithm uses 4/(3ε) passes; • they also give a (1 − ε)-approximation algorithm that uses O(log n · poly(1/ε)) passes, where n is the number of vertices of the input graph; although our algorithm is (2/3 − ε)approximation, our number of passes do not depend on n. Earlier multi-pass algorithms either have a large constant inside big-O notation for the number of passes [9] or the constant cannot be determined due to the involved analysis [22, 1], so our multi-pass algorithm should use much fewer passes for approximation ratios bounded slightly below 2/3. 1 Introduction Maximum matching is a well-studied problem in a variety of computational models. We consider it in the semi-streaming model formalized by Feigenbaum, Kannan, McGregor, Suri, and Zhang [13] that is inspired by generation of ginormous graphs in recent times. A graph stream is an (adversarial) sequence of the edges of a graph, and a semi-streaming algorithm must access the edges in the given order and use O(n polylog n) space only, where n is the number of vertices; note that a matching can have size Ω(n), so Ω(n log n) space is necessary. The number of times an algorithm goes over a stream of edges is called the number of passes. A trivial (1/2)approximation algorithm that can be easily implemented as a one-pass semi-streaming algorithm is to output a maximal matching. Since the formalization of the semi-streaming model more than a decade ago, the problem of finding a better than (1/2)-approximation algorithm or proving that one cannot do better has baffled researchers [21]. In a step towards resolving this, Goel, Kapralov, and Khanna [14] proved that for any ε > 0, a one-pass semi-streaming (2/3 + ε)-approximation algorithm does not exist; Kapralov [16], building on those techniques, showed non-existence of one-pass semi-streaming (1 − 1/e + ε)-approximation algorithms for any ε > 0. A natural next question is: Can we do better in, say, two passes or three passes? In answering that, Konrad, Magniez, and Mathieu [20] gave three-pass and two-pass algorithms that output matchings that are better than (1/2)-approximate. In this work, we give algorithms that improve their approximation ratios for two-pass and three-pass algorithms. We also give a multi-pass algorithm that does better than the best known multi-pass algorithms for at least initial few passes. We are able to bound the number of passes precisely: we give a (2/3 − ε)-approximation algorithm that uses 2/(3ε) passes for triangle-free graphs and 4/(3ε) passes for general graphs. Earlier works either have a large constant inside the big-O notation for the number of passes [9] or the constant cannot be determined due to the involved analysis [22, 1]. For example, the (1 − ε)-approximation algorithm by Eggert et al. [9] potentially uses 288/ε5 passes, and for the (1 − ε)-approximation algorithms by McGregor [22] and Ahn and Guha [1], the constants inside the big-O bound cannot be determined due to the involved analysis. The (2/3 − ε)-approximation algorithm by Feigenbaum et al. [13] uses O(log(1/ε)/ε) passes, which is O(log(1/ε)) factor larger than the number of passes we use to get the same approximation ratio. Our algorithms are simple and combinatorial, use O(n log n) space, and (can be implemented to) have O(1) update time per edge. We also give an explicit and tight analysis of the three-pass algorithm by Konrad et al. [20] that is reminiscent of Feigenbaum et al.’s [13] multi-pass algorithm. Technical overview: If we can find a matching M such that there are no augmenting paths of length 3 in M ∪ M ∗ , where M ∗ is a maximum matching, then M is (2/3)-approximate, i.e., (1/2 + 1/6)-approximate. This is because, in each connected component of M ∪ M ∗ , the ratio of Medges to M ∗ -edges is at least 2/3. This is the basis for the (2/3 − ε)-approximation algorithm by Feigenbaum et al. [13] that uses O(log(1/ε)/ε) passes. The same idea is used by Konrad et al. [20] in the analysis of their two-pass algorithms. In the first pass, they find a maximal matching M0 and some subset of support edges, say S. If M0 is so bad that M0 ∪ M ∗ is almost entirely made up of augmenting paths of length 3 (i.e., | M0 | ≈ | M ∗ |/2), then by the end of the second pass, they manage to augment (using length-3 augmentations) a constant fraction of M0 using S and a fresh access to the edges, resulting in a better than (1/2)-approximation. On the other hand, if M0 is not so bad, then they already have a good matching. One limitation this idea faces is that a fraction of the edges in S may become useless for an augmentation if both its endpoints get matched in M0 by the end of the first pass. Our main result is a two-pass algorithm (described in Section 5) that differs in two ways from the former approach. Firstly, in the first pass, we only find a maximal 1 matching M0 so that in the second pass, where we maintain a set S of support edges, S would not contain “useless” edges. Secondly, any augmentation in our algorithm happens immediately when an edge arrives if it forms an augmenting path of length 3 with edges in M0 and S. Our results: In light of the discussion so far, one way to evaluate an algorithm is how much advantage it gains over the (1/2)-approximate maximal matching found in the first pass. We summarize our two-pass and three-pass results in Table 1 and multi-pass results in Table 2. We stress that we are able to bound the number of passes precisely, without big-O notation. For general graphs, our multi-pass algorithm improves the best known deterministic algorithms in terms of number of passes—see the third multi-row of Table 2. We note that our multi-pass algorithm is not just a repetition of the second pass of our two-pass algorithm. Such a repetition will give an asymptotically worse number of passes (see, for example, the multi-pass algorithm due to Feigenbaum et al. [13]; the first row of Table 2). We carefully choose the parameters for each pass to get the required number of passes. Also note that Table 1 shows advantages over a maximal matching—an algorithm is said to have advantage α if it is a (1/2 + α)-approximation algorithm (because a maximal matching is (1/2)-approximate). Table 1: Advantages over a maximal matching—advantage α means (1/2 + α)-approximation. Problem Bipartite two-pass Bipartite three-pass Triangle-free two-pass Triangle-free three-pass General two-pass General three-pass Previous work Esfandiari et al. [11] Esfandiari et al. [11] Advantage 1/12 1/9.52 Not considered separately Konrad et al. [20] 1/140 Not considered separately Advantage in this work Not considered separately 1/16 (in Section 5) 1/10 (in Section A) 1/32 (in Section 5) 1/19.753 (in Section B) Table 2: Multi-pass algorithms—see Section 6. Graph Bipartite Triangle free General Results Feigenbaum et al. [13] Eggert et al.[9] Ahn and Guha [1] This work (in Section 6) McGregor [22] randomized Ahn and Guha [1] Ahn and Guha [1] This work (in Section 6) Approx 2/3 − ε 1−ε 1−ε 2/3 − ε 1−ε 2/3 − ε 1−ε 2/3 − ε # Passes O(log(1/ε)/ε) 288/ε5 O(log log(1/ε)/ε2 ) 2/(3ε) O((1/ε)1/ε ) O(log(1/ε)/ε2 ) O(log n · poly(1/ε)) 4/(3ε) Note of independent work The work of Esfandiari et al. [11] who claim better approximation ratios for bipartite graphs in two passes and three passes is independent and almost concurrent. Our work differs in several aspects. We consider triangle-free graphs (superset of bipartite graphs) and general graphs, and we additionally consider multi-pass algorithms. Also, their algorithm has a 2 √ post-processing step that uses time O( n · | E|), whereas our algorithms can be implemented to have O(1) update time per edge. One further detail about this appears in Section D. 1.1 Related Work Karp, Vazirani, and Vazirani [18] gave the celebrated (1 − 1/e)-competitive randomized online algorithm for bipartite graphs in the vertex arrival setting. Goel et al. [14] gave the first one-pass deterministic algorithm with the same approximation ratio, i.e., 1 − 1/e, in the semi-streaming model in the vertex arrival setting. For the rest of this section, results involving ε hold for any ε > 0. As mentioned earlier, Goel, Kapralov, and Khanna [14] proved nonexistence of one-pass (2/3 + ε)-approximation semi-streaming algorithms, which was extended to (1 − 1/e + ε)-approximation algorithms by Kapralov [16]. On the algorithms side, nothing better than outputting a maximal matching, which is (1/2)-approximate, is known. Closing this gap is considered an outstanding open problem in the streaming community [21]. On the multi-pass front, in the semi-streaming model, Feigenbaum et al. [13] gave a (2/3 − ε)-approximation algorithm for bipartite graphs that uses O(log(1/ε)/ε) passes; McGregor [22] improved it to give a (1 − ε)-approximation algorithm for general graphs that uses O((1/ε)1/ε ) passes. For bipartite graphs, this was again improved by Eggert et al. [9] who gave a (1 − ε)approximation O((1/ε)5 )-pass algorithm. Ahn and Guha [1] gave a linear-programming based (1 − ε)-approximation O(log log(1/ε)/ε2 )-pass algorithm for bipartite graphs. For general graphs, their (1 − ε)-approximation algorithm uses number of passes proportional to log n, so it is worse than that of McGregor [22]. For the problem of one-pass weighted matching, there is a line of work starting with Feigenbaum et al. [13] giving a 6-approximation semi-streaming algorithm. Subsequent results improved this approximation ratio: see McGregor [22], Zelke [24], Epstein et al. [10], Crouch and Stubbs [8], Grigorescu et al.[15], and most recently in a breakthrough, giving a (2 + ε)-approximation semistreaming algorithm, Paz and Schwartzman [23]. The multi-pass version of the problem was considered first by McGregor [22], then by Ahn and Guha [1]. Chakrabarti and Kale [5] and Chekuri et al. [6] consider a more general version of the matching problem where a submodular function is defined on the edges of the input graph. The problem of estimating the size of a maximum matching (instead of outputting the actual matching) has also been considered. We mention Kapralov et al. [17], Esfandiari et al. [12], Bury and Schwiegelshohn [4], and Assadi et al. [2]. In the dynamic streams, edges of the input graph can be removed as well. The works of Konrad [19], Assadi et al. [3], and Chitnis et al. [7] consider the maximum matching problem in dynamic streams. 1.2 Organization of the Paper After setting up notation in Section 2, we give a tight analysis of the three-pass algorithm for bipartite graphs by Konrad et al. [20] in Section 3. In Section 4, we see our simple two-pass algorithm for triangle-free graphs. Then in Section 5, we see our main result—the improved two-pass algorithm, and then we see the multi-pass algorithm in Section 6. The results that are not covered in the main sections are covered in the appendix. 3 2 Preliminaries We work on graph streams. The input is a sequence of edges (stream) of a graph G = (V, E), where V is the set of vertices and E is the set of edges; a bipartite graph is denoted as G = ( A, B, E). A streaming algorithm may go over the stream a few times (multi-pass) and use space O(n polylog n), where n = |V |. In this paper, we give algorithms that make two, three, or a few more passes over the input graph stream. A matching M is a subset of edges such that each vertex has at most one edge in M incident to it. The maximum cardinality matching problem, or maximum matching, for short, is to find a largest matching in the given graph. Our goal is to design streaming algorithms for maximum matching. For a subset F of edges and a subset U of vertices, we denote by U ( F ) ⊆ U the set of vertices in U that have an edge in F incident on them. Conversely, we denote by F (U ) ⊆ F the set of edges in F that have an endpoint in U. For a subset F of edges and a vertex v ∈ V ( F ), we denote by NF (v) the set of v’s neighbors in the graph (V ( F ), F ), and we define degF (v) := | NF (v)|. In the first pass, our algorithms compute a maximal matching which we denote by M0 . We use M ∗ to indicate a matching of maximum cardinality. Assume that M0 and M ∗ are given. For i ∈ {3, 5, 7, . . . }, a connected component of M0 ∪ M ∗ that is a path of length i is called an i-augmenting path (nonaugmenting otherwise). We say that an edge in M0 is 3-augmentable if it belongs to a 3-augmenting path, otherwise we say that it is non-3-augmentable. Lemma 2.1 (Lemma 1 in [20]). Let α > 0, M0 be a maximal matching in G, and M ∗ be a maximum matching in G such that | M0 | 6 (1/2 + α)| M ∗ |. Then the number of 3-augmentable edges in M0 is at least (1/2 − 3α)| M ∗ |, and the number of non-3-augmentable edges in M0 is at most 4α| M ∗ |. Proof. Let the number of 3-augmentable edges in M0 be k. For each 3-augmentable edge in M0 , there are two edges in M ∗ incident on it. Also, each non-3-augmentable edge in M0 lies in a connected component of M0 ∪ M ∗ in which the ratio of the number of M ∗ -edges to the number of M0 -edges is at most 3/2. Hence, 3 | M ∗ | 6 2k + (| M0 | − k) 2    1 3 ∗ + α |M | − k 6 2k + 2 2   1 3 3 = k+ + α | M∗ | , 2 4 2 since there are | M0 | − k non-3-augmentable edges , because | M0 | 6 (1/2 + α)| M ∗ | , which, after simplification, gives k > (1/2 − 3α)| M ∗ |. And the number of non-3-augmentable edges in M0 is | M0 | − k 6 | M0 | − (1/2 − 3α)| M ∗ | 6 (1/2 + α − 1/2 + 3α)| M ∗ | = 4α| M ∗ |. We make the following simple, yet crucial, observation. Observation 2.1. Let M0 be a maximal matching. Then V ( M0 ) is a vertex cover, and there is no edge between any two vertices in V \ V ( M0 ). Therefore, even if the input graph is not a bipartite graph, the set of edges incident on V \ V ( M0 ), i.e., E(V \ V ( M0 )) give rise to a bipartite graph with bipartition (V \ V ( M0 ), V ( M0 )). For all the algorithms in this paper, it can be verified that their space complexity is O(n log n) and update time per edge is O(1). We also ignore floors and ceilings for the sake of exposition. 4 B \ B ( M0 ) A ( M0 ) B ( M0 ) A \ A ( M0 ) M0 M∗ MA MB Figure 1: Example: state of variables in an execution of Algorithm 1. 3 Analyzing the Three Pass Algorithm for Bipartite Graphs We analyze the three-pass algorithm for bipartite graphs given by Konrad et al. [20], i.e., Algorithm 1 by considering the distribution of lengths of augmenting paths. We also give a tight example. Algorithm 1 Three-pass algorithm for bipartite graphs due to Konrad et al. [20] 1: In the first pass, find a maximal matching M0 . 2: In the second pass, find a maximal matching • M A in F2 := {ab : a ∈ A( M0 ), b ∈ B \ B( M0 )} (see Figure 1). 3: In the third pass, find a maximal matching • MB in F3 := {ab : a ∈ A \ A( M0 ) and ∃ a′ ∈ A( M A ) such that a′ b ∈ M0 }. 4: Augment M0 using edges in M A and MB and return the resulting matching M. Theorem 3.1. Algorithm 1 is a three-pass, semi-streaming, (1/2 + 1/10)-approximation algorithm for maximum matching in bipartite graphs. Proof. Without loss of generality, let M ∗ be a maximum matching such that all nonaugmenting connected components of M0 ∪ M ∗ are single edges. For i = {3, 5, 7, . . . }, let ki denote the number of i-augmenting paths in M0 ∪ M ∗ , and let k = | M0 ∩ M ∗ |. Then | M0 | = k + ∑ i i−1 ki 2 and | M∗ | = k + ∑ i i+1 ki . 2 (1) Consider an i-augmenting path b1 a1 b2 a2 b3 · · · b(i+1)/2 a(i+1)/2 in M0 ∪ M ∗ , where for each j, we have a j ∈ A and b j ∈ B. We call the vertex a(i−1)/2 a good vertex, because an edge in M A incident to a(i−1)/2 can potentially be augmented using the edge b(i+1)/2 a(i+1)/2 . To elaborate, consider the set of all edges in M A incident on good vertices; call it M ′A . Consider the set of edges of the type b(i+1)/2 a(i+1)/2 from each i-augmenting path; call it M F . Note that M F is a matching. Then we can augment M0 using M ′A and M F by as much as | M ′A |. There is a matching of size ∑i ki in F2 formed by edges of the type b1 a1 from each i-augmenting path. Since M A is maximal in F2 , we have | M A | > (∑i ki )/2. Now, the number of good vertices is ∑i ki ; therefore, the number of bad (i.e., not good) vertices is | M0 | − ∑i ki . So the number of edges 5 B \ B ( M0 ) A ( M0 ) B ( M0 ) A \ A ( M0 ) M0 M∗ MA Figure 2: Tight example for Algorithm 1: M A has only one edge that lands on a bad vertex and cannot be augmented in the third pass. So | M | = | M0 | = 3 and | M ∗ | = 5. in M A incident on good vertices (see Figure 2) ∑ k | M ′A | > i i − 2 | M0 | − ∑ k i i ! = 3 k i − | M0 | . 2∑ i Let BG := {b ∈ B : ∃ a ∈ A( M ′A ) such that ab ∈ M0 }. Let M ′F ⊆ M F be defined as M ′F := {ba ∈ M F : b ∈ BG }. Then we know that | M ′F | = | M ′A | and M ′F ⊆ M F ⊆ F3 . Since we select a maximal matching in F3 in the third pass, | MB | > | M ′A | | M ′F | 3 | M0 | = = ∑ ki − . 2 2 4 i 2 (2) So the output size | M | = | M0 | + | M B | | M0 | 3 > | M0 | + ∑ k i − 4 i 2 = by (1) and (2), | M0 | 3 + (| M ∗ | − | M0 |) 2 4 by (1), ∑ ki = | M ∗ | − | M0 | , i i.e., | M | > 3| M ∗ |/4 − | M0 |/4, but we also have | M | > | M0 |, hence   3 ∗ 1 | M | > max | M0 |, | M | − | M0 | . 4 4 So the bound is minimized when | M0 | = 3| M ∗ |/4 − | M0 |/4 = 3| M ∗ |/5 = (1/2 + 1/10)| M ∗ |. As we can see in the proof above, the worst case happens when | M | = | M0 | = 3| M ∗ |/5. Setting k3 = k5 > 1, k = 0, and ki = 0 for i > 5 gives us the tight example shown in Figure 2. 4 A Simple Two Pass Algorithm for Triangle Free Graphs Before seeing our main result, we see a simple two pass algorithm for triangle-free graphs. The function S EMI () in Algorithm 2 greedily computes a subset of edges such that each vertex in X has degree at most one and each vertex in Y has degree at most λ; we call such a subset a (λ, X, Y )semi-matching (Konrad et al. [20] call this a λ-bounded semi-matching). In Algorithm 2, we find a maximal matching M0 in the first pass, and, in the second pass, we find a (λ, V ( M0 ), V \ V ( M0 ))semi-matching S. After the second pass, we greedily augment edges in M0 one by one using edges in S. 6 Algorithm 2 Two-pass algorithm for triangle-free graphs 1: In the first pass: M0 ← maximal matching 2: In the second pass: S ← S EMI ( λ, V ( M0 ), V \ V ( M0 )) (see Figure 3). 3: After the second pass, augment M0 greedily using edges in S to get M; output M. 4: function S EMI (λ, X, Y) ⊲ based on Algorithm 7 in Konrad et al. [20] 5: S←∅ 6: foreach edge xy such that x ∈ X, y ∈ Y do 7: if degS ( x) = 0 and degS (y) 6 λ − 1 then 8: S ← S ∪ { xy} u6 v6 u5 v5 u4 v4 u3 v3 M0 u2 v2 u1 v1 S Figure 3: Example showing M0 and S at the end of the second pass of Algorithm 2 with λ = 2. When we greedily augment M0 after the second pass, we may choose to augment u5 v5 and lose two possible augmentations of edges u4 v4 and u6 v6 . Theorem 4.1. Algorithm 2 is a two-pass, semi-streaming, (1/2 + 1/20)-approximation algorithm for maximum matching in triangle-free graphs. Proof. As in the proof of Theorem 3.1, let M ∗ be a maximum matching such that all nonaugmenting connected components of M0 ∪ M ∗ are single edges. For i = {3, 5, 7, . . . }, let ki denote the number of i-augmenting paths in M0 ∪ M ∗ , and let k denote the number of edges in M ∗ ∩ M0 . Now, we define good vertices. Consider an i-augmenting path x1 y1 x2 y2 x3 · · · x(i+1)/2 y(i+1)/2 in M0 ∪ M ∗ . We call the vertices y1 ∈ V ( M0 ) and x(i+1)/2 ∈ V ( M0 ) good vertices, because the edges x1 y1 ∈ M ∗ and x(i+1)/2 y(i+1)/2 ∈ M ∗ can potentially be added to S by our algorithm. Denote by VG the set of good vertices and by VB := V ( M0 ) \ VG the set of bad vertices. Then |VG | = 2 ∑ i ki . Note that VG ∩ VB = ∅ and VG ∪ VB = V ( M0 ) by definition. Let VNC := VG \ V (S) be the set of good vertices not covered by S. An edge uv ∈ M ∗ with u ∈ V \ V ( M0 ) and v ∈ VNC was not added to S, because degS (u) = λ. Hence λ|VNC | 6 |V ( M0 )| − |VNC | i.e., 7 |VNC | 6 2 | M0 | , λ+1 (3) because at most |V ( M0 )| − |VNC | vertices in V ( M0 ) are covered by S. Now, |V ( M0 ) \ V (S)| = |VG \ V (S)| + |VB \ V (S)| ∵ VG ∩ VB = ∅ and VG ∪ VB = V ( M0 ), 6 |VNC | + |VB | ∵ VNC = VG \ V (S) and |VB \ V (S)| 6 |VB |, 2 6 | M0 | + |V ( M0 )| − |VG | by (3) and the fact |VB | = |V ( M0 )| − |VG |, λ+1 2 | M0 | + |V ( M0 )| − 2 ∑ ki because |VG | = 2 ∑ ki . = λ+1 i i Using |V ( M0 )| = |V ( M0 ) \ V (S)| + |V ( M0 ) ∩ V (S)| and the above, we get 2 |V ( M0 ) ∩ V (S)| > |V ( M0 )| − | M0 | + |V ( M0 )| − 2 ∑ ki λ+1 i ! 1 = 2 ∑ ki − | M0 | . λ+1 i ! (4) We observe that at most | M0 | vertices in V ( M0 ) (one endpoint of each edge) can be covered by S without having both endpoints of an edge in M0 covered. Hence, at least |V ( M0 ) ∩ V (S)| − | M0 | edges in M0 have both their endpoints covered by S, which, by (4), is at least ! 1 λ+3 2 ∑ ki − | M0 | − | M0 | = 2 ∑ k i − | M0 | . (5) λ+1 λ+1 i i After the second pass, when we greedily augment an edge from the above edges, i.e., edges whose both endpoints are covered by S, we may potentially lose 2(λ − 1) other augmentations (see Figure 3). To see this, consider uv ∈ M0 such that u, v ∈ V (S) and au ∈ S and vb ∈ S. The graph is triangle free, so we know that a 6= b, and we can augment M0 using the 3-augmenting path auvb; but we may lose at most λ − 1 edges incident to a in S and at most λ − 1 edges incident to b in S. Therefore the number of augmentations c we get after the second pass is at least 1/(2λ − 1) times the right hand side of (5), i.e., c> 2 λ+3 | M0 | . ki − ∑ 2λ − 1 i (2λ − 1)(λ + 1) So the output size | M | = | M0 | + c, and using the above bound on c and simplifying we get: | M| > 2 2( λ 2 − 2) | M0 | ; ki + ∑ 2λ − 1 i (2λ − 1)(λ + 1) substituting ∑i ki = | M ∗ | − | M0 |, by (1), in the above, | M| > 2( λ 2 − λ − 3) 2 | M∗ | + | M0 | . 2λ − 1 (2λ − 1)(λ + 1) Using λ = 3 and the fact that M0 is 2-approximate, we get 2 3 2 3 11 | M | > | M ∗ | + | M0 | > | M ∗ | + | M ∗ | = | M ∗ | = 5 10 5 20 20 8  1 1 + 2 20  | M∗ | . 5 Improved Two Pass Algorithm We present our main result that is a two pass algorithm in this section. In the first pass, we find a maximal matching M0 . In the second pass, we maintain a set S of support edges xy, such that x ∈ V \ V ( M0 ), y ∈ V ( M0 ), and degS (y) 6 λM and degS ( x) 6 λU , where λM > 1 and λU > 1 are parameters denoting maximum degree allowed in S for matched and unmatched vertices (with respect to M0 ), respectively. Whenever a new edge forms a 3-augmenting path with an edge in M0 and an edge in S, we augment. Algorithm 3 gives a formal description. Algorithm 3 Improved two-pass algorithm: input graph G 1: In the first pass, find a maximal matching M0 . 2: if G is triangle-free then 3: Return I MPROVE -M ATCHING( M0 , 2, 1) 4: else 5: Return I MPROVE -M ATCHING( M0 , 4, 2) 6: function I MPROVE -M ATCHING(M0 , λU , λM ) 7: M ← M0 , S ← ∅, I ← ∅ and IB ← ∅ 8: foreach edge xy in the stream do 9: if x or y ∈ I ∪ IB then 10: Continue, i.e., ignore xy. 11: else if x ∈ V ( M0 ) and y ∈ V ( M0 ) then 12: Continue, i.e., ignore xy. 13: else if there exist v and b such that yv ∈ M0 and vb ∈ S then 14: M ← M \ {yv} ∪ { xy, vb} ⊲ a 3-augmentation Let Ix ← {u x , vx : xu x ∈ S and u x vx ∈ M0 }. Let Ib ← {ub , vb : ub vb ∈ M0 and vb b ∈ S}. 15: Then I ← I ∪ { x, y, v, b} and IB ← IB ∪ Ix ∪ Ib . 16: else Without loss of generality, assume that x ∈ V \ V ( M0 ) and y ∈ V ( M0 ). 17: if degS ( x) < λU and degS (y) < λM then ⊲ See Figure 4. 18: S ← S ∪ { xy} ⊲ Note: Once an edge is added to S, it is never removed from it. 19: Return M. Setting up a charging scheme to lower bound the number of augmentations We first lay the groundwork and give a charging scheme. Observation 5.1. For general graphs (that are possibly not triangle-free), we need to set λM > 2. To see why, suppose λM = 1. Let uv be a 3-augmentable edge in M0 . Then, for the edge uv, we might end up storing the edges ub and vb in S, and the edge uv would not get augmented. If λM > 2, and we store at least λM edges incident to u, then an edge incident to v will not form a triangle with at least one of those and uv would get augmented. So, for general graphs, we need to set λM > 2. Let | M0 | = (1/2 + α)| M ∗ |. For a 3-augmentable edge uv ∈ M0 , let auvb be the 3-augmenting path such that au, vb ∈ M ∗ . Without loss of generality, assume that au arrived before vb. Then we make the following observation. 9 a6 a4 u6 v6 u5 v5 u4 v4 u3 v3 u2 v2 u1 v1 b6 b4 M0 some M ∗ edges some S edges Figure 4: Example showing M0 and some of the edges in M ∗ and S during the second pass of Algorithm 3 for triangle-free graphs with λU = 2 and λM = 1. At most one of ui and vi can have positive degree in S, because we would rather augment ui vi instead of adding the latter edge to S. By our convention, a4 u4 arrived before v4 b4 , and a6 u6 arrived before v6 b6 . Since a4 u4 was not added to S, we have degS ( a4 ) = λU (S edges incident to a4 are not shown). Observation 5.2. When au arrived, it may not be added to S for one of the following reasons: • The vertex a was already matched. • There were λM edges incident to u in S. • There were λU edges incident to a in S. We call some edges in M0 good, some partially good, and some bad. An edge is good if it got augmented. An edge uv ∈ M0 is bad if it is 3-augmentable, not good, and vertex a or b had λU edges incident to them in S when edge au or vb arrived. An edge uv ∈ M0 is partially good if it is 3-augmentable, but neither good nor bad (“partially” good because, as we will see later, we can hold some good edge u′ v′ ∈ M0 responsible for uv not getting augmented). Note that all 3-augmentable edges get some label according to our labeling. We require the following Lemma to describe the charging scheme. Lemma 5.1. Suppose au was not added to S because there were already λM edges incident to u in S. If, later, uv did not get augmented when vb arrived, then • b was already matched via augmenting path a′′ u′′ v′′ b, or • there exists a′ u ∈ S and u′ v′ ∈ M0 such that a′ was matched via augmenting path a′ u′ v′ b′ . Proof. When au arrived, | NS (u)| > λM . If b was unmatched when vb arrived, then some a′ ∈ NS (u) \ {b} must have been matched, otherwise we would have augmented uv. Now for trianglefree graphs b ∈ / NS (u), so | NS (u) \ {b}| = | NS (u)| > 1, and for general graphs, by Observation 5.1, λM > 2, so | NS (u) \ {b}| > λM − 1 > 1. 10 a u v b v′ y partially good a′′ u′ M0 some M ∗ edges good u′′ bad some S edges v′′ b′′ a stream edge Figure 5: Example showing a good edge, a bad edge, and a partially good edge. We use parameters λU = 2 and λM = 1, so we are in the triangle-free case. The edge u′ v′ is not 3-augmentable but was augmented using a′′ u′ v′ y, so u′ v′ is a good edge. The edge u′′ v′′ is a 3-augmentable edge that was not augmented and when a′′ u′′ arrived, degS ( a′′ ) = 2, so u′′ v′′ is a bad edge. For uv, we did not take au in S, because degS (u) = 1, so uv is a partially good edge, and we can charge uv to u′ v′ using Lemma 5.1. Charging Scheme. As alluded to earlier, we charge a partially good edge to some good edge. Recall that for a 3-augmentable edge uv ∈ M0 , we denote by au, vb ∈ M ∗ the edges that form the 3-augmenting path with uv such that au arrived before vb. We use Observation 5.2 and consider the following cases. See Figure 5. • Suppose au was not added to S because a was already matched. Then, let u′ v′ ∈ M0 was augmented using au′ v′ b′ . If degS ( a) 6 λU − 1, then we charge uv to u′ v′ . Otherwise, uv is bad. • Suppose au was not added to S because degS (u) = λM . Then we use Lemma 5.1. We either charge uv to u′ v′ , or if degS (b) 6 λU − 1, then we charge uv to u′′ v′′ . Otherwise, uv is bad. • Suppose au was not added to S because degS ( a) = λU , then uv is bad. • Otherwise, au was added to S, but uv did not get augmented when vb arrived. Then: – Either there exists a′ ∈ NS (u) that was matched via augmenting path a′ u′ v′ b′ (note that a′ may be same as a), then we charge uv to u′ v′ ; – or b was already matched via augmenting path a′′ u′′ v′′ b, and vb was ignored; in this case, if degS (b) 6 λU − 1, then we charge uv to u′′ v′′ , otherwise, uv is bad. We now bound the number of bad edges in M0 from above. Lemma 5.2. The number of bad edges is at most λM | M0 |/λU . Proof. We claim that for any uv ∈ M0 , degS (u) + degS (v) 6 λM , hence |S| 6 λM | M0 |. A short argument is that the (λM + 1)th edge would cause an augmentation and will not be added to S. Let us assume the claim. By the definition of a bad edge, λU edges in S are “responsible” for one bad edge in M0 . Also, an edge au′ (or v′′ b, resp.) in S can be responsible for at most one bad edge that can only be uv if au ∈ / S (or if vb ∈ / S, resp.; considering the 3-augmenting path auvb). Hence, the total number of bad edges is at most |S|/λU 6 λM | M0 |/λU . Now we prove the claim. We first prove the claim for triangle-free graphs by contradiction. Let degS (u) + degS (v) > λM , and let vy ∈ S be the (λM + 1)th edge incident to one of u and v that was added to S. Since λM > 1 and degS (v) 6 λM , we have degS (u) > 1, i.e. NS (u) 6= ∅. Now when vy arrived: 11 • the vertex y was unmatched, otherwise vy would not be added to S; • no vertex x ∈ NS (u) was matched, otherwise u, v ∈ IB , and vy would not be added to S. The above implies that when vy arrived, due to some x ∈ NS (u) the if condition on Line 14 became true, and we augmented uv via xuvy instead of adding vy to S. This is a contradiction. For general graphs, we argue by contradiction slightly informally for the sake of brevity. By Observation 5.1, for general graphs, λM > 2. Let degS (u) + degS (v) > λM > 2. Let vy be the second edge incident to one of u and v that was added to S; the first edge can be xu or vy′ . Suppose xu was the first edge. If x 6= y, then we would have augmented uv via xuvy instead of adding vy to S—a contradiction. If x = y, then after vy was processed, NS (u) = NS (v) = {y}, and a third edge incident to one of u and v would not be added to S, because it would have formed a 3-augmenting path with either yu or vy, resulting in a contradiction that degS (u) + degS (v) = 2. Otherwise, suppose vy′ was the first edge; then NS (v) = {y, y′ } after vy was processed. Since eventually degS (u) + degS (v) > λM + 1 > 3 and degS (u), degS (v) 6 λM , we would eventually have degS (u) > 1, so let xu ∈ S. When xu arrived, it would have formed an 3-augmenting path with either vy or vy′ (here, taking care of the fact that one of y and y′ can be same as x), resulting in a contradiction that xu was not added to S. Thus, we get the claim and complete the proof. As a consequence, we get the following. Observation 5.3. In any call to I MPROVE -M ATCHING(), we need to set λU > λM , i.e., λU > 2. To see why, suppose λU 6 λM . Then by Lemma 5.2, potentially all 3-augmentable edges in M0 could become bad edges. Recall that a 3-augmentable edge is good, partially good, or bad; so by Lemmas 2.1 and 5.2,   1 λ | M0 | − 3α | M ∗ | − M # good or partially good edges > 2 λU     λM 1 1 ∗ − 3α | M | − + α | M∗ | = 2 λU 2     λU − λM 3λU + λM = − α | M∗ | . (6) 2λU λU In the following Lemma, we bound the number of partially good edges in M0 that are charged to one good edge. Lemma 5.3. At most 2λU − 1 partially good edges in M0 are charged to one good edge in M0 . Proof. Suppose uv ∈ M0 was augmented by edges xu and vy such that xu arrived before vy, then xu ∈ S. Now | NS ( x)|, | NS (y)| 6 λU . Since xu ∈ S, we have | NS ( x) \ {u}| 6 λU − 1. Let B := ( NS ( x) \ {u}) ∪ NS (y), then | B| 6 2λU − 1. Now, the set of partially good edges that are charged to uv is a subset of M0 ( B). Observing that | M0 ( B)| 6 | B| 6 2λU − 1 finishes the proof. The following lemma characterizes the improvement given by I MPROVE -M ATCHING(). Lemma 5.4. Let | M0 | = (1/2 + α)| M ∗ | and M = I MPROVE -M ATCHING ( M0 , λU , λM ), then       1 λU − λM 3λU + λM 1 λU − λM ∗ + + + 1− α |M | > | M∗ | . | M| > 2 2 4λ2U 2λ2U 4λ2U 12 Proof. By (6) and Lemma 5.3, the total number of augmentations during one call to I MPROVE M ATCHING() is at least         3λU + λM 3λU + λM 1 λU − λM λU − λM ∗ − α | M∗ | . − α |M | = 2λU 2λU λU 4λ2U 2λ2U Hence, we get the following bound on the size of the output matching M:   λU − λM 3λU + λM | M | > | M0 | + − α | M∗ | 4λ2U 2λ2U     3λU + λM 1 λU − λM + + 1− α | M∗ | because | M0 | = (1/2 + α)| M ∗ | , = 2 4λ2U 2λ2U   1 λU − λM > | M∗ | since λU > 2 (see Observation 5.3) . + 2 4λ2U Now we state and prove our main result. Theorem 5.5. Algorithm 3 uses two passes and has an approximation ratio of 1/2 + 1/16 for triangle-free graphs and an approximation ratio of 1/2 + 1/32 for general graphs for maximum matching. Proof. By Lemma 5.4, after the second pass, the output size | M | > (1/2 + (λU − λM )/(4λ2U ))| M ∗ |; we use λU = 2 and λM = 1 for triangle-free graphs and λU = 4 and λM = 2 (see Observation 5.1) for general graphs to get the claimed approximation ratios. 6 Multi Pass Algorithm We run the function I MPROVE -M ATCHING() in Algorithm 3 with increasing values of λU , and the approximation ratio converges to 1/2 + 1/6. We note that this multi-pass algorithm is not just a repetition of the function I MPROVE -M ATCHING(). Such a repetition will give an asymptotically worse number of passes (see, for example, the multi-pass algorithm due to Feigenbaum et al. [13]). We carefully choose the parameter λU for each pass to get the required number of passes. Algorithm 4 Multi-pass algorithm: input graph G 1: In the first pass, find a maximal matching M1 . 2: M ← M1 3: if G is triangle-free then 4: for i = 2 to ⌈2/(3ε)⌉ do 5: M ← Improve-Matching( M, i, 1) 6: 7: 8: 9: else for i = 2 to ⌈4/(3ε)⌉ do M ← Improve-Matching( M, i + 1, 2) Return M. Theorem 6.1. For any ε > 0, Algorithm 4 is a semi-streaming (1/2 + 1/6 − ε)-approximation algorithm for maximum matching that uses 2/(3ε) passes for triangle-free graphs and 4/(3ε) passes for general graphs. 13 Proof. We prove the theorem for triangle-free case; the general case is similar. Let Mi be the matching computed by Algorithm 4 after ith pass, and let p := ⌈2/(3ε)⌉, so ε 6 2/(3p). Since M1 is maximal, it is (1/2)-approximate. Let α1 := 0, and for i ∈ {2, 3, . . . , p}, let   3i + 1 i−1 + 1− α i−1 αi := 4i2 2i2 (see Lemma 5.4 with λU = i and λM = 1). Then, by Lemma 5.4 and the logic of Algorithm 4, for i ∈ [ p], the matching Mi is (1/2 + αi )-approximate (by a trivial induction). Now we bound α p by induction. We claim that for i ∈ [ p], 1 2 αi > − , 6 3i which we prove by induction on i. Base case: For i = 1, we have 1/6 − α1 = 1/6 − 0 = 1/6 6 2/(3 · 1). For inductive step, we want to show that   1 i−1 3i + 1 1 2 − αi = − 2 − 1 − α i−1 6 , 2 6 6 4i 2i 3i which is implied by the following (using inductive hypothesis)    2 1 3i + 1 1 i−1 − 2 + 1− − 6 6 4i 2i2 3( i − 1) 6  2   2i − 3i − 1 4−i+1 1 i−1 − 2 + 6 6 4i 2i2 6( i − 1) which is implied by 2 , 3i 2 , 3i multiplying both sides by 12i2 (i − 1), we then need to show that, which is implied by which is implied by 2i2 (i − 1) − 3(i − 1)2 + (2i2 − 3i − 1)(−i + 5) 6 8i (i − 1) 2i3 − 2i2 − 3(i2 − 2i + 1) + (−2i3 + 10i2 + 3i2 − 15i + i − 5) 6 8i2 − 8i which is implied by 2i3 − 5i2 + 6i − 3 + (−2i3 + 13i2 − 14i − 5) 6 8i2 − 8i 8i2 − 8i − 8 6 8i2 − 8i which is true, so we get the claim. Therefore α p > 1/6 − 2/(3p) > 1/6 − ε, and by our earlier observation, M p is (1/2 + α p )-approximate, and this finishes the proof for triangle-free case. The proof for general case is very similar. We define p := ⌈4/(3ε)⌉ and α1 := 0, and for i ∈ {2, 3, . . . , p}, we define   i−1 3( i + 1) + 2 αi := + 1− α i−1 , 4( i + 1) 2 2( i + 1) 2 i.e., we use λU = i + 1 and λM = 2. The corresponding claim then is that for i ∈ [ p], αi > 4 1 − , 6 3i which can be verified by induction on i. Acknowledgements. We thank Sundar Vishwanathan and Ashish Chiplunkar for helpful discussions. The first author would like to thank his advisor Amit Chakrabarti and Andrew McGregor for helpful discussions. 14 References [1] Kook Jin Ahn and Sudipto Guha. Linear programming in the semi-streaming model with application to the maximum matching problem. Inf. Comput., 222:59–79, January 2013. [2] Sepehr Assadi, Sanjeev Khanna, and Yang Li. On estimating maximum matching size in graph streams. In Proc. 28th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1723–1742, 2017. [3] Sepehr Assadi, Sanjeev Khanna, Yang Li, and Grigory Yaroslavtsev. Maximum matchings in dynamic graph streams and the simultaneous communication model. In Proc. 27th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1345–1364, 2016. [4] Marc Bury and Chris Schwiegelshohn. Sublinear estimation of weighted matchings in dynamic data streams. In Proc. 23rd Annual European Symposium on Algorithms, pages 263–274, 2015. [5] Amit Chakrabarti and Sagar Kale. Submodular maximization meets streaming: matchings, matroids, and more. Mathematical Programming, 154(1):225–247, 2015. [6] Chandra Chekuri, Shalmoli Gupta, and Kent Quanrud. Streaming algorithms for submodular function maximization. In Proc. 42nd International Colloquium on Automata, Languages and Programming, pages 318–330, 2015. [7] Rajesh Chitnis, Graham Cormode, Hossein Esfandiari, MohammadTaghi Hajiaghayi, Andrew McGregor, Morteza Monemizadeh, and Sofya Vorotnikova. Kernelization via sampling with applications to finding matchings and related problems in dynamic graph streams. In Proc. 27th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1326–1344, 2016. [8] Michael Crouch and Daniel M. Stubbs. Improved streaming algorithms for weighted matching, via unweighted matching. In Proc. 17th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems, volume 28, pages 96–104, 2014. [9] Sebastian Eggert, Lasse Kliemann, Peter Munstermann, and Anand Srivastav. matching in the semi-streaming model. Algorithmica, 63(1):490–508, 2012. Bipartite [10] Leah Epstein, Asaf Levin, Julian Mestre, and Danny Segev. Improved approximation guarantees for weighted matching in the semi-streaming model. SIAM Journal on Discrete Mathematics, 25(3):1251–1265, 2011. [11] H. Esfandiari, M. Hajiaghayi, and M. Monemizadeh. Finding large matchings in semistreaming. In 2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW), pages 608–614, Dec 2016. [12] Hossein Esfandiari, Mohammad T. Hajiaghayi, Vahid Liaghat, Morteza Monemizadeh, and Krzysztof Onak. Streaming algorithms for estimating the matching size in planar graphs and beyond. In Proc. 26th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1217–1233, 2015. [13] Joan Feigenbaum, Sampath Kannan, Andrew McGregor, Siddharth Suri, and Jian Zhang. On graph problems in a semi-streaming model. Theor. Comput. Sci., 348(2):207–216, December 2005. 15 [14] Ashish Goel, Michael Kapralov, and Sanjeev Khanna. On the communication and streaming complexity of maximum bipartite matching. In Proc. 23rd Annual ACM-SIAM Symposium on Discrete Algorithms, pages 468–485, 2012. [15] Elena Grigorescu, Morteza Monemizadeh, and Samson Zhou. Streaming weighted matchings: Optimal meets greedy. CoRR, abs/1608.01487, 2016. [16] Michael Kapralov. Better bounds for matchings in the streaming model. In Proc. 24th Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM, 2013. [17] Michael Kapralov, Sanjeev Khanna, and Madhu Sudan. Approximating matching size from random streams. In Proc. 25th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 734–751, 2014. [18] Richard M. Karp, Umesh V. Vazirani, and Vijay V. Vazirani. An optimal algorithm for on-line bipartite matching. In Proc. 22nd Annual ACM Symposium on the Theory of Computing, pages 352–358, 1990. [19] Christian Konrad. Maximum matching in turnstile streams. In Proc. 23rd Annual European Symposium on Algorithms, pages 840–852, 2015. [20] Christian Konrad, Frédéric Magniez, and Claire Mathieu. Maximum matching in semistreaming with few passes. CoRR, abs/1112.0184, 2014. [21] Andrew McGregor. Problem 60: Single-pass unweighted matchings. http://sublinear.info/index.php?title=Open_Problems:60. Accessed: 2017-02-16. [22] Andrew McGregor. Finding graph matchings in data streams. In Proc. 8th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems, pages 170–181, 2005. [23] Ami Paz and Gregory Schwartzman. A (2 + ε)-approximation for maximum weight matching in the semi-streaming model. In Proc. 28th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 2153–2161, 2017. [24] Mariano Zelke. Weighted matching in the semi-streaming model. In Proc. 25th International Symposium on Theoretical Aspects of Computer Science, pages 669–680, 2008. A Three Pass Algorithm for Triangle Free Graphs For completeness, we present our three-pass algorithm for triangle-free graphs. Theorem A.1. Algorithm 5 is a three-pass, semi-streaming, (1/2 + 1/10)-approximation algorithm for maximum matching in triangle-free graphs, and the analysis is tight. Proof. Let | M0 | = (1/2 + α)| M ∗ |. The number of edges in M ∗ incident on V ( M ∗ ) \ V ( M0 ) is |V ( M ∗ ) \ V ( M0 )| > |V ( M ∗ )| − |V ( M0 )| = 2| M ∗ | − 2| M0 | = (1 − 2α)| M ∗ | ; (7) and these edges also belong to F1 . Since M1 is a maximal matching in F1 , | M1 | > (1 − 2α)| M ∗ |/2 = (1/2 − α)| M ∗ | . 16 (8) Algorithm 5 Three-pass algorithm for triangle-free graphs 1: In the first pass, find a maximal matching M0 . 2: In the second pass, find a maximal matching M1 in F1 : = {uv : u ∈ V \ V ( M0 ), v ∈ V ( M0 )}. 3: After the second pass: • M1′ ← arbitrary largest subset of M1 such that there is no 3-augmenting path in M1′ ∪ M0 with respect to M0 • V2 ← { x ∈ V ( M0 ) : ∃v, w such that vw ∈ M1′ and wx ∈ M0 } • For x ∈ V2 , denote by P( x) the vertex v such that there exists w with vw ∈ M1′ and wx ∈ M0 . See x and P( x) in Figure 6. 4: 5: 6: 7: 8: 9: 10: In the third pass: F2 := { xy : x ∈ V2 , y ∈ V \ V ( M0 )} M2 ← ∅ for edge xy ∈ F2 do if x, and y are unmarked then M2 ← M2 ∪ { xy}; since the graph is triangle free, y 6= P( x), and we can augment M0 using xy. Mark P( x), x, y, and P−1 (y) (if exists). Let M be largest of M3 and M3′ which are computed below. • Augment M0 using edges in M1 to get M3 . • Augment M0 using edges in M1′ and M2 to get M3′ . 11: Output M. Let c be the number of 3-augmenting paths in M1 ∪ M0 , so | M1′ | = | M1 | − c by the definition of M1′ . By Lemma 2.1, there are at most 4α| M ∗ | non-3-augmentable edges in M0 . So at least | M1 | − c − 4α| M ∗ | edges of M1′ are incident on 3-augmentable edges of M0 . Therefore there is a matching of size at least | M1 | − c − 4α| M ∗ | in F2 ; consider one such matching M F . We claim that | M2 | > | M F |/4. See Figure 6. Let xy ∈ M2 ; we note that xy disallows at most four edges in M F from being added to M2 due to the (at most) four marks that it adds, because a marked vertex can disallow at most one edge in M F (due to it being a matching), which shows the claim. Hence: P( x) y x M0 P −1 ( y) MF M1′ a stream edge Figure 6: An edge xy ∈ M2 disallows at most four edges in M F from being added to M2 . 17 B \ B ( M0 ) A ( M0 ) B ( M0 ) A \ A ( M0 ) M0 M∗ M1 Figure 7: Tight example for Algorithm 5: M1 has only two edges that land on bad vertices and cannot be augmented in the third pass. So | M | = | M0 | = 3 and | M ∗ | = 5. | MF | 4 | M1 | − c − 4α| M ∗ | >   4  1 1 ∗ ∗ > − α | M | − c − 4α| M | 4 2    1 1 ∗ − 5α | M | − c . = 4 2 | M2 | > by (8) , Now, each edge in M2 gives one augmentation after the second pass. To see this, we observe that for any x ∈ V2 , at any point in the algorithm, x and P( x) are either both marked or both unmarked. So when an edge xy ∈ M2 arrives, x and y are unmarked, and P( x) and P−1 (y) (if it exists) are also unmarked, otherwise one of x and y would have been marked and xy would not have been added to M2 . Since both P( x) and P−1 (y) were unmarked, we can use the augmenting path { M1′ ({ P( x)}), M0 ({ x}), xy}. Hence we get at least     1 1 ∗ − 5α | M | − c max c, 4 2 augmentations after the third pass. This is minimized by setting    1 1 ∗ c= − 5α | M | − c 4 2    1 1 ∗ = − 5α | M | 5 2   1 = − α | M∗ | . 10 So we get the following bound:         1 1 1 1 1 ∗ ∗ ∗ | M | > | M0 | + − α |M | > + α |M | + − α |M | = + | M∗ | . 10 2 10 2 10 The tight example is shown in Figure 7. 18 B Three Pass Algorithm for General Graphs We find a maximal matching M1 in the first pass. Then we use I MPROVE -M ATCHING() function from Algorithm 3, i.e., • in the second pass, M2 ← I MPROVE - MATCHING ( M1 , 4, 2), and • in the third pass, M3 ← I MPROVE - MATCHING ( M2 , 5, 2). We observe that M1 is (1/2)-approximate. Then by double application of Lemma 5.4, we get that M3 is (1/2 + 81/1600) ≈ (1/2 + 1/19.753)-approximate. C Three Pass Algorithm for Bipartite Graphs: Suboptimal Analysis We now give an analysis of Algorithm 1 that shows approximation ratio of only 1/2 + 1/18 that is based on Konrad et al.’s [20] analysis for their two-pass algorithm for bipartite graphs. Afterward, we demonstrate that by not considering the distribution of lengths of augmenting paths, we may prove an approximation ratio of at most 1/2 + 1/14. The better and tight analysis appears in Section 3. Theorem C.1. Algorithm 1 is a three-pass, semi-streaming, (1/2 + 1/18)-approximation algorithm for maximum matching in bipartite graphs. Proof. As usual, let | M0 | = (1/2 + α)| M ∗ |. Since M0 is a maximal matching, there are | B( M ∗ ) \ B( M0 )| edges of M ∗ that are also in F2 . We have | B( M ∗ ) \ B( M0 )| > | B( M ∗ )| − | B( M0 )| = | M ∗ | − | M0 | , and since M A is maximal, we then get the following:      1 1 1 1 1 1 ∗ ∗ ∗ +α |M | = − α | M ∗ |. (9) 1− | M A | > | B( M ) \ B( M0 )| > (| M | − | M0 |) = 2 2 2 2 2 2 By Lemma 2.1, there are at most 4α| M ∗ | non-3-augmentable edges in M0 . Which means that at least | M A | − 4α| M ∗ | edges of M A are incident on 3-augmentable edges of M0 ; therefore there is a matching of size at least | M A | − 4α| M ∗ | in F3 . Since we output a maximal matching in F3 , we get at least (1/2)(| M A | − 4α| M ∗ |) augmentations after the third pass. So we get the following bound: 1 | M | > | M0 | + (| M A | − 4α| M ∗ |) 2    1 1 1 − α − 4α | M ∗ | > | M0 | + 2 2 2   1 9 − α | M∗ | = | M0 | + 8 4     1 1 9 ∗ = + α |M | + − α | M∗ | 2 8 4   1 1 5 = + − α | M∗ | . 2 8 4 by (9) , because | M0 | = (1/2 + α)| M ∗ | , We also have | M | > | M0 | = (1/2 + α)| M ∗ |. As α increases, the former bound deteriorates and the latter improves, so the worst case α is when these two bounds are equal, which happens at α = 1/18, and the approximation ratio we get is 1/2 + 1/18. 19 C.1 Improved Analysis Without Considering Longer Augmenting Paths We can analyze Algorithm 1 better if we bound | M A | more carefully. The claim is that at least (1/2 − 7α)| M ∗ |/2 edges of M A are incident on 3-augmentable edges of M0 . Let A G ⊆ A( M0 ) be the set of vertices in A that are endpoints of 3-augmentable edges of M0 ; also, let A N = A( M0 ) \ A G . So there is a matching of size at least | A G | in F2 that covers A G . Any maximal matching in F2 has at least (| A G | − | A N |)/2 edges that are incident on A G . To see the claim, we use the facts | A G | > (1/2 − 3α)| M ∗ | and | A N | 6 4α| M ∗ |. So there is a matching of size at least (1/2 − 7α)| M ∗ |/2 in F3 . We output a maximal matching in F3 ; hence we get at least (1/2 − 7α)| M ∗ |/4 augmentations after the third pass. So we get the following bound:   1 1 | M | > | M0 | + − 7α | M ∗ | 4 2     1 1 1 ∗ + α |M | + − 7α | M ∗ | = 2 4 2   1 1 3 = + − α | M∗ | . 2 8 4 where the second inequality is by (9). We also have | M | > | M0 | = (1/2 + α)| M ∗ |, so the worst case α is when these two bounds are equal, which happens at α = 1/14 and the approximation ratio we get is 1/2 + 1/14, and we get the following theorem. Theorem C.2. Algorithm 1 is a three-pass, semi-streaming, (1/2 + 1/14)-approximation algorithm for maximum matching in bipartite graphs. D A Note on the Analysis by Esfandiari et al. [11] We demonstrate with an example that the analysis of the algorithm by Esfandiari et al. [11] given for bipartite graphs cannot be extended for triangle-free graphs to get the same approximation ratio. See Figure 8. Lemma 6 in their paper, as they correctly claim, holds only for bipartite graphs and not for triangle-free graphs. Our algorithm in Section 4 is essentially the same algorithm except for the post-processing step; we augment the maximal matching computed in the first pass greedily, whereas they use an offline maximum matching algorithm. We have highlighted some other comparison points in Section 1. 20 a7 u7 v7 a6 u6 v6 u5 v5 b7 u4 v4 b2 u3 v3 a1 M a2 u2 some M ∗ edges v2 golden edges u1 v1 Figure 8: Example demonstrating that Lemma 6 in Esfandiari et al. [11] does not hold when the input graph is not bipartite but is triangle-free. We use k = 3. For an M edge ui vi , there are two M ∗ edges incident on it, which are ai ui and vi bi , and some of the M ∗ edges are not shown, but all golden edges are shown, which we call support edges or denote by S in our terminology. It can be seen from this example that their algorithm is not a (1/2 + 1/12)-approximation algorithm for triangle free graphs, because out of the seven 3-augmentable edges in M, only one will get augmented, thereby giving a worse approximation ratio. 21
8cs.DS
Higher-Order Linearisability Andrzej S. Murawski1 and Nikos Tzevelekos2 arXiv:1610.07965v1 [cs.PL] 25 Oct 2016 1 University of Warwick 2 Queen Mary University of London Abstract. Linearisability is a central notion for verifying concurrent libraries: a given library is proven safe if its operational history can be rearranged into a new sequential one which, in addition, satisfies a given specification. Linearisability has been examined for libraries in which method arguments and method results are of ground type, including libraries parameterised with such methods. In this paper we extend linearisability to the general higher-order setting: methods can be passed as arguments and returned as values. A library may also depend on abstract methods of any order. We use this generalised notion to show correctness of several higher-order example libraries. 1 Introduction Computer programs often take advantage of libraries, which are collections of routines, often of specialised nature, implemented to facilitate software development and, among others, code reuse and modularity. To support the latter, libraries should follow their specifications, which describe the range of expected behaviours the library should conform to for safe and correct deployment. Adherence to given specifications can be formalised using the classic notion of contextual approximation (refinement), which scrutinises the behaviour of code in any possible context. Unfortunately, the quantification makes it difficult to prove contextual approximations directly, which motivates research into sound techniques for establishing it. In the concurrent setting, a notion that has been particularly influential is that of linearisability [12]. Linearisability requires that, for each history generated by a library, one should be able to find another history from the specification (its linearisation), which matches the former up to certain rearrangements of events. In the original formulation by Herlihy and Wang [12], these permutations were not allowed to disturb the order between library returns and client calls. Moreover, linearisations were required to be sequential traces, that is, sequences of method calls immediately followed by their returns. This notion of linearisability only applies to closed, i.e. fully implemented, libraries in which both method arguments and results are of ground types. The closedness limitation was lifted by Cerone, Gotsman and Yang [3], who extended the techniques to parametric libraries, whereby methods were divided into available routines (public methods) and unimplemented ones (abstract methods). However, both public and abstract methods were still restricted to first-order functions of type int → int. In this paper, we make a further step forward and present linearisability for general higher-order concurrent libraries, where methods can be of arbitrary higher-order types. In doing so, we also propose a corresponding notion of sequential history for higher-order histories. We examine libraries L that can interact with their environments by means of public and abstract methods: a library L with abstract methods of types Θ = θ1 , ⋯, θn and public methods Θ′ = θ1′ , ⋯, θn′ ′ is written as L ∶ Θ → Θ′ . We shall work with arbitrary higher-order types generated from the ground types unit and int. Types in Θ, Θ′ must always be function types, i.e. their order is at least 1. A library L may be used in K L L' computations by placing it in a PL context that will keep on calling Θ Θ' M1 Θ' PK Θ OL OK its public methods (via a client K) as well as providing impleΘ'' MN Θ'' mentations for the abstract ones (via a parameter library L′ ). The ′ comprising setting is depicted in Figure 1. Fig. 1. A library L ∶ Θ′ → Θ in environment ′′ a parameter library L ∶ ∅ → Θ, Θ and a client K of the Note that, as the library L inform Θ′ , Θ′′ ⊢ M1 ∥⋯∥MN . ′ teracts with K and L , they exchange functions between each other. Consequently, in addition to K making calls to public methods of L and L making calls to its abstract methods, K and L′ may also issue calls to functions that were passed to them as arguments during higher-order interactions. Analogously, L may call functions that were communicated to it via library calls. Our framework is operational in flavour and draws upon concurrent [15,7] and operational game semantics [13,16,8]. We shall model library use as a game between two participants: Player (P), corresponding to the library L, and Opponent (O), representing the environment (L′ , K) in which the library was deployed. Each call will be of the form call m(v) with the corresponding return of the shape ret m(v), where v is a value. As we work in a higher-order framework, v may contain functions, which can participate in subsequent calls and returns. Histories will be sequences of moves, which are calls and returns paired with thread identifiers. A history is sequential just if every move produced by O is immediately followed by a move by P in the same thread. In other words, the library immediately responds to each call or return delivered by the environment. In contrast to classical linearisability, the move by O and its response by P need not be a call/return pair, as the higher-order setting provides more possibilities (in particular, the P response may well be a call). Accordingly, linearisable higher-order histories can be seen as sequences of atomic segments (linearisation points), starting at environment moves and ending with corresponding library moves. In the spirit of [3], we are going to consider two scenarios: one in which K and L′ share an explicit communication channel (the general case) as well as a situation in which they can only communicate through the library (the encapsulated case). Further, in the encapsulated case, we will handle the case in which extra closure assumptions can be made about the parameter library (the relational case). The restrictions can deal with a variety of assumptions on the use of parameter libraries that may arise in practice. In each of the three cases, we shall present a candidate definition of linearisability and illustrate it with tailored examples. The suitability of each kind of linearisability will be demonstrated by showing that it implies the relevant form of contextual approximation (refinement). We shall also examine compositionality of the proposed concepts. One of our examples will discuss the correctness of an implementation of the flat-combining approach [11,3], adapted to higher-order types. 2 2 Higher-order linearisability As mentioned above, we examine libraries interacting with their context by means of abstract and public methods. In particular, we consider higher-order types given by the grammar on the left below, and let Meths be a set of method names. θ ∶∶= unit ∣ int ∣ θ × θ ∣ θ → θ Meths = ⊎θ,θ′ Methsθ,θ′ v ∶∶= () ∣ i ∣ m ∣ (v, v) Methods are ranged over my m (and variants), and each set Methsθ,θ′ contains names for methods of type θ → θ′ . Finally, we let v range over computational values, which include a unit value, integers, methods, and pairs of values. The framework of a higher-order library and its environment is depicted in Figure 1. Given Θ, Θ′ ⊆ Meths, a library L is said to have type Θ → Θ′ if it defines public methods with names (and types) as in Θ′ , using abstract methods Θ. The environment of L consists of a client K (which invokes the public methods of Θ′ ), and a parameter library L′ (which provides the code for the abstract methods Θ). In general, K and L′ may be able to interact via a disjoint set of methods Θ′′ ⊆ Meths, to which L has no access. In the rest of this paper we will be implicitly assuming that we work with a library L operating in an environment presented in Figure 1. The client K will consist of a fixed number N of concurrent threads. Next we introduce a notion of history tailored to the setting and define how histories can be linearised. In Section 3 we present the syntax for libraries and clients; and in Section 4 we define their semantics in terms of histories (and co-histories). 2.1 Higher-order histories The operational semantics of libraries will be given in terms of histories, which are sequences of method calls and returns each decorated with a thread identifier and a polarity index XY , where X ∈ {O, P } and Y ∈ {L, K}: (t, call m(v))XY (t, ret m(v))XY We refer to decorated calls and returns like above as moves. Here, m is a method name and v is a value of matching type. The index XY is specifying which of the three entities (L, L′ , K) produces the move, and towards whom it is addressed. – If X = P then the move is issued by L. Moreover, if Y = L then it is addressed to L′ ; otherwise, if Y = K then it is addressed to K. – If XY = OL then the move is issued by L′ , and is addressed to L. – If XY = OK then the move is issued by K, and is addressed to L. We can justify the choice of indices: the moves can be seen as defining a 2-player game between the library (L), which represents the Proponent player in the game, and its enviroment (L′ , K) that represents the Opponent. Moves played between L and L′ are moreover decorated with L; whereas those between L and K have K. Note that the possible interaction between L′ and K is invisible to L and is therefore not accounted for in the game (but we will later see how it can affect it). We use O to refer to either OK or OL, and P to refer to either P K or P L. 3 Definition 1. [Prehistories] We define prehistories as sequences of moves derived by one of the following grammars, P ′ O HO pre ∶∶=  ∣ call m(v)OY Hpre ret m(v )P Y Hpre P O ′ Hpre ∶∶=  ∣ call m(v)P Y Hpre ret m(v )OY HP pre where, in each line, the two occurrences of Y ∈ {K, L} and m ∈ Meths must each match. Moreover, if m ∈ Methsθ,θ′ , the types of v, v ′ must match θ, θ′ respectively. The elements of HO pre are patterns of actions starting with an O-move, while those in start with a P -move. Note that, in each case, the polarities alternate and the polarities of calls and matching returns always match the pattern (XY, X ′ Y ) for X =/ X ′ . Histories will be interleavings of prehistories tagged with thread identifiers (natural O/P numbers) that satisfy a number of technical conditions. Given h ∈ Hpre and t ∈ N, we write t × h for h in which each call or return is decorated with t. We refer to such moves with (t, call m(v))XY or (t, ret m(v))XY respectively. If we only want to stress the X or Y membership, we shall drop Y or X respectively. Moreover, when no confusion arises, we may sometimes drop a move’s polarity altogether. HP pre Definition 2. [Histories] Given Θ, Θ′ , the set of histories over Θ → Θ′ is defined by: HΘ,Θ′ = ⋃N >0 ⋃h1 ,⋯,hN ∈HO (1 × h1 ) ∣ ⋯ ∣ (N × hN ) pre where (1 × h1 ) ∣ ⋯ ∣ (N × hN ) is the set of all interleavings of (1 × h1 ), ⋯, (N × hN ), satisfying the following conditions. 1. For any s1 (t, call m(v))XY s2 ∈ HΘ,Θ′ : – either m ∈ Θ′ and XY = OK, or m ∈ Θ and XY = P L, – or there is a move (t′ , x′ )X ′ Y in s1 with X ≠ X ′ , such that x′ ∈ {call m′ (v), ret m′ (v)} and v contains m. 2. For any s1 (t, x)XY s2 ∈ HΘ,Θ′ , where x ∈ {call m(v), ret m(v)} and v includes some m′ ∈ Meths, m′ must not occur in s1 . Condition 1 in the definition above requires that any call must refer to Θ or Θ′ , or be introduced earlier as a higher-order argument or result. If the method is from Θ′ , the call must be tagged with OK (i.e. issued by K). Dual constraints apply to Θ. If a method name does not come from Θ or Θ′ , in order for the call (t, call m(v))XY to be valid, m must be introduced in an earlier action with the same tag Y but with the opposite tag X. Moreover, as specified by Condition 2, any action involving a higher-order value (i.e. a method name) in its argument or result must label it with a fresh name, one that has not been used earlier. This is done to enable the history to refer unambiguously to each method name encountered during the interaction. We shall range over HΘ,Θ′ using h, s. The subscripts Θ, Θ′ will often be omitted. Remark 3. Histories will be used to define the semantics JLK of libraries (cf. Section 4). In particular, for each library L ∶ Θ → Θ′ , we shall have JLK ⊆ HΘ,Θ′ . 4 1 public count, update; 2 3 4 Lock lock; F := λx.0; 11 12 5 6 13 count = λi. (! F)i 14 7 8 9 10 15 update = λ(i, g ). upd r( i , g ,(! F)i ) 16 upd r = λ(i, g, j ). let y = | gj | in lock . acquire (); let f = !F in if ( j==fi ) then { F := λx. if (x==i) then y else fx ; lock . release (); y } else { lock. release (); upd r( i ,g, fi ) } Fig. 2. Multiset library Lmset . [count ∶ int → int, update ∶ int × (int → int) → int] Example 4. Let Θ = {m ∶ int → int} and Θ′ = {m′ ∶ int → int}. Note that singlethreaded library histories from H∅,Θ′ must have the shape (1, call m′ (i1 ))OK (1, ret m′ (j1 ))P K ⋯ (1, call m′ (ik ))OK (1, ret m′ (jk ))P K . In this case, the definition coincides with [3]. However, in general, our notion of histories is more liberal. For example, the sequence (1, call m′ (1))OK (1, call m(2))P L (1, call m′ (3))OK (1, ret m′ (4))P K (1, ret m(5))OL (1, ret m′ (6))P K is in HΘ,Θ′ , even though it is not allowed by Definition 1 of [3]. The sequence represents a scenario in which the public method m′ is called for the second time before the first call is answered. In our higher-order setting, this scenario may arise if the parameter library communicates with the client and the communication includes a function that can issue a call to the public method m′ .1 Finally, we present a notion of sequential history, which generalises that of [12]. Definition 5. We call a history h ∈ HΘ,Θ′ sequential if it is of the form h = (t1 , x1 )OY1 (t1 , x′1 )P Y1′ ⋯ (tk , xk )OYk (tk , x′k )P Yk′ seq for some ti , xi , x′i , Yi , Yi′ . We let HΘ,Θ ′ contain all sequential histories of HΘ,Θ ′ . Next we consider our first example. Example 6 (Multiset). Consider a concurrent multiset library Lmset that uses a private reference for storing the multiset’s characteristic function. The implementation is given in Figure 2. This is a simplified version of the optimistic set algorithm of [10,19], albeit extended with a higher-order update method. The method computes the new value for element i without acquiring a lock on the characteristic function in the hope that when the lock is acquired the value at i will still be the same and the update can proceed (otherwise another attempt to update the value has to be made). The use of a single reference instead of a linked list means that memory safety is no longer problematic, so we focus on linearisability instead. Note we write ∣j∣ for the absolute value of j. Our verification goal will be to prove linearisability of Lmset to a specification seq Amset ⊆ H∅,Θ , where Θ = {count, update} (the method upd r is private). Amset certifies that Lmset correctly implements some integer multiset I whose elements change over time according to the moves in h. That is, for each history h ∈ Amset there is a multiset I that is empty at the start of h (i.e. I(i) = 0 for all i) and:2 1 2 By comparison, in [3], each (1, call m(v))P L must be followed by some (1, ret m(v ′ ))OL . For a multiset I and a natural number i, we write I(i) for the multiplicity of i in I; moreover, we set I[i ↦ j] to be I with its multiplicity of i set to j. 5 – If I changes value between two consecutive moves in h then the second move is a P -move. In other words, the client cannot update the elements of I. – Each call to count on argument i must be immediately followed by a return with value I(i), and with I remaining unchanged. – Each call to update on (i, m) must be followed by a call to m on i, with I unchanged. Then, m must later return with some value j. Assuming at that point the multiset will have value I ′ , if I(i) = I ′ (i) then the next move is a return of the original update call, with value j; otherwise, a new call to m on i is produced, and so on. We formally define the specification next. ○ Let H∅,Θ contain extended histories over ∅ → Θ, which are histories where each move is accompanied by a multiset (i.e. the sequence consists of elements of the form ○ (t, x, I)XY ). For each s ∈ H∅,Θ , we let π1 (s) be the history extracted by projection, i.e. π1 (s) ∈ H∅,Θ . For each t, we let s ↾ t be the subsequence of s of elements with first component t. Writing ⊑ for the prefix relation, and dropping the Y index from moves (Y is always K here), we define Amset = {π1 (s) ∣ s ∈ A○mset } where: seq ○ A○mset = {s ∈ H∅,Θ ∣ π1 (s) ∈ H∅,Θ ∧ (∀s′ ( , I)P ( , J)O ⊑ s. I = J) ∧ ∀t. s ↾ t ∈ S} and, for each t, the set of t-indexed histories S is given by the following grammar. S →  ∣ (t, call cnt(i), I)O (t, ret cnt(I(i)), I)P S ∣ (t, call upd(i, m), I)O Mi,j I,J (t, ret upd(∣j∣), J[i ↦ ∣j∣])P S Mi,j I,J → (t, call m(I(i)), I)P S (t, ret m(j), J)O (if I(i) = J(i)) i,j ′ ′ Mi,j I,J → (t, call m(I(i)), I)P S (t, ret m(j ), J )O MJ ′ ,J (if I(i) =/ J ′ (i)) By definition, the histories in Amset are all sequential. The elements of A○mset carry along the multiset I that is being represented. The conditions on A○mset stipulate that O cannot change the value of I, while the rest of the conditions above are imposed by the grammar for S. With the notion of linearisability to be introduced next, it will be possible to show that JLmset K indeed linearises to Amset (see Section 5.1). 2.2 General linearisability We begin with a definition of reorderings on histories. Definition 7. Suppose X, X ′ ∈ {O, P } and X ≠ X ′ . Let ◁XX ′ ⊆ HΘ,Θ′ × HΘ,Θ′ be the smallest binary relation over HΘ,Θ′ satisfying s1 (t′ , x′ )X ′ (t, x)s2 ◁XX ′ s1 (t, x)(t′ , x′ )X ′ s2 s1 (t′ , x′ )(t, x)X s2 ◁XX ′ s1 (t, x)X (t′ , x′ )s2 where t ≠ t′ . Intuitively, two histories h1 , h2 are related by ◁XX ′ if the latter can be obtained from the former by swapping two adjacent moves from different threads in such a way that, after the swap, an X-move will occur earlier or an X ′ -move will occur later. Note that, because of X ≠ X ′ , the relation always applies to pairs of moves of the same polarity. On the other hand, we cannot have s1 (t, x)X (t′ , x′ )X ′ s2 ◁XX ′ s1 (t′ , x′ )X ′ (t, x)X s2 . 6 Definition 8. [General Linearisability] Given h1 , h2 ∈ HΘ,Θ′ , we say that h1 is linearised by h2 , written h1 ⊑ h2 , if h1 ◁∗P O h2 . Given libraries L, L′ ∶ Θ → Θ′ and set of sequential histories A ⊆ HΘ,Θ′ we write L ⊑ A, and say that L can be linearised to A, if for any h ∈ JLK there exists h′ ∈ A such seq that h ⊑ h′ . Moreover, we write L ⊑ L′ if L ⊑ JL′ K ∩ HΘ,Θ ′. Remark 9. The definition above follows the classic definition from [12] and allows us to (first) express linearisability in terms of a given library L and a sequential specification A. The definition given in [3], on the other hand, expresses linearisability as a relation between two libraries. This is catered for at the end of Definition 8. Explicitly, we have that L ⊑ L′ if for all h ∈ JLK there is some sequential h′ ∈ JL′ K such that h ⊑ h′ . Example 10. Let Θ = {m ∶ int → int} and Θ′ = {m′ ∶ int → int}. First, consider h, h1 , h2 ∈ H∅,Θ′ given by: h = (1, call m(1))OK (1, ret m(2))P K (2, call m(3))OK (2, ret m(4))P K h1 = (1, call m(1))OK (2, call m(3))OK (1, ret m(2))P K (2, ret m(4))P K h2 = (2, call m(3))OK (1, call m(1))OK (2, ret m(4))P K (1, ret m(2))P K The histories are related in the following ways: for any i = 1, 2, we have h ◁∗OP hi and hi ◁∗P O h. Moreover, h1 ◁∗XX ′ h2 and h2 ◁∗XX ′ h1 for any X ≠ X ′ . Note that we do not have h ◁∗P O hi or hi ◁∗OP h. Consider now h3 , h4 ∈ HΘ,Θ′ given respectively by: h3 = (1, call m(1))OK (1, call m′ (2))P L (1, ret m′ (3))OL (1, ret m(4))P K (2, call m(5))OK (2, call m′ (6))P L (2, ret m′ (7))OL (2, ret m(8))P K h4 = (1, call m(1))OK (2, call m(5))OK (1, call m′ (2))P L (1, ret m′ (3))OL (2, call m′ (6))P L (2, ret m′ (7))OL (2, ret m(8))P K (1, ret m(4))P K Observe that h3 ◁∗OP h4 (and, thus, h4 ◁∗P O h3 ). However, we do not have h4 ◁∗OP h3 or h3 ◁∗P O h4 . Regarding linearisability, we can make the following remarks. – Observe that in histories from H∅,Θ′ , we shall have the following actions: call m′ (i)O and ret m′ (j)P . Thus, ◁∗P O cannot swap (t, ret m′ (j)) with (t′ , call m′ (i)), as in the standard definition of linearisability [12]. – When HΘ,Θ′ is considered, the available actions are call m′ (i)O , ret m(j)O and call m(i)P , ret m′ (j)P . Then ◁∗P O coincides with Definition 2 of [3] for secondorder libraries. 2.3 Encapsulated linearisability A different notion of linearisability will be applicable in cases where the parameter library L′ of Figure 1 is encapsulated, that is, the client K can have no direct access to it (i.e. Θ′′ = ∅). In particular, we shall impose an extra condition on histories in order to reflect the more restrictive nature of interaction. Specifically, in addition to sequentiality in every thread, we shall disallow switches between L and K components by O. 7 Definition 11. We call a history h ∈ HΘ,Θ′ encapsulated if, for each thread t, if h = s1 (t, x)P Y s2 (t, x′ )OY ′ s3 and moves from t are absent from s2 then Y = Y ′ . Moroever, enc enc ′ we set HΘ,Θ ′ = {h ∈ HΘ,Θ ′ ∣ h encapsulated} and JLKenc = JLK ∩ HΘ,Θ ′ (if L ∶ Θ → Θ ). We define the corresponding linearisability notion as follows. Definition 12. [Enc-linearisability] Let ◇ ⊆ HΘ,Θ′ × HΘ,Θ′ be the smallest binary relation on HΘ,Θ′ such that s1 (t, m)Y (t′ , m′ )Y ′ s2 ◇ s1 (t′ , m′ )Y ′ (t, m)Y s2 for any Y, Y ′ ∈ {K, L} such that Y ≠ Y ′ and t ≠ t′ . enc Given h1 , h2 ∈ HΘ,Θ ′ , we say that h1 is enc-linearised by h2 , and write h1 ⊑enc h2 , ∗ if h1 (◁P O ∪ ◇) h2 and h2 is sequential. seq A library L ∶ Θ → Θ′ can be enc-linearised to A, written L ⊑enc A, if A ⊆ HΘ,Θ ′ ∩ enc ′ ′ HΘ,Θ′ and for any h ∈ JLKenc there exists h ∈ A such that h ⊑enc h . Moreover, we write seq L ⊑enc L′ if L ⊑enc JL′ Kenc ∩ HΘ,Θ ′. Remark 13. Recall Θ, Θ′ from Example 10. Note that histories may contain the following actions only: call m′ (i)OK , ret m(i)OL , call m(i)P L , ret m′ (i)P K . Then (◁P O ∪ ◇)∗ preserves the order between call m(i)P L and ret m(i)OL as well as that between ret m′ (i)P K and call m′ (i)OK , i.e. it coincides with Definition 3 of [3]. Example 14 (Parameterised multiset). We revisit the multiset library of Example 6 and extend it with an abstract method foo and a corresponding update method update enc which performs updates using foo as the value-updating function. In contrast to the update method of Lmset , the method update enc is not optimistic: it retrieves the lock upon its call, and only releases it before return. In particular, the method calls foo while it retains the lock. We call this library Lmult2 ∶ ∅ → Θ′ , with Θ′ = {count, update, update enc}. Observe that, were foo able to externally call update, we would reach a deadlock: foo would be keeping the lock while waiting for the return of a method that requires the lock. On the other hand, if the library is encapsulated then the latter scenario is not plausible. In such a case, Lmult2 linearises to the specification Amult2 which is defined as follows (cf. Example 6). Let Amset2 = {π1 (s) ∣ s ∈ A○mset2 } where: seq ○ ′ A○mset2 = {s ∈ H∅,Θ ′ ∣ π1 (s) ∈ H ∅,Θ ′ ∧ (∀s ( , I)P ( , J)O ⊑ s. I = J) ∧ ∀t. s ↾ t ∈ S} and the set S is now given by the grammar of Example 6 extended with the rule: S → (t, call upd enc(i), I)OK (t, call foo(i), I)P L (t, ret foo(j), I)OL (t, ret upd enc(∣j∣), I ′ )P K S with I ′ = I[i ↦ ∣j∣]. Linearisability is shown in Section 5.1. 2.4 Relational linearisability Finally, we consider a special subcase of the encapsulated case, in which parameter libraries must satisfy additional constraints, specified through relational closure. This notion is desirable in cases when unconditional encapsulated linearisability does not hold, yet one may want to show that it would hold conditionally on the parameter library. The condition on the parameter library L′ ∶ ∅ → Θ is termed as a relation R ⊆ H∅,Θ × H∅,Θ . 8 1 2 3 4 public count, update, update enc; abstract foo; Lock lock; F := λx.0; 5 6 20 21 22 count = λi. (! F)i 23 update = λ(i, g ). upd r( i , g ,(! F)i ) 25 7 8 9 10 24 upd r = λ(i, g, j ). ... update enc = λi. lock . acquire (); let y = | foo ((! F)i )| in let f = !F in F := λx. if (x==i) then y else fx ; lock . release (); y Fig. 3. Parameterised multiset library Lmset2 . [foo , update enc∶ int → int, lines 10-16 as in Fig. 2] Definition 15. Given a history h ∈ HΘ,Θ′ and X ∈ {K, L}, we write h ↾ X for the subsequence of h with moves whose second index is X. Given a sequence of moves s, we write s for the sequence obtained from s by simply swapping the O/P indexes in its moves (e.g. xOL yP K zP K = xP L yOK zOK ). Definition 16. [Relational linearisability] Let R ⊆ H∅,Θ × H∅,Θ be a set closed under enc permutations of names in Meths ∖ Θ. Given h1 , h2 ∈ HΘ,Θ ′ , we say that h1 is Rlinearised by h2 , written h1 ⊑R h2 , if h2 is sequential and (h1 ↾K) ◁∗P O (h2 ↾K) and (h1 ↾ L) R (h2 ↾ L). seq enc Given L ∶ Θ → Θ′ and A ⊆ H∅,Θ ∩ H∅,Θ , we write L ⊑R A if for any h ∈ JLKenc seq ′ ′ there exists h ∈ A such that h ⊑R h . Moreover, L ⊑R L′ if L ⊑ JL′ Kenc ∩ HΘ,Θ ′. Note the above permutation-closure requirement: if h R h′ and π is a (type-preserving) permutation on Meths ∖ Θ then π(h) R π(h′ ). The requirement adheres to the fact that, apart from the method names from a library interface, the other method names in its semantics are arbitrary and can be freely permuted without any observable effect. Thus, R in particular should not be distinguishing between such names. Our third example extends the flat-combining case study from [3] by lifting it to higher-order types. Example 17. Flat combining [11] is a synchronisation paradigm that advocates the use of single thread holding a global lock to process requests of all other threads. To facilitate this, threads share an array to which they write the details of their requests and wait either until they acquire a lock or their request has been processed by another thread. Once a thread acquires a lock, it executes all requests stored in the array and the outcomes are written to the shared array for access by the requesting threads. The authors of [3] analysed a parameterised library that reacts to concurrent requests by calling the corresponding abstract methods subject to mutual exclusion. In Figure 4, we present the code adapted to arbitrary higher-order types, which is possible thanks to the presence of higher-order references in our framework. We assume: Θ = {mi ∈ Methsθi ,θi′ ∣ 1 ≤ i ≤ k} Θ′ = {m′i ∈ Methsθi ,θi′ ∣ 1 ≤ i ≤ k} Thus the setup of [3] corresponds to θi = θi′ = int. The library Lfc is built following the flat combining approach and, on acquisition of a lock, the winning thread acts as a combiner of all registered requests. Note that the requests will be attended to one after 9 1 2 3 4 abstract mi ; public m′i ; . . . ; Lock lock; struct {op, parm, wait , retv } requests [N]; 7 8 9 12 13 14 5 6 10 11 m′i = λz. requests [tid ].op := i; requests [tid ].parm := z; requests [tid ].wait := 1; 15 16 17 18 while ( requests [tid ].wait) if ( lock . tryacquire ()) { for ( t=0; t<N; t++) if ( requests [ t ]. wait) { let j = requests [ t ]. op in requests [ t ]. retv := mj (requests[t].parm); requests [ t ]. wait := 0 }; lock . release () }; requests [tid ].retv; Fig. 4. Flat combination library Lfc . another (thus guaranteeing mutual exclusion) and only one lock acquisition will suffice to process one array of requests. In Section 5.2, we shall show that Lfc can be R-linearised to the specification given by the sequential histories of the library Lspec that implements m′i as follows: 6 m′i = λz. ( lock.acquire(); let result = mi (z) in lock.release(); result ) Thus, each abstract call in Lspec is protected by a lock. 3 Library syntax We now look at the concrete syntax of libraries and clients. Libraries comprise collections of typed methods. The order3 of their argument and result types is unrestricted and will adhere to the grammar: θ ∶∶= unit ∣ int ∣ θ → θ ∣ θ × θ. We shall use three disjoint enumerable sets of names, referred to as Vars, Meths and Refs, to name respectively variables, methods and references. x, f (and their decorated variants) will be used to range over Vars; m will range over Meths and r over Refs. Methods and references are implicitly typed, that is, we assume Meths = ⊎θ,θ′ Methsθ,θ′ Refs = Refsint ⊎ ⊎θ,θ′ Refsθ,θ′ where Methsθ,θ′ contains names for methods of type θ → θ′ , Refsint contains names of integer references and Refsθ,θ′ contains names for references to methods of type θ → θ′ . We write ⊎ to stress the disjointness of sets in their union. The syntax for building libraries is defined in Figure 5. Thus, each library L begins with a series of method declarations (public or abstract) followed by a block B consisting of method implementations (m = λx.M ) and reference initialisations (r ∶= i or r ∶= λx.M ). Our typing rules will ensure that each public method must be implemented within the block, in contrast to abstract methods. On the other hand, a client K consists of a parallel composition of closed terms. Terms M specify the shape of allowable method bodies. () is the skip command, i ranges over integers, tid is the current thread identifier and ⊕ represents standard arithmetic operations. Thanks to higher-order references, we can simulate divergence 3 Type order is defined by ord(unit) = ord(int) = 0, ord(θ1 × θ2 ) = max(ord(θ1 ), ord(θ2 )), ord(θ1 → θ2 ) = max(ord(θ1 ) + 1, ord(θ2 )). 10 L ∶∶= B ∣ abstract m; L ∣ public m; L Libraries Clients K ∶∶= M ∥⋯∥ M Blocks B ∶∶=  ∣ m = λx.M ; B ∣ r ∶= λx.M ; B ∣ r ∶= i; B Terms M ∶∶= () ∣ i ∣ tid ∣ x ∣ m ∣ M ⊕ M ∣ if M then M else M ∣ ⟨M, M ⟩ ∣ π1 M ∣ π2 M ∣ λxθ.M ∣ xM ∣ mM ∣ let x = M in M ∣ r ∶= M ∣ !r Γ ⊢ () ∶ unit Γ ⊢ i ∶ int Γ ⊢ tid ∶ int Γ ⊢ M ∶ int Γ ⊢ M0 , M1 ∶ θ Γ ⊢ if M then M1 else M0 ∶ θ Γ ⊢ !r ∶ θ → θ′ Γ ⊢ M ∶ θ Γ, x ∶ θ ⊢ N ∶ θ′ Γ ⊢ let x = M in N ∶ θ′ m ∈ Methsθ,θ′ ⊢B  ∶ ∅ ′ Γ ⊢ xM ∶ θ ′ ⊢B B ∶ Θ ⊢B B ∶ Θ Meths(B) ⊢L B ∶ ∅ → Θ ′ m ∉ Θ′′ Θ ⊢L abstract m; L ∶ Θ ⊎ {m} → Θ′′ Γ ⊢ M1 , M2 ∶ int Γ ⊢ M1 ⊕ M2 ∶ int Γ ⊢ ⟨M1 , M2 ⟩ ∶ θ1 × θ2 Γ ⊢ πi M ∶ θi (i = 1, 2) Γ ⊢M ∶θ ⊢B m = λx.M ; B ∶ Θ ⊎ {m} Θ ⊎ {m} ⊢L L ∶ Θ′ → Θ′′ Γ ⊢ m ∶ θ → θ′ r ∈ Refsθ,θ′ Γ ⊢ M ∶ θ → θ′ Γ ⊢ r ∶= M ∶ unit Γ (x) = θ → θ x ∶ θ ⊢ M ∶ θ′ r ∈ Refsint ⊢B B ∶ Θ ⊢B r ∶= i; B ∶ Θ m ∈ Methsθ,θ′ Γ ⊢ Mi ∶ θi (i = 1, 2) Γ ⊢ ⟨M1 , M2 ⟩ ∶ θ1 × θ2 r ∈ Refsθ,θ′ r ∈ Refsint Γ ⊢ M ∶ int Γ ⊢ r ∶= M ∶ unit Γ (x) = θ Γ ⊢x∶θ m ∈ Methsθ,θ′ r ∈ Refsint Γ ⊢ !r ∶ int Γ, x ∶ θ ⊢ M ∶ θ′ Γ ⊢ λxθ.M ∶ θ → θ′ Γ ⊢M ∶θ Γ ⊢ mM ∶ θ′ r ∈ Refsθ,θ′ x ∶ θ ⊢ M ∶ θ′ ⊢B B ∶ Θ ⊢B r ∶= λx.M ; B ∶ Θ Θ ⊎ {m} ⊢L L ∶ Θ′ → Θ′′ m ∈ Θ′′ ′ Θ ⊢L public m; L ∶ Θ → Θ′′ Meths(Mj ) ⊆ Θ ⊢ Mj ∶ unit (j = 1, ⋯, N ) Θ ⊢K M1 ∥⋯∥MN ∶ unit Fig. 5. Library syntax and typing rules for terms (⊢), blocks (⊢B ), libraries (⊢L ), clients (⊢K ). by (!r)(), where r ∈ Refsunit,unit is initialised with λxunit .(!r)(). Similarly, after r ∶= λxunit .let y = M in (if y then (N ; (!r)()) else ()), while M N can be simulated by (!r)(). We shall also use the standard derived syntax for sequential composition, i.e. M ; N stands for let x = N in M , where x does not occur in M . Remark 18. In Section 2 we used lock-related operations in our example libraries (acquire, tryacquire, release), on the understanding that they can be coded using shared memory. Similarly, the array of Example 17 can be constructed using references. For each term M , we write Meths(M ) for the set of method names occurring in M . We also use the same notation for method names in blocks and libraries. Terms are typed in environments Γ = {x1 ∶ θ1 , ⋯, xn ∶ θn } assigning types to their free variables. Method blocks are typed through judgements ⊢B B ∶ Θ, where Θ ⊆ Meths. The judgments collect the names of methods defined in a block as well as making sure that the definitions respect types and are not duplicated. Also, any initialisation statements will be scrutinised for type compliance. Finally, we type libraries using statements of the form Θ ⊢L L ∶ Θ′ → Θ′′ , where Θ, Θ′ , Θ′′ ⊆ Meths and Θ′ ∩ Θ′′ = ∅. The judgment ∅ ⊢L L ∶ Θ′ → Θ′′ guarantees that any method occurring in L is present either in Θ′ or Θ′′ , that all methods in Θ′ have been declared as abstract and not implemented, while all methods in Θ′′ have been declared 11 as public and defined. Thus, ∅ ⊢L L ∶ Θ → Θ′ stands for a library in which Θ, Θ′ are the abstract and public methods respectively. In this case, we also write L ∶ Θ → Θ′ . Remark 19. For simplicity, we do not include private methods but the same effect could be achieved by storing them in higher-order references. As we explain in the next section, references present in library definitions are de facto private to the library. Note also that, according to our definition, sets of abstract and public methods are disjoint. However, given m, m′ ∈ Refsθ,θ′ , one can define a “public abstract” method with: public m; abstract m′ ; m = λxθ .m′ x . 4 Semantics The semantics of our system will be given in several stages. First, we define an operational semantics for sequential and concurrent terms that may draw methods from a function repository. We then adapt it to capture interactions of concurrent clients with libraries that do not feature abstract methods. The extended notion is then used to define contextual approximation (refinement) for arbitrary libraries. Subsequently, we introduce a trace semantics of arbitrary libraries that will be used to define higher-order notions of linearisability and, ultimately, to relate them to contextual refinement. 4.1 Library-client evaluation Libraries, terms and clients are evaluated in environments comprising: – A method environment R, called own-method repository, which is a finite partial map on Meths assigning to each m in its domain, with m ∈ Methsθ,θ′ , a term of the form λy.M (we omit type-superscripts from bound variables for economy). – A finite partial map S ∶ Refs ⇀ (Z ∪ Meths), called store, which assigns to each r in its domain an integer (if r ∈ Refsint ) or name from Methsθ,θ′ (if r ∈ Refsθ,θ′ ). The evaluation rules are presented in Figure 6. Remark 20. We shall assume that reference names used in libraries are library-private, i.e. sets of reference names used in different libraries are assumed to be disjoint. Similarly, when libraries are being used by client code, this is done on the understanding that the references available to that code do not overlap with those used by libraries. Still, for simplicity, we shall rely on a single set Refs of references in our operational rules. First we evaluate the library to create an initial repository and store. This is achieved by the first set of rules in Figure 6, where we assume that Sinit is empty. Note that m in the last rule is any fresh method name of the appropriate type. Thus, library evaluation produces a tuple (, R0 , S0 ) including a method repository and a store, which can be used as the initial repository and store for evaluating M1 ∥⋯∥MN using the (KN ) rule. We shall call the latter evaluation semantics for clients (denoted by Ô Ô⇒) the multi-threaded operational semantics. Reduction rules rely on evaluation contexts E, defined along with values v in the third group in Figure 6. Finally, rules for closed-term reduction (Ð →t ) are given in the last group, where t is the current thread index. Note that the rule for E[λx.M ] involves the creation of a new method name m, which is used to put the function in the repository R. We define termination for clients linked with libraries that have no abstract methods. 12 (L) Ð→lib (L, ∅, Sinit ) (r ∶= i; B, R, S) Ð→lib (B, R, S[r ↦ i]) (abstract m; L, R, S) Ð→lib (L, R, S) (m = λx.M ; B , R, S) Ð→lib (B, R ⊎ (m ↦ λx.M ), S) (public m; L, R, S) Ð→lib (L, R, S) (r ∶= λx.M ; B, R, S) Ð→lib (B, R ⊎ (m ↦ λx.M ), S∗ ) (M, R, S) Ð →t (M ′ , R′ , S ′ ) (M1 ∥⋯∥Mt−1 ∥M ∥Mt+1 ∥⋯∥MN , R, S) Ô Ô⇒ (M1 ∥⋯∥Mt−1 ∥M ′ ∥Mt+1 ∥⋯∥MN , R′ , S ′ ) (KN ) E ∶∶= ● ∣ E ⊕ M ∣ i ⊕ E ∣ if E then M else M ∣ πj E ∣ ⟨E, M ⟩ ∣ ⟨v, E⟩ ∣ mE ∣ let x = E in M ∣ r ∶= E v ∶∶= () ∣ i ∣ m ∣ ⟨v, v⟩ (Evaluation Contexts and Values) (E[tid ], R, S) Ð →t (E[t], R, S) (E[i1 ⊕ i2 ], R, S) Ð →t (E[i], R, S) (E[!r], R, S) Ð →t (E[S(r)], R, S) (E[if 0 then M1 else M0 ], R, S) Ð →t (E[M0 ], R, S) (i = i1 ⊕ i2 ) (E[r ∶= v], R, S) Ð →t (E[()], R, S∗∗ ) (E[if i∗ then M1 else M0 ], R, S) Ð →t (E[M1 ], R, S) (E[λx.M ], R, S) Ð →t (E[m], R∗ , S) (E[mv], R∗∗ , S) Ð →t (E[M {v/x}], R∗∗ , S) (E[πj ⟨v1 ,v2 ⟩], R, S) Ð →t (E[vj ], R, S) (E[let x = v in M ], R, S) Ð →t (E[M {v/x}], R, S) Fig. 6. Evaluation rules for libraries (Ð→lib ), clients (Ô Ô⇒), and terms (Ð →t ). Here, S∗ = S[r ↦ m], S∗∗ = S[r ↦ v], R∗ = R ⊎ (m ↦ λx.M ); and i∗ ≠ 0, R∗∗ (m) = λx.M . Definition 21. Let L ∶ ∅ → Θ′ and Θ′ ⊢K M1 ∥⋯∥MN ∶ unit.4 We say that M1 ∥⋯∥MN terminates with linked library L if (M1 ∥⋯∥MN , R0 , S0 ) Ô Ô⇒∗ (()∥⋯∥(), R, S) for some R, S, where (L) Ð→∗lib (, R0 , S0 ). Then we write link L in (M1 ∥⋯∥MN ) ⇓. We shall build a notion of contextual approximation of libraries on top of termination: one library will be said to approximate another if, whenever the former terminates when composed with any parameter library and client, so does the latter. There are several ways of composing libraries. Here we will be considering the notions of union and sequencing. The latter is derived from the former with the aid of a third construct, called hiding. Below, we denote a library L as L = D; B, where D contains all the (public/abstract) method declarations of L, and B is its method block. Definition 22. [Library union, hiding, sequencing] Let L1 ∶ Θ1 → Θ2 be a library of the form D1 ; B1 . – Given library L2 ∶ Θ1′ → Θ2′ (= D2 ; B2 ) which accesses disjoint parts of the store from L1 and such that Θ2 ∩ Θ2′ = ∅, we define the union of L1 and L2 as: L1 ∪L2 ∶ (Θ1 ∪Θ1′ )∖(Θ2 ∪Θ2′ ) → Θ2 ∪Θ2′ = (D1 ; B1 )∪(D2 ; B2 ) = D1′ ; D2′ ; B1 ; B2 where D1′ is D1 with any “abstract m” declaration removed for m ∈ Θ2′ ; dually for D2′ . 4 Recall our convention (Remark 20) that L and M1 , ⋯, MN must access disjoint parts of the store. Terms M1 , ⋯, MN can share reference names, though. 13 – Given some Θ = {m1 , ⋯, mn } ⊆ Θ2 , we define the Θ-hiding of L1 as: L1 ∖ Θ ∶ Θ1 → (Θ2 ∖ Θ) = (D1 ; B1 ) ∖ Θ = D1′ ; B1′ {!r1 /m1 }⋯{!rn /mn } where D1′ is D1 without public m declarations for m ∈ Θ and, for each i, ri is a fresh reference matching the type of mi , and B1′ is obtained from B1 by replacing each definition mi = λx.M by ri ∶= λx.M . The sequencing of L′ ∶ ∅ → Θ1 , Θ′ with L1 is: L′ ; L1 ∶ ∅ → Θ2 , Θ′ = (L′ ∪ L1 ) ∖ Θ1 . Thus, the union of two libraries L1 and L2 as above corresponds to merging their code and removing any abstract method declarations for methods defined in the union. On the other hand, the hiding of a public method simply renders it private via the use of references. These notions are used in defining contextual notions for libraries, that is, notions that require quantification over all possible contexts. Definition 23. Given L1 , L2 ∶ Θ → Θ′ , we say that L1 contextually approximates L2 , written L1 ⊏∼ L2 , if for all L′ ∶ ∅ → Θ, Θ′′ and Θ′ , Θ′′ ⊢K M1 ∥⋯∥MN ∶ unit, if link L′ ; L1 in (M1 ∥⋯∥MN ) ⇓ then link L′ ; L2 in (M1 ∥⋯∥MN ) ⇓. In this case, we also say that L2 contextually refines L1 . Note that, according to this definition, the parameter library L′ may communicate directly with the client terms through a common interface Θ′′ . We shall refer to this case as the general case. Later on, we shall also consider more restrictive testing scenarios in which this possibility of explicit communication is removed. Moreover, from the disjointness conditions in the definitions of sequencing and linking we have that Li , L′ and M1 ∥⋯∥MN access pairwise disjoint parts of the store. Remark 24. Our definition of contextual approximation models communication between the client and the parameter library explicitly through the shared interface Θ′′ . This is different in style (but not in substance) from [3], where the presence of public abstract methods inside the tested library provides such a communication channel. 4.2 Trace semantics Building on the earlier operational semantics, we next introduce a trace semantics of libraries, in the spirit of game semantics [1]. As mentioned in Section 2, the behaviour of a library will be represented as an exchange of moves between two players called O and P , representing the library (P ) and the corresponding context (O) respectively. The context consists of the client of the library as well as the parameter library, with an index on each move specifying which of them is involved in the move (K or L respectively). In contrast to the semantics of the previous section, we will handle scenarios in which methods need not be present in the repository R. Calls to such undefined methods will be represented by labelled transitions – calls to the context made on behalf of the library (P ). The calls can later be responded to with labelled transitions corresponding to returns, made by the context (O). On the other hand, O will be able to invoke methods in R, which will also be represented through suitable labels. Because we work in a higher-order setting, calls and returns made by both players may involve methods as arguments or results. Such methods also become available for future calls: function 14 arguments/results supplied by P are added to the repository and can later be invoked by O, while function arguments/results provided by O can be queried in the same way as abstract methods. After giving semantics to libraries, we shall also define a semantics for contexts, i.e. clients paired with parameter libraries where the main library is missing. More precisely, given a parameter library L′ ∶ ∅ → Θ, Θ′ and client Θ′ , Θ′′ ⊢K M1 ∥⋯∥MN ∶ unit, we will define the semantics of M1 ∥⋯∥MN when paired with L′ . In such a scenario, the roles of O and P will be reversed: P K will own moves played by the client, P L will be the parameter library, while O will correspond to the missing main library (OK when interacting with the client, and OL when talking to the parameter library). Recall the notions of history (Def. 2) and history complementation (Def. 15). We next define a dual notion of history that is used for assigning semantics to contexts. co Definition 25. The set of co-histories over Θ → Θ′ is: HΘ,Θ ′ = {h ∣ h ∈ HΘ,Θ ′ }. co We shall range over HΘ,Θ ′ again using variables h, s. We can show the following. co Lemma 26. – For all h ∈ HΘ,Θ′ we have h ↾ L ∈ H∅,Θ and h ↾ K ∈ H∅,Θ′ . co co – For all h ∈ HΘ,Θ′ we have h ↾ L ∈ H∅,Θ and h ↾ K ∈ H∅,Θ ′. The trace semantics will utilise configurations that carry more components than the previous semantics, in order to compensate for the fact that we need to keep track of the evaluation history that is currently processed, as well as the method names that have been passed between O and P . We define two kinds of configurations: O-configurations (E, −, R, P, A, S) and P-configurations (E, M, R, P, A, S) , where the component E is an evaluation stack, that is, a stack of the form [X1 , X2 , ⋯, Xn ] with each Xi being either an evaluation context or a method name. On the other hand, P = (PL , PK ) with PL , PK ⊆ dom(R) being sets of public method names, and A = (AL , AK ) is a pair of sets of abstract method names. P will be used to record all the method names produced by P and passed to O: those passed to OK are stored in PK , while those leaked to OL are kept in PL . Inside A, the story is the opposite one: AK (AL ) stores the method names produced by OK (resp. OL) and passed to P . Consequently, the sets of names stored in PL , Pk , AL , Ak will always be disjoint. Given a pair P as above and a set Z ⊆ Meths, we write P ∪K Z for the pair (PL , PK ∪ Z). We define ∪L in a similar manner, and extend it to pairs A as well. Moreover, given P and A, we let φ(P, A) be the set of fresh method names for P, A: φ(P, A) = Meths ∖ (PL ∪ PK ∪ AL ∪ AK ). We next give the rules generating the trace semantics. Note that the rules are parameterised by Y ∈ {K, L}. This parameter will play a role in our treatment of the encapsulated case, as it allows us to track the labels related to interactions with the client and the parameter library respectively. In all of the rules below, whenever we write m(v) or m(v ′ ), we assume that the type of v matches the argument type of m. Internal rule First we embed the internal rules, introduced earlier: if (M, R, S) Ð →t (M ′ , R′ , S ′ ) and dom(R′ ∖ R) consists of names that do not occur in E, A, we have: (E, M, R, P, A, S) Ð →t (E, M ′ , R′ , P, A, S ′ ) (I NT) This includes internal method calls (i.e. (E[mv], R, S) Ð →t ⋯ with m ∈ dom(R) ). 15 P calls In the next family of rules, the library (P ) calls one of its abstract methods (either the original ones or those acquired via interaction). Thus, the rule applies to m ∈ AY . call m(v ′ )P Y (E, E[mv], R, P, A, S) ÐÐÐÐÐÐÐ→t (m ∶∶ E ∶∶ E, −, R′ , P ′ , A, S) (PQ Y) If v does not contain any method names then v ′ = v, R′ = R, P ′ = P. Otherwise, if v contains the (pairwise distinct) names m1 , ⋯, mk , a fresh name m′i ∈ φ(P, A) is created for each method name mi (for future reference to the method), and replaced for mi in v. That is, v ′ = v{m′i /mi ∣ 1 ≤ i ≤ k}. We must also have that the m′i ’s are pairwise distinct (the rule fires for any such m′i ’s), and also R′ = R ⊎ {m′i ↦ λx.mi x ∣ 1 ≤ i ≤ k} and P ′ = P ∪Y {m′1 , ⋯, m′k }. P returns Analogously, the library (P ) may return a result to an earlier call made by the context. This rule is applicable provided m ∈ PY . ret m(v ′ )P Y (m ∶∶ E, v, R, P, A, S) ÐÐÐÐÐÐ→t (E, −, R′ , P ′ , A, S) (PAY) v ′ , R′ , P ′ are subject to the same constraints as in (PQ Y ). O calls The remaining rules are dual and represent actions of the context. Here the context calls a public method: either an original one or one that has been made public later (by virtue of having been passed to the context by the library). Here we require m ∈ PY and R(m) = λx.M . call m(v)OY (E, −, R, P, A, S) ÐÐÐÐÐÐ→t (m ∶∶ E, M {v/x}, R, P, A′ , S) (OQ Y) If v contains the names m1 , ⋯, mk ∈ Meths, it must be the case that mi ∈ φ(P, A), for each i, and A′ = A ∪Y {m1 , ⋯, mk }. O returns Finally, we have rules corresponding to values being returned by the context in response to calls made by the library. In this case we insist on m ∈ AY . ret m(v)OY (m ∶∶ E ∶∶ E, −, R, P, A, S) ÐÐÐÐÐÐ→t (E, E[v], R, P, A′ , S) (OAY) As in the previous case, if m ∈ Meths is present in v then we need m ∈ φ(P, A) and A′ is calculated in the same way. Finally, we extend the trace semantics to a concurrent setting where a fixed number of N -many threads run in parallel. Each thread has separate evaluation stack and term components, which we write as C = (E, X) (where X is a term or “−”). Thus, a configuration now is of the following form, and we call it an N -configuration: (C1 ∥⋯∥CN , R, P, A, S) where, for each i, Ci = (Ei , Xi ) and (Ei , Xi , R, P, A, S) is a sequential configuration. We shall abuse notation a little and write (Ci , R, P, A, S) for (Ei , Xi , R, P, A, S). Also, ⃗ ↦ C ′ ] = C1 ∥⋯∥Ci−1 ∥C ′ ∥Ci+1 ∥⋯∥CN . below we write C⃗ for C1 ∥⋯∥CN and C[i 16 The concurrent traces are produced by the following two rules with the proviso that ⃗ the names freshly produced internally in (PI NT ) are fresh for the whole of C. (Ci , R, P, A, S) Ð →i (C ′ , R′ , P, A, S ′ ) (PI NT) ⃗ R, P, A, S) Ô ⃗ ↦ C ′ ], R′ , P, A, S ′ ) (C, Ô⇒ (C[i xXY (Ci , R, P, A, S) ÐÐ→i (C ′ , R′ , P ′ , A′ , S ′ ) (i,x)XY ⃗ ↦ C ′ ], R′ , P ′ , A′ , S ′ ) ⃗ R, P, A, S) ÔÔÔ⇒ (C[i (C, (PE XT) We can now define the trace semantics of a library L. We call a configuration component Ci final if it is in one of the following forms: Ci = ([], −) or Ci = ([], ()) ⃗ R, P, A, S) final just if C⃗ = for O- and P -configurations respectively. We call (C, C1 ∥⋯∥CN and each Ci is final. Definition 27. For each L ∶ Θ → Θ′ , we define the N -trace semantics of L to be: JLKN = { s ∣ (C⃗0 , R0 , (∅, Θ′ ), (Θ, ∅), S0 ) Ô Ô⇒∗ ρ ∧ ρ final } s where C⃗0 = ([], −)∥⋯∥([], −) and (L) Ð→∗lib (, R0 , S0 ). For economy, in the sequel we might be dropping the index N from JLKN . We conclude this section by providing a semantics for library contexts. Recall that in the definition of contextual approximation the library L ∶ Θ → Θ′ is deployed in a context consisting of a parameter library L′ ∶ ∅ → Θ, Θ′′ and a concurrent composition of client threads Θ′ , Θ′′ ⊢ Mi ∶ unit (i = 1, ⋯, N ). This context makes internal use of methods defined in the parameter library, while it recurs to the trace system for the methods in Θ′ . At the same time, the context provides the methods in Θ in the trace system. We shall write link L′ ; − in (M1 ∥⋯∥MN ), or simply C, to refer to such contexts. We give the following semantics to contexts. Definition 28. Let Θ′ , Θ′′ ⊢K M1 ∥⋯∥MN ∶ unit and L′ ∶ ∅ → Θ, Θ′′ . We define the semantics of the context formed by L′ and M1 , ⋯, MN to be: Jlink L′ ; − in (M1 ∥⋯∥MN )K = { s ∣ (C⃗0 , R0 , (Θ, ∅), (∅, Θ′ ), S0 ) Ô Ô⇒∗ ρ ∧ ρ final } s where (L′ ) Ð→∗lib (, R0 , S0 ) and C⃗0 = ([], M1 )∥⋯∥([], MN ). Lemma 29. For any L ∶ Θ → Θ′ , L′ ∶ ∅ → Θ, Θ′′ and Θ′ , Θ′′ ⊢K M1 ∥⋯∥MN ∶ unit co we have JLKN ⊆ HΘ,Θ′ and Jlink L′ ; − in (M1 ∥⋯∥MN )K ⊆ HΘ,Θ ′. Proof. The relevant sequences of moves are clearly alternating and well-bracketed, when projected on single threads, because the LTS is bipartite (O- and P -configurations) and separate evaluation stacks control the evolution in each thread. Other conditions for histories follow from the partitioning of names into AK , AL , PK , PL and suitable initialisation: Θ, Θ′ are inserted into AL , PK respectively (for JLK) and into PL , AK for JCK. 17 5 Examples We now revisit the example libraries from Section 2 and show they each linearise to their respective specification, according to the relevant notion of linearisability (general/encapsulated/relational). 5.1 Multiset examples Recall the multiset library Lmset and the specification Amset of Example 6 and Figure 2. We show that Lmset ⊑ Amset . More precisely, taking an arbitrary history h ∈ JLmset K we show that h can be rearranged using ◁∗P O to match an element of Amset . We achieve this by identifying, for each O-move (t, x)O and its following P -move (t, x′ )P in h, a linearisation point between them, i.e. a place in h to which (t, x)O can moved right and to which (t, x′ )P can be moved left so that they become consecutive and, moreover, the resulting history is still produced by Lmset . After all these rearrangements, we obtain a sequential history ĥ such that h ⊑ ĥ and ĥ is also produced by Lmset . It then suffices to show that ĥ ∈ Amset . Lemma 30 (Multiset). Lmset linearises to Amset . Proof. Taking an arbitrary h ∈ JLmset K, we demonstrate the linearisation points for pairs of (O, P ) moves in h, by case analysis on the moves (we drop K indices from moves as they are ubiquitous). Let us assume that h has been generated by a sequence ρ1 ⇒ ρ2 ⇒ ⋯ ⇒ ρk of atomic transitions and that the variable F of Lmset is instantiated with the reference rF . Line numbers used below will refer to Figure 2. 1. h = ⋯ (t, call cnt(i))O s (t, ret cnt(j))P ⋯ . Here the linearisation point is the configuration ρi that dereferences rF as per line 6 in Lmset (the !F expression). 2. h = ⋯ (t, call upd(i, m))O s (t, call m(j))P ⋯ . The linearisation point is the dereferencing of rF in line 8. 3. h = ⋯ (t, ret m(j ′ ))O s (t, ret upd(∣j ′ ∣))P ⋯ . The linearisation point is the update to rF in line 14. 4. h = ⋯ (t, ret m(j ′ ))O s (t, call m(j ′′ ))P ⋯ . The linearisation point is the dereferencing of rF in line 12. Each of the linearisation points above specifies a P O-rearrangement of moves. For instance, for h = s0 (t, call cnt(i))O s (t, ret cnt(j))P s′ , let s = s1 s2 where s0 s1 is the prefix of h produced by ρ1 ⇒ ρ2 ⇒ ⋯ ⇒ ρi . The rearrangement of h is then the history ĥ = s0 s1 (t, call upd(i, m))O (t, call m(j))P s2 s′ . We thus obtain h ◁∗P O ĥ. The selection of linearisation points is such that it guarantees that ĥ ∈ JLmset K. E.g. in case 1, the transitions occurring in thread t between the configuration that follows (t, call cnt(i))O and ρi do not access rF . Hence, we can postpone them and fire them in sequence just ρi . After ρi+1 and until (t, ret cnt(j))P there is again no access of rF in t and we can thus bring forward the corresponding transitions just after ρi+1 . Similar reasoning applies to case 2. In case 4, we reason similarly but also take into account that rendering the acquisition of the lock by t atomic is sound (i.e. the semantics can produce the rearranged history). Case 3 is similar, but we also use the fact that the access to rF in lines 11-16 is inside the lock, and hence postponing dereferencing (line 12) to occur in sequence before update (line 14) is sound. 18 Now, any transition sequence α which produces ĥ (in JLmult K) can be used to derive an extended history h○ ∈ A○mult , by attaching to each move in ĥ the multiset represented x in the configuration that produces the move (ρ produces the move x if ρ Ô Ô⇒ ρ′ in α). By projection we then obtain ĥ ∈ Amult . On the other hand, the multiset library of Example 14 and Figure 3 requires encapsulation in order to linearise (cf. Example 14). Lemma 31 (Parameterised multiset). Lmset2 enc-linearises to Amset2 . Proof. Again, we identify linearisation points, this time for given h ∈ JLmult2 Kenc . For cases 1-4 as above we reason as in Lemma 30. For upd enc we have the following case. h = s (t, call upd enc(i))OK s1 (t, call foo(j))P L s2 (t, ret foo(j ′ ))OL s3 (t, ret upd enc(∣j ′ ∣))P K ⋯ Here, we need a linearisation point for all four moves above. We pick this to be the point corresponding to the update of the multiset reference F on line 24 (Figure 3). We now transform h to ĥ so that the four moves become consecutive, in two steps: – Let us write s3 as s3 = s13 s23 , where the split is at the linearisation point. Since the lock is constantly held by thread t in s2 s13 , there can be no calls or returns to foo in s2 s13 . Hence, all moves in s2 s13 are in component K and can be transposed with the L-moves above, using ◇∗ , to obtain h′ = s (t, call upd enc(i))OK s1 s2 s13 (t, call foo(j))P L (t, ret foo(j ′ ))OL s23 (t, ret upd enc(∣j ′ ∣))P K ⋯ – Next, by P O-rearrangement we obtain ĥ = s s1 s2 s13 (t, call upd enc(i))OK (t, call foo(j))P L (t, ret foo(j ′ ))OL (t, ret upd enc(∣j ′ ∣))P K s23 ⋯ . Thus, h(◁P O ∪ ◇)∗ ĥ. To prove that ĥ ∈ Amult2 we work as in Lemma 30, i.e. via showing that ĥ ∈ JLmult2 Kenc . For the latter, we rely on the fact that the linearisation point was taken at the reference update point (so that any dereferencings from other threads are preserved), and that the dereferences of lines 22 and 23 are within the same lock as the update. 5.2 Flat combining Recall the libraries Lfc and Lspec from Example 17 and Figure 4. We shall investigate the impact of introducing higher-order types to the flat combining algorithm, which will lead to several surprising discoveries. First of all, let us observe that, even if θi = θi′ = int, higher-order interactions of both Lspec and Lfc with a client and parameter library according to Definition 23 (general case) may lead to deadlock. In this case, a client can communicate with the parameter library (via Θ′′ ) and, for example, he could supply it with a function that calls a public method of the library, say, m′i . That function could then be used to implement mi and, consequently, a client call to m′i would result in lock acquisition, then a call to mi , which would trigger another call to m′i and an attempt to acquire the same lock, while mi cannot return (cf. Example 4). Deadlock can also arise in the encapsulated case (Definition 37) if the library contains a public method, say m′i , with functional arguments. Then the client can pass a function that calls m′i as an argument to m′i and, if the abstract method mi subsequently calls the argument, deadlock would follow in the same way as before. Correspondingly, in these cases there exist sequences of transitions induced by our transition system that cannot 19 be extended to a history, e.g. for θi = unit → unit this happens after (1, call m′ (v))OK (1, call m(v ′ )P L ) (1, call v ′ ())OL (1, call v())P K (1, call m′ (v ′′ ))OK . Consequently, the protocol should not really be used in an unrestricted higher-order setting. However, the phenomenon described above cannot be replicated in the encapsulated case provided the argument types are ground (θi = unit, int). In this case, without imposing any restrictions on the result types θi′ , we shall show that Lfc ⊑R Lspec , where R stands for thread-name invariance. Note that this is a proper extension of the result in [3], where θi′ had to be equal to unit or int. It is really necessary to use the finer notion of ⊑R here, as we do not have Lfc ⊑enc Lspec (a parameter library that is sensitive to thread identifiers may return results that allow one to detect that a request has been handled by a combiner thread which is different from the original one). Lemma 32 (Flat combining). Let Θ = {mi ∈ Methsint,θi′ ∣ 1 ≤ i ≤ k} and Θ′ = {m′i ∈ Methsint,θi′ ∣ 1 ≤ i ≤ k} be such that Θ ∩ Θ′ = ∅. Let R consist of pairs (h1 , h2 ) ∈ H∅,Θ × H∅,Θ that are identical once thread identifiers are ignored. Then Lfc ⊑R Lspec . Proof. Observe that histories from JLspec Kenc feature threads built from segments of one of the three forms (we suppress integer arguments for economy): – (t, call m′i )OK (t, call mi )P L (t, ret mi (v))OL (t, ret m′i (v ′ ))P K , or – (t′ , call w(v))OY (t′ , call w′ (v ′ ))P Y ′ for Y ≠ Y ′ , where w is a name introduced in an earlier move (t′′ , x)P Y and w′ is a corresponding name introduced by the move preceding (t′′ , x)P Y , or – (t′ , ret w′ (v ′′ ))OY ′ (t′ , ret w(v ′′′ ))P Y such that a segment (t′ , call w(v))OY (t′ , call w′ (v ′ ))P Y ′ already occurred earlier. We shall refer to moves in the second and third kind of segments as inspection moves and write φ to refer to sequences built exclusively from such sequences (we will use φK and φL if we want to stress that the moves are exclusively tagged with K or L). Let us write X for the subset of JLspec Kenc containing plays of the form: (t0 , call m′i0 )(t0 , call mi0 )(t0 , ret mi0 (v0 ))(t0 , ret m′i0 (v0′ )) φ0 (t1 , call m′i1 )(t1 , call mi1 ) φ1 (t1 , ret mi1 (v1 ))(t1 , ret m′i1 (v1′ )) φ2 ⋯ (tk , call m′ik )(tk , call mik ) φ2k−1 (tk , ret mik (vk ))(tk , ret m′ik (vk )) φ2k where each φi contains moves alternating between O and P . Let h1 ∈ JLfc Kenc . Threads in h1 are built from blocks of the shapes: (t, call m′i )OK ((t, call mj )P L (t, ret mj (v)OL ))∗ (t, ret m′i (v ′ ))P K or (t′ , call w(v))OY (t′ , call w′ (v ′ ))P Y ′ or (t′ , ret w′ (v ′′ ))OY (t′ , ret w(v ′′′ ))P Y ′ . In the first case, the j’s and v’s are meant to represent different values in each iteration. In the second kind of block, w needs to be introduced earlier by some (t′′ , x)P Y move and w′ is then a name introduced by the preceding move. For the third kind, an earlier calling sequence of the second kind must exist in the same thread. Note that, due to locking and sequentiality of loops, h1 ↾ L takes the shape: L L (t0 , call mi0 )P (t0 , ret mi0 (v0 ))O φL 0 (t1 , call mi1 )P φ1 (t1 , ret mi1 (v1 ))O φ2 L L ⋯ (tk , call mik )P φ2k−1 (tk , ret mik (vk ))O φ2k 20 Each segment Sj = (tj , call mij ) φL 2j−1 (tj , ret mij (vj )) in h1 ↾ L must be preceded (in h1 ) by a corresponding public call (t′j , call m′j )OK and followed by a matching return (t′j , ret m′j (wj ))P K , where t′j need not be equal to tj . Note that there can be no other moves from t′j separating the two moves in h1 ↾ K. Let h′1 be obtained from h1 via the following operations around each segment Sj : – move (t′j , call m′ij ) right to precede (tj , call mij ), – move (t′j , ret m′ij (vj′ )) left to follow (tj , ret mi (vj )), – rename (tj , call mi )P L (tj , ret mij (vj ))OL to (t′j , call mij )P L (t′j , ret mij (vj ))OL . We stress that the changes are to be performed simultaneously for each segment Sj . The rearrangements result in a library history, because they bring forward the points at which various vj , vj′ have been introduced and, thus, inspection moves are legal. Then we have h1 ⊑R h′1 , i.e. (h1 ↾ L)R(h′1 ↾ L) and (h1 ↾ K) ◁∗P O (h′1 ↾ K). The former follows from the multiple renaming of the moves originally tagged with tj to t′j and the fact that their order in h1 ↾ L is unaffected. The latter holds, because O-moves move right and P -moves move left past other moves in h1 ↾ K that are not from the same thread. To conclude, we show that there exists h2 ∈ X such that (h′1 ↾ L) = (h2 ↾ L) and (h′1 ↾ K) ◁∗P O (h2 ↾ K). We can obtain h2 by rearranging inspection moves in different threads of h′1 so that they alternate between O and P . Since the inspection moves come in pairs this can simply be done by bringing the paired moves next to each other. Because one of them is always from K and the other from L, this can be achieved by rearranging moves in h′1 ↾ K only: if the O-move is from K it can be moved to the right, otherwise the P -move from K can be moved left. Then we have h2 ∈ X ⊆ JLspec Kenc with h1 ⊑R h2 , as required. 6 Correctness In this section we prove that the linearisability notions we introduce are correct: linearisability implies contextual approximation. The approach is based on showing that, in each case, the semantics of contexts is saturated relatively to conditions that are dual to linearisability. Hence, linearising histories does not alter the observable behaviour of a library. We start by presenting two compositionality theorems on the trace semantics, which will be used for relating library and context semantics. 6.1 Compositionality The semantics we defined is compositional in the following ways: – To compute the semantics of a library L inside a context C, it suffices to compose the semantics of C with that of L, for a suitable notion of context-library composition (JCK ⊘ JLK). – To compute the semantics of a union library L1 ∪ L2 , we can compose the semantics of L1 and L2 , for a suitable notion of library-library composition (JL1 K ⊗ JL2 K). The above are proven using bisimulation techniques for connecting syntactic and semantic compositions, and are presented in Appendices C and D respectively. The latter correspondence is used in Appendix E for proving that linearisability is a congruence for library composition. From the former correspondence we obtain the following result, which we shall use for showing correctness. 21 Theorem 33. Let L ∶ Θ → Θ′ , L′ ∶ ∅ → Θ, Θ′′ and Θ′ , Θ′′ ⊢K M1 ∥⋯∥MN ∶ unit, with L, L′ and M1 ; ⋯; MN accessing pairwise disjoint parts of the store. Then: link L′ ; L in (M1 ∥⋯∥MN ) ⇓ ⇐⇒ ∃h ∈ JLKN . h̄ ∈ Jlink L′ ; − in (M1 ∥⋯∥MN )K 6.2 General linearisability Recall the general notion of linearisability defined in Section 2.2, which is based on move-reorderings inside histories. In Def.s 27 and 28 we have defined the trace semantics of libraries and contexts. The semantics turns out to be closed under ◁∗OP . Lemma 34 (Saturation). Let X = JLK (Def. 27) or X = Jlink L′ ; − in (M1 ∥⋯∥MN )K (Def. 28). Then if h ∈ X and h ◁∗OP h′ then h′ ∈ X. Proof. Recall that the same labelled transition system underpins the definition of X in either case. We make several observations about the single-threaded part of that system. – The store is examined and modified only during -transitions. – The only transition possible after a P -move is an O-move. In particular, it is never the case that a P -move is separated from the following O-move by an -transition. Let us now consider the multi-threaded system and t ≠ t′ . (t′ ,m′ )P ∗ (t,m) – Suppose ρ ÔÔÔÔ⇒ ρ1 Ô Ô⇒ ρ2 ÔÔ⇒ ρ3 . Then the (t′ , m′ )P -transition can be ∗ (t,m) (t′ ,m′ )P delayed inside t′ until after (t, m), i.e. ρ Ô Ô⇒ ρ′1 ÔÔ⇒ ρ′2 ÔÔÔÔ⇒ ρ3 for some ′ ′ ′ ρ1 , ρ2 . This is possible because the ((t , m′ )P -labelled) transition does not access or modify the store, and none of the -transitions distinguished above can be in t′ , thanks to our earlier observations about the behaviour of the single-threaded system. (t′ ,m′ ) ∗ (t,m)O – Analogously, suppose ρ ÔÔÔ⇒ ρ1 Ô Ô⇒ ρ2 ÔÔÔ⇒ ρ3 . Then the (t, m)O -transition (t′ ,m′ ) (t,m)O ∗ can be brought forward, i.e. ρ ÔÔÔ⇒ ρ′1 ÔÔÔ⇒ ρ′2 Ô Ô⇒ ρ3 , because it does not access or modify the store and the preceding -transitions cannot be from t. This, along with the fact that h1 ◁XX ′ h2 ⇐⇒ h2 ◁X ′ X h1 ⇐⇒ h1 ◁X ′ X h2 . lead us to the notion of linearisability defined in Def. 8. Theorem 35. L1 ⊑ L2 implies L1 ⊏∼ L2 . Proof. Consider C such that C[L1 ] ⇓. We need to show C[L2 ] ⇓. Because C[L1 ] ⇓, Theorem 33 implies that there exists h1 ∈ JL1 K such that h1 ∈ JCK. Because L1 ⊑ L2 , there exists h2 ∈ JL2 K with h1 ◁∗P O h2 . Note that h1 ◁∗OP h2 . By Lem. 34, h2 ∈ JCK. Because h2 ∈ JL2 K and h2 ∈ JCK, using Theorem 33 we can conclude C[L2 ] ⇓. Theorem 36. If L1 ⊑ L2 then, for suitably typed L accessing disjoint part of the store than L1 and L2 , we have L ∪ L1 ⊑ L ∪ L2 . 22 6.3 Encapsulated linearisability In this case libraries are being tested by clients that do not communicate with the parameter library explicitly. Definition 37. [Encapsulated ⊏∼ ] Given libraries L1 , L2 ∶ Θ → Θ′ , we write L1 ⊏∼ enc L2 when, for all L′ ∶ ∅ → Θ and Θ′ ⊢K M1 ∥⋯∥MN ∶ unit, if link L′ ; L1 in (M1 ∥⋯∥MN ) ⇓ then link L′ ; L2 in (M1 ∥⋯∥MN ) ⇓. We shall call contexts of the above kind encapsulated, because the parameter library L′ can no longer communicate directly with the client, unlike in Def. 23, where they shared methods in Θ′′ . Consequently, Jlink L′ ; − in (M1 ∥⋯∥MN )K can be decomposed via parallel composition into two components, whose labels correspond to L (parameter library) and K (client) respectively. Lemma 38 (Decomposition). Suppose L′ ∶ ∅ → Θ and Θ′ ⊢K M1 ∥⋯∥MN ∶ unit, where Θ ∩ Θ′ = ∅. Then, setting C ′ ≡ link ∅ ; − in (M1 ∥⋯∥MN ), we have: co ′ ′ Jlink L′ ; − in (M1 ∥⋯∥MN )K = { h ∈ HΘ,Θ ′ ∣ (h ↾ L) ∈ JL K, (h ↾ K) ∈ JC K } . Remark 39. Consider parameter library L′ ∶ ∅ → {m} and client {m′ } ⊢K M ∶ unit with m, m′ ∈ Methsunit→(unit→unit) , and suppose we insert in their context a “copycat” library L which implements m′ as m′ = λx.mx . Then the following scenario may seem to contradict encapsulation: – M calls m′ (); – L calls m(); – L′ returns with m(m′′ ) to L; – and finally L copycats this return to M . However, by definition the latter copycat is done by L returning m′ (m′′′ ) to M , for some fresh name m′′′ , and recording internally that m′′′ ↦ λx.m′′ x. Hence, no methods of L′ can leak to M and encapsulation holds. Because of the above decomposition, the context semantics satisfies a stronger closure property than that already specified in Lem. 34, which in turn leads to the notion of encapsulated linearisability of Def. 12. The latter is defined in term of the symmetric reordering relation ◇, which allows for swaps (in either direction) between moves from different threads if they are tagged with K and L respectively. Moreover, we can show the following. Lemma 40 (Encapsulated saturation). Consider X = Jlink L′ ; − in (M1 ∥⋯∥MN )K (Definition 28). Then: – If h ∈ X and h (◁OP ∪ ◇)∗ h′ then h′ ∈ X. – Let s1 (t, x)OY s2 (t, x′ )P Y ′ s3 ∈ X be such that no move in s2 comes from thread t. Then Y = Y ′ , i.e. inside a thread only O can switch between K and L. Due to Theorem 33, the above property of contexts means that, in order to study termination in the encapsulated case, one can safely restrict attention to library traces satisfying a dual property to the one above, i.e. to elements of JLKenc . Note that JLKenc can be obtained directly from our labelled transition system by restricting its single-threaded part to reflect the switching condition. Observe that Theorem 33 will still hold for JLKenc (instead of JLK), because we have preserved all the histories that are compatible with context histories. We are ready to prove correctness of encapsulated linearisability. 23 Theorem 41. L1 ⊑enc L2 implies L1 ⊏∼ enc L2 . Proof. Similarly to Theorem 35, except we invoke Lemma 40 instead of Lemma 34. We next examine the behaviour of ⊑enc with respect to library composition. In contrast to Section 6.2, we need to restrict composition for it to be compatible with encapsulation. Remark 42. The general case of union does not conform with encapsulation in the sense that encapsulated testing of L ∪ Li (i = 1, 2) according to Def. 37 may subject Li to unencapsulated testing. For example, because method names of L and Li are allowed to overlap, methods in L may call public methods from Li as well as implementing abstract methods from Li . This amounts to L playing the role of both K and L, which in addition can communicate with each other, as both are inside L. Even if we make L and Li non-interacting (i.e. without common abstract/public methods), if higher-order parameters are still involved, the encapsulated tests of L ∪ Li can violate the encapsulation hypothesis for Li . For instance, consider the methods m2 , m′1 , m′2 ∈ Methsunit,unit and m1 ∈ Meths(unit→unit),unit , and libraries L1 , L2 ∶ {m1 } → {m2 } and L ∶ {m′1 } → {m′2 }, as well as the unions L ∪ Li ∶ {m1 , m2 } → {m′1 , m′2 }. A possible trace in JL ∪ Li Kenc is this one: hi = (1, call m2 ())OK (1, call m1 (v))P L (1, ret m1 ())OL (1, ret m2 ())P K (1, call m′2 ())OK (1, call m′1 ())P L (1, call v())OL which decomposes as hi = h′ ⩕σΠ,∅ h′i , with Π = {m1 , m2 , m′1 , m′2 }, σ = 2222112, h′ = (1, call m′2 ())OK (1, call m′1 ())P L and: h′i = (1, call m2 ())OK (1, call m1 (v))P L (1, ret m1 ())OL (1, ret m2 ())P K (1, call v())OL We now see that h′i ∉ JLi Kenc as in the last move O is changing component from K to L. We therefore look at compositionality for two specific cases: encapsulated sequencing (e.g. of L∶ Θ → Θ′ with L′ ∶ Θ′ → Θ′′ ) and disjoint union for first-order methods. Given L ∶ Θ1 → Θ2 and L′ ∶ Θ1′ → Θ2′ , we define their disjoint union L ⊎ L′ = L ∪ L′ ∶ (Θ1 ∪ Θ1′ ) → (Θ2 ∪ Θ2′ ) under the assumption that (Θ1 ∪ Θ2 ) ∩ (Θ1′ ∪ Θ2′ ) = ∅. Theorem 43. Let L1 , L2 ∶ Θ1 → Θ2 and L ∶ Θ1′ → Θ2′ . If L1 ⊑enc L2 then: – assuming Θ2′ = Θ1 , we have L ; L1 ⊑enc L ; L2 and L1 ; L ⊑enc L2 ; L; – if Θ1 , Θ2 , Θ1′ , Θ2′ are first-order then L ⊎ L1 ⊑enc L ⊎ L2 . 6.4 Relational linearisability Finally, we examine relational linearisability (Def. 16). We begin with a suitable notion of relation R. We next restrict encapsulated contextual testing to R-closed contexts. Definition 44. Let R ⊆ H∅,Θ × H∅,Θ be a set closed under permutation of names in Meths ∖ Θ. We say that L ∶ ∅ → Θ is R-closed if, for any h, h′ such that h R h′ , if h ∈ JLK then h′ ∈ JLK. Definition 45. [R-closed encapsulated ⊏∼ ] Given L1 , L2 ∶ Θ → Θ′ , we write L1 ⊏∼ R L2 if, for all R-closed L′ ∶ ∅ → Θ and for all Θ′ ⊢K M1 ∥⋯∥MN ∶ unit, whenever link L′ ; L1 in (M1 ∥⋯∥MN ) ⇓ then we also have link L′ ; L2 in (M1 ∥⋯∥MN ) ⇓. 24 Theorem 46. L1 ⊑R L2 implies L1 ⊏∼ R L2 . We conclude by showing that ⊑R is also compositional in the sense proposed in [3]. )-closed if, for all k ∈ H and h′ ∈ JLKenc , (h′ ↾ Given R, G ⊆ H × H, we say that L is (R G K) R k implies that there is h′′ ∈ JLKenc with (h′′ ↾ K) = k and (h′ ↾ L) G (h′′ ↾ L). Theorem 47. Let R, G ⊆ H × H, L1 , L2 ∶ Θ1 → Θ2 and L ∶ Θ1′ → Θ2′ , such that L1 ⊑R L2 . If L is suitably typed: )-closed, we have L ; L1 ⊑G L ; L2 ; – L1 ; L ⊑R L2 ; L; – if L is (R G – if Θ1 , Θ2 , Θ1′ , Θ2′ are first-order then L ⊎ L1 ⊑R+ L ⊎ L2 , where R+ = {(s, s′ ) ∈ H∅,Θ1 ∪Θ1′ × H∅,Θ1 ∪Θ1′ ∣ (s↾Θ1 )R(s′ ↾Θ1 ), (s↾Θ1′ ) = (s′ ↾Θ1′ )}and s ↾ Θ is the largest subsequence of s belonging to H∅,Θ1 . 7 Related and future work Since the work of Herlihy and Wing [12], linearisability has been consistently used as a correctness criterion for concurrent algorithms on a variety of data structures [18], and has yielded a variety of proof methods [5]. As mentioned in the Introduction, the field has focussed on libraries with methods of base-type inputs and outputs, with Cerone et al. recently catering for the presence of abstract as well as public methods [3]. An explicit connection between linearisability and refinement was made by Filipovic et al. in [6], where it was shown that, in base-type settings, linearisability and refinement coincide. Similar results have been proved in [4,9,17,3]. Our contributions herein are notions of linearisability that can serve as correctness criteria for libraries with methods of arbitrary higher-order types. Moreover, we relate them to refinement, thus establishing the soundness of linearisability, and demonstrate they are well-behaved with respect to library composition. Verification of concurrent higher-order programs has been extensively studied outside of linearisability; we next mention works most closely related to linearisability reasoning. At the conceptual level, [6] proposed that the verification goal behind linearisability is observational refinement. In the same vein, [24] utilised logical relations as a direct method for proving refinement in a higher-order concurrent setting, while [23] introduced a program logic that builds on logical relations. On the other hand, proving conformance to a history specification has been addressed in [20] by supplying historyaware interpretations to off-the-shelf Hoare logics for concurrency. Other logic-based approaches for concurrent higher-order libraries, which do not use linearisability or any other notion of logical atomicity, include Higher-Order and Impredicative Concurrent Abstract Predicates [21,22]. One possible avenue for expansion of this work, following the example of [6], would be to identify language fragments where higher-order linearisability coincides with observational refinement. Based on the game semantic results of [7], such a correspondence may be possible to demonstrate already in the language examined herein. The higher-order language we examined used memory in the form of references, which were global and moreover followed the standard memory model (sequential consistency). Therefore, future research also includes enriching the setting with dynamically allocated memory and expanding its reach to weak memory models. In the latter direction, our traces could need to be strengthened towards truly-concurrent structures, such as event-structures, following the recent examples of [2,14]. 25 References 1. S. Abramsky and G. McCusker. Game semantics. In H. Schwichtenberg and U. Berger, editors, Logic and Computation. Springer-Verlag, 1998. Proceedings of the 1997 Marktoberdorf Summer School. 2. S. Castellan. Weak memory models using event structures. In JFLA, 2016. 3. A. Cerone, A. Gotsman, and H. Yang. Parameterised linearisability. In Proceedings of ICALP’14, volume 8573 of Lecture Notes in Computer Science, pages 98–109. Springer, 2014. 4. J. Derrick, G. Schellhorn, and H. Wehrheim. Mechanically verified proof obligations for linearizability. ACM Trans. Program. Lang. Syst., 33(1):4, 2011. 5. B. Dongol and J. Derrick. Verifying linearisability: A comparative survey. ACM Comput. Surv., 48(2):19, 2015. 6. I. Filipovic, P. W. O’Hearn, N. Rinetzky, and H. Yang. Abstraction for concurrent objects. Theor. Comput. Sci., 411(51-52):4379–4398, 2010. 7. D. R. Ghica and A. S. Murawski. Angelic semantics of fine-grained concurrency. In Proceedings of FOSSACS, volume 2987 of Lecture Notes in Computer Science, pages 211– 225. Springer-Verlag, 2004. 8. D. R. Ghica and N. Tzevelekos. A system-level game semantics. Electr. Notes Theor. Comput. Sci., 286:191–211, 2012. 9. A. Gotsman and H. Yang. Liveness-preserving atomicity abstraction. In Automata, Languages and Programming - 38th International Colloquium, ICALP 2011. Proceedings, Part II, pages 453–465, 2011. 10. S. Heller, M. Herlihy, V. Luchangco, M. Moir, W. N. S. III, and N. Shavit. A lazy concurrent list-based set algorithm. In OPODIS, pages 3–16, 2005. 11. D. Hendler, I. Incze, N. Shavit, and M. Tzafrir. Flat combining and the synchronizationparallelism tradeoff. In Proceedings of SPAA 2010, pages 355–364, 2010. 12. M. Herlihy and J. M. Wing. Linearizability: A correctness condition for concurrent objects. ACM Trans. Program. Lang. Syst., 12(3):463–492, 1990. 13. A. Jeffrey and J. Rathke. A fully abstract may testing semantics for concurrent objects. Theor. Comput. Sci., 338(1-3):17–63, 2005. 14. A. S. A. Jeffrey and J. Riely. On thin air reads: Towards an event structures model of relaxed memory. In Proceedings of LICS, 2016. 15. J. Laird. A game semantics of Idealized CSP. In Proceedings of MFPS’01, pages 1–26. Elsevier, 2001. ENTCS, Vol. 45. 16. J. Laird. A fully abstract trace semantics for general references. In Proceedings of ICALP, volume 4596 of Lecture Notes in Computer Science, pages 667–679. Springer, 2007. 17. H. Liang and X. Feng. Modular verification of linearizability with non-fixed linearization points. In ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’13. Proceedings, pages 459–470, 2013. 18. M. Moir and N. Shavit. Concurrent data structures. In Handbook of Data Structures and Applications. Chapman and Hall/CRC, 2004. 19. P. W. O’Hearn, N. Rinetzky, M. T. Vechev, E. Yahav, and G. Yorsh. Verifying linearizability with hindsight. In PODC, pages 85–94, 2010. 20. I. Sergey, A. Nanevski, and A. Banerjee. Specifying and verifying concurrent algorithms with histories and subjectivity. In Programming Languages and Systems - 24th European Symposium on Programming, ESOP 2015. Proceedings, pages 333–358, 2015. 21. K. Svendsen and L. Birkedal. Impredicative concurrent abstract predicates. In Programming Languages and Systems - 23rd European Symposium on Programming, ESOP 2014. Proceedings, pages 149–168, 2014. 26 22. K. Svendsen, L. Birkedal, and M. J. Parkinson. Joins: A case study in modular specification of a concurrent reentrant higher-order library. In ECOOP 2013 - Object-Oriented Programming - 27th European Conference. Proceedings, pages 327–351, 2013. 23. A. Turon, D. Dreyer, and L. Birkedal. Unifying refinement and hoare-style reasoning in a logic for higher-order concurrency. In ACM SIGPLAN International Conference on Functional Programming, ICFP’13, pages 377–390, 2013. 24. A. J. Turon, J. Thamsborg, A. Ahmed, L. Birkedal, and D. Dreyer. Logical relations for finegrained concurrency. In The 40th Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’13, pages 343–356, 2013. 27 A Big-step vs small-step reorderings [3] defines linearisation in the general case using a “big-step” relation that applies a single permutation to the whole sequence. This contrasts with our definition as ◁∗P O , in which we combine multiple adjacent swaps. We show that the two definitions are equivalent. Definition 48. [[3]] Let h1 , h2 ∈ HΘ,Θ′ of equal length. We write h1 ◁big P O h2 if there is a permutation π ∶ {1, ⋯, ∣h1 ∣} → {1, ⋯, ∣h2 ∣} such that, writing hi (j) for the j-th element of hi : for all j, we have h1 (j) = h2 (π(j)) and, for all i < j: ((∃t. h1 (i) = (t, −) ∧ h1 (j) = (t, −)) ∨(∃t1 , t2 . h1 (i) = (t1 , −)P ∧ h1 (j) = (t2 , −)O )) Ô⇒ h2 (i) < h2 (j) In other words, h2 is obtained from h1 by permuting moves in such a way that their order in threads is preserved and whenever a O-move occurred after an P -move in h1 , the same must apply to their permuted copies in h2 . ∗ Lemma 49. ◁big P O = ◁P O . Proof. It is obvious that ◁∗P O ⊆ ◁big P O , so it suffices to show the converse. Suppose h1 ◁big h . Consider the set Xh1 ,h2 = {h ∣ h1 ◁∗P O h, h ◁big PO 2 P O h2 }. Note that Xh1 ,h2 is not empty, because h1 ∈ Xh1 ,h2 . For two histories h′ , h′′ , define δ(h′ , h′′ ) to be the length of the longest common prefix of h′ and h′′ . Let N = max{δ(h, h2 ) ∣ h ∈ Xh1 ,h2 }. Note that N ≤ ∣h1 ∣ = ∣h2 ∣. h – If N = ∣h2 ∣ then we are done, because N = ∣h2 ∣ implies h2 ∈ Xh1 ,h2 and, thus, h1 ◁∗P O h2 . – Suppose N < ∣h2 ∣ and consider h such that N = δ(h, h2 ). We are going to arrive at a contradiction by exhibiting h′ ∈ Xh1 ,h2 such that δ(h′ , h2 ) > N . Because N = δ(h, h2 ) and N < ∣h2 ∣, we have h2 = a1 ⋯aN (t, m)u h = a1 ⋯aN (t1 , m1 )⋯(tk , mk )(t, m)u′ , where ti ≠ t, because order in threads must be preserved. Consider h′ = a1 ⋯aN (t, m)(t1 , m1 )⋯(tk , mk )u′ . Clearly δ(h′ , h2 ) > N so, for a contradiction, it suffices to show that h′ ∈ Xh1 ,h2 . big ′ Note that because h ◁big P O h2 , we must also have h ◁P O h2 , because the new P O ′ dependencies in h (wrt h) caused by moving (t, m) forward are consistent with h2 . Hence, we only need to show that h ◁∗P O h′ . Let us distinguish two cases. ● If (t, m) is a P -move then, clearly, h ◁∗P O h′ (P -move moves forward). ● If (t, m) is an O-move then, because h ◁big P O h2 , all of the (ti , mi ) actions must be O-moves (otherwise their position wrt (t, m) would have to be preserved in h2 and it isn’t). Hence, h ◁∗P O h′ , as required. 28 B Proofs from Section 6 Proof of Lemma 40. For the first claim, closure under ◁OP (resp. ◇) follows from Lemma 34. (resp. Lemma 38). Suppose h = s1 (t, x)OY s2 (t, x′ )P Y ′ s3 violates the second claim and (t, x), (t, x′ ) is the earliest such violation in h, i.e. no violations occur in s1 . Observe that then h restricted to moves of the form (t, z)XY ′ would not be alternating, which contradicts the fact that h ↾ Y ′ is a history (Lemma 38). Proof of Theorem 46. Consider C such that C[L1 ] ⇓. We need to show C[L2 ] ⇓. Since C[L1 ] ⇓, by Theorem 33 there exists h1 ∈ JL1 Kenc such that h1 ∈ JCKenc . Also, by Lemma 38, (h1 ↾ K) ∈ JC ′ Kenc and (h1 ↾ L) ∈ JL′ Kenc for C ′ , L′ specified in that lemma. Because L1 ⊑R L2 , there exists h2 ∈ JL2 Kenc such that (h1 ↾ K) ◁∗P O (h2 ↾ K) and (h1 ↾ L) R (h2 ↾ L). Note that the former implies (h1 ↾ K) ◁∗OP (h2 ↾ K). Because (h1 ↾ L) ∈ JL′ Kenc , (h1 ↾ L) R (h2 ↾ L) and L′ is R-closed, we have (h2 ↾ L) ∈ JL′ Kenc . On the other hand, because (h1 ↾ K) ∈ JC ′ Kenc and (h1 ↾ K) ◁∗OP (h2 ↾ K) Lemma 40 implies (h2 ↾ K) ∈ JC ′ Kenc . Consequently, (h2 ↾ K) ∈ JC ′ Kenc and (h2 ↾ L) ∈ JL′ Kenc , so Lemma 38 entails h2 ∈ JCKenc . Hence, since h2 ∈ JL2 K and h2 ∈ JCKenc , we can conclude C[L2 ] ⇓ by Theorem 33. C Trace compositionality In this section we demonstrate how the semantics of a library inside a context can be drawn by composing the semantics of the library and that of the context. The result played a crucial role in our arguments about linearisability and contextual refinement in Section 6. Let us divide (reachable) evaluation stacks into two classes: L-stacks, which can be produced in the trace semantics of a library; and C-stacks, which appear in traces of a context. EL ∶∶= [] ∣ m ∶∶ E ∶∶ EL′ EC ∶∶= [] ∣ m ∶∶ EC′ EL′ EC′ ∶∶= m ∶∶ E ∶∶ EC ∶∶= m ∶∶ EL From the trace semantics definition we have that N -configurations in the semantics of a library feature evaluation stacks of the forms EL (in O-configurations) and EL′ (in P -configurations): these we will call L-stacks. On the other hand, those produced from a context utilise C-stacks which are of the forms EC (in P -configurations) and EC′ (in O-configurations). From here on, when we write E we will mean an L-stack or a C-stack. Moreover, we ⃗ ⋯) will call an N -configuration ρ an L-configuration (or a C-configuration), if ρ = (C, and, for each i, Ci = (Ei , ⋯) with Ei an L-stack (resp. a C-stack). ⃗ R, P, A, S) is a C-configuration Let ρ, ρ′ be N -configurations and suppose ρ = (C, ′ ′ ′ ′ ′ ′ ⃗ and ρ = (C , R , P , A , S ) an L-configuration. We say that ρ and ρ′ are compatible, written ρ ≍ ρ′ , if S and S ′ have disjoint domains and, for each i: – Ci = (EC , M ) and Ci′ = (EL , −), or Ci = (EC′ , −) and Ci′ = (EL′ , M ). – If the public and abstract names of Ci are (PL , PK ) and (AL , AK ) respectively, and ′ those of Ci′ are (PL′ , PK ) and (A′L , A′K ), then PL = A′L , PK = A′K , AL = PL′ and ′ AK = PK . 29 – The private names of ρ (i.e. those in dom(R) ∖ PL ∖ PK ) do not appear in ρ′ , and dually for the private names of ρ′ . – If Ci = (E, ⋯) and Ci′ = (E ′ , ⋯) then E and E ′ are in turn compatible, that is: ● either E = m ∶∶ E ∶∶ E1 , E ′ = m ∶∶ E1′ and E1 , E1′ are compatible, ● or E = m ∶∶ E1 , E ′ = m ∶∶ E ∶∶ E1′ and E1 , E1′ are compatible, or E = E ′ = []. Note, in particular, that if ρ ≍ ρ′ then ρ must be a context configuration, and ρ′ a library configuration. We next define a trace semantics on compositions of compatible such N -configurations. We use the symbol ⊘ for configuration composition: we call this external composition, to distinguish it from the composition of ρ and ρ′ we can obtain by merging their components, which we will examine later. ρ1 Ô Ô⇒ ρ′1 I NT1 ρ1 ⊘ ρ2 Ð→ ρ′1 ⊘ ρ2 ρ2 Ô Ô⇒ ρ′2 I NT2 ρ1 ⊘ ρ2 Ð→ ρ1 ⊘ ρ′2 (t,call m(v)) (t,call m(v)) (t,ret m(v)) (t,ret m(v)) ρ1 ÔÔÔÔÔ⇒ ρ′1 ρ2 ÔÔÔÔÔ⇒ ρ′2 C ALL ρ1 ⊘ ρ2 Ð→ ρ′1 ⊘ ρ′2 ρ1 ÔÔÔÔÔ⇒ ρ′1 ρ2 ÔÔÔÔÔ⇒ ρ′2 R ETN ρ1 ⊘ ρ2 Ð→ ρ′1 ⊘ ρ′2 The I NT rules above have side-conditions imposing that the resulting pairs of configurations are still compatible. Concretely, this means that the names created fresh in internal transitions do not match the names already present in the configurations of the other component. Note that external composition is not symmetric, due to the context/library distinction we mentioned. Our next target is to show a correspondence between the above-defined semantic composition and the semantics obtained by (syntactically) merging compatible configurations. This will demonstrate that composing the semantics of two components is equivalent to first syntactically composing them and then evaluating the result. In order to obtain this correspondence, we need to make the semantics of syntactically composed configurations more verbose: in external composition methods belong either to the context or the library, and when e.g. the client wants to evaluate mm′ , with m a library method, the call is made explicit and, more importantly, m′ is replaced by a fresh method name. On the other hand, when we compose syntactically such a call will be done internally, and without refreshing m′ . To counter-balance the above mismatch, we extend the syntax of terms and evaluation contexts, and the operational semantics of closed terms as follows. The semantics will now involve quadruples of the form: ⃗ S) (E[M ], R1 , R2 , S) written also (E[M ], R, where the two repositories correspond to context and library methods respectively, so in particular dom(R1 ) ∩ dom(R2 ) = ∅. Moreover, inside E[M ] we tag method names and lambda-abstractions with indices 1 and 2 to record which of the two components (context or library) is enclosing them: the tag 1 is used for the context, and 2 for the 30 library. Thus e.g. a name m1 signals an occurrence of method m inside the context. Tagged methods are passed around and stored as ordinary methods, but their behaviour changes when they are applied. Moreover, we extend (tagged) evaluation contexts by explicitly marking return points of methods: E ∶∶= ● ∣ ⋯ ∣ let x = E in M ∣ mE ∣ r ∶= E ∣ ⟨mi ⟩ E In particular, E[M ] may not necessarily be a (tagged) term, due to the return annotations. The new reduction rules are as follows (we omit indices when they are not used in the rules). ⃗ S) Ð ⃗ S′) (E[i1 ⊕ i2 ], R, →′t (E[i], R, ⃗ S) Ð ⃗ S′) (E[tid ], R, →′ (E[t], R, (i = i1 ⊕ i2 ) t ⃗ S) Ð ⃗ S′) (E[πj ⟨v1 , v2 ⟩], R, →′t (E[vj ], R, ⃗ S) Ð ⃗ S) (j = (i > 0)) (E[if i then M0 else M1 ], R, →′t (E[Mj ], R, ⃗ S) Ð ⃗ ⊎i (m ↦ λx.M ), S) (E[λi x.M ], R, →′ (E[mi ], R t ⃗ S) Ð ⃗ S) if Ri (m) = λx.M (E[mi v], R, →′t (E[M {v/x}i ], R, i ′ i ′ ⃗ S) Ð ⃗ ′ , S) if R3−i (m) = λx.M with (E[m v], R, →t (E[⟨m ⟩ M {v /x}3−i ], R ⃗′ = R ⃗ ⊎i {m′ ↦ λy.mj y ∣ 1 ≤ j ≤ k} Meths(v) = {m1 , ⋯, mk }, v ′ = v{ m′j /mj ∣ 1 ≤ j ≤ k}, R j i ′ ′i ′ ′ ⃗ ⃗ (E[⟨m ⟩ v], R, S) Ð → (E[v ], R ⊎3−i {m ↦ λy.mj y}, S) with mj , m and v ′ as above t j j ⃗ S) Ð ⃗ S) (E[let x = v in M ], R, →′t (E[M {v/x}], R, ′ ⃗ S) Ð ⃗ S) (E[!r], R, → (E[S(r)], R, t ⃗ S) Ð ⃗ S[r ↦ i]) (E[r ∶= i], R, →′t (E, R, ⃗ S) Ð ⃗ S[r ↦ mi ]) (E[r ∶= mi ], R, →′ (E, R, t Above we write M i for the term M with all its methods and lambdas tagged (or re⃗ ⊎1 (m ↦ λx.M ) = (R1 ⊎ (m ↦ tagged) with i. Moreover, we use the convention e.g. R λx.M ), R2 ). Note that the repositories need not contain tags as, whenever a method is looked up, we subsequently tag its body explicitly. Thus, the computationally observable difference of the new semantics is in the rule for reducing E[mi v] when m is not in the domain of Ri : this corresponds precisely to the case where e.g. a library method is called by the context with another method as argument. A similar behaviour is exposed when such a method is returning. However, this novelty merely adds fresh method names by η-expansions and does not affect the termination of the reduction. Defining parallel reduction Ô Ô⇒′ in an analogous way to Ô Ô⇒, we can show the following. We let a quadruple (M1 ∥⋯∥MN , R, S) be final if Mi = () for all i, and we write (M1 ∥⋯∥MN , R, S) ⇓ if (M1 ∥⋯∥MN , R, S) can reduce to some final quadruple; these notions are defined for (M1 ∥⋯∥MN , R1 , R2 , S) in the same manner. Lemma 50. For any legal (M1 ∥⋯∥MN , R1 , R2 , S), we have that (M1 ∥⋯∥MN , R1 , R2 , S) ⇓ iff (M1 ∥⋯∥MN , R1 ∪ R2 , S) ⇓. 31 We now proceed to syntactic composition of N -configurations. Given a pair ρ1 ≍ ρ2 , we define a single quadruple corresponding to their syntactic composition, called their in⃗ R1 , P1 , A1 , S1 ) and ρ2 = (C⃗′ , R2 , P2 , A2 , S2 ) ternal composition, as follows. Let ρ1 = (C, ′ and, for each i, Ci = (Ei , Xi ) and Ci = (Ei′ , Xi′ ), with {Xi , Xi′ } = {Mi , −}, and we let ki = 1 just if Xi = Mi . We let the internal composition of ρ1 and ρ2 be the quadruple: kN ′ ρ1 ⩕ ρ2 = ((E1 ⩕ E1′ )[M1k1 ]∥⋯∥(EN ⩕ EN )[MN ], R1 , R2 , S1 ⊎ S2 ) where compatible evaluation stacks E, E ′ are composed into a single evaluation context E ⩕ E ′ , as follows. (m ∶∶ E ∶∶ E) ⩕ (m ∶∶ E ′ ) = (E ⩕ E ′ )[E[⟨m⟩ ●]1 ] (m ∶∶ E ′ ) ⩕ (m ∶∶ E ∶∶ E) = (E ⩕ E ′ )[E[⟨m⟩ ●]2 ] and [] ⩕ [] = ●. Unfolding the above, we have that, for example: [mk , Ek , mk−1 , mk−2 , Ek−2 , ⋯, m1 , E1 ] ′ ′ ⩕ [mk , mk−1 , Ek−1 , mk−2 , ⋯, m1 ] = E11 [⟨m11 ⟩ E22 [⋯Ekk [⟨mkk ⟩ ●]⋯]] where k ′ = 2 − (k mod 2). We proceed to fleshing out the correspondence. We observe that an L-configuration ρ can be the final configuration of a trace just if all its components are O-configurations with empty evaluation stacks. On the other hand, for C-configurations, we need to reach P -configurations with terms (). Thus, we call an N -configuration ρ final if ρ = ⃗ R, P, A, S) and either Ci = ([], −) for all i, or Ci = ([], ()) for all i. (C, Let us write (S1 , ↪1 , F1 ) for the transition system induced from external composition, and (S2 , ↪2 , F2 ) be the transition system derived from internal composition: – S1 = {ρ ⊘ ρ′ ∣ ρ ≍ ρ′ }, F1 = {ρ ⊘ ρ′ ∈ S1 ∣ ρ, ρ′ final}, and ↪1 the transition relation Ð→ defined previously. ⃗ S) ∣ (M1 ∥⋯∥MN , R1 ⊎ R2 , S) valid}, F2 = {x ∈ S2 ∣ – S2 = {(M1 ∥⋯∥MN , R, x final}, and ↪2 the transition relation Ô Ô⇒′ defined above. A relation R ⊆ S1 × S2 is called a bisimulation if, for all (x1 , x2 ) ∈R: – x1 ∈ F1 iff x2 ∈ F2 , – if x1 ↪1 x′1 then x2 ↪2 x′2 and (x′1 , x′2 ) ∈ R, – if x2 ↪2 x′2 then x1 ↪1 x′1 and (x′1 , x′2 ) ∈ R. Given (x1 , x2 ) ∈ S1 × S2 , we say that x1 and x2 are bisimilar, written x1 ∼ x2 , if (x1 , x2 ) ∈ R for some bisimulation R. Lemma 51. Let ρ ≍ ρ′ be compatible N -configurations. Then, (ρ ⊘ ρ′ ) ∼ (ρ ⩕ ρ′ ). Recall we write h̄ for the O/P complement of the history h. We can now prove Theorem 33, which states that the behaviour of a library L inside a context C can be deduced by composing the semantics of L and C. 32 Theorem 33 Let L ∶ Θ → Θ′ , L′ ∶ 1 → Θ, Θ1 and Θ′ , Θ1 ⊢ M1 , ⋯, MN ∶ unit, with L, L′ and M1 ; ⋯; MN accessing pairwise disjoint parts of the store. Then, link L′ ; L in (M1 ∥⋯∥MN ) ⇓ iff there is h ∈ JLKN such that h̄ ∈ Jlink L′ ; − in (M1 ∥⋯∥MN )K. Proof. Let C be the context link L′ ; − in (M1 ∥⋯∥MN ), and suppose (L) Ð→∗lib (, R0 , S0 ) and (L′ ) Ð→∗lib (, R′0 , S0′ ) with dom(R0 ) ∩ dom(R′0 ) = dom(S0 ) ∩ dom(S0′ ) = ∅. We set: ρ0 = (([], −)∥⋯∥([], −), R0 , (∅, Θ′ ), (Θ, ∅), S0 ) ρ′0 = (([], M1 )∥⋯∥([], MN ), R′0 , (Θ, ∅), (∅, Θ′ ), S0′ ) We pick these as the initial N -configurations for JLKN and JCK respectively. Moreover, ⃗ ∣ we have that (L′ ; L) Ð→∗lib (, R′′0 , S0′′ ) where R′′0 = {(m, (R0 ⊎ R′0 )(m){!⃗ r/m}) ⃗ ⊎s {(ri , mi ) ∣ i = 1, ⋯, n}, assuming m ∈ dom(R0 ⊎ R′0 )} and S0′′ = (S0 ⊎ S0′ ){!⃗ r/m} Θ = {m1 , ⋯, mn } and r1 , ⋯, rn are fresh references of corresponding types. Hence, the initial triple for JC[L]K is taken to be φ0 = (([], M1 )∥⋯∥([], MN ), R′′0 , S0′′ ). On the other hand, ρ′0 ⩕ ρ0 = (([], M1 )∥⋯∥([], MN ), R′0 , R0 , S0 ⊎ S0′ ) and, using also Lemma 50, we have that φ0 ⇓ iff ρ′0 ⩕ ρ0 ⇓. Then, for the forward direction of the claim, from φ0 ⇓ we obtain that ρ′0 ⩕ρ0 ⇓. From the previous lemma, we have that so does ρ′0 ⊘ρ0 . From the latter reduction we obtain the required common history. Conversely, suppose h ∈ JLKN and h̄ ∈ JCK. WLOG, assume that Meths(h) ∩ (dom(R0 ) ∪ dom(R′0 )) ⊆ Θ ∪ Θ1 ∪ Θ′ (we can appropriately alphacovert R0 and R′0 for this). Then, ρ0 and ρ′0 both produce h, with opposite polarities. By definition of the external composite reduction, we then have that ρ′0 ⊘ ρ0 reduces to some final state. By the previous lemma, we have that ρ′0 ⩕ ρ0 reduces to some final quadruple, which in turn implies that φ0 ⇓, i.e. link L′ ; L in (M1 ∥⋯∥MN ) ⇓. C.1 Lemma 50 We purpose to show that, for any legal (M1 ∥⋯∥MN , R1 , R2 , S), (M1 ∥⋯∥MN , R1 , R2 , S) ⇓ iff (M1 ∥⋯∥MN , R1 ∪ R2 , S) ⇓. We prove something stronger. For any repository R whose entries are of the form (m, λx.m′ x), we define a directed graph G(R) where vertices are all methods appearing in R, and (m, m′ ) is a (directed) edge just if R(m) = λx.m′ x. In such a case, we call R an expansion class if G(R) is acyclic and all its vertices have at most one outgoing edge. Moreover, given an expansion class R, we define the method-for-method substitution {R} that assigns to each vertex m of G(R) the (unique) leaf m′ such that there is a directed path from m to m′ in G(R). Let us write L(R) for the set of leaves of G(R). For any quadruple φ = (E1 [M1 ]∥⋯∥EN [MN ], R1 , R2 , S) and expansion class R ⊆ R1 ∪ R2 , we define the triple: φ#R = (E1 [M1 ]∥⋯∥EN [MN ], R1 ∪ R2 , S){R} = (E1 [M1 ]{R}∥⋯∥EN [MN ]{R}, (R1 ∪ R2 ){R}, S{R}) where R′ {R} = {(m, R′ (m){R}) ∣ m ∈ dom(R′ ∖ R) ∪ L(R)}, S{R} = (S ↾ Refsint ) ∪ {(r, S(r){R}) ∣ r ∈ dom(S) ∖ Refsint }, and E[M ] is the term obtained from E[M ] by removing all tagging. 33 We next define a notion of indexed bisimulation between the transition systems produced from quadruples and triples respectively. Given an expansion class R, a relation RR between quadruples and triples is called an R-bisimulation if, whenever φ1 R R φ2 : – – – – φ1 final implies φ2 final φ2 final implies φ2 ⇓ φ1 Ô Ô⇒′ φ′1 implies φ2 Ô Ô⇒= φ′2 and φ′1 RR′ φ′2 for some expansion class R′ ⊇ R ′ φ2 Ô Ô⇒ φ2 implies φ1 Ô Ô⇒′∗ φ′1 and φ′1 RR′ φ′2 for some expansion class R′ ⊇ R. Thus, Lemma 50 directly follows from the next result. Lemma 52. For all expansion classes R, the relation RR = ⃗ S) legal ∧ R ⊆ R1 ∪ R2 } {(φ, φ#R) ∣ φ = (E1 [M1 ]∥⋯∥EN [MN ], R, is a bisimulation. Proof. Suppose φRR φ#R . We note that finality conditions are satisfied: if φ is final then so is φ#R ; while if φ#R is final then all its contexts are from the grammar: E ′ ∶∶= ● ∣ ⟨mi ⟩ E ′ so φ ⇓ by acyclicity of G(R). Suppose now φ Ô Ô⇒ φ′ , say due to (E1 [M1 ], R1 , R2 , S) Ð →′1 (E1′ [M1′ ], R′1 , R′2 , S ′ ). In case the reduction is not a function call or return, then it can be clearly simulated by φ#R . Otherwise, suppose: ⃗ S) Ð ⃗ S). If m ∉ dom(R) then, writing R12 for – (E1 [mi v], R, →′1 (E1 [M {v/x}i ], R, R1 ∪R2 , the above can be simulated by (E1 [mv], R12 , S){R} Ð →1 (E1 [M {v/x}], R12 , S){R}. If, on the other hand, m ∈ dom(R), suppose Ri (m) = λx.m′ x, then M = m′ x and m{R} = m′ {R} so we have: E1 [M {v/x}i ]{R} = E1 [(m′ v)i ]{R} = E1 [(mv)i ]{R} and E1 [(mv)i ] = E1 [mi v] by the way the semantics was defined, so φ′#R = φ#R . ⃗ S) Ð ⃗ ′ , S), with R3−i (m) = λx.M , – (E1 [mi v], R, →′1 (E1 [⟨mi ⟩ M {v ′ /x}3−i ], R ′ ′ ′ ⃗ = R⊎ ⃗ i {m′ ↦ λx.mj x ∣ 1 ≤ j ≤ k}. ⃗ /m} ⃗ and R Meths(v) = {m1 , ⋯, mk }, v = {m j ′ ′ Let R = R ⊎ {mj ↦ λx.mj x ∣ 1 ≤ j ≤ k} ⊆ R′1 ∪ R′2 . If m ∉ dom(R) then (E1 [mv], R12 , S){R} Ð →1 (E1 [M {v/x}], R12 , S){R}, and we have: E1 [⟨mi ⟩ M {v ′ /x}3−i ]{R′ } = E1 [M {v ′ /x}]{R′ } = E1 [M {v/x}]{R} Moreover, R12 {R} = (R′1 ∪R′2 ){R′ } and S{R} = S{R′ }, so φ′#R = (E1 [M {v/x}], R12 , S){R}. On the other hand, if R(m) = λx.m′′ x then: E[⟨mi ⟩ M {v ′ /x}3−i ]{R′ } = E[m′′ v ′ ]{R′ } = E[m′′ v]{R} = E[mv]{R} so φ#R = φ′#R . ′ 34 – Finally, the cases for method-return reductions are treated similarly as above. ⃗ S). Suppose now φ#R Ô Ô⇒ φ′ , where recall that we write φ as (E1 [M1 ]∥⋯∥EN [MN ], R, We show by induction on sizeR (E1 [M1 ], ⋯, EN [MN ]) that φ Ô Ô⇒′ φ′′ and φ′ RR′ φ′′ for some R′ ⊇ R. The size-function we use measures the length of G(R)-paths that appear inside its arguments: sizeR (E1 [M1 ], ⋯, EN [MN ]) = sizeR (E1 [M1 ]) + ⋯ + sizeR (EN [MN ]) sizeR (E[M ]) = ∑ 2∣m∣R + ∑ 1 m∈X1 m∈X2 where X1 is the multiset containing all occurrences of methods m ∈ dom(R) inside E[M ] in call position (e.g. mM ′ ), and X2 contains all occurrences of methods m ∈ dom(R) inside E[M ] in return position (i.e. ⟨mi ⟩ ⋯). We write ∣m∣R for the length of the unique directed path from m to a leaf in G(R). The fact that X1 , X2 are multisets reflects that we count all occurrences of m in call/return positions. Suppose WLOG that the reduction to φ′ is due to some (E1 [M1 ], R12 , S){R} Ð →1 (E ′ [M ′ ], R′ , S ′ ). If the reduction happens inside M1 {R} (this case also encompasses the base case of the induction) then the only case we need to examine is that of the reduction being a method call. In such a case, suppose we have E1 [M1 ]{R} = E[mv], E ′ = E, M ′ = M {v/x} and R12 {R}(m) = λx.M . Then, E1 [M1 ] = Ẽ[m̃i ṽ] for some Ẽ, m̃, ṽ such that m̃{R} = m, ṽ{R} = v and Ẽ{R} = E. If m =/ m̃ then, supposing R(m̃) = λx.m̃′ x we have the following cases: ⃗ S) Ð ⃗ S) = φ′′ – (Ẽ[m̃i ṽ], R, →′1 (Ẽ[m̃′i ṽ], R, 1 i ′ i ⃗ ⃗ ′ , S) = φ′′ , with R ⃗′ = R ⃗ ⊎3−i {m′ ↦ – (Ẽ[m̃ ṽ], R, S) Ð →1 (Ẽ[⟨m̃ ⟩(m̃′ v ′ )3−i ], R 1 j λx.mj x ∣ 1 ≤ j ≤ k}, etc. Let φ′′ be the extension of φ′′1 to an N -quadruple by using the remaining Ei [Mi ]’s of φ, so that φ Ô Ô⇒′ φ′′ . In the first case above we have that φ′′#R = φ, and in the latter that ′ φ′′#R = φ (with R′ = R ⊎ {m′j ↦ λx.mj x ∣ 1 ≤ j ≤ k}), and we appeal to the IH. Suppose now that m̃ = m and R12 (m) = λx.M̃ . Then, one of the following is the case: ⃗ S), R, ⃗ S) Ð ⃗ S) = φ′′ – (Ẽ[m̃i ṽ], R, →′1 (Ẽ[M̃ {ṽ/x}i ], R, 1 i ′ i ′ 3−i ⃗ ⃗ ′ = R⊎ ⃗ 3−i {m′ ↦ ⃗ ′ , S) = φ′′ , with R – (Ẽ[m̃ ṽ], R, S) Ð →1 (Ẽ[⟨m̃ ⟩ M̃ {v /x} ], R 1 j λx.mj x ∣ 1 ≤ j ≤ k}, etc. Extending φ′′1 to φ′′ as above, in the former case we then have that φ′′#R = φ′ , and in ′ the latter that φ′′#R = φ′ , as required. Finally, let us suppose that M1 is some value v. Then, we can write E1 as E1 = E2 [E ′ ], with E ′ coming from the grammar E ′ ∶∶= ● ∣ ⟨mi ⟩ E ′ and E2 not being of the form E ′′ [⟨mi ⟩ ●]. Observe that E1 = E2 . If E ′ = ● then by a case analysis on E1 ⃗ S) Ð we can see that φ#R can simulate the reduction. Otherwise, (E2 [E ′ [v]], R, →′1 ′′ ′i ′ ′ ′′ i ′ ′ ⃗ ⃗ ⃗ (E2 [E [v ]], R , S) whereby E = E [⟨m ⟩ ●] and R = R ⊎3−i {mj ↦ λx.mj x ∣ 1 ≤ j ≤ k}, etc. We have that ⃗ ′ , S){R′ } = (E2 [E ′ [v]], R, ⃗ S){R} φ′′1 = (E2 [E ′′ [v ′i ]], R and hence, extending φ′′1 to φ′′ , we have φ′′#R = φ#R . We can now appeal to the IH. ′ 35 C.2 Lemma 51 Let ρ ≍ ρ′ be compatible N -configurations. Then, (ρ ⊘ ρ′ ) ∼ (ρ ⩕ ρ′ ). We prove that the relation R = {(ρ1 ⊘ ρ2 , ρ1 ⩕ ρ2 ) ∣ ρ1 ≍ ρ2 } is a bisimulation. Let us suppose that (ρ1 ⊘ ρ2 , ρ1 ⩕ ρ2 ) ∈ R. – Suppose ρ1 ⊘ρ2 ↪1 ρ′1 ⊘ρ′2 . If the transition is due to (I NT 1) then ρ2 = ρ′2 and we can see that ρ1 ⩕ ρ2 Ô Ô⇒′ ρ′1 ⩕ ρ2 . Similarly if the transition is due to (I NT 2). Suppose (1,call m(v)) (1,call m(v)) now we used instead (C ALL), e.g. ρ1 ÔÔÔÔÔ⇒ ρ′1 and ρ2 ÔÔÔÔÔ⇒ ρ′2 , and let us consider the case where v ∈ Meths (the other case is simpler). Then, assuming ρ1 = (C11 ∥⋯, R1 , P1 , A1 , S1 ) and ρ2 = (C12 ∥⋯, R2 , P2 , A2 , S2 ), we have that either of the following scenarios holds, for some x ∈ {K, L}: C11 = (E1 , E[mm′ ]), C12 = (E2 , −) and call m(v) (E1 , E[mm′ ], R1 , P1 , A1 , S1 ) ÐÐÐÐÐ→1 (m ∶∶ E ∶∶ E1 , R1 ⊎ (v ↦ λx.m′ x), P1 ∪x {v}, A1 , S1 ) call m(v) (E2 , −, R2 , P2 , A2 , S2 ) ÐÐÐÐÐ→1 (m ∶∶ E2 , M {v/x}, R2 , P1 , A1 ∪x {v}, S2 ) or its dual, where ρ2 contains the code initiating the call. Focusing WLOG in the former case and setting S = S1 ⊎ S2 : ρ1 ⩕ ρ2 = ((E1 ⩕ E2 )[E[m1 m′ ]]∥⋯, R1 , R2 , S) ↪2 ((E1 ⩕ E2 )[E[⟨m1 ⟩ M {v/x}2 ]]∥⋯, R′1 , R2 , S) = ρ′1 ⩕ ρ′2 (R′1 = R1 ⊎ (v ↦ λx.m′ x)) The case for (R ETN) is treated similarly. ⃗ S) ↪2 (E[M ′ ]∥ M2 ∥⋯∥MN , R ⃗ ′, S′) – Suppose ρ1 ⩕ ρ2 = (E[M1 ]∥M2 ∥⋯∥MN , R, 1 ′′ and let ρ1 = ((E1 , M1 )∥⋯, R1 , P1 , A1 , S1 ) and ρ2 = ((E2 , −)∥⋯, R2 , P2 , A2 , S2 ), where (E1 ⩕ E2 )[M1′′ ] = E[M1 ]. If the redex M1 is not of the forms M1 = m1 v or M1 = ⟨m1 ⟩ v, with m ∈ dom(R2 ), then the reduction can clearly be simulated by ρ1 ⊘ ρ2 (internally, by ρ1 ). Otherwise, similarly as above, the reduction can be simulated by a mutual call/return of m. Finally, it is clear that ρ1 ⊘ ρ2 is final iff ρ1 ⩕ ρ2 is final. D General compositionality This compositionality result will allow us to compose histories of component libraries in order to obtain those of their composite library. Let L1 ∶ Θ1 → Θ2 and L2 ∶ Θ1′ → Θ2′ . The semantic composition will be guided by two sets of names Π, P. Π contains method names that are shared between by the respective libraries and their context. Thus Π ⊇ Θ1 ∪ Θ1′ ∪ Θ2 ∪ Θ2′ . The names in P, on the other hand, will be used for private communication between L1 and L2 . Consequently, Π ∩ P consists of names that can be used both for internal communication between L1 and L2 , and for contextual interactions, i.e. Π ∩ P = (Θ1 ∪ Θ1′ ) ∩ (Θ2 ∪ Θ2′ ). 36 Given hi ∈ JLi K(i = 1, 2), we define the composition of h1 and h2 , written h1 ⩕σΠ,P h2 , as a partial operation depending on Π, P and an additional parameter σ ∈ {0, 1, 2}∗ which we call a scheduler. It is given inductively as follows. We let  ⩕Π,P  =  and: σ (t, call m(v))s1 ⩕0σ Π,P (t, call m(v))s2 = s1 ⩕Π,P′ s2 σ (t, ret m(v))s1 ⩕0σ Π,P (t, ret m(v))s2 = s1 ⩕Π,P′ s2 σ (t, call m(v))P Y s1 ⩕1σ Π,P s2 = (t, call m(v))P Y (s1 ⩕Π ′ ,P s2 ) σ (t, ret m(v))P Y s1 ⩕1σ Π,P s2 = (t, ret m(v))P Y (s1 ⩕Π ′ ,P s2 ) σ (t, call m(v))OY s1 ⩕1σ Π,P s2 = (t, call m(v))OY (s1 ⩕Π ′ ,P s2 ) σ (t, ret m(v))OY s1 ⩕1σ Π,P s2 = (t, ret m(v))OY (s1 ⩕Π ′ ,P s2 ) along with the dual rules for the last four cases (i.e. where we schedule 2 in each case). Note that the definition uses sequences of moves that are suffixes of histories (such as si ). The above equations are subject to the following side conditions: – Meths(v) ∩ (Π ∪ P) = ∅, Π ′ = Π ⊎ Meths(v) and P′ = P ⊎ Meths(v); – m ∈ P in the 0-scheduling cases; – m ∈ Π in the 1-scheduling cases and, also, m ∈ Π ∖ P in the third case (the P -call); – in the 1-scheduling cases, we also require that the leftmost move with thread index t in s2 is not a P -move. History composition is a partial function: if the conditions above are not met, or h1 , h2 , σ are not of the appropriate form, then the composition is undefined. The above conditions ensure that the composed histories are indeed compatible and can be produced by composing actual libraries. For instance, the last condition corresponds to determinacy of threads: there can only be at most one component starting with a P -move in each thread t. We then have the following correspondence. Theorem 53. If L1 ∶ Θ1 → Θ2 and L2 ∶ Θ1′ → Θ2′ access disjoint parts of the store then JL1 ∪ L2 KN = { h ∈ H ∣ ∃σ, h1 ∈JL1 KN, h2 ∈JL2 KN. h = h1 ⩕σΠ0 ,P0 h2 } with Π0 = Θ1 ∪ Θ2 ∪ Θ1′ ∪ Θ2′ and P0 = (Θ1 ∪ Θ1′ ) ∩ (Θ2 ∪ Θ2′ ). The rest of this section is devoted in proving the Theorem. Recall that we examine library composition in the sense of union of libraries. This scenario is more general than the one of Appendix C as, during composition via union, the calls and returns of each of the component libraries may be caught by the other library or passed as a call/return to the outer context. Thus, the setting of this section comprises given libraries L1 ∶ Θ1 → Θ2 and L2 ∶ Θ1′ → Θ2′ , such that Θ2 ∩ Θ2′ = ∅, and relating their semantics to that of their union L1 ∪ L2 ∶ (Θ1 ∪ Θ1′ ) ∖ (Θ2 ∪ Θ2′ ) → Θ2 ∪ Θ2′ . Given configurations for L1 and L2 , in order to be able to reduce them together we need to determine which of their methods can be used for communication between them, and which for interacting with the external context, which represents player O in the game. We will therefore employ a set of method names, denoted by Π and variants, to register those methods used for interaction with the external context. Another piece of 37 information we need to know is in which component in the composition was the last call played, or whether it was an internal call instead. This is important so that, when O (or P ) has the choice to return to both components, in the same thread, we know which one was last to call and therefore has precedence. We use for this purpose sequences w = (w1 , ⋯, wN ) where, for each i, wi ∈ {0, 1, 2}∗ . Thus, if e.g. w1 = 2w1′ , this would mean that, in thread 1, the last call to O, was done from the second component; if, on the other hand, w1 = 0w1′ then the last call in thread 1 was an internal one between the two components. Given such a w and some j ∈ {0, 1, 2}, for each index t, we write j +t w for w[t ↦ (jwt )]. Let us fix libraries L1 ∶ Θ1 → Θ2 and L2 ∶ Θ1′ → Θ2′ . Let ρ1 , ρ2 be N -configurations, ⃗ R, P, A, S) and ρ2 = and in particular L-configurations, and suppose that ρ1 = (C, ′ ′ ′ ′ ′ ′ ′ ⃗ (C , R , P , A , S ). Moreover, let Θ1 ∪ Θ2 ∪ Θ1 ∪ Θ2 ⊆ Π. We say that ρ1 and ρ2 are ′ (w, Π)-compatible, written ρ1 ≍w Π ρ2 , if S, S have disjoint domains and, for each i; – Ci = (EL′ , M ) and Ci′ = (EL , −), or Ci = (EL , −) and Ci′ = (EL′ , M ), or Ci = (EL1 , −) and Ci′ = (EL2 , −). ′ – We have Θ1 ⊆ Al , Θ2 ⊆ PK , Θ1′ ⊆ A′L , Θ2′ ⊆ PK and, setting ′ P = (PK ∩ A′L ) ⊎ (PL ∩ A′K ) ⊎ (PK ∩ AL ) ⊎ (Pl′ ∩ AK ) we also have: ′ ● (PL ⊎ PK ⊎ Al ⊎ AK ) ∩ (PL′ ⊎ PK ⊎ A′l ⊎ A′K ) = P ⊎ (Θ1 ∩ Θ1′ ), ′ ′ ● Π ∩ P = (Θ2 ∪ Θ2 ) ∩ (Θ1 ∪ Θ1 ), ′ ● Π ∪ P = PL ∪ PK ∪ Pl′ ∪ PK ∪ AL ∪ AK ∪ A′L ∪ A′K . – The private names of R do not appear in ρ2 , and dually for the private names of R′ . – If Ci = (E, ⋯) and Ci′ = (E ′ , ⋯) then E and E ′ are wi -compatible, that is, either E = E ′ = [] or: ● E = m ∶∶ E1 and E ′ ∈ EL , with m ∈ Π, wi = 1u and E1 , E ′ are u-compatible, ● or E = m ∶∶ E1 and E ′ = m ∶∶ E ∶∶ E2 , with m ∈ P, wi = 0u and E1 , E2 are u-compatible, ● or E = m ∶∶ E ∶∶ E1 and E ′ ∈ EL , with m ∈ Π ∖ P, wi = 1u and E1 , E ′ are u-compatible, or the dual of one of the three conditions above holds. w Given ρ1 ≍w Π ρ2 , we let their external composition be denoted as ρ1 ⊗Π ρ2 (and note that now the notation is symmetric for ρ1 and ρ2 ) and define the semantics for external 38 composition by these rules: ρ1 Ô Ô⇒ ρ′1 I NT1 ′ w ρ1 ⊗ w Π ρ2 Ð→ ρ1 ⊗Π ρ2 (t,call m(v)) ρ1 ÔÔÔÔÔ⇒ ρ′1 (t,call m(v)) ρ2 ÔÔÔÔÔ⇒ ρ′2 0+t w ′ ′ ρ 1 ⊗w ρ2 Π ρ2 Ð→ ρ1 ⊗Π (t,ret m(v)) ρ1 ÔÔÔÔÔ⇒ ρ′1 C ALL (m ∈ P) (t,ret m(v)) ρ2 ÔÔÔÔÔ⇒ ρ′2 ′ tw ρ1 ⊗0+ ρ2 Ð→ ρ′1 ⊗w Π ρ2 Π R ETN (m ∈ P) (t,call m(v))P Y ρ1 ÔÔÔÔÔÔÔ⇒ ρ′1 ρ 1 ⊗w Π (t,call m(v))P Y ρ2 ÐÐÐÐÐÐÐÐ→ ρ′1 tw ⊗1+ Π′ PC ALL1 (m ∈ Π ∖ P) ρ2 (t,ret m(v))P Y ρ1 ÔÔÔÔÔÔ⇒ ρ′1 (t,ret m(v))P Y tw ρ1 ⊗1+ ρ2 ÐÐÐÐÐÐÐÐ→ ρ′1 ⊗w Π ′ ρ2 Π PR ETN1 (m ∈ Π) (t,call m(v))OY ρ1 ÔÔÔÔÔÔÔ⇒ ρ′1 (t,call m(v))OY 1+t w ′ ρ 1 ⊗w ρ2 Π ρ2 ÐÐÐÐÐÐÐÐ→ ρ1 ⊗Π ′ OC ALL1 (m ∈ Π) (t,ret m(v))OY ρ1 ÔÔÔÔÔÔÔ⇒ ρ′1 (t,ret m(v))OY tw ρ1 ⊗1+ ρ2 ÐÐÐÐÐÐÐÐ→ ρ′1 ⊗w Π ′ ρ2 Π OR ETN1 (m ∈ Π ∖ P) along with their dual counterparts (I NT2 , XC ALL2 , XR ETN2 ). The internal rules above have the same side-conditions on name privacy as before. Moreover, in (XR ETNi) and (XC ALLi), for X=O,P, we let Π ′ = Π ⊎t Meths(v) and impose that the t-th component of ρ3−i be an O-configuration and Meths(v) ∩ Meths(ρ3−i ) = ∅. We can now show the following. s w ′ Lemma 54. Let ρ1 ≍w Ð→∗ ρ′1 ⊗w Π ρ2 and suppose ρ1 ⊗Π ρ2 Ð Π ′ ρ2 for some sequence s ′ w′ ′ of moves. Then, ρ1 ≍Π ′ ρ2 . ′ We next juxtapose the semantics of external composition to that obtained by internally composing the libraries and then deriving the multi-threaded semantics of the result. As before, we call the latter form internal composition. The traces we obtain are produced from a transition relation, written Ô Ô⇒′ , between configurations of the ⃗ R, ⃗ P, A, S). In particular, in each form (C1 ∥⋯∥CN , R1 , R2 , P, A, S), also written (C, Ci = (Ei , Xi ) with Xi = Ei [Mi ] or Xi = −, Ei is selected from the extended evaluation contexts and Ei is an extended L-stack, that is, of either of the following two forms: ′ Eext ∶∶= [] ∣ mi ∶∶ E ∶∶ Eext ′ Eext ∶∶= m ∶∶ Eext where E is again from the extended evaluation contexts. First, given u-compatible evaluation stacks E, E ′ , we construct a pair E ⩕u E ′ consisting of an extended evaluation context and an extended L-stack, as follows. Given 39 E ⩕u E ′ = (E ′ , E ′′ ): (m ∶∶ E ∶∶ E) ⩕0u (m ∶∶ E ′ ) = (E ′ [E[⟨m⟩ ●]1 ], E ′′ ) (m ∶∶ E) ⩕0u (m ∶∶ E ∶∶ E ′ ) = (E ′ [E[⟨m⟩ ●]2 ], E ′′ ) (m ∶∶ E) ⩕1u E ′ = E ⩕2u (m ∶∶ E ′ ) = (●, m ∶∶ E ′ ∶∶ E ′′ ) (m ∶∶ E ∶∶ E) ⩕1u E ′ = E ⩕2u (m ∶∶ E ∶∶ E ′ ) = (●, m ∶∶ E ′ [E] ∶∶ E ′′ ) if E ′ ∈ EL and [] ⩕ [] = (●, []). For each pair ρ1 ≍w Π ρ2 , we define a configuration corresponding to their syntactic ′ composition as follows. Let ρ1 = (C1 ∥⋯∥CN , R1 , P1 , A1 , S1 ) and ρ2 = (C1′ ∥⋯∥CN , R2 , P2 , A2 , S2 ) ′ ′ ′ u ′ ′′ and, for each i, Ci = (Ei , Xi ) and Ci = (Ei , Xi ). If Ei ⩕ Ei = (Ei , Ei ), we set: Ci ⩕ u Ci′ ⎧ (Ei′′ , Ei [M 1 ]) ⎪ ⎪ ⎪ ⎪ ′′ = ⎨(Ei , Ei [M 2 ]) ⎪ ⎪ ′′ ⎪ ⎪ ⎩(Ei , −) if Xi = M and Xi′ = − if Xi = − and Xi′ = M if Xi = Xi′ = − We then let the internal composition of ρ1 and ρ2 be: w1 ′ wN ′ ρ1 ⩕w CN , R1 , R2 , P ′ , A′ , S1 ⊎ S2 ) Π ρ2 = (C1 ⩕ C1 ∥⋯∥CN ⩕ where we set P ′ = ((P1L ⊎ P2L ) ∩ Π, (P1K ⊎ P2K ) ∩ Π) and A′ = ((A1L ∪ A2L ) ∩ (Π ∖ P), (A1K ⊎ A2K ) ∩ Π). Now, as expected, the definition of Ô Ô⇒′ builds upon Ð →′t . The definition of the latter is given by the following rules. ⃗ S) Ð ⃗ ′, S′) (E[M ], R, →′t (E ′ [M ′ ], R (I NT′ ) ⃗ P, A, S) Ð ⃗ ′ , P, A, S ′ ) (E, E[M ], R, →′ (E, E ′ [M ′ ], R t ⃗ P, A, S) (E, E[m v], R, i call m(v ′ )P Y ÐÐÐÐÐÐÐ→t′ ⃗ ′ , P ′ , A, S) (mi ∶∶E ∶∶ E, −, R (PQ Y′ ) ′ ret m(v )P Y ′ ⃗ P, A, S) ÐÐÐÐÐÐ→ ⃗′ ′ (m ∶∶ E, v, R, t (E, −, R , P , A, S) (PAY′ ) call m(v)OY ′ i ⃗ ′ ⃗ P, A, S) ÐÐÐÐÐÐ→ (E, −, R, t (m ∶∶ E, M {v/x} , R, P, A , S) (OQ Y′ ) ⃗ P, A, S) ÐÐÐÐÐÐ→t′ (E, E[v i ], R, ⃗ P, A′ , S) (mi ∶∶E ∶∶ E, −, R, (OAY′ ) ret m(v)OY The side-conditions are similar to those for the relation Ð →t between ordinary configurations, with the following exceptions: in (PQ Y′ ), if Meths(v) = {m1 , ⋯, mk } then ⃗′ = R ⃗ ⊎i {m′ ↦ λx.mj x}; and in v ′ = v{m′j /mj ∣ 1 ≤ j ≤ k}, for fresh m′j ’s, and R j ′ ′ ′ ⃗ ⃗ (PAY ), if m ∈ dom(Ri ) then R = R ⊎i {mj ↦ λx.mj x}, etc. Moreover, in (OQ Y′ ) we have that m ∈ dom(Ri ). Finally, we let (t,x)XY ′ ⃗ R, ⃗ P, A, S) ÔÔÔ⇒ ⃗ ↦ C ′ ], R ⃗ ′ , P ′ , A′ , S ′ ) (C, (C[t ⃗ P, A, S) ÐÐÐ→t′ (C ′ , R ⃗ ′ , P ′ , A′ , S ′ ). just if (Ct , R, xXY 40 We next relate the transition systems induced by external (via ⊗) and internal composition (via ⩕). Let us write (S1 , ↪1 , F1 ) for the transition system induced by external composition of compatible N -configurations (so ↪1 is Ð→), and (S2 , ↪2 , F2 ) be the one for internal composition (so ↪2 is Ô Ô⇒′ ). Finality of extended N -configurations ⃗ ⋯) is defined as expected: all Ci ’s must be ([], −). A relation R ⊆ S1 ×S2 (C1 ∥⋯∥CN , R, is called a bisimulation if, for all (x1 , x2 ) ∈ R: – x1 ∈ F1 iff x2 ∈ F2 , – if x1 ↪1 x′1 then x2 ↪2 x′2 and (x′1 , x′2 ) ∈ R, (t,x)XY (t,x)XY (t,x)XY (t,x)XY – if x1 ,ÐÐÐÐ→1 x′1 then x2 ,ÐÐÐÐ→2 x′2 and (x′1 , x′2 ) ∈ R, – if x2 ↪2 x′2 then x1 ↪1 x′1 and (x′1 , x′2 ) ∈ R, – if x2 ,ÐÐÐÐ→2 x′2 then x1 ,ÐÐÐÐ→1 x′1 and (x′1 , x′2 ) ∈ R. Again, we say that x1 and x2 are bisimilar, and write x1 ∼ x2 , if there exists a bisimulation R such that (x1 , x2 ) ∈ R. ′ w ′ w ′ Lemma 55. Let ρ ≍w Π ρ be compatible N -configurations. Then, (ρ ⊗Π ρ ) ∼ (ρ ⩕Π ρ ). w w Proof. We prove that the relation R = {(ρ1 ⊗w Π ρ2 , ρ1 ⩕Π ρ2 ) ∣ ρ1 ≍Π ρ2 } is a bisimuw w lation. Let us suppose that (ρ1 ⊗Π ρ2 , ρ1 ⩕Π ρ2 ) ∈ R. (t,x) ′ ÐÐ→ ρ′1 ⊗w – Let ρ1 ⊗w Π ρ2 Ð Π ′ ρ2 with the transition being due to (XC ALL 1 ), e.g. ′ (1,call m(v)) ρ1 ÔÔÔÔÔ⇒ ρ′1 and ρ′2 = ρ2 , w′ = 1 +1 w and Π ′ = Π ⊎1 Meths(v), Meths(v) = {m′1 , ⋯, m′j }, and recall that Meths(v) ∩ Meths(ρ2 ) = ∅. Then, assuming ρ1 = (C11 ∥⋯, R1 , P1 , A1 , S1 ), we have that one of the following holds, for some x ∈ {K, L}: call m(v) C11 = (E1 , E[mv ′ ]) and (C11 , R1 , P1 , A1 , S1 ) ÐÐÐÐÐ→1 (m ∶∶ E ∶∶ E1 , −, R1 ⊎ {m′j ↦ λx.mj x ∣ 1 ≤ j ≤ k}, P1 ∪x Meths(v), A1 , S1 ) call m(v) C11 = (E1 , −) and (C11 , R1 , P1 , A1 , S1 ) ÐÐÐÐÐ→1 (m1 ∶∶ E1 , mv, R1 , P1 , A1 ∪x Meths(v), S1 ) In the former case, if ρ2 = ((E2 , −)∥⋯, R2 , P2 , A2 , S2 ) with E1 ⩕w1 E2 = (E ′ , E), we get: ′ ′ 1 ρ1 ⩕w Π ρ2 = ((E, E [E[mv ] ])∥⋯, R1 , R2 , P, A, S) (1,call m(v)) ÔÔÔÔÔ⇒′ (m1 ∶∶ E ′ [E 1 ] ∶∶ E, −)∥⋯, R1 ⊎ {m′j ↦ λx.mj x ∣ 1 ≤ j ≤ k}, R2 , P ′, A, S) with P, A as in the definition of composition and P ′ = P ∪x Meths(v), and the latter ′ N -configuration equals ρ′1 ⩕w Π ′ ρ2 . The other case is treated in the same manner, and we work similarly for (R ETN1 ). – On the other hand, if the transition is due to (C ALL) or (R ETN) then we work as in the proof of Lemma 51. 41 (1,call m(v)) ′ ′ ⃗ ⃗′ ′ ′ – Suppose ρ1 ⩕w Π ρ2 = (C1 ∥⋯, R, P, A, S) ÔÔÔÔÔ⇒ (C1 ∥⋯, R , P , A , S). Then, assuming WLOG that v ∈ Meths, one of the following must be the case, for some x ∈ {K, L} and i ∈ {1, 2}: call m(v) ′ ⃗ P, A, S) Ð C1 = (E, E[mi v ′ ]) and (C1 , R, ÐÐÐÐ→1 i ′ ⃗ (m ∶∶ E ∶∶ E, R ⊎i {mj ↦ λx.mj x ∣ 1 ≤ j ≤ k}, P ∪x Meths(v), A, S) ⃗ P, A, S) ÐÐÐÐÐ→1′ C1 (E, −) and (C1 , R, call m(v) ⃗ P, A ∪x Meths(v), S) (m ∶∶ E, M {v/x}i , R, We only examine the former case, as the latter one is similar, and suppose that i = 1. Taking ρj = (C1j ∥⋯, Rj , Pj , Aj , Si ), for j = 1, 2, we have that (C11 , C12 ) = ((E1 , E ′ [mv ′ ], (E2 , −)), for some E, E1 , E2 such that E1 ⩕w1 E2 = (E ′′ , E) and E = E ′′ [E ′1 ]. Moreover, taking R′1 = R1 ⊎{m′j ↦ λx.mj x ∣ 1 ≤ j ≤ k}, P1′ = P1 ⊎x {v}, w′ = 1 +1 w and Π ′ = Π ⊎ Meths(v) (note Meths(v) = {m′1 , ⋯, m′k }), (1,call m(v)) ′ ′ ′ ′ ′ w ′ w ρ1 ⊗ w Π ρ2 ÐÐÐÐÐÐÐ→ ((m ∶∶ E ∶∶ E1 , −)∥⋯, R1 , P1 , A1 , S1 ) ⊗Π ′ ρ2 = ρ1 ⊗Π ′ ρ2 ′ ′ ⃗′ ′ ′ and ρ′1 ⩕w Π ′ ρ2 = (C1 ∥⋯, R , P , A , S) as required. The case for return transitions is similar. – On the other hand, if the transition out of ρ1 ⩕w Π ρ2 does not have a label then we work as in the proof of Lemma 51. w Moreover, by definition of syntactic composition, ρ1 ⊗w Π ρ2 is final iff ρ1 ⩕Π ρ2 is. h Given an N -configuration ρ and a history h, let us write ρ ⇓ h if ρ Ô Ô⇒ ρ′ for ′ ⃗ ⃗ some final configuration ρ . Similarly if ρ is of the form (C, R, P, A, S). We have the following connections in history productions. The next lemma is proven in a similar fashion as Lemma 50. Lemma 56. For any legal (M1 ∥⋯∥MN , R1 , R2 , P, A, S) and history h, we have that (M1 ∥⋯∥MN , R1 , R2 , P, A, S) ⇓ h iff (M1 ∥⋯∥MN , R1 ∪ R2 , P, A, S) ⇓ h. w Lemma 57. For any compatible N -configurations ρ1 ≍w Π ρ2 and history h, (ρ1 ⊗Π ρ2 ) ⇓ h iff: ∃h1 , h2 , σ. ρ1 ⇓ h1 ∧ ρ2 ⇓ h2 ∧ h = h1 ⩕σΠ,P h2 where P is computed from ρ1 , ρ2 and Π as before. Proof. We show that, for any compatible N -configurations ρ1 ≍w Π ρ2 and history suffix s, (ρ1 ⊗w Π ρ2 ) ⇓ s iff: ∃s1 , s2 , σ. ρ1 ⇓ s1 ∧ ρ2 ⇓ s2 ∧ s = s1 ⩕σΠ,P s2 where P is computed from ρ1 , ρ2 and Π as in the beginning of this section. The left-to-right direction follows from straightforward induction on the length of the reduction that produces s. For the right-to-left direction, we do induction on the 42 length of σ. If σ =  then s1 = s2 = s = . Otherwise, we do a case analysis on the first element of σ. We only look at the most interesting subcase, namely of σ = 0σ ′ . Then, for some m ∈ P: s1 = (t, call m(v))s′1 s2 = (t, call m(v))s′2 w ′ w ′ By ρi ⇓ si and ρ1 ≍w Π ρ2 we have that ρ1 ⊗Π ρ2 Ð→ ρ1 ⊗Π ρ2 , where w = 0 +t w and ′ w′ ′ ′ ′ ′ σ′ ′ ′ w′ ′ ρ1 ≍Π ρ2 . Also, ρi ⇓ si and s = s1 ⩕Π,P′ s2 so, by IH, (ρ1 ⊗Π ρ2 ) ⇓ s. ′ We can now prove the correspondence between the traces of component libraries and those of their union. Theorem 53 Let L1 ∶ Θ1 → Θ2 and L2 ∶ Θ1′ → Θ2′ be libraries accessing disjoint parts of the store. Then, JL1 ∪ L2 KN = {h ∈ HL ∣ ∃σ, h1 ∈JL1 KN, h2 ∈JL2 KN. h = h1 ⩕σΠ0 ,P0 h2 } with Π0 = Θ1 ∪ Θ2 ∪ Θ1′ ∪ Θ2′ and P0 = (Θ1 ∪ Θ1′ ) ∩ (Θ2 ∪ Θ2′ ). Proof. Let us suppose (Li ) Ð→∗lib (, Ri , Si ), for i = 1, 2, with dom(R1 ) ∩ dom(R2 ) = dom(S1 ) ∩ dom(S2 ) = ∅. We set: ρ1 = (([], −)∥⋯∥([], −), R1 , (∅, Θ2 ), (Θ1 , ∅), S1 ) ρ2 = (([], −)∥⋯∥([], −), R2 , (∅, Θ2′ ), (Θ1′ , ∅), S2 ) We pick these as the initial configurations for JL1 KN and JL2 KN respectively. Then, (L1 ∪ L2 ) Ð→∗lib (, R0 , S0 ) where R0 = R1 ⊎ R2 and S0 = S1 ⊎ S2 , and we take ρ0 = (([], −)∥⋯∥([], −), R0 , (∅, Θ2 ∪ Θ2′ ), ((Θ1 ∪ Θ1′ ) ∖ P0 , ∅), S0 ) as the initial N -configuration for JL1 ∪ L2 KN . On the other hand, we have ρ1 ⩕Π0 ρ2 = (([], −)∥⋯∥([], −), R1 , R2 , (∅, Θ2 ∪ Θ2′ ), ((Θ1 ∪ Θ1′ ) ∖ P0 , S0 ). From Lemma 56, we have that ρ0 ⇓ h iff ρ1 ⩕Π0 ρ2 ⇓ h, for all h. Pick a history h. For the forward direction of the claim, ρ0 ⇓ h implies ρ1 ⩕Π0 ρ2 ⇓ h which, from Lemma 55, implies ρ1 ⊗Π0 ρ2 ⇓ h. We now use Lemma 57 to obtain h1 , h2 , σ such that ρi ⇓ hi and h = h1 ⩕σΠ0 ,P0 h2 . Conversely, suppose that hi ∈ JLi KN and h = h1 ⩕σΠ0 ,P0 h2 . WLOG assume that (Meths(h1 ) ∪ Meths(h2 )) ∩ (dom(R1 ) ∪ dom(R2 )) ⊆ Π0 (or we appropriately alpha-covert R1 and R2 ). Then, ρi ⇓ hi , for i = 1, 2, and therefore ρ1 ⊗Π0 ρ2 ⇓ h by Lemma 57. By Lemma 55 we have that ρ1 ⩕Π0 ρ2 ⇓ h, which in turn implies that ρ0 ⇓ h, i.e. h ∈ JL1 ∪ L2 KN . E E.1 Composition congruence Proof of Theorem 36 Proof. Assume L1 ⊑ L2 and suppose h1 ∈ JL ∪ L1 K. By Theorem 53, h1 = h′ ⩕σΠ,P h′1 , where h′ ∈ JLK and h′1 ∈ JL1 K. Because L1 ⊑ L2 , there exists h′2 ∈ JL2 K such that h′1 ⊑ h′2 , i.e. h′1 ◁∗P O h′2 . Note that some of the rearrangements necessary to transform h′1 into h′2 may concern actions shared by h′1 and h′ ; their polarity will then be different in h′ . Let h′′ be obtained by applying such rearrangements to h′ . We claim that h′ ◁∗OP h′′ . 43 Indeed, suppose that (t′ , x′ )(t, x)P are consecutive in h′1 , but swapped in order to obtain h′2 , and (t, x)P appears in h′ as (t, x)O . Now, the move (t′ , x′ ) either appears in h1 , or it appears in h′ and gets hidden in h1 . In every case, let s contain the moves of h′ that are after (t′ , x′ ) in the composition to h1 , and before (t, x)O . We have that s(t, x)O is a subsequence of h′ and h′ ◁∗OP h′′ holds just if s contains no moves from t. But, if s contained moves from t then the rightmost one such would be some (t, y)P . Moreover, in the composition towards h1 , the move would be scheduled with 1. The latter would break the conditions for trace composition as, at that point, the corresponding subsequence of h′1 has as leftmost move in t the P-move (t, x)P . We can show similarly that h′ ◁∗OP h′′ holds in the case that the permutation in h′1 is on consecutive moves (t, x)O (t′ , x′ ). Finally, the rearrangements in h′1 that do not affect moves shared with h′ can be treated in a simpler way: e.g. in the case of (t′ , x′ )(t, x)P consecutive in h′1 and swapped in h′2 , if (t, x)P does not appear in h′ then we can check that h′ cannot contain any t-moves between (t′ , x′ ) and (t, x) as the conditions for trace composition impose that only O is expected to play in that part of h′ (and any t-move would swap this polarity). ′ Now, since h′ ∈ JLK, Lemma 34 implies h′′ ∈ JLK. Take h2 to be h′′ ⩕σΠ,P h′2 , where σ ′ is obtained from σ following these move rearrangements. We then have h2 ∈ JL ∪ L2 K. Moreover, h1 ⊑ h2 thanks to h′1 ⊑ h′2 . Hence, h2 ∈ JL ∪ L2 K and h1 ⊑ h2 . Thus, L ∪ L1 ⊑ L ∪ L2 . E.2 Proof of Theorem 43 Proof. Let us consider the first sequencing case (the second one is dual), and assume that L1 , L2 ∶ Θ → Θ′ and L ∶ Θ′′ → Θ. Assume L1 ⊑enc L2 and suppose h1 ∈ JL ; L1 Kenc . By Theorem 53, h1 = h′ ⩕σΠ,P h′1 , where h′ ∈ JLK, h′1 ∈ JL1 K and method calls from Θ are always scheduled with 0. The fact that O cannot switch between L/K components in (threads of) h1 implies that the same holds for h′ , h′1 , hence h′ ∈ JLKenc and h′1 ∈ JL1 Kenc . Because L1 ⊑enc L2 , there exists h′2 ∈ JL2 Kenc such that h′1 ⊑ h′2 , i.e. h′1 (◁P O ∪ ◇)∗ h′2 . As before, some of the rearrangements necessary to transform h′1 into h′2 may concern actions shared by h′1 and h′ ; we need to check that these can lead to compatible h′′ ∈ JLKenc . Let h′′ be obtained by applying such rearrangements to h′ . We claim that h′ ◁∗OP h′′ . The transpositions covered by ◁P O are treated as in Lemma 36. Suppose now that (t′ , x′ )P K (t, x)OL are consecutive in h′1 but swapped in order to obtain h′2 , and (t, x)OL appears in h′ as (t, x)P K . Now, the move (t′ , x′ ) cannot appear in h′ as it is in L1 ’s K-component (L is the L-component of L1 ). Let s contain the moves of h′ that are after (t′ , x′ ) in the composition to h1 , and before (t, x)P K . We claim that s contains no moves from t, so h′ can be directly composed with h′2 as far as this transposition is concerned. Indeed, if s contained moves from t then, taking into account the encapsulation conditions, the leftmost one such would be some (t, y)OK . But the K-component of L is L1 , which contradicts the fact that the moves we consider are ′ consecutive in h′1 . Hence, taking h2 to be h′′ ⩕σΠ,P h′2 , where σ ′ is obtained from σ following the ◁P O move rearrangements, we have h2 ∈ JL ; L2 Kenc and h1 ⊑enc h2 . Thus, L ; L1 ⊑enc L ; L2 . The case of L ⊎ L1 ⊑enc L ⊎ L2 is treated in a similar fashion. In this case, because of disjointness, the moves transposed in h′1 do not have any counterparts in h′ . Again, we consider consecutive moves (t′ , x′ )P K (t, x)OL in h′1 that are swapped in order to 44 obtain h′2 . Let s contain the moves of h′ that are after (t′ , x′ ) in the composition to h1 , and before (t, x). As Θ1 , Θ1′ is first-order, (t, x)OL must be a return move and the t-move preceding it in h1 must be the corresponding call. The latter is a move in h′1 , which therefore implies that there can be no moves from t in s. Similarly for the other transposition case. E.3 Proof of Theorem 47 )-closed and L1 ⊑R L2 . Consider h1 ∈ Proof. For the first claim, suppose L is (R G JL ; L1 Kenc . By Theorem 53, h1 = h′ ⩕σΠ,P h′1 , where h′ ∈ JLKenc and h′1 ∈ JL1 Kenc . Since L1 ⊑R L2 , there exists h′2 ∈ JL2 Kenc such that (h′1 ↾ K) ⊑ (h′2 ↾ K) and (h′1 ↾ L) R (h′2 ↾ L). By the permutation-closure of R, we can pick h′2 not to contain any common names with h′ apart from those in the common moves of h′1 and h′ . Because )-closed, h′ ∈ JLKenc and (h′ ↾ L) = (h′1 ↾ L) and (h′1 ↾ L) R (h′2 ↾ L), L is (R G we can conclude that there exists h′′ ∈ JLKenc such that (h′′ ↾ K) = (h′2 ↾ L) and (h′ ↾ L) G (h′′ ↾ L). Applying the corresponding rearrangements to σ, we have that h′′ ′ ′ and h′2 are compatible, i.e. (h′′ ⩕σΠ,P h′2 ) ∈ JL ; L2 Kenc . Let h2 = h′′ ⩕σΠ,P h′2 . We want to show h1 ⊑R h2 . To that end, it suffices to make the following observations. – We have (h1 ↾ K) ⊑ (h2 ↾ K) because (h1 ↾ K) = (h′1 ↾ K), (h′1 ↾ K) ⊑ (h′2 ↾ K) and (h′2 ↾ K) = (h2 ↾ K). – We have (h1 ↾ L) G (h2 ↾ L) because (h1 ↾ L) = (h′ ↾ L), (h′ ↾ L) G (h′′ ↾ L) and (h′′ ↾ L) = (h2 ↾ L). Consequently L ; L1 ⊑G L ; L2 . Suppose now L1 ⊑R L2 . Consider h1 ∈ JL1 ; LKenc , i.e. h1 = h′1 ⩕σΠ,P h′ , where h′1 ∈ JL1 Kenc and h′ ∈ JLKenc . Because L1 ⊑R L2 , there exists h′2 ∈ JL2 Kenc such that h′1 ⊑R h′2 , i.e. (h′1 ↾ K) ⊑ (h′2 ↾ K) and (h′1 ↾ L)R(h′2 ↾ L). Define h′′ to be h′ in which (h′ ↾ L) = (h′1 ↾ K) was modified by applying the same rearrangements as those witnessing (h′1 ↾ K) ⊑ (h′2 ↾ K). Consequently h′ ◁∗P O h′′ . By Lemma 40, h′′ ∈ JLKenc . Moreover, (h′′ ↾ L) = (h′2 ↾ K). Consequently, h′2 and h′′ are compatible ′ for the corresponding σ ′ . Let h2 = h′2 ⩕σΠ,P h′′ ∈ JL2 ; LK. Then we get: – (h1 ↾ K) = (h′ ↾ K) = (h′′ ↾ K) = (h2 ↾ K); – (h1 ↾ L) = (h′1 ↾ L), (h′1 ↾ L)R(h′2 ↾ L), (h′2 ↾ L) = (h2 ↾ L). Consequently h1 ⊑R h2 and, hence, L1 ; L ⊑R L2 ; L. For the last claim, we observe that because of the type-restrictions, the elements of JL ⊎ Li Kenc are interleavings of histories from JLKenc and JLi Kenc . Consider now h1 ∈ JL ⊎ L1 Kenc , i.e. h1 = h′ ⩕σΠ,P h′1 , where h′1 ∈ JL1 Kenc and h′ ∈ JLKenc , and let h′2 ∈ JL2 Kenc be such that h′1 ⊑R h′2 . From our previous observation, we have that h′ can ′ still be composed with h′2 , for appropriate σ ′ . Thus, taking h2 = h′ ⩕σΠ,P h′2 , we have ′ ′ h2 ∈ JL ⊎ L2 Kenc and, moreover, h1 ⊑R h2 implies h1 ⊑R+ h2 . 45
6cs.PL
arXiv:0804.4116v1 [cs.SE] 25 Apr 2008 Under consideration for publication in Theory and Practice of Logic Programming 1 Design and Implementation of a Tracer Driver: Easy and Ecient Dynamic Analyses of Constraint Logic Programs∗ Ludovic Langevine† Mission Critical IT, Boulevard de France, 9, Bât A, 1420 - Braine-l'Alleud, Belgium (e-mail: [email protected]) Mireille Ducassé IRISA/INSA of Rennes, Campus de Beaulieu, 35 042 Rennes Cedex, France (e-mail: [email protected]) submitted 5 July 2006; revised 5 June 2007, 20 December 2007; accepted 23 April 2008 Abstract Tracers provide users with useful information about program executions. In this article, we propose a tracer driver. From a single tracer, it provides a powerful front-end enabling multiple dynamic analysis tools to be easily implemented, while limiting the overhead of the trace generation. The relevant execution events are specied by exible event patterns and a large variety of trace data can be given either systematically or on demand. The proposed tracer driver has been designed in the context of constraint logic programming; experiments have been made within GNU-Prolog. Execution views provided by existing tools have been easily emulated with a negligible overhead. Experimental measures show that the exibility and power of the described architecture lead to good performance. The tracer driver overhead is inversely proportional to the average time between two traced events. Whereas the principles of the tracer driver are independent of the traced programming language, it is best suited for high-level languages, such as constraint logic programming, where each traced execution event encompasses numerous low-level execution steps. Furthermore, constraint logic programming is especially hard to debug. The current environments do not provide all the useful dynamic analysis tools. They can signicantly benet from our tracer driver which enables dynamic analyses to be integrated at a very low cost. To appear in Theory and Practice of Logic Programming (TPLP), Cambridge University Press. KEYWORDS: Software Engineering, Debugging, Execution Trace Analysis, Execution Monitoring, Execution Tracing, Execution Visualization, Programming Environment, Constraint Logic Programming ∗ This work has been partly supported by the French RNTL OADymPPaC project and the ERCIM fellowship programme. † Most of the work has been done while Ludovic Langevine was at INRIA Rocquencourt, France L. Langevine and M. Ducassé 2 1 Introduction Dynamic program analysis is the process of analyzing program executions. It is generally acknowledged that dynamic analysis is complementary to static analysis; see for example the discussion of Ball (Ball 1999). Dynamic analysis tools include, in particular, tracers, debuggers, monitors and visualizers. Analysis 1 (e.g. a tracer) Development of ad hoc Analysis 2 (e.g. a monitor) instrumentation parts ... Analysis n (e.g. a visualizer) Fig. 1. The usual case: all dynamic analysis tools implement a dedicated instrumentation part In order to be able to analyze executions, some data must be gathered and some sort of instrumentation mechanisms must be implemented. The state-of-thepractice, illustrated by Fig. 1, is to re-implement the instrumentation for each new dynamic analysis tool. The advantages are, rstly, that the instrumentation is naturally and tightly connected to the analysis, and secondly, that it is specialized for the targeted analysis and produces relevant information. The drawback, however, is that this implementation usually requires much tedious work which has to be repeated by each tool's writer for each environment. This acts as a brake upon development of dynamic analysis tools. 1.1 A Tracer Driver to Eciently Share Instrumentations In this article we suggest that standard tracers can be used to give information about executions to several dynamic analysis tools. Indeed, Harrold et al. have shown that a trace consisting of the sequence of program statements traversed as the program executes subsumes a number of interesting other representations such as the set of conditional branches or the set of paths (Harrold et al. 1998). trace Analysis 1 (e.g. a standard tracer) Analysis 2 (e.g. a monitor) basic tracer ... Analysis n (e.g. a visualizer) Fig. 2. The generate-and-dump approach: the instrumentation part is shared but the amount of data is huge However, the separation of the extraction part from the analysis part requires Tracer Driver and Dynamic Analyses of CLP(FD) 3 requests relevant trace Analysis 1 (e.g. a standard tracer) Analysis 2 (e.g. a monitor) basic tracer with a tracer driver ... Analysis n (e.g. a visualizer) Fig. 3. Our supervised generation approach: with a tracer driver only the relevant part of the execution information is generated care. Indeed, as illustrated by Fig. 2, if execution information is systematically generated and dumped to the analysis tools, the amount of information that ows from the tracer to the analysis modules can be huge, namely several gigabytes for a few seconds of execution. Whether the information ows through a le, a pipe, or even main memory, writing such an amount of information takes so much time that the tools are not usable interactively. This is especially critical for debugging, even when it is automated, because users need to interact in real-time with the tools. Reiss and Renieris propose to encode and compact the trace information (Reiss and Renieris 2001). Their approach is used in a context where multiple tracing sources send information to the same analysis module. In this article, we propose another approach, more accurate when a single source sends information to (possibly) several analysis modules. As illustrated by Fig. 3, we have designed what we call a tracer driver, whose primary function is to lter the data on the y according to requests sent by the analysis modules. Only the necessary trace information is actually generated. This often drastically reduces the amount of trace data, and signicantly improves the performance. Therefore, the instrumentation module is shared among several analysis tools and there is very little slowdown compared to the solution where each analysis has its dedicated instrumentation. From a single tracer, the tracer driver provides a powerful front-end for multiple dynamic analysis tools while limiting the overhead of the trace generation. The consequence is that specifying and implementing dynamic analysis tools is much easier, without negative impact on the end-user. 1.2 Interactions between a Tracer and Analyzers In the following, we call a module that is connected to a tracer an analyzer . In its simplest form the analyzer is only the standard output, or a le, in which traces are written by a primitive tracer. Another form of analyzers is traditional debuggers, which are mere interactive tracers of executions. They handle the interaction once the execution is stopped at interesting points. They show some trace information and react on users' commands. More sophisticated debugging tools exhibit more sophisticated analyzers. A trace querying mechanism with a database avor can be connected to a tracer and let users investigate executions in a more thorough way. This has been done for example for C in the Coca tool (Ducassé 1999a) and for Prolog in the Opium tool (Ducassé 1999b). A real database can even be used L. Langevine and M. Ducassé 4 if on the y performance is not a big issue. This has been done for a distributed system in Hy + (Consens et al. 1994). Algorithmic debugging traverses an execution tree in an interactive way. Users answer queries and an algorithm focuses on nodes which seem erroneous to users while their children seem correct (Shapiro 1983). Note that the declarative debugger of Mercury is explicitly built on top of the Mercury tracer (MacLarty et al. 2005). Monitoring tools can be connected to tracers in order to supervise executions and collect data. For example, the Morphine tool for Mercury is able to assess the quality of a test set (Jahier and Ducassé 2002). The EMMI tool, for Icon, is able to detect some programming mistakes (Jeery and Griswold 1994). A number of visualization tools use traces to generate graphical views such as the DiSCiPl views (Deransart et al. 2000) for constraint logic programming. The latest trend in fault localization consists in mining sets of program executions to cross check execution traces, see for example (Jones et al. 2002; Jones and Harrold 2005; Denmat et al. 2005). The interaction modes between the tracer and the analyzers exhibited by the previous examples are all dierent and specic. For primitive tracers, simple visualization and trace mining, the tracer simply outputs trace information into a given channel. Traditional debuggers, trace query systems and declarative debuggers output information about executions and get user requests. When information is displayed and until the user sends a request, the execution is blocked. Monitors process the trace information on the y. They also block the execution until they have nished processing the current trace information but without any interaction with users. At present, all these tools are disjoint and dicult to merge. Therefore, further mechanisms are required in order to share a tracer among analyzers of dierent types. When the tracer has sent trace information to the analyzers, the above examples exhibit two behaviors: 1) the execution is blocked waiting for an answer from the analyzers; this is called synchronous interaction in the following; 2) the execution proceeds; this is called asynchronous interaction in the following. Our tracer driver includes both mechanisms. It enables dierent interaction modes between a tracer and analyzers to be integrated in one single tool. This has several advantages. Firstly, users do not switch tools to achieve dierent aims. They use a single tool to trace, debug, monitor and visualize executions. Secondly, integrating all the possible usages results in a more powerful tool than the mere juxtaposition of dierent tools. For example, one can, simultaneously, check for known bug patterns, and collect data for visualization. Whenever a bug is encountered the tool can switch to a synchronous debugging session, using the already collected visualization data. The visualization tool can also change the granularity of the collected data depending on the current context. 1.3 Debugging of Constraint Logic Programs The proposed tracer driver has been designed in the context of constraint logic programming. Experiments have been made within GNU-Prolog (Diaz 2003). Pro- Tracer Driver and Dynamic Analyses of CLP(FD) 5 grams with constraints are especially hard to debug (Meier 1995). The numerous constraints and variables involved make the state of the execution dicult to grasp. Moreover, the complexity of the ltering algorithms as well as the optimized propagation strategies lead to a tortuous execution. As a result, when a program gives incorrect answers, misses expected solutions, or has disappointing performance, the developer gets very little support from the current programming environments to improve the program. This issue is critical because it increases the expertise required to develop constraint programs. Some previous papers have addressed this critical issue. Most of them are based on dynamic analyses. During the execution, some data are collected in the execution so as to display some graphical views, compute some statistics and other abstraction of the execution behavior. Those data are then examined by the programmer to gain a better understanding of the execution. For instance, a display of the search-tree shows users how the search heuristics behave (Fages 2002). Adding some visual clues about the domain propagation helps users locate situations where propagation is not strong enough (Carro and Hermenegildo 2000; Bracchi et al. 2001). A structured inspection of the store helps to investigate constraints behavior (Goualard and Benhamou 2000). A more detailed view of the propagation in specic nodes of the search-tree gives a good insight to nd redundant constraints or select dierent ltering algorithms (Simonis and Aggoun 2000). A common observation is that there is no ultimate tool that would meet all the debugging needs. A large variety of complementary tools already exists, ranging from coarse-grained abstraction of the whole execution to very detailed views of small sub-parts, and even application-specic displays. As a matter of fact, none of the current environments in CLP contain all of the interesting features that have already been identied. Each of these features requires a dedicated instrumentation of the execution, or a dedicated annotation of the traced program, to collect the data they need. Those instrumentations are often hard to make. Yet, many of those dynamic analyses could be built on top of low-level tracers. The generated traces can be structured and abstracted by analyzers in order to produce high-level views. With our tracer driver, it is easy to develop and explore dynamic analyses with diverse abstraction levels. For instance, we can rst compute a general view of the search-tree, tracing only the execution events related to the search-tree construction and ignoring the propagation events or just computing some basics statistics about propagation stages. Such an analysis quickly gives a general picture of the execution. Then, a more specic analysis of, say, a subset of variables or constraints may provide further details about a sub-part of the program but may need a more voluminous execution trace. 1.4 Contributions The contributions of this article are threefold. Firstly, it justies the need for a tracer driver in order to be able to eciently integrate several dynamic analyses within a single tool. In particular, it emphasizes that both synchronous and asynchronous communications are required between the tracer and the analyzer. L. Langevine and M. Ducassé 6 Secondly, it describes in breadth and in some depth the mechanisms needed to implement such a tracer driver: 1) the patterns to specify what trace information is needed, 2) the language of interaction between the tracer driver and the analyzers and 3) the mechanisms to eciently lter trace information on the y. Lastly, an implementation has been achieved inside GNU-Prolog. The paper assumes propagation based solvers only for purposes of exposition. The mechanisms of the tracer driver are independent of the traced language, they are still applicable to solvers that do not use propagators. Experimental measurements on CLP(FD) executions show that this architecture increases the trace relevance and drastically speeds up trace generation and communication. More precisely, the experiments show that 1. The overhead of the core tracer mechanisms is small, therefore the core tracer can be permanently activated 2. The tracer driver overhead is inversely proportional to the average time between two traced events. It is acceptable for CLP(FD). 3. There is no overhead in the ltering mechanisms when searching simultaneously for several patterns. 4. The tracer driver overhead is predictable for given patterns. 5. The tracer driver approach that we propose is more ecient than sending over a default trace, even to construct sophisticated graphical views. 6. Answering queries is orders of magnitude more ecient than displaying traces. 7. There is no need to restrict the trace information a priori. 8. The performance of our tool is comparable to the state-of-the-practice while being more powerful and more generic. Whereas the principles of the tracer driver are independent of the traced programming language, it is best suited for high-level languages, such as constraint logic programming, where each traced execution event encompasses numerous low-level execution steps. In the following, Section 2 gives an overview of the tracer driver and in particular the interactions it enables between a tracer and an analyzer. Section 3 species the nature of patterns. Section 4 presents the requests that an analyzer can send to our tracer and how they are processed. Section 5 describes in detail our ltering mechanism and its implementation. Section 6 discusses the requirements on the tracer for the overall architecture to be ecient. Section 7 gives experimental results and shows the eciency of the tracer driver mechanism. Section 8 discusses related work. 2 Overview of the tracer driver This section presents an overview of the tracer driver architecture and, in particular, the interactions it enables between a tracer and analyzers. The tracer and the analyzers are run as two concurrent processes. The tracer spies the execution and can output, when needed, certain trace events with some attached data. Each analyzer listens to the trace and processes it. Tracer Driver and Dynamic Analyses of CLP(FD) Execution Observed by tracer Event i Tracer Driver 7 Analyzer Mediator Event patterns Event Filtering Event i+1 Event i+2 Event i+3 Trace data Event i+4 Asynchronous Handler Event i+5 Trace a nd sync hr. data Send more data! Data Retrieving Synchronous Additional Trace Data Modify patterns! Handler Go! Event i+6 ... Fig. 4. Asynchronous and synchronous interactions between the tracer and the analyzer mediator As already mentioned in the introduction, both synchronous and asynchronous interactions are necessary between the tracer and the analyzers. On the one hand, some analyzers are highly interactive. Users may ask for more information about the current event than what is provided by default. In that cases, it is important that the execution is blocked until all the analyzers notify it that it can proceed. On the other hand, if the analyzers passively collect information there is no need to block the execution. An execution trace is a sequence of observed execution events that have at- event patterns. An event pattern is a condition on the attributes of an event (see details tributes. The analyzers specify the events to be observed by the means of in Section 3). The tracer driver manages a set of active event patterns. Each execution event is checked against the set of active patterns. An event matches an L. Langevine and M. Ducassé 8 event pattern if and only if the pattern condition is satised by the attributes of this event. An asynchronous pattern species that, at matching trace events, some trace data synchronous pattern are to be sent to analyzers without freezing the execution. A species that, at matching trace events, some trace data are to be sent to analyzers. The execution is frozen until the analyzers order the execution to resume. An handler event is a procedure dened in an analyzer, which is called when a matching event is encountered. Fig. 4 illustrates the treatment of the two types of patterns. The execution is presented as a sequence of elementary blocks (the execution events). An analyzer mediator gathers the patterns requested by the analyzers and distributes the trace data sent by the tracer driver to the analyzers (see detailed description Section 4). At each trace event, the tracer driver is called to lter the event. If the current event does not match any of the specied active patterns, the execution continues (events i, i + 1, i + 2, i + 4, i + 6). If the current event matches an active pattern, some trace data are sent to the analyzer mediator (events i + 3, i + 5). If the matched pattern is asynchronous the data is processed by the relevant analyzer in an asynchronous way (event i + 3). If the pattern is synchronous the execution is frozen, waiting for a i + 5). The analyzer processes the sent data and can query from the analyzer (event ask for more data about the state of the execution. The tracer driver can retrieve useful data about the execution state and send them to the analyzer on demand. The analyzer can also request some modications of the active patterns: add new patterns or remove existing ones. When no analyzer has any further request to make about the current event, the analyzer mediator sends the resuming command to the tracer driver (go command). The tracer then resumes the execution until the next matching event. The architecture enables the management of several active patterns. Each pattern is identied by a label. A given execution event may match several patterns. When sending the trace data, the list of (labels of ) matched patterns is added to the trace. Then, the analyzer mediator calls a specic handler for each matched pattern and dispatches relevant trace data to it. If at least one matched pattern is synchronous, the analyzer mediator waits for every synchronous handler to nish before sending the resuming command to the tracer driver. From the point of view of a given event handler, the activation of other handlers on the same execution event is transparent. This article focuses more on the tracer driver than on the analyzer mediator. On the one hand, the design and implementation of the tracer driver is critical with respect to response time. Indeed it is called at each event and executions of several millions of events (see Section 7) are very common. Every overhead, even the tiniest, is therefore critical. On the other hand, the implementation of the analyzer mediator is much less critical because it is called only on matching events. Furthermore, its implementation is much easier. Tracer Driver and Dynamic Analyses of CLP(FD) 9 3 Event Patterns As already mentioned, an event pattern is a condition on the attributes of events. It consists of a logical formula combining elementary conditions on the attributes. This section summarizes the information attached to trace events, species the format of the event patterns and gives examples of patterns. 3.1 Trace Events Some information is attached to each trace event. This section summarizes the format of trace event information used in this article; a more detailed description can be found in (Langevine et al. 2004). The actual format of trace event information has no inuence on the tracer driver mechanisms. The important issue is that events have attributes and that some attributes are specic to the type of events. Note that the pattern language is independent of the traced language. A constraint program manipulates variables and constraints on these variables. Each variable has a domain, a nite set of possible values. The aim of a constraint program is to nd a valuation (or the best valuation, given an objective function) of the variables such that every constraint is satised. To do so, constraint solvers implement numerous algorithms coming from various research areas, such as operation research. Traditionally in logic programming, the type of events is called a port. There are 14 possible event types in the tracer we use. Those types of events are partitioned into two classes: control events and propagation events. The control events are related to the handling of the constraint store and the search: • new variable species that a new variable is introduced; • new constraint species that the solver declares a new constraint; • post species that a declared constraint is introduced into the store as the active one; • new child species that the current solver state corresponds to a new node of the search-tree; • jump to species that the solver back-jumps from its current state to a previous choice-point; • solution species that the current solver state is a solution (a toplevel success); • failure species that the current state is inconsistent. Six ports describe the domain reductions and the constraint propagation: • reduce species that a domain is being reduced, this generates domain updates which have to be propagated; • suspend species that the active constraint cannot reduce any more domains and is thus suspended; • • • • entail species that the active constraint is true; reject species that the active constraint is unsatisable; schedule species that a domain update is selected by the propagation loop; awake species that a constraint which depends on the scheduled domain update is awakened; L. Langevine and M. Ducassé 10 1 2 3 4 5 6 7 8 9 10 11 12 newVariable v1=[0-268435455] newVariable v2=[0-268435455] newConstraint c1 fd_element([v1,[2,5,7],v2]) reduce c1 v1=[1,2,3] delta=[0,4-268435455] reduce c1 v2=[2,5,7] delta=[0-1,3-4,6,8-268435455] suspend c1 newConstraint c4 x_eq_y([v2,v1]) reduce c4 v2 =[2] delta=[5,7] reduce c4 v1 =[2] delta=[1,3] suspend c4 awake c1 reject c1 Fig. 5. A portion of trace • end of trace noties the end of the tracing process. 3.2 Event Attributes Each event has common and specic attributes. Attributes are data about the execution event. The common attributes are: the port, a chronological event number, the depth of the current node in the search-tree, the solver state (containing all the domains, the full constraint store and the propagation queue), and the time user spent since the beginning of the execution. The specic attributes depend on the port. For example, the specic attributes for port new variable are the variable identier and its initial domain. For the ports related to the search-tree the only specic attribute is the node label. Specic attributes for other ports are described in (Langevine et al. 2004). Fig. 5 presents the beginning of a trace of a toy program in order to illustrate the events described above. This program, fd_element(I, [2,5,7],A), (A#=I ; A#=2), species that A is a nite domain variable which is in {2, 5, 7} and I is the index of the value of A in this list; moreover A is either equal to I or equal to 2. The second alternative is the only feasible one. The trace can be read as follows. The rst two events are related to the introduction of two variables v1 and v2, corresponding respectively to I and A. variables are always created with the maximum domain (from 0 to the rst constraint is created: fd_element In Gnu-Prolog, 228 − 1 ). Then (event #3). This constraint makes two domain reductions (events #4 and #5): the values removed from the domain of the rst variable (I) are listed in of A becomes {2, 5, 7}, delta, the domain becomes {1, 2, 3} the constraint is suspended (event #6). The next constraint, #7). Two reductions are done on variables and I to be equal is 2 and the domain the only consistent values so far. After these reductions, A and I, A#=I, is added (event A the only possible value for (events #8 and #9). After these reductions, the constraint A and I 2, I cannot be the rank of A. Indeed, the rank of 2 is 1 and the value at rank 2 is 5. The constraint is therefore rejected (event #12). The execution continues and nds the solution (A=2, I=1) This requires 20 other events not shown is suspended (event #10). The rst constraint is awoken (event #11). If are both equal to here. Tracer Driver and Dynamic Analyses of CLP(FD) 11 pattern ::= label: when evt_pattern op_synchro action_list op_synchro::= do | do_synchro action_list::= action , action_list | action action ::= current(list_of_attributes ) | call(procedure ) evt_pattern::= evt_pattern or evt_pattern | evt_pattern and evt_pattern | not evt_pattern | ( evt_pattern ) | condition condition ::= attribute op2 value | op1 (attribute) op2 ::= < | > | = | \= | >= | =< | in | notin contains | op1 ::= | (1) (2) (3) (4) (5) | true notcontains isNamed Fig. 6. Grammar of event patterns 3.3 Patterns We use patterns similar to the path rules of Bruegge and Hibbard (Bruegge and Hibbard 1983). Fig. 6 presents the grammar of patterns. A pattern contains four parts: a label, an event pattern, a synchronization operator and a list of actions. An event pattern is a composition of elementary conditions using logical conjunction, disjunction and negation. It species a class of execution event. A synchronization operator tells whether the pattern is asynchronous (do) or synchronous (do_synchro). An action species either to ask the tracer driver to collect at- list_of_attributes )), tribute values (current( cedure call(procedure ). or to ask the analyzer to call a pro- Note that the procedure is written in a language that the analyzer is able to execute. This language is independent of the tracer driver. An elementary condition concerns an attribute of the current event. There are several kinds of attributes. Each kind has a specic set of operators to build elementary conditions. For example, most of the common attributes are integer (chrono, depth, node label). Classical operators can be used with those attributes: equality, disequality (6=), inequalities (<, ≤, > and ≥). The port attribute is the type of the current event. It has a small set of possible values. The following operators can be used with the port attribute: equality and disequality (= and and two set operators, in and notin. 6=) Constraint solvers manipulate a lot of con- straints and variables. Often, a trace analysis is only interested in a small subset of them. Operators in and notin, applied to identiers of entities or name of the variables, can specify such subsets. Operators contains and notcontains are used to express conditions on domains. This set of operators is specic to the type of execution we trace. It could be extended to cope with other types of attributes. 12 L. Langevine and M. Ducassé visu_tree: when port in [choicePoint, backTo, solution, failure] do current(port=P and node=N and depth=D and usertime=Time), call search_tree(P,N,D,Time) visu_cstr: when port = post do current(cstr=C and cstrRep=Rep and varC(cstr)=VarC), call new_cstr(C, Rep, VarC) visu_prop: when port = reduce and isNamed(var) and (not cstrType='assign') and delta notcontains [maxInt] do current(cstr=C and var=V), call spy_propag(C,V) leaf: when port in [solution, failure] do_synchro current(port=P and node=N and depth=D), call new_leaf(P,N,D) symbolic: when port in [reduce,suspend] and (cstrType = 'fd_element_var' or cstrType = 'fd_exactly') do_synchro call symbolic_monitor Fig. 7. Examples of patterns for visualization and monitoring Fig. 8. Two trace-based views based on the pattern visu_cstr of Figure 7. 3.4 Examples of Patterns Fig. 7 presents ve patterns that can be activated simultaneously. The rst three patterns are visualization oriented: the rst one, visu_tree, aims at constructing the search-tree (creation of a choice-point, backtracking, failure and solution are the relevant ports) with few details about each node: the label of the node, its depth Tracer Driver and Dynamic Analyses of CLP(FD) 13 and the chronological time stamp of the event. This enables the display of nodes to be weighted by the time spent in their sub-tree. The second pattern, visu_cstr, requests the trace of each constraint-posting with the identier of the constraint (a unique integer), its representation (the name of the constraint with its parameters) and the list of the involved variables varC ). Fig. 8 gives two screen-shots of visualizations that are built using pattern ( visu_cstr. The rst picture is generated by CLPGUI (Fages 2002). The second picture is generated by Pavot, a graphical tool developed at INRIA Rocquencourt and connected to GNU-Prolog tracer driver The third pattern, visu_prop, 1 . requests the trace of all the domain reductions made by constraints that do not come from the assignment procedure and that do not remove the maximal integer value. It stores the reducing constraint and the reduced variable. Those data can be used to compute some statistics and to visualize the impact of each constraint on its variables. Those three patterns are asynchronous: the requested data are sucient for the visualization and the patterns do not have to be modied. The fourth pattern, leaf, synchronizes the execution at each leaf of the search- tree (solution or failure). At those events, the new_leaf function can interact with the tracer to investigate the execution state. The last pattern, symbolic, is more monitoring-oriented: it freezes the execu- tion at each domain reduction made by a symbolic constraint such as variables) or element (on exactly. This pattern allows the monitoring of the ltering algorithms used for these two constraints. 4 Analyzer Mediator The analyzer mediator is the interface between the tracer driver and the analyzers. It species to the tracer driver what events are needed and may execute specic actions for each class of relevant events. The mediator can supervise several analyses at a time. Each analysis has its own purpose and uses specic pieces of trace data. The independence of the concurrent analyses is ensured by the mediator that centralizes the communication with the tracer driver and distributes the trace data to the ongoing analyses. The advantage is that if a piece of trace information is needed by several analyses, it is sent over the interface only once. When a synchronous event has been sent to the mediator, the requests that an analyzer can send to the driver are of three kinds. Firstly, the analyzer can ask for additional data about the current event. Secondly, the analyzer can modify the set of active event patterns, to be checked by the tracer driver. Thirdly, the analyzer can notify the tracer driver that the execution can be resumed. The actual requests are as follows. current 1 species a list of event attributes to retrieve in the current execution http://contraintes.inria.fr/arnaud/pavot/ L. Langevine and M. Ducassé 14 step :reset, add([step:when true dosynchro call(tracer_toplevel)]), go. skip_reductions :current(cstr = CId and port = P), reset, ( P == awake -> add([sr:when cstr = CId and port in [suspend,reject,entail] dosynchro call(tracer_toplevel)]), ; add([step:when true do_synchro call(tracer_toplevel)]) ), go. Fig. 9. Implementation of two tracing commands event. The tracer retrieves the requested pieces of data. It sends the data as a list of pairs (attribute, value). reset deletes all the active event patterns and their labels. remove deletes the active patterns whose labels are specied in the parameter. add inserts in the active patterns, the event patterns specied in the parameter, following the grammar described in Figure 6. go noties the tracer driver that the traced execution is to be resumed. Fig. 9 illustrates the use of the primitives to implement two tracing commands. Let us assume that the analyzer is a (possibly simplied) Prolog interpreter, as for example in Opium (Ducassé 1999b). Command step enables execution to go to the very next event. It simply resets all patterns and adds one that will match any event and call, with synchronous interactions, the tracer toplevel. Command skip_reductions enables execution to skip the details of variable domain reduc- tions when encountering the awakening of a constraint. It rst checks the current port. If it is awake it asks to go to the suspension of this constraint. There, the user will, for example, be able to check the value of the domains after all the reductions. If the command is called on an event of another type, it simply acts as step. 5 Filtering Mechanism This section describes in detail the critical issue of the ltering mechanism. At each execution event, it is called to test the relevance of the event with respect to the active patterns. Notice that the execution of a program with constraints can lead to several millions of execution events per second. Therefore, the eciency of the event ltering is a key issue. In the following, we rst describe the algorithm of the tracer driver. Then we specify the automata which drive the matching of events against active patterns. We discuss some specialisation issues. We give some details about the incremental handling of patterns. Lastly, we emphasize that event attributes are computed only upon demand. Tracer Driver and Dynamic Analyses of CLP(FD) 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 15 proc tracerDriver(P : set of active patterns) tagged ← ∅ for each p ∈ P do if match(p) then tagged ← tagged ∪ {p} end for T ← {requested_data(p) | p ∈ tagged} send_trace_data(T, label(tagged)) if synchronous(tagged) 6= ∅ then notify(synchronous(tagged)) repeat request ← receive_from analyzer() execute(request) until request = go end if end proc Fig. 10. Algorithm of the Tracer Driver 5.1 Tracer Driver Algorithm When an execution event occurs, the tracer is called. The tracer collects some data to maintain its own data structures and then calls the tracer driver. The algorithm of the tracer driver is given in Fig. 10. The ltering mechanism can handle several active event patterns. For each pattern, if the current event matches the pattern the latter is tagged as activated; whatever the matching result, the next pattern is checked (lines 3-5). When no more patterns have to be checked, the tagged patterns are processed; the union of requested pieces of data is sent as trace data with the labels of the tagged patterns (lines 6-7). If at least one synchronous pattern is tagged, a signal is sent to the analyzer; the tracer driver waits for requests coming from the analyzer and processes them until the go primitive is sent by the analyzer (lines8-14). 5.2 Pattern Automata The matching of an event against a pattern is driven by an automaton where each state is labeled by an elementary condition with two possible transitions: true or false. The automaton has two nal states, true and false. If the true state is reached, the event is said to match the pattern. Each automaton results from the compilation of an event pattern. This compilation is inspired by the evaluation of Boolean expressions in imperative languages. It has been proven that it minimizes the number of conditions to check (Wilhelm and Maurer 1995). Examples of automata are given in Figure 11, section 5.4. 5.3 Specialization According to the Port As seen in Section 3, the port is a special attribute since it denotes the type of the execution event being traced. A port corresponds to specic parts of the solver code where a call to a function of the tracer has been hooked. For example, in the tracer we use, there are four hooks for the reduce port, embedded into four specic L. Langevine and M. Ducassé 16 Ports choicePoint backTo solution failure post reduce awake suspend entail reject schedule newConstraint newVariable endOfTrace visu_tree isNamed(var) Yes No port depth current node time cstrType=’assign’ Yes Fail Yes No [268435455] No Asyn c hro n o us visu_cstr delta cstr cstrRep varC(cstr) current leaf cstrType=’fd_element_var’ Yes No cstrType=’fd_exactly’ Yes No Fail current port node depth Asyn c hro n o us visu_prop S yn c hro n o us symbolic current S yn c hro n o us cstr var Asyn c hro n o us Fig. 11. Internal representation of the ve patterns of Figure 7 functions that make domain reductions in four dierent ways. Furthermore, specic attributes depend on the port. As a consequence, the port is central in the pattern specication. For most patterns, a condition on the port will be explicit. When an event occurs, it is useless to call the tracer driver if no pattern is relevant for the port of this event. Therefore, for each port, a ag in the related hooks indicates whether the port appears in at least one pattern. This simple mechanism avoids useless calls to the tracer driver. 5.4 Examples of Pattern Automata Fig. 11 shows the internal representation of the ve patterns presented in Figure 7, Section 3.4. The 14 ports are represented on the left-hand side. The irrelevant ports are in italic. The relevant ports are linked to their corresponding patterns. Only two automata are necessary since three of the patterns check the port only. A set of actions is assigned to each automaton. This set is attached to synchronous or asynchronous and the label of the pattern. 5.5 Incremental Pattern Management Since each active pattern is a specic automaton (or a list of specic automata when split), the add primitive has just to compile the new n patterns into m automata, linked them with their respective ports and store the labels with the lists of resulting automata. The remove primitive has just to delete the automata associated to the specied labels and to erase the dead links. After each operation, the port-ltering ags are updated so as to take into account the new state of the active patterns. 6 Prototype Implementation In this Section, we briey present the prototype implementation. In particular, in order for the overall architecture to be ecient, it is essential that the tracer is lazy. Trace information must not be computed if it is not explicitly required by a pattern. Indeed, an execution has many events and events potentially have many Tracer Driver and Dynamic Analyses of CLP(FD) 17 attributes. Most of them are not straightforwardly available, they have to be computed from the execution state or from the debugging data of the tracer. Systematically computing all the attributes at all the execution events would be terribly inecient. Fortunately, not all the attributes need to be computed at each event. According to the active patterns, only a subset of the attributes is needed: rstly, the attributes necessary to check the relevance of the current event with respect to the patterns, and secondly, the attributes requested by the patterns in case of matching. Therefore, the tracer must not compute any trace attribute before it is needed. When a specic attribute is needed, it is computed and its value is stored until the end of the checking of the current event. If an attribute is used in several conditions, it is computed only once. The tracer implemented in the current prototype, Codeine, strictly follows this guideline. Some core tracer mechanisms are needed to handle the debugging information (see (Langevine et al. 2003) for more details). As shown in Sec. 7, the overhead induced by these mechanisms is marginal, even though constraint solvers do manipulate large and complex data. Currently, Codeine is implemented in 6800 lines of C including comments. The tracer driver, including the communication mechanisms, is 1700 lines of C. The codeine tracer including the tracer driver is available under GNU Public Licence. The set of debugging and visualization tools called Pavot has been developed by 2 Arnaud Guillaume and Ludovic Langevine. Both systems are available on line . 7 Experimental Results This section assesses the performance of the tracer driver and its eects on the cost of the trace generation and communication. Sections 7.1 and 7.2 describe the methodology of the experiments and the experimental setting. Section 7.3 lists the benchmark programs. Sections 7.4, 7.5 and 7.6 respectively discuss the tracer overhead, the tracer driver overhead and the communication overhead. 7.1 Methodology of the Experiments When tracing a program, some time is spent in the program execution (Tprog ), some time is spent in the core mechanisms of the tracer (∆tracer ), some time is spent in the tracer driver (∆driver ), some time is spent generating the requested trace and sending it to the analysis process (∆gcom ), and lastly some time is spent in the analyses (∆ana ). Hence, if we call T the execution time of a traced and analyzed program, we approximately have: T ' Tprog + ∆tracer + ∆driver + ∆gcom + ∆ana . 2 They can be retrieved at http://contraintes.inria.fr/langevin/codeine/ http://contraintes.inria.fr/arnaud/pavot/. and L. Langevine and M. Ducassé 18 The mediator is a simple switch. The time taken by its execution is negligible compared to the time taken by the simplest analysis, namely the display of trace information. Trace analysis takes a time which can vary considerably according to the nature of the analysis. The focus of this article is not to discuss which analyses can be achieved in reasonable time but to show that a exible analysis environment can be oered at a low overhead. Therefore, in the following measurements ∆ana = 0. 7.2 Experimental Setting The experiments have been run on a PC, with a 2.4 GHz Pentium iv , 512 Kb of cache, 1 Gb of RAM, running under the GNU/Linux 2.4.18 operating system. We used the most recent stable release of GNU -Prolog (1.2.16). The tracer is an instru- mentation of the source code of this version and has been compiled by The execution times have been measured with the GNU gcc-2.95.4. -Prolog proling facility whose accuracy is said to be 1 ms. The measured executions consist of a batch of executions such that each measured time is at least 20 seconds. The measured time is the sum of system and user times. Each experimental time given below is the average time of a series of ten measurements. In each series, the maximal relative deviation was smaller than 1 %. 7.3 Benchmark Programs The 9 benchmark programs 3 are listed in Table 1, sorted by increasing number of trace events. Magic(100), square(4), golomb(8) and golfer(5,4,4) are part of CSPLib, a benchmark library for constraints by Gent and Walsh (Gent and Walsh 1999). The golomb(8) program is executed with two strategies which exhibit very dierent response times. Those four programs have been chosen for their signicant execution time and for the variety of constraints they involve. Four other programs have been added to cover more specic aspects of the solver mechanisms: Pascal Van Hentenryck's bridge problem, implementation of (Diaz 2003); two instances of the n-queens problem; and propag, which proves the infeasibility of 1 ≤ x ∧ y ≤ 70000000 ∧ x < y ∧ y < x. The interest of the latter is the long stage of propagation involving one of the simplest and the most optimized constraints GNU -Prolog provides: bound consis- tency for a strict inequality. Therefore, this program the worst kind of case for the propagation instrumentation. The benchmark programs have executions large enough for the measurements to be meaningful. They range from 200,000 events to about 400 millions events. Furthermore, they represent a wide range of CLP(FD) programs. The third column gives the size of the traces of the benchmarked programs for the default trace model. All executions but the smallest one exhibit more than a 3 Their source code is available at http://contraintes.inria.fr/langevin/codeine/benchmarks Tracer Driver and Dynamic Analyses of CLP(FD) Program #events 6 (10 ) Trace Size Tprog (ms) ε (ns) Rtr. Dev. for (Gb) bridge queens(256) magic(100) square(24) golombF golomb golfer(5,4,4) propag queens(14) 19 Tx in % 0.2 0.1 14 72 1.21 0.8 1.5 173 210 1.14 3.2 1.4 215 66 1.03 4.2 20.8 372 88 1.05 15.5 3.4 7,201 464 1.01 38.4 7.9 1,721 45 1.00 1.05 61.0 >30 3,255 53 280.0 >30 3,813 14 1.28 394.5 >30 17,060 43 1.08 ≤ 0.4 ≤ 0.2 ≤ 0.2 ≤ 0.6 ≤ 0.4 ≤ 0.5 ≤ 0.7 ≤ 1.0 ≤ 0.4 Table 1. Benchmark Programs and tracer overhead gigabyte, for executions sometimes less than a second. It is therefore not feasible to systematically generate such an amount of information. As a matter of fact measuring these sizes took us hours and, in the last three cases, exhausted our patience! Note that the size of the trace is not strictly proportional to the number of events because dierent sets of attributes are collected at each type of events. For example, for domain reductions, several attributes about variables, constraints and domains are collected while other types of events simply collect the name of the corresponding constraint. The fourth column gives run by GNU T ε= Tprog , the execution time in ms of the program simply -Prolog. The fth column shows the average time of execution per prog #events . It is between 14 ns and 464 ns per event. For most of the suite is around 50ns. The three notable exceptions are (ε = 14 ns), event (ε = 210 ns) and golombF (ε = 464 propag ns). The low ε ε queens(256) is due to the eciency of the propagation stage for the constraints involved in this computation. The large εs are due to a lower proportion of ne-grained events. 7.4 Tracer Overhead The sixth column of Table 1 also gives the results of the measurements of the overhead of the core tracer mechanisms, Rtr(acer) , Rtr(acer) = which is dened as the ratio: Ttracer Tprog . where the measure of Ttracer ' Tprog + ∆tracer is the execution time of the program run by the tracer without any pattern activated. The tracer maintains its own data for all events. However, no attribute is calculated and no trace is generated. The seventh column gives the maximum deviation for Tprog and Ttracer . L. Langevine and M. Ducassé 20 1a. when port=post and isNamed(cname) do current(port,chrono,cident). 2a. when port=reduce and (isNamed(vname) and isNamed(cname)) do current(port,chrono,cident). 3a. when chrono=0 do current(chrono). 4a. when depth=50000 or (chrono>=1 and node=9999999) do current(chrono,depth). 5a: patterns 1a, 2a, 3a and 4a activated simultaneously. Fig. 12. Patterns used to measure the tracer driver overhead Core tracer mechanisms can be permanently activated. ecutions Rtracer For all the measured ex- is less than 30% in the worst case, and less than 5% for ve traced Rtracer programs. The results for are very positive; they mean that the core mech- anisms of the tracer can be systematically activated. Users will hardly notice the overhead. Therefore, while developing programs, users can directly work in traced mode; they do not need to switch from untraced to traced environments. This is a great comfort. As soon as they need to trace they can immediately get information. 7.5 Tracer Driver Overhead The measure of Tdriver ' Tprog + ∆trace + ∆driver is the execution time of the program run by the tracer with the ltering procedure activated for generic patterns. Only the attributes necessary for the requested patterns are calculated at relevant events. In order for ∆gcom to be zero, the patterns are designed such that no event matches them. One run is done per pattern. The patterns are listed in Figure 12. Pattern 1a 2a is reduce events trace is checked on few events and on one costly attribute only. Pattern checked on two costly attributes and on numerous events. Indeed, the main mechanism of the propagation and they are signicantly more numerous than the other types of events. Pattern cheap attribute. Pattern 4a 3a is checked on all events and on one is checked on all events and systematically on three attributes. In order to measure the overhead of the tracer driver, for each of the 5 patterns a ratio Rpattern is computed for all benchmark programs : Rpattern = Tdriver Tprog =1+ ∆pattern driver Tprog . Figure 13 displays all the ratios compared to the average time per event (ε) of the programs. The smallest value of ε, 14, corresponds to program propag and the biggest value, 464, corresponds to program golombF. In addition, the gure shows the curve Sigma R = R1a + R2a + R3a + R4a − 3, that adds the overheads of the four separated patterns. Tracer Driver and Dynamic Analyses of CLP(FD) R5a R1a R2a R3a R4a Sigma R 8 7 Performance Ratio 21 6 5 4 3 2 1 0 50 100 150 200 250 300 350 Time per execution event (ns) 400 450 500 Fig. 13. Cost of the tracer driver mechanisms for patterns 1a to 5a compared to Tracer driver overhead is acceptable. Rpattern 5a ε, ε In Figure 13, for all but one program, is negligible for the very simple patterns and less than 3.5 for pattern which is the combination of the other four patterns. For programs with a large even searching for pattern 5a is negligible. In the worst case, for propag, an overhead ratio of 8 is still acceptable. No overhead for simultaneous search for patterns. When n patterns are checked (n − 1)Ttracer compared to the search in sequence program to be executed n times instead of one time. Figure 13 the curve Sigma R = R1a + R2a + R3a + R4a − 3, is above the simultaneously they already save which requires the further shows that curve of R5a . Hence (1|2|3|4)a 2a 3a 4a ∆1a driver + ∆driver + ∆driver + ∆driver > ∆driver . This means that not only is there no overhead in the ltering mechanism induced by the simultaneous search, but there is even a minor gain, due to the factorization in the automata described Section 5.2. Tracer driver overhead is predictable. The measured points of Figure 13 can be interpolated with curves of the form Rdriver = a + b/ε. L. Langevine and M. Ducassé 22 1e+09 R5a nb. of evts. (log) propag 8 1e+08 7 Performance Ratio 6 1e+06 5 100000 4 10000 queens(14) 3 golfer(5,4,4) Number of Events 1e+07 1000 golomb square(24) magic(100) bridge 2 1 0 50 100 queens(256) 100 golombF 150 200 250 300 350 Time per execution event (ns) 400 450 10 500 Fig. 14. Cost of the tracer driver mechanisms compared to the number of events Figure 14 recalls the curve for pattern 5a and gives the number of events. It shows that there is no correlation between the size of the trace and the tracer driver overhead. Those results mean that the tracer and tracer driver overheads per event can be independent of the traced program. Indeed, let us assume that that ∆tracer = N δtracer approximated to constants depending on the patterns and these constants are and ∆driver = N δdriver δdriver where N is the number of events of an execution, δtracer and are the average time per event taken respectively by the core tracer mech- anism and the tracer driver, for all the programs. We have also already assumed that Tdriver ' Tprog + ∆tracer + ∆driver , and we have Rdriver = Tdriver Tprog , and Tprog = N ε therefore Rdriver ' Tprog +∆tracer +∆driver Tprog Rdriver ' 1 + N (δtracer +δdriver ) Nε Rdriver ' 1 + δtracer +δdriver ε Tracer Driver and Dynamic Analyses of CLP(FD) 23 6b. cstr: when port=post do current(chrono,cident,cinternal). tree: when port in [failure,backTo, choicePoint,solution] do current(chrono,node,port). 7b. newvar: when port=newVariable do current(chrono, vident, vname). dom: when port in [choicePoint,backTo,solution] do current(chrono,node,port,named_vars,full_dom). 8b. propag1: when port=reduce do current(chrono). 9b. propag2: when port=awake do current(chrono). Fig. 15. Event patterns used to assess the trace generation and the communication overhead The measured Rdriver for pattern 5a is Rdriver = 0.95 + For pattern 5a, the average time per event taken by the core tracer mechanism and the tracer driver (δtracer The Rdriver 100ns ε . + δdriver ) can therefore be approximated to 100ns. overhead could thus be made predictable. For a given program, it ε, the average time of execution per event. For δtracer + δdriver can be computed for each pattern. We have is easy to automatically measure a library of patterns shown above that the overhead of the simultaneous search for dierent patterns can be over approximated by the sum of all the overheads. Our environment could therefore provide estimation mechanisms. When δtracer + δdriver ε would be too small compared to the user would be warned that the overhead may become large. 7.6 Communication Overhead The measure of Tgcom ' Ttracer + ∆driver + ∆gcom is the execution time of the program run by the tracer. A new set of patterns are used so that some events match the patterns, the requested attributes of the matched events are generated and sent to a degenerated version of the mediator: a C program that simply reads the trace data on its standard input. We show the result of program golomb(8) which has a median number of events and has a median ε. The patterns are listed in Figure 15. Pattern 6b, composed of two basic patterns, allows a bare search tree to be constructed, as shown by most debugging tools. Pattern 7b (two basic patterns) allows the display of 3D views of variable updates as shown in Figure 8. Pattern 8b and pattern 9b provide two dierent execution details to decorate search trees. Depending on the tool settings, three dierent visual clues can be displayed. One is shown in Figure 8, Section 3.4. Table 2 gives the results for the above patterns and some of their combinations. All combinations correspond to existing tools. For example, combining or/and 4b 6b with 8b allows a Christmas tree as shown in Figure 8 to be constructed with L. Langevine and M. Ducassé 24 Program: golomb(8) Patterns ε = 45ns Tprog=1.73s Matched events Elapsed time (s) Rdriver Rgcom (106 ) XML Trace size (Mbytes) 6b 7b 8b 9b 0.36 0.13 5.04 14.58 21 111 141 394 4.50 16.17 33.57 89.40 1.03 1.02 1.14 1.32 2.6 9.35 19.40 51.68 (6|7)b (6|8)b (6|9)b (6|8|9)b (6|7|8|9)b 0.36 5.40 14.94 19.97 19.97 124 162 415 556 660 17.47 36.08 92.71 122.72 136.80 1.04 1.15 1.33 1.44 1.44 10.09 20.85 53.59 70.93 79.07 default trace 38.36 7,910 393.08 1.96 227.21 Table 2. Cost of the trace generation and communication for program golomb(8) two dierent parametrization. The 2 rd match the pattern. The 3 th is sent to the tool. The 4 the ratio Rgcom = Rdriver , nd column gives the number of events which column gives the size of the resulting XML trace as it 4 column gives the elapsed time . The 5 recomputed for each pattern. The th 6 th column gives column gives the ratio Tgcom Tprog . Filtered trace is more ecient and more accurate than default trace. The last line gives results for the default trace. On the one hand, the default trace contains twice as many events as the trace generated by pattern (6|7|8|9)b; it also contains more attributes than requested by the pattern; as a result, its size is ten times larger and its Rgcom overhead is three times larger. On the other hand, the default trace does not contain all the attributes. In that particular case, some relevant attributes are missing in the default trace while they are present in the trace generated by pattern (6|7|8|9)b. These attributes can be reconstructed by the analysis module, but this requires further computation and memory resources. As a consequence, the tracer driver approach that we propose is more ecient than sending over a default trace, even to construct sophisticated graphical views. The accuracy and the lower volume of the trace ease its post-processing by the debugging tools. Answering queries is more ecient than displaying traces. larger than 4 Rdriver , from 2.6 to 79.07 Rgcom is always much in our example. Therefore, queries using Here system and user time are not sucient because two processes are involved. Tprog has been re-measured in the same conditions. Tracer Driver and Dynamic Analyses of CLP(FD) 25 patterns that drastically lter the trace have signicantly better response time than queries that rst display the trace before analyzing it. When debugging, programmers often know what they want to check. In that case they are able to specify queries that demand a simple answer. In such a case our approach is signicantly better than systematically sending the whole trace information to an analyzer. No need to restrict the trace information a priori. Many tracers restrict the trace information a priori in order to reduce the volume of trace sent to an analyzer. This restricts the possibilities of the dynamic analyses without preventing the big size and time overhead as shown above with the default trace which does not contain important information while being huge. With our approach, trace information which is not requested does not cost much, therefore our trace model can aord to be very rich. This makes it easier to add new dynamic analyses. Performance is comparable to the state-of-the-practice. Rgcom varies from 2.6 to 79.07. To give a comparison, the Mercury tracer of Somogyi and Henderson (Somogyi and Henderson 1999) is regularly used by Mercury developers. For executions of size equivalent to those of our measurements, the Mercury tracer overhead has been measured from 2 to 15, with an average of 7 (Jahier and Ducassé 2002). Hence the ratios for patterns 6b, 7b and 6|7b are quite similar to the state-of-the-practice debuggers. The other patterns show an overhead that can discourage interactive usage. However, these patterns are thought of more for monitoring than debugging when the interaction does not have to be done in real time. Note, furthermore, that for the measured programs, the absolute response time is still on the range of two minutes for the worst case. When debugging, this is still acceptable, especially if one considers that in the relevant cases the alternatives to the uses of tools like ours are likely to be much worse. Our approach therefore allows to have the tracer present but idle by default. When a problem is encountered, simple queries can be set to localize roughly the source of the problem. Then, more costly patterns can be activated on smaller parts of the program. This is similar to what experienced programmers do. The dierence with our approach is that they do not have to either change tools, or reset the parametrization of the debugger. 8 Related Work Kraut (Bruegge and Hibbard 1983) implements a nite state machine to nd sequences of execution events that satisfy some patterns, called path rules. Several patterns are allowed and they can be enabled or disabled during the execution, using a labeling policy. Specied actions are triggered when a rule is satised but they are limited to some debugger primitives, such as a message display or incrementing a counter. The main interest of this tool is to abstract the trace and to allow the L. Langevine and M. Ducassé 26 easy development of monitors. The trace analysis is necessarily synchronous and does not benet from the power of a complete programming language. Reiss and Renieris (Reiss and Renieris 2001) have an approach similar to ours. They also structure their dynamic analyses into three dierent modules: 1) extraction of trace, 2) compaction and ltering and 3) visualization. They provide a number of interesting compaction functions which should be integrated in a further version of our system. They, however, rst dump the whole trace information in les before any ltering is processed. With our tracer driver, ltering is done on the y, and Section 7 has shown that this is much more ecient than rst storing in les. Their approach, however, is able to deal with partially ordered execution threads; adapting our framework to languages with partially ordered threads would require some technical work. Coca (Ducassé 1999b) and Opium (Ducassé 1999b) provide a trace query mechanism, respectively for C and Prolog. This mechanism is synchronous and does not allow concurrent analyses. It can be easily emulated with our tracer driver and an analyzer mediator written in Prolog. Hy + (Consens et al. 1994) writes the trace into a real relational database to query it with SQL. This is even slower than writing the trace simply into a le. However, when on the y performance is not an issue, for example for post mortem analysis, this is a very powerful and elegant solution which is straightforward to connect to a tracer with our tracer driver. Dalek (Olsson et al. 1990) is a powerful extension of gdb. It allows users to as- sociate sequences of execution events to specic synchronous handlers written in a dedicated imperative language. This language includes primitives to retrieve additional trace data and to synchronize the execution. The management of handlers is not incremental. A key feature of Dalek, especially useful in an imperative language, is the explicit queue of events that stores the achieved execution events. The user can explicitly remove events from this queue and add higher-level events. This approach requires an expensive storage of a part of the trace but enables both monitoring, debugging and proling of programs. In EBBA (Bates 1995), expected program behaviors are modeled as relationships between execution events. Those models are then compared to the actual behavior during execution. EBBA tries to recognize relevant sequences of events and to check some constraints about such sequences. A kind of automaton is built to nd instantiations of the models. The events are rst generated by the tracer before being ltered according to the automata. Our approach allows ltering execution events directly inside the tracer, which is more ecient. Nevertheless, EBBA recognizes sequences of events whereas we lter one event at time. Our approach could be used upstream of the sequence recognition. The incrementality of the event patterns could be used to adapt the relevant events to the states of the automata. UFO (Auguston et al. 2002) oers a more powerful language to specify patterns and monitors than EBBA. The patterns can involve several events, not necessarily consecutive. In our framework, the monitors have to be implemented in the analyzer with a general programming language. A further extension should allow at least the implementation of monitors in the trace driver to improve eciency. UFO, however, Tracer Driver and Dynamic Analyses of CLP(FD) 27 does not allow the same exibility as our tracer driver, and is heavier to use for interactive debugging. So far, our framework applies only to a single execution and does not easily scale to compare numerous executions as is done in batch mode by (Jones et al. 2002). It seems, however, possible to extend our framework so that two executions can be run in parallel with two tracer drivers. This would allow the implementation of the debugging analyses of (Zeller and Hildebrandt 2002) and (Sosic and Abramson 1997) which compare two executions at a given moment. For some applications, it is important to be able to rewind the execution. The necessary mechanisms are orthogonal to the ones presented here, and can be merged with them. Interested readers are referred to the Mercury mechanisms (MacLarty and Somogyi 2006) or the survey of (Ronsse et al. 2000). The Ilog Christmas Tree (Bracchi et al. 2001) is built by processing an XML trace produced by the debugger. This tracer is generic: it can be specialized to feed a specic tool. This specialization requires, however, a good understanding of the solver behavior and cannot be modied during the execution. Moreover, the amount of data available at each event is very limited compared to the full state our approach allows. For instance, the set of constraints and variables cannot be inspected. A debugging library for SICStus Prolog has been implemented (Ågren et al. 2002; Hanák et al. 2004). Its main quality is the explanations it provides about events that narrow domains. This helpful information needs a dicult and costly instrumentation of SICStus constraints: only a few ones have actually been instrumented. No performance results are available. Some tuning of the trace display is possible but the tracer is based on a complete storage of the trace and a postmortem investigation: this is impractical with real-sized executions. The lazy generation of the trace our tracer driver enables leads to the same kind of trace data in a more ecient and practical way. Some C(L)P debugging tools enable users to interact with the execution states. User of Oz Explorer (Schulte 1997) can act on the current state of the execution to drive the search-tree exploration. Users of CLPGUI (Fages 2002) can add new constraints on a partial solution. They can recompute a former state in both systems. Those features are helpful. Our approach is complementary, it addresses the communication from the traced execution to the debugging tools. 9 Conclusion In this paper we have presented a tracer driver which, with limited development eorts, provides a powerful front-end for complex debugging and monitoring tools based on trace data. We have dened an expressive language of event patterns where relevant events are described by logical formulæ involving most of the data the tracer can access. Specic primitives enable the retrieval of large pieces of data on demand and the adaptation of the event patterns to the evolving needs of trace analyzers. Experiments for CLP(FD) have shown that the overhead of the core tracer mech- L. Langevine and M. Ducassé 28 anisms is small, therefore the core tracer can be permanently activated; the tracer driver overhead is acceptable; there is no overhead in the ltering mechanisms when searching simultaneously for several patterns; the tracer driver overhead is predictable for given patterns; the tracer driver approach that we propose is more ecient than sending over a default trace, even to construct sophisticated graphical views; answering queries is orders of magnitude more ecient than displaying traces; there is no need to restrict the trace information a priori; last but not least, the performance of our tool is comparable to the state-of-the-practice while being more powerful and more generic. Traditionally, tracer designers decide on a static basis what the observed events should be. As a result, compromises regarding the amount of information to trace are made once for all or, at best, before each execution. With our approach the trace contents can be much richer because only what is needed is retrieved. Hence there is less chance that important information is missing. The tracer driver overhead is inversely proportional to the average time between two traced events. Whereas the principles of the tracer driver are independent of the traced programming language, it is best suited for high-level languages, such as constraint logic programming, where each traced execution event encompasses numerous low-level execution steps. The current C(L)P environments do not provide all the useful dynamic analysis tools. They can signicantly benet from our tracer driver which enables dynamic analyses to be integrated at a very low cost. Acknowledgment The authors thank Pierre Deransart and their OADymPPaC part- ners for fruitful discussions, as well as Guillaume Arnaud for the screenshots of his Pavot tool and his careful beta-testing of Codeine. Ludovic Langevine is indebted to Mats Carlsson and the SICStus team at Uppsala who helped to implement a preliminary tracer driver prototype connected to the FDBG SICStus Prolog tracer (Ågren et al. 2002). It contributed to the validation of the principles described in this article. References Ågren, M., Szeredi, T., Beldiceanu, N., and Carlsson, M. 2002. explaining execution of clp(fd) programs. In Environments'02, Tracing and Proc. of Workshop on Logic Programming A. Tessier, Ed. Computer Research Repository cs.SE/0207047. Auguston, M., Jeffery, C., and Underwood, S. 2002. A framework for automatic debugging. In Proceedings fo the 17th International Conference on Automated Software Engineering (ASE'02), W. Emmerich and D. Wile, Eds. IEEE Press, 217222. Ball, T. 1999. The concept of dynamic analysis. In ESEC / SIGSOFT FSE, O. Nierstrasz and M. Lemoine, Eds. Lecture Notes in Computer Science, vol. 1687. Springer, 216231. Bates, P. C. 1995. models of behavior. Debugging heterogeneous distributed systems using event-based ACM Transactions on Computer Systems 13, 1 (Feb.), 131. Bracchi, C., Gefflot, C., and Paulin, F. 2001. Combining propagation information and search-tree visualization using opl studio. In Proceedings of WLPE'01, A. Kusalik, Ed. Computer Research Repository cs.PL/0111040, Cyprus, 2739. Bruegge, B. and Hibbard, P. 1983. Generalized path expressions: A high-level debug- ging mechanism. The Journal of Systems and Software 3, 265276. Elsevier. Tracer Driver and Dynamic Analyses of CLP(FD) Carro, M. and Hermenegildo, M. 2000. The VIFID/TRIFID tool. 29 See Deransart et al. (2000), Chapter 10. Consens, M., Hasan, M., and Mendelzon, A. 1994. tributed event traces with Hy+. In ference, Visualizing and querying dis- Applications of Databases, First International Con- W. Litwin and T. Risch, Eds. Springer, Lecture Notes in Computer Science, Vol. 819, 123141. Denmat, T., Ducassé, M., and Ridoux, O. 2005. Data mining and cross-checking of execution traces. a re-interpretation of Jones, Harrold and Stasko test information Proceedings of the 20th IEEE/ACM International Conference on Automated Software Engineering, T. Ellman and A. Zisman, Eds. ACM Press. Deransart, P., Hermenegildo, M., and Maluszynski, J., Eds. 2000. Analysis and visualization tools for constraint programming. Lecture Notes in Computer Science, vol. visualization. In 1870. Springer-Verlag. Diaz, D. 2003. Gnu Prolog, a free Prolog compiler with constraint solving over nite domains. http://gprolog.sourceforge.net/ Distributed under the GNU license. Ducassé, M. 1999a. Coca: An automated debugger for C. International Conference on Software Engineering. Ducassé, M. 1999b. In Proceedings of the 21st ACM Press, 504513. Opium: An extendable trace analyser for Prolog. of Logic programming 39, 177223. The Journal Special issue on Synthesis, Transformation and Analysis of Logic Programs, A. Bossi and Y. Deville (eds). Fages, F. 2002. Clpgui: a generic graphical user interface for constraint logic programming over nite domains. In Proceedings of WLPE'02, A. Tessier, Ed. Computer Research Repository cs.SE/0207048, Copenhagen. Gent, I. and Walsh, T. 1999. CSPLib: a benchmark library for constraints. Tech. rep., Technical report APES-09-1999. Available from http://csplib.cs.strath.ac.uk/. A shorter version appears in the Proceedings of CP-99. Goualard, F. and Benhamou, F. 2000. Debugging constraint programs by store in- spection. See Deransart et al. (2000), Chapter 11. Hanák, D., Szeredi, T., and Szeredi, P. 2004. FDBG, the CLP(FD) debugger library of SICStus Prolog. In Proc. of Iclp'04, B. Demoen and V. Lifschitz, Eds. Poster. LNCS 3132. Harrold, M. J., Rothermel, G., Wu, R., and Yi, L. 1998. An empirical investigation of program spectra. In Proceedings of the 1998 ACM SIGPLAN-SIGSOFT workshop on Program analysis for software tools and engineering. ACM Press, 8390. Jahier, E. and Ducassé, M. 2002. Generic program monitoring by trace analysis. Theory and Practice of Logic Programming 2, 4-5 (July-September), 611643. Jeffery, C. and Griswold, R. 1994. A framework for execution monitoring in icon. Software-Practice and Experience 24, 11 (November), 10251049. Jones, J. A. and Harrold, M. J. 2005. Empirical evaluation of the Tarantula auto- Proceedings of the 20th IEEE/ACM international Conference on Automated Software Engineering. ACM Press, 273282. matic fault-localization technique. In Jones, J. A., Harrold, M. J., and Stasko, J. 2002. Visualization of test information to assist fault localization. In Engineering. Proceedings of the 24th International Conference on Software ACM Press, 467477. Langevine, L., Deransart, P., and Ducassé, M. 2004. A generic trace schema for the portability of CP(FD) debugging tools. In Recent advances in Constraint Programming, J. Vancza, K. Apt, F. Fages, F. Rossi, and P. Szeredi, Eds. Springer-Verlag, Lecture Notes in Articial Intelligence 3010. Langevine, L., Ducassé, M., and Deransart, P. 2003. A propagation tracer for Gnu-Prolog: from formal denition to ecient implementation. In Proc. of the 19th L. Langevine and M. Ducassé 30 Int. Conf. in Logic Programming, C. Palamidessi, Ed. Springer-Verlag, Lecture Notes in Computer Science 2916. MacLarty, I. and Somogyi, Z. 2006. Controlling search space materialization in a prac- tical declarative debugger. In PADL, P. V. Hentenryck, Ed. Lecture Notes in Computer Science, vol. 3819. Springer, 3144. MacLarty, I., Somogyi, Z., and Brown, M. 2005. Divide-and-query and subterm de- pendency tracking in the Mercury declarative debugger. In AADEBUG'05: Proceedings of the sixth international symposium on Automated analysis-driven debugging. ACM Press, New York, NY, USA, 5968. Proceedings of the First International Conference on Principles and Practice of Constraint Programming, U. Montanari and Meier, M. 1995. Debugging constraint programs. In F. Rossi, Eds. Lecture Notes in Computer Science, vol. 976. Springer-Verlag, 204221. Olsson, R., Crawford, R., and Ho, W. 1990. Dalek: A GNU, improved programmable debugger. In Proceedings of the Summer 1990 USENIX Conference: June 1115, 1990. 221232. Reiss, S. and Renieris, M. 2001. Encoding program executions. In 23rd International Conference on Software Engineering, Proceedings of the M.-J. Harrold and W. Schäfer, Eds. IEEE Press, 221230. Ronsse, M., Bosschere, K. D., and Chassin de Kergommeaux, J. 2000. Execution In Proceedings of the International Workshop on Automated Debugging (AADEBUG2000), M. Ducassé, Ed. http://xxx.lanl.gov/abs/cs.SE/0011006. Schulte, C. 1997. Oz explorer: a visual constraint programming tool. In Proc. of the 14th Int. Conf. on Logic Programming, L. Naish, Ed. MIT Press, 286300. Shapiro, E. 1983. Algorithmic Program Debugging. MIT Press, Cambridge, MA. ISBN replay and debugging. 0-262-19218-7. Simonis, H. and Aggoun, A. 2000. Search-tree visualisation. See Deransart et al. (2000), Chapter 7. Somogyi, Z. and Henderson, F. 1999. cury debugger. vironments. In The implementation technology of the Mer- Proceedings of the Tenth Workshop on Logic Programming En- Vol. 30(4). Elevier, Electronic Notes in Theoretical Computer Science. http://www.elsevier.nl/cas/tree/store/tcs/free/entcs/store/tcs30/cover.sub.sht. Sosic, R. and Abramson, D. 1997. and Experience 27, Guard: A relative debugger. Software - Practice 2, 185206. Wilhelm, R. and Maurer, D. 1995. Compiler design. Addison-Wesley. ISBN: 0-201- 42290-5. Zeller, A. and Hildebrandt, R. 2002. Simplifying and isolating failure-inducing input. IEEE Transactions on Software Engineering 28, 2, 183200.
6cs.PL
IDENTIFICATION OF DYNAMIC FUNCTIONAL BRAIN NETWORK STATES THROUGH TENSOR DECOMPOSITION arXiv:1410.0446v1 [cs.NE] 2 Oct 2014 Arash Golibagh Mahyari, Selin Aviyente [email protected], [email protected] Department of Electrical and Computer Engineering Michigan State University East Lansing, MI 48823 USA ABSTRACT With the advances in high resolution neuroimaging, there has been a growing interest in the detection of functional brain connectivity. Complex network theory has been proposed as an attractive mathematical representation of functional brain networks. However, most of the current studies of functional brain networks have focused on the computation of graph theoretic indices for static networks, i.e. long-time averages of connectivity networks. It is well-known that functional connectivity is a dynamic process and the construction and reorganization of the networks is key to understanding human cognition. Therefore, there is a growing need to track dynamic functional brain networks and identify time intervals over which the network is quasi-stationary. In this paper, we present a tensor decomposition based method to identify temporally invariant ’network states’ and find a common topographic representation for each state. The proposed methods are applied to electroencephalogram (EEG) data during the study of error-related negativity (ERN). Index Terms— Graphs, Dynamic Networks, Tensor Decomposition, Electroencephalography. 1. INTRODUCTION With the advance in noninvasive imaging modalities such as fMRI, EEG, and MEG, it is important to develop computational methods capable of giving a succinct description of the functional brain networks [1]. Functional connectivity describes the coordinated activation of segregated groups of neurons. Traditionally, functional connectivity has been quantified through linear [2] and nonlinear measures [3]. Synchronization of neuronal oscillations has been suggested as one plausible mechanism in the interaction of spatially distributed neural populations and has been quantified using phase synchrony [4]. Although phase synchrony is successful at quantifying pairwise interactions [5], it cannot completely This work was in part supported by the National Science Foundation under Grant No. CCF-1218377. describe the complex relationship between function and organization of the brain. Recently, research in the area of complex networks has led to fundamental insights into the organization of the healthy and diseased brain [5, 6]. However, the current studies are limited to the analysis of static brain networks obtained through averaging long-term functional connectivity and thus, neglect possible time-varying properties of the topologies. There is growing evidence that functional networks dynamically reorganize and coordinate on millisecond scale for the execution of mental processes [7]. For this reason, there has been an interest in characterizing the dynamics of functional networks using high temporal resolution EEG recordings. The early work in this area was an extension of static network analysis to the dynamic case by extracting graph theoretic features from graphs across time and tracking the evolution of these parameters [8]. However, these approaches lose the spatial information provided by the graphs and cannot identify which parts of the brain contributed to the observed changes in the network. More recently, a ”network state” framework has been proposed [9, 7], where each state is defined as periods of time during which the network topology is quasi-stationary. In this paper, we adopt this framework for tracking the topology of brain networks across time and representing each state with a common topographic map. The current work differs from the existing approaches in a couple of ways. First, we take into account the full structure of the network at each time point. Second, we consider extracting network states common across time and subjects unlike current work which considers individual subjects. Finally, the current work offers a compressed spatial representation of each network state through tensor-tensor projection unlike current approaches which use averaging. Tensor-to-tensor projection proposed in this paper projects the information across subjects and time into a lower dimensional ’signal’ subspace whereas averaging assigns equal weights to all subjects. It is also important to note that the proposed framework is closely tied to dynamic network tracking. The most common approaches to network tracking have been to identify anomalies using subspace projection methods such as in [10, 11] or through sliding window estimates with time independent or dependent weighting factors [12, 13]. More recently, adaptive evolutionary clustering [14] was proposed to track the cluster changes over time. However, all of these methods either detect time points where events of interest happen, or find different clustering structures at each time point. In this paper, we propose a comprehensive framework that first identifies time intervals during which the network topology is stationary and then summarizes each interval by a single lower dimensional network through tensor-tensor projection. 2. BACKGROUND 2.1. Time-Varying Network Construction The time-varying functional brain networks are constructed from multichannel EEG data with the nodes corresponding to the different brain regions and the edges to the connectivity between these regions. In this paper, we quantify the connectivity using a recently introduced phase synchrony measure based on RID-Rihaczek distribution [15]. The first step in quantifying phase synchrony is to estimate the time and hfrequency i dependent phase, Φi (t, ω), of a signal, si , arg Ci (t,ω) |Ci (t,ω)| 1 Therefore, {G(t)}t=1,2,...,T is a time series of N × N weighted and undirected graphs corresponding to the functional connectivity network at time t for a fixed frequency band, where T is the total number of time points and N is the number of nodes within the network. 2.2. Tensor Subspace Analysis Linear data models such as Principal Component Analysis (PCA) and Independent Component Analysis (ICA) are widely used for the decomposition of matrices. Depending on the criteria, different types of basis vectors are extracted and appropriate lower dimensional features are determined through projection. Multiway data analysis extends these linear methods to capture multilinear structures and underlying correlations in higher-order datasets, also known as tensors. Some exemplary methods include PARAFAC, Tucker decomposition, and Higher-Order Singular Value Decomposition (HOSVD) [16, 17]. The Tucker decomposition is a higher order generalization of Singular Value Decomposition (SVD) [16]. Let X ∈ Rm1 ×m2 ...×md be a d -mode array, then its Tucker decomposition can be expressed as: , where Ci (t, ω) is the complex RID-Rihaczek distribution : Ci (t, ω) = Z Z  θτ (θτ )2 exp(j )Ai (θ, τ )e−j(θt+τ ω) dτ dθ exp − σ 2 | {z } | {z }  Choi-Williams kernel Rihaczek kernel (1) R and Ai (θ, τ ) = si (u + τ2 )s∗i (u − τ2 )ejθu du is the ambiguity function of the signal s i . The phase synchrony between nodes i and j at time t and frequency ω is computed using Phase Locking Value (PLV): P LVij (t, ω) = L  1 X exp jΦkij (t, ω) L where L is the number of trials and Φkij (t, ω) = |Φki (t, ω) − Φkj (t, ω)| is the phase difference between the two channels for the k th trial. Once the pairwise synchrony values are computed at each time and frequency point, we can construct a time-varying graph {G(t)}t=1,2,...,T with ωb 1 X P LVij (t, ω) Ω ω=ω X = C ×1 U(1)×2 U(2) ... ×d U(d)+ E (1) (2) (d) Ci1 ,i2 ,...,id ui1 ◦ ui2 ... ◦ uid + Ei1 ,i2 ,...,id (4) where C ∈ Rr1 ×r2 ...×rd is the core tensor, and U(1) ∈ Rm1 ×r1 , U(2) ∈ Rm2 ×r2 , ..., U(d) ∈ Rmd ×rd , where r1 ≤ m1 , r2 ≤ m2 , ..., rd ≤ md , are the projection matrices whose columns are orthogonal. E ∈ Rm1 ×m2 ...×md is the residual error, and ×k is the product of a tensor and a matrix along mode-k. Reconstruction of the original tensor, X̃ ∈ Rm1 ×m2 ...×md , from a lower dimensional representation is obtained as: i1 ,i2 ,...,id (2) k=1 Gij (t) = = P (3) a where Gij (t) ∈ [0, 1] represents the connectivity strength between the nodes i and j within the frequency band of interest, [ωa , ωb ], and Ω is the number of frequency bins in that band. 1 The details of the RID-Rihaczek distribution and the corresponding synchrony measure are given in [15].       † † † X˜ = X ×1 U(1) U(1) ×2 U(2) U(2) ... ×d U(d) U(d) (5) where † is the transpose of the matrix. 3. THE PROPOSED METHOD 3.1. Temporal Tracking for Network State Identification In the proposed work, the time-varying functional connectivity graphs across subjects are considered as a 4-mode tensor G ∈ RN ×N ×T ×S constructed as channel × channel × time × subject, with N being the number of channels, T the total number of time points and S the number of subjects. The Tucker decomposition of this connectivity tensor yields: G = C ×1 U(1) ×2 U(2) ×3 U(3) ×4 U(4) + E (6) where C ∈ RN ×N ×T ×S is the core tensor, and E N ×N ×T ×S is the residual error. To obtain an approximation of G, G̃ ∈ RN ×N ×T ×S , we first consider the full Tucker decomposition with U(1) ∈ RN ×N , U(2) ∈ RN ×N , U(3) ∈ RT ×T , and U(4) ∈ RS×S . The singular values along each mode are ordered by fixing the index of all of the other modes to 1. Since first singular values along each mode represent the largest variance of the data along that mode, we choose that to order the remaining mode. To get the approximation tensor, the appropriate number of singular vectors along first and second modes N̄ is defined as N̄ = jk , where jk is the highest index for which |Cjk ,1,1,1 | ≥ 0. Similarly, the number of singular vectors along the fourth mode is S̄ = sk , where sk is the highest index for which |C1,1,1,sk | ≥ 0. The time mode is not projected to a lower dimensional space since all time points are necessary to identify the exact boundaries of the network states. The lower dimensional projection matrices are de(2) (2) (2) (1) (1) (1) fined as: Ũ(1) = [u1 u2 ...uÑ ], Ũ(2) = [u1 u2 ...uÑ ], (4) (4) (4) Ũ(4) = [u1 u2 ...uS̃ ]. The reconstructed tensor G̃ ∈ RN ×N ×T ×S is obtained as:     † † G̃ = G ×1 Ũ(1) Ũ(1) ×2 Ũ(2) Ũ(2)     † † ×3 Ũ(3) Ũ(3) ×4 Ũ(4) Ũ(4) . (7) The 4-mode approximation tensor G̃ can be written as a sequence of 3-mode tensors G̃t ; t = 1, 2, ..., T . To detect the boundaries of network states, we propose a new temporal clustering algorithm. Unlike regular data clustering, the proposed method considers both the similarity of the lower dimensional representation of the networks as well as their closeness in time. The similarity of two networks at time t1 and t2 is quantified through a cosine similarity metric between G̃t1 and G̃t2 as follows: ∆(t1 , t2 ) = hG̃t1 ,G̃t2 i kG̃t1 kkG̃t2 k ; t1 , t2 = 1, 2, ..., T (8) where ha, bi is the inner product of a and b, and k a k is the Frobenius Norm. Similarly, the temporal closeness between −(t1 −t2 )2 two graphs is quantified as Θ(t1 , t2 ) = e 2σ2 ; t1 , t2 = 1, 2, ..., T , where σ is a parameter which determines the weighting for different time separations, and depends on the sampling frequency. The combined similarity matrix is defined as: Ψ(t1 , t2 ) = λΘ(t1 , t2 ) + (1 − λ)∆(t1 , t2 ); t1 , t2 = 1, 2, ..., T (9) where λ ∈ (0, 1) determines the trade-off between tensor similarity and time proximity. This similarity matrix is input to a standard spectral clustering algorithm combined with k-means to identify the boundaries of the network states [18]. 3.2. Topographic Compression for Network State Representation Once the time boundaries of the different network states are identified, each state has to be summarized with a single topographic map. Previously, this was commonly addressed by averaging the edges over the time interval [19]. This method has the drawback of emphasizing all of the edges equally and resulting in very dense network representations. For a given time interval (T1 , T2 ) and the 3-mode tensor sequence corresponding to this interval GT1 , GT1 +1 , GT1 +2 , ..., GT2 , the goal is to extract the topographic map Ĝ ∈ RN ×N which best represents that network state. The 3-mode tensors corresponding to the time interval (T1 , T2 ) can be rewritten as a 4-mode tensor by taking the time modality into account G ′ ∈ RN ×N ×(T2 −T1 +1)×S , and decomposed using (1) (2) the full Tucker decomposition, U′ ∈ RN ×N , U′ ∈ (3) (4) RN ×N , U′ ∈ R(T2 −T1 +1)×(T2 −T1 +1) , U′ ∈ RS×S , similar to Equation 6. In order to summarize the subject information to find a general unique model which fits all subjects, the 4-mode tensor G ′ ∈ RN ×N ×(T2 −T1 +1)×S is projected by the singular (4) vector u′l corresponding to the lth largest singular value in this mode. Likewise, to summarize the time information of the resulting 3-mode tensor, it is projected to the singular vec(3) tor u′k corresponding to the k th largest singular value of the (3) (4) time mode, Ĝ = G ′ ×3 u′k ×4 u′l . The values of k and l are usually equal to 1 but may change depending on the data. 4. EXPERIMENTAL RESULTS 4.1. EEG Data The proposed framework is applied to a set of EEG data containing the error-related negativity (ERN) 2 . The ERN is a brain potential response that occurs following performance errors in a speeded reaction time task usually 25-75 ms after the response [20]. Previous work [21] indicates that there is increased coordination between the lateral prefrontal cortex (lPFC) and medial prefrontal cortex (mPFC) within the theta frequency band (4-8 Hz) and ERN time window (25-75 ms), supporting the idea that frontal and central electrodes are functionally integrated during error processing. EEG data from 62-channels was collected in accordance with the 10/20 system on a Neuroscan Synamps2 system (Neuroscan, Inc.). A speeded-response flanker task was employed, and response-locked averages were computed for each subject. All EEG epochs were converted to current source density (CSD) using published methods [22]. Data were averaged across trials (∼ 200 trials) for the purpose of ERN and timefrequency analysis.In this paper, we analyzed data from 91 subjects corresponding to the error responses. 2 We thank Dr. Edward Bernat from the University of Maryland for sharing his EEG dataset with us. 4.2. Network State Identification and Summarization The connectivity matrices are constructed by computing the pairwise average PLV between 62 channels in the theta frequency band for all time (2 seconds) and all subjects using Eq. 2. The time-varying graphs {G(t)}t=1,2,...,T for all subjects and all time will be treated as a 4-mode tensor, which is decomposed using Tucker decompostion. The approximation tensors G̃t ; t = 1, 2, ..., T with N̄ = 2, S̄ = 3 are used to obtain the 256 × 256 similarity matrix Ψ. The matrix Ψ is computed with λ = 0.4 and σ = 2500 as shown in Fig. 1. The values of λ and σ are empirically chosen to obtain the best separation between clusters. Once the matrix Ψ is ob- long-range relationships we are interested in. In order to show the most significant edges in the summarized graph, the edges with values in the top 1% are selected and plotted in Fig. 2. Most of the significant edges are in the frontal and central areas where node AF 8 acts as a hub, with the highest degree equal to 59. C1 has the second highest degree (53) followed by F P 2 (52). These nodes correspond to the right lateral prefrontal cortex (lPFC) and medial prefrontal cortex (mPFC) in accordance with previous findings which indicate increased synchronization between these regions during ERN [21]. 60 FP1 FPZ FP2 AF8 AF7 AF3 AFZ F7 F5 F8 F1 F3 FZ F2 F4 F6 FT8 FT7FC5FC3 FC1FCZFC2FC4FC6 T7 C5 C3 C1 CZ C2 C4 C6 T8 CPZCP2CP4CP6 TP7CP5CP3CP1 TP8 PZ P2 P4 P6 P1 P7 P5 P3 P8 PO3 POZPO4 PO6 PO5 PO8 PO7 O1 OZ O2 50 40 30 20 10 Fig. 2. The topographic map with the most signifcant edges of the compressed connectivity matrix for the network state (−132ms, 188ms). Fig. 1. Similarity matrix, Ψ, computed for 2 seconds of EEG data across 91 subjects in the theta frequency band. 5. CONCLUSIONS tained, the critical time points are detected using the spectral clustering with K = 5. The number of clusters K is selected based on the eigenspectrum of the similarity matrix [23]. The detected time intervals are (−1000, −703) ms, (−703, −132) ms, (−132, 188) ms,(188, 736) ms,(736, 1000) ms. As expected, the first two time intervals correspond to the prestimulus part where there is less change in the network configuration. The third time interval (−132ms, 188ms) is of particular interest since it includes the time interval right before a response is made as well as the ERN interval (0 − 150ms). This is extracted as a separate network state indicating a reorganization of the functional network configuration. Similarly, the time interval (188, 736)ms contains the P300 event which is expected to result in a distinct topographic map. In this paper, we will focus on extracting the common topographic map for the time interval (−132, 188) ms since it coincides with the ERN [21]. To obtain a single network representation for this time interval, we selected k = l = 2 instead of the singular vectors corresponding to the highest singular values since the projection to the subspace spanned by the largest singular value mostly contains edges between physically adjacent nodes. This is a side effect of volume conduction affecting PLV values and does not convey the actual In this paper, we proposed a tensor-based method for monitoring dynamic functional connectivity networks to identify quasi-stationary network states and represent the common topographic distribution of each state. Network states were identified computing a similarity metric which takes both the similarity of the reconstructed tensors and their proximity in time to obtain a partitioning of the dynamic networks into contiguous time intervals. After identifying the boundaries of the network states, a topographical map for each time interval was obtained by a tensor-tensor projection. The application of the proposed algorithm to ERN data yields time intervals that closely correspond to events of interest and topographic maps that are consistent with previous hypotheses with regard to error monitoring in the brain. The proposed method is time consuming for large data sets due to the complexity of Tucker decomposition. Future work will consider extensions of this framework to partitioning time and frequency dependent connectivity networks by considering higher order tensor representations. Moreover, the choice of optimal parameters (λ and σ) will be considered using cost functions such as modularity for evaluating the quality of the different partitions. 6. REFERENCES [1] B. He, T. Coleman, G.M. Genin, G. Glover, X. Hu, N. Johnson, T. Liu, S. Makeig, P. Sajda, and K. Ye, “Grand challenges in mapping the human brain: NSF workshop report,” IEEE transactions on Biomedical engineering, 2013. in press. [2] M. Rubinov and O. Sporns, “Complex network measures of brain connectivity: uses and interpretations,” Neuroimage, vol. 52, no. 3, pp. 1059–1069, 2010. [3] E. Pereda, R.Q. Quiroga, and J. Bhattacharya, “Nonlinear multivariate analysis of neurophysiological signals,” Progress in Neurobiology, vol. 77, no. 1-2, pp. 1–37, 2005. [4] F. Varela, J.P. Lachaux, E. Rodriguez, and J. Martinerie, “The brainweb: Phase synchronization and large-scale integration,” Nature Reviews Neuroscience, vol. 2, no. 4, pp. 229–239, 2001. [5] E. Bullmore and O. Sporns, “Complex brain networks: graph theoretical analysis of structural and functional systems,” Nature Reviews Neuroscience, vol. 10, no. 3, pp. 186–198, 2009. [6] C.J. Stam and J.C. Reijneveld, “Graph theoretical analysis of complex networks in the brain,” Nonlinear biomedical physics, vol. 1, no. 1, pp. 3, 2007. [7] S.I. Dimitriadis, N.A. Laskaris, and A. Tzelepi, “On the quantization of time-varying phase synchrony patterns into distinct functional connectivity microstates (fcµstates) in a multi-trial visual ERP paradigm,” Brain topography, pp. 1–13, 2013. [8] S.I. Dimitriadis, N.A. Laskaris, V. Tsirka, M. Vourkas, S. Micheloyannis, and S. Fotopoulos, “Tracking brain dynamics via time-dependent network analysis,” Journal of Neuroscience Methods, vol. 193, no. 1, pp. 145– 155, 2010. [9] R.F. Betzel, M.A. Erickson, M. Abell, B.F. O’Donnell, W.P. Hetrick, and O. Sporns, “Synchronization dynamics and evidence for a repertoire of network states in resting eeg,” Frontiers in computational neuroscience, vol. 6, 2012. [10] A. Lakhina, M. Crovella, and C. Diot, “Diagnosing network-wide traffic anomalies,” in ACM SIGCOMM Computer Communication Review. ACM, 2004, vol. 34, pp. 219–230. [11] B.A. Miller, M.S. Beard, and N.T. Bliss, “Eigenspace analysis for threat detection in social networks,” in 2011 Proceedings of the 14th International Conference on Information Fusion (FUSION), 2011, pp. 1–7. [12] H. Tong, S. Papadimitriou, P. S. Yu, and C. Faloutsos, “Proximity tracking on time-evolving bipartite graphs,” in 8th SIAM International Conference on Data Mining, Applied Mathematics. [13] K. Xu, M. Kliger, and A. Hero, “A shrinkage approach to tracking dynamic networks,” in Proceedings of the IEEE Statistical Signal Processing Workshop (SSP ’11), 2011, p. 517520. [14] K.S. Xu, M. Kliger, and A.O. Hero III, “Adaptive evolutionary clustering,” arXiv preprint arXiv:1104.1990, 2011. [15] S. Aviyente and A.Y. Mutlu, “A time-frequency-based approach to phase and phase synchrony estimation,” IEEE Transactions on Signal Processing, vol. 59, no. 7, pp. 3086–3098, 2011. [16] E. Acar and B. Yener, “Unsupervised multiway data analysis: A literature survey,” Knowledge and Data Engineering, IEEE Transactions on, vol. 21, no. 1, pp. 6– 20, 2009. [17] L. De Lathauwer, B. De Moor, and J. Vandewalle, “A multilinear singular value decomposition,” SIAM journal on Matrix Analysis and Applications, vol. 21, no. 4, pp. 1253–1278, 2000. [18] U. Von Luxburg, “A tutorial on spectral clustering,” Statistics and computing, vol. 17, no. 4, pp. 395–416, 2007. [19] A. G. Mahyari and S. Aviyente, “Two-dimensional svd for event detection in dynamic functional brain networks,” in Proceedings of IEEE Global Conference on Signal and Information Processing, 2013. [20] J. R. Hall, E. M. Bernat, and C. J. Patrick, “Externalizing psychopathology and the error-related negativity,” Psychological Science, vol. 18, no. 4, pp. 326–333, 2007. [21] J.F. Cavanagh, M.X. Cohen, and J.J.B. Allen, “Prelude to and resolution of an error: EEG phase synchrony reveals cognitive control dynamics during action monitoring,” The Journal of Neuroscience, vol. 29, no. 1, pp. 98–105, 2009. [22] J. Kayser and C.E. Tenke, “Principal components analysis of laplacian waveforms as a generic method for identifying ERP generator patterns: I. evaluation with auditory oddball tasks,” Clinical Neurophysiology, vol. 117, no. 2, pp. 348–368, 2006. [23] A.Y. Ng, M.I. Jordan, and Y. Weiss, “On spectral clustering: Analysis and an algorithm,” Advances in neural information processing systems, vol. 2, pp. 849–856, 2002.
9cs.NE